OPTIMALITY OF THE TRIANGULAR LATTICE FOR A PARTICLE SYSTEM WITH WASSERSTEIN INTERACTION D.P. BOURNE, M.A. PELETIER, AND F. THEIL Abstract. Weprovestrongcrystallizationresultsintwodimensionsforanenergythatarises inthetheoryofblockcopolymers. Theenergyisdefinedonsetsofpointsandtheirweights,or equivalentlyonthesetofatomicmeasures. Itconsistsoftwoterms;thefirsttermisthesumof the square root of the weights, and the second is the quadratic optimal transport cost between the atomic measure and the Lebesgue measure. We prove that this system admits crystallization in several different ways: (1) the energy is bounded from below by the energy of a triangular lattice (called T); (2) when the energy equals that of T, then the measure is a rotated and translated copy of T; (3) when the energy is close to that of T, then the measure is close to a rotated and translated copy of T. These 3 three results require the domain to be a polygon with at most six sides. A fourth result states 1 that the energy of T can be achieved in the limit of large domains, for domains with arbitrary 0 boundaries. 2 The proofs make use of three ingredients. First, the optimal transport cost associates to n each point a polygonal cell; the energy can be bounded from below by a sum over all cells of a a function that depends only on the cell. Second, this function has a convex lower bound that is J sharpatT. Third,Euler’spolytopeformulalimitstheaveragenumberofsidesofthepolygonal 1 cells to six, where six is the number corresponding to the triangular lattice. ] P A . h t 1. Introduction a m 1.1. The setting. Many materials achieve their state of lowest energy with a periodic arrange- [ ment of the atoms: their ground states are crystalline. Many other systems also favor ordered, 2 periodic structures; examples are packed spheres [HHM+10], convection cells (e.g. [KP74]), v reaction-diffusion systems [Hoy06], higher-order variational systems (e.g. [LSAC08]) and also 3 block copolymers, the system that inspired the energy that we study in this paper. On the 7 9 other hand, there are also many examples of deviation from periodicity: entropy may over- 6 rule order, defects may appear, and even non-periodic ground states exist, as in the case of . 2 quasicrystals [Bar09]. 1 It follows that the question whether and why a given system favors periodicity is a non- 2 1 trivial one. It is also an important one, since many material properties depend strongly on the : microscopic arrangement of atoms or particles. And it is a surprisingly hard question to answer. v i In two and three dimensions, the strongest results are available for two-point interaction X (cid:80) energies of the form V(x −x ). In the case of hard-sphere repulsion the triangular ar- r i(cid:54)=j i j a rangement in two dimensions is easily recognized as optimal, but the highest-density stacking of spheres in three dimensions was computed by Hales in 1998 in a proof that is still being formal- ized [HHM+10]. For various Lennard-Jones-like interaction potentials V with sufficiently short range it has been proved that global minimizers in two dimensions are triangular under appro- priate boundary conditions [Rad81, The06, AYFS12]. E and Li show that addition of suitable three-pointinteractionsshiftsthegroundstatefromthetriangulartoahexagonal lattice[EL09]. For systems with more general interactions between the particles, however, we know of no rigorous results; in this paper we study a system in this class, and prove several strong crystal- lization results. LetΩ ⊂ R2 befixedsuchthat|Ω| = 1. Thesystemisdescribedbyafinitenumberofpoints z i in Ω and their masses v , or equivalently by an atomic measure, i.e. a positive measure µ of the i 1 form (cid:88) (cid:88) µ = v δ , with v > 0 and v = 1, (1) z z z z z∈Z z∈Z where Z is any finite subset of Ω. The (unscaled) energy of the system is Eˆλ(µ) = λ(cid:88)µ({z})12 +W(LΩ,µ). z∈Z Here L is the Lebesgue measure on Ω. The function W is the quadratic optimal transport Ω cost; see [Vil03] for an extensive introduction to this topic. For our purposes it is sufficient to define W(L ,µ) for µ of the form (1): Ω (cid:26)(cid:90) (cid:27) W(L ,µ) := inf |x−T(x)|2dx : T: Ω → Z, and T−1(z) = µ({z}) for all z . (2) Ω Ω By, e.g., [Vil03, Theorem 2.12] there exists an optimal map T. This system arises as a highly stylized model for block copolymer melts. The copolymers consist of two parts, called the A and B parts; the A and B parts strongly repel each other, leading to phase separation, but since they are connected to each other by a covalent bond, the phases have to be microscopically mixed. In the regime described here, the B parts have much larger volume than the A parts, and therefore the A parts congregate into small balls represented by the points z; the B parts fill the remaining volume. The masses v = µ({z}) are z the relative amount of A at the point z. The two terms in Eˆ represent the two important contributions to the energy. The first term λ measures the (rescaled) interfacial area separating the two phases; since the A phase resembles a 1/2 smallballofvolumev ,itsinterfacialareaisproportionaltov . Thesecondtermisanenergetic z z penalty for a large separation between the A and B parts: the map T maps a B particle to its corresponding A particle, and |x−T(x)|2 measures the energy of the covalent bond (modeled by a linear spring) connecting the two particles. We discuss the modeling background of this system in more detail in [BP]. This system has a number of distinguishing features. (1) It is a system of ‘particles’ that interact with each other via the nonlocal functional W. This nonlocal functional potentially allows each particle to interact with all other parti- cles simultaneously. This makes it different from particle systems with two-, three-, or four-particle interactions. (2) Each particle carries a ‘weight’ µ({z}) that influences the interaction. (3) There is no imposed length scale: the length scale is determined in the competition between the two terms, much as in the case of other block copolymer models [ACO09, Cho01, CPW09, Mur10]. (4) The number of particles is not fixed in advance: it also arises from the trade-off between the terms. Wewillseebelowthatforminimizersthenumberofparticlesscalesapproximatelyasλ−2/3. In the limit λ → 0, therefore, the typical number of particles for a minimizer becomes unbounded. Numerical calculations suggest that in this limit the particles organize themselves in a regular triangular pattern, as illustrated in Figure 1. The aim of this paper is to characterize and prove this phenomenon of crystallization. Tobeconcreteweprovefourresultsthateachcharacterizesthephenomenonofcrystallization in a different way. We assume that Ω is a polygon with at most six sides. (1) An energy bound: We show that for any λ > 0 the energy of an arbitrary configuration is bounded from below by the energy of an optimal triangular lattice (Theorem 1). (2) The energy bound is sharp: In the limit λ → 0 this bound can be obtained; or equiva- lently, that for fixed λ, the bound can be reached in the limit of large domains (Theo- rem 2). 2 5 points 7 points 10 points 1 1 1 0 0 0 0 1 0 1 0 1 64 points 128 points 256 points 1 1 1 0 0 0 0 1 0 1 0 1 Figure 1. An illustration of the appearance of a hexagonal pattern as λ → 0. As λ decreases, the optimal number of points z ∈ Z increases as λ−2/3, and they organize in a nearly-triangular lattice. The polygons surrounding the points are the cells T−1(Z) (see Section 2), and they approximate regular hexagons as λ → 0. (These figures are not calculated by minimizing Eˆ , which is computationally λ challenging. Instead, the number of points in Z is fixed beforehand, and the reduced energy W(L ,µ), which we study in Section 2.5, is minimized over all Ω µ of the form (1) using Lloyd’s algorithm. The connection between these two minimization problems is discussed in more detail in the companion paper [BP]. However, Theorems 1–3 give a rigorous version of the convergence suggested by these figures.) (3) Exact crystallization: If the energy bound is achieved exactly, then the structure is exactly triangular with the optimal separation between the points (Theorem 3). (4) Geometricstability: Iftheenergyboundisnotexactlyachieved,butthegapinthebound is small in the limit λ > 0, then the structure is asymptotically triangular (Theorem 3). Some of these results also hold for other domains. For the precise statement of these results we first introduce some notation. 1.2. Setting up the results: Rescaled energy. We scale space in such a way that small λ and largedomains become the samething. The new domainwill have(two-dimensional) volume √ (cid:18) (cid:19)2 2c6 3 5 3 V := , where c = = 0.160375... (3) λ 6 λ 54 The constant c is central in this work, and we will comment on it later. For fixed Ω ⊂ R2 with 6 |Ω| = 1 we therefore define the scaled domain 1/2 Ω := V Ω, (4) λ λ 3 −1/2 and for given µ ∈ P(Ω) we define a rescaled measure µ ∈ M (Ω ) by µ(A) := V µ (V A) 0 ≥0 λ λ 0 λ for any Borel set A. Under this rescaling the energy Eˆ becomes, up to a factor λ4/3(2c )−4/3, λ 6 E : M (Ω ) → R, λ ≥0 λ (cid:40) (cid:80) 1 E (µ) := 2c6 z∈Zµ({z})2 +W(LΩλ,µ) provided µ is atomic and µ(Ωλ) = |Ωλ| = Vλ, λ ∞ otherwise. This is the energy that we shall consider throughout this paper. In this scaling, we expect µ to consist of O(V ) points, each with O(1) mass, and spaced at distance O(1). The crystallization λ results below are a much stronger version of this statement. 1.3. Results. Throughout the rest of this paper, the energy functionals will always be defined with respect to a set Ω which is constructed as in (4) out of a unit-area set Ω and a parameter λ λ > 0. Theorem 1. Let Ω be a polygon with at most six sides with |Ω| = 1. Then for all λ > 0 we have the lower bound E ≥ 3c V . (5) λ 6 λ Asweshowbelow,theright-handsidein(5)istheenergyofastructureofV regularhexagons λ of area 1. This lower bound can also be achieved, and this is even possible for more general sets Ω: Theorem2. LetΩbeabounded, connecteddomaininR2 with|Ω| = 1andsuchthat∂Ω = ϕ(K) for some Lipschitz function ϕ : R → R2 and some compact set K ⊂ R. Then lim V−1 infE = 3c . (6) λ λ 6 λ→0 The triangular lattice T with density 1 is defined as (cid:26) (cid:18) (cid:19) (cid:27) 1 2 1 T = √ k : k ∈ Z2 ⊂ R2. (7) 121/4 0 3 Whentheinequality (5)issaturatedornearlysaturated, thenthestructureisexactlytriangular or nearly so: Theorem 3. Assume the conditions of Theorem 1. (a) If E (µ) = 3c V , then µ is an atomic measure with all weights equal to 1, and suppµ λ 6 λ a translated and rotated copy of the triangular lattice T. (b) Define the dimensionless defect of a measure µ on Ω as λ d(µ) := V−1E (µ)−3c . λ λ 6 There exists C > 0 such that for λ < C−1 and for all µ with d := d(µ) ≤ C−1, suppµ is O(d1/6) close to a triangular lattice, in the following sense: after eliminating CV d1/6 λ points, the remaining points have six neighbors whose distance lies between (1−Cd1/6) and (1+Cd1/6) of the optimal distance 21/23−3/4. Part(a)ofTheorem3isanaturalcounterpartofpart(b), whichshouldapplywhend(µ) = 0. In this case, since Ω is a polygon with at most six sides, this assertion is nearly empty: the only domain Ω for which equality can be achieved is the case when Ω is a regular hexagon of area λ 1 and λ = 1, and Ω ∩T is a single point at the origin. λ However, the methods of this paper can be extended to the case of ‘periodic domains’, and we give an example here. Let us define T to be a rectangle [0,γ)×[0,γ−1) with area one and periodic boundary conditions, or more precisely, as the two-dimensional Riemannian manifold T = R2/(γZ×γ−1Z). As before, T is the blown-up version of T, and the energy E has a natural analogue Eper on λ λ λ T . λ 4 Theorem 4. If Eper(µ) = 3c V , then µ is an atomic measure with all weights equal to 1, and λ 6 λ suppµ a translated and rotated subset of the triangular lattice T. Naturally, equality can only be achieved if the size and aspect ratio of T are commensurate λ with the periodicity of the triangular lattice. 1.4. Discussion. In this section we comment on a number of similarities and differences with other results. Exact and approximate crystallization. IntheintroductionwementionedtheresultsofRadin, Theil, and Yeung-Friesecke-Schmidt [Rad81, The06, AYFS12] on exact crystallization for sys- tems of points in the plane. In one dimension there are many more results that prove that minimizers of some functional are exactly periodic; examples are the block copolymer-inspired systemsstudiedbyMu¨ller[Mu¨l93]andRenandWei[RW00],theSwift-Hohenbergenergy[PT96], (cid:80) and two-point interaction systems of the form V(x −x ) [Ven78, VN79]. i,j:i(cid:54)=j i j An important class of related functionals in two dimensions arises from the ‘location problem’ or ‘optimal configurations of points’ (see, e.g., [BSS]). An example of such a problem is (cid:90) (cid:2) (cid:3) inf G (Z), where G (Z) = minf(|x−z|) dx. (8) Ω Ω Z⊂Ω Ω z∈Z |Z|=n Here f : [0,∞) → R is a given non-decreasing function and n is given. When f(r) = r2, then this problem is in fact identical to (cid:88) inf I (χ ) Ω i χ∈C(Ω) i in the notation of Section 2 (see equations (11) and (12)). For problem (8), a variety of different crystallizationresultsexist. L.FejesT´othshowedthatifthedomainΩisapolygonwithatmost sixsides, theexpression(8)isboundedfrombelowbyntimesthesameexpressioncalculatedfor aregularhexagonofarea|Ω|/n(Theorem11belowisaversionofthis;see[FT72,Gru99,MB02]). G. Fejes To´th gave an improved version that includes a stability statement [FT01], which we include below as Lemma 8. Although ‘optimality’ in this location problem is defined differently than optimality for the energy E of this paper, the two ‘energies’ are close enough to allow λ the results by the two Fejes T´oth’s to be applied to the structures of this paper. These results therefore figure centrally in the arguments below. Boundaries have positive energy. OneinterpretationofTheorems1and2isthatanimperfect boundary contributes a positive energy to the system, provided it does not have too many sides; ontheotherhand,asweshallseebelow,curvedboundariescanactuallybebetterthanpolygonal ones. This boundary penalization is similar to the case of Lennard-Jones-type potentials, but different from the case of fully repulsive potentials. Neighborsandtheconnectivitygraph. ForLennard-Jonessystemsoneoftendefines‘neighbors’ of a point x as those points x such that V(x −x ) is close to minV. Although this definition i j i j containsanarbitrarychoiceof‘closeness’, itworkswellbecauseflatgeometrycreateshardlimits on how many neighbors there may be (six in two dimensions, twelve in three). In the system of this paper, no such limit exists; a point can have an arbitrarily large number of neighbors, and indeed this is energetically favorable for that point (but not for the others), as we shall see below. Instead of a local limit on the number of neighbors, there is a global limit of a graph-theoretic nature: Euler’s polyhedral formula limits the average number of neighbors to six. For this property to hold, the boundary ∂Ω should not introduce too many vertices and sides, and this is the origin of the restriction in Theorems 1 and 3 on the number of sides of Ω. The Abrikosov lattice. Our problem and the location problem also have strong links with vor- tex lattice problems like the Abrikosov lattice, which is observed in superconducting materials. More precisely, we say that a function h is in the admissible class A2 if −∆h = µ−1 in R2, (9) 5 for some positive measure µ of the form (cid:88) µ = δ , z z∈Z where Z ⊂ R2 is countable. In [SS12] a renormalized Coulomb energy Sˆ2(h) is associated to h Ω and a domain Ω ⊂ R2. The renormalized energy has the property that (cid:16) (cid:17) lim (cid:107)∇hε(cid:107)2 −(cid:107)∇hε(cid:107)2 = Sˆ2(h )−Sˆ2(h ), 1 L2(Ω) 2 L2(Ω) Ω 1 Ω 2 ε→0 if h ,h are admissible, hε is a mollification of h and #(Z ∩Ω) = #(Z ∩Ω). 1 2 1 2 Itisconjecturedin[SS12]thattherenormalizedenergydensitylim R−2Sˆ2 isminimized R→∞ RΩ by a triangular lattice (Z = T), which is interpreted as the Abrikosov lattice in the context of superconductivity. This conjecture admits a natural generalization where (9) is replaced by the p-Laplacian (∆ h = div(|∇h|p−2∇h)). We say that for an atomic measure µ with µ(Ω) = |Ω| p the function h ∈ W1,p(Ω) is in the admissibility class Ap(µ) if −∆ h = µ−1 in Ω and ∂h = 0 on ∂Ω if 2 < p < ∞, p ∂ν h(x) = |x−T(x)| if p = ∞ and there exists a map T : Ω → suppµ such that |T−1({z})| = µ({z}) for z ∈ suppµ. The energy of h ∈ Ap is defined by (cid:26) 1(cid:107)∇h(cid:107)p if 2 < p < ∞, Sp(h) = p(cid:48) Lp(Ω) Ω (cid:82) h(x)dx if p = ∞. Ω We conjecture that the infimum of lim min R−2Sp (·) is realized by the triangular R→∞ h∈Ap RΩ (cid:80) lattice T, i.e. if µ = δ , then T z∈T z lim R−2 min Sp (h) RΩ R→∞ h∈Ap(αRΩµT) (cid:26) (cid:27) =inf lim R−2 min Sp (h) : µ = (cid:88)δ , Z ⊂ R2 countable and lim α (µ) = 1 , RΩ z RΩ R→∞ h∈Ap(αRΩµ) z∈Z R→∞ |Ω| where α (µ) = is a normalization factor. Fejes T´oth’s result in the case f(r) = |r| in Ω µ(Ω) equation (8) implies that the conjecture is true if p = ∞ and Ω is a polygonal domain with at most 6 sides. The definition of S∞ is motivated by the observation that the minimum of Sp over Ap(µ) Ω Ω admits an unconstrained variational characterization if p > 2. Proposition 5. Let µ be an atomic measure such that µ(Ω) = |Ω|. Then min Sp(h) = Γp(µ) (10) Ω Ω h∈Ap(µ) holds for all 2 < p ≤ ∞, where (cid:110) (cid:111) sup (cid:82) φdµ− 1 (cid:82) |∇φ|pdx : φ ∈ W1,p(Ω), (cid:82) φ = 0 2 < p < ∞, Γp(µ) = Ω p Ω Ω Ω sup(cid:8)(cid:82) φdµ : (cid:107)∇φ(cid:107) ≤ 1, (cid:82) φ = 0(cid:9) p = ∞. Ω L∞ Ω Proposition 5 provides a homotopic connection between the physically interesting functional S2 andthefunctionalS∞ forwhichmathematicallyrigorousanalysisoftheasymptoticbehavior Ω Ω of minimizers is available. The presence of the connection suggests that the Abrikosov lattice is optimalforsufficientlylargepandoffersastrategyfortheconstructionofrigorousmathematical proofs. The proof is given in Section 5. 6 2. Preliminaries 2.1. Cells and an alternative formulation. A central concept in this work is that of cells, which can be seen in Figure 1. These cells arise from the definition (2) of W: the cell associated with any z ∈ Z is the set T−1(z), where T is the optimal map in (2). In Lemma 6 we show that for any µ, these cells are separated by straight lines, and Figure 1 illustrates this. Note that when the cells are exactly hexagonal, the points z are arranged in a triangular lattice, and vice versa. In fact there is a useful alternative formulation of this system in terms of the cells themselves. Define the set of partitions C(Ω) of Ω by (cid:40) n (cid:41) (cid:88) C(Ω) = χ ∈ L∞(Ω;{0,1})n for some n ≥ 1 : χ = 1 on Ω . (11) i i=1 Now define the alternative energy functional n (cid:34) (cid:18)(cid:90) (cid:19)1/2 (cid:35) (cid:88) F : C(Ω ) → R, F (χ) = 2c χ +I (χ ) , λ λ λ 6 i Ωλ i Ω i=1 λ where I (χ) is defined for any set U ⊆ R2 and function χ ∈ L∞(U;{0,1}) by U (cid:90) I (χ) = inf |x−ξ|2χ(x)dx. (12) U ξ∈U U The formulations in terms of µ and of χ are strongly related. One can construct one out of the other as follows: • Given µ with support Z and transport map T, define the partition χ by setting, for each z ∈ Z, χ to be the characteristic function of the set T−1(z), so that (cid:82) χ = µ(z); z z (cid:80) (cid:0)(cid:82) (cid:1) • Given χ = {χ } , let z achieve the optimum in (12); then set µ = χ δ . i i∈I i i∈I i zi There is loss of information going from one to the other, and in general this transformation does notpreserveenergy. However,minimizersaremappedtominimizers,asthefollowingcalculation shows. Given a µ, construct the corresponding χ as above; then (cid:90) (cid:88) E (µ) = 2c µ({z})1/2+ |x−T(x)|2dx λ 6 Ω z∈Z λ (cid:88)(cid:16)(cid:90) (cid:17)1/2 (cid:88)(cid:90) = 2c χ + |x−z|2χ (x)dx 6 z z Ω Ω z∈Z λ z∈Z λ (cid:88)(cid:16)(cid:90) (cid:17)1/2 (cid:88) (cid:90) ≥ 2c χ + inf |x−ξ|2χ (x)dx = F (χ). (13) 6 z z λ z∈Z Ωλ z∈Zξ∈Ωλ Ωλ The inequality above becomes an identity if we minimize the left-hand side over all choices of the support points Z of µ. It follows that inf E = inf F , and that minimizers are converted µ λ χ λ into minimizers. 2.2. Cells can be assumed to be polygonal. Thefollowinglemmainoptimaltransportation theory shows how the minimization in the definition (2) of W causes cells to be polygonal. Lemma 6 (Cells are polygonal). Let µ ∈ M (Ω ) be atomic with µ(Ω ) = |Ω | and define ≥0 λ λ λ Z = supp(µ). Let T be the optimal transport map for W(L ,µ). Then there exists numbers Ω λ (cid:96) ∈ R such that for all z ∈ Z z T−1(z) = {x ∈ Ω : (cid:96) +|x−z|2 ≤ (cid:96) +|x−z(cid:48)|2 for all z(cid:48) ∈ Z}. (14) λ z z(cid:48) Moreover, if µ minimizes E , then (cid:96) = c µ({z})−1/2. λ z 6 Thecharacterization(14)impliesthatforgivenµ,thecorrespondingcellscanbecharacterized astheintersectionofΩ withafinitenumberofhalf-planes. Cellsthatdonotmeettheboundary λ 7 ∂Ω are therefore convex polygons; cells adjacent to a piece of curved boundary have a mixture λ of straight and curved sides. In this paper we refer to both cases as convex polygons. A characterization related to (14) appears in a number of places [AHA98, M´11], and can be proved using Brenier’s theorem characterizing optimal transport [Bre91]. It shows that the transport cells T−1(z) form the power diagram of the set of points Z with weights −(cid:96) , z and provides a link between optimal transportation theory and computational geometry. Since Lemma 6 is a slightly stronger statement, we give an independent proof in Section 5. 2.3. Optimal energy for polygons. We first discuss the minimum energy for polygonal do- mains. Define the number (cid:8) (cid:9) c = inf I (χ) : χ is the characteristic function of a polygonal with n sides and area 1 n R2 χ (cid:26) (cid:90) (cid:27) = inf min |x−ξ|2dx : P is a polygon with n sides and area 1 . P ξ∈P P (15) A classical result by L. Fejes T´oth [FT72, p. 198] states that the minimizing n-gon is a regular n-gon: Lemma 7 (Regular polygons are optimal). The minimum in (15) is attained by a regular polygon with n sides, and in particular (cid:18) (cid:19) 1 1 π π c = tan +cot . (16) n 2n 3 n n The minimum is unique up to rotation and translation. Note that the number c , defined in (3), equals c for n = 6. If χ is the characteristic function 6 n of a regular n-gon with volume v contained in a domain Ω, then I (χ) = v2c . By Lemma 7, Ω n if χ is the characteristic function of an irregular n-gon with volume v contained in a domain Ω, then I (χ) ≥ v2c . (17) Ω n G. Fejes T´oth proved a stability result for a large number of polygons that applies to the situation at hand. We reproduce a consequence of the main theorem of [FT01] here: Lemma 8 (Geometric stability). Let Ω be a polygon of unit area with at most six sides, and let λ > 0. Let χ = {χ } be a polygonal partition of Ω . Set i i=1,...,N λ (cid:80) I (χ )−Nc ε := i Ωλ i 6. Nc 6 There exists ε and C > 0 such that if 0 < ε < ε then the following holds. Except for at most 0 0 Cε1/3 indices i, all χ are O(ε1/3) close to unit-area regular hexagons, in the sense that suppχ is i i a hexagon and the distances from the center of mass to the vertices and to the sides are between (1−Cε1/3) and (1+Cε1/3) of the corresponding values for a unit-area regular hexagon. Note that this lemma implies a similar statement on the centers of mass: if z is the center of i mass of χ , thus achieving the minimum in (12), then apart from a fraction Cε1/3, all of the z i i have exactly six neighbours at distance (1±Cε1/3) of the optimal lattice spacing. 2.4. The average number of edges of a polygonal cell. Lemma 6 shows that the optimal transport map T gives rise to a partition of Ω by convex polygons. Therefore Euler’s polytope λ formula applies: vertices−edges+faces = 2. In the proofs of Theorems 1 and 3 we will use the following lemma, which follows from Euler’s polytope formula: Lemma 9 (Bound on the average number of edges of the polygons). Assume that Ω ⊂ R2 is a polygon with at most six sides. Consider a partition of Ω by convex polygons. Then the average number of edges per polygon is less than or equal to six. 8 Proof. A proof of this is given in Morgan & Bolton (2002, Lemma 3.3) for the case where Ω is a square. The proof for 3-, 5- and 6-gons is almost identical, and we only give it here for completeness. Let S ∈ {3,4,5,6} denote the number of sides of Ω. Let the tiling of Ω consist of n convex polygons P. Denote the number of edges of polygon P by N(P). Let N denote the number of 0 exterior edges. Alltheinterioredgesmeettwofaces,whereastheexterioredgesmeetonlyoneface. Therefore the total number of edges e can be written as (cid:88) N(P) N0 e = + . (18) 2 2 P Since the tiles are convex, each interior vertex lies on at least three faces. The exterior vertices, except possibly the S corners of Ω, lie on at least two faces. Therefore we can bound the total number of vertices v by (cid:88) N(P) N0 S v ≤ + + . (19) 3 3 3 P Euler’s formula gives (cid:32) (cid:33) (cid:32) (cid:33) (cid:88) N(P) N0 S (cid:88) N(P) N0 2 = v−e+(n+1) ≤ + + − + +(n+1). (20) 3 3 3 2 2 P P Since N ≥ S it follows that 0 1 (cid:88) 6−S N(P) ≤ 6− ≤ 6. (21) n n P (cid:3) For the proof of Theorem 2, where ∂Ω is the image of a Lipschitz function, we will need a different version of Lemma 9: Lemma 10 (Bound on average number of edges for a planar graph). Let G be a planar graph such that the degree of each vertex is at least three. Then the average number of edges per face is less than six. Proof. This is proved by simply taking S = 0 in (19)–(21) in the proof of Lemma 9. (cid:3) 2.5. L. Fejes T´oth’s Theorem. For pedagogical purposes we consider a simpler setting where the surface energy, the first term of E , is dropped, i.e., the case λ = 0. Minimizers of this λ energy are shown in Figure 1, for the case where the number of points in Z is fixed beforehand. We give a short proof of a special case of a classic result by L. Fejes T´oth [FT72]. Theorem 11. Let Ω ⊂ R2 be a polygon with at most 6 sides such that |Ω| = 1. Then #supp(µ)W(L ,µ) ≥ c Ω 6 for all atomic probability measures µ that are supported on a finite set. Moreover inf{#supp(µ)W(L ,µ) : µ atomic probability measure with finite support} = c . (22) Ω 6 Proof of Theorem 11. Let µ be an atomic measure. By Lemma 6 the characteristic functions χ are supported on polygonal domains, T−1(z). Let n ∈ {3,4,...} be the number of sides of z z T−1(z). Lemma 7 implies that we can reduce the energy of µ by replacing each polygon T−1(z) with a regular polygon with the same number of sides and the same area: (cid:90) (cid:88) (cid:88) (cid:88) W(L ,µ) = |x−z|2χ dx ≥ I (χ ) ≥ v2c Ω z Ω z z nz Ω z∈Z z∈Z z∈Z 9 √ by equation (17), where v = (cid:82) χ . Define κ = ∂cn| = 2π − 5 3 < 0. Define g(v,n) = v2c . z z ∂n n=6 243 324 n By computing the Hessian of g one can show that g is convex in (v,n): 8π2v2sec2(cid:0)π(cid:1) ∂2g det(D2g) = n > 0, = 2c > 0. 9n6 ∂v2 n Hence for each v ≥ 0 one finds that 0 v2c ≥ c v2+2v c (v−v )+κv2(n−6) (23) n 6 0 0 6 0 0 for all v ≥ 0, n ∈ {3,4,...}. This implies that W(L ,µ) ≥(cid:88)(cid:0)c v2+2v c (v −v )+κv2(n −6)(cid:1) Ω 6 0 0 6 z 0 0 z z∈Z (cid:32) (cid:33) (cid:88) ≥c v2|Z|+2v c −2v2c |Z|+κv2 n −6|Z| , (24) 6 0 0 6 0 6 0 z z where we have used that (cid:80)v = |Ω| = 1. Substituting v = |Z|−1 into (24) gives z 0 (cid:32) (cid:33) W(L ,µ) ≥ c6 +κv2 (cid:88)n −6|Z| . Ω |Z| 0 z z Lemma 9 implies that (cid:80) n ≤ 6|Z|. Recall also that κ < 0. Therefore W(L ,µ) ≥ c6 as z z Ω |Z| required. To prove the upper bound we define χ to be the characteristic functions of the Voronoi- z tessellation of R2 that is associated with the set Zm = m−12T ⊂ R2, m ∈ N, where T is the triangularlatticedefinedinequation(7). Wewillcheckthatthefollowingsequenceofprobability measures (µ )∞ achieves the infimum in (22): m m=1 (cid:90) (cid:88) µ = δ χ . m z z Ω z∈Zm It is easy to check that v := (cid:82) χ = m−1 if suppχ ⊂ Ω and v < m−1 otherwise. Also, for all z Ω z z z z ∈ Z , we have m (cid:90) c |x−z|2χ dx ≤ 6 , (25) z m2 Ω with equality if suppχ ⊂ Ω. Furthermore it can be shown that z 1 b(m) = #{z ∈ Z : ∅ =(cid:54) suppχz ∩Ω (cid:54)= suppχz} ≤ Cm2 (26) for some universal constant C (which depends on H1(∂Ω)). Therefore by (25) we obtain (cid:90) W(L ,µ ) ≤ (cid:88) |x−z|2χ dx ≤ (m+b(m)) c6 ≤ c6 +C c6 . Ω m z m2 m m3/2 Ω z∈Zm Since #supp(µ ) ≤ m+b(m) this proves that the lower bound (22) can be achieved with the m sequence µ . (cid:3) m Remark. WewillseethattheproofofTheorem1mimicstheproofofTheorem11. Theimportant difference is that for E the function f(v,n) that corresponds to g(v,n) in the proof of Theorem λ 11 is not convex (see equation (28) for the definition of f). We circumvent this lack of convexity byprovingthataconvexityinequalityoftheform(23)stillholdsifv issufficientlylarge: v ≥ m 1 (Lemma 12). Then we prove in Lemma 13 that if µ is a minimizer of E , then v > m for all λ z 1 z and so the convexity inequality applies. 10