# A note on the strong maximum principle

## A note on the strong maximum principle

Available online at www.sciencedirect.com ScienceDirect J. Differential Equations 259 (2015) 4356–4375 www.elsevier.com/locate/jde A note on the str...

Available online at www.sciencedirect.com

ScienceDirect J. Differential Equations 259 (2015) 4356–4375 www.elsevier.com/locate/jde

A note on the strong maximum principle Michiel Bertsch a,b , Flavia Smarrazzo c , Alberto Tesei b,∗ a Dipartimento di Matematica, Università di Roma “Tor Vergata”, Via della Ricerca Scientifica, 00133 Roma, Italy b Istituto per le Applicazioni del Calcolo “M. Picone”, CNR, Roma, Italy c Università Campus Bio-Medico di Roma, Via Alvaro del Portillo 21, 00128 Roma, Italy

Received 28 July 2014; revised 23 April 2015 Available online 30 June 2015

Abstract We give a necessary and sufficient condition for the validity of the strong maximum principle in one space dimension. © 2015 Elsevier Inc. All rights reserved. MSC: 35B50; 35D30; 35J15 Keywords: Strong maximum principle; Singular elliptic equations; Weak solutions; Differential inequalities

1. Introduction The maximum principle is an important feature of second-order elliptic equations. Although its weak form is sufficient for many purposes (e.g., see [12,19]), the strong maximum principle, which excludes the assumption of a nontrivial interior maximum, is often important. Its classical form can be stated as follows. Let  ⊆ RN be a connected, open set with smooth boundary ∂. Consider a uniformly elliptic operator of the form Lu ≡ −

N 

aij

i,j =1

N  ∂ 2u ∂u + bi + cu ∂xi ∂xj ∂xi i=1

* Corresponding author.

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4357

with coefficients aij = aj i , bi , c ∈ C() ∩ L∞ () (i, j = 1, . . . , N ). Let c ≥ 0 in , and let u ∈ C 2 () ∩ C() satisfy the inequality Lu ≤ 0 in . Then the strong maximum principle asserts that, if there exists x0 ∈  such that u(x0 ) = sup u =: M ≥ 0, then u ≡ M in . It follows that, if Lu ≥ 0 in  and there exists x0 ∈  such that u(x0 ) = inf u =: m ≤ 0, then u ≡ m in . In particular, by the strong maximum principle, if u ≥ 0 and Lu ≥ 0 in , then either u > 0 in , or u ≡ 0 in  . Much effort has been made to extend the above result in two main directions: (i) addressing degenerate elliptic operators, for which the assumptions of uniform ellipticity, and/or regularity and boundedness of the coefficients aij , bi , c are not satisfied; (ii) relaxing the assumptions concerning the boundary ∂. In connection with (i), weak or distributional formulations of the differential inequality Lu ≤ 0 have been used. Let us recall some results concerning degenerate elliptic operators. Let aij ∈ C 2 (), bi ∈ 1 C (), D α aij ∈ L∞ () for |α| ≤ 2, D α bi ∈ L∞ () for |α| ≤ 1; let u ∈ C 2 () satisfy Lu ≤ 0 in . For any x0 ∈  such that u(x0 ) = sup u > 0, consider the propagation set P(x0 ) := {x ∈  | u(x) = u(x0 )}. As proven in , P(x0 ) contains the closure (in the relative topology) of the set P (x0 ) consisting of points, which can be joined to x0 by a finite number of subunitary and/or drift trajectories (see [6,9,16,21] for the proof in particular cases; see also ). Remarkably, the set P (x0 ) coincides with the support of the Markov process corresponding to the operator L – namely, with the closure of the collection of all trajectories of a Markovian particle, starting at x0 , with generator L (see [23,24]). Therefore, only the attainable points of the boundary ∂ (see [10,11]) belong to P (x0 ). A similar idea underlies the so-called refined maximum principle proven in , where the operator L is uniformly elliptic, but no smoothness of the boundary ∂ is assumed. Consider the subset of E ⊆ ∂ consisting of attracting boundary points, where the minimal positive solution U0 of the first exit time equation −

N  i,j =1

N  ∂ 2u ∂u aij + bi = 1 in  ∂xi ∂xj ∂xi

(1.1)

i=1

can be prolonged to zero (see [13,14]). It was proven in  that, if u is bounded from below in , u ≥ 0 in E and Lu ≥ 0 in , then u ≥ 0 in  (in particular, in the usual situation the weak maximum principle is recovered). In the above literature boundedness of the coefficients aij , bi , c is always assumed. Concerning unbounded coefficients, while referring the reader to [8,15,20] and the references therein, let us focus on the following result. It is known that the strong maximum principle holds for √ p the operator Lu = −u + cu, if c ≥ 0 is measurable and c ∈ Lloc () for some p > N (see [22, Corollary 8.1]). More generally, the following holds (see , [7, Theorem 1], [17, Proposition 5.20]): Theorem 1.1. Let  ⊆ RN be connected and open. Let c ∈ L1 (), c ≥ 0 a.e. in , and let u ∈ C() ∩ L1 () be nonnegative in . If −u + cu ≥ 0 in D () and u = 0 in a set of positive Newtonian capacity, then u ≡ 0 in . √ It was asked in  whether the condition c ∈ L1 () can be relaxed, e.g. by assuming c ∈ L1loc (). It is proven below that this is not the case (see Remark 2.2). This follows from the main

4358

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

result of this note (see Theorem 2.1): a necessary and sufficient condition on c for the validity of the strong maximum principle for the operator Lu = −u

+ cu, if N = 1. Apart from the intrinsic interest of the problem, the present study is motivated by a nonlinear evolution problem considered in [4,5], where a sharp version of the strong maximum principle is needed. Our proof is based on the classical Jacobi method of multiplicative variation, which relies on the construction of a proper solution U of the equation U

+ c = 0 in I ,

(1.2)

where I ⊆  is a bounded interval. This solution is a supersolution of the problem 

−u

+ cu = 0 in I u=0 on ∂I

(see Subsection 3.2, where U = −ψ). Similar methods have been used to study the wellposedness of boundary value problems for elliptic and parabolic equations with singular coefficients (see [18,20] and the references therein). Observe that, by the connection between the operator −u

+ cu, with c unbounded, and the degenerate elliptic operator − 1c u

, equation (1.2) is the counterpart of the first exit time equation (1.1). Let us stress that both Theorem 2.1 and its proof heavily depend on assuming N = 1. If N > 1, the problem of the optimal condition on the coefficient c, which ensures the strong maximum principle for the operator −u + cu to hold, is to our knowledge completely open. 2. The main result In the sequel we shall always assume that 

 ⊆ R is a nonempty open interval , c :  → R is measurable, and c ≥ 0 a.e. in  .

(H )

Let us state the following definitions. Definition 2.1. Let assumption (H ) be satisfied. A function u ∈ C(), u ≥ 0 is said to satisfy the inequality −u

+ cu ≥ 0 in D () ,

(2.1)

if cu ∈ L1loc () and 



uρ dx ≤ 

cuρ dx

for any ρ ∈ Cc∞ (), ρ ≥ 0 .

(2.2)



Definition 2.2. Let assumption (H ) be satisfied. The strong maximum principle is said to hold in  for the operator Lu = −u

+ cu, if for any nonnegative function u ∈ C() inequality (2.1) implies that either u ≡ 0, or u > 0 in .

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

For any measurable function c :  → R, c ≥ 0 set   N := x0 ∈  |  δ > 0 such that c ∈ L1 (Iδ (x0 )) ,

4359

(2.3)

where Iδ (x0 ) ≡ (x0 − δ, x0 + δ) with δ > 0 sufficiently small. Observe that  \ N is open, and N = ∅ if and only if c ∈ L1loc (). Let us denote by dN (x) the distance dist(x, N ), for any x ∈ . The main result of this note is the content of the following Theorem 2.1. Let assumption (H ) be satisfied, and let the set N be defined by (2.3). The following statements are equivalent: (i) the strong maximum principle holds in  for the operator Lu = −u

+ cu; (ii) either N = ∅, or N = ∅ and for any connected component (α, β) of  \ N the map ⎧ 1 if β ∈ ∂ , ⎪ ⎨ Lloc ([α, β)) 1 x → dN (x)c(x) belongs to Lloc ((α, β]) if α ∈ ∂ , ⎪ ⎩ 1 1 Lloc ([α, β)) ∪ Lloc ((α, β]) if [α, β] ⊂  .

(2.4)

Remark 2.1. It follows from the proof of Theorem 2.1 in Section 4 that if N = ∅ and condition (2.4) is satisfied for any connected component of  \ N , not only the strong maximum principle holds, but every function u ∈ C(), u ≥ 0 satisfying inequality (2.1) is identically zero in . −λ Remark 2.2. Let  = (−1, 1) and c(x) = x −2 log |x| . Then N = {0} and  \ N = (−1, 0) ∪ (0, 1). If λ > 1, the map x → |x|c(x) belongs to L1loc () whence, by Theorem 2.1, the strong √ maximum principle holds. On the other hand, c ∈ L1loc () if and only if λ > 2. Therefore, for √ λ ∈ (1, 2) the strong maximum principle holds, yet the function c does not belong to L1loc (). √ Hence the condition c ∈ L1loc () is not necessary for the validity of the strong maximum principle. √ The condition c ∈ L1loc () is not sufficient either. In fact, the coefficient  n4 if n ≤ y ≤ n + n14 (n ∈ N) b log |x| c(x) = , where b(y) := 1 otherwise, x2 1+y 4 satisfies

√ c ∈ L1loc () with  = (−1, 1), since 1 2

1 

c(x) dx =

−1

∞ 0

 ∞  1 dy b(y) dy ≤ +  < ∞. 2 n 1 + y4 n=1 ∞

0

On the other hand, by Theorem 2.1 the strong maximum principle fails, since the map x → |x|c(x) does not belong to L1loc ([0, 1)): 1/e ∞ xc(x) dx = b(y) dy = ∞ . 0

1

4360

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

Remark 2.3. In Theorem 2.1 the inequality −u

+ cu ≥ 0 is meant in the sense of distributions (see Definition 2.1), but in principle its definition could be weakened. In particular, one could avoid the condition cu ∈ L1loc () by saying that the inequality −u

+ cu ≥ 0 is satisfied, if for any function ρ ∈ Cc∞ (), ρ ≥ 0 either cuρ ∈ / L1 (),   or cuρ ∈ L1 () and uρ

dx ≤ cuρ dx . 

(2.5)



To check that (2.5) is weaker than Definition 2.1, choose  = (−1, 1), c(x) = |x|−1−α and u(x) = |x|α for some α ∈ (0, 1). Therefore cu ∈ / L1loc (), thus inequality (2.1) is not satisfied in the sense of Definition 2.1. On the other hand, condition (2.5) is satisfied: if cuρ = |x|−1 ρ ∈ L1 (), then ρ(0) = 0 and, since ρ ≥ 0 and ρ ∈ Cc∞ (), there holds ρ(x) ≤ Cx 2 for some constant C > 0. Therefore, 1

|x| [−ρ (x) + |x| α

−1−α

1 ρ(x)] dx = α

−1

x|x|

1

α−2

ρ (x) dx +

−1

|x|−1 ρ(x) dx =

−1

1 = α(1 − α)

1 |x|

α−2

ρ(x) dx +

−1

|x|−1 ρ(x) dx ≥ 0 .

−1

It also follows from the above example that Theorem 2.1 is false if we replace (2.1) by (2.5) in Definition 2.2 . In fact, the function u(x) = |x|α satisfies u(0) = 0 and u ≡ 0 in , whereas |x|c(x) = |x|−α with α ∈ (0, 1), hence condition (2.4) is satisfied with N = {0}. In view of Remark 2.3, we introduce another condition on u: for any compact interval K ⊂ , u(x) ≤ C|x − y| if u(y) = 0 and x, y ∈ K .

(L)

Let us also state the following definition. Definition 2.3. A nonempty subset N ⊂  is called locally finite, if its intersection N ∩ K with any compact interval K ⊂  is a finite set. Then Theorem 2.1 can be modified as follows. Theorem 2.2. Let (H ) be satisfied, and let N be defined by (2.3). The following statements are equivalent: (i) for all u ∈ C(), u ≥ 0 which satisfy (2.5) and condition (L), either u ≡ 0, or u > 0 in ;

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4361

(ii) either N = ∅, or N is locally finite and the map x → dN (x)c(x) belongs to L1loc () .

(2.6)

Observe that: a) condition (2.6) is stronger than (2.4); b) condition (L) excludes the counterexample u(x) = |x|α (α ∈ (0, 1)) discussed in Remark 2.3. 3. Preliminaries 3.1. Notation We denote by M() the space of finite Radon measures on . For every μ ∈ M(), we denote by μs the singular part of μ, and by μr the density of the absolutely continuous part μac of μ – namely, according to the Radon–Nikodym Theorem, μr is the unique function in L1 () such that  μac (E) = μr dx for all Borel sets E ⊆  . E

We denote by ·, · the duality map between M() and the space Cc () of continuous functions with compact support. If μ ∈ M() and ρ ∈ L1 (, μ) we set, by abuse of notation,  μ, ρ :=

ρ(x) dμ(x) .

(3.1)



For any Borel set E ⊆ , we denote by L1loc (E) the set of functions f : E → R such that f ∈ L1 (K) for any compact subset K ⊆ E. 3.2. Some auxiliary functions Set I = (−1, 1), I0 = I \ {0}. Let the function a : I → R satisfy 1 a ≥ 0,

x → |x|a(x) ∈ L (I ),

|x|a(x) dx < 1 .

1

(A)

−1

Observe that assumption (A) implies that a ∈ L1loc (I0 ). Consider the problem ⎧

⎨ −ψ + a = 0 in I0 ψ(0) = 1 ⎩ ψ(±1) = 0 .

(3.2)

2,1 Definition 3.1. A function ψ ∈ C(I¯) ∩ Wloc (I0 ) is called a solution of problem (3.2) if it satisfies (3.2) in the strong sense.

4362

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

Lemma 3.1. Let assumption (A) be satisfied. Then the unique solution of problem (3.2) is the function ⎧ 1 ⎪ ⎪ ⎪ ⎪ ⎪ (y − x)a(y) dy if x ∈ [0, 1] , C(1 − x) + ⎪ ⎪ ⎨ x ψ(x) := x ⎪ ⎪ ⎪ ⎪ ⎪ D(1 + x) + (x − y)a(y) dy if x ∈ [−1, 0) , ⎪ ⎪ ⎩ −1

with 1 C := 1 −

0 D := 1 +

ya(y) dy ,

ya(y) dy .

−1

0

In particular, 1

ψ (x) = −C −

a(y) dy

(x ∈ (0, 1)) ,

x

xψ (x) → 0 as x → 0 , ψ (x) → −∞

as x → 0+ if and only if a ∈ / L1 (I ) .

(3.3)

Proof. By condition (A), C and D are positive constants, so ψ < 0 in (0, 1) and ψ > 0 in (−1, 0). Below we show that 1 a(y) dy = 0 ,

lim x

x→0+

(3.4)

x

thus xψ (x) → 0 as x → 0. The remainder of the proof is straightforward, thus we omit it. Consider the function 1 φ(x) :=

(y − x)a(y) dy

(x ∈ (0, 1)) .

x

Since

1

φ (x) = −

a(y) dy ≤ 0 in (0, 1) , x

the function φ is decreasing, hence there exists limx→0+ φ(x) =: φ(0). Moreover, by (3.5), 1 0 ≤ φ(0) ≤

ya(y) dy < ∞ . 0

(3.5)

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4363

Letting x → 0+ in equality (3.5), we find that 1 k := lim x x→0+

1 a(y) dy = −φ(0) +

x

ya(y) dy 0

is well-defined and finite. By contradiction, let us suppose that k > 0. Then there exists ε > 0 such that 1 a(y) dy ≥

k 2x

for every x ∈ (0, ε) ,

x

whence ε

1 a(y) dy ≥

dz x

k ε log 2 x

(x ∈ (0, ε)) .

z

By letting x → 0+ in the above inequality we obtain ε xa(x) dx = ∞ , 0

which contradicts (Aε ). Hence equality (3.4) follows.

2

Let ψ the solution of problem (3.2), and let η ∈ (0, 1) be fixed. Consider the problem 

ψ 2ξ + χ = 0 ξ(±η) = 0 ,

in (−η, η)

(3.6)

where χ ∈ C(I ), χ ≥ 0 and supp χ ⊂ (−η, η). The unique classical solution of problem (3.6) is given by x ξ(x) := −η

⎫ ⎧ y  ⎬ dy ⎨ χ(z) dz + B − ⎭ ψ 2 (y) ⎩

(x ∈ [−η, η]) ,

(3.7)

−η

where η B :=

dy  y −η ψ 2 (y) −η χ(z) dz  η dy −η ψ 2 (y)

(3.8)

4364

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

(observe that both ξ and ψ 2 ξ belong to C 1 ([−η, η])). In particular, η 2

(ψ ξ )(−η) = B > 0 ,

2

(ψ ξ )(η) = −

dy  η −η ψ 2 (y) y χ(z) dz  η dy −η ψ 2 (y)

=: −B˜ < 0 .

(3.9)

Now let us define ζ : I → R by setting  ζ :=

ξ

in [−η, η]

0

in I \ [−η, η] .

(3.10)

By the properties of ξ , the function ψ 2 ζ is continuously differentiable in the set I \ {±η} with derivative −χ , whereas it has a jump discontinuity at ±η. Therefore, its derivative (ψ 2 ζ ) belongs to M(I ), and by equalities (3.9) there holds (ψ 2 ζ ) = −χ + B δ−η + B˜ δη

in M(I )

(3.11)

(here δx denotes the Dirac measure concentrated at x ∈ I ; by abuse of notation, we have identified the absolutely continuous part of (ψ 2 ζ ) with its density). Proposition 3.2. Let assumption (A) be satisfied, and let w ∈ C(I ), w ≥ 0. Given η ∈ (0, 1), let ζ be defined by (3.10). If   (ψ 2 ζ ) , w ≤ 0 ,

(3.12)

I

then w(η)

w(x) ≥ W (x) :=

x

 η dy dy −η ψ 2 (y) + w(−η) x ψ 2 (y)  η dy −η ψ 2 (y)

for any x ∈ [−η, η] .

(3.13)

Proof. The function W in the right-hand side of inequality (3.13) is a classical solution of the problem 

ψ 2 W = 0 in (−η, η) W (±η) = w(±η) .

By (3.6), (3.9) and (3.11),    (ψ 2 ζ ) , W = − χ W dx + (ψ 2 ξ )(−η) W (−η) − (ψ 2 ξ )(η) W (η) = η

I

−η

η = −η

 

ψ 2 ξ W dx + (ψ 2 ξ )(−η) W (−η) − (ψ 2 ξ )(η) W (η) =

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

η =−

w(−η) − w(η) ψ ξ W dx =  η dy 2

−η

−η

ψ 2 (y)

4365

ξ dx = 0 .

−η

On the other hand, 

2

(ψ ζ ) , w − W

 I

η =−

˜ χ (w − W ) dx + B[w(−η) − W (−η)] + B[w(η) − W (η)] =

−η

η =−

χ (w − W ) dx .

−η

In view of (3.12), we have that    − (ψ 2 ζ ) , w = χ (w − W ) dx ≥ 0 for all χ ∈ Cc (−η, η), χ ≥ 0 . η

I

−η

Then the conclusion follows from the arbitrariness of χ . 2 We conclude this subsection with the proof of the following result. Lemma 3.3. Assume that a ∈ L1loc ((0, 1]) ,

a ≥ 0,

x → xa(x) ∈ / L1 ((0, 1)) .

(B)

Then there exists u ∈ C 1 ([0, 1]) ∩ W 2,1 ((0, 1)) such that −u

+ au = 0 a.e. in (0, 1), u > 0 in (0, 1) and u(0) = u (0+ ) = 0. Proof. Let uk ∈ C 1 ((0, 1]) be the solution of 

in (0, 1) u

= au u(1) = 1 , u (1) = k (k > 0) .

(3.14)

If k ≤ 1, there holds uk (x) ≥ x > 0 for any x ∈ (0, 1]. On the other hand, it is easily seen that ˜ = 0 for some x˜ ∈ (0, 1) if we choose k > 1 sufficiently large. Hence, by setting uk (x) A0 := {k > 1 | ∃ x˜ ∈ (0, 1) such that uk (x) ˜ = 0, u k (x) ˜ > 0} , we have that k ∗ := inf A0 ∈ (1, ∞). Since uk depends continuously on k (locally in (0, 1]), the / A0 . It follows at once from the definition of k ∗ that uk ∗ (0) = 0 and set A0 is open and k ∗ ∈

+ uk ∗ (0 ) ≥ 0. It remains to prove that u k ∗ (0+ ) = 0. By contradiction, suppose that u k ∗ (0+ ) > 0. Since uk ∗ is convex, u

k ∗ (x) = a(x)uk ∗ (x) ≥ u k ∗ (0+ )xa(x) for x ∈ (0, 1) .

4366

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

By integrating between x and 1, we find that

u k ∗ (x) ≤ k ∗

− u k ∗ (0+ )

1 ya(y) dy

for x ∈ (0, 1) .

x

Hence u k ∗ (0+ ) = −∞, which gives a contradiction. 2 3.3. Approximation of a in L1 () Let  = I , and let the function a satisfy assumption (A). Since the solution ψ of problem / L1 (I ) (see (3.3)), we consider the solutions ψn of the ap(3.2) satisfies ψ (0± ) = ∓∞ if a ∈ proximating problems 

−ψn

+ an = 0 in I0 ψn (0) = 1 , ψn (±1) = 0

(n ∈ N) ,

(3.15)

where {an } ⊂ L1 (I ) is any sequence such that 0 ≤ an ≤ a in I, and the maps x → xan (x) converge to x → xa(x) in L1 (I ) .

(An )

An example is an = aχ(−1,− 1 )∪( 1 ,1) n

(n ∈ N) ,

n

where χE is the characteristic function of a Borel set E. Since Lemma 3.1 can be applied to problem (3.15), we have that ψn ∈ C(I¯) ∩ C 1 (I¯ \ {0}) and ψn (0+ ) − ψn (0− ) := lim ψn (x) − lim ψn (x) = x→0+

x→0−



(1 − |y|) an (y) dy =: −jn < 0 .

=− I

In particular, ψn

∈ M(I ) and 

ψn

 s

= −jn δ0 ,



ψn

 r

= an ∈ L1 (I ) .

If we identify, by abuse of notation, the absolutely continuous part of ψn

with its density, this implies that ψn

= an − jn δ0

in M(I ) .

(3.16)

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4367

Observe that, by assumption (A) and Lemma 3.1, sup |ψn (x) − ψ(x)| = x∈I

1 1 = sup −(1 − x) y(a − an )(y) dy + (y − x)(a − an )(y) dy ≤ x∈(0,1) x

0

1 ≤2

y(a − an )(y) dy → 0 as n → ∞ . 0

Therefore, ψn → ψ

in C(I ) .

(3.17)

Finally, let η ∈ (0, 1), let χ be as in (3.6), and let ζn : I → R be defined by  ζn :=

ξn 0

in [−η, η] in I \ [−η, η]

(3.18)

(n ∈ N) ,

where ξn is the solution of the problem 

ψn2 ξ + χ = 0

in (−η, η)

ξ(±η) = 0 . By (3.7) and (3.8), x ξn (x) = −η

⎫ ⎧ y  ⎬ ⎨ dy χ(z) dz + B − n ⎭ ψn2 (y) ⎩

for x ∈ [−η, η]

−η

with dy  y −η ψn2 (y) −η χ(z) dz  η dy −η ψn2 (y)

η Bn :=

,

and dy  η −η ψn2 (y) y χ(z) dz  η dy −η ψn2 (y)

η (ψn2 ξn )(−η) = Bn > 0 ,

(ψn2 ξn )(η) = −

=: −B˜ n < 0 .

Clearly, the function ζn has the same regularity as ζ , and (ψn2 ζn ) = −χ + Bn δ−η + B˜ n δη

in M(I ) .

(3.19)

4368

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

By (3.17) there holds ζn → ζ in C(I¯) ,

(3.20)

and Bn → B ,

B˜ n → B˜ .

(3.21)

Proposition 3.4. Let the function c : (0, 1) → R satisfy 1 x → xc(x) ∈ L ([0, 1)),

c ≥ 0,

1

xc(x) dx < 1 .

(A )

0

Let u ∈ C([0, 1)), u ≥ 0, satisfy cu ∈ L1loc ((0, 1)) and −u

+ cu ≥ 0 in D ((0, 1)) .

(3.22)

If u(0) = u (0+ ) = 0, then u ≡ 0 in (0, 1). Proof. By contradiction, suppose that u(η) > 0 for some η ∈ (0, 1). We extend u and c to I = (−1, 1) by setting 



u in [0, 1) u˜ := 0 in (−1, 0) ,

a :=

c 0

in [0, 1) in (−1, 0) .

Then u˜ ∈ C([−1, 1]), and the function a satisfies assumption (A) (see Subsection 3.2). Since u (0+ ) = 0, it follows from a standard approximation procedure and (3.22) that 1

1

uρ dx ≤ 0

cuρ dx

for any ρ ∈ Cc2 ([0, 1)), ρ ≥ 0 .

0

Therefore, a u˜ ∈ L1loc (I ) and −u˜

+ a u˜ ≥ 0 in D (I ) .

(3.23)

Set w :=

u˜ ψ

in I ,

(3.24)

where ψ is the solution of problem (3.2). Since by assumption u(η) > 0, we also have that w(η) > 0. Moreover, if w satisfies (3.12) for any ζ defined by (3.10),

(3.25)

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4369

it follows from Proposition 3.2 that w(x) > 0 for any x ∈ [−η, η]. In particular, there holds u(0) ˜ = u(0) > 0, and we have found a contradiction. To prove (3.25) we set wn :=

u˜ ψn

in I ,

where ψn is the solution of problem (3.15). First we prove that for any n ∈ N    2

(ψn ζn ) , wn ≤ (a − an )u ζn dx for any n ∈ N , I

(3.26)

I

where ζn is defined by (3.18). To this purpose, observe that     (ψn ζn )

, u˜ I = ψn

ζn + 2ψn ζn + ψn ζn

, u˜ I =    = an u˜ ζn dx + 2ψn ψn ζn + ψn2 ζn

, wn = I

I



  an u˜ ζn dx + (ψn2 ζn ) , wn ;

=

I

(3.27)

I

here we have used the equality 

 ψn

ζn , u˜ I =



 an u˜ ζn dx − jn δ0 , u ˜ I=

I

an u˜ ζn dx , I

which follows from (3.16) since, by assumption, u(0) ˜ = u(0) = 0. Recall that, by inequality (3.23), 

u(−ρ ˜ + aρ) dx ≥ 0 for any ρ ∈ Cc∞ (I ), ρ ≥ 0 . I

It follows by a standard approximation procedure that we can relax the regularity of ρ:  

 − ρ , u˜ I + a uρ ˜ dx ≥ 0 for any ρ ∈ Cc (I ), ρ ≥ 0, with ρ ∈ BVloc (I ) .

(3.28)

I

In fact, let j ∈ Cc∞ (I ), j ≥ 0, denote a standard mollifier ( > 0). Then for any ρ as in (3.28) (j ∗ ρ)

= j

∗ ρ → ρ

in D () as  → 0+ , and, by (2.1),



u(−j ˜  ∗ ρ + aj ∗ ρ) dx ≥ 0 for any  > 0 .



By letting  → 0+ we obtain (3.28).

4370

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

In particular, we can choose ρ = ψn ζn in (3.28) (n ∈ N). By (3.27), this gives       0 ≤ − (ψn ζn )

, u˜ I + a uψ ˜ n ζn dx ≤ − (ψn2 ζn ) , wn + (a − an )u˜ ζn dx , I

I

I

since ψn ≤ 1 in I . Hence we have proven (3.26). By (3.11), (3.17), (3.19), (3.21) and the definition of u, ˜ ⎫ ⎧ ⎨  χ u˜   u(η) ⎬ lim (ψn2 ζn ) , wn = lim − dx + B˜ n = n→∞ n→∞ ⎩ I ψn ψn (η) ⎭  =−

I

χ u˜ u(η)  2  dx + B˜ = (ψ ζ ) , w . I ψ ψ(η)

I

On the other hand, by assumption (An ) and since a u˜ ∈ L1loc (I ), it follows from the Monotone Convergence Theorem and (3.20) that    0 ≤ (a − an )u˜ ζn dx ≤ (a − an )ζ u˜ dx + a|ζn − ζ |u˜ dx → 0 as n → ∞ . I

I

I

Then letting n → ∞ in (3.26) gives (3.12). This proves the result.

2

Remark 3.1. We shall use Proposition 3.4 in the proof of Theorem 2.1, where 0 can be thought of as a point of the set N and I ⊂  as a neighbourhood of 0 such that I ∩ N = {0} (a similar result can be stated and proven in the interval (−1, 0]). In this connection observe that the choice of the interval [0, 1) in Proposition 3.4, of (−1, 1) in Lemma 3.1, and the integral conditions in (A), (A ) are not restrictive. In fact, if a ≥ 0 and the map x → |x|a(x) is integrable in some neighbourhood of 0, we can always choose δ > 0 such that a satisfies the condition δ x → |x|a(x) ∈ L ((−δ, δ)),

a ≥ 0,

|x|a(x) dx < 1 .

1

−δ

Then the unique solution of 

−ψ

+ a = 0 in (−δ, 0) ∪ (0, δ) ψ(0) = 1, ψ(±δ) = 0

is given by

ψδ (x) :=

⎧ δ ⎪ ⎪ ⎪ ⎪ ⎪ Cδ (δ − x) + (y − x)a(y) dy ⎪ ⎪ ⎨

if x ∈ [0, δ]

x

x ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ Dδ (δ + x) + (x − y)a(y) dy ⎪ ⎩ −δ

if x ∈ [−δ, 0) ,

(Aδ )

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4371

where ⎞ ⎛ δ 1⎝ 1 − ya(y) dy ⎠ , Cδ = δ 0

⎞ ⎛ 0 1⎝ Dδ = 1 + ya(y) dy ⎠ . δ −δ

4. Proof of the main result To prove Theorem 2.1 we shall make use of the following lemma. Lemma 4.1. Let u ∈ C(), u ≥ 0 satisfy inequality (2.1). If u(x) ¯ = 0 for some x¯ ∈ , the derivative u (x) ¯ exists, and there holds u (x) ¯ = 0. Proof. By inequality (2.1) the distribution −u

+ cu is nonnegative, hence it is a nonnegative Radon measure over . Since cu ∈ L1loc () (see Definition 2.1), it follows that u

is a Radon measure in , and its singular part (u

)s is nonpositive. Since u

is a Radon measure in , there holds u ∈ BVloc (). Let u(x) ¯ = 0 for some x¯ ∈ . Since by assumption u ≥ 0 in , and u ∈ BVloc (), both derivatives u (x¯ + ) ≥ 0 and u (x¯ − ) ≤ 0 exist and are finite. Hence there holds (u

)s ({x}) ¯ ≥ 0. By the above remarks, it follows that (u

)s ({x}) ¯ = 0. This implies that u (x¯ + ) = u (x¯ − ) = 0, hence u (x) ¯ exists and u (x) ¯ = 0. 2 Proof of Theorem 2.1. (ii) ⇒ (i): Let us show that, if either N = ∅, or N = ∅ and for any connected component of  \ N condition (2.4) is satisfied, the strong maximum principle holds in . Let N = ∅. Let u(x0 ) > 0 for some x0 ∈ , and let (x1 , x2 ) be the connected component of the open set {x ∈  |u(x) > 0} containing x0 . We must show that u > 0 in , namely (x1 , x2 ) = . By contradiction, suppose that x1 ∈  (the case x2 ∈  is similar). Then u(x1 ) = 0, whence by Lemma 4.1 u (x1 ) = 0. Moreover, since N = ∅, there holds c ∈ L1 (Iδ (x1 )) for some δ > 0, thus the assumptions of Proposition 3.4 (with [0, 1) replaced by [x1 , x1 + δ)) are satisfied. It follows that u ≡ 0 in [x1 , x1 + δ), which contradicts the definition of x1 . This proves the claim in this case. Now suppose that N = ∅, and for any connected component of  \ N condition (2.4) is satisfied. Let us show that every nonnegative function u ∈ C() satisfying inequality (2.1) is identically zero in , thus in particular the strong maximum principle holds. In fact, let u ∈ C(), u ≥ 0 satisfy inequality (2.1), thus by Definition 2.1 there holds cu ∈ L1loc (). By the continuity of u, this implies that u(x) ¯ = 0 for every x¯ ∈ N (if u(x) ¯ > 0 for some ¯ where u > 0 strictly, whence c ∈ L1 (Iδ (x)), ¯ x¯ ∈ N , there would exist a neighbourhood Iδ (x) a contradiction). By Lemma 4.1, it follows that u (x) ¯ = 0. On the other hand, by (2.4) and the symmetry between α and β, we may assume that the map x → dN (x)c(x) belongs to L1 ([x, ¯ x¯ + δ)) for some δ > 0. Then by Proposition 3.4 (with [0, 1) replaced by [x, ¯ x¯ + δ)) there holds u ≡ 0 in [x1 , x1 + δ), whence the claim follows as in the case N = ∅. (i) ⇒ (ii): Suppose that the set N does not satisfy (2.4). Then there exists a connected component (α, β) of  \ N such that one of the following holds:

4372

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

(1) α, β ∈ , (x − α)c(x) ∈ / L1loc ([α, β)) and (β − x)c(x) ∈ / L1loc ((α, β]); 1 (2) α ∈ ∂, β ∈  and (β − x)c(x) ∈ / Lloc ((α, β]); / L1loc ([α, β)). (3) α ∈ , β ∈ ∂ and (x − α)c(x) ∈ Let us prove that in this case the strong maximum principle does not hold. Consider case (1), the argument being analogous in the other cases. Let η1 , η2 > 0 satisfy / L1loc ([α, β)), by Lemma 3.3 there exists u1 ∈ C 1 ([α, α + η1 + η2 < β − α. Since (x − α)c(x) ∈ 2,1 η1 ]) ∩ Wloc ((α, α + η1 )) such that −u

1 + cu1 = 0 a.e. in (α, α + η1 ), u1 > 0 in (α, α + η1 ) and u1 (α) = u 1 (α + ) = 0. 2,1 Similarly, since (β − x)c(x) ∈ / L1loc ((α, β]), there exists u2 ∈ C 1 ([β − η2 , β]) ∩ Wloc ((β − η2 , β)) such that

−u

2 + cu2 = 0 a.e. in (β − η2 , β), u2 > 0 in (β − η2 , β) and u2 (β) = u 2 (β − ) = 0. Without loss of generality we may assume that u1 (α + η1 ) = u2 (β − η2 ) =: l. Set ⎧ 0 ⎪ ⎪ ⎪ ⎪ ⎪ ⎪ ⎨ u1 (x) u(x) = l ⎪ ⎪ ⎪ u2 (x) ⎪ ⎪ ⎪ ⎩ 0

if x < α , if α ≤ x ≤ α + η1 , if α + η1 < x < β − η2 , if β − η2 ≤ x ≤ β , if x > β .

It is easily seen that u ∈ C(), u ≥ 0 in , u > 0 in (α, β) in cu ∈ L1 (), (u

)s is nonpositive and −(u

)r + cu ≥ 0 a.e. in  . Hence inequality (2.1) is satisfied, and the conclusion follows. 2 Proof of Theorem 2.2. (ii) ⇒ (i): Let N be nonempty, locally finite and satisfying (2.6). Let u ∈ C(), u ≥ 0 satisfy (2.5) and condition (L), and assume that u(x0 ) > 0 for some x0 ∈ . Let (x1 , x2 ) be the connected component of the open set {x ∈  |u(x) > 0} containing x0 . We must show that u > 0 in , namely (x1 , x2 ) = . By contradiction, suppose that x1 ∈  (the case x2 ∈  is similar). Then u(x1 ) = 0, and by condition (L) there holds u(x) ≤ C|x − x1 | for any x ∈  . By the local finiteness of N and condition (2.6), this implies that there exists a neighbourhood Iδ (x1 ) ≡ (x1 − δ, x1 + δ) such that cu ∈ L1 (Iδ (x1 )). In fact,   0≤ cu dx ≤ C |x − x1 |c(x) dx < ∞ . Iδ (x1 )

Iδ (x1 )

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4373

It follows that for any function ρ ∈ Cc∞ (), ρ ≥ 0 such that supp ρ ⊆ Iδ (x1 ) 



 cuρ dx =



cuρ dx ≤ ρ∞

Iδ (x1 )

cu dx < ∞ .

Iδ (x1 )

Then by (2.5) for any ρ as above 

dx ≤



 cuρ dx , 

namely −u

+ cu ≥ 0 in D (Iδ (x1 )) .

(4.1)

By the above remarks, we can adapt the proof of Proposition 3.4 to the present case. Hence we obtain that u ≡ 0 in the interval (x1 , x1 + δ), which contradicts the definition of x1 . This completes the proof of the claim in this case. If N is empty, we can make use of the above argument at any point of  where u vanishes; hence the claim follows. (i) ⇒ (ii): Let (i) be satisfied, and let N = ∅. We argue by contradiction, by distinguishing two cases: either (C1 ) N is locally finite, and there exists a compact interval K = [α, β] ⊂  such that dN c ∈ / L1 (K), or (C2 ) N is not locally finite. In case (C1 ), there exist ξ ∈ K and η0 > 0 such that |x − ξ |c(x) ∈ / L1 (Uη0 ) and c ∈ \ {ξ }), where Uη0 = (ξ − η0 , ξ + η0 ). Without loss of generality we may assume that |x − ξ |c(x) ∈ / L1 ((ξ, ξ + η)). By Lemma 3.3, with I replaced by Uη0 , there exists u ∈ 1 C ([ξ, ξ + η0 ]) ∩ W 2,1 ((ξ, ξ + η0 )) such that u > 0 in (ξ, ξ + η0 ), u(ξ ) = u (ξ ) = 0, and −u

+ cu = 0 in D (Uη0 ). Extending u to  by setting u(x) = 0 for x < ξ and u(x) = u(ξ + η0 ) for x > ξ + η0 we obtain a contradiction (it is enough to follow the proof of part (ii) of Theorem 2.1, without the request that cu ∈ L1loc ()). Next we consider case (C2 ). Without loss of generality we may assume that for some ξ ∈ 

L1loc (Uη0

c∈ / L1loc (Vη ) for each right neighbourhood Vη = (ξ, ξ + η) of ξ .

(4.2)

We fix η > 0. Then there exists x1 ∈ (ξ, ξ + η) such that c ∈ / L1 ((x1 , ξ + η)). By (4.2) with 1 Vη = (ξ, x1 ), there exists x2 ∈ (ξ, x1 ) such that c ∈ / L ((x2 , x1 )). By iteration, we find a strictly decreasing sequence {xn } converging to ξ such that xn−1

c(x)dx = ∞ for any n ∈ N . xn

Then for all n ∈ N there exists kn > 0 such that

4374

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375 xn−1

(Tkn c)(x)dx ≥ xn

1 , n(xn − ξ )

where (Tk c)(x) = min{c(x), k} for x ∈ (ξ, ξ + η). Let c(x) :=

∞  (Tkn c)(x)χ(xn ,xn−1 ) (x) for a.e. x ∈ (ξ, ξ + η) . n=1

Then c ∈ L1loc ((ξ, ξ + η)), 0 ≤ c ≤ c a.e. in (ξ, ξ + η) and (x − ξ )c ∈ / L1 (ξ, ξ + η): x ξ +η  ∞ n−1  (x − ξ )c(x)dx = (x − ξ )(Tkn c)(x)dx ≥ n=1 xn

ξ

xn−1 ∞ ∞   1 (xn − ξ ) (Tkn c)(x)dx ≥ ≥ = ∞. n n=1

n=1

xn

Arguing as in the proof of Lemma 3.3 (with the interval (0, 1) replaced by (ξ, ξ + η) and c by c) shows that there exists u ∈ C 1 ([ξ, ξ + η]) such that u > 0 in (ξ, ξ + η) and u is a strong solution of the problem ⎧

in (ξ, ξ + η) ⎨ u = cu u(ξ + η) = 1 ⎩ u(ξ ) = u (ξ ) = 0 . We extend the definition of u to  \ [ξ, ξ + η]:  u(x) =

0 if x ∈ , x < ξ 1 if x ∈ , x > ξ + η .

Then u ∈ C(), u

∈ M() and u

= c χ(ξ,ξ +η) − u ((ξ + η)− ) δη

in M() .

Therefore, for any ρ ∈ Cc∞ (), ρ ≥ 0,  

  u(−ρ

+ cρ) dx ≥ − u

, ρ  +



c χ(ξ,ξ +η) ρu dx = u ((ξ + η)− )ρ(ξ + η) ≥ 0 .



To summarise, we have constructed a function u which is continuous and nonnegative in  and satisfies (2.5) and condition (L), for which u = 0 for x ≤ ξ and u > 0 for x > ξ . In view of (i) we have found a contradiction. 2

M. Bertsch et al. / J. Differential Equations 259 (2015) 4356–4375

4375

Acknowledgments A stimulating discussion with Joost Hulshof concerning Remark 2.3 is gratefully acknowledged. References  K. Amano, Maximum principles for degenerate elliptic-parabolic operators, Indiana Univ. Math. J. 28 (1979) 545–557.  A. Ancona, Une propriété d’invariance des ensembles absorbants par perturbation d’un opérateur elliptique, Comm. Partial Differential Equations 4 (1979) 321–337.  H. Berestycki, L. Nirenberg, S.R.S. Varadhan, The principal eigenvalue and the maximum principle for second-order elliptic operators in general domains, Comm. Pure Appl. Math. 47 (1994) 47–92.  M. Bertsch, F. Smarrazzo, A. Tesei, Pseudo-parabolic regularization of forward–backward parabolic equations: bounded nonlinearities, preprint, 2015.  M. Bertsch, F. Smarrazzo, A. Tesei, On a forward–backward parabolic equation, preprint, 2015.  J.M. Bony, Principe du maximum, inégalité de Harnack et unicité du problème de Cauchy pour les opérateurs elliptiques dégénérés, Ann. Inst. Fourier 19 (1969) 277–304.  H. Brezis, A.C. Ponce, Remarks on the strong maximum principle, Differential Integral Equations 16 (2003) 1–12.  S. Cerrai, Second Order PDE’s in Finite and Infinite Dimension. A Probabilistic Approach, Springer, 2001.  C. Denson Hill, A sharp maximum principle for degenerate elliptic-parabolic equations, Indiana Univ. Math. J. 20 (1970) 213–229.  M. Freidlin, Functional Integration and Partial Differential Equations, Princeton University Press, 1985.  A. Friedman, Stochastic Differential Equations and Applications, I, II, Academic Press, 1976.  D. Gilbarg, N.S. Trudinger, Maximum Principles in Differential Equations, Prentice Hall, 1967.  I. Guikhman, A. Skorokhod, Introduction à la théorie des processus aléatoires, Editions MIR, 1977.  R.Z. Khas’minskii, Diffusion processes and elliptic differential equations degenerating at the boundary of the domain, Theory Probab. Appl. 4 (1958) 400–419.  L. Lorenzi, M. Bertoldi, Analytical Methods for Markov Semigroups, CRC Press, 2006.  O.A. Oleinik, E.V. Radkevic, Second Order Equations with Nonnegative Characteristic Form, Amer. Math. Soc., 1973.  A.C. Ponce, Selected problems on elliptic equations involving measures, Mémoire présenté au concours annuel de 2012 de la Classe des Sciences de l’Académie Royale de Belgique, Groupe I, Mathématiques, Université Catholique de Louvain, 2012.  M.A. Pozio, F. Punzo, A. Tesei, Criteria for well-posedness of degenerate elliptic and parabolic problems, J. Math. Pures Appl. 90 (2008) 353–386.  M.H. Protter, H.F. Weinberger, Maximum Principles in Differential Equations, Prentice Hall, 1967.  F. Punzo, A. Tesei, Uniqueness of solutions to degenerate elliptic problems with unbounded coefficients, Ann. Inst. H. Poincaré Anal. Non Linéaire 26 (2009) 2001–2024.  R.M. Redheffer, The sharp maximum principle for nonlinear inequalities, Indiana Univ. Math. J. 21 (1971) 227–248.  G. Stampacchia, Le problème de Dirichlet pour les équations elliptiques du second ordre à coefficients discontinus, Ann. Inst. Fourier 15 (1965) 189–258.  D. Stroock, S.R.S. Varadhan, On degenerate elliptic-parabolic operators of second order and their associate diffusions, Comm. Pure Appl. Math. 25 (1972) 651–713.  K. Taira, Diffusion Processes and Partial Differential Equations, Academic Press, 1998.