Maximum Principles for Inhomogeneous Elliptic Inequalities on Complete Riemannian Manifolds
Dimitri Mugnai, Patrizia Pucci ∗
Dipartimento di Matematica e Informatica
Universit`a degli Studi di Perugia , Via Vanvitelli 1, 06123 Perugia, Italy e-mails: mugnai@dmi.unipg.it, pucci@dmi.unipg.it
Received 24 July 2008 Communicated by Paul Rabinowitz
Abstract
We prove some maximum principle results for weak solutions of elliptic inequali- ties, possibly inhomogeneous, on complete Riemannian manifolds.
1991 Mathematics Subject Classification . Primary: 58 J 05, 35 B 50; Secondary: 35 B 45, 35 J 60 .
Key words . Elliptic inequalities on manifolds, Maximum Principles.
1 Introduction
In this paper we are concerned with weak solutions of differential inequalities on a complete Riemannian manifold M of dimension n. More precisely, our aim is to prove maximum principles for inequalities governed by operators which may be inhomogeneous.
Let us start with three examples of general interest: the first concerns the mean curvature operator, say with global growth p = 1, while the latter two involve the p–Laplace–Beltrami operator, that is ∆ p u :=div(|∇u| p−2 ∇u), p > 1, where ∇u is the Riemannian gradient of u on M. Take the mean curvature equation
div
à ∇u
p 1 + |∇u| 2
!
= nH(x)
∗
Research supported by the National Project Metodi Variazionali ed Equazioni Differenziali Non Lineari.
1
in a domain Ω ⊂ M. Putting A(ξ) = ξ ±p
1 + |ξ| 2 , after calculation, we find
hA(ξ), ξi = |ξ| + Ã
−|ξ| + |ξ| 2 p 1 + |ξ| 2
!
≥ |ξ| − s
5 √ 5 − 11
2 .
This means that it is not possible to control hA(ξ), ξi from below by a positive quantity, as it is usual and useful to do in obtaining some kind of maximum principle.
We also recall that such an operator is natural while studying general relativity (for example see [18]).
For the second prototype of inhomogeneous elliptic operator, we consider the intriguing inequality
∆ p u − b 1 ∆ p−1 u + b 2 |u| p−2 u + b ≥ 0, p > 2, (1.1) as well as the homogeneous soliton inequality (see, for example, [3])
∆ p u + b 1 ∆ p−1 u + b 2 |u| p−2 u + b ≥ 0, p > 2,
where in both cases b 1 > 0, b 2 , b ∈ R. Indeed, in (1.1) we have A(ξ) = |ξ| p−2 ξ − b 1 |ξ| p−3 ξ, so that
hA(ξ), ξi ≥ 1
p (|ξ| p − b p 1 ) .
For these problems the theory of the paper applies and general maximum prin- ciples still hold true for their weak solutions.
Our first results are, actually, semi–maximum principles, in the sense that we prove inequalities of the form
u ≤ Ckuk p
p ≥ 1. Indeed, we underline the fact that the mean curvature operator has linear growth, so that L 1 –norms are natural and expected. Secondly, from these semi–
maximum principles we derive classical maximum principles of the form u ≤ C,
with C independent of u, but depending only on the data of the problem, without growth conditions on the coefficients, as required in [4].
We are interested in maximum principles and for existence of weak solutions on compact Riemannian manifolds we refer to the recent survey [6] and the references therein, while for the noncompact case we recall [19].
The first results about maximum principles in the Euclidean setting go back to the sixties with Serrin ([21], based on earlier work for homogeneous linear equa- tions by Stampacchia [22], Maz’ya [11], Moser [12]), Gilbarg and Trudinger [7].
More recently, general results were proved in [15], [16], [20], and in [13] and [14] for
the Riemannian setting. Finally, for inhomogeneous inequalities, in [17] very gen-
eral results are given in the Euclidean case, and we take those results as inspiring
motivations for inequalities on manifolds. Actually, we extend to the Riemannian
setting the semi–maximum principles and the maximum principles proved in [17]
for inhomogeneous inequalities in the Euclidean case. For further comments on the results, we refer to Section 3, where we discuss some applications after stating the main theorems.
2 Preliminaries
In this section we introduce the main notation. From now on M denotes a smooth complete Riemannian n–manifold, with metric tensor g ∈ C ∞ (M , T ∗ M N
T ∗ M ).
Definition 2.1 The fibered product bundle of two bundles (E, π 1 , M ) and (F, π 2 , M ) is the manifold
E × M F = {(e, f ) ∈ E × F : π 1 (e) = π 2 (f )} , with the induced vector bundle structure.
In the sequel, Ω will be any regular domain of M and, for brevity, we write T Ω × Ω R in place of T Ω × Ω (Ω × R) to denote the fibered product bundle T Ω × Ω
(Ω × R). Of course T Ω × Ω (Ω × R) ∼ = T Ω × R, and in turn the notation is not ambiguous. In analogy with the Euclidean case, elements of T Ω × Ω R are denoted by (x, z, ξ), where (x, ξ) ∈ T Ω and (x, z) ∈ Ω × R.
Integrals are taken with respect to the natural Riemannian measure. For ex- ample, if (U, Φ) is a coordinate chart and u is a continuous function compactly supported in U , we define
Z
U
u dM = Z
Φ(U )
( √
Gu) ◦ Φ −1 dx,
where dx stands for the Lebesgue measure on R n and G is the absolute value of the determinant of the metric tensor in the coordinate chart (U, Φ). With the help of smooth partitions of unity, the construction above defines a canonical positive Radon measure on M , which is the natural Lebesgue measure on M , denoted simply by | · |. In particular, we can deal with measurable vector fields, i.e. measurable sections of the tangent bundle. For a section V defined on Ω and p ≥ 1, we introduce the usual Lebesgue p–norm as
kV k p,Ω = µZ
Ω
|V | p dM
¶ 1/p ,
where |V |(x) = p
g(V (x), V (x)). In fact, we use the notation |·| to denote, according to the cases if no ambiguity occurs, the real modulus, the Riemannian norm of tangent vectors and the measure of measurable subsets of M .
Let H 1,p (Ω) be the closure of C ∞ (Ω) in the Sobolev norm kuk = kuk p,Ω + k∇uk p,Ω , where kuk p,Ω = kuk L
p(Ω) and let
H loc 1,p (Ω) = ©
u : Ω → R : u |Ω
0∈ H 1,p (Ω 0 ) for all open sets Ω 0 ⊂⊂ Ω ª
.
Finally, as usual, denote by H 0 1,p (Ω) the closure of C c ∞ (Ω) with respect to the Sobolev norm k · k.
We recall that, if u ∈ L 1 loc (Ω), a locally integrable vector field H ∈ L 1 loc (Ω, T Ω) is a weak gradient for u if
Z
Ω
hH, V i dM = − Z
Ω
u div V dM
for every vector field V ∈ C c ∞ (Ω, T Ω) (see [10]). Since H is unique, we set H = ∇u.
Of course, if u is a smooth function its usual Riemannian gradient is also the weak gradient. We shall also use the following fact:
Lemma 2.1 ([8], Proposition 2.4) If u : Ω → R is a Lipschitz function, then u ∈ H loc 1,p (Ω) for every p ≥ 1.
3 Maximum principles for inhomogeneous inequalities
In this section we extend to a Riemannian setting the results concerning Maximum Principles for solutions of inhomogeneous elliptic inequalities proved in [17, Chap- ter 6] for the Euclidean case. Therefore, we cover their results, but at the same time we establish more precise a priori estimates, in which we make explicit the dependence on all the coefficients appearing in the inequality.
Our results apply to a large class of differential inequalities, including, in particu- lar, those involving the p–Laplace–Beltrami operator defined for a smooth function u as ∆ p u := div (|∇u| p−2 ∇u), p > 1, or the mean curvature operator given by div
à ∇u
p 1 + |∇u| 2
!
; but they also apply to more general and sophisticated differ- ential operators on Riemannian manifolds, which are elliptic according to the new definition of ellipticity proposed in [1], to which we refer.
From now on, we assume Ω to be a bounded and smooth domain of M , so that Ω is a smooth manifold with boundary. However, we can also treat the case of Ω unbounded but with finite measure; in this case the boundary condition “u ≤ M on ∂Ω” is replaced by
lim sup
|x|→∞
u(x) ≤ M,
while by u ≤ M on ∂Ω we mean that for every δ > 0 there exists a neighborhood of ∂Ω in which u ≤ M + δ. In fact, u will be assumed only of class H loc 1,p (Ω), so that it may have no trace on ∂Ω.
We consider inequalities of the form
divA(x, u, ∇u) + B(x, u, ∇u) ≥ 0 in Ω, (3.2)
where divergence and gradient are taken with respect to the Riemannian structure.
We assume that A : T Ω × Ω R → T Ω, where T Ω × Ω R stands for T Ω × Ω (Ω × R) as already mentioned, and A(x, z, ξ) ∈ T x M for all x ∈ Ω, z ∈ R and ξ ∈ T x M , while B is a real function defined in T Ω × Ω R. We also suppose that there exist p ≥ 1, a 1 > 0, a 2 , a, b 1 , b 2 , b ≥ 0 such that for all (x, z, ξ) ∈ T Ω × Ω R there hold
hA(x, z, ξ), ξi ≥ a 1 |ξ| p − a 2 |z| p − a p , B(x, z, ξ) ≤ b 1 |ξ| p−1 + b 2 |z| p−1 + b p−1
(3.3)
if p > 1, and
hA(x, z, ξ), ξi ≥ a 1 |ξ| − a 2 |z| − a, B(x, z, ξ) ≤ b (3.4) if p = 1. Of course, by a rescaling argument it is enough to consider only the case a 1 = 1, so without loss of generality we assume a 1 = 1 throughout the rest of the paper.
Definition 3.1 A (weak) solution of (3.2) is a function u ∈ H loc 1,p (Ω) such that A(·, u, ∇u) ∈ L 1 loc (Ω; T Ω), B(·, u, ∇u) ∈ L p loc
0(Ω),
where p 0 = p/(p − 1) if p > 1 and p 0 = ∞ if p = 1, and such that Z
Ω
hA(x, u, ∇u), ∇φi dM ≤ Z
Ω
B(x, u, ∇u)φ dM (3.5) for all nonnegative φ ∈ H 1,p (Ω) such that φ = 0 a.e in some neighborhood of ∂Ω.
Furthermore, we say that u is a p–regular solution if u is a solution of (3.2) and in addition
A(·, u, ∇u) ∈ L p loc
0(Ω; T Ω). (3.6) Remark 3.1 Condition (3.6) reads A(·, u, ∇u) ∈ L ∞ loc (Ω; T Ω) when p = 1. Of course, we can substitute our results for p–regular solutions, p ≥ 1, assuming for example that u ∈ H loc 1,p (Ω) and, for some nonnegative constants γ i , i = 1, 2, 3,
|A(x, s, ξ)| ≤ γ 1 + γ 2 |s| p−1 + γ 3 |ξ| p−1 for all (x, z, ξ) ∈ T Ω × Ω R.
We now list our main results, where we denote by C generic constants which may depend only on p, n and |Ω|.
Theorem 3.1 (Semi–Maximum principle) Let u be a p–regular solution of (3.2) with A and B satisfying (3.3) or (3.4). Assume also that u ≤ M on ∂Ω for some constant M ≥ 0. Then u + ∈ L ∞ (Ω) and there exists a universal constant C = C(n, p, |Ω|) > 0 such that
u ≤ M + C ¡
ku + k p + a + b + K ¢
a.e. in Ω, where K = K(a 2 , b 1 , b 2 ) is given by
K =
( [b 1 + (a 2 + b 2 ) 1/p ] n/p + (a 1/p 2 + b 1/(p−1) 2 )M, if p > 1,
a n 2 + a 2 M, if p = 1.
The same result can be given if we let the coefficients in (3.3) and (3.4) depend on the x–variable with some regularity. More precisely, denoting simply by kf k the norm in L q (Ω) of f when q is assigned, we have the following generalization.
Theorem 3.2 Let a, a 2 , b, b 1 , b 2 belong to some Lebesgue spaces, in particular a, b 1 ∈ L αp (Ω), b ∈ L α(p−1) (Ω), a 2 , b 2 ∈ L α (Ω),
α = max{n/p, 1}
1 − ε , ε ∈ (0, 1],
(3.7)
and let u be a p–regular solution of (3.2) with A and B satisfying (3.3) or (3.4).
If u ≤ M on ∂Ω for some constant M ≥ 0, then u + ∈ L ∞ (Ω) and there exists a universal constant C = C(n, p, |Ω|, ε) > 0 such that
u ≤ M + C
( (1 + kb 1 k + ka 2 + b 2 k 1/p ) ν ¡
ku + k p + K ¢
, if p > 1, (1 + ka 2 + bk) ν ¡
ku + k 1 + kak + M ka 2 k ¢
, if p = 1, (3.8) where ν = n/εp if n > p and ν = 4/ε if p ≥ n, and for brevity
K = kak + kbk + (ka 2 k 1/p + kb 2 k 1/(p−1) )M.
The limit case ε = 1 is also allowed, provided that α is replaced by ∞.
The embedding of H 0 1,1 (Ω) in L n/(n−1) (Ω) is used below, with an obvious mean- ing when n = 1, and S = S(1, n) denotes the corresponding Sobolev constant. With the aid of the previous estimates (3.8), one can prove the following general results.
Theorem 3.3 (Maximum principle) Let u be a p–regular solution of (3.2) in Ω, where A and B satisfy (3.3), with b 1 = b 2 = 0, or (3.4). Suppose u ≤ M on ∂Ω for some constant M ≥ 0 and, if p = 1, we also assume that
a 2 + b ≤ |Ω| −1/n (1 − δ)/S, (3.9) where δ ∈ (0, 1). Then u + ∈ L ∞ (Ω) and there exists C = C(n, p, |Ω|) > 0 such that
u ≤ M + C ¡
a + b + M a 1/p 2 ¢
e a
(ν+1)/p2a.e. in Ω, (3.10)
if p > 1, while
u ≤ M + Ca
δ (1 + a 2 + b) a.e. in Ω (3.11) if p = 1.
Theorem 3.4 (Maximum principle: case p > 1) Let u be a p–regular solution of (3.2) in Ω, where A and B satisfy (3.3), with a 2 = b 2 = 0. If u ≤ M on ∂Ω for some constant M ≥ 0, then u + ∈ L ∞ (Ω) and
u ≤ M + (a + b)e C(1+b
1)
ν+1a.e. in Ω. (3.12)
Again, the result can be generalized if the coefficients in (3.3) or (3.4) belong to some suitable Lebesgue spaces.
Theorem 3.5 Theorems 3.3 and 3.4 continue to be valid if the coefficients a, b, a 2
and b 1 are functions in the following Lebesgue spaces:
a, b 1 ∈ L βp (Ω), b ∈ L β(p−1) (Ω), a 2 ∈ L β (Ω), β =
(
n/p(1 − ε), if 1 < p ≤ n,
1, if p > n, ε ∈ (0, 1], (3.13)
and a, b, a 2 ∈ L σ (Ω) for some σ > n if p = 1. In the latter case also assume that ka 2 + bk n ≤ 1 − δ
S . (3.14)
Under these hypotheses, (3.10) becomes u ≤ M + C ¡
kak + kbk + M ka 2 k 1/p ¢
e ka
2k
(ν+1)/pa.e. in Ω, (3.15) while (3.11) is replaced by
u ≤ M + Ckak 1
δ (1 + ka 2 + bk σ ) a.e. in Ω, (3.16) and (3.12) changes into
u ≤ M + (kak + kbk)e C(1+kb
1k)
ν+1a.e. in Ω, (3.17) with ν in (3.15) and (3.17) defined as in Theorem 3.2.
Let us note that the Lebesgue spaces allowed for the coefficients in Theorems 3.2 and 3.5 are different.
For the applications and examples of Theorem 3.5 in the Euclidean case when p ≥ 1, we refer to [17, Sections 6.1 and 6.5]. In the context of n–dimensional Cartan–Hadamard manifolds 1 Croke establishes in [5], see also [9, Section 8.2], the explicit formula for an upper bound of the best Sobolev constant S = S(1, n), that is
C(n) := ω n−2 n−2 ω n−1 n−1
ÃZ π/2
0
(cos t) n/(n−2) (sin t) n−2 dt
! n−2
≥ S(1, n),
where ω n denotes the n–dimensional surface measure of the unit sphere S n of R n+1 . Moreover, Croke shows that C(4) = S(1, 4) = 2 −7 π −2 , and for the main prototype when p = 1 and a 2 = 0, that is
div
à ∇u
p 1 + |∇u| 2
!
+ B(x, u, ∇u) ≥ 0, (3.18)
1