MORREY SPACES AND GENERALIZED CHEEGER SETS

Transcription

1 MORREY SPACES AND GENERALIZED CHEEGER SETS QINFENG LI AND MONICA TORRES Abstract. We maximize the functional E h(x)dx, where E Ω is a set of finite perimeter, Ω is an open bounded set with Lipschitz boundary and h is nonnegative. Solutions to this problem are called generalized Cheeger sets in Ω. We show that the Morrey spaces L 1,λ (Ω), λ n 1, are natural spaces for h to study this problem. For the case n = 2, we prove that if h L 1,λ (Ω), λ > 1, then generalized Cheeger sets exist. We also study the embedding of Morrey spaces into L p spaces. We show that, for any dimension n and for any 0 < λ < n, the Morrey space L 1,λ (Ω) is not contained in any L q (Ω), 1 < q < p = n n λ. We also show that if h L 1,λ (Ω), λ > n 1, then the reduced boundary in Ω of a generalized Cheeger set is C 1,α and the singular set has Hausdorff dimension at most n 8 (empty if n 7). For the critical case h L 1,n 1 (Ω), we demonstrate that this strong regularity fails. We also prove a structure result for generalized Cheeger sets in R n ; namely, a bounded generalized Cheeger set E R n with h L 1 (R n ) is always pseudoconvex, and any pseudoconvex set is a generalized Cheeger set for some h L 1 (R n ), h nonnegative and not equivalent to zero. A similar structure theorem holds for generalized Cheeger sets in Ω. 1. Introduction In this paper we study the existence and regularity of solutions of the problem (1.1) v1 h := sup V1 h (E), V1 h h (x) dx E (E) = E Ω H n 1 ( E), where Ω is an open bounded set with Lipschitz boundary and h is an integrable function that belongs to the Morrey spaces. We consider in particular the case h 0. If h 1, (1.1) reduces to sup E Ω L n (E) H n 1 ( E) := M 1(Ω), ( whose solutions are called Cheeger sets. Cheeger established the inequality λ 1 (Ω) 2, 1 2M 1(Ω)) where λ 1 (Ω) is the first eigenvalue of the Laplacian under Dirichlet boundary conditions. References to Cheeger sets include Caselles-Chambolle-Novaga [18, 19], Alter-Caselles [2], Figalli-Maggi- Pratelli [25], Parini [41], Alter-Caselles-Chambolle [3], Carlier-Comte-Peyre [16], Carlier-Comte [17], Butazzo-Carlier-Comte [14], Kawohl-Friedman [32], Kawohl-Novaga [34] and Kawohl-Lachand [33]. A set E where the supremum (1.1) is attained (i.e. v1 h = V h Cheeger set in Ω. sup E Ω 1 (E) = 1/J h (E)) is called a generalized Indeed, generalized Cheeger sets refer to more general problems of the type h1 dx E. For an application of generalized Cheeger sets to landslides see Ionescu- E h2 dhn 1 Lachand-Robert [31]. Generalized Cheeger have been studied when h 1 L and h 2 continuous (see Buttazo-Carlier-Comte [14] and the references therein). We will show in this paper that it is natural to study (1.1) with h belonging to the Morrey spaces. Date: June Key words and phrases. function spaces, geometric measure theory, calculus of variation, minimal surfaces. 1

2 2 QINFENG LI AND MONICA TORRES Since a necessary condition for generalized Cheeger sets to exist is that (1.2) sup V1 h (E) <, E Ω we start our study by analyzing the space S(Ω), consisting of all functions g that satisfy, for each set E R n of finite perimeter, E Ω (1.3) sup g H n 1 ( E) <. We note that if h 0 and Ω is convex, then h satisfies (1.2) if and only if h S(Ω). Moreover, for n = 2 and h 0, h satisfies (1.2) if and only if h S(Ω) (see Remark??). Therefore, since S(Ω) is contained in the Morrey space L 1,n 1 (Ω) (see Theorem 3.6), it follows that, at least for Ω convex, a necessary condition to guarantee the existence of generalized Cheeger sets in Ω is that h L 1,n 1 (Ω). This motivates our work with the Morrey spaces. In order to study characterizations of S(Ω) in terms of Morrey spaces we first note that the isoperimetric inequality implies that E Ω g H n 1 ( E) is bounded above by E Ω g dx E 1 1 n. This observation suggests that S(Ω) can be compared to the Morrey space L 1,n 1 (Ω), whose definition uses cubes Q, and Q 1 1 n instead of E 1 1 n (see Definition 2.10 and note our notation M p (Ω) := L 1,λ (Ω), λ > 0, λ = n(1 1 p )). Indeed, for n = 2, we show that S(Ω) = M 2 (Ω) (see Theorem 3.6) and S(Ω) L 1,n 1 (Ω) for n > 2. Moreover, if h 0 and n = 2, we prove that generalized Cheeger sets exist whenever h belongs to the Morrey spaces L 1,λ (Ω), λ > n 1 = 1 (see Corollary 4.6). In order to obtain these results, we prove a covering theorem for convex sets which actually holds in any number of dimensions (see Theorem 3.7). This covering theorem allows to control the perimeters of the elements of the cover in a universal way, and it has its own interest. For n = 2, it gives a new characterization of Morrey spaces in terms of sets of finite perimeter. More specifically, the proof of the existence of generalized Cheeger sets essentially requires the argument that, for any polytope E, there exists a universal constant C(n) and countable convex sets E i such that (1.4) E i E i, and Σ i C(n)P (E i ). For n = 2, (1.4) follows from the covering Theorem 3.7 for convex sets and the fact that the convex hull of an indecomposable (see Definition 2.3) set E R 2 can not have larger perimeter than E. However, for n 3, the convex hull does not have this property. Moreover, we can not use the minimal hull E m of E (see Definition 7.2) because, even though it satisfies the desired property P (E m ), the set E m is not necessarily convex in dimensions greater than 2 (see Remark 7.6 for a counterexample). For this reason, we restrict in this paper our existence result to the case n = 2 (see also Remark 8.3). Another space that naturally arises in connection to problem (1.1) is M p (Ω) (see definition 2.11), which is actually the same as the weak L p space, denoted as L p,w (Ω) (see Remark 2.12). The existence of generalized Cheeger sets for h L n (Ω) was proven in Bright-Li-Torres [12]. In particular, the spaces L p,w (Ω), p > n, are contained in L n (Ω). Our existence result in Corollary 4.6 truly generalizes the existence result in [12, Corollary 6.2] since the spaces M p (Ω), p > n, are not contained in L n (Ω). This follows from Section 4, where we show that M p (Ω), p > n, can not be embbeded into any space L q (Ω), with 1 < q < p. In the second part of this paper we investigate the regularity of generalized Cheeger sets in Ω. For convenience, we study the equivalent problem (1.5) C0 h := inf J h (F ), J h (F ) = P (F ) F Ω is a set of finite perimeter, hdx, F with h M n (Ω), h 0 and h not equivalent to the zero function.

3 MORREY SPACES AND GENERALIZED CHEEGER SETS 3 We study the regularity of a generalized Cheeger set E in Ω by noticing that E is also a minimizer of the functional (1.6) I H (F ) = P (F ) + H(x)dx, F Ω, where H(x) = C0 h h(x). The minimization (1.6), with Ω replaced by R n, is called the variational mean curvature problem, and H is called the variational mean curvature of E. The regularity of this problem, for h L p (Ω), has been studied by several authors including De Giorgi [22], Massari [37, 38], Barozzi-Gonzalez-Tamanini [9], Gonzalez-Massari-Tamanini [28] and Gonzalez-Massari [27]. The regularity of more general classes of quasi minimizers of perimeter were studied in Tamanini [43], Bombieri [15], Paolini [40] and Ambrosio-Paolini [6]. Using the regularity theory for quasi minimizers developed in Tamanini [43], we show that, if E is a generalized Cheeger set for some h M p (Ω), p > n, then E Ω is a C 1,α -hypersurface and the singular set has dimension at most n 8 (empty if n 7). In particular, if h M p (Ω), p > n, the same regularity result holds. This strong regularity, for h L p (Ω), p > n, corresponds to the classical results for the variational mean curvature problem mentioned above. We also study the structure of generalized Cheeger sets in R n (i.e. Ω replaced by R n in the minimization (1.5)). For the variational mean curvature problem, it is well known that, for any set of finite perimeter E, there exists an L 1 (R n ) function H such that E has mean curvature H (see [9] and [27]). This implies that we can not expect to have regularity for sets with L 1 mean curvature, since a set of finite perimeter can have a wild boundary (see, for example, [36, Example 12.25]). Motivated by the close relationship between our generalized Cheeger set problem and the variational mean curvature problem, we ask if this is also true in our situation, namely, if for any set of finite perimeter E with 0 < E <, there exists an L 1 nonnegative function h such that E is a generalized Cheeger set in R n corresponding to this h. We answer negatively this question by showing a structure theorem (see Theorem 7.7) which gives a necessary and sufficient condition for E to be a generalized Cheeger set. More specifically, we show that if E is a generalized Cheeger set in R n for some h L 1 (R n ), then E is a pseudoconvex set (see Definition 7.2), and if E is a pseudoconvex set, then a non negative function in L 1 (R n ) can be constructed so that E minimizes J h in R n. Theorem 7.7 has some interesting consequences, in particular the fact that each indecomposable component (see Definition 2.9) of a generalized Cheeger set for some h L 1 (R n ) is a generalized Cheeger set for the same h, and hence each indecomposable component of a pseudoconvex set is also pseudoconvex (see Theorem 8.1). A similar structure theorem holds for generalized Cheeger sets in Ω (see Theorem 7.9). For the variational mean curvature problem, De Giorgi conjectured in 1992 that if H L n, then the boundary of a minimizer E of I H is locally parametrizable (out of a singular set, if n 8) by a bi-lipschitz map defined on an open ball of R n 1. De Giorgi also proposed an example of a quasi minimizer in the plane having a singular point at the origin. Gonzalez-Massari-Tamanini [28] proved that the example of De Giorgi, whose boundary is the union of two bilogarithmic spirals, is indeed a set with mean curvature in L 2 (R 2 ). The full conjecture is still an open problem, but the case n = 2 has been solved by Ambrosio-Paolini [6, Theorem 5.2]. Also, for n > 2, Paolini showed in [40] that the boundary of E is locally parametrizable for any α < 1 with a map τ such that both τ and τ 1 are C α. This result was extended to quasi minimizers of perimeter by Ambrosio-Paolini [6, Theorem 4.10]. For our problem, since it was proven in Bright-Li-Torres [12, Corollary 6.2] that generalized Cheeger sets exist for h L n (Ω), it follows from [40, 6] that, in this case, the boundary of a generalized Cheeger set E in Ω is parametrizable by a bi-hölder continuous map and, if n = 2, by a bi-lipschitz map. Motivated by the De Giorgi conjecture, we would like to know if the regularity obtained in [40, 6] also holds for generalized Cheeger sets corresponding to our critical case h M n (Ω). Although this assumption on h is weaker than h L n (Ω), our structure theorem for generalized Cheeger sets (see Theorem 7.7) suggests that one could expect some type of partial regularity in the critical case F

4 4 QINFENG LI AND MONICA TORRES h M n (Ω). In this paper we prove that the strong regularity fails by showing that there exists a function h M n (Ω) M n (Ω), and a generalized Cheeger set in Ω whose E contains a large set (of Hausdorff dimension n) consisting of points of zero density. As an application of our results, we consider in the last section the following averaged shape optimization problem (see [12]): (1.7) v := inf V (E), V (E) = V 2 (E) + V g 1 (E) = inf E f(x, ν E(x))dH n 1 (x) + E gdx E Ω E Ω H n 1 (, E) This problem was studied in [12] under the condition g L n (Ω). It was proven there that if there exists a minimizing sequence E i such that E i 0 or H n 1 (E i ), then the infimum can be approximated by convex polytopes with n+1 faces shrinking to a point. In this paper we extend this result to the case g M p (Ω), p > n = 2. In this case, we show in Corollary 9.4 that, if there exists a minimizing sequence E i of (1.7) satisfying H n 1 ( E i ) or P (E i ) 0, then v can also be approximated with a sequence V ( i ), where i is a sequence of convex polytopes with n + 1 faces that are shrinking to a point. We note that this result does not require g to be negative, although we are interested in the case g 0, to be consistent with our existence results of generalized Cheeger sets (see Remark 4.9). We organize the paper as follows. In section 2 we introduce some preliminaries needed for the paper. In section 3 we prove a covering theorem for convex sets that holds in any dimension and a characterization of Morrey spaces. In section 4 we use our covering theorem to prove the existence of generalized Cheeger sets for n = 2 and h M p (Ω), p > 2. In section 5 we study the embedding of Morrey spaces into L p spaces. In section 6 we study the regularity of generalized Cheeger sets in Ω. In section 7 we prove a structure theorem of generalized Cheeger sets and in section 8 we study the critical case h M n (Ω). Finally, in section 9 we present an application to averaged shape optimization. 2. Preliminaries In this paper we work with sets of finite perimeter. We refer the reader to the standard references by Maggi [36], Ambrosio-Fusco-Pallara [5] and Evans [23] for the relevant definitions and properties of sets of finite perimeter used in this paper. Also, H n 1 denotes the (n 1)-dimensional Hausdorff measure in R n, and L n is the Lebesgue measure in R n. We will use the notation L n (E) = E. For a set of finite perimeter E, we use indistinctly the notation = H n 1 ( E). Remark 2.1. Throughout this paper, Ω denotes a bounded open set with Lipschitz boundary. Remark 2.2. In particular, we note that we can write E Ω or E Ω since Ω is Lipschitz and hence E is L n -equivalent to E Ω. Definition 2.3. A set of finite perimeter E is said to be decomposable if there exists a partition of E into two measurable sets A, B with strictly positive measure such that (2.1) = P (A) + P (B) If no such a partition exists, we call E a indecomposable set. Remark 2.4. The notion of indecomposable set extends the topological notion of connectedness. It can be easily shown that a connected open set of finite perimeter is indecomposable. Definition 2.5. For H L 1 (R n ), let I H be the functional defined as (2.2) I H (F ) = P (F ) + H(x)dx, F R n is a set of finite perimeter. F A set E of finite perimeter is said to have variational mean curvature H if E is a minimizer of (2.2).

5 MORREY SPACES AND GENERALIZED CHEEGER SETS 5 Remark 2.6. It can be easily seen (by computing the first variation of the functional I H ) that if E has variational mean curvature H, H is continuous at x E and E is smooth near x, then the value of the (classical) mean curvature of E at x is given by H(x) n 1. The fact that every set of finite perimeter has a variational mean curvature was observed for the first time in Barozzi-Gonzalez-Tamanini [9]. Moreover, we have the following Theorem 2.7. Let E be a set of finite perimeter with E <, then there exists a function H E L 1 (R n ), such that H E is the variational mean curvature for E. Furthermore, (2.3) H E < 0 a.e. on E, H E > 0 a.e. on R n \ E, and (2.4) H L1 (E) =, H E L1 (R n ) = 2. Proof. See [27, Theorem 2.3] for the proof of (2.4). The property (2.3) is proven in [27, (2.16) and (2.17)]. The following theorem by Ambrosio-Caselles-Masnou-Morel [4, Proposition 3.5, Theorem 1 (section 4)], is an important tool we will use in this article. Theorem 2.8. Let E be a set of finite perimeter in R n. Then there exists a unique countable family of pairwise disjoint indecomposable sets E i such that E i > 0, E = i E i, = i P (E i), E = i E i (mod H n 1 ) and H n 1 ( E i E j ) = 0. Moreover, if F E is an indecomposable set then F is contained (mod H n ) in some set E i. Remark 2.9. We call each E i in the previous theorem a indecomposable component of E. We now proceed to review the definition of Morrey spaces, which will be shown to be the natural spaces to study the existence and regularity of generalized Cheeger sets in Ω. The spaces L λ,z (Ω) were introduced by Morrey in [39] and are defined as (2.5) L λ,z (Ω) = { u : sup r>0,x 0 Ω 1 r λ B r(x 0) Ω u(y) z dy < When λ = 0, the Morrey space is the same as the usual L z space. When λ = n, the spatial dimension, the Morrey space is equivalent to L, due to the Lebesgue differentiation theorem. When λ > n, the space contains only the 0 function. In this paper we study the Morrey space L 1,λ (Ω), where 0 < λ < n. Recall that L 1,λ = L 1,λ where the latter is the Companato space. (See Definition 4.1 and [42, Theorem 4.3]). If p > 1 is such that λ = n(1 1 p ) we have that u L1,λ (Ω) if and only if or equivalently, Definition L 1,λ (Ω) := sup r>0,x 0 Ω 1 B r (x 0 ) 1 1 p B r(x 0) Ω u(y) dy <, { { u L 1 Q Ω (Ω) : sup u dx } } : Q is a cube < +. Q 1 1/p (2.6) M p (Ω) := L 1,λ (Ω), λ = n(1 1 ), p > 1. p We now define the spaces of functions M p (Ω), p > 1, which are the counterparts of the Morrey spaces M p, p > 1, as follows: }

6 6 QINFENG LI AND MONICA TORRES Definition For p > 1, let (2.7) M p (Ω) := { g Lebesgue measurable : sup A Ω,A measurable A g dx < + A 1 1/p We note that M p (Ω) L 1 (Ω) by choosing A = Ω in the definition above. Moreover, if g M p (Ω), we define A (2.8) g Mp(Ω) := sup g dx A Ω,A measurable A. 1 1/p Remark The standard weak L p space, denoted as L p,w (Ω), satisfies (see [12, Lemma 8.3]): (2.9) L p,w (Ω) = M p (Ω), p > 1. Remark Remark 2.12 implies that M p (Ω), p > 1, can be embedded in to any space L q (Ω), with 1 q < p. Remark Notice that (2.10) M p (Ω) M q (Ω), 1 < q < p, and (2.11) M p (Ω) M q (Ω), 1 < q < p. One goal of this paper is to study the relationship between M p (Ω) and M p (Ω). Clearly, we have (2.12) M p (Ω) M p (Ω), p > 1, and it is natural to ask whether M p (Ω) = M p (Ω), since they are defined in a similar manner. If this is not true, we still ask the question whether M p (Ω), for p large enough, can be embedded into M q (Ω), for some q close enough to 1. We will show that the answer to both questions is false, by demostrating in section 5 that M p (Ω), p > 1, cannot be embedded into any space L q (Ω), for 1 < q < p. 3. A covering theorem and a characterization of Morrey spaces Since the necessary condition for generalized Cheeger sets to exist is that (3.1) v1 h E = sup hdx E Ω H n 1 ( E) <, in this section we study appropriate spaces for h to guarantee (3.1). We first introduce the following spaces which are related to the functional in (3.1). Definition 3.1. (3.2) S(Ω) := and (3.3) S c (Ω) := }. { { g L 1 E Ω (Ω) : sup g dx } } : E is a set of finite perimeter < +. { { g L 1 K Ω (Ω) : sup g dx } } : K is a bounded convex set < +. P (K) Remark 3.2. We define the semi-norm of u with respect to S(Ω) by { E Ω g S(Ω) := sup g dx } : E is a set of finite perimeter. We similarly define g Sc(Ω). Remark 3.3. Since a bounded convex set has finite perimeter (see Excercise in [36]), S(Ω) S c (Ω).

7 MORREY SPACES AND GENERALIZED CHEEGER SETS 7 From the definition of S(Ω), we have the following: Remark 3.4. The following observations motivate our analysis of the space S(Ω): (a) If h S(Ω) then (3.1) is satisfied. (b) For n = 2 and h 0, then (3.1) is satisfied if and only if h S(Ω). (c) For n 2, h 0 and Ω convex, (3.1) is satisfied if and only if h S(Ω). Clearly, (a) is true. Since P (E Ω) if Ω is convex, then (c) is also true. To show (b), it suffices to show that if v h <, then E Ω hdx is uniformly bounded for any connected polytope E with small perimeter. Indeed, this is because h is in L 1, and we can approximate sets of finite perimeter by open polytopes, which is a countable disjoint union of connected components E i, with estimate (3.4) E Ω h sup i E h i Ω P (E i ). Now for any connected polytope E, by the covering Theorem 3.7, there exist cubes Q k such that E Ê kq k with Σ k P (Q k ) CP (Ê) C. Since Ω is bounded and Lipschitz, we can choose {B δi (x i )} N i=1 a finite covering of Ω such that Ω is a Lipschitz graph in B δ i (x i ). Therefore, when is small, P (Q k ) is small and thus diam(q k ) is small. Hence, by the Lebesgue s number lemma, for those Q k such that Q k Ω, Q k is either contained in Ω or in some B δi (x i ). Since Ω is the graph of a Lipschitz function in B δi (x i ), it follows that P (Q k Ω) P (Q k ) if Q k Ω. Therefore, (3.5) (3.6) and thus (b) is true. E Ω hdx sup k sup k Q k Ω hdx P (Q k ) Q k Ω hdx P (Q k Ω) v h, Remark 3.5. If our conjecture (3.22) were true for any polytope E then (b) would be true for n > 2. We would like to characterize the space S(Ω) in terms of Morrey spaces. It is clear that M n (Ω) S(Ω), since g S(Ω) C(n) g Mn(Ω), by the isoperimetric inequality. However, the space M n(ω) is not the optimal space to guarantee (3.1), see Example 5.1. It turns out that its counterpart M n (Ω) gives the exact characterization of S(Ω), at least for n = 2. The goal of this section is to prove the following: Theorem 3.6. (3.7) S(Ω) S c (Ω) = M n (Ω) For n = 2, (3.7) can be improved to (3.8) S(Ω) = S c (Ω) = M 2 (Ω). Before proving Theorem 3.6, we need the following: Theorem 3.7. If K is a bounded convex set in R n. Then there exists a finite collection of n- dimensional cubes Q k, and a universal constant C(n), such that K k Q k and (3.9) Σ k P (Q k ) C(n)P (K). Remark 3.8. One may ask whether there exists a partition, instead of a covering, that satisfies (3.9). For n = 2 and K a rectangle, it is true. It is not hard to prove that there exists a universal constant

8 8 QINFENG LI AND MONICA TORRES C (C can be chosen to be 6), such that K can be partitioned into a countable collection of squares {Q k } k=1 such that K = kq k, Qk Q l =, and Σ k=1p (Q k ) CP (K). For general convex set K, however, it is not possible to use cubes as a partition as in the theorem above, even for n = 2 and K be a triangle. Indeed, let the base [a, b] of the triangle K lies on the axis OX. Now project all squares onto OX. We claim that all but countably many points in the segment [a, b], are covered infinitely many times. Indeed, assume that the point c (a, b) is not a projection of any vertexes of any square, whose cardinality is countable, then the vertical section {x = c} of our triangle is covered by infinitely many squares, hence the claim is true. Then this means that total sum of projections is infinite by Fubini s Theorem, hence total sum of perimeters of squares is infinite, which definitely cannot be bounded by the perimeter of P (K) up to a universal constant. Remark 3.9. Theorem 3.7 is the key to prove Theorem 4.5 later in this paper, which directly implies the existence of the generalized Cheeger set for n = 2 under very weak conditions (see Corollary 4.6). To prove Theorem 3.7, we need the following two lemmas. Lemma Let 1 k n 1. Let R be an n-dimensional rectangle of the form R = [0, a] k [0, a k+1 ] [0, a n ], a a k+1... a n ; i.e. the smallest k edge(s) of R have the same length. Then there exists a finite number, say m, of n-dimensional rectangle {R i } m i=1, and a universal constant C 1 (n) such that R m i=1 Q i and Σ m i=1p (R i ) C 1 (n)p (R). Moreover, for each R i, the smallest k + 1 edges have the same length. Proof. We have R = [0, a] k [0, a k+1 ] [0, a n ]. By assumption, there exists m 1 such that (3.10) Let R = [0, a k+1 m (3.11) P ( R) = 2 We estimate P ( R) P (R) (3.12) Now we define [ R i := 1 m + 1 < a 1 a k+1 m. ]k [0, a k+1 ] [0, a n ]. Then the right half of (3.10) implies R R. Hence [ k( a k+1 m )k 1 a k+1... a n + Σ n j=k+1( a k+1 a ] m )k k+1... a n. a j P ( R) 2(a k 1 a k+1... a n ) [ ] = k( a k+1 m )k 1 a k+1... a n + Σ n j=k+1 ( a k+1 m )k a k+1...a n a j a k 1 a k+1... a n, by (3.11) = k( a k+1 am )k m (a k+1 ma )k 1 Σ n a k+1 a k+1... a n j=k+1 a j a k+1... a n k( m + 1 m )k 1 + (n k) 1 m (m + 1 m )k 1, since a j a k+1 and the left half of (3.10) n(1 + 1 m )k 1 since 1 m 1 n 2 n 2 since m 1 and k n 1 0, a k+1 m ] [ k i 1 m a k+1, i ] m a k+1 [0, a k+2 ] [0, a n ], i = 0,..., m.

9 MORREY SPACES AND GENERALIZED CHEEGER SETS 9 Note that R = m i=1 R i, and the smallest k + 1 edges of R i have the same length. Note also that, for i = 1,..., m 1, R i and R i+1 have the common n-dimensional face [ C i = 0, a ] { k k+1 i k+1} m m a [0, a k+2 ] [0, a n ], i = 1,..., m 1. Clearly H n 1 (C i ) = H n 1 (C j ) := c. Let F be the face {0} [0, a k+1 m ]k 1 [0, a k+1 ] [0, a n ] of R, then clearly (3.13) m c = H n 1 (F ) < P ( R). Hence we have the following estimate (3.14) Therefore, from (3.12) we conclude Σ m i=1 P (R i) P ( R) = P ( R) + (m 1) c P ( R) 2P ( R) P ( R) = 2, by (3.13) Σ m i=1p (R i ) 2P ( R) n 2 n 1 P (R). Lemma Let R be an n-dimensional rectangle in R n, then there exists a finite collection of n-dimensional cubes {Q i } and a universal constant C 2 (n) such that R i Q i and Σ i P (Q i ) C 2 (n)p (R). Proof. Step 1: By Lemma 3.10, there exists m 1 number of n-dimensional rectangle {R i1 } m1 i, and 1=1 a universal constant C 1 (n) such that R m1 i R 1=1 i 1 and Σ m1 i 1=1 P (R i 1 ) C 1 (n)p (R). Moreover, for each R i1, at least two of its edges have the same lenght. Step 2: Now for each R i1, again by Lemma 3.10, there exists m 2 = m 2 (i 1 ) number of n- dimensional rectangle {R i1,i 2 } m2 i such that R 2=1 i 1 m2 i R 2=1 i 1,i 2 and Σ m2 i 2=1 P (R i 1,i 2 ) C 1 (n)p (R i1 ). Moreover, for each R i1,i 2, at least three of its edges have the same length. Also, R m1 i 1=1 m2 i R 2=1 i 1,i 2 and Σ m1 i 1=1 Σm2 i P (R 2=1 i 1,i 2 ) C 1 (n) 2 P (R). If we continue with this process, after n 1 steps we have {R i1,...,i n 1 } 1i1m 1,...,1i n 1m n 1 such that each R i1,...,i n 1 has n edges with the same length, i.e., each R i1,...,i n 1 is an n-dimensional cube. Moreover, R m1 i 1=1 mn 1 i R n 1=1 i 1,...,i n 1 and Σ 1i1m 1,...,1i n 1m n 1 P (R i1,...,i n 1 ) C 1 (n) n 1 P (R). This finishes the proof. Proof of Theorem 3.7. Let w(u) be the width of K in direction u (where u is a unit vector), i.e., w(u) = max u, x y x,y K By [12, Lemma 4.6], w(u) is a continuous function and thus we can define h = min u S n 1 w(u) = w(u 0 ), which is the minimal height of K. If we let π to be the projection along the u 0 direction,

10 10 QINFENG LI AND MONICA TORRES then π(k) is also convex set. Thus, by [35, Lemma 1], there exists an (n 1)-dimensional rectangle R such that π(k) R and (3.15) H n 1 (π(k)) 1 (n 1)! Hn 1 ( R). Therefore, K is contained in an n-dimensional rectangle R with base R and height h. Let l 1,..., l n 1 denote the length of the edges of R. Then by our choice of h we have h l i, i = 1,..., n 1. Hence ( (3.16) P (R) = 2 H n 1 ( R) ) + Σ n 1 j=1 Π i jl i h 2nH n 1 ( R) since, for each j, Π i j l i h P ( R). Hence by (3.15), (3.16) and the fact H n 1 (π(k)) P (K), we have (3.17) P (R) 2n(n 1)!P (K) Then, by Lemma 3.11 applied to R, there exists a finite collection of n-dimensional cubes {Q i } and a universal constant C 2 (n) such that R i Q i and (3.18) Σ i P (Q i ) C 2 (n)p (R) Let C(n) = 2n(n 1)!, then K i Q i, and by (3.17) and (3.18), we have Σ i P (Q i ) C(n)P (K). This finishes the proof of Theorem 3.7. We now proceed to give a proof of Theorem 3.6. Proof of Theorem 3.6. Note that for any cube Q there exists α(n) such that Q 1 1/n = α(n)p (Q). Q Ω Hence g dx α(n)p (Q), which immediately implies S c(ω) M n (Ω). Q Ω g dx = Q 1 1/n Suppose now that g M n (Ω). For any convex set K, by Proposition 3.7, there exists a finite collection of n-dimensional cubes Q k and a universal constant C(n) which depends only on the dimension n, such that K k Q k and Σ k P (Q k ) C(n)P (K). Therefore we have the following estimate K Ω g Σ k Q k Ω g P (K) 1 C(n) Σ kp (Q k ) = C(n)α(n)Σ k Q k Ω g Σ k Q k n 1 n (3.19) Q α(n)c(n) sup k Ω g α(n)c(n) g k Q k n 1 M n (Ω) <, n which shows that g S c (Ω). This implies M n (Ω) S c (Ω), hence we finish the proof of (3.7). We now proceed to prove (3.8). By (3.7), it suffices to prove that S c (Ω) S(Ω). Suppose now that g S c (Ω). By the same argument contained in the explanation in Remark 3.4, it suffices to show that { (3.20) sup E Ω g } : E is a connected polytope <. But this is immediate since P (Ê) for n = 2. The proof is finished.

11 MORREY SPACES AND GENERALIZED CHEEGER SETS 11 Remark For the case n > 2, the result (3.21) S(Ω) = M n (Ω) = S c (Ω) would follow if there existed universal constant C(n) such that we could cover a connected nonconvex polytope E with convex sets {K i } with the property that (3.22) P (K i ) C(n). Indeed, we would have (3.23) E Ω g i=1 C(n) Σ i K g i Ω Σ i P (K i ) C(n) sup i K i Ω g P (K i ) C(n) g Sc(Ω), which implies S c (Ω) S(Ω), and thus (3.21) would be true. However, we are not able to prove or disprove our conjecture (3.22) at this time. We first introduce the following spaces: Definition 4.1. We define (4.1) S(Ω) := {g L 1 (Ω) : lim sup 0 and (4.2) Sc (Ω) := 4. Existence of generalized Cheeger sets { g L 1 (Ω) : lim sup P (K) 0 E Ω g dx = 0, E is a set of finite perimeter K Ω g dx P (K) = 0, K is a convex set Remark 4.2. The arguments used in (3.20) and (3.4) imply that { S(Ω) = g L 1 E Ω (Ω) : lim sup g dx } = 0, E is a connected polytope. 0 Remark 4.3. Clearly S(Ω) S c (Ω) and, for n = 2, S(Ω) = S c (Ω). Indeed, for a connected polytope E, P (Ê), where Ê is the convex hull of E. Thus, E Ω g dx Ê Ω g dx P (Ê). Since 0 implies P (Ê) 0 then Remark 4.2 immediately implies that S c (Ω) S(Ω). Hence, we conclude (4.3) S(Ω) = Sc (Ω), n = 2. We now proceed to show the relation between the Morrey spaces introduced in Section 3 and the spaces S(Ω) and S c (Ω). Theorem 4.4. If n 2, (4.4) M p (Ω) S c (Ω), p > n. }. }.

12 12 QINFENG LI AND MONICA TORRES Proof. Let g M p (Ω), p > n. For a convex set K, we apply Theorem 3.7 to obtain cubes Q k and a universal constant C(n) such that (3.9) holds. Then, (4.5) Q k = (α(n)p (Q k )) n n 1 (C(n)α(n)P (K)) n n 1, by (3.9). Therefore, we have the following estimates: K Ω g Σ k Q k Ω g P (K) 1 C(n) Σ kp (Q k ), by (3.9), = C(n)α(n)Σ k Q k Ω g Σ k Q k n 1 n Q C(n)α(n) sup k Ω g k Q k n 1 n [ ] = Q α(n)c(n) sup k Ω g Q k Q k p 1 k 1 n 1 p p (4.6) Since 1 n 1 p > 0, result. K Ω g P (K) Immediately, from Remark 4.3, we also have Theorem 4.5. If n = 2, α(n)c(n) g M p (Ω) (C(n)α(n)P (K)) n n 1 ( 1 n 1 p ), by (4.5). 0 if P (K) 0. Hence, from Definition 4.1, we conclude the desired (4.7) M p (Ω) S(Ω), p > 2. In particular, Theorem 4.5 implies the existence of generalized Cheeger sets: Corollary 4.6. (Existence of generalized Cheeger sets) Let n = 2 and let h M p (Ω), p > 2, with h 0 and h not equivalent to zero function. A bounded open set Ω with Lipschitz boundary contains a generalized Cheeger set maximizing (4.8) v h 1 = sup E Ω E h(x)dx Proof. We first note that v1 h > 0. Let E i be a minimizing sequence of (4.8), thus P (E i ) are uniformly bounded for otherwise we would have v1 h = 0. We also have that lim inf i P (E i ) := P >0, for otherwise we would have that, up to a subsequence, P (E i ) 0 and Theorem 4.5 would imply v1 h = 0. Therefore, by the standard compactness theorem for sets of finite perimeter there exists a set E 0 such that, up to a subsequence, E i E 0 in L 1 (Ω) and P (E i ) P. If P (E 0 ) = 0 then E 0 = 0 and hence the dominated convergence theorem yields E i h(x)dx E 0 h(x)dx = 0, P (E i ) P that is, v1 h = 0, which is not possible since v1 h > 0. We conclude that P (E 0 ) > 0 and hence (4.9) v1 h E = lim i h(x)dx = P (E 0) E 0 h(x)dx. i P (E i ) P P (E 0 )

13 MORREY SPACES AND GENERALIZED CHEEGER SETS 13 The lower semicontinuity of the perimeter gives P (E 0 ) P, and hence E 0 h(x)dx P (E 0 ) = P P (E 0 ) v h 1 v1 h. Therefore, P (E 0 ) = P and v h 1 is attained by E 0 in the maximization (4.8). The following example shows that in the critical case h M n (Ω), there is no definite conclusion as to whether generalized Cheeger set exists. Example 4.7. Let Ω = B 1 (0). Let h 1 (x) = n 1 x and h 2 (x) = n 1 x 1. Hence, h 1, h 2 0, and it is easy to check that h 1, h 2 M n (Ω) \ q>n M q (Ω). The arguments in [12, Example 7.6] justify the following computation (4.10) E h 1(x)dx = E div x x dx = E x x ν E(x)dH n 1 (x) 1, where "=" holds if and only if E is equivalent to a ball centered at the origin. Hence, any ball B r (0) E h2(x)dx is a maximizer of E (4.11) h 2(x)dx = We note that lim r 0 E h2(x)dx. Similarly, E div x x dx Br (0) h2(x)dx P (B r(0)) E = E x x ν E(x)dH n 1 (x) E 1 E < 1. = lim r 0 (1 C(n)r) = 1. Therefore, by (4.11), the maximum of the functional, which is 1, cannot be obtained. We showed the existence of generalized Cheeger sets assuming that h 0. We now proceed to explain why we can not expect to have existence if this condition is not satisfied. Lemma 4.8. Let g L 1 (Ω), and Λ := {x Ω : lim sup r 0 B r(x) g r n 1 > 0}, then dim Λ n 1, where dim means Hausdorff dimension. If g M n (Ω), then dim Λ = 0. Proof. If g L 1, then by [23, Theorem 2.4.3], H n 1 (Λ) = 0, and hence dim Λ n 1. For g M n (Ω), since M n (Ω) can be embedded into any L p (Ω), 1 p < n, by applying Holder s inequality, it follows that H n p (Λ) = 0, for all 1 p < n. Hence, dim Λ = 0. Remark 4.9. Lemma 4.8 implies that if h L 1 (Ω), and h 0 in a neighborhood of x, then for H n 1 P (B -a.e. y in the neighborhood, lim inf r(y)) r 0 =, and hence a generalized Cheeger set h Br (y) with h 0 in most cases cannot exist. We end this section by pointing out that, if conjecture (3.22) in Remark 3.12 can be justified, the results in this section also hold for n > 2. We believe that this covering result is also true in higher dimensions. However, at this point, we can only point the reader to our Remark 8.3, which provides a weaker condition than could replace (3.22). 5. On the embedding problem of Morrey spaces into L p spaces The Morrey space M p (Ω) is also defined in [26, Page 164]. It has applications in the study of Sobolev spaces and PDE theory. We recall that M p (Ω), p > 1 (i.e., the weak L p space), can be embedded into any space L q (Ω), 1 q < p (see Remark 2.13). Thus, it is natural to ask whether M p (Ω) has the same property, since both are defined in a similar manner. If this is not true, since both L p (Ω) and M p (Ω) become larger as p gets smaller, we ask the question whether it is possible

14 14 QINFENG LI AND MONICA TORRES to embed M p (Ω) into some L q (Ω) for p is sufficiently large and q sufficiently close to 1. Clearly, M p (Ω) M p (Ω) for any p > 1, and thus it is also natural to ask whether M p (Ω) = M p (Ω). In this section, we construct an example which will rigorously show that for any Ω R n and 1 < q < p, there exists g M p (Ω) \ M q (Ω). This answers negatively the questions raised in the previous paragraph. Example 5.1. We let Ω be an open bounded set with Lipschitz boundary. For any 1 < q < p, there exists 0 < t < s < 1 such that q = 1 1 t and p = 1 1 s. Now for such t and s, we choose α such that (5.1) 1 s s < α < 1 s s t. By the second inequality of (5.1), we have α(1 t) < (α + 1)(1 s), and hence we can choose β such that (5.2) α(1 t) < β (α + 1)(1 s). By the first inequality of (5.1), (α + 1)(1 s) < α, and thus (5.2) yields (5.3) β < α. Therefore, we can choose γ such that (5.4) 0 < γ < α β. We let Q k be the n-dimensional cube with Q k = k γ 1, then diam( Q k ) 1 and Σ k=1 Q k <. By the well known Auerbach-Banach-Mazur-Ulam theorem (see [35]), there is a cube Q, such that { Q k } can be put into the cube Q in a manner that no two of the cubes { Q k } k=1 intersect. Without loss of generality, up to rescaling the size of the cubes Q k by a constant, we may assume that Ω contains a neighborhood of Q. We now let Q k be the n-dimensional cube such that Q k = k α 1. As γ < α, Q k < Q k, and thus we can put Q k into Q k. We let the center of each Q k coincide with the center of each Q k, and the faces of each Q k be parallel to the faces of the corresponding Q k. We now define (5.5) g := k β on Q k and g = 0 otherwise. Clearly g is integrable, since g L 1 (Ω) = k kβ α 1 <. We will show that for any cube Q, <. Given any Q, we note that there are three possible situations. First, if Q does not Q Q k g Q s intersect any of the cubes Q k, then Q Ω g Q s = 0. Second, if Q Q k for some k, then Q Q k g Q s kβ Q Q k Q Q k s (5.6) = k β Q Q k 1 s k β Q k 1 s = k β (α+1)(1 s) <, by the right half of (5.2). Finally, if Q is not contained in any of the cubes Q k, we let I Q = {k : Q Q k > 0}. For k I Q, by the geometry of Q k and Q k, there exists a universal constant C such that (5.7) Q Q k Q Q k C k γ 1 k α 1 = Ckα γ.

15 MORREY SPACES AND GENERALIZED CHEEGER SETS 15 Therefore Q Ω g Q s k I Q k β Q Q k ( k I Q Q Q k ) s k I c Q k β Q Q k ( k I Q k α γ Q Q k ) s, c = 1, by (5.7) Cs k I c Q k β Q Q k ( k I Q k β Q Q k ) s, since α γ > β c( k β Q Q k ) 1 s k I Q (5.8) c( k k β α 1 ) 1 s < Therefore, for any n-dimensional cube Q, Q Ω g dx Q Ω = g Q 1 1/p Q s < C(n) <. That is, g M p (Ω). However, let E K = k=k Q k E K g (5.9) E K 1 1 q = E K g E K t = Σ k=k kβ k α 1 (Σ k=k k α 1 ) t Kβ α K αt = K β α(1 t) as K, by the first inequality of (5.2). Therefore, g / M q (Ω). Remark 5.2. Example 5.1 will also serve as a model example in the study of the critical case g M n (Ω) in section on the regularity of generalized Cheeger sets for h M p (Ω), p > n. In this section, we study the regularity of generalized Cheeger sets in an open bounded Lipschitz domain Ω. A a generalized Cheeger set in Ω is a minimizer of the problem (6.1) C0 h := inf J h (F ), J h (F ) = P (F ) F Ω is a set of finite perimeter. F Ω hdx, We recall that C h 0 = 1/v h 1. F Remark 6.1. From Remark?? and Theorem 3.6 we see that, for n = 2, C0 h > 0 if and only if h M n (Ω). For n 3 and Ω convex, C0 h > 0 if and only if h M n (Ω). Hence, at least when Ω is convex, a necessary condition to guarantee the existence of generalized Cheeger sets in Ω is that h M n (Ω). This motivates our analysis with Morrey spaces.

16 16 QINFENG LI AND MONICA TORRES We observe that, if the set of finite perimeter E minimizes J h, that is, J h (E) = C0 h, then E is also the minimizer of the restriction of the functional (2.2) to Ω, that is, (6.2) I H (E) I H (F ) = P (F ) + H(x)dx, F Ω, where H(x) = C h 0 h(x). Throughout the rest of the article we continue to assume that sptµ E = E (see [36, Remark and Proposition 12.9]). We now state the first result concerning the regularity of generalized Cheeger sets. Theorem 6.2. If Ω is a Lipschitz domain in R n, h M p (Ω), p > n, and J h (E) = inf F Ω J h (F ), where J h is defined as above, then Ω E is a C 1,α -hypersurface, where α = p n 2p. Moreover, the singular set has dimension at most n 8 (and it is empty if n 7). Proof. Let C0 h = inf F Ω J h (F ). Clearly C0 h > 0. We consider the functional I H (F ) = P (F ) + F H(x)dx with H = Ch 0 h(x). We note that I H is nonnegative and I H (E) = 0, and hence E is a minimizer of I H. Moreover, H M p (Ω). Let F be a set such that E F B r (x) Ω. By the minimality of E, (6.3) P (E; B r (x)) + H P (F ; B r (x)) + H, and using that H M p (Ω) we obtain (6.4) P (E; B r (x)) P (F ; B r (x))+ E B r(x) B r(x) F F B r(x) H P (F ; B r (x))+cr n n p = P (F ; Br (x))+cr 2α r n 1, p n where C = ω n H M p (Ω) and 0 < α = 2p < 1 2. It was shown in Tamanini [43] and Bombieri [15] that if a set E is a quasi minimizer in the sense that (6.5) P (E; B r (x)) P (F ; B r (x)) + η(r)r n 1 for all variations F of the set E such that F E B r (x) where η : (0, r 0 ) [0, ) satisfies (6.6) η(0+) = 0, η(r) is increasing on (0, r 0 ) r and r0 η(r) (6.7) dr <, r 0 then E can be split into the union of a C 1 relatively open hypersurface and a closed singular set with Hausdorff dimension at most n 8 (empty if n 7). Moreover, given x E, there exist C, r > 0 such that ( x y ) η(r) (6.8) ν E (x) ν E (y) C + x y 1/2, for all y B(x, r) E. r 0 In our situation, from (6.4) it follows that our generalized Cheeger set E is a quasi minimizer in the sense of Tamanini (6.5) with η(r) cr 2α. Moreover, we note that η satisfies the required conditions (6.6) and (6.7). Furthermore, from (6.8) we conclude that E can be split into the union of a C 1,α relatively open hypersurface and a closed singular set with Hausdorff dimension at most n 8 (empty if n 7). Corollary 4.6 immediately implies the following: Theorem 6.3. If Ω is a Lipschitz domain in R 2 and h M p (Ω), p > 2, then inf F Ω P (F ) F h(x)dx is attained by a set E of finite perimeter. Moreover, Ω E is a C 1,α -hypersurface, where α = p n 2p.

17 MORREY SPACES AND GENERALIZED CHEEGER SETS 17 Corollary 6.4. If h L p P (F ) (Ω), p > n, then inf F Ω is attained by a set E of finite perimeter. h(x)dx F Moreover, there exists a closed set Σ(E) with Hausdorff dimension not greater than n 8 (empty if p n 1, n 7), such that ( E Ω) \ Σ(E) is a C 2p (n 1)-dimensional hypersurface. Proof. That the infimum is attained by a set E Ω follows from [12, Corollary 6.2], where the existence for generalized Cheeger sets was proven for h L n (Ω). The regularity follows from the classical theory of variational mean curvatures (see, for example, Gonzalez-Massari [27] and the references therein) or from Theorem 6.2 since L p (Ω) M p (Ω). Corollary 6.5. If h L n (Ω), then generalized Cheeger sets E exist. Moreover, there exists a closed set Σ(E) with Hausdorff dimension not greater than n 8, such that ( E Ω) \ Σ(E) is a C 0,α (n 1)-dimensional manifold for all α < 1. If n = 2, then E Ω is a Lipschitz 1-dimensional manifold. Proof. The existence follows from [12, Corollary 6.2]. The regularity for the critical case h L n (Ω) was proven by Ambrosio-Paolini [6, Theorem 4.10, Theorem 5.2] (see also Paolini [40]). Motivated by the conjecture of De Giorgi discussed in the introduction (see Ambrosio-Paolini [6]) concerning the regularity of boundaries with prescribed mean curvature in the critical space L n, we would like to know if the same regularity holds for generalized Cheeger sets with h in the critical space M n (Ω). We first note that a generalized Cheeger set is also a quasi minimizer in the sense of Ambrosio-Paolini [6] since, from (6.5), it follows that (6.9) P (E; B r (x)) (1 + ω(r))p (F ; B r (x)), where ω(r) = c(n)r2α 1 c(n)r, α = n p 2α 2p. However, for the critical case p = n, ω(r) becomes a constant and hence the condition (6.10) lim r 0 ω(r) = 0 is not satisfied. Thus, we can not proceed as is [6], where this condition was critical to obtain the regularity described in Corollary 6.5. Indeed, the dimension estimates for the singularities of the minimizers in Corollary 6.5 depend heavily on the fact that, if h is L n integrable, then the blow-up limit E is a minimal set with mean curvature H = 0 (see [28, Theorem 1.1]). Then, the dimension estimates for the singularities follow from the non-existence of minimizing cones for n 7 and Federer s dimension reduction argument (see, for example, [36, Chapter 28]). However, in order to show that the blow-up limit has zero mean curvature, ω(r) must be infinitesimal, that is, the condition w(0+) = 0 is needed. We finish this section with a discussion about the critical case h M n (Ω). The next Example 6.6 shows that the strong regularity described in Theorem 6.2 is not true in general for the critical case h M n (Ω). We will also show in the next section (see Lemma 7.16) that our structure theorem for generalized Cheeger sets implies upper density estimates on points of the boundary of generalized Cheeger sets. However, the following Example 6.6 also shows that the lower density bounds can fail for h M n (Ω). We will provide in Section 8 a stronger example (see Example 8.5), which shows that, even for h M n (Ω) M n (Ω), the set of points where the lower density estimate fails can be very large, and could even have Hausdorff dimension n. Although Example 8.5 gives more information, Example 6.6 provides motivations for the structure theorem (Theorem 7.7) of generalized Cheeger sets and gives understanding about their behavior in the critical case h M n (Ω). Example 6.6. This example is a continuation of Example 5.1. Let Ω be a bounded open set with Lipschitz boundary and let n = 2. Let s = 1 1 n. Then, for any 1 < q < n, we can choose 0 < t < s < 1 with t = 1 1 q. 1 s As in Example 5.1, we let s < α < 1 s s t and choose

18 18 QINFENG LI AND MONICA TORRES β = (α + 1)(1 s) = α+1 n. Thus, we choose 0 < γ < α β and define Q k and Q k as in Example 5.1. We now let (6.11) h := k k β χ Ck Then, proceeding as in Example 5.1, h / M q (Ω), h M n (Ω) and (5.7) holds. Let C Q be the Cheeger constant of the unit cube, and C k be the classical Cheeger set in Q k. Thus, by scaling of perimeter and volume, (6.12) P (C k ) C k P (F ) = inf = C Q F Q k F l(q k ), where the function l(q k ) denotes the length of the edge of the cube. We note that l(q k ) = k α+1 n = k β, and hence P (C k ) C k h = P (C k) k β = k β C Q C k l(q k ) = C Q. Let E = k C k, then J h (E) = k P (C k) C k h = C Q. k For any indecomposable set F E, by Theorem 2.8, F Q k (mod H n ) for some k. Hence (6.13) P (F ) P (F ) = k β h F F k β P (C k) C k = C Q. For general F E, let {F k } be the indecomposable components of F. Hence F k E, P (F k) F k C Q, and hence J h (F ) = k P (F k) F k h C Q. k Therefore J h (E) = inf F E J h (F ). We now claim that E satisfies (7.6). Indeed, this is because each C k is convex (from which we can apply [36, Exercise 15.14]) and the distance between any two of the sets C k is large compared to the diameter of each C k, and hence, if F connects two cubes Q i, Q j, the part of F whose projections to the direction of the edges of each Q k intersect Q k, but are not contained in any cube Q k, has much larger length outside the union of the cubes Q k. Here we are not going to write carefully, but one can use the projection argument as in [12, Example 7.3] to verify the aforementioned steps carefully, for n = 2. Therefore, since h 0 on E and h = 0 on E c, by Lemma 7.12, E satisfies (7.13). Hence Proposition 7.10 implies that E satisfies (7.5), and thus E also satisfies (7.7). Thus, we conclude that E is a generalized Cheeger set for this h in Ω. Next we proceed to show the strong regularity of E fails by finding a point x 0 spt Dχ E E 0. The x 0 will be constructed as follows. Let x k be the center of Q k. Since all the cubes Q k are contained in a cube Q Ω, we can find x 0 such that, up to a subsequence, x k x 0. Clearly x 0 spt Dχ E. Let Q r be the cube centered at x 0 with edge lenght r. Then, for any r, by the geometry of Q k and Q k there exists a universal constant C 1 such that Qr Q k Q r Q k C 1k α γ (see (5.7)). Let I(r) = {k : Q r Q k } and

19 MORREY SPACES AND GENERALIZED CHEEGER SETS 19 m(r) = inf I(r), then m(r) as r 0. We have the following E B r r n E Q r Q r = k I(r) C k Q r Σ k I(r) Q k Q r C k Q r sup k I(r) Q k Q r Q k Q r < sup k I(r) Q k Q r (6.14) 1 C 1 m(r) γ α 0 This implies x 0 E structure of generalized Cheeger sets The previous section was concerned with the differentiability properties of the boundaries of generalized Cheeger sets. In this section, we are concerned about the shape of the sets, and in particular their property of being pseudoconvex. We recall that, for the variational mean curvature problem, given any set of finite perimeter E, there exists a function H E L 1 such that E has mean curvature H E. Therefore, since sets of finite perimeter can have very rough boundaries, the same is true for minimizers of this problem. However, for the generalized Cheeger set problem, we will show in Theorem 7.7 that minimizers have the extra property of being pseudoconvex. Before stating the main result of this section, we make some definitions: Definition 7.1. Let Ω R n be a bounded open set with Lipschitz boundary or Ω = R n, and let E Ω be a bounded set of finite perimeter. We say that F is a set with least perimeter in Ω containing E if (7.1) P ( F ) = inf{p (F ) : Ω F E}, or, equivalently P ( F ) P (F ), for all Ω F E. Definition 7.2. A bounded set E R n of finite perimeter is called pseudoconvex if E is a set with least perimeter in R n containing E, that is (7.2) = inf{p (F ) : R n F E}. A pseudoconvex set E is also called a subsolution to the least area problem. If E R n is a bounded set of finite perimeter, then its minimal hull E m is defined as E m = S J E S, where J E denotes the family of pseudoconvex sets in R n that contain E. Remark 7.3. From [8, Proposition 1.1], if each E i, i = 1, 2,.., is a pseudoconvex set, then i E i is also a pseudoconvex set and, if in addition E i E i+1, then i E i is pseudoconvex. Hence, if E is a convex set, then E is a pseudoconvex set (since a convex set is the intersection of hyperplanes). We note that the disjoint union of two pseudoconvex sets is not necessarily pseudoconvex (consider for example two disjoint parallel cubes sufficiently close to each other). Remark 7.4. The previous remark implies that E m is a pseudoconvex set (see [10, Proposition 2.2]) containing E and contained in the convex hull of E, denoted as Ê. Moreover, E m (see [10, Proposition 2.2]) is a set of least perimeter containing E, that is, (7.1) holds: (7.3) P (E m ) = min{p (F ) : E F },

20 20 QINFENG LI AND MONICA TORRES which implies (7.4) P (E m ) and P (E m ) P (Ê). We refer the reader to [24, Lemma 4] for the proof of the following theorem, which is very useful in dealing with the case n = 2. This theorem is not true in higher dimensions, as shown in Remark 7.6. Theorem 7.5. Let E R 2 be an indecomposable bounded set. If E is not equivalent to a convex open set, then there exists a bounded set F such that E F and P (F ) < Remark. Theorem 7.5 implies that Ê = E m in R 2. Moreover, E m = E if and only if E is a convex set. Remark 7.6. We note that Theorem 7.5 is not true for n 3. In order to see this, we consider, in dimension 3, an open connected set E such that the perimeter of its convex hull Ê is strictly greater than the perimeter of E. Such a set does not exists in dimension 2, but in dimension 3 one may consider a set with the shape of a banana. We consider now the minimal hull E m of E. If Theorem 7.5 were true in dimension 3 then E m would be convex, but if E m were convex then E m would contain Ê and therefore [36, Exercise 15.14] and (7.4) would yield P (Ê) P (E m Ê) = P (E m), which that P (Ê) >. Therefore, in dimension 3, when considering a set E with P (Ê) >, the minimal hull E m is not convex. In this section our main result is the following: Theorem 7.7. (Structure Theorem) Let E be a bounded set of finite perimeter with E > 0. Then, E minimizes (7.5) J h (E) = inf J h (F ), J h (F ) = P (F ) F R n F h(x)dx for some h L 1 (R n ), h 0, if and only if E is pseudoconvex, that is, (7.6) = inf{p (F ) : F E}. Moreover, if E is a minimizer for some J h defined on R n then, each indecomposable component E i of E is also a minimizer for the same J h and satisfies P (E i ) = inf{p (F ) : E i F }. In particular, if n = 2, then E i is H 2 -equivalent to a convex open set. Remark 7.8. We recall that a set E satisfying (7.5) is called a generalized Cheeger set. If instead we consider our problem in a bounded Lipschitz domain Ω, we similarly have Theorem 7.9. Let E Ω be a set of finite perimeter with E > 0, where Ω is a bounded Lipschitz domain. Then, E minimizes (7.7) J h (E) = inf J h (F ), J h (F ) = P (F ) F Ω F h(x)dx for some h L 1 (Ω), h 0, if and only if E satisfies (7.8) = inf{p (F ) : Ω F E}. Moreover, if E is a minimizer for some J h defined on Ω then, each indecomposable component E i of E also minimizes the same J h and satisfies P (E i ) = inf{p (F ) : Ω F E i }. In particular, if Ω is convex and n = 2, then E i is H 2 -equivalent to a convex open set. In order to prove Theorem 7.7, we need some preliminary results. Proposition Let J h be defined as in (7.5). If J h (E) J h (F ) for either F E or F E, then J h (E) = inf F R n J h (F ).