notes on the p-laplace equation (second edition)usually p 1. at the critical points ( r u = 0) the...

UNIVERSITY OF JYVASKYLADEPARTMENT OF MATHEMATICS

AND STATISTICS

REPORT 161

UNIVERSITAT JYVASKYLAINSTITUT FUR MATHEMATIK

UND STATISTIK

BERICHT 161

NOTES ON THE p-LAPLACE EQUATION(SECOND EDITION)

PETER LINDQVIST

JYVASKYLA2017

Editor: Pekka KoskelaDepartment of Mathematics and StatisticsP.O. Box 35 (MaD)FI–40014 University of JyvaskylaFinland

ISBN 978-951-39-7119-9 (print)ISBN 978-951-39-7120-5 (pdf)ISSN 1457-8905

Copyright c© 2017, Peter Lindqvistand University of Jyvaskyla

University Printing HouseJyvaskyla 2017

Contents

1. Introduction 1

2. The Dirichlet problem and weak solutions 5

3. Regularity theory 15

3.1. The case p > n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2. The case p = n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.3. The case 1 < p < n . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4. Differentiability 27

5. On p-superharmonic functions 35

5.1. Definition and examples . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.2. The obstacle problem and approximation . . . . . . . . . . . . . . . . 38

5.3. Infimal convolutions . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

5.4. The Poisson modification . . . . . . . . . . . . . . . . . . . . . . . . . 46

5.5. Summability of unbounded p-superharmonic functions . . . . . . . . . 47

5.6. About pointwise behaviour . . . . . . . . . . . . . . . . . . . . . . . . 49

5.7. Summability of the gradient . . . . . . . . . . . . . . . . . . . . . . . 51

6. Perron’s method 55

7. Some remarks in the complex plane 65

8. The infinity Laplacian 68

9. Viscosity solutions 76

10. Asymptotic mean values 86

11. Some open problems 93

12. Inequalities for vectors 95

First edition 2006. These notes are written up after my lectures at the SummerSchool in Jyvaskyla in August 2005. I am grateful to Xiao Zhong for his valuableassistance with the practical arrangements. Juan Manfredi has read the entire orig-inal manuscript and contributed with valuable comments and improvements. I alsowant to thank Karoliina Kilpelainen for the typsetting of my manuscript.

The most important partial differential equation of the second order is the cele-brated Laplace equation. This is the prototype for linear elliptic equations. It is lesswell-known that it also has a non-linear counterpart, the so-called p-Laplace equa-tion, depending on a parameter p. The p-Laplace equation has been much studiedduring the last fifty years and its theory is by now well developed. Some chal-lenging open problems remain. The p-Laplace equation is a degenerate or singularelliptic equation in divergence form. It deserves a treatise of its own, without anyextra complications and generalizations. This is my humble attempt to write sucha treatise. Perhaps the interested reader wants to consult the monograph NonlinearPotential Theory of Degenerate Elliptic Equations by J. Heinonen, T. Kilpelainenand O. Martio, when it comes to more advanced and general questions.

Second edition 2017. Two new chapters have been added, one about viscositysolutions and one about the Asymptotic Mean Value Property. Some misprints anddiscrepancies have been corrected. Yet, a good account of the implications fromStochastic Game Theory is a missing chapter. I thank Fredrik Hoeg, Erik Lindgren,and Eero Ruosteenoja for help with proofreading.

1. Introduction

The Laplace equation ∆u = 0 or

∂2u

∂x21

+∂2u

∂x22

+ · · ·+ ∂2u

∂x2n

= 0

is the Euler-Lagrange equation for the Dirichlet integral

D(u) =

∫

Ω

|∇u|2dx =

∫· · ·∫

Ω

[( ∂u∂x1

)2

+ · · ·+( ∂u∂xn

)2]dx1 . . . dxn .

If we change the square to a pth power, we have the integral

I(u) =

∫

Ω

|∇u|pdx =

∫· · ·∫

Ω

[( ∂u∂x1

)2

+ · · ·+( ∂u∂xn

)2] p

2

dx1 . . . dxn .

The corresponding Euler-Lagrange equation is

div(|∇u|p−2∇u) = 0 .

This is the p-Laplace equation and the p-Laplace operator is defined as

∆pu = div(|∇u|p−2∇u)

= |∇u|p−4|∇u|2∆u+ (p− 2)

n∑

i,j=1

∂u

∂xi

∂u

∂xj

∂2u

∂xi∂xj

.

Usually p ≥ 1. At the critical points (∇u = 0) the equation is degenerate for p > 2and singular for p < 2. The solutions are called p-harmonic functions.

There are several noteworthy values of p.

p = 1 ∆1u = div( ∇u|∇u|

)= −H ,

where H is the Mean Curvature Operator. In only two variables we have thefamiliar expression

H =u2yuxx − 2uxuyuxy + u2

xuyy

(u2x + u2

y)32

The formula ∆1ϕ(u) = ∆1u holds for general functions ϕ in one variable,indicating that solutions are determined by their level sets.

1

p = 2 We have the Laplace operator

∆2u = ∆u =n∑

i=1

∂2u

∂x2i

.

p = n The borderline case. When n is the number of independent variables,the integral

∫

Ω

|∇u|ndx =

∫· · ·∫

Ω

( ∂u∂x1

)2

+ · · ·+( ∂u∂xn

)2n2dx1 . . . dxn

is conformally invariant. The n-harmonic equation ∆nu = 0 in n variables istherefore invariant under Mobius transformations. For example, the coordinatefunctions of the inversion (a Mobius transformation)

y = a+x− a|x− a|2

are n-harmonic. The borderline case is important in the theory of quasicon-formal mappings.

p =∞ As p→∞ one encounters the equation ∆∞u = 0 or

n∑

i,j=1

∂u

∂xi

∂u

∂xj

∂2u

∂xi∂xj= 0 .

This is the infinity Laplace equation. It has applications for optimal Lipschitzextensions and has been used in image processing.

The case p > 2 is degenerate and the case 1 < p < 2 is singular 1. In the classicaltheory of the Laplace equation several main parts of mathematics are joined in afruitful way: Calculus of Variations, Partial Differential Equations, Potential Theory,Function Theory (Analytic Functions), not to mention Mathematical Physics and

1The terminology comes from fluid dynamics. An equation of the type

∂v

∂t= div(ρ(|∇v|)∇v)

is called singular, when it so happens that the diffusion ρ =∞ and degenerate when ρ = 0. Noticealso the expansion

ρ(|∇v|) = |∇v|p−2(c1 + c2|∇v|+ c2|∇v|2 · · ·

).

2

Calculus of Probability. This is the strength of the classical theory. It is veryremarkable that the p-Laplace equation occupies a similar position, when it comesto non-linear phenomena. Much of what is valid for the ordinary Laplace equationalso holds for the p-harmonic equation, except that the Principle of Superpositionis naturally lost. Even the Mean Value Formula holds, though only infinitesimallyin small balls. A non-linear potential theory has been created with all its requisites:p-superharmonic functions, Perron’s method, barriers, Wiener’s criterion and soon. In the complex plane a special structure related to quasiconformal mappingsappears. Last but not least, the p-harmonic operator appears in physics: rheology,glacelogy, radiation of heat, plastic moulding etc. Some advances indicate that eventhe Brownian motion has its counterpart and a mathematical game ”Tug of War”leads to the case p =∞. For finite p a sophisticated stochastic game is a counterpartto the Brownian motion.

Needless to say, the equation ∆pu = 0 has numerous generalizations. For example,one may start with variational integrals like

∫|∇u|pωdx ,

∫|∇u(x)|p(x)dx ,

∫ ∣∣∣∣n∑

i,j=1

aij∂u

∂xi

∂u

∂xj

∣∣∣∣p2

dx ,

∫ (∣∣∣∣∂u

∂x1

∣∣∣∣p

+ · · ·+∣∣∣∣∂u

∂xn

∣∣∣∣p )dx

and so on. The non-linear potential theory has been developed for rather generalequations

div Ap(x,∇u) = 0 .

However, one may interpret Polya’s Paradox2 as indicating that the special caseis often more difficult than the general case. In these lecture notes I resist thetemptation of including any generalizations. Thus I stick to the pregnant formulation∆pu = 0.

The p-harmonic operator appears in many contexts. A short list is the following.

• The non-linear eigenvalue problem

∆pu+ λ|u|p−2u = 0

• The p-Poisson equation∆pu = f(x)

2”The more ambitious plan may have more chances of success”, G.Polya, How to Solve It,Princeton University Press, 1945.

3

• Equations like∆pu+ |u|αu = 0 ,

which are interesting when the exponent α is ”critical”.

• Parabolic equations like∂v

∂t= ∆pv ,

where v = v(x, t) = v(x1, . . . , xn, t)

• So-called p-harmonic maps u = (u1, u2, . . . , un) minimizing the ”p-energy”

∫|Du|pdx =

∫ ∑

i,j

(∂uj∂xi

)2 p2dx ,

perhaps with some constraints. A system of equations appears.

• The fractional p-Laplace operator

(−∆p)su(x) = cp,s

∫

Ω

|u(y)− u(x)|p−2(u(y)− u(x))

|x− y|n+spdy.

These additional topics are very interesting but cannot be treated here.

The reader is supposed to know some basic facts about Lp-spaces and Sobolevspaces, especially the first order spaces W 1,p(Ω) and W 1,p

0 (Ω). The norm is

‖u‖W 1,p(Ω) =

∫

Ω

|u|pdx+

∫

Ω

|∇u|pdx 1

p

.

Ω is always a domain (= an open connected set) in the n-dimensional Euclidean spaceRn. Text books devoted entirely to Sobolev spaces are no good for our purpose.Instead we refer to [GT, Chapter 7], which is much to the point, [G, Chapter 3]or [EG]. The reader with an apt to estimates will enjoy the chapter ”AuxiliaryPropositions” in the classical book [LU].

4

2. The Dirichlet problem and weak solutions

The natural starting point is a Dirichlet integral

(2.1) I(u) =

∫

Ω

|∇u|pdx

with the exponent p, 1 < p < ∞, in place of the usual 2. Minimizing the integralamong all admissible functions with the same given boundary values, we are led tothe condition that the first variation must vanish, that is

(2.2)

∫

Ω

〈|∇u|p−2∇u,∇η〉dx = 0

for all η ∈ C∞0 (Ω). This is the key to the concept of weak solutions. Under suitableassumptions this is equivalent to

(2.3)

∫

Ω

η div(|∇u|p−2∇u)dx = 0 .

Since (2.3) has to hold for all test functions η, we must have

(2.4) ∆pu ≡ div(|∇u|p−2∇u) = 0

in Ω. In other words, the p-Laplace equation is the Euler-Lagrange equation for thevariational integral I(u).

It turns out that the class of classical solutions is too narrow for the treatment ofthe aforementioned Dirichlet problem. (By a classical solution we mean a solutionhaving continuous second partial derivatives, so that the equation can be pointwiseverified.) We define the concept of weak solutions, requiring no more diffenrentiabil-ity than that they belong to the first order Sobolev space W 1,p(Ω). Even the localspace W 1,p

loc (Ω) will do.

5

2.5. Definition. Let Ω be a domain in Rn. We say that u ∈ W 1,ploc (Ω) is a weak

solution of the p-Laplace equation in Ω, if

(2.6)

∫〈|∇u|p−2∇u,∇η〉dx = 0

for each η ∈ C∞0 (Ω). If, in addition, u is continuous, then we say that u is ap-harmonic function.

We naturally read |0|p−20 as 0 also when 1 < p < 2. As we will see in section3, all weak solutions are continuous. In fact, every weak solution can be redefinedin a set of zero Lebesgue measure so that the new function is continuous. Whenappropriate, we assume that the redefinition has been performed.

We have the following basic result.

2.7. Theorem. The following conditions are equivalent for u ∈ W 1,p(Ω):

(i) u is minimizing:

∫|∇u|pdx ≤

∫|∇v|pdx , when v − u ∈ W 1,p

0 (Ω).

(ii) the first variation vanishes:

∫〈|∇u|p−2∇u,∇η〉dx = 0 , when η ∈ W 1,p

0 (Ω).

If, in addition, ∆pu is continuous, then the conditions are equivalent to the pointwiseequation ∆pu = 0 in Ω.

Remark. If (2.6) holds for all η ∈ C∞0 (Ω), then it also holds for all η ∈ W 1,p0 (Ω), if

we know that u ∈ W 1,p(Ω). Thus the minimizers are the same as the weak solutions.

Proof: ”(i) ⇒ (ii)”. We use a device due to Lagrange. If u is minimizing, select

v(x) = u(x) + εη(x) ,

where ε is a real parameter. Since

J(ε) =

∫

Ω

|∇(u+ εη)|pdx

6

attains its minimum for ε = 0, we must have J ′(0) = 0 by the infinitesimal calculus.This is (ii).

”(ii) ⇒ (i)” The inequality

|b|p ≥ |a|p + p〈|a|p−2a, b− a〉

holds for vectors (if p ≥ 1) by convexity. It follows that

∫

Ω

|∇v|pdx ≥∫

Ω

|∇u|pdx+ p

∫

Ω

〈|∇u|p−2∇u,∇(v − u)〉dx .

If (ii) is valid, take η = v − u to see that the last integral vanishes. This is (i).

Finally, the equivalence of (ii) and the extra condition is obtained from (2.3).

Before proceeding, we remark that the operator

∆pu = |∇u|p−4

|∇u|2∆u+ (p− 2)

∑ ∂u

∂xi

∂u

∂xj

∂2u

∂xi∂xj

is not well defined at points where ∇u = 0 in the case 1 < p < 2, at least not forarbitrary smooth functions. In the case p ≥ 2 one can divide out the crucial factor.Actually, the weak solutions u ∈ C2(Ω) are precisely characterized by the equation

(2.8) |∇u|2∆u+ (p− 2)∑ ∂u

∂xi

∂u

∂xj

∂2u

∂xi∂xj= 0

for all p in the range 1 < p < ∞. The proof for p < 2 is difficult, cf [JLM]. Thereader may think of the simpler problem: Why are the equations |∇u|∆u = 0 and∆u = 0 equivalent for u ∈ C2 ?

Let us return to Definition 2.5 and derive some preliminary estimates from theweak form of the equation. The art is to find the right test function. We will oftenuse the notation

Br = B(x0, r) , B2r = B(x0, 2r)

for concentric balls of radii r and 2r, respectively.

7

2.9. Lemma. (Caccioppoli) If u is a weak solution in Ω, then

(2.10)

∫

Ω

ζp|∇u|pdx ≤ pp∫

Ω

|u|p|∇ζ|pdx

for each ζ ∈ C∞0 (Ω), 0 ≤ ζ ≤ 1. In particular, if B2r ⊂ Ω, then

(2.11)

∫

Br

|∇u|pdx ≤ ppr−p∫

B2r

|u|pdx .

Proof: Useη = ζpu ,

∇η = ζp∇u+ pζp−1u∇ζ .

By the equation (2.6) and Holder’s inequality

∫

Ω

ζp|∇u|pdx = −p∫

Ω

ζp−1u〈|∇u|p−2∇u,∇ζ〉dx

≤ p

∫

Ω

|ζ∇u|p−1|u∇ζ|dx

≤ p

∫

Ω

ζp|∇u|pdx1− 1

p∫

Ω

|u|p|∇ζ|pdx 1

p

.

The estimate follows.

Finally, if B2r ⊂ Ω, we may choose ζ as a radial function satisfying ζ = 1 inBr, |∇ζ| ≤ r−1 and ζ = 0 outside B2r. This is possible by approximation. Thisconcludes the proof.

Occasionally, it is useful to consider weak supersolutions and weak subsolutions.As a mnemonic rule, ”∆pv ≤ 0” for supersolutions and ”∆pu ≥ 0” for subsolutions.

2.12. Definition. We say that v ∈ W 1,ploc (Ω) is a weak supersolution in Ω, if

(2.13)

∫

Ω

〈|∇v|p−2∇v,∇η〉dx ≥ 0

for all nonnegative η ∈ C∞0 (Ω). For weak subsolutions the inequality is reversed.

8

In the a priori estimate below it is remarkable that the majorant is independentof the weak supersolution itself.

2.14. Lemma. If v > 0 is a weak supersolution in Ω, then

∫

Ω

ζp|∇ log v|pdx ≤ (p

p− 1)p∫

Ω

|∇ζ|pdx

whenever ζ ∈ C∞0 (ζ), ζ ≥ 0.

Proof: One may add constants to the weak supersolutions. First, prove the esti-mate for v(x) + ε in place of v(x). Then let ε→ 0 in

∫

Ω

ζp|∇v|p(v + ε)p

dx ≤ (p

p− 1)p∫

Ω

|∇ζ|pdx .

Hence we may assume that v(x) ≥ ε > 0. Next use the test function η = ζpv1−p.Then

∇η = pζp−1v1−p∇ζ − (p− 1)ζpv−p∇vand we obtain

(p− 1)

∫

Ω

ζpv−p|∇v|pdx ≤ p

∫

Ω

ζp−1v1−p〈|∇v|p−2∇v,∇ζ〉dx

≤ p

∫

Ω

ζp−1v1−p|∇v|p−1|∇ζ|dx

≤ p

∫

Ω

ζpv−p|∇v|pdx1− 1

p∫

Ω

|∇ζ|pdx 1

p

,

from which the result follows.

The Comparison Principle, which in the linear case is merely a restatement of theMaximum Principle, is one of the cornerstones in the theory.

2.15. Theorem. (Comparison Principle) Suppose that u and v are p-harmonicfunctions in a bounded domain Ω. If at each ζ ∈ ∂Ω

lim supx→ζ

u(x) ≤ lim infx→ζ

v(x) ,

excluding the situation ∞ ≤∞ and −∞ ≤ −∞, then u ≤ v in Ω.

9

Proof: Given ε > 0, the open set

Dε = x|u(x) > v(x) + ε

is empty or Dε ⊂⊂ Ω . Subtracting the equations we get

∫

Ω

〈|∇v|p−2∇v − |∇u|p−2∇u,∇η〉dx = 0

for all η ∈ W 1,p0 (Ω) with compact support in Ω. The choice

η(x) = minv(x)− u(x) + ε, 0

yields ∫

Dε

〈|∇v|p−2∇v − |∇u|p−2∇u,∇v −∇u〉dx = 0 .

This is possible only if ∇u = ∇v a.e. in Dε, because the integrand is positive when∇u 6= ∇v. Thus u(x) = v(x) +C in Dε and C = ε because u(x) = v(x) + ε on ∂Dε.Thus u ≤ v + ε in Ω. It follows that u ≤ v.

Remark. The Comparison Principle also holds when u is a weak subsolution andv a weak supersolution. Then u ≤ v is valid a.e. in Ω.

The next topic is the existence of a p-harmonic function with given boundaryvalues. One can use the Lax-Milgram theorem,but I prefer the direct method in theCalculus of Variations, due to Lebesgue in 1907. The starting point is the variationalintegral (2.1), the Dirichlet integral with p.

2.16. Theorem. Suppose that g ∈ W 1,p(Ω), where Ω is a bounded domain in Rn.There exists a unique u ∈ W 1,p(Ω) with the boundary values u − g ∈ W 1,p

0 (Ω) suchthat ∫

Ω

|∇u|p dx ≤∫

Ω

|∇v|p dx

for all similar v. Thus u is a weak solution. In fact, u ∈ C(Ω) after a redefinition.If, in addition, g ∈ C(Ω) and if the boundary ∂Ω is regular enough, then u ∈ C(Ω)and u|∂Ω = g|∂Ω.

10

Proof: Let us begin with the uniqueness, which is a consequence of strict convexity.If there were two minimizers, say u1 and u2, we could choose the competing functionv = (u1 + u2)/2 and use

∣∣∣∣∇u1 +∇u2

2

∣∣∣∣p

≤ |∇u1|p + |∇u2|p2

.

If ∇u1 6= ∇u2 in a set of positive measure, then the above inequality is strict thereand it would follow that

∫

Ω

|∇u2|p dx ≤∫

Ω

∣∣∣∣∇u1 +∇u2

2

∣∣∣∣p

dx ≤∫

Ω

|∇u1|p + |∇u2|p2

dx

<1

2

∫

Ω

|∇u1|p dx +

∫

Ω

|∇u2|p dx =

∫

Ω

|∇u2|p dx,

which is a clear contradicition. Thus ∇u1 = ∇u2 a.e. in Ω and hence u1 =u2+Constant. The constant of integration is zero, because u2 − u1 ∈ W 1,p

0 (Ω).This proves the uniqueness.

The existence of a minimizer is obtained through the so-called direct method, see[D] and [G]. Let

I0 = inf

∫

Ω

|∇v|pdx ≤∫

Ω

|∇g|pdx <∞ .

Thus 0 ≤ I0 <∞. Choose admissible functions vj such that

(2.17)

∫

Ω

|∇vj|pdx < I0 +1

j, j = 1, 2, 3, . . .

We aim at bounding the sequence ‖vj‖W 1,p(Ω). The inequality

‖w‖Lp(Ω) ≤ CΩ‖∇w‖Lp(Ω)

holds for all w ∈ W 1,p0 (Ω), and in particular for w = vj − g. We obtain

‖vj − g‖Lp(Ω) ≤ CΩ‖∇vj‖Lp(Ω) + ‖∇g‖Lp(Ω)≤ CΩ(I0 + 1)

1p + ‖∇g‖Lp(Ω)

Now it follows from the triangle inequality that

(2.18) ‖vj‖Lp(Ω) ≤M (j = 1, 2, 3, . . . )

11

where the constant M is independent of the index j. Together 2.17 and 2.18 consti-tute the desired bound.

By weak compactness there exist a function u ∈ W 1,p(Ω) and a subsequence suchthat

vjν u , ∇vjν ∇u weakly in Lp(Ω) .

We have u − g ∈ W 1,p0 (Ω), because this space is closed under weak convergence.

Thus u is an admissible function. We claim that u is also the minimizer sought for.By weak lower semicontinuity

∫

Ω

|∇u|pdx ≤ limν→∞

∫

Ω

|∇vjν |pdx = I0

and the claim follows. (This can also be deduced from

∫

Ω

|∇vjν |pdx ≥∫

Ω

|∇u|pdx+ p

∫

Ω

〈|∇u|p−2∇u,∇vjν −∇u〉dx

since the last integral approaches zero. Recall that

|b|p ≥ |a|p + p〈|a|p−2a, b− a〉, p ≥ 1 ,

holds for vectors.) We remark that, a posteriori, one can verify that the minimizingsequence converges strongly in the Sobolev norm.

For the rest of the proof we mention that the continuity will be treated in section3 and the question about classical boundary values is postponed till section 6.

A retrospect of the previous proof of existence reveals that we have avoided somedangerous pitfalls. First, if we merely assume that the boundary values are contin-uous, say g ∈ C(Ω), it may so happen that I(v) = ∞ for each reasonable functionv ∈ C(Ω) with these boundary values g. Indeed, J. Hadamard has given such anexample for p = n = 2. If we take Ω as the unit disc in the plane and define

g(r, θ) =∞∑

j=1

rj! cos(j!θ)

n2

in polar coordinates, we have the example. The function g(r, θ) is harmonic whenr < 1 and continuous when r ≤ 1 (use Weierstrass’s test for uniform convergence).The Dirichlet integral of g is infinite. –Notice that we have avoided the phenomenon,encountered by Hadamard, by assuming that g belongs to a Sobolev space.

12

The second remark is a celebrated example of Weierstrass. He observed that theone-dimensional variational integral

I(u) =

1∫

−1

x2u′(x)2dx

has no continuous minimizer with the ”boundary values” u(−1) = −1 and u(+1) =+1. The weight function x2 is catastrophical near the origin. The example can begeneralized. This indicates that some care is called for, when it comes to questionsabout existence.

We find it appropriate to give a quantitative formulation of the continuity of theweak solutions, although the proof is postponed.

2.19. Theorem. Suppose that u ∈ W 1,ploc (Ω) is a weak solution to the p-harmonic

equation. Then

|u(x)− u(y)| ≤ L|x− y|α

for a.e. x, y ∈ B(x0, r) provided that B(x0, 2r) ⊂⊂ Ω. The exponent α > 0 dependsonly on n and p, while L also depends on ‖u‖Lp(B2r).

We shall deduce the theorem from the so-called Harnack inequality, given belowand proved in section 3. We write Br = B(x0, r).

2.20. Theorem. (Harnack’s inequality) Suppose that u ∈ W 1,ploc (Ω) is a weak

solution and that u ≥ 0 in B2r ⊂ Ω. Then the quantities

m(r) = ess infBr

u , M(r) = ess supBr

u

satisfy

M(r) ≤ Cm(r)

where C = C(n, p).

The main feature is that the same constant C will do for all weak solutions.

Since one may add constants to solutions, the Harnack inequality implies Holdercontinuity. To see this, first apply the inequality to the two non-negative weaksolutions u(x)−m(2r) and M(2r)− u(x), where r is small enough. It follows that

M(r)−m(2r) ≤ C(m(r)−m(2r)),

M(2r)−m(r) ≤ C(M(2r)−M(r)) .

13

Hence

ω(r) ≤ C − 1

C + 1ω(2r)

where ω(r) = M(r) − m(r) is the (essential) oscillation of u over B(x0, r). It isdecisive that

λ =C − 1

C + 1< 1 .

Iterating ω(r) ≤ λω(2r), we get ω(2−kr) ≤ λkω(r). We conclude that

ω(%) ≤ A(%

r)αω(r), 0 < ρ < r

for some α = α(n, p) > 0 and A = A(n, p).

Thus we have proved that Harnack’s inequality implies Holder continuity3, pro-vided that we already know that also sign changing solutions are locally bounded.The possibility ω(r) =∞ is eliminated in Corollary 3.8.

Finally, we point out a simple but important property, the Strong MaximumPrinciple.

2.21. Corollary. (Strong Maximum Principle) If a p-harmonic function attainsits maximum at an interior point, then it reduces to a constant.

Proof: If u(x0) = maxx∈Ω

u(x) for x0 ∈ Ω, then we can apply the Harnack inequality

on the p-harmonic function u(x0) − u(x), which indeed is non-negative. It followsthat u(x) = u(x0), when 2|x − x0| < dist(x0, ∂Ω). Through a chain of intersectingballs the identity u(x) = u(x0) is achieved at an arbitrary point x in Ω.

Remark. Of course, also the corresponding Strong Minimum Principle holds.However, a strong version of the Comparison Principle is not known in several di-mensions, n ≥ 3, when p 6= 2.

3 ”Was it Plato who made his arguments by telling a story with an obvious flaw, and allowingthe listener to realize the error?”

14

3. Regularity theory

The weak solutions of the p-harmonic equation are, by definition, members ofthe Sobolev space W 1,p

loc (Ω). In fact, they are also of class Cαloc(Ω). More precisely,

a weak solution can be redefined in a set of Lebesgue measure zero, so that thenew function is locally Holder continuous with exponent α = α(n, p). Actually, adeeper and stronger regularity result holds. In 1968 N.Ural’tseva proved that eventhe gradient is locally Holder continuous; we refer to [Ur], [Db1], [E], [Uh], [Le2],[To] for this C1,α

loc result.

To obtain the Holder continuity of the weak solutions one had better distinguishbetween three cases, depending on the value of p. Recall that n is the dimension.

1) If p > n, then every function in W 1,p(Ω) is continuous.

2) The case p = n (the so-called borderline case) is rather simple, but requires aproof. We will present a proof based on ”the hole filling technique”of Widman.

3) The case p < n is much harder. Here the regularity theory of elliptic equationsis called for. There are essentially three methods, developed by

– E. DeGiorgi 1957

– J. Nash 1958

– J. Moser 1961

to prove the Holder continuity in a wide class of partial differential equations.While DeGeorgi’s method is the most robust, we will, nevertheless, use Moser’sapproach, which is very elegant. Thus we will present the so-called Moseriteration, which leads to Harnack’s inequality. A short presentation for p = 2can be found in [J]. See also [Mo2]. The general p is in [T1]. DeGiorgi’smethod is in [Dg], [G] and [LU]. For an alternative proof of the case p > n− 2and p ≥ 2 see the remark after the proof of Theorem 4.1.

3.1. The case p > n

In this case all functions in the Sobolev space W 1,ploc (Ω) are continuous. Indeed, if

p > n and v ∈ W 1,p(B) where B is a ball (or a cube) in Rn, then

(3.1) |v(y)− v(x)| ≤ C1|x− y|1−np ‖∇v‖Lp(B)

15

when x, y ∈ B, cf [GT, Theorem 7.17]. The Holder exponent is α = 1− np. If u is a

positive weak solution or supersolution, Lemma 2.14 implies

(3.2) ‖∇ log u‖Lp(Br) ≤ C2rn−pp

assuming that u > 0 in B2r. For v = log u we obtain

(3.3)

∣∣∣∣ logu(y)

u(x)

∣∣∣∣ ≤ C1C2 .

This is Harnack’s inequality (see Theorem 2.20) with the constant C(n, p) = eC1C2 .

In the favourable case p > n a remarkable property holds for the Dirichlet problem:all the boundary points of an arbitrary domain are regular. Indeed, if Ω is a boundeddomain in Rn and if g ∈ C(Ω)∩W 1,p(Ω) is given, there exists a p-harmonic functionu ∈ C(Ω)∩W 1,p(Ω) such that u = g on ∂Ω. The boundary values are attained, notonly in Sobolev’s sense, but also in the classical sense. This follows from the generalinequality

|v(y)− v(x)| ≤ CΩ|x− y|1−np ‖∇v‖Lp(Ω)

valid for all v ∈ W 1,p0 (Ω). Hence v ∈ Cα(Ω) and v = 0 on ∂Ω. The argument is to

apply the inequality to a minimizing sequence. See also section 6.

3.2. The case p = n

The proof of the Holder continuity is based on the so-called hole filling technique(due to Widman, see [Wi]) and the following elementary lemma. We do not seemto reach Harnack’s inequality this way.

3.4. Lemma. (Morrey) Assume that u ∈ W 1,p(Ω), 1 ≤ p <∞. Suppose that

(3.5)

∫

Br

|∇u|pdx ≤ Krn−p+pα

whenever B2r ⊂ Ω. Here 0 < α ≤ 1 and K are independent of the ball Br. Thenu ∈ Cα

loc(Ω). In fact,

oscBr

(u) ≤ 4

α

(K

ωn

) 1p

rα , B2r ⊂ Ω .

16

Proof: See [LU, Chapter 2, Lemma 4.1, p.56] or [GT, Theorem 7.19].

For the continuity proof, we let B2r = B(x0, 2r) ⊂⊂ Ω. Select a radial testfunction ζ such that 0 ≤ ζ ≤ 1, ζ = 1 in Br, ζ = 0 outside B2r and |∇ζ| ≤ r−1.Choose

η(x) = ζ(x)n(u(x)− a)

in the n-harmonic equation. This yields

∫

Ω

ζn|∇u|ndx = −n∫

Ω

ζn−1(u− a)〈|∇u|n−2∇u,∇ζ〉dx

≤ n

∫

Ω

|ζ∇u|n−1|(u− a)∇ζ|dx

≤ n∫

Ω

ζn|∇u|ndx1− 1

n∫

Ω

|u− a|n|∇ζ|ndx 1n.

It follows that ∫

Br

|∇u|ndx ≤ nnr−n∫

B2r\Br

|u− a|ndx .

The constant a is at our disposal. Let a denote the average

a =1

|H(r)|

∫

H(r)

u(x)dx

of u taken over the annulus H(r) = B2r \Br. The Poincare inequality

∫

H(r)

|u(x)− a|ndx ≤ Crn∫

H(r)

|∇u|ndx

yields ∫

Br

|∇u|ndx ≤ Cnn∫

H(r)

|∇u|ndx .

Now the trick comes. Add Cnn∫Br|∇u|ndx to both sides of the last inequality.

This fills the hole in the annulus and we obtain

(1 + Cnn)

∫

Br

|∇u|ndx ≤ Cnn∫

B2r

|∇u|ndx .

17

In other wordsD(r) ≤ λD(2r) , λ < 1 ,

holds for the Dirichlet integral

D(r) =

∫

Br

|∇u|ndx

with the constant

λ =Cnn

1 + Cnn< 1 .

By iterationD(2−kr) ≤ λkD(r) , k = 1, 2, 3, . . .

A calculation reveals that

D(%) ≤ 2δ(%

r)δD(r) , 0 < % < r ,

with δ = log(1/λ) : log 2, when B2r ⊂ Ω. This is the estimate called for in Morrey’slemma. The Holder continuity follows.

Remark. A careful analysis of the above proof shows that it works for all p in asmall range (n− ε, n], where ε = ε(n, p).

3.3. The case 1 < p < n

This is much more difficult than the case p ≥ n. The idea of Moser’s proof is toreach the Harnack inequality

ess supB

u ≤ C ess infB

u

through the limits

ess supB

u = limq→∞

∫

B

uqdx 1q

ess infB

u = limq→−∞

∫

B

uqdx 1q

18

The equation is used to deduce reverse Holder inequalities like

∫

Br

up2dx 1p2 ≤ K

∫

BR

up1dx 1p1

where −∞ < p1 < p2 <∞ and 0 < r < R. The ”constant”K will typically blow upas r → R, and, since one does not reach all exponents at one stroke, one has to payattention to this, when using the reverse Holder inequality infinitely many times.

Several lemmas are needed and it is convenient to include weak subsolutions andsupersolutions. In the first lemma we do not assume positivity, because we need itto conclude that arbitrary solutions are locally bounded.

3.6. Lemma. Let u ∈ W 1,ploc (Ω) be a weak subsolution. Then

(3.7) ess supBr

(u+) ≤ Cβ

1

(R− r)n∫

BR

uβ+dx

1β

for β > p− 1 when BR ⊂⊂ Ω. Here u+ = maxu(x), 0 and Cβ = C(n, p, β).

Proof: The proof has two major steps. First, the test function η = ζpuβ−(p−1)+ is

used to produce the estimate

∫

Br

uκβ+ dx 1κβ ≤ C

1β

(2β − p+ 1

β − p+ 1

) pβ 1

(R− r)pβ

∫

BR

uβ+dx 1β

where κ = n/(n− p) and β > p− 1. Second, the above estimate is iterated so thatthe exponents κβ, κ2β, κ3β, . . . are reached, while the radii shrink.

Write α = β − (p− 1) > 0. We insert

∇η = pζp−1uα+∇ζ + αuα−1+ ζp∇u+

into the equation. This yields

α

∫

Ω

ζpuα−1+ |∇u+|pdx = −p

∫

Ω

ζp−1uα+〈|∇u+|p−2∇u+,∇ζ〉dx

since ∇u+ = ∇u a.e. in the set where u ≥ 0.

19

where S = S(n, p). Recall, that, as usual |∇ζ| ≤ 1/(R − r) and ζ = 1 in Br. Itfollows that

∫

Br

uκβdx

1κβ

≤(

S2β − p+ 1

β − p+ 1

1

R− r)p ∫

BR

uβdx

1β

.

We have accomplished the first step, a reverse Holder inequality.

Next, let us iterate the estimate. Fix a β, say β0 > p− 1 and notice that

2β − p+ 1

β − p+ 1≤ 2β0 − p+ 1

β0 − p+ 1= b

when β ≥ β0. Start with β0 and the radii r0 = R and r1 = r + (R − r)/2 in theplace of R and r. This yields

‖u‖κβ0,r1 ≤ (Sb)p/β0( 2

R− r) pβ0 ‖u‖β0,r0

with the notation

‖u‖q,% =

∫

B%

uqdx

1q

.

Then use r1 and r2 = r + 2−2(R− r) to improve κβ0 to κ2β0. Hence

‖u‖κ2β0,r2 ≤ (Sb)puβ0

( 4

R− r) pκβ0 ‖u‖κβ0,r1

≤ (Sb)pβ0

+ pκβ0

2pβ0

+ 2pκβ0

(R− r)pβ0

+ pκβ0

‖u‖β0,r0

Here we can discern a pattern. Continuing like this, using radii rj = r+ 2−j(R− r),we arrive at

‖u‖κj+1β0,rj+1≤( Sb

R− r)pβ−1

0

∑κ−k

2pβ−10

∑(k+1)κ−k‖u‖β0,r0

where the index k is summed over 0, 1, 2, . . . , j. The sums in the exponents areconvergent and, for example,

∑κ−k =

1− κ−j−1

1− κ−1→ n

p

as j →∞. To conclude the proof, use

‖u‖κj+1β0,r ≤ ‖u‖κj+1β0,rj+1

and let j →∞. The majorant contains (R− r) to the correct power n/β0.

21

3.8 . Corollary. The weak solutions to the p-harmonic equation are locallybounded.

Proof: Let β = p and apply the lemma to u and −u.

The next lemma is for supersolutions. It is decisive that one may take the exponentβ > p− 1, which is possible since κ > 1. Hence one can combine with Lemma 3.6,because of the overlap.

3.9. Lemma. Let v ∈ W 1,ploc (Ω) be a non-negative weak supersolution. Then

(3.10)

1

(R− r)n∫

Br

vβdx

1β

≤ C(ε, β)

1

(R− r)n∫

BR

vεdx

1ε

,

when 0 < ε < β < κ(p− 1) = n(p− 1)/(n− p) and BR ⊂⊂ Ω.

Proof: We may assume that v(x) ≥ σ > 0. Otherwise, first prove the lemma forv(x) + σ and let σ → 0 at the end. Use the test function

η = ζpvβ−(p−1)

This yields

∫

Br

vκβdx

1κβ

≤ C1β

( p− 1

p− 1− β) pβ 1

(R− r)p/β∫

BR

vβdx

1β

for 0 < β < p−1. Notice that we can reach an exponent κβ > p−1. The calculationsare similar to those in Lemma 3.6 and are omitted.

An iteration of the estimate leads to the desired result. The details are skipped.

In the next lemma the exponent β < 0.

3.11. Lemma. Suppose that v ∈ W 1,ploc (Ω) is a non-negative supersolution. Then

(3.12)

1

(R− r)n∫

BR

vβdx

1β

≤ C ess infBr

v

when β < 0 and BR ⊂⊂ Ω. The constant C is of the form c(n, p)−1/β.

22

Proof: Use the test function η = ζpvβ−(p−1) again, but now β < 0. First we arriveat ∫

Ω

|∇(ζvβ/p)|pdx ≤(p− 1− 2β

p− 1− β)p ∫

Ω

vβ|∇ζ|pdx

after some calculations, similar to those in the proof of Lemma 3.6. The constant isless than 2p. Using the Sobolev inequality we can write

∫

Ω

ζκpvκβdx

1κ

≤ (2S)p∫

Ω

vβ|∇ζ|pdx

where S = S(n, p) and κ = n/(n− p). The estimate

∫

Br

vκβdx

1κ

≤( 2S

R− r)p ∫

BR

vβdx

follows. An iteration of the estimate with the radii r0 = R, r1 = r+2−1(R−r), r2 =r + 2−2(R− r), . . . yields, via the exponents β, κβ, κ2β, . . .

∫

Brj

vκjβdx

κ−j≤( 2S

R− r)p∑κ−k

2p∑

(k+1)κ−k∫

BR

vβdx

where ∑κ−k = 1 + κ−1 + · · ·+ κ−(j−1) .

As j →∞ we obtain

ess supBr

(vβ) ≤( 2S

R− r)n

2n2

p

∫

BR

vβdx

Taking into account that β < 0 we have reached the desired estimate.

Combining the estimates achieved so far in the case 1 < p < n, we have thefollowing bounds for non-negative weak solutions:

ess supBr

u ≤ C1(ε, n, p)

1

(R− r)n∫

BR

uεdx

1ε

,

ess infBr

u ≥ C2(ε, n, p)

1

(R− r)n∫

BR

u−εdx

− 1ε

23

for all ε > 0. Take R = 2r. The missing link is the inequality

∫

BR

uεdx

1ε

≤ C

∫

BR

u−εdx

− 1ε

for some small ε > 0. The passage from negative to positive exponents is delicate.The gap in the iteration scheme can be bridged over with the help of the John-Nirenberg theorem, which is valid for functions in L1. Its proof is in [JN] or [G,Section 2.4]. The weaker version given in [GT,Theorem 7.21] will do.

3.13. Theorem. (John-Nirenberg) Let w ∈ L1loc(Ω). Suppose that there is a

constant K such that

(3.14)

∫

Br

|w(x)− wBr |dx ≤ K

holds whenever B2r ⊂ Ω. Then there exists a constant ν = ν(n) > 0 such that

(3.15)

∫

Br

eν|w(x)−wBr |/Kdx ≤ 2

whenever B2r ⊂ Ω. (It also holds when Br ⊂ Ω.)

The notation

wBr =

∫Brw(x)dx∫Brdx

=

∫

Br

wdx

was used.

The two inequalities ∫

Br

e±ν(w(x)−wBr )/Kdx ≤ 2

follow immediately. Multiplying them we arrive at

(3.16)

∫

Br

eνw(x)/Kdx

∫

Br

e−νw(x)/Kdx ≤ 4

since the constant factors e±νwBr/K cancel.

Next we use w = log u for the passage from negative to positive exponents. Firstwe show that w = log u satisfies (3.14). Then we can conclude from (3.16) that

∫

Br

uν/Kdx ·∫

Br

u−ν/Kdx ≤ 4 .

24

Writing ε = ν/K we have ”the missing link”

(3.17)

∫

Br

uεdx

1ε

≤ 41ε

∫

Br

u−εdx

− 1ε

when B2r ⊂⊂ Ω.

To complete the first step, assume to begin with that u > 0 is a weak solution.Combining the Poincare inequality

∫

Br

| log u(x)− (log u)Br |pdx ≤ C1rp

∫

Br

|∇ log u|pdx

with the estimate ∫

Br

|∇ log u|pdx ≤ C2rn−p

from lemma 2.14, we obtain for B2r ⊂⊂ Ω

∫

Br

|w − wBr |pdx ≤ C1C2ω−1n = K .

This is the bound needed in the John-Nirenberg theorem. Finally, to replace u > 0by u ≥ 0, it is sufficient to observe that, if (3.17) holds for the weak solutionsu(x) + σ, then it also holds for u(x).

We have finished the proof of the Harnack inequality

M(r) ≤ Cm(r), when B4r ⊂ Ω .

Remark. It is possible to avoid the use of the John-Nirenberg inequality in theproof. To accomplish the zero passage one can use the equation in a more effectiveway by a more refined testing. Powers of log u appear in the test function and anextra iteration procedure is used. The original idea is in [BG]. See also [SC], [HL,Section 4.4, pp. 85-89] and [T2].

We record an inequality for weak supersolutions.

25

3.18. Corollary. Suppose that v ∈ W 1,ploc (Ω) is a non-negative supersolution.

Then

(3.19)

∫

Br

vβdx

1β

≤ C(n, p, β) ess infBr

v , β <n(p− 1)

n− 1,

whenever B2r ⊂ Ω.

Proof: This is a combination of (3.10), (3.12), and (3.17).

In fact, weak supersolutions are lower semicontinuos, after a possible redefinitionin a set of measure zero. They are then pointwise defined.

3.20. Proposition. A weak supersolution v ∈ W 1,p(Ω) is lower semicontinuous(after a redefinition in a set of measure zero). We can define

v(x) = ess lim infy→x

v(y)

pointwise. This representative is a p-superharmonic function.

Proof: The case p > n is clear, since then the Sobolev space contains only contin-uous functions (Morrey’s inequality). In the range p < n we claim that


v(y)

at a.e. every x in Ω. The proof follows from this, since the right-hand side is alwayslower semicontinuous. For simplicity, we assume that v is locally bounded. Supposethat p > 2n/(n + 1) so that we may take q = 1 in the weak Harnack inequality(3.19). Use the function v(x)−m(2r), where

m(r) = ess infBr

v.

We have

0 ≤∫

B2r

v dx − m(2r)

=

∫

B2r

(v(x)−m(2r)

)dx ≤ C

(m(r)−m(2r)

).

Since m(r) is monotone, m(r)−m(2r)→ 0 as r → 0. It follows that

ess lim infy→x0

v(y) = limr→0

m(2r) = limr→0

∫

B(x0,2r)

v(x) dx

at each point x0. Lebesgue’s Differentiation Theorem states that the limit of theaverage on the right-hand side coincides with v(x0) at almost every point x0.

If we are forced to take q < 1 in the weak Harnack inequality, a slight modificationof the above proof will do.

26

4. Differentiability

We have learned that the p-harmonic functions are Holder continuous. In fact,much more regularity is valid. Even the gradients are locally Holder continuous. Insymbols, the function is of class C1,α

loc (Ω). More precisely, if u is p-harmonic in Ω andif D ⊂⊂ Ω, then

|∇u(x)−∇u(y)|α ≤ LD|x− y|

when x, y ∈ D. Here α = α(n, p) and LD depends on n, p, dist(D, ∂Ω) and ‖u‖∞.This was proved in 1968 by N. Uraltseva, cf. [Ur]. We also refer to [E], [Uh], [Le2],[Db1] and [To] about this difficult regularity question4. Here we are content with aweaker, but much simpler, result:

1) If 1 < p ≤ 2, then u ∈ W 2,ploc (Ω); that means that u has second Sobolev

derivatives.

2) If p ≥ 2, then |∇u|(p−2)/2∇u belongs to W 1,2loc (Ω). Thus the Sobolev derivatives

∂

∂xj

(|∇u| p−2

2∂u

∂xi

)

exist, but the passage to ∂2u∂xi∂xj

is very difficult at the critical points (∇u = 0).

According to Lemma 5.1 on page 20 in [MW] the second derivatives exist whenthe Cordes Condition5

1 < p < 3 +2

n− 2

holds. Then u ∈ W 2,2loc (Ω). To this one may add that u is real analytic (=is repre-

sented by the Taylor expansion) in the open set where ∇u 6= 0, cf. [Le1, p.208].

4The second Russian edition of the book [LU] by Ladyzhenskaya and Uraltseva includes theproof.

5For the equation∑

aij(x, u,∇u)∂2u

∂xi∂xj= a0(x, u,∇u)

the Cordes Condition reads

( n∑

j=1

ajj

)2≥ (n− 1 + δ)

n∑

i,j=1

a2ij .

27

We begin with the study of

F (x) = |∇u(x)|(p−2)/2∇u(x)

in the case p ≥ 2. It is plain that∫

Ω

|F |2dx =

∫

Ω

|∇u|pdx .

4.1. Theorem. (Bojarski - Iwaniec) Let p ≥ 2. If u is p-harmonic in Ω, thenF ∈ W 1,2

loc (Ω). For each subdomain G ⊂⊂ Ω,

(4.2) ‖DF‖L2(G) ≤C(n, p)

dist(G, ∂Ω)‖F‖L2(Ω) .

Proof: The proof is taken from [BI1]. It is based on integrated difference quotients.Let ζ ∈ C∞0 (Ω) be a cutoff function so that 0 ≤ ζ ≤ 1, ζ|G = 1 and |∇ζ| ≤Cn/ dist(G, ∂Ω). (If required, replace Ω by a smaller domain Ω1, G ⊂⊂ Ω1 ⊂⊂ Ω.)We aim at difference quotients. Take |h| < dist(supp ζ, ∂Ω). Notice that also uh =u(x + h) is p-harmonic, when x + h ∈ Ω, h denoting a constant vector. The testfunction

η(x) = ζ(x)2(u(x+ h)− u(x))

will do in the equations∫

Ω

〈|∇u(x)|p−2∇u(x),∇η(x)〉dx = 0 ,

∫

Ω

〈|∇u(x+ h)|p−2∇u(x+ h),∇η(x)〉dx = 0 .

Hence, after subtraction,

(4.3)

∫

Ω

〈|∇u(x+ h)|p−2∇u(x+ h)− |∇u(x)|p−2∇u(x),∇η(x)〉dx = 0 .

It follows that∫

Ω

ζ(x)〈|∇u(x+ h)|p−2∇u(x+ h)− |∇u(x)|p−2∇u(x),∇u(x+ h)−∇u(x)〉dx

= −2

∫

Ω

ζ(x)(u(x+ h)− u(x))〈|∇u(x+ h)|p−2∇u(x+ h)− |∇u(x)|p−2∇u(x),∇ζ(x)〉dx

≤ 2

∫

Ω

ζ(x)|u(x+ h)− u(x)|∣∣|∇u(x+ h)|p−2∇u(x+ h)− |∇u(x)|p−2∇u(x)

∣∣|∇ζ(x)|dx

28

To continue we need the ”elementary inequalities”

4

p2

∣∣|b| p−22 b− |a| p−2

2 a∣∣2 ≤ 〈|b|p−2b− |a|p−2a, b− a〉 ,

∣∣|b|p−2b− |a|p−2a∣∣ ≤ (p− 1)

(|a| p−2

2 + |b| p−22

)∣∣|b| p−22 b− |a| p−2

2 a∣∣

given in section 12. We obtain

4

p2

∫

Ω

ζ2(x)|F (x+ h)− F (x)|2dx

≤2(p− 1)

∫

Ω

|u(x+ h)− u(x)||∇ζ(x)|(|∇u(x+ h)| p−2

2 + |∇u(x)| p−22

)ζ(x)|F (x+ h)− F (x)|dx

≤2(p− 1)

∫

Ω

|u(x+ h)− u(x)|p|∇ζ(x)|pdx 1

p∫

Ω

ζ2(x)|F (x+ h)− F (x)|2dx 1

2

· ∫

supp ζ

(|∇u(x+ h)| p−2

2 + |∇u(x)| p−22

) 2pp−2dx

p−22p

.

At the last step Holder’s inequality with the three exponents p, 2 and 2p/(p − 2)was used; indeed, they match

1

p+

1

2+p− 2

2p= 1

as required. The last integral factor is majorized by

(∫

Ω

|∇u(x+ h)|pdx) p−2

2p

+

(∫

Ω

|∇u(x)|pdx) p−2

2p

≤ 2

(∫

Ω

|∇u(x)|pdx) p−2

2p

= 2

(∫

Ω

|F |2dx) p−2

2p

according to Minkowski’s inequality, when |h| is small. Dividing out the commonfactor (=the square root of the integral containing F (x+ h)− F (x)) we arrive at

(4.4)

1

p2

∫

Ω

ζ2(x)

∣∣∣∣F (x+ h)− F (x)

h

∣∣∣∣2

dx

12

≤ (p− 1)

∫

Ω

|F |2dx p−2

2p∫

Ω

∣∣∣∣u(x+ h)− u(x)

h

∣∣∣∣p

|∇ζ(x)|pdx 1

p

29

Recall the characterization of Sobolev spaces in terms of integrated differencequotients (see for example section 7.11 in [GT] or [LU, Chapter 2, Lemma 4.6,p.65]). We conclude that

∫

Ω

∣∣∣∣u(x+ h)− u(x)

h

∣∣∣∣p

|∇ζ(x)|pdx 1

p

≤ Cndist(G, ∂Ω)

∫

Ω

|∇u(x)|pdx 1

p

Hence (4.3) yields

∫

G

∣∣∣∣F (x+ h)− F (x)

h

∣∣∣∣2

dx

12

≤ C(n, p)

dist(G, ∂Ω)

∫

Ω

|F (x)|2dx 1

2

This is sufficient to guarantee that F ∈ W 1,2(G) and the desired bound follows.

Remark. A pretty simple proof of the Holder continuity of u is available, whenp > n − 2 and p ≥ 2. It is based on Theorem 4.1. The reasoning is as follows.Since the differential DF belongs to L2

loc(Ω) by the theorem, Sobolev’s inbedding

theorem assures that F ∈ L2n/(n−2)loc (Ω), that is∇u ∈ Lnp/(n−2)

loc (Ω). This summabilityexponent is large. Indeed

np

n− 2> n when p > n− 2 .

We conclude that u ∈ Cαloc(Ω), with α = 1 − (n − 2)/p, since it belongs to some

W 1,sloc (Ω) where s is greater than the dimension n.

This was the case p ≥ 2.

In the case 1 < p < 2 the previous proof does not work. However, an ingenioustrick, mentioned in [G, Section 8.2], leads to a stronger result. We start with asimple fact.

4.5. Lemma. Let f ∈ L1loc(Ω). Then

∫

Ω

ϕ(x)f(x+ hek)− f(x)

hdx = −

∫

Ω

∂ϕ

∂xk

( 1∫

0

f(x+ thek)dt

)dx

holds for all ϕ ∈ C∞0 (Ω).

30

Proof: For a smooth function f the identity holds, because

∂

∂xk

1∫

0

f(x+ thek)dt =f(x+ hek)− f(x)

h

by the infinitesimal calculus. The general case follows by approximation.

Regarding the xk-axis as the chosen direction, we use the abbreviation

∆hf = ∆hf(x) =f(x+ hek)− f(x)

h

By the lemma the formula

∆h(|∇u|p−2∇u) =∂

∂xk

1∫

0

|∇u(x+ thek)|p−2∇u(x+ thek)dt

can be used in Sobolev’s sense.

4.6. Theorem. Let 1 < p ≤ 2. If u is p-harmonic in Ω, then u ∈ W 2,ploc (Ω).

Moreover ∫

D

∣∣∣∣∂2u

∂xi∂xj

∣∣∣∣p

dx ≤ CD

∫

Ω

|∇u|pdx

when D ⊂⊂ Ω.

Proof: Use formula (4.1) again. In our new notation the identity next after (4.1)can be written as

∫ζ2〈∆h(|∇u|p−2∇u),∆h(∇u)〉dx

=− 2

∫ζ∆hu〈∆h(|∇u|p−2∇u),∇ζ〉dx

=2

∫〈

1∫

0

|∇u(x+ thek)|p−2∇u(x+ thek)dt,∂

∂xk(∆hu · ζ∇ζ)〉dx

The last equality was based on Lemma 4.5. This was ”the ingenious trick”. We have

∂

∂xk(∆hu · ζ∇ζ) = ζ∇ζ∆huxk + ∆hu(ζxk∇ζ + ζ∇ζxk)

31

by direct differentiation. Let us fix a ball B3R ⊂⊂ Ω and select a cutoff function ζvanishing outside B2R, ζ = 1 in BR, 0 ≤ ζ ≤ 1 such that

|∇ζ| ≤ R−1, |D2ζ| ≤ CR−2

For simplicity, abbreviate

Y (x) =

∫

B2R

|∇u(x+ thek)|p−1dt .

The estimate

(4.7)

∫

Ω

ζ2〈∆h(|∇u|p−2∇u),∆h(∇u)〉dx

≤ 2

R

∫

Ω

ζY |∆huxk |dx+c

R2

∫

B2R

|∆hu|Y dx

follows. Since 1 < p < 2, the inequality

〈|b|p−2b− |a|p−2a, b− a〉 ≥ (p− 1)|b− a|2(1 + |a|2 + |b|2)p−22

is available, see VII in section 10, and we can estimate the left hand side of (4.7)from below. With the further abbreviation

W (x)2 = 1 + |∇u(x)|2 + |∇u(x+ hek)|2

we write, using also |∆h(∇u)| ≥ |∆huxk |,

(p− 1)

∫

Ω

ζ2W p−2|∆h(∇u)|2dx ≤ 2

R

∫

Ω

ζY |∆h(∇u)|dx

+c

R2

∫

B2R

|∆hu|Y dx

The first term in the right hand member has to be absorbed (the so-called Peter-Paul Principle). To this end, let ε > 0 and use

2R−1ζY |∆h(∇u)| = 2ζW (p−2)/2|∆h(∇u)|W (2−p)/2Y R−1

≤ εζ2W p−2|∆h(∇u)|2 + ε−1R−2W 2−pY 2 .

32

For example, ε = (p− 1)/2 will do. The result is then

p− 1

2

∫

BR

W p−2|∆h(∇u)|2dx ≤ 2

p− 1R−2

∫

B2R

W 2−pY 2dx

+ cR−2

∫

B2R

|∆hu|Y dx .

Incorporating the elementary inequalities

|∆h(∇u)|p ≤ W p−2|∆h(∇u)|2 +W p ,

W 2−pY 2 ≤ W p + Y p/(p−1) ,

|∆hu|Y ≤ |∆hu|p + Y p/(p−1) ,

the estimate takes the form

∫

BR

|∆h(∇u)|pdx ≤ c1

∫

B2R

W pdx+ c2

∫

B2R

Ypp−1dx+ c3

∫

B2R

|∆hu|pdx

where the constants also depend on R. It remains to bound the three integrals ash→ 0. First, it is plain that

∫

B2R

W pdx ≤ CRn + C

∫

B3R

|∇u|pdx .

Second, the middle integral is bounded as follows:

∫

B2R

Ypp−1dx =

∫

B2R

( 1∫

0

|∇u(x+ thek)|p−1dt

) pp−1

dx

≤∫

B2R

1∫

0

|∇u(x+ thek)|pdtdx ≤∫

B3R

|∇u|pdx

for h small enough. For the last integral the bound

∫

B2R

|∆hu|pdx ≤∫

B3R

|∇u|pdx

33

follows from the characterization of Sobolev’s space in terms of integrated differencequotients.

Collecting the three bounds, we have the final estimate

∫

BR

|∆h(∇u)|pdx ≤ C(n, p,R)

∫

B3R

|∇u|pdx

and the theorem follows.

34

5. On p-superharmonic functions

In the classical potential theory the subharmonic and superharmonic functionsplay a central role. The gravitational potential predicted by Newton’s theory isthe leading example. It is remarkable that the mathematical features of this lin-ear theory are, to a great extent, preserved when the Laplacian is replaced by thep-Laplace operator or by some more general differential operator with a similarstructure. Needless to say, the principle of superposition is naturally lost in thisgeneralization. This is the modern non-linear potential theory, based on partialdifferential equations. –This chapter is taken from [L2].

5.1. Definition and examples

The definition is based on the Comparison Principle. (In passing, we mention thatthere is an equivalent definition used in the modern theory of viscosity solutions andthe p-superharmonic functions below are precisely the viscosity supersolutions, cf.[JLM].)

5.1. Definition. A function v : Ω→ (−∞,∞] is called p-superharmonic in Ω, if

(i) v is lower semi-continuous in Ω

(ii) v 6≡ ∞ in Ω

(iii) for each domain D ⊂⊂ Ω the Comparison Principle holds: if h ∈ C(D) isp-harmonic in D and h|∂D ≤ v|∂D, then h ≤ v in D

A function u : Ω→ [−∞,∞) is called p-subharmonic if v = −u is p-superharmonic.

It is clear that a function is p-harmonic if and only if it is both p-subharmonicand p-superharmonic, but Theorem ?? is needed for a proof.

For p = 2 this is the classical definition of F. Riesz. We emphasize that not eventhe existence of the gradient ∇v is required in the definition. (A very attentivereader might have noticed that the definition does not have a local character.) Aswe will learn, it exists in Sobolev’s sense. For sufficiently regular p-superharmonicfunctions we have the following, more practical, characterization.

5.2. Theorem. Suppose that v belongs to C(Ω) ∩W 1,ploc (Ω). Then the following

conditions are equivalent

35

(i)∫D|∇v|pdx ≤

∫D|∇(v + η)|pdx whenever D ⊂⊂ Ω and η ∈ C∞0 (D) is non-

negative

(ii)∫〈|∇v|p−2∇v,∇η〉dx ≥ 0 whenever η ∈ C∞0 (Ω) is non-negative

(iii) v is p-superharmonic in Ω.

Proof: The equivalence of (i) and (ii) is well-known in the Calculus of Variations.If (ii) is valid, so is (i) because

|∇(v + η)|p ≥ |∇v|p + p〈|∇v|p−2∇v,∇η〉 .

If (i) holds, then the function

J(ε) =

∫

D

|∇(v(x) + εη(x))|pdx

satisfies J(0) ≤ J(ε), when ε ≥ 0. Here the domain D ⊂⊂ Ω contains the supportof η. By the infinitesimal calculus J ′(0) ≥ 0. This is (ii).

It remains to show that (ii) and (iii) are equivalent. First, suppose that (ii) holds.Let D ⊂⊂ Ω and suppose that h ∈ C(D) is p-harmonic in D and h ≤ v on ∂D. Thetest function

η = maxh− v, 0produces the inequality

∫

v≤h

|∇v|pdx ≤∫

v≤h

〈|∇v|p−2∇v,∇h〉dx

≤ ∫

v≤h

|∇v|pdx1− 1

p ∫

v≤h

|∇h|pdx 1

p

.

Hence ∫

v<h

|∇v|pdx ≤∫

v<h

|∇h|pdx .

In other words, v is a minimizer in (each component of) the open set v < h. Theboundary values are v = h. The minimizer is unique and so v = h in this set. Thiscontradiction proves that v ≥ h. Thus (ii) or (i) implies (iii).

The proof of the sufficiency of (iii) seems to require the introduction of an obstacleproblem. It will be given in Corollary 5.8, which does not rely on ”(iii)⇒ (ii)” whenit comes to its proof.

36

Remark. The continuity of v is not needed for the equivalency of (i) and (ii).The whole theorem holds for lower semicontinuous functions in the Sobolev space.More could be said about this.

It is instructive to consider some examples. The one-dimensional situation isenlighting. The p-harmonic functions in one variable are just the line segmentsh(x) = ax + b. Now p has no bearing. The p-superharmonic functions are exactlythe concave functions of one variable. The comparison principle is the familiar”arc above chord” condition. – In several dimensions, the concave functions arep-superharmonic, simultaneously for all p, but there are many more of them.

The leading example of a p-superharmonic function is

(n− p)|x|p−np−1 (p 6= n) , − log |x| (p = n) ,

usually multiplied by a positive normalizing constant. Outside the origin the func-tion is p-harmonic. Notice that the function is not of class W 1,p(Ω), if Ω contains theorigin. Therefore it is not a weak supersolution in the sense of Definition 2.12! Wecannot resist mentioning that, although the principle of superposition is not valid,the function

(5.3) v(x) =

∫

Ω

%(y)dy

|x− y|(n−p)/(p−1)(1 < p < n)

is, indeed, p-superharmonic for %(y) ≥ 0. This follows from an interesting calculationby Crandall and Zhang done for the corresponding Riemann sums, cf [CZ]. Of course,this remarkable representation formula cannot directly give all the p-superharmonicfunctions.

It is useful that the pointwise minimium of two p-superharmonic functions is againp-superharmonic as a direct consequence of the definition.

Before going further we had better make a simple comment. Assumption (ii) inDefinition 5.1 means that v is finite at least at one point. In fact, it follows easilythat the set v <∞ is dense in Ω. (As we will later see, v <∞ a.e..)

5.4. Proposition. If v is p-superharmonic in Ω, then the set where v = ∞ doesnot contain any ball.

Proof: Suppose to begin with that v ≥ 0 in Ω. Assume that v ≡ +∞ in someball Br = B(x0, r) and that BR = B(x0, R) ⊂⊂ Ω, where R > r. We claim thatv ≡ +∞ also in the larger ball BR. The function

h(x) =

∫ R|x−x0| t

−(n−1)/(p−1)dt∫ Rrt−(n−1)/(p−1)dt

37

is p-harmonic when x 6= x0, in particular it is p-harmonic in the annulus r <|x − x0| < R. It takes the boundary values 0 on ∂BR and 1 on ∂Br. Considerthe p-harmonic function kh(x). The comparison principle shows that

v(x) ≥ kh(x), k = 1, 2, 3, . . .

in the annulus. We conclude that v ≡ ∞ in the annulus. In other words v ≡ ∞ inBR.

To get rid of the restriction v ≥ 0, we consider the function v− inf v instead of v.Again the conclusion is that v|BR ≡ ∞ if v|Br ≡ ∞.

Repeating the procedure through a suitable chain of balls, we finally arrive at thecontradiction v ≡ ∞ in Ω.

5.2. The obstacle problem and approximation

As we have seen, the p-harmonis functions come from a minimization problem inthe Calculus of Variations. If one adds a restriction on the admissible functions,when minimizing, weak supersolutions of the p-harmonic equation are produced.The restrictive condition is nothing more than that the functions have to lie abovea given function, which acts as a fixed obstacle.

Suppose, as usual, that Ω is a bounded domain in Rn. Given a function ψ ∈C(Ω) ∩W 1,p(Ω) we consider the problem of minimizing the integral

∫

Ω

|∇v|pdx

among all functions in the class

Fψ(Ω) = v ∈ C(Ω) ∩W 1,p(Ω)∣∣v ≥ ψ in Ω and v − ψ ∈ W 1,p

0 (Ω) .

This is the obstacle problem with ψ acting as an obstacle from below. Also theboundary values are prescribed by ψ. (One could also allow other boundary values,but we do not discuss this variant.)

5.5. Theorem. Given ψ ∈ C(Ω)∩W 1,p(Ω), there exists a unique minimizer vψ inthe class F(Ω), i.e. ∫

Ω

|∇vψ|pdx ≤∫

Ω

|∇v|pdx

38

for all similar v. The function vψ is p-superharmonic in Ω and p-harmonic in theopen set vψ > ψ. If in addition, Ω is regular enough and ψ ∈ C(Ω), then alsovψ ∈ C(Ω) and vψ = ψ on ∂Ω.

Proof: The existence of a unique minimizer is easily established, except for thecontinuity; only some functional analysis is needed. Compare with the problemwithout obstacle in section 2. It is the continuity that is difficult to prove in thecase 1 < p ≤ n. We refer to [MZ] for the proof of the continuity of vψ.

Next we conclude that

(5.6)

∫

Ω

〈|∇vψ|p−2∇vψ,∇η〉dx ≥ 0

when η ∈ C∞0 (Ω), η ≥ 0, according to Theorem 5.2, which also assures that vψ isp-superharmonic in Ω.

We have come to the important property that vψ is p-harmonic in the set wherethe obstacle does not hinder, say

S = x ∈ Ω|vψ(x) > ψ(x) .

In fact, we can conclude that (5.6) is valid for all η ∈ C∞0 (Ω), positive or not,satisfying

vψ(x) + εη(x) ≥ ψ(x)

for small ε > 0. Consequently, we can remove the sign restriction on η in the set S.Indeed, if η ∈ C∞0 (S) it suffices to consider ε so small that

ε‖η‖∞ ≤ min(vψ − ψ)

the minimum being taken over the support of η. Here η can take also negativevalues. We conclude that v is p-harmonic in S.

For the question about classical boundary values in regular domains we refer to[F].

I take myself the liberty to hint that it is a good excercise to work out the previousproof in the one-dimensional case, where no extra difficulties obscure the matter,and pictures can be drawn.

Remark. More advanced regularity theorems hold for the solution. If the obsta-cle is smooth, then vψ is of class C1,α

loc (Ω). Of course, the regularity cannot be any

39

better than for general p-harmonic functions. We refer to [CL], [S] and [L3] aboutthe gradient ∇vψ.

In the sequel we will use a sequence of obstacles to study the differentiability prop-erties of p-superharmonic functions. The proof that an arbitrary p-superharmonicfunction v has Sobolev derivatives requires several steps:

1) v is pointwise approximated from below by smooth functions ψj.

2) The obstacle problem with ψj acting as an obstacle is solved. It turns out that

ψj(x) ≤ vψj(x) ≤ v(x) .

3) Since the vψj ’s are supersolutions, they satisfy expedient a priori estimates.

4) The a priori estimates are passed over to v = lim vψj , first in the case when vis bounded.

5) For an unbounded v one goes via the bounded p-superharmonic functionsminv(x), k and an estimate free of k is reached at the end.

To this end, we assume that v is p-superharmonic in Ω. Because of the lowersemicontinuity of v, there exists an increasing sequence of functions ψj ∈ C∞(Ω)such that

ψ1(x) ≤ ψ2(x) ≤ · · · ≤ v(x) , limj→∞

ψj(x) = v(x)

at each x ∈ Ω. Next, fix a regular bounded domain D ⊂⊂ Ω. Let vj = vψj denotethe solution of the obstacle problem in D, the function ψj acting as an obstacle.Thus vj ∈ Fψj(D) and vj ≥ ψj in D. We claim that

v1 ≤ v2 ≤ . . . , ψj ≤ vj ≤ v

pointwise in D. To see that vj ≤ v, we notice that this is true except possibly in theopen set Aj = vj > ψj, where the obstacle does not hinder. By Theorem 5.5 vjis p-harmonic in Aj (provided that Aj is not empty) and on the boundary ∂Aj weknow that vj = ψj. Hence vj ≤ v on ∂Aj and so the comparison principle, which vis assumed to obey, implies that vj ≤ v in Aj. This was the main point in the proof,here the comparison principle was used. We have proved that vj ≤ v at each pointin D.

The inequalities vj ≤ vj+1, j = 1, 2, 3, . . . , have a similar proof, because vj+1

satisfies the comparison principle according to Theorem 5.5.

We have established the first part of the next theorem.

40

5.7 . Theorem. Suppose that v is a p-superharmonic function in the domainΩ. Given a subdomain D ⊂⊂ Ω there are such p-superharmonic functions vj ∈C(D) ∩W 1,p(D) that

v1 ≤ v2 ≤ . . . and v = limj→∞

vj

at each point in D. If, in addition, v is (locally) bounded from above in Ω, then alsov ∈ W 1,p

loc (Ω), and the approximants vj can be chosen so that

limj→∞

∫

D

|∇(v − vj)|pdx = 0 .

Proof: Fix D and choose a regular domain D1, D ⊂⊂ D1 ⊂⊂ Ω. By the previousconstruction there are p- superharmonic functions vj in D1 such that v1 ≤ v2 ≤ . . . ,vj → v pointwise in D1 and vj ∈ C(D1) ∩W 1,p(D1).

For the second part of the theorem we know that

C = supD1

v − infD1

ψ1 <∞

if v is locally bounded. Theorem 5.2 and a simple modification of Lemma 2.9 toinclude weak supersolutions lead to the bound

∫

D

|∇vj|pdx ≤ ppCp

∫

D1

|∇ζ|pdx = M (j = 1, 2, 3, . . . ) .

By a standard compactness argument v ∈ W 1,p(D) and ‖∇v‖Lp(D) ≤ M . For asubsequence we have that ∇vk ∇v weakly in Lp(D). We also conclude thatv ∈ W 1,p

loc (Ω).

To establish the strong convergence of the gradients, it is enough to show that

limj→∞

∫

Br

|∇v −∇vj|pdx = 0

whenever Br is such a ball in D that the concentric ball B2r (with double radius) iscomprised in D1. As usual, let ζ ∈ C∞0 (B2r), 0 ≤ ζ ≤ 1 and ζ = 1 in Br. Next, usethe non-negative test function ηj = ζ(v − vj) in the equation

∫

B2r

〈|∇vj|p−2∇vj,∇ηj〉dx ≥ 0

41

to find that

Jj =

∫

B2r

〈|∇v|p−2∇v − |∇vj|p−2∇vj,∇(ζ(v − vj))〉dx

≤∫

B2r

〈|∇v|p−2∇v,∇(ζ(v − vj))〉dx

By the weak convergence of the gradients

lim supj→∞

Jj ≤ 0 .

We split Jj in two parts:

Jj =

∫

B2r

ζ〈|∇v|p−2∇v − |∇vj|p−2∇vj,∇v −∇vj〉dx

+

∫

B2r

(v − vj)〈|∇v|p−2∇v − |∇vj|p−2∇vj,∇ζ〉dx

The last integral is bounded in absolute value by the majorant

∫

B2r

(v − vj)pdx 1

p( ∫

B2r

|∇v|pdx)1− 1

p

+

( ∫

B2r

|∇vj|pdx)1− 1

p

max |∇ζ|

≤ 2M1− 1p max |∇ζ|

∫

B2r

(v − vj)pdx 1

p

and hence it approaches zero as j →∞. Collecting results, we see that

limj→∞

∫

B2r

ζ〈|∇v|p−2∇v − |∇vj|p−2∇vj,∇v −∇vj〉dx ≤ 0

at least for a subsequence. The integrand is non-negative. For p ≥ 2 we can use theinequality

ζ〈|∇v|p−2∇v − |∇vj|p−2∇vj,∇v −∇vj〉 ≥ 22−p|∇v −∇vj|p

in Br to conclude the proof. The reader might find it interesting to complete theproof for 1 < p < 2.

42

With this approximation theorem it is easy to prove that bounded p-superhar-monic functions are weak supersolutions. Also the opposite statement is true, pro-vided that the issue about semicontinuity be properly handled.

5.8. Corollary. Suppose that v is p-superharmonic and locally bounded in Ω.Then v ∈ W 1,p

loc (Ω) and v is a weak supersolution:

∫

Ω

〈|∇v|p−2∇v,∇η〉dx ≥ 0

for all non-negative η ∈ C∞0 (Ω).

Proof: We have to justify the limit procedure

∫

Ω

〈|∇v|p−2∇v,∇η〉dx = limj→∞

∫

Ω

〈|∇vj|p−2∇vj,∇η〉dx ≥ 0

where the v′js are the approximants in Theorem 5.7. By their construction theysolve an obstacle problem and hence they are weak supersolutions (Theorem 5.2).In the case p ≥ 2, one can use the inequality

∣∣|∇v|p−2∇v − |∇vj|p−2∇vj∣∣ ≤

(p− 1)|∇v −∇vj|(|∇v|p−2 + |∇vj|p−2)

and then apply Holder’s inequality. In the case 1 < p ≤ 2 one has directly that

∣∣|∇v|p−2∇v − |∇vj|p−2∇vj∣∣ ≤ γ(p)|∇v −∇vj|p−1 .

The strong convergence in Theorem 5.7 is needed in both cases.

We make a discursion and consider the convergence of an increasing sequence ofp-harmonic functions.

5.9. Theorem. (Harnack’s convergence theorem) Suppose that hj is p-harmonicand that

0 ≤ h1 ≤ h2 ≤ . . . , h = limhj

pointwise in Ω. Then, either h ≡ ∞ or h is a p-harmonic function in Ω.

43

Proof: Recall the Harnack inequality (Theorem 2.20)

hj(x) ≤ Chj(x0) , j = 1, 2, 3, . . .

valid for each x ∈ B(x0, r), when B(x0, 2r) ⊂⊂ Ω. The constant C is independentof the index j. If h(x0) < ∞ at some point x0, then h(x) < ∞ at each x ∈ Ω.This we can deduce using a suitable chain of balls. It also follows that h is locallybounded in this case.

The Caccioppoli estimate

∫

Br

|∇hj|pdx ≤ Cr−p∫

B2r

|hj|pdx ≤ Cr−p∫

B2r

|h|pdx

≤ c1Cprn−ph(x0)p

allows us to conclude that h ∈ W 1,ploc (Ω). Finally,

∫

Ω

〈|∇h|p−2∇h,∇η〉dx = limj→∞

∫

Ω

〈|∇hj|p−2∇hj,∇η〉dx = 0

for each η ∈ C∞0 (Ω) follows from a repetition of the corresponding argument in theproof of Theorem 5.7.

5.3. Infimal convolutions

Instead of using the approximation with solutions of obstacle problems, as in theprevious subsection, one can directly use so-called infimal convolutions. They inheritthe comparison principle. See [LMa].

Suppose that v is lower semicontinuous and bounded in Ω :

0 ≤ v(x) ≤ L

and define

(5.10) vε(x) = infy∈Ω

v(y) +

|x− y|22ε

, ε > 0.

Then

44

• vε(x) v(x) as ε→ 0+

• vε(x)− |x|2/2ε is locally concave in Rn

• vε is locally Lipschitz continuous in Ω

• The Sobolev gradient ∇vε exists and belongs to L∞loc(Ω)

• The second derivatives D2vε exist in Alexandrov’s sense. See Section ?

The assertion about the Sobolev derivatives follow from Rademacher’s theoremabout Lipschitz functions, see [EG]. A most remarkable feature is that the infi-mal convolutions preserve the property of p-superharmonicity.

5.11. Lemma. If v is a p-superharmonic function in Ω and 0 ≤ v ≤ L, the infimalconvolution vε is a p-superharmonic function in the open subset of Ω where

dist(x, ∂Ω) >√

2Lε.

Similarly, the local weak supersolutions of the p-Laplace equation are preserved.

Proof: We assume that v is p-superharmonic in

Ωε =x ∈ Ω| dist(x, ∂Ω) >

√2Lε

.

First, notice that for x in Ωε, the infimum is attained at some point y = x? com-prised in Ω. The possibility that x? escapes to the boundary of Ω is hindered by theinequalities

|x− x∗|22ε

≤ |x− x∗|2

2ε+ v(x∗) = vε(x) ≥ v(x) ≤ L,

|x− x∗| ≤√

2Lε < dist(x, ∂Ω).

This explains why the domain shrinks a little.

We have to verify the comparison principle for vε in an arbitrary subdomainD ⊂⊂ Ωε. Suppose that h ∈ C(D) is a p-harmonic function such that vε(x) ≥ h(x)on the boundary ∂D or, in other words,

|x− y|22ε

+ v(y) ≥ h(x) when x ∈ ∂D, y ∈ Ω.

45

Thus, writing y = x+ z, we have

w(x) ≡ v(x+ z) +|z|22ε≥ h(x), x ∈ ∂D

whenever z is a small fixed vector. But also w = w(x) is a p-superharmonic functionin Ωε. By the comparison principle, w(x) ≥ h(x) in the whole D. Given any point x0

in D, we may choose z = x?0−x0. This yields vε(x0) ≥ h(x0). Since x0 was arbitrary,we have verified that

vε(x) ≥ h(x) when x ∈ D.

This concludes the proof of the comprison principle.

The statement about (weak) p-supersolutions is immediate.

5.4. The Poisson modification

This subsection, based on [GLM], is devoted to a simple but useful auxiliarytool, generalizing Poisson’s formula in the linear case p = 2. The so-called Poissonmodification of a p-superharmonic function v is needed for instance in connexionwith Perron’s method. Given a regular subdomain D ⊂⊂ Ω it is defined as thefunction

V = P (v,D) =

v in Ω\Dh in D

where h is the p-harmonic function in D with boundary values v on ∂D. One verifieseasily that V ≤ v and that V is p-superharmonic, if the original v is continuous.Otherwise, the interpretation of h = v on ∂D requires some extra considerations.In the event that v is merely semicontinuous one goes via the approximants vj inTheorem 5.7 and defines

V = limVj = limP (vj, D)

where we have tacitly assumed that vj → v in the whole Ω (here this is no re-striction). Now we use the Harnack convergence theorem (Theorem 5.9) on thefunctions hj to conclude that the limit function h = limhj is p-harmonic in D.(Since hj ≤ vj ≤ v the case h ≡ ∞ is out of the question. Also the situation hj ≥ 0is easy to arrange by adding a constant to v.) With this h it is possible to verify thatV is p-superharmonic. It is the limit of an increasing sequence of p-superharmonicfunctions.

46

5.12. Proposition. Suppose that v is p-superharmonic in Ω and that D ⊂⊂ Ω.Then the Poisson modification V = P (v,D) is p-superharmonic in Ω, p-harmonicin D, and V ≤ v. Moreover, if v is locally bounded, then

∫

G

|∇V |pdx ≤∫

G

|∇v|pdx

for D ⊂ G ⊂⊂ Ω.

Proof: It remains to prove the minimization property. This follows from the ob-vious property ∫

G

|∇Vj|pdx ≤∫

G

|∇vj|pdx .

In fact, the case G = D is the relevant one.

5.5. Summability of unbounded p-superharmonic functions

We have seen that the so-called polar set

Ξ = x ∈ Ω| v(x) =∞

of a p-superharmonic function v cannot contain any open set (Proposition 5.4).Much more can be assured. Ξ is empty, when p > n, and it has Lebesgue measurezero is all cases. The key is to study the p-superharmonic functions

vk = vk(x) = minv(x), k , k = 1, 2, 3, . . . .

Since they are locally bounded, they satisfy the inequality

∫

Ω

〈|∇vk|p−2∇vk,∇η〉dx ≥ 0

for each non-negative η ∈ C∞0 (Ω) and so the estimates for weak supersolutions areavailable.

47

5.13. Theorem. If v is p-superharmonic in Ω, then

∫

D

|v|qdx <∞

whenever D ⊂⊂ Ω and 0 ≤ q < n(p− 1)/(n− p) in the case 1 n the function v is continuous.

Proof: Because the theorem is of a local nature, we may assume that v > 0 byadding a constant. Then also vk > 0.

First, let 1 < p < n. According to Corollary 3.18

∫

Br

vqkdx

1q

≤ C(p, n, q)ess infBr

vk

whenever q < n(p− 1)/(n− p) and B2r ⊂⊂ Ω. The constant is independent of theindex k. Since vk ≤ v we obtain

(∫

Br

vqdx

) 1q

≤ C(p, n, q)ess infBr

v

It remains to prove that

ess infBr

v <∞

This is postponed till Theorem 5.14, the proof of which does not rely upon thepresent section.

Next, consider the case p > n. Here the situation v(x) = ∞ for a.e. x will beexcluded without evoking Theorem 5.14. The estimate

∫

Ω

ζp|∇ log vk|pdx ≤( p

p− 1

)p∫

Ω

|∇ζ|pdx

in Lemma 2.14 yields, as usual,

‖∇ log vk‖Lp(Br) ≤ C1r(n−p)/p

if B2r ⊂ Ω. According to (3.1)

| log vk(x)− log vk(y)| ≤ C2|x− y|(p−n)/p‖∇ log vk‖Lp(Br)

48

when x, y ∈ Br. It follows that vk(x) ≤ Kvk(y) and

v(x) ≤ Kv(y)

when x, y ∈ Br and B2r ⊂ Ω; K = eC1C2 . Thus we have proved the Harnackinequality for v.

We can immediately conclude that v(x) <∞ at each point in Ω, because there isat least one such point. As we know, the Harnack inequality implies continuity. Infact, v ∈ Cα

loc(Ω).

Finally, we have the borderline case p = n. It requires some special considerations.We omit the proof that v ∈ Lqloc(Ω) for each q <∞.

Remark. The previous theorem has been given a remarkable proof by T. Kilpelai-nen and J. Maly, cf [KM1]. The use of an ingenious test function makes it possibleto avoid the Moser iteration.

5.6. About pointwise behaviour

Although we know that v <∞ in a dense subset, the conclusion that ess inf v <∞requires some additional considerations. We will prove a result about pointwisebehaviour from which this follows. In order to appreciate the following investigationwe should be aware of that in the linear case p = 2 there exists a superharmonicfunction v defined in Rn such that v(x) = +∞ when all the coordinates of x arerational numbers, yet v < ∞ a.e.. (Actually, the polar set contains more points,since it has to be a Gδ-set.) The example is

v(x) =∑

q

cq|x− q|n−2

,

where the cq > 0 are chosen to create convergence. It is astonishing that this functionhas Sobolev derivatives! – A similar ”monster” can be constructed for 1 < p < n.

Recall that a p-superharmonic function v is lower semicontinuous. Thus

v(x) ≤ lim infy→x

v(y) ≤ ess lim infy→x

v(y)

at each point x ∈ Ω. ”Essential limes inferior” means that any set of n-dimensionalLebesgue measure zero can be neglected, when limes inferior is calculated. Thedefinition is given in [Brelot, II.5]. In fact, the reverse inequality also holds.

49

5.14. Theorem. If v is p-superharmonic in Ω, then


v(y)

at each point x in Ω.

The following lemma is the main step in the proof. A pedantic formulation cannotbe avoided.

5.15. Lemma. Suppose that v is p-superharmonic in Ω. If v(x) ≤ λ at each pointx in Ω and if v(x) = λ for a.e. x in Ω, then v(x) = λ at each x in Ω.

Proof: The idea is that v is its own Poisson modification and for a continuousfunction the theorem is obvious. Therefore fix a regular subdomain D ⊂⊂ Ω andconsider the Poisson modification V = P (v,D). We have

V ≤ v ≤ λ

everywhere. We claim that V = λ at each point in D. Since v is locally bounded itis a weak supersolution and as such it belongs to W 1,p

loc (Ω). According to Proposition5.12 ∫

G

|∇V |pdx ≤∫

G

|∇v|pdx =

∫

G

|∇λ|pdx = 0

for D ⊂ G ⊂⊂ Ω. Hence ∇V = 0 and so V is constant in G. It follows that V = λa.e. in G. But in D the function V is p-harmonic. It follows that V (x) = λ at eachpoint x in D. Since D was arbitrary, the theorem follows.

5.16. Lemma. If v is p-superharmonic in Ω and if v(x) > λ for a.e. x in Ω, thenv(x) ≥ λ for every x in Ω.

Proof: If λ = −∞ there is nothing to prove. Applying Lemma 5.15 to the p-superharmonic function defined by

minv(x), λ

we obtain the result in the case λ > −∞.

50

Proof of Theorem 5.14: Fix any x in Ω. We must show that

λ = ess lim infy→x

v(y) ≤ v(x) .

Given any ε > 0, there is a radius δ > 0 such that v(y) > λ− ε for a.e. y ∈ B(x, δ),where δ is small enough. By Lemma 5.16 v(y) ≥ λ − ε for each y ∈ B(x, δ). Inparticular, v(x) ≥ λ − ε. Because ε > 0 was arbitrary, we have established thatλ ≤ v(x).

5.7. Summability of the gradient

We have seen that locally bounded p-superharmonic functions are of class W 1,ploc (Ω).

They have first order Sobolev derivatives. For unbounded functions the summabilityexponent p has to be decreased, but it is important that the exponent can be taken≥ p− 1. The following fascinating theorem is easy to prove at this stage.

5.17. Theorem. Suppose that v is a p-superharmonic function defined in thedomain Ω in Rn, p > 2− 1

n. Then the Sobolev derivative

∇v =

(∂v

∂x1

, . . . ,∂v

∂xn

)

exists an the local summability result

∫

D

|∇v|qdx <∞ , D ⊂⊂ Ω ,

holds whenever 0 < q < n(p − 1)/(n − 1) in the case 1 n. Furthermore,

∫

Ω

〈|∇v|p−2∇v,∇φ〉 dx ≥ 0

whenever φ ≥ 0 φ ∈ C∞0 (Ω).

Remark. The fundamental solution

|x|(p−n)/(p−1) (p < n) , − log |x| (p = n)

51

shows that the exponent q is sharp. – The case p = 2 can be read off from the Rieszrepresentation formula. – The restriction p > 2 − 1

nis not essential but guarantees

that one can take q ≥ 1. The interpretation of ∇v would demand some care if1 < p ≤ 2− 1

n.

Proof: Suppose first that v ≥ 1. Fix D ⊂⊂ Ω and q < n(p − 1)/(n − 1). Thecutoff functions

vk = minv, k , k = 1, 2, 3, . . . ,

are bounded p-superharmonic functions and by Corollary 5.8 they are weak super-solutions. Use the test function η = ζpv−αk , α > 0, in the equation

∫

Ω

〈|∇vk|p−2∇vk,∇η〉dx ≥ 0

to obtain ∫

Ω

ζpv−1−αk |∇vk|pdx ≤

( pα

)p∫

Ω

vp−1−αk |∇ζ|pdx .

Here ζ ∈ C∞0 (Ω), 0 ≤ ζ ≤ 1 and ζ = 1 in D. By Holder’s inequality

∫

D

|∇vk|qdx =

∫

D

v(1+α)q/pk |v−(1+α)/p

k ∇vk|qdx

≤∫

D

v(1+α)q/(p−q)k dx

1− qp∫

D

v−1−αk |∇vk|pdx

qp

≤( pα

)q∫

D

v(1+α)q/(p−q)dx

1− qp∫

Ω

vp−1−α|∇ζ|pdx q

p

for any small α > 0. A calculation shows that

q

p− q <n(p− 1)

n− p

and hence we can fix α so that also

(1 + α)q

p− q <n(p− 1)

n− p .

52

Inspecting the exponents we find out that, in virtue of Theorem 5.13, the sequence‖∇vk‖Lq(D), k = 1, 2, 3, . . . is bounded. A standard compactness argument showsthat ∇v exists in D and

∫

D

|∇v|qdx ≤ limk→∞

∫

D

|∇vk|qdx .

Since D was arbitrary, we conclude that v ∈ W 1,qloc (Ω).

Finally, the restriction v ≥ 1 is locally removed by adding a constant to v. Thisconcludes our proof.

The equation with measure data. The equation

∇·(|∇v|p−2∇v) = −Cn,pδ

is satisfied by the fundamental solution

|x|(n−p)/(p−1) (p 6= n) log(|x|), (p = n)

in the distributional sense:

∫

Rn

〈|∇v|p−2∇v,∇φ〉 dx = φ(0)

whenever φ ∈ C∞0 (Rn). Thus Dirac’s delta is the right-hand side for the funda-mental solution. Theorem 5.17 enables us to produce a Radon measure for eachp-superharmonic function:

∇·(|∇v|p−2∇v) = −µ.

Recall that the summability exponent q > p− 1 for the gradient. This is decisive.

5.18. Theorem. Let v be a p-superharmonic function in Ω. Then there exists anon-negative Radon measure µ such that

∫

Ω

〈|∇v|p−2∇v,∇φ〉 dx =

∫

Ω

φ dµ

for all φ ∈ C∞0 (Rn).

53

Proof: We already know that v and ∇v belong to Lp−1loc (Ω). In order to use Riesz’s

Representation Theorem we define the linear functional

Λv : C∞0 (Ω)→ R,

Λv(φ) =

∫

Ω

〈|∇v|p−2∇v,∇φ〉 dx.

Now Λv(φ) ≥ 0 for φ ≥ 0 according to Theorem 5.17. Thus the functional is positiveand the existence of the Radon measure follows from Riesz’s theorem, cf. [EG, 1.8].

Some further results can be found in [KM1]. See also [KKT] and [KM].

54

6. Perron’s method

In 1923 O. Perron published a method for solving the Dirichlet boundary valueproblem

∆h = 0 in Ω ,

h = g on ∂Ω

and it is of interest, especially if ∂Ω or g are irregular. The same method workswith virtually no essential modifications for many other partial differential equationsobeying a comparison principle. We will treat it for the p-Laplace equation. Thep-superharmonic and p-subharmonic functions are the building blocks. This chapteris based on [GLM].

Suppose for simplicity that the domain Ω is bounded in Rn. Let g : ∂Ω →[−∞,∞] denote the desired boundary values. To begin with, g does not even haveto be a measurable function. In order to solve the boundary value problem, we willconstruct two functions, the upper Perron solution h and the lower Perron solutionh. Always, h ≤ h and the situation h = h is important; in this case we write h for

the common function h = h.

These functions have the following properties:

1) h ≤ h in Ω

2) h and h are p-harmonic functions, if they are finite

3) h = h, if g is continuous

4) If there exists a p-harmonic function h in Ω such that

limx→ξ

h(x) = g(ξ)

at each ξ ∈ ∂Ω, then h = h = h.

5) If, in addition, g ∈ W 1,p(Ω) and if h is the p-harmonic function with Sobolevboundary values h− g ∈ W 1,p

0 (Ω), then h = h = h.

There are more properties to list, but we stop here. Notice that 5) indicates thatthe Perron method is more general than the Hilbert space method.

We begin the construction by defining two classes of functions: the upper class Ugand the lower class Lg. The upper class Ug consists of all functions v : Ω→ (−∞,∞]such that

55

(i) v is p-superharmonic in Ω,

(ii) v is bounded below,

(iii) lim infx→ξ

v(x) ≥ g(ξ), when ξ ∈ ∂Ω.

The lower class Lg has a symmetric definition. We say that u ∈ Lg if

(i) u is p-subharmonic in Ω,

(ii) u is bounded above,

(iii) lim supu(x)x→ξ

5 g(ξ), when ξ ∈ ∂Ω.

It is a temptation to replace the third condition by lim v(x) = g(ξ), but that doesnot work. Neither is the requirement lim inf v(x) = g(ξ) a good one. The reason isthat we must be able to guarantee that the class is non-empty.

Notice that if v1, v2, . . . , vk ∈ Ug, then also the pointwise minimum

minv1, v2, . . . , vk

belongs to Ug. (A corresponding statement about maxu1, u2, . . . , uk holds forLg.) This is one of the main reasons for not assuming any differentiability of p-superharmonic functions in their definition. (However, when it comes to Perron’smethod it does no harm to assume continuity.) It is important that the Poissonmodification is possible: if v belongs to the class Ug, so does its Poisson modificationV ; recall subsection 5.4.

After these preliminaries we define at each point in Ω

the upper solution hg(x) = infv∈Ug

v(x) ,

the lower solution hg(x) = sup

u∈Lgu(x) .

Often, the subscript g is omitted. Thus we write h for hg. Before going further, letus examine an example for Laplace’s equation.

Example. Let Ω denote the punctured unit disc 0 < r < 1, r =√x2 + y2, in the

xy-plane. The boundary consists of a circle and a point (the origin). We prescribethe (continuous) boundary values

g(0, 0) = 1; g = 0 when r = 1 .

56

We have

0 ≤ h(x, y) ≤ h(x, y) ≤ ε log(1/√x2 + y2)

for ε > 0, because 0 ∈ Lg and ε log(1/r) ∈ Ug and always h ≤ h. Letting ε→ 0, weobtain

h = h = 0 .

Although the Perron solutions coincide, they take the wrong boundary values at theorigin! In fact, the harmonic function sought for does not exists, not with boundaryvalues in the classical sense. However, h− g ∈ W 1,2

0 (Ω) if g is smoothly extended.

A similar reasoning applies to the p-Laplace equation in a punctured ball in Rn,when 1 n the solution is

1− |x|(p−n)/(p−1)

and now it attains the right boundary values.

The next theorem is fundamental.

6.1. Theorem. The function h satisfies one of the conditions:

(i) h is p-harmonic in Ω,

(ii) h ≡ ∞ in Ω,

(iii) h ≡ −∞ in Ω.

A similar result holds for h.

The cases (ii) and (iii) require a lot of pedantic attention in the proof. For asuccinct presentation we assume from now on that

(6.2) m ≤ g(ξ) ≤M, when ξ ∈ ∂Ω .

Now the constants m and M belong to Lg and Ug respectively. Thus m ≤ h ≤ h ≤M . If v ∈ Ug, so does the cut function minv,M. Cutting off all functions, we mayassume that every function in sight takes values only in the interval [m,M ]. Theproof of the theorem relies on a lemma.

6.3. Lemma. If g is bounded, hg and hg

are continuous in Ω.

57

Proof: Let x0 ∈ Ω and B(x0, R) ⊂⊂ Ω. Given ε > 0 we will find a radius r > 0such that

|h(x1)− h(x2)| < 2ε when x1, x2 ∈ B(x0, r).

Suppose that x1, x2 ∈ B(x0, r). We can find functions vi ∈ U such that

limi→∞

vi(x1) = h(x1), limi→∞

vi(x2) = h(x2) .

Indeed, if v1i (x1)→ h(x1) and v2

i (x2)→ h(x2) we can use vi = minv1i , v

2i . Consider

the Poisson modifications

Vi = P (vi, B(x0, R)) .

It is decisive that Vi ∈ U. By Proposition 5.12 h ≤ Vi ≤ vi in Ω. Take i so large that

vi(x1) < h(x1) + ε , vi(x2) < h(x2) + ε .

It follows thath(x2)− h(x1) < Vi(x2)− Vi(x1) + ε

≤ oscB(x0,r)

Vi + ε .

Recall that Vi is p-harmonic in B(x0, R). The Holder continuity (Theorem 2.19)yields

oscB(x0,r)

Vi ≤ L( rR

)αosc

B(x0,R)Vi ≤ L

( rR

)α(M −m)

when 0 < r < R/2. Thus

h(x2)− h(x1) < ε+ ε = 2ε

when r is small enough. By symmetry, h(x1) − h(x2) < 2ε. The continuity of hfollows.

A similar proof goes for h.

Proof: of Theorem 6.1. We claim that h is a solution, having assumed (6.2) forsimplicity. Let q1, q2, . . . , qν , . . . be the rational points in Ω. We will first constructa sequence of functions in the upper class U converging to h at the rational points.Given qν we can find vν1 , v

ν2 , . . . in U such that

h(qν) ≤ vνi (qν) < h(qν) +1

i, i = 1, 2, 3, . . . .

58

Define

wi = minv11, v

12, . . . , v

1i , v

21, v

22, . . . , v

2i , . . . , v

i1, v

i2, . . . , v

ii

Then wi ∈ U, w1 ≥ w2 ≥ . . . and

h(qν) ≤ wi(qν) ≤ vνi (qν) when i ≥ ν .

Hence limwi(qν) = h(qν) at each rational point, as desired.

Suppose that B ⊂⊂ Ω and consider the Poisson modification

Wi = P (wi, B) .

Since also Wi ∈ U, we have

h ≤ Wi ≤ wi .

Thus limWi(qν) = h(qν) at the rational points. In other words, Wi is better than wi.We also conclude that W1 ≥ W2 ≥ W3 ≥ . . . . According to Harnack’s convergencetheorem (Theorem 5.9)

W = limi→∞

Wi

is p-harmonic in B. By the construction W ≥ h and W (qν) = h(qν) at the rationalpoints. We have two continuous functions, the p-harmonic W and h (Lemma 6.3),that coincide in a dense subset. Then they coincide everywhere. The conclusion isthat in B we have h = the p-harmonic function W . Thus h is p-harmonic in B. Itfollows that h is p-harmonic also in Ω.

A similar proof applies to h.

We have learned that the Perron solutions are p-harmonic functions, if they takefinite values. Always

−∞ ≤ h ≤ h ≤ ∞

but the situation h 6= h is possible. When h = h we denote the common functionwith h.

6.4. Theorem. (Wiener’s resolutivity theorem). Suppose that g : ∂Ω → R iscontinuous. Then h

g= hg in Ω.

59

Proof: Our proof is taken from [LM]. For the proof we need to know that there isan exhaustion of Ω with regular domains Dj ⊂⊂ Ω,

Ω =∞⋃

j=1

Dj, D1 ⊂ D2 ⊂ . . .

The domain Dj can be constructed as a union of cubes or as a domain with a smoothboundary.

We first do a reduction. If we can prove the theorem for smooth g’s, we are done.Indeed, given ε > 0 there is a smooth ϕ such that

ϕ(ξ)− ε < g(ξ) < ϕ(ξ) + ε , when ξ ∈ ∂Ω .

Thus,

hϕ − ε ≤ hϕ−ε ≤ hg≤ hg ≤ hϕ+ε = hϕ + ε ,

if hϕ

= hϕ. Since ε > 0 was arbitrary, we conclude that hg

= hg. Thus we can

assume that g ∈ C∞(Rn). What we need is only g ∈ W 1,p(Ω) ∩ C(Ω).

The proof, after the reduction to the situation g ∈ C∞(Rn), relies on the unique-ness of the solution to the Dirichlet problem with boundary values in Sobolev’s sense.In virtue of Theorem 2.16 there is a unique p-harmonic function h ∈ C(Ω)∩W 1,p(Ω)in Ω with boundary values h − g ∈ W 1,p

0 (Ω). Nothing has to be assumed aboutthe domain Ω, except that it is bounded, of course. We claim that h ≥ h andh ≤ h, which implies the desired resolutivity h = h. To this end, let v denote thesolution to the obstacle problem with g acting as obstacle. See Theorem 5.5. Thenv − g ∈ W 1,p

0 (Ω) and v ≥ g in Ω. Since v is a weak supersolution, v ∈ Ug. (Thereason for introducing the auxiliary function v is that one cannot guarantee that hitself belongs to the upper class! However, the obstacle causes v ≥ g.)

Construct the sequence of Poisson modifications

V1 = P (v,D1), V2 = P (v,D2) = P (V1, D2) ,

V3 = P (v,D3) = P (V2, D3), . . .

Then V1 ≥ V2 ≥ V3 ≥ . . . and Vj ∈ Ug. Also vj − g ∈ W 1,p0 (Ω) and

(6.5)

∫

Ω

|∇Vj|pdx ≤∫

Ω

|∇v|pdx ≤∫

Ω

|∇g|pdx .

60

We have hg ≤ Vj. Using Harnack’s convergence theorem (Theorem 5.9) we see that

V = limj→∞

Vj

is p-harmonic in D1, in D2, . . . , and hence in Ω. But (6.5) and the fact that Vj−g ∈W 1,p

0 (Ω) shows that V − g ∈ W 1,p0 (Ω). Thus V solves the same problem as h. The

aforementioned uniqueness implies that V = h in Ω.

We have obtained the result

hg ≤ limVj = h ,

as desired. The inequality hg≥ h has a similar proof. The theorem follows.

As a byproduct of the proof we obtain the following.

6.6. Proposition. If g ∈ W 1,p(Ω) ∩ C(Ω), then the p-harmonic function withboundary values in Sobolev’s sense coincides with the Perron solution hg.

The question about at which boundary points the prescribed continuous boundaryvalues are attained (in the classical sense) can be restated in terms of so-calledbarriers, a kind of auxiliary functions. Let Ω be a bounded domain. We say thatξ ∈ ∂Ω is a regular boundary point, if

limx→ξ

hg(x) = g(ξ)

for all g ∈ C(∂Ω).

Remark. There is an equivalent definition of a regular boundary point ξ. Theequation

∆pu = −1

has a unique weak solution u ∈ W 1,p0 (Ω) ∩ C(Ω). The point ξ is regular if and only

iflimx→ξ

u(x) = 0 .

The advantage is that only one function is involved. The proof of the equivalenceof the definitions is difficult.

6.7. Definition. A point ξ0 ∈ ∂Ω has a barrier if there exists a function w : Ω→ Rsuch that

61

(i) w is p-superharmonic in Ω,

(ii) lim infx→ξ w(x) > 0 for all ξ 6= ξ0, ξ ∈ ∂Ω,

(iii) limx→ξ0 w(x) = 0.

The function w is called a barrier.

6.8. Theorem. Let Ω be a bounded domain. The point ξ0 ∈ ∂Ω is regular if andonly if there exists a barrier at ξ0.

Proof: The proof of that the existence of a barrier is sufficient for regularity iscompletely analogous to the classical proof. Let ε > 0 and M = sup |g|. We can usethe assumptions to find δ > 0 and λ > 0 such that

|g(ξ)− g(ξ0)| < ε, when |ξ − ξ0| < δ;

λw(x) ≥ 2M, when |x− ξ0| ≥ δ .

This has the consequence that the functions g(ξ0) + ε+λw(x) and g(ξ0)− ε−λw(x)belong to the classes Ug and Lg, respectively. Thus

g(ξ0)− ε− λw(x) ≤ hg(x) ≤ hg(x) ≤ g(ξ0) + ε+ λw(x)

or

|hg(x)− g(ξ0)| ≤ ε+ λw(x)

Since w(x) → 0 as x → ξ0, we obtain that hg(x) → g(ξ0) as x → ξ0. Thus ξ0 is aregular boundary point.

For the opposite direction we assume that ξ0 is regular. In order to construct thebarrier we take

g(x) = |x− ξ0|pp−1 .

An easy calculation shows that ∆pg(x) is a positive constant, when x 6= ξ0. Weconclude that g is p-subharmonic in Ω. Let hg denote the corresponding upperPerron solution, when g are the boundary values. By the comparison principlehg ≥ g in Ω. Because ξ0 is assumed to be a regular boundary point, we have

limx→ξ0

hg(x) = g(ξ0) = 0 .

Hence w = hg will do as a barrier.

62

Example. 1 < p ≤ n. Suppose that Ω satisfies the well-known exterior spherecondition. Then each boundary point is regular. For the construction of a barrierat ξ0 ∈ ∂Ω we assume that B(x0, R) ∩ Ω = ξ0. The function

w(x) =

|x−x0|∫

R

t−n−1p−1 dt

will do as a barrier.

Example. p > n. Without any hypothesis

w(x) = |x− ξ0|p−np−1

will do as a barrier at ξ0. Thus every boundary point of an arbitrary domain isregular, when p > n.

An immediate consequence of Theorem 6.6 is the following result, indicating thatthe more complement the domain has, the better the regularity is. If Ω1 ⊂ Ω2 and ifξ0 ∈ ∂Ω1 ∩ ∂Ω2, then, if ξ0 is regular with respect to Ω2, so it is with respect to Ω1.The reason is that the barrier for Ω2 is a barrier for Ω1.

The concept of a barrier is rather implicit in a general situation. A much moreadvanced characterization of the regular boundary points is the celebrated Wienercriterion, originally formulated for the Laplace equation in 1924 by N. Wiener. Heused the electrostatic capacity. We need the p-capacity.

The p-capacity of a closed set E ⊂⊂ Br is defined as

Capp(E,Br) = infζ

∫

Br

|∇ζ|pdx

where ζ ∈ C∞0 (Br), 0 ≤ ζ ≤ 1 and ζ = 1 in E. The Wiener criterion can now bestated.

6.9. Theorem. The point ξ0 ∈ ∂Ω is regular if and only if the integral

1∫

0

[Capp(B(ξ0, t) ∩ E,B(ξ0, 2t))

Capp(B(ξo, t), B(ξ0, 2t))

] 1p−1 dt

t=∞

diverges, where E = Rn\Ω.

63

The Wiener criterion with p was formulated in 1970 by V. Mazja. He proved thesufficiency, [Ma]. For the necessity, see [KM2]. The case p > n − 1 has a simplerproof, written down for p = n in [LM]. The proofs are too difficult to be given here.

One can say the following when p varies but the domain is kept fixed. The greaterp is, the better for regularity. If ξ0 is p1-regular, then ξ0 is p2-regular for all p2 ≥ p1.This deep result can be extracted from the Wiener criterion. The Wiener criterion isalso the fundament for the so-called Kellogg property: The irregular boundary pointsof a given domain form a set of zero p-capacity. Roughly speaking, this means thatthe huge majority of the boundary points is regular.

It would be nice to find simpler proofs when it comes to the Wiener criterion!

64

7. Some remarks in the complex plane

For elliptic partial differential equations it is often the case that in only twovariables the theory is much richer than in higher dimensions. Indeed, also thep-Laplace equation

(7.1)∂

∂x

(|∇u|p−2∂u

∂x

)+

∂

∂y

(|∇u|p−2∂u

∂y

)= 0

in two variables, x and y, exhibits an interesting structure, not known of in space. Itlives a life of its own in the plane! Among other things a remarkable generalizationof the Cauchy-Riemann equations is possible. The hodograph method can be usedto obtain many explicit solutions.

In the plane the advanced theory of quasiconformal mappings is available forequations of the type

∂

∂x

(%(|∇u|)∂u

∂x

)+

∂

∂y

(%(|∇u|)∂u

∂y

)= 0

and is described in the book ”Mathematical Aspects of Subsonic and Transonic GasDynamics” by L. Bers. The p-Laplace equation presents some difficulties at thecritical points (∇u = 0). It was shown by B. Bojarski and T. Iwaniec in 1983 that

f =∂u

∂x− i∂u

∂y(i2 = −1)

is a quasiregular mapping (=quasiconformal, except injective). The most importantconsequence is that the zeros of f , that is, the critical points of the p-harmonicfunction u, are isolated. Thus they are points, as the name suggests.

7.2. Theorem. (Bojarski-Iwaniec) Let u be a p-harmonic function in the domainΩ in the plane. Then the complex gradient f = ux − iuy is a quasiregular mapping,that is:

(i) f is continuous in Ω

(ii) ux, uy ∈ W 1,2loc (Ω)

(iii)∣∣∂f∂z

∣∣ ≤∣∣1− 2

p

∣∣∣∣∂f∂z

∣∣ a.e. in Ω.

65

Remark. It is essential that |1− 2/p| < 1. The notation

∂

∂z=

1

2

(∂

∂x− i ∂

∂y

),

∂

∂z=

1

2

(∂

∂x+ i

∂

∂y

)

is convenient. The proof is given in [BI1] where also the formula

∂f

∂z=

(1

p− 1

2

)(f

f

∂f

∂z+f

f

(∂f

∂z

))

is established.

By the general theory, the zeros of a quasiregular mapping are isolated, exceptwhen the mapping is identically zero. We infer that the critical points

S = (x, y)| ∇u(x, y) = 0

of a p-harmonic function u are isolated, except when the function is a constant.Outside the set S the function is real-analytic. According to a theory due to Y.Reshetnyak an elliptic partial differential equation is associated to a quasiregularmapping, cf [Re] and [BI2]. In the plane this equation is always a linear one. In thepresent case ux, uy and log |∇u| are solutions to the same linear equation. However,this equation depends on ∇u itself! A different approach to find an equation forlog |∇u| has been suggested by Alessandrini, cf [Al].

Next, let us consider a counterpart to the celebrated Cauchy-Riemann equations.If u is p-harmonic in a simply connected domain Ω, then there is a function v, uniqueup to a constant, such that

vx = −|∇u|p−2uy , vy = |∇u|p−2ux

or, equivalently,ux = |∇v|q−2vy , uy = −|∇v|q−2vx

in Ω. For smooth functions this is evident from (7.1) but the general case is harder.In particular, |∇u|p = |∇v|q and 1/p + 1/q = 1. The conjugate function v is q-harmonic in Ω, q being the conjugate exponent:

1

p+

1

q= 1 .

A most interesting property is that

〈∇u,∇v〉 = 0 .

66

Therefore the level curves of u and v are orthogonal to each other, apart from thesingular set S. A good example is

u+ iv =p− 1

p− 2

(|z|

p−2p−1 − 1

)+ i arg z ,

where z = x+ iy and Ω is the complex plane (with a slit from 0 to ∞). We refer to[AL] for this kind of function theory.

A lot of explicit examples are given in [A5]. The optimal regularity of a p-harmonicfunction in the plane has been determined by T. Iwaniec and J. Manfredi.

7.3. Theorem. (Iwaniec-Manfredi) Every p-harmonic function, p 6= 2, is of classCk,α

loc (Ω) ∩W k+2,qloc (Ω), where the integer k > 1 and the exponent α, 0 < α ≤ 1, are

determined by the formula

k + α = 1 +1

6

(1 +

1

p− 1+

√1 +

14

p− 1+

1

(p− 1)2

).

The summability exponent q is any number in the range

1 ≤ q <2

2− α .

Proof: The proof is based on a hodograph representation, see [IM].

Remark.

1) Notice that always u ∈ W 3,1loc (Ω). Therefore u has Sobolev derivatives of order

three.

2) As p → ∞, the above formula does not produce the correct regularity classfor the limit equation. The reason is subtle.

3) As p→ 1, k →∞. However ”1-harmonic functions” are not of class C∞.

There are several properties that have been established in the plane but, so far aswe know, not in space. A few of them are:

The Principle of Unique Continuation. Suppose that u is a p-harmonic func-tion in Ω and that u ≡ 0 in a ball B ⊂ Ω. Then u ≡ 0 in Ω.

The Strong Comparison Principle. Suppose that u and v are p-harmonic func-tions and that u ≤ v in Ω. If u(x0) = v(x0) at some point x0 ∈ Ω, then u ≡ vin Ω. – For a proof we refer to [M1].

67

8. The infinity Laplacian

The limit equation of the p-Laplace equation as p→∞ is a very fascinating one.In two variables it is the equation

u2xuxx + 2uxuyuxy + u2

yuyy = 0 ,

which was found in 1967 by G. Aronsson, cf [A1]. It provides the best Lipschitzextension of given boundary values and has applications in image processing. Itrequires the modern concept of viscosity solutions, originally developed for equationsof the first order (Hamilton-Jacobi equations). The equation has a connection toStochastic Game Theory.

The ∞-Laplace operator

∆∞u =n∑

i,j=1

∂u

∂xi

∂u

∂xj

∂2u

∂xi∂xj=

1

2〈∇u,∇|∇u|2〉

comes from the following consideration. Start with

∆pu = |∇u|p−4|∇u|2∆u+ (p− 2)∆∞u

= 0 ,

divide out the factor |∇u|p−4, and let p→∞ in

|∇u|2∆u

p− 2+ ∆∞u = 0 .

This leads to the equation∆∞u = 0 .

However, this derivation of the∞-Laplace equation leaves much to be desired. Nev-ertheless, the equation is the correct one.

For a finite p the equation ∆pu = 0 is the Euler-Lagrange equation for the varia-tional integral

‖∇u‖p =

∫

Ω

|∇u|pdx 1

p

.

Hence one may expect the equation ∆∞u = 0 to be the Euler-Lagrange equation forthe ”functional”

‖∇u‖∞ = limp→∞‖∇u‖p = ess sup

x∈Ω|∇u(x)| .

68

Thus the minimax problemminu

maxx|∇u(x)|

is, as it were, involved here. All this can be done rigorously.

The equation has an interesting geometric interpretation, though valid only forrather smooth functions. To explain it via the ”gradient flow” we consider the curvex = x(t) = (x1(t), . . . , xn(t)) in Rn. Follow |∇u|2 along the curve. Differentiating|∇u(x(t))|2 we obtain

d

dt|∇u|2 = 2

∑ ∂u

∂xi

∂2u

∂xi∂xj

dxjdt.

We observe that if the curve is a solution of the dynamical system (the so-calledgradient flow)

dx

dt= ∇u(x(t))

we obtain, replacingdxjdt

by ∂u∂xj

,

d

dt|∇u|2 = 2∆∞u

taken along the curve. So far, u is arbitrary. Thus, if the original u was a solution of∆∞u = 0, we conclude that |∇u| is a constant along the curve. Since ∇u representsthe normal direction to the level surfaces of u, we have the following interpretation.Along a stream line |∇u| is constant. However, different stream lines usually havedifferent constants. This property is useful for applications to image processing.

The ∞-Laplacian also appears in an amusing formula. In the Taylor expansion

u(x+ h) = u(x) + 〈∇u(x), h〉+1

2〈h,D2u(x)h〉+ . . .

we take h = t∇u(x). We arrive at

u(x+ t∇u(x)) = u(x) + t|∇u(x)|2 +1

2t2∆∞u(x) + . . .

to our pleasure. The t2-term contains the ∞-Laplacian. The resulting formula

u(x+ t∇u(x))− 2u(x) + u(x− t∇u(x))

t2= ∆∞u(x) + . . .

can be utilized in a numerical scheme.

69

A few explicit solutions are

a√x2

1 + · · ·+ x2k + b (1 ≤ k ≤ n)

a1x1 + · · ·+ anxn + b

a1x4/31 + · · ·+ anx

4/3n (

∑a3j = 0)

as well as all angles in spherical coordinates like

arctan(x2

x1

), arctan

( x3√x2

1 + x22

).

Expressions in disjoint variables like

5√x2

1 + x22 + 3

√x2

3 + x24 + (x

4/35 − x4/3

6 )

can be added to the list. Finally, we mention the solutions of the eikonal equation|∇u|2 = 1. – Many examples in two variables are constructed in [A3].

The Dirichlet problem is to find a solution to

∆∞u = 0 in Ω ,

u = g on ∂Ω

in a bounded domain Ω. (In two variables the equation is formally classified as aparabolic one but the boundary values are prescribed as for elliptic equations!) Thedifficulty here is the concept of solutions, because u is not always of class C2. Wewill return to the concept of solutions later. Suppose now that g : ∂Ω → R is aLipschitz continuous function, that is

|g(ξ1)− g(ξ2)| ≤ L|ξ1 − ξ2|

when ξ1, ξ2 ∈ ∂Ω. We may extend g to be defined in Ω using one of the formulas

g(x) = maxξ∈∂Ω

(g(ξ)− L|x− ξ|

)or g(x) = min

ξ∈∂Ω

(g(ξ) + L|x− ξ|

).

The extended function has the same Lipschitz constant L. By Rademacher’s theorem∇g exists a.e. and |∇g| ≤ L. Therefore we may assume that g ∈ C(Ω) ∩W 1,∞(Ω).Now we want to construct the solution by letting p→∞. Let p > n. As we know,

70

there is a unique p-harmonic function up ∈ C(Ω)∩W 1,p(Ω) such that up = g on ∂Ω.(Since p > n, the regularity of Ω plays no role now.) We have

∫

Ω

|∇up|sdx 1

s

≤∫

Ω

|∇up|pdx 1

p

≤∫|∇g|pdx

1p

≤ |Ω|− 1pL

as soon as p > s. Using some compactness arguments we can conclude the following.

8.1. Proposition. There is a subsequence upk and a function u∞ ∈ C(Ω)∩W 1,∞(Ω)such that upk → u∞ uniformly in Ω and ∇upk ∇u∞ weakly in each Ls(Ω). Inparticular, u∞ = g on ∂Ω.

The so obtained u∞ is called a variational solution of the equation. Several ques-tions arise. Is u∞ unique or does it depend on the particular subsequence chosen?How is u∞ related to the limit equation ∆∞u = 0? Is it a ”solution”? At least itfollows directly from the construction that u∞ has a minimizing property:

8.2. Lemma. If D ⊂ Ω is a subdomain and if v ∈ C(D)∩W 1,∞(D) has boundaryvalues v = u∞ on ∂D, then

‖∇u∞‖L∞(D) ≤ ‖∇v‖L∞(D) .

In view of the mean value theorem in the differential calculus, the lemma saysthat the Lipschitz constant of u∞ cannot be locally improved. It is the best one.

Let us discuss the concept of solutions. In two variables the theorem below easilyenables one to conclude that there are ”solutions” not having second continuousderivatives.

8.3. Theorem. (Aronsson) Suppose that u ∈ C2(Ω) where Ω is a domain in Rn.If ∆∞u = 0 in Ω, then ∇u 6= 0 in Ω, except when u reduces to a constant.

Proof: See [A2] for n = 2. The cases n ≥ 2 are in [Y].

It turns out that the ∞-Laplace equation does not have a weak formulation withthe test functions under the integral sign. Indeed, multiplying the equation with atest function and integrating leads to

∫

Ω

η∆∞udx = 0 ,

71

an expression from which one cannot eliminate the second derivatives of u. Actually,integrations by part seem to make the situation worse!

The way out of this dead end is to use viscosity solutions as in [BDM]. It has tobe written in terms of viscosity supersolutions and subsolutions.

8.4 . Definition. We say that the lower semicontinuous function v is ∞-superharmonic in Ω, if whenever x0 ∈ Ω and ϕ ∈ C2(Ω) are such that

(i) ϕ(x0) = v(x0)

(ii) ϕ(x) < v(x), when x 6= x0

then we have ∆∞ϕ(x0) ≤ 0.

Notice that each point x0 needs its own family of test functions (which may beempty) and that ∆∞ϕ is evaluated only at the point of contact. By the infinitesimalcalculus ∇ϕ(x0) = ∇v(x0), provided that the latter exists at all. It is known thatv ∈ C(Ω) and that could have been incorporated in the definition.

The definition of an ∞-subharmonic function is similar. A function is definedto be ∞-harmonic if it is both ∞-superharmonic and ∞-subharmonic. Thus the∞-harmonic functions are the viscosity solutions of the equation.

Example. The function v(x) = 1 − |x| is ∞-harmonic when x 6= 0. It is ∞-superharmonic in Rn. At the origin there is no test function touching v from below.Thus there is no requirement to verify.

Example. The interesting function

x4/3 − y4/3

in two variables belongs to a family of solutions discovered by G. Aronsson [A3].The reader may verify that it is ∞-harmonic, indeed. This function belongs toC

1,1/3loc (R2) and to W

2,3/2−εloc (R2) for each ε > 0. It does not have second continuous

derivatives on the coordinate axes. See also [Sa].

We have now three concepts of solutions to deal with: classical solutions, varia-tional solutions and viscosity solutions. The inclusions

classical solutions ⊂ variational solutions ⊂ viscosity solutions

are not very difficult to prove. – In fact, all solutions are variational solutions. Thisfollows from R. Jensen’s remarkable uniqueness theorem.

72

8.5. Theorem. (Jensen) Let Ω be an arbitrary bounded domain. Given a Lipschitzcontinuous function g : ∂Ω → R there exists one and only one viscosity solutionu∞ ∈ C(Ω) of the equation ∆∞u∞ = 0 in Ω with boundary values u∞ = g on ∂Ω.

Proof: The existence is essentially Proposition 8.1. The uniqueness is proved in[J]. Jensen’s uniqueness proof uses several auxiliary equations and the method ofdoubling the variables. Another proof is given in [BB]. In [AS] a simple proof basedon comparison with cones is given.

We mention a characterization in terms of comparison with cones, cf [CEG] and[CZ]. For smooth functions the property follows from Taylor’s expansion.

8.6. Theorem. (Comparison with Cones) Let v be continuous in Ω. Then v is an∞-superharmonic function if and only if the comparison with cones holds: if D ⊂ Ωis any subdomain, a > 0 and x0 ∈ Rn\D, then v(ξ) ≤ a|ξ − x0| on ∂D implies thatv(x) ≤ a|x− x0| in D.

The apex x0 is outside the domain. The ∞-harmonic functions are preciselythose that obey the comparison with cones, both from above and below! Thisproperty has been used by O. Savin to prove that ∞-harmonic functions in theplane are continuously differentiable. In higher dimensions ∞-harmonic functionsare, of course, differentiable at almost every point. In fact, they are differentiable atevery point. The proof of this property in [ES] is of an unusual kind: no estimategiving continuity is produced.

We cannot resist mentioning that the function

V (x) =

∫|x− y|%(y)dy

is ∞-subharmonic for % ≥ 0, cf [CZ].6 It is a curious coincidence that the funda-mental solution of the biharmonic equation ∆∆u = −δ in three dimensional space(n = 3) is |x−x0|

8πso that ∆∆V (x) = −8π%(x). Given V, this tells us how to find a

suitable %.

Finally, we mention that the property of unique continuation does not hold. Thereis an example with a domain Ω and two∞-harmonic functions u1 and u2 in Ω, suchthat u1 = u2 in an open subset of Ω but u1 6≡ u2 in Ω. We do not know, whetherthis phenomenon can occur for u1 ≡ 0.

6This result due to Grandall and Zhang has the consequence that the ”mysterious inequality”∫∫∫ |x− c|2〈x− a, x− b〉 − 〈x− a, x− c〉〈x− b, x− c〉

|x− a||x− b||x− c|3 %(a)%(b)%(c)dadbdc ≥ 0

has to hold for all compactly supported densities %.

73

Tug-of-War. A marvellous connection between the ∞-Laplace Equation andStochastic Game Theory was discovered in 2009 by Y. Peres, O. Schramm, S.Sheffield, and D. Wilson. A mathematical game called Tug-of-War lead to theequation. We refer directly to [PSW] about this fascinating discovery. See [Ro] foran introductory account. We shall only give a sketch7, describing the game withoutproper mathematical terms, naturally without laying claim to any exactness.

Recall the Brownian Motion for the ordinary Laplace Equation. Consider abounded domain Ω, sufficiently regular for the Dirichlet Problem. Suppose thatthe boundary values

g(x) =

1, when x ∈ C0, when x ∈ ∂Ω \ C

are prescribed, where C ⊂ ∂Ω is closed. Let u denote the solution of ∆u = 0 inΩ with boundary values g (one may take the Perron solution). If a particle startsits Brownian motion at the point x ∈ Ω, then u(x) = the probability that theparticle (first) exits through C. In other words, the harmonic measure is related toBrownian motion. One may also consider more general boundary values. So muchabout Laplace’s Equation.

Let us now describe ”Tug-of-War”. Consider a game played by two players. Atoken is placed at the point x ∈ Ω. One player tries to move it so that it leaves thedomain via the boundary part C, the other one aims at the complement ∂Ω\C. Therules are:

• A fair coin is tossed.

• The player who wins the toss moves the token less than ε units in the mostfavourable direction.

• Both players play optimally.

• The game ends when the token hits the boundary ∂Ω.

Let us denote

uε(x) = the probability that the exit is through C.

Then the ‘Dynamic Programming Principle‘

uε(x) =1

2

(sup|e|<1

uε(x+ εe) + inf|e|<1

uε(x+ εe))

7This is taken from my lecture notes in [L4]

74

holds. The abbreviation is DPP. At a general point x there are usually only twooptimal directions, viz. ±∇uε(x)/|∇uε(x)|, one for each player. The directions areopposite, whence the name ’Tug-of-War’. The reader may consult [Ro]. As the stepsize ε goes to zero, one obtains a function

u(x) = limε→0

uε(x).

The sequence converges almost surely. The spectacular result is that the so obtainedfunction u is the solution to the ∞-Laplace Equation with boundary values g.

While the Brownian motion does not favour any direction, the Tug-of-War does.There is also a stochastic game for the p-harmonic Equation, though the rules aremore complicated, cf. [MPR3]. See [KMP] for p < 2.

75

9. Viscosity solutions

The modern theory of viscosity solutions was developed by Crandall, Evans,Jensen, Ishii, Lions, and others. First, it was designed for first order equations.Later, it was extended to second order equations. The solutions must obey a Com-parison Principle. We refer to [K] and [CIL] for some background.8 For the ∞-Laplacian the concept appeared in [BDM]. The viscosity solutions of the p-LaplaceEquation are the same as the p-harmonic functions, which was established in [JLM].We shall give a simple proof of this fact, based on [JJ], though only for p ≥ 2.

We begin with some infinitesimal calculus. Suppose that v : Ω → R is a givenfunction. Assume that φ ∈ C2(Ω) is touching v from below at some point x0 ∈ Ω :

φ(x0) = v(x0),

φ(x) < v(x) when x 6= x0.

If v happens to be smooth, then

∇φ(x0) = ∇v(x0), D2φ(x0) ≤ D2v(x0)

by the infinitesimal calculus. Here

D2φ(x0) =

(∂2φ(x0)

∂xi∂xj

)

n×n

is the Hessian matrix evaluated at the touching point x0. For symmetric real matriceswe use the ordering

X ≤ Y ⇐⇒n∑

i,j

xijξiξj ≤n∑

i,j

yijξiξj for all ξ = (ξ1, ξ2, · · · , ξn).

In particular, it follows that

∇v(x0) = ∇φ(x0),∂2v(x0)

∂x2i

≥ ∂2φ(x0)

∂x2i

, i = 1, 2, · · · , n,

and so

∆v(x0) ≥ ∆φ(x0), ∆∞v(x0) ≥ ∆∞φ(x0), ∆pv(x0) ≥ ∆pφ(x0).

8A few chapters from [K] are enough for our purpose.

76

If v is a p−superharmonic function and v ∈ C2(Ω), then ∆pv ≤ 0 and hence

∆pφ(x0) ≤ 0, if p > 2.

Notice carefully that the last inequality makes sense, even if v does not have anyderivatives! The function φ has. The next definition is based on this importantobservation.

9.1. Definition. Let p ≥ 2. We say that v : Ω → (−∞,∞] is a viscositysupersolution of the p-Laplace Equation, if

(i) v is finite in a dense subset

(ii) v is lower semicontinuous

(iii) whenever φ ∈ C2(Ω) touches v from below at the point x0 ∈ Ω, we have

∆pφ(x0) ≤ 0.

9.2. Definition. Let p ≥ 2. We say that u : Ω → [−∞,∞) is a viscositysubsolution of the p-Laplace Equation, if

(i) u is finite in a dense subset

(ii) u is upper semicontinuous

(iii) whenever ψ ∈ C2(Ω) touches u from above at the point x0 ∈ Ω, we have

∆pψ(x0) ≥ 0.

9.3. Definition. We say that h ∈ C(Ω) is a viscosity solution in Ω if it is both aviscosity subsolution and a viscosity supersolution.

Remark.

• The operator ∆pφ or ∆pψ is evaluated only at the point x0 of contact (thetouching point). Each point has its own family of test functions. If there isno test function touching at x0, then there is no requirement to be verified;the point passes for free. (It is not difficult to prove that the possible touchingpoints are dense in Ω.)

77

• The strict touching φ(x) < v(x), x 6= x0, can be replaced by φ(x) ≤ v(x).

• The viscosity solution9 is assumed to be continuous, a property that requireda proof for weak solutions!

If 1 < p < 2, the operator

∆pφ = |∇φ|p−2∆φ+ (p− 2)|∇φ|p−4∆∞φ

is undefined at the critical points, the points where ∇φ = 0. For this equation thetheory works, if one adds the requirement that ∇φ(x0) 6= 0 at the point of contact,cf. [JLM]. Thus there is no condition to be verified at the critical points. For p ≥ 2this amendment is of no influence, since ∆pφ(x0) = 0 now.

The definition is consistent. If v ∈ C2(Ω) is a viscosity supersolution then it isalso a weak supersolution. This is clear if p ≥ 2 since v itself then can serve as atest function.

9.4. Theorem. A p-superharmonic function is a viscosity supersolution.

Proof: Let v be a p-superharmonic function in the domain Ω. In order to provethat it satisfies Definition 9.1 we use an indirect reasoning. Assume that there is atest function φ touching v from below at the point x0 and satisfying ∆pφ(x0) > 0.By continuity

∆pφ(x) > 0 in B(x0, δ)

for some small δ. We may also assume that the touching is strict: v(x) > φ(x) whenx 6= x0. Now φ is a p-subharmonic function in B(x0, δ) and by adding the constant

m =1

2min

∂B(x0,δ)(v − φ) > 0

we obtain the p-subharmonic function φ(x) +m satisfying

φ(x) +m ≤ v(x) on ∂B(x0, δ).

By the Comparison Principle φ(x)+m ≤ v(x) in B(x0, δ), but this is a contradictionat the point x = x0.

9Here the word ”viscosity” is only a label. It comes from the method of vanishing viscosity.In our case one would replace ∆pu = 0 by ∆puε + ε∆uε = 0 and send ε to zero, so that theartificial viscosity term ε∆uε vanishes. So limuε = u is reached. Properly arranged, this is thesame concept.

78

Jets. We shall need an equivalent formulation for viscosity solutions, using so-called jets. We say that a pair (ξ,X), where ξ is a vector in Rn and X is a symmetricn× n-matrix, belongs to the subjet J2,−v(x) if

v(y) ≥ v(x) + 〈ξ, y − x〉+1

2〈y − x,X(y − x)〉+ o(|y − x|2)

as y → x. See [K], section 2.2, p. 17. If it so happens that v has continuous secondderivatives, then we can take ξ = ∇v(x), X = D2v(x) = the Hessian matrix. Inother words,

(∇v(x),D2v(x)) ∈ J2,−v(x), if v ∈ C2(Ω),

and the polynomial in y is Taylor’s. As we shall see later, it is important thatthe second Alexandrov derivatives will do as members of the subjet. We need anecessary10 condition for subjets.

9.5. Lemma. Let p ≥ 2. If ∆pv ≤ 0 in the viscosity sense, then

|ξ|p−2Trace(X) + (p− 2)|ξ|p−4〈ξ,X ξ〉 ≤ 0

when (ξ,X) ∈ J2,−v(x) and x ∈ Ω.

Infimal Convolutions. Assume that the function v is lower semicontinuous andthat

0 ≤ v(x) ≤ L.

Define again the infimal convolution

vε(x) = infy∈Ω

v(y) +

|y − x|22ε

.

and recall its properties in Section 5.3. We prove Lemma 5.11 again, but this timefor viscosity supersolutions.

9.6. Theorem. Assume that 0 ≤ v ≤ L in Ω. If v is a viscosity supersolution inΩ, so is the infimal convolution vε in the open set

Ωε =x ∈ Ω| dist(x, ∂Ω) >

√2Lε

.

10It is also sufficient provided that the closures J2,−v of the subjets are evoked. See [K]

79

Proof: If x belongs to Ωε then the infimum in the definition of vε is attained atsome point x∗ ∈ Ω. That x∗ cannot escape to the boundary was shown in the proofof Lemma 5.11.

Fix x0 ∈ Ωε. Assume that the test function φ touches vε from below at x0. Thenx∗0 is in Ω. Using

φ(x0) = vε(x0) =|x− x0|2

2ε+ v(x∗0)

φ(x) ≤ vε(x) ≤ |x− y|2

2ε+ v(y)

we can verify that the test function

ψ(x) = φ(x+ x0 − x∗0)− |x− x∗0|2

2ε

touches the original function v from below at x∗0. At this interior point ∆pψ(x∗0) ≤ 0by assumption. Because

∇ψ(x∗0) = ∇φ(x0), D2ψ(x∗0) = D2ψ(x0)

we also have∆pφ(x0) = ∆pψ(x∗0) ≤ 0.

Thus v fulfills the requirement in Definition 9.1.

We aim at proving that the viscosity supersolution v is a p-superharmonic func-tion. The key step is to achieve the following result in the shrunken domain Ωε forthe infimal convolution vε.

9.7. Proposition. If v is a viscosity supersolution and 0 ≤ v(x) ≤ L in Ω, thenvε is a weak supersolution in Ωε, i .e.

(9.8)

∫

Ωε

|∇vε|p−2〈∇vε,∇η〉dx ≥ 0

when η ∈ C∞0 (Ωε), η ≥ 0.

The proof in [JJ] relies on the fact that the infimal convolutions possess secondderivatives in the sense of Alexandrov, which can be used in the testing with subjets.We turn to the preparations for the proof, which we shall provide only for p ≥ 2.

80

9.9. Theorem (Alexandrov). A concave function f : Rn → R has second deriva-tives in the sense of Alexandrov: at a.e. point x there is a symmetric n× n-matrixA = A(x) such that the expansion

f(y) = f(x) + 〈∇f(x), y − x〉+1

2〈y − x,A(x)(y − x)〉+ o(|y − x|2)

is valid as y → x.

Proof: We refer to [EG], Section 6.4, pp. 242–245. Some details in [GZ], Lemma7.11, p. 199, are helpful to understand the singular part in the Lebesgue decompo-sition.

The first derivatives are Sobolev derivatives and ∇f ∈ L∞loc. The problem is withthe second ones. We use the notation D2f = A, although the Alexandrov deriva-tives may contain a singular Radon measure (for example, Dirac’s measure) so thatintegration by parts can fail. The proof in [EG] establishes that a.e. we have

A = limε→0

(D2(f ? ρε)

)

where ρε is Friedrich’s mollifier. This is an important ingredient in the proof. Alexan-drov’s Theorem is applicable to the concave function

(9.10) fε(x) = vε(x)− |x|2

2ε,

which is defined in the whole space (although the infimum is only over a subset).The quadratic part has no influence. Thus

(9.11) D2vε(x) = limσ→0

(D2(vε ? ρσ)(x)

)

a.e. in Rn.

We learned that the Alexandrov derivatives D2vε(x) exist at a.e. x. It followsthat (

∇vε(x),D2vε(x))∈ J2,−vε(x)

almost everywhere. By Lemma 9.5 the inequality

∆pvε(x) = |∇vε(x)|p−2∆vε(x)

+(p− 2)|∇vε(x)|p−4∆∞vε(x)

≤ 0

81

is valid a. e. in Ωε. Here ∆vε = Trace(D2vε).

We need a further mollification of the function fε in (9.10). Define the convolution

fε,j = fε ? ρεj , ρεj(x) =

Cnεnj

exp( −ε2jε2j−|x|2

), |x| < εj

0, otherwise.

The smooth functions vε,j = vε ? ρεj satisfy

∫

Ωε

〈|∇vε,j|p−2∇vε,j,∇η〉 dx =

∫

Ωε

η(−∆pvε,j

)dx

when η ∈ C∞0 (Ωε). Keep η ≥ 0. We want to extend this to vε. Notice that thefunctions vε,j are not viscosity supersolutions themselves. By (9.11)

limj→∞

D2vε,j(x) = D2vε(x)

almost everywhere. Thus also

limj→∞

∆pvε,j(x) = ∆pvε(x)

at almost every x in Ωε. The convolution has preserved the concavity so that D2fε,j ≤0. Since D2|x|2 = +2In, we obtain

D2vε,j ≤Inε, ∆vε,j ≤

n

ε,

where In = (δij) is the unit matrix. It is immediate that

|∇vε,j| ≤ ‖∇vε‖∞,Ωε = Cε

in the support of η.

Together these inequalities yield the bound

−∆pvε,j ≥ −Cp−2ε

n+ p− 2

ε

82

valid almost everywhere in the support of η. This lower bound justifies the use ofFatou’s Lemma below:

∫

Ωε

〈|∇vε|p−2∇vε,∇η〉 dx = limj→∞

∫

Ωε

〈|∇vε,j|p−2∇vε,j,∇η〉 dx

= limj→∞

∫

Ωε

η(−∆pvε,j

)dx

≥∫

Ωε

lim infj→∞

(η(−∆pvε,j

))dx

=

∫

Ωε

η(−∆pvε

)dx

≥∫

Ωε

η 0 dx = 0.

In the very last step we used the pointwise inequality −∆pvε ≥ 0, which we recallthat needed Alexandrov’s Theorem for its proof.

Now the proof of Proposition 9.7 is accomplished. It remains to pass to the limitas ε→ 0.

9.12. Theorem. Suppose that v is a bounded viscosity supersolution of ∆pv ≤ 0in Ω. Then the Sobolev gradient ∇v exists and v ∈ W 1,p

loc (Ω). Furthermore,

∫

Ω

〈|∇v|p−2∇v,∇η〉 dx ≥ 0

for all η ≥ 0, η ∈ C∞0 (Ω).

Proof: Choose ζ ≥ 0, ζ ∈ C∞0 (Ω). We can assume 0 ≤ v(x) ≤ L in the support ofζ. Use

η(x) = (L− vε(x))ζ(x)p

in Proposition 9.7. When ε is small enough for the inclusion supp(η) ⊂ Ωε weproceed as in the proof of Theorem 5.7. The Caccioppoli estimate for vε now reads

∫

Ω

ζp|∇vε|p dx ≤ (pL)p∫

Ω

|∇ζ|p dx.

83

By a compactness argument (at least for a subsequence)∇vε w weakly in Lploc(Ω).Since vε v we can identify w = ∇v. The conclusion is that ∇v exists and by weaklower semicontinuity of convex integrals also

∫

Ω

ζp|∇v|p dx ≤ (pL)p∫

Ω

|∇ζ|p dx.

To conclude the proof we show that the convergence ∇vε → ∇v is strong inLploc(Ω) so that one may pass to the limit under the integral sign in (9.8). Theprocedure is almost a repetition of the proof of Theorem 5.7. To this end, fix afunction θ ∈ C∞0 (Ω), 0 ≤ θ ≤ 1, and use the test function η = (v − vε)θ in theequation for vε. Then

∫

Ω

〈|∇v|p−2∇v − |∇vε|p−2∇vε,∇ ((v − vε)θ)〉 dx

≤∫

Ω

〈|∇v|p−2∇v,∇ ((v − vε)θ)〉 dx −→ 0,

where the last integral approaches zero due to the weak convergence.

We split the first integral into two parts

∫

Ω

θ 〈|∇v|p−2∇v − |∇vε|p−2∇vε,∇v −∇vε〉 dx

+

∫

Ω

(v − vε)〈|∇v|p−2∇v − |∇vε|p−2∇vε,∇θ〉 dx

and notice that the second integral approaches zero, because its absolute value isless than

‖v − vε‖Lp(D)

(‖∇v‖p−1

Lp(D) + ‖∇vε‖p−1Lp(D)

)‖∇θ‖,

where D contains the support of θ and ‖v−vε‖Lp(D) → 0. Recall also that ‖∇vε‖Lp(D)

is uniformly bounded.

Thus we have established that

limε→0

∫

Ω

θ 〈|∇v|p−2∇v − |∇vε|p−2∇vε,∇v −∇vε〉 dx −→ 0

at least for a subsequence. Now the strong convergence of the gradients follows.Indeed, the case p ≥ 2 follows from inequality (I) in Section 12. The singular case

84

p < 2 requires some work.11 Therefore we can proceed to the limit under the integralsign in (9.8).

Unbounded Viscosity Solutions. Let v(x) ≥ 0 be a viscosity solution. So isthe function

vL(x) = minv(x), L

cut at height L. The point is that vL is bounded. Now Theorem 9.12 is applicableand we conclude that vL is a p-superharmonic function. So is v itself as a limit of anincreasing sequence: v = lim vL. Indeed, to verify the Comparison Principle for v, wetake D ⊂⊂ Ω and assume that h ∈ C(D) is a p-harmonic function with boundaryvalues h|∂D ≤ v|∂D. Take L > maxh. Then vL ≥ h on ∂D. By the ComparisonPrinciple, which is valid for vL, one has vL ≥ h in D. Since v ≥ vL, we conclude thatv ≥ h in D. This proves the Comparison Principle for v.

We assumed that v ≥ 0 but since the theorem is local, this is no restriction. Wehave now proven the converse of Theorem 9.4.

9.13. Theorem. A viscosity supersolution of ∆pv ≤ 0 is a p-superharmomicfunction.

The Theorem has an interesting consequence. A viscosity supersolution v has agradient ∇v in Sobolev’s sense and ∇v ∈ Lsloc for some exponent s described inTheorem 5.17. This gradient was not mentioned in Definition 9.1!

11

∫|∇v −∇vε|p dx

=

∫|∇v −∇vε|p

(1 + |∇v|2 + |∇vε|2

)p(p−2)/4(1 + |∇v|2 + |∇vε|2

)p(2−p)/4dx

≤∫|∇v −∇vε|2

(1 + |∇v|2 + |∇vε|2

)(p−2)/2 p

2∫ (

1 + |∇v|2 + |∇vε|2)p/2

dx1− p

2

≤

1p−1

∫〈|∇v|p−2∇v − |∇vε|p−2∇vε,∇v −∇vε〉 dx

p2∫ (

1 + |∇v|2 + |∇vε|2)p/2

dx 2−p

2

where inequality (VII) in Section 12 was used at the last step.

85

10. Asymptotic mean values

The celebrated Mean Value Property, discovered by Gauss for harmonic func-tions, has a sophisticated replacement for p-harmonic functions. Naturally, a nonlinear extra term is needed. Furthermore, the new formula is valid in a particularasymptotic (infinitesimal) sense. The mean values

∫

B(x,ε)

u(y) dy ≡ 1

|B(x, ε)|

∫

B(x,ε)

u(y) dy

are taken over balls with radii ε shrinking to 0. W. Blaschke observed in 1916 thata continuous function u is harmonic in Ω if and only if

u(x) =

∫

B(x,ε)

u(y) dy + o(ε2) as ε→ 0

when x ∈ Ω. Here ε−2o(ε2)→ 0. (We deliberately ignore the fact that here the errorterm o(ε2) ≡ 0 a posteriori.)

The Fundamental Asymptotic Formula

(10.1) u(x) =p− 2

p+ n

supB(x,ε)

u+ infB(x,ε)

u

2+

2 + n

p+ n

∫

B(x,ε)

u(y) dy + o(ε2)

was given by Manfredi, Parviainen, and Rossi in [MPR1].12 As we shall see, properlyinterpreted in the viscosity sense, it characterizes the p-harmonic functions. The firstterm counts for the nonlinearity and the second one (the mean value) is linear. Whenp =∞ one should read

(10.2) u(x) =1

2

(supB(x,ε)

u+ infB(x,ε)

u)

+ o(ε2).

For an interpretation in Stochastic Game Theory it is essential that the coefficientssum up to 1, i.e.

p− 2

p+ n+

2 + n

p+ n= 1.

They represent probabilities, at least for p > 2.

12This Section is based on [MPR1].

86

Smooth Functions. Let us first derive some formulas for sufficiently smooth func-tions, say φ ∈ C2(Ω). Integrating the Taylor formula

φ(y) = φ(x) +∇φ(x) · (y − x) +1

2

n∑

i,j=1

∂2φ(x)

∂xi∂xj(yi − xi)(yj − xj)

+o(|x− y|2)

with respect to y over B(x, ε) we get

∫

B(x,ε)

φ(y) dy = φ(x) + 0 +1

2

n∑

i,j=1

∂2φ(x)

∂xi∂xj

∫

B(x,ε)

(yi − xi)(yj − xj) dy + o(ε2).

By symmetric cancellation the integrals vanish for i 6= j, and for i = j we have

∫

B(x,ε)

(yi − xi)2dy =1

n

∫

B(x,ε)

|y − x|2dy =ε2

n+ 2.

Thus we have arrived at the asymptotic formula in the lemma below. (It is atruncated version of the Pizzetti formula from 1909.)

10.3. Lemma. Let φ ∈ C2(Ω). When x ∈ Ω

(10.4) φ(x) =

∫

B(x,ε)

φ(y) dy − ε2

n+ 2∆φ(x) + o(ε2)

as ε→ 0.

In the next lemma the critical points must be avoided. There the normalized∞-Laplacian appears:

∆♦∞φ ≡ |∇φ|−2∆∞φ =⟨D2φ

∇φ|∇φ| ,

∇φ|∇φ|

⟩.

Later the normalized p-Laplace operator

∆♦p φ ≡ |∇φ|2−p∆pφ = ∆φ + (p− 2)∆♦∞φ

will be useful.13

13Clarification: In some recent literature, careless use of the traditional symbol ∆p also for thenormalized operator creates confusion about the proper meaning! Often ∆N

p or ∆Gp are used to

distinguish the normalized operator.

87

10.5. Lemma. Let φ ∈ C2(Ω). If x ∈ Ω and ∇φ(x) 6= 0, then

(10.6) φ(x) =1

2

(maxB(x,ε)

φ+ minB(x,ε)

φ)− ∆∞φ(x)

2 |∇φ(x)|2 ε2 + o(ε2)

as ε→ 0.

Proof: Let x0 ∈ Ω with ∇φ(x0) 6= 0. Choose ε > 0 so small that

∇φ(y) 6= 0 when |y − x0| ≤ ε.

The extremal values are attained at points x on ∂B(x0, ε) where

(10.7) x = x0 ± ε∇φ(x)

|∇φ(x)| , |x− x0| = ε.

The plus sign corresponds to the maximum and the minus sign to the minimum. Tosee this, use Lagrange multipliers. They are approximatively opposite endpoints ofa diameter14. It follows from

φxj(y) = φxj(x0) + O(ε)

that

∇φ(y)

|∇φ(y)| =∇φ(x0)

|∇φ(x0)| + O(ε), when |y − x0| ≤ ε.

Let x, x∗ ∈ ∂B(x0, ε) be exactly the endpoints of a diameter, i.e.

x+ x∗

2= x0.

Add the two Taylor expansions

φ(y) = φ(x0) + 〈∇φ(x0), y − x0〉 +1

2

⟨D2φ(x0)(y − x0), y − x0

⟩

+ o(|y − x0|2)

for y = x and y = x∗. The first order gradient terms cancel so that

φ(x) + φ(x0) = 2φ(x0) +⟨D2φ(x0)(y − x0), y − x0

⟩+ o(|y − x0|2).

14The notation hides the dependence on ε, i.e. x = xε.

88

In the quadratic term we substitute

x− x0 = ±ε ∇φ(x)

|∇φ(x)| = ±ε ∇φ(x0)

|∇φ(x0)| + O(ε2).

It follows that

φ(x) + φ(x0) = 2φ(x0) + ε2∆♦∞φ(x0) + o(ε2),

where x+ x∗ = 2x0 and x satisfies (10.7).

First, let x be the maximum point:

φ(x) = maxB(x0,ε)

φ.

Then

maxB(x0,ε)

φ + minB(x0,ε)

φ =φ(x) + minB(x0,ε)

φ

≤φ(x) + φ(x∗)

≤ 2φ(x0) + ε2∆♦∞φ(x0) + o(ε2).

We can derive the opposite inequality by selecting x as the minimum point, sincenow

maxB(x0,ε)

φ + minB(x0,ε)

φ ≥ φ(x∗) + φ(x).

This proves formula (10.6) with x0 in the place of x.

10.8. Lemma. Let φ ∈ C2(Ω). If ∇φ(x) 6= 0 at the point x ∈ Ω, then theasymptotic expansion

φ(x) =p− 2

p+ n

maxB(x,ε)

φ+ minB(x,ε)

φ

2+

2 + n

p+ n

∫

B(x,ε)

φ(y) dy

− 1

2ε2∆♦p φ(x) + o(ε2)(10.9)

is valid as ε→ 0.

Proof: Combine the asymptotic formulas (10.4) and (10.6) in a suitable way tosee this.

89

Mean values in the viscosity sense. We can read off from formula (10.9) thata function u ∈ C2(Ω) such that ∇u 6= 0 in Ω is p-harmonic in Ω if and only if thefundamental asymptotic formula (10.1) holds in Ω. Unfortunately, in the presence ofcritical points (i.e. zeros of the gradient) the above calculation is not valid. Recallthat ∆pu = 0 in the viscosity sense if and only if

|∇u|2∆u + (p− 2)∆∞u = 0

in the viscosity sense. No testing is needed at the critical points of touching testfunctions. On the other hand, we know that the p-harmonic functions are exactlythe viscosity solutions of the p-Laplace Equation. This suggests to interpret also thefundamental asymptotic formula (10.1) in the viscosity sense.

10.10. Definition. A function u satisfies the formula

u(x) =p− 2

p+ n

maxB(x,ε)

u+ minB(x,ε)

u

2+

2 + n

p+ n

∫

B(x,ε)

u(y) dy + o(ε2)

in the viscosity sense, if the two conditions:

• If x0 ∈ Ω and if φ ∈ C2(Ω) touches u from below at x0, then

φ(x) ≥ p− 2

p+ n

maxB(x,ε)

φ+ minB(x,ε)

φ

2+

2 + n

p+ n

∫

B(x,ε)

φ(y) dy + o(ε2)

as ε→ 0. Furthermore, if it so happens that ∇φ(x0) = 0 then the test functionmust obey the rule D2φ(x0) ≤ 0.

• If x0 ∈ Ω and if ψ ∈ C2(Ω) touches u from below at x0, then

ψ(x) ≤ p− 2

p+ n

maxB(x,ε)

ψ + minB(x,ε)

ψ

2+

2 + n

p+ n

∫

B(x,ε)

ψ(y) dy + o(ε2)

as ε→ 0. Furthermore, if it so happens that ∇ψ(x0) = 0 then the test functionmust obey the rule D2ψ(x0) ≥ 0.

hold.

90

Remark. The extra restrictions on the test functions at critical points are usedto conclude that

limy→x0

φ(y)− φ(x0)

|y − x0|2≤ 0, lim

y→x0

ψ(y)− ψ(x0)

|y − x0|2≥ 0.

10.11. Theorem (Manfredi - Parviainen -Rossi). Let u ∈ C(Ω). Then ∆pu = 0 inthe viscosity sense if and only if the Asymptotic Mean Value Formula (10.1) holdsin the viscosity sense.

Proof: Let us consider the case of subsolutions. Thus we assume that ψ ∈ C2(Ω)touches u from above at the point x0 ∈ Ω. If ∇ψ(x0) 6= 0, then the desired conclu-sions are contained in Lemma 10.8.

In the situation ∇ψ(x0) = 0 we proceed as follows. First, we assume that ∆pu ≥ 0in the viscosity sense. We have to verify that

(10.12) ψ(x0) ≤ p− 2

p+ n

maxB(x0,ε)

ψ + minB(x0,ε)

ψ

2+

2 + n

p+ n

∫

B(x0,ε)

ψ(y) dy + o(ε2).

Now only the situation D2ψ(x0) ≥ 0 is permissible. In particular, ∆ψ(x0) ≥ 0 andhence

(10.13) ψ(x0) ≤∫

B(x0,ε)

ψ(y) dy + o(ε2)

by (10.4). The extra condition

limy→x0

ψ(y)− ψ(x0)

|y − x0|2≥ 0

is at our disposal. Denoting

ψ(xε) = min|y−x0|≤ε

ψ(y)

we have

lim infε→0

1

ε2

1

2

(maxB(x0,ε)

ψ + minB(x0,ε)

ψ)− ψ(x0)

= lim infε→0

1

ε2

1

2

(maxB(x0,ε)

ψ − ψ(x0))

+1

2

(minB(x0,ε)

ψ − ψ(x0))

≥ lim infε→0

1

ε2

1

2

(maxB(x0,ε)

ψ − ψ(x0))

=1

2lim infε→0

(ψ(xε)− ψ(x0)

|xε − x0|2)( |xε − x0|2

ε2

)

≥ 0

91

since |xε − x0|2/ε2 ≤ 1. This shows that

(10.14) ψ(x0) ≤ 1

2

(maxB(x0,ε)

ψ + minB(x0,ε)

ψ)

+ o(ε2).

Combining (10.13) and (10.14) we see that (10.12) is valid, as desired.

Second, assume that (10.12) holds. But the mere fact that ∇ψ(x0) = 0 meansthat there is nothing to prove in Definition 9.2 concerning ∆pψ(x0) ≤ 0, sincecritical contact points of the test function pass for free. (Therefore the AsymptoticMean Value Formula was not even used at this step.) This concludes the case ofsubsolutions.

Finally, the case when the test functions touch from below is similar.

Comments. In the case p =∞ the Asymptotic Mean Value Formula is not validpointwise. The example with the ∞-harmonic function

u(x, y) = x43 − y 4

3

exhibits this fact, see [MPR1] for the calculations. However, in the cases 1 < p <∞the Asymptotic Mean Value Formula holds pointwise in the plane (n = 2), cf. [LiM]and [ALl]. The proofs in the plane are based on the hodograph method. To the bestof my knowledge the situation in higher dimensions n ≥ 2 is an open problem.

Dynamic Programming Principle. An interpretation in Stochastic Game The-ory comes from the Dynamic Programming Principle or DPP that is satisfied by thefunction of the game:

uε(x) =α

2

(supB(x0,ε)

uε + infB(x0,ε)

uε(x))

+ β

∫

B(x,ε)

uε(y) dy

where the ’probabilities’ are

α =p− 2

p+ n, β =

2 + n

p+ n.

The set up and how uε approaches a p-harmonic function is described in [MPR2].See also [KMP] for p < 2. Actually, the Stochastic Game was found first, then camethe Asymptotic Mean Value Formula.

92

11. Some open problems

As a challenge we mention some problems which, to the best of our knowledge,are open for the p-Laplace equation, when p 6= 2. In general, the situation in theplane is better understood than in higher dimensional spaces.

The Problem of Unique Continuation. Can two different p-harmonic functionscoincide in an open subset of their common domain of definition? The most pregnantversion is the following. Suppose that u = u(x1, x2, x3) is p-harmonic in R3 and thatu(x1, x2, x3) = 0 at each point in the lower half-space x3 < 0. Is u ≡ 0 then? Theplane case n = 2 is solved in [BI1]. In the extreme case p = ∞, the Principle ofUnique Continuation does not hold.

The Strong Comparison Principle. Suppose that u1 and u2 are p-harmonic func-tions satisfying u2 ≥ u1 in the domain Ω. If u2(x0) = u1(x0) at some interior pointx0 of Ω, does it follow that u2 ≡ u1? The plane case is solved in [M]. The StrongComparison Principle does not hold for p = ∞. One may add that, if one of thefunctions is identically zero, then this is the Strong Maximum Principle, which,indeed, is valid for 1 < p ≤ ∞.

Very Weak Solutions. Suppose that u ∈ W 1,p−1(Ω) and that

∫

Ω

〈|∇u|p−2∇u,∇ϕ〉dx = 0

for all ϕ ∈ C∞0 (Ω). Does this imply that u is (equivalent to) a p-harmonic function?Please, notice that the assumption

∫

Ω

|∇u|p−1dx <∞

with the exponent p − 1 instead of the natural exponent p is not strong enoughto allow test functions like ζpu. When p = 2 a stronger theorem (Weyl’s lemma)holds. T. Iwaniec and G. Martin have proved that the assumption u ∈ W 1,p−ε(Ω) issufficient for some small ε > 0, cf [I]. J. Lewis has given a simpler proof in [Le3].

Second Derivatives. For 1 2 ? The two dimensional case n = 2 was settled in Section 7. Therange 1 < p < 3 + 2

n−2, in which the Cordes condition is valid, was settled in [MW].

The C1-regularity for p =∞. Does an∞-harmonic function belong to C1loc? What

about C1,αloc ? Recently, O. Savin proved that in the plane all ∞-harmonic functions

93

have continuous gradients, cf [Sa]. An educated guess is that the optimal regularity

class is C1,1/3loc in the plane. The Holder exponent 1/3 for the gradient is attained

for the function x4/3− y4/3. In space Evans and Smart have proved in [ES] that the∞-harmonic functions are (totally) differentiable at every point. The proof providesno modulus of continuity.

The Asymptotic Mean Value Property. Does the Asymptotic Mean Value Propertyhold pointwise in space? It holds in the viscosity sense. In the plane it is known tobe valid pointwise, cf. [ALl] and [LiM].

High regularity as p is close to 1. In the plane case a p-harmonic function is ofsome differentiability class C

k(p)loc where k(p)→∞ as p→ 1 + 0. (However, solutions

of the limit equation are, in general, not of class C∞.). Does the regularity increasealso for n ≥ 3 when p→ 1 + 0?

There are many more problems. ”Luck and chance favours the prepared mind.”15

15Dans les champs de l’observation le hasard ne favorise que les esprits prepares. Louis Pasteur.

94

12. Inequalities for vectors

Some special inequalities are helpful in the study of the p-Laplace operator. Ex-pressions like

〈|∇v|p−2∇v − |∇u|p−2∇u,∇v −∇u〉

are ubiquitous and hence inequalities for

〈|b|p−2b− |a|p−2a, b− a〉

are needed, a and b denoting vectors in Rn. As expected, the cases p > 2 and p < 2are different. Let us begin with the identity

〈|b|p−2b− |a|p−2a, b− a〉 =|b|p−2 + |a|p−2

2|b− a|2

+(|b|p−2 − |a|p−2)(|b|2 − |a|2)

2,

which is easy to verify by a calculation. We can read off the following inequalities

(I)

〈|b|p−2b− |a|p−2a, b− a〉 ≥ 2−1(|b|p−2 + |a|p−2)|b− a|2

≥ 22−p|b− a|p ,

if p ≥ 2.

(II)

〈|b|p−2b− |a|p−2a, b− a〉 ≤ 1

2(|b|p−2 + |a|p−2)|b− a|2 ,

if p ≤ 2.

However, the second inequality in (I) cannot be reversed for p ≤ 2, as the first one,not even with a poorer constant than 22−p. Nevertheless, we have

(III)

〈|b|p−2b− |a|p−2a, b− a〉 ≤ γ(p)|b− a|p , p ≤ 2,

95

according to [Db2].16

The formula

|b|p−2b− |a|p−2a =

1∫

0

d

dt|a+ t(b− a)|p−2(a+ t(b− a))dt

yields

(IV)

|b|p−2b− |a|p−2a = (b− a)

1∫

0

|a+ t(b− a)|p−2dt

+ (p− 2)

1∫

0

|a+ t(b− a)|p−4〈a+ t(b− a), b− a〉(a+ t(b− a))dt

and consequently we have

〈|b|p−2b− |a|p−2a, b− a〉 = |b− a|21∫

0


+ (p− 2)

1∫

0

|a+ t(b− a)|p−4(〈a+ t(b− a), b− a〉

)2dt .

16By conjugation (III) follows from (I). To see this, let 1 2.By (I)

22−q|B −A|q ≤ 〈|B|q−2B − |A|q−2A,B −A〉.

Use a = |A|q−2A, A = |a|p−2a and the same for B to obtain

22−q∣∣|b|p−2b− |a|p−2a

∣∣q ≤ 〈|b|p−2b− |a|p−2a, b− a〉 ≤ |b− a|∣∣|b|p−2b− |a|p−2a

∣∣ .

It follows that

22−q∣∣|b|p−2b− |a|p−2a

∣∣q−1 ≤ |b− a|.

Thus, since (p− 1)(q − 1) = 1,

∣∣|b|p−2b− |a|p−2a∣∣ ≤ 22−p|b− a|p−1, 1 < p < 2.

This directly implies (III) with γ(p) = 22−p.

96

To proceed further, we notice that the last integral has the estimate

0 ≤1∫

0

|a+ t(b− a)|p−4(〈a+ t(b− a), b− a〉

)2dt

≤ |b− a|21∫

0

|a+ t(b− a)|p−2dt .

We begin with p ≥ 2. First we get

〈|b|p−2b− |a|p−2a, b− a〉 ≥ |b− a|21∫

0


and hence

∣∣|b|p−2b− |a|p−2a∣∣ ≥ |b− a|

1∫

0


by the Cauchy-Schwarz inequality. We also have

∣∣|b|p−2b− |a|p−2a∣∣ ≤ (p− 1)|b− a|

1∫

0

|a+ t(b− a)|p−2dt ,

where p ≥ 2. Continuing, we obtain replacing p by (p+ 2)/2:

∣∣|b| p−22 b− |a| p−2

2 a∣∣2 ≤

(p2

)2|b− a|2( 1∫

0

|a+ t(b− a)| p−22 dt

)2

≤(p

2

)2|b− a|21∫

0

|a+ t(b− a)|p−2dt ≤(p

2

)2〈|b|p−2b− |a|p−2a, b− a〉

We have arrived at

(V)∣∣|b| p−2

2 b− |a| p−22 a∣∣2 ≤

(p2

4

)〈|b|p−2b− |a|p−2a, b− a〉

if p ≥ 2

97

This is one of the inequalities used by Bojarski and Iwaniec (see Chapter 4). Wealso have, keeping p ≥ 2,

∣∣|b|p−2b− |a|p−2a∣∣ ≤ (p− 1)|b− a|

1∫

0


≤ (p− 1)|b− a|(|b| p−2

2 + |a| p−22

) 1∫

0

|a+ t(b− a)| p−22 dt

≤ (p− 1)(|b| p−2

2 + |a| p−22

)∣∣|b| p−22 b− |a| p−2

2 a∣∣ .

At the intermediate step |a+ t(b− a)|p−2 was factored and then

|a+ t(b− a)| p−22 ≤ |a| p−2

2 + |b| p−22

was used. We have arrived at

(VI) ∣∣|b|p−2b− |a|p−2a∣∣ ≤

(p− 1)(|b| p−2

2 + |a| p−22

)∣∣|b| p−22 b− |a| p−2

2 a∣∣ , if p ≥ 2

Also this inequality was used by Bojarski and Iwaniec in their differentiability proof.

Let us return to the formula below IV and consider now 1 < p ≤ 2. We obtain

〈|b|p−2b− |a|p−2a, b− a〉 ≥ (p− 1)|b− a|21∫

0

|a+ t(b− a)|p−2dt .

A simple estimation, taking into account that now p− 2 < 0, yields

(VII)

〈|b|p−2b− |a|p−2a, b− a〉 ≥ (p− 1)|b− a|2(1 + |a|2 + |b|2)p−22

if 1 ≤ p ≤ 2.

Recall

(VIII) ∣∣|b|p−2b− |a|p−2a∣∣ ≤ 22−p|b− a|p−1

if 1 ≤ p ≤ 2.

98

We remark that for many purposes the simple fact

〈|b|p−2b− |a|p−2a, b− a〉 > 0 , a 6= b ,

valid for all p, is enough.

Finally we just mention that the inequality

|b|p ≥ |a|p + p〈|a|p−2a, b− a〉 , p ≥ 1 ,

expressing the convexity of the function |x|p can be sharpened. In the case p ≥ 2the inequality

|b|p ≥ |a|p + p〈|a|p−2a, b− a〉+ C(p)|b− a|p

holds with a constant C(p) > 0. The case 1 < p < 2 requires a modification of thelast term.

99

Literature

[A1] G. Aronsson: Extension of functions satisfying Lipschitz conditions,Arkiv for Matematik 6 (1967), pp. 551–561.

[A2] G. Aronsson: On the partial differential equation u2xuxx + 2uxuyuxy +

u2yuyy = 0, Arkiv for Matematik 7 (1968), pp. 395–435.

[A3] G. Aronsson: On certain singular solutions of the partial differentialequation u2

xuxx + 2uxuyuxy + u2yuyy = 0, Manuscripta Mathematica 47

(1984), pp. 133–151.

[A4] G. Aronsson: Construction of singular solutions to the p-harmonic equa-tion and its limit equation for p =∞, Manuscripta Mathematica 56 (1986),pp. 135–158.

[A5] G. Aronsson: On certain p-harmonic functions in the plane, ManuscriptaMathematica 61 (1988), pp. 79–101.

[A6] G. Aronsson: Representation of a p-harmonic function near a criticalpoint in the plane, Manuscripta Mathematica 66 (1989), pp. 73–95.

[Al] G. Alessandrini: Critical points of solutions to the p-Laplace equationin dimension two. Bollettino della Unione Matematica Italiana, Sezione A,Serie VII, 1 (1987), pp. 239–246.

[AL] G. Aronsson & P. Lindqvist: On p-harmonic functions in the planeand their stream functions, Journal of Differential Equations 74 (1988), pp.157–178.

[ALl] A. Arroyo & J. Llorente: On the asymptotic mean value property forplanar p-harmonic functions, Proceedings of the American MathematicalSociety 144 (2016), pp. 3859–3868.

[AS] S. Armstrong & Ch. Smart: An easy proof of Jensen’s theorem onthe uniqueness of infinity harmonic functions, Calculus of Variations andPartial Differential Equations 37 (2010), pp. 381–384.

[B] M. Brelot: Elements de la Theorie Classique du Potentiel (2e edition),Paris 1961.

[Be] L. Bers: Mathematical Aspects of Subsonic and Transonic Gas Dynamics,Surveys in Applied Mathematics III, John Wiley & Sons, New York 1958.

100

[BB] G.Barles & J. Busca: Existence and comparison results for fully non-linear degenerate elliptic equations without zeroth-order term, Communi-cations on Partial Differential Equations 26 (2001), pp. 2323–2337.

[BDM] T. Bhattacharya, E. DiBenedetto & J. Manfredi: Limits as p→∞ of ∆pu = f and related problems, Rendiconti del Seminario MatematicoUniversita e Polytecnico di Torino (1989), pp. 15–68.

[BG] E. Bombieri & E. Giusti: Harnack’s inequality for elliptic differentialequations on minimal surfaces, Inventiones Mathematicae 15 (1972), pp.24–46.

[BI1] B. Bojarski & T. Iwaniec: p-harmonic equation and quasiregular map-pings. Partial Differential Equations (Warsaw 1984), pp. 25–38, BanachCenter Publications 19 (1987).

[BI2] B. Bojarski & T. Iwaniec: Analytical foundations of the theory ofquasiconformal mappings in Rn, Annales Academiae Scientiarum Fennicae,Ser.A.I., 8 (1983), pp. 257–324.

[CEG] M. Crandall, L. Evans & R. Gariepy: Optimal Lipschitz extensionsand the infinity Laplacian, Calculus of Variations and Partial DifferentialEquations 13 (2001), pp. 123–129.

[CL] H. Choe & J. Lewis: On the obstacle problem for quasilinear ellipticequations of p-Laplacian type, SIAM Journal on Mathematical Analysis 22(1991), pp. 623–638.

[CIL] M. Crandall, H. Ishii & P.-L. Lions: User’s guide to viscosity solu-tions of second order partial differential equations, Bulletin of the AmericanMathematical Society 27 (1992), pp. 1–67.

[CZ] M. Crandall & J. Zhang: Another way to say harmonic, Transactionsof the American Mathematical Society 355 (2003), 241–263.

[D] B. Dacorogna: Direct Methods in the Calculus of Variations, Springer,Heidelberg 1989.

[Db1] E. Di Benedetto: C1,α local regularity of weak solutions of degenerateelliptic equations, Nonlinear Analysis 7 (1983), pp. 827–850.

[Db2] E. DiBenedetto: Degenerate Parabolic Equations, Springer-Verlag, NewYork 1993.

[Dg] E. De Giorgi: Sulla differenziabilita e l’analiticita delle estremali degliintegrali multipli regolari. Mem. Accad. Sci. Torino (Classe di Sci.mat., fis.e nat.) (3) 3 (1957), 25–43.

101

[E] L. Evans: A new proof of local C1,α regularity for solutions of certaindegenerate elliptic P.D.E., Journal of Differential Equations 45 (1982), pp.356–373.

[EG] L. Evans & R. Gariepy: Measure Theory and Fine Properties of Func-tions, CRC Press, Boca Raton 1992.

[ES] L. Evans & Ch. Smart: Everywhere differentiability of infinity har-monic functions, Calculus of Variations and Partial Differential Equations42 (2011), pp. 289–299.

[F] M. Fuchs: p-harmonic obstacle problems. III. Boundary regularity, Annalidi Matematica Pura ed Applicata (Series IV) 156 (1990), pp. 159–180.

[GZ] R. Gariepy & W. Ziemer: Modern Real Analysis, PWS PublishingCompany, Boston 1994.

[G] E. Giusti: Metodi Diretti nel Calcolo delle Variazioni, UMI, Bologna 1994.English translation: Direct Methods in the Calculus of Variations, WorldScientfic Publ.Co., River Edge 2003.

[GLM] S. Granlund, P. Lindqvist & O. Martio: Note on the PWB-methodin the nonlinear case, Pacific Journal of Mathematics 125 (1986), pp. 381–395.

[GT] D. Gilbarg & N. Trudinger: Elliptic Partial Differential Equations ofSecond Order,2nd Edition, Springer, Berlin 1983.

[HL] Q. Han & F. Ling: Elliptic Partial Differential Equations (Courant Lec-ture Notes), New York 1997.

[I] T. Iwaniec & G. Martin: Quasiregular mappings in even dimensions,Acta Mathematica 170 (1993), pp. 29–81.

[IM] T. Iwaniec & J. Manfredi: Regularity of p-harmonic functions in theplane. Revista Matematica Iberoamericana 5 (1989), pp. 1–19.

[J] R. Jensen: Uniqueness of Lipschitz extensions minimizing the sup-normof the gradient, Archive for Rational Mechanics and Analysis 123 (1993)pp. 51–74.

[JLM] P. Juutinen, P. Lindqvist & J. Manfredi: On the equivalence ofviscosity solutions and weak solutions for a quasilinear equation, SIAMJournal on Mathematical Analysis 33 (2001), pp. 699–717.

102

[JN] F. John & L. Nirenberg: On functions of bounded mean oscillation,Communications on Pure and Applied Matmematics 14 (1961), pp. 415–426.

[Jo] J. Jost: Partielle Differentialgleichungen, Springer, Berlin 1998.

[JJ] V. Julin & P. Juutinen: A new proof of the equivalence of weak andviscosity solutions for the p-Laplace equation, Communications in PartialDifferential Equations 37 (2012), pp. 934–946.

[KMP] B. Kawohl, J. Manfredi & M. Parviainen: Solutions of nonlinearPDEs in the sense of averages, Journal de Mathematiques Pures et Ap-pliquees (9) 97 (2012), pp. 173–188.

[KKT] T. Kilpelainen, T. Kuusi & A. Tuhola-Kujanpaa: Superharmonicfunctions are locally renormalized solutions, Annales de l’institut HenriPoincare, Analyse Non Lineaire 28 (2011), pp. 775–795.

[KM1] T. Kilpelainen & J. Maly: Degenerate elliptic equations with measuredata and nonlinear potentials, Annali della Scuola Normale Superiore diPisa (Science Fisiche e Matematiche), Serie IV, 19 (1992), pp. 591–613.

[KM2] T. Kilpelainen & J. Maly: The Wiener test and potential estimates forquasilinear elliptic equations, Acta Mathematica 172 (1994), pp. 137–161.

[K] S. Koike: A Beginner’s Guide to the Theory of Viscosity Solutions (MSJMemoirs 13, Mathematical Society of Japan), Tokyo 2004.

[KM] T. Kuusi & G. Mingione: Linear potentials in nonlinear potential the-ory, Archive for Rational Mechanics and Analysis 207 (2013), pp. 215–246.

[Le1] J. Lewis: Capacitary functions in convex rings, Archive for Rational Me-chanics and Analysis 66 (1977), pp. 201–224.

[Le2] J. Lewis: Regularity of the derivatives of solutions to certain degenerateelliptic equations, Indiana University Math. J. 32 (1983), pp. 849–858.

[Le3] J. Lewis: On very weak solutions of certain elliptic systems, Communica-tions in Partial Differential Equations 18 (1993), pp. 1515–1537.

[L1] P. Lindqvist: On the growth of the solutions of the equationdiv(|∇u|p−2∇u) = 0 in n-dimensional space, Journal of Differential Equa-tions 58 (1985), pp. 307–317.

[L2] P. Lindqvist: On the definition and properties of p-superharmonic func-tions, Journal fur die reine und angwandte Mathematic (Crelles Journal)365 (1986), pp. 67–79.

103

[L3] P. Lindqvist: Regularity for the gradient of the solution to a nonlinearobstacle problem with degenerate ellipticity, Nonlinear Analysis 12 (1988),pp. 1245–1255.

[L4] P. Lindqvist: Notes on the Infinity Laplace Equation (SpringerBriefs inMathematics), Springer 2016.

[LMa] P. Lindqvist & J. Manfredi: Viscosity solutions of the evolutionary p-Laplace equation, Differential and Integral Equations 20 (2007), pp. 1303–1319.

[LiM] P. Lindqvist & J. Manfredi: On the mean value property of the p-Laplace equation in the plane, Proceedings of the American MathematicalSociety 144 (2016), pp. 143–149.

[LM] P. Lindqvist & O. Martio: Two theorems of N. Wiener for solutions ofquasilinear elliptic equations, Acta Mathematica 155 (1985), pp. 153–171.

[LU] O. Ladyzhenskaya & N. Uraltseva: Linear and Quasilinear EllipticEquations, Academic Press, New York 1968.

[M1] J. Manfredi: p-harmonic functions in the plane, Proceedings of theAmerican Mathematical Society 103 (1988), pp. 473–479.

[M2] J. Manfredi: Isolated singularities of p-harmonic functions in the plane,SIAM Journal on Mathematical Analysis 22 (1991), pp. 424–439.

[MPR1] J. Manfredi, M. Parviainen & J. Rossi: An asymptotic mean valuecharacterization for p-harmonic functions, Proceedings of the AmericanMathematical Society 138 (2010), pp. 881–889.

[MPR2] J. Manfredi, M. Parviainen & J. Rossi: On the definition and prop-erties of p-harmonious functions, Annali della Scuola Normale Superiore diPisa (Classe di Scienze), Serie V, 11 (2012), pp. 215–241.

[MPR3] J. Manfredi, M. Parviainen & J. Rossi: Dynamic programming prin-ciple fot tug-of-wae games with noise, ESAIM, Control, Optimisation andCalculus of Variations 18 (2012), pp. 81–90.

[MW] J. Manfredi & A. Weitsman: On the Fatou theorem for p-harmonicfunctions, Communications in Partial Differential Equations 13 (1988), pp.651–668.

[Ma] V. Maz’ja: On the continuity at a boundary point of solutions of quasilin-ear elliptic equations. Vestnik Leningradskogo Universiteta 13 (1970), pp.42-55. In Russian.

104

[Mo1] J. Moser: A new proof of De Giorgi’s theorem concerning the regularityproblem for elliptic differential equations. Communications on Pure andApplied Mathematics, 13 (1960), pp. 457-468.

[Mo2] J. Moser: On Harnack’s theorem for elliptic differential equations. Com-munications on Pure and Applied Mathematics (1961), pp. 577-591.

[MZ] J. Michel & P. Ziemer: Interior regularity for solutions to obstacleproblems, Nonlinear Analysis 10 (1986), pp. 1427-1448.

[PSW] Y. Peres, O. Schramm, S. Sheffield & D. Wilson: Tug-of-war andthe infinity Laplacian, Journal of the American Mathematical Society 22(2009), pp. 167-210.

[R] T. Rado: Subharmonic Functions, New York 1949.

[Re] J. Reshetnyak: Extremal properties of mappings with bounded distor-tion. Sibirskij Matematicheskij Zhurnal 10 (1969), pp.1300-1310. In Rus-sian.

[Ro] J. Rossi: Tug-of-war games and PDEs, Proceedings of the Royal Societyof Edinburgh, Section A Mathematics, 141 (2011), pp. 319-369.

[S] S. Sakaguchi: Coincidence sets in the obstacle problem for the p-harmonic operator, Proceedings of the American Mathematical Society 95(1985), pp. 382-386.

[Sa] O. Savin: C1 regularity for infinity harmonic functions in two dimensions,Archive for Rational Mechanics and Analysis 176 (2005), pp. 351-361.

[SC] L. Saloff-Coste: Aspects of Sobolev-Type Inequalities, London Math-ematical Society Lecture Note Series 289, Cambridge 2002.

[T1] N. Trudinger: On Harnack type inequalities and their application toquasilinear elliptic equations, Communications on Pure and Applied Math-ematics 20 (1967), pp. 721-747.

[T2] N. Trudinger: On the regularity of generalized solutions of linear, non-uniformly elliptic equations, Archive for Rational Mechanics and Analysis42 (1971), pp. 50–62.

[To] P. Tolksdorf: Regularity for a more general class of quasilinear ellipticequations, Journal of Differential Equations 51 (1984), pp. 126–150.

[Uh] K. Uhlenbeck: Regularity for a class of nonlinear elliptic systems, ActaMathematica 138 (1977), pp. 219–240.

105

[Ur] N. Ural’ceva: Degenerate quasilinear elliptic systems, Zap. Naucn. Sem.Leningrad. Otdel. Mat. Inst. Steklov 7 (1968), pp. 184–192. (In Russian).

[Wi] K.-O. Widman: Holder continuity of solutions of elliptic equations,Manuscripta Mathematica 5 (1971), pp. 299–308.

[Y] Y. Yu: A remark on C2 infinity-harmonic functions, Electronic Journal ofDifferential Equations 2006 no 122 (2006), pp. 1–4.

Department of Mathematical SciencesNorwegian University of Science and Technology, NO-7491 Trondheim, [email protected]

106

REP. UNIV. JYVASKYLA DEPT. MATH. STAT.

130. KEISALA, JUKKA, Existence and uniqueness of p(x)-harmonic functions for bounded andunbounded p(x). (56 pp.) 2011

131. SEPPALA, HEIKKI, Interpolation spaces with parameter functions and L2-approximations ofstochastic integrals. (18 pp.) 2011

132. TUHOLA-KUJANPAA, ANNA, On superharmonic functions and applications to Riccati typeequations. (17 pp.) 2012

133. JIANG, RENJIN, Optimal regularity of solutions to Poisson equations on metric measure spacesand an application. (13 pp.) 2012

134. TORMAKANGAS, TIMO, Simulation study on the properties of quantitative trait modelestimation in twin study design of normally distributed and discrete event-time phenotypevariables. (417 pp.) 2012

135. ZHANG, GUO, Liouville theorems for stationary flows of generalized Newtonian fluids. (14 pp.)2012

136. RAJALA, TUOMAS, Use of secondary structures in the analysis of spatial point patterns.(27 pp.) 2012

137. LAUKKARINEN, EIJA, On Malliavin calculus and approximation of stochastic integrals forLevy processes. (21 pp.) 2012

138. GUO, CHANGYU, Generalized quasidisks and the associated John domains. (17 pp.) 2013

139. AKKINEN, TUOMO, Mappings of finite distortion: Radial limits and boundary behavior.(14 pp.) 2014

140. ILMAVIRTA, JOONAS, On the broken ray transform. (37 pp.) 2014141. MIETTINEN, JARI, On statistical properties of blind source separation methods based on joint

diagonalization. (37 pp.) 2014142. TENGVALL, VILLE, Mappings of finite distortion: Mappings in the Sobolev space W 1,n−1 with

integrable inner distortion. (22 pp.) 2014143. BENEDICT, SITA, Hardy-Orlicz spaces of quasiconformal mappings and conformal densities.

(16 pp.) 2014144. OJALA, TUOMO, Thin and fat sets: Geometry of doubling measures in metric spaces. (19 pp.)

2014145. KARAK, NIJJWAL, Applications of chaining, Poincare and pointwise decay of measures.

(14 pp.) 2014

146. JYLHA, HEIKKI, On generalizations of Evans and Gangbo’s approximation method and L∞transport. (20 pp.) 2014

147. KAURANEN, AAPO, Space-filling, energy and moduli of continuity. (16 pp.) 2015148. YLINEN, JUHA, Decoupling on the Wiener space and variational estimates for BSDEs.

(45 pp.) 2015

149. KIRSILA, VILLE, Mappings of finite distortion on generalized manifolds. (14 pp.) 2015150. XIANG, CHANG-LIN, Asymptotic behaviors of solutions to quasilinear elliptic equations with

Hardy potential. (20 pp.) 2015151. ROSSI, EINO, Local structure of fractal sets: tangents and dimension. (16 pp.) 2015152. HELSKE, JOUNI, Prediction and interpolation of time series by state space models. (28 pp.)

2015153. REINIKAINEN, JAAKKO, Efficient design and modeling strategies for follow-up studies with

time-varying covariates. (36 pp.) 2015154. NUUTINEN, JUHO, Maximal operators and capacities in metric spaces. (22 pp.) 2016155. BRANDER, TOMMI, Calderon’s problem for p-Laplace type equations. (21 pp.) 2016

156. ARJE, JOHANNA, Improving statistical classification methods and ecological status assesmentfor river macroinvertebrates. (30 pp.) 2016

157. HELSKE, SATU, Statistical analysis of life sequence data. (40 pp.) 2016158. RUOSTEENOJA, EERO, Regularity properties of tug-of-war games and normalized equations.

(16 pp.) 2017159. ZHANG, YI, Planar Sobolev extension domains. (13 pp.) 2017160. YLITALO, ANNA-KAISA, Statistical inference for eye movement sequences using spatial and

spatio-temporal point processes. (45 pp.) 2017

ISBN 978-951-39-7119-9 (print)

ISBN 978-951-39-7120-5 (pdf)

ISSN 1457-8905

Peter Lindqvist: Notes on the p-Laplace equation (Second edition)

notes on the p-laplace equation (second edition)usually p 1. at the critical points ( r u = 0) the...

Documents