Crystallization in the Winterbottom Shape and Sharp Fluctuation Laws Manuel Friedrich, Leonard Kreutz, Ulisse Stefanelli Abstract We address finite crystallization in two dimensions in the presence of a flat crystalline substrate. Particles interact through short-range two- and three-body potentials favoring local square-lattice arrangements. An additional interaction term of relative strength β>0 couples the particles and the substrate. Our first main result proves crystallization for all β>0, corresponding to the onset of discrete Winterbottom configurations. The proof relies on a stratification technique from [], characterizing the topology of the bond graph of minimizing configurations. Our second main result concerns fluctuations estimates for β∈(0,1). We obtain bounds on the distance between distinct minimizers with the same number N of particles, showing a sharp scaling law N3/4 when β is rational, and N1/3 when β is irrational and algebraic. This reveals a genuine substrate-driven effect on fluctuation laws. As a corollary, we derive a discrete-to-continuum convergence of minimizers towards the Winterbottom equilibrium shape in the large-particle limit. Introduction At low temperatures, matter typically exhibits crystalline order. In this regime, the interactions between atoms and molecules can be well approximated by configurational potentials. In the zero-temperature limit, it is usually assumed that matter organizes into optimal configurations, minimizing such configurational energies. Proving mathematically that such optimal configurations are periodic (crystalline) is precisely the aim of the so-called crystallization problem []. Despite considerable attention, rigorous crystallization results remain scarce. To date, crystallization of a finite number of particles (finite crystallization) has been established in one and two spatial dimensions, under various assumptions on the interaction potentials, see Sect. below. In this paper, we investigate the discrete Winterbottom problem, namely, the determination of equilibrium particle configurations in a crystal in contact with a substrate []. Specifically, we establish the first crystallization result in the presence of a flat substrate. Restricting to two space dimensions, we consider a finite system of N particles interacting at short range through two- and three-body potentials. These interactions are designed to favor local arrangements with up to four nearest-neighbors bonds forming π/2 angles between them. In addition, the particles interact with a fixed flat crystalline substrate. We are interested in characterizing optimal configurations of such particles. This framework is inspired by epitaxial growth where the crystal develops layer by layer on a substrate. The relative intensity of the particle-particle and particle-substrate interaction is described by a parameter β>0. Our results are twofold. At first, we prove crystallization for all β>0, see Theorem . To the best of our knowledge, this constitutes the first two-dimensional finite-crystallization result relative to a prescribed substrate. The crystallization proof relies on the stratification technique introduced in [], see Sect. below. Due to the substrate interaction, the minimizing Winterbottom configurations differ from minimizers in the absence of a substrate, in agreement with the predictions of the Winterbottom problem. Note that for β≥1 the system falls into the so-called wetting regime, see Remark , where minimizers flatten against the substrate and hence are less interesting. Secondly, for β∈(0,1) we prove sharp fluctuations estimates, see Theorem . As is typical in crystallization problems, minimizers are nonunique for most values of N [, , ]. We prove matching upper and lower bounds on the distance between minimizers, depending on β. More precisely, if β is rational, we prove that minimizers can differ up to CN3/4 particles for some constant C>0. This scaling coincides with that observed in other two-dimensional lattice regimes without a substrate [, , , ]. In contrast, for β irrational but algebraic, minimizers may differ up to Cβ,δN1/3+δ particles, for arbitrarily small δ>0 with Cβ,δ→∞ as δ→0. This a genuine effect of the presence of the substrate of fluctuation laws. In a much simplified one-dimensional setting, the sensitivity of ground-state configurations on the value of β has been already observed in []. An immediate consequence of our sharp fluctuation results is the characterization of the macroscopic shape of minimizers in the large-system limit N→∞. We establish a quantitative discrete-to-continuum convergence of Winterbottom configurations towards the corresponding macroscopic Winterbottom shape, that is, the equilibrium shape of a macroscopic crystal in contact with the substrate. Relation with the literature In two dimensions, finite crystallization of hard spheres into the triangular lattice was first established by Heitman & Radin [], building on a result by Harborth []. The same result was later reobtained independently by De Luca & Friesecke [] through an entirely different approach, see Sect. below. The case of soft spheres was addressed by Radin [] and Wagner [], and subsequently extended by Del Nin & De Luca []. Two-body interactions in combination with three-body ones can lead either to the square lattice [] or the hexagonal one [], depending on the specific choices. Among these, [] introduces a new proof technique via stratification, which is the one used in this paper as well, see Sect. below. Finite crystallization of ionic dimers systems has been obtained both in the square case [] and in the hexagonal case []. De Luca, Ninno, & Ponsiglione [] deal with a finite-crystallization setting including orientations. Apart from the very special setting considered in [], no finite-crystallization result in three dimensions is currently available. To the best of our knowledge, our result in Theorem is the first finite-crystallization theorem in which the interaction with a substrate is taken into account. Note that the crystallization problem takes another flavor in the thermodynamic limit N→∞. One-dimensional results can be found in [, , ], while the stability of periodic one-dimensional configuration, or lack thereof, is discussed in []. In two dimensions, Theil [] proved that certain long-range two-body interactions lead to crystallization in the triangular lattice. Crystallization as N→∞ has been obtained by Bétermin, De Luca, & Petrache [] in the square-lattice case and by Farmer, Esedoḡlu, & Smereka in the hexagonal one []. In three dimensions, Flatley & Theil [] proved that the face-centered-cubic lattice arises when a specific selection mechanism on next-to-nearest neighbors interactions is imposed, see also []. A related computation of the corresponding Wulff shape for N→∞ is in []. The defective case has also been considered. The emergence of rigid polycrystaline structures has been tackled in []. Dislocations in discrete structures and their coarse-graining have been studied under different assumptions on the interaction energy in [–, ], among others. Yet another different setting is that of lattice-crystallization, where one considers the best lattice for a given lattice energy. Here, the literature is vast and the reader is referred to [, , , ] for references. Before the present work, the only study dealing with finite crystallization in the presence of a substrate was the already mentioned []. There, a one-dimensional hard-sphere finite-crystallization problem is addressed, under the influence of a periodic background modeling substrate interaction. In contrast to the setting considered here, [] assumes that the substrate and the crystallizing particles favor different lattice parameters. Uniqueness in finite-crystallization problems occurs only for specific values of N. In case of nonuniqueness, the study of the distance of two distinct minimizers (up to lattice translations) has led to different results, depending on the underlying lattice structure. For the triangular lattice, Schmidt [] proved that minimizers can differ by CN3/4 particles, confirming a conjecture from []. This has been revisited in [], where a sharp constant C is identified. Cicalese & Leonardi [] proved the same law by a different approach based on quantitative isoperimetric inequalities [, , ]. Their method extends to Zd, as well, providing the upper bound CN1-1/2d. This bound, however, is not sharp for d>2, as was shown by identifying the sharp regime, first in the cubic lattice Z3 [] and eventually by Mainini & Schmidt [] in any dimension. The fluctuation law CN3/4 has also been shown to hold for the square lattice [], the hexagonal lattice [], and ionic dimer systems [, ]. To the best of our knowledge, Theorem provides the first fluctuation estimates in the presence of a substrate. Under the assumption of crystallinity, the Winterbottom shape has already been identified in the setting of the discrete double-bubble problem in the square lattice in [], see also [], and the continuous counterparts [, , ]. A discrete-to-continuous justification of the emergence of a hexagonal Winterbottom shape, emerging from a discrete model with two crystals with mismatched lattice parameters, was given by Piovano & Velčić in [, ]. The stability of Winterbottom shapes was investigated in [–]. In contrast to these works, the novelty of our contribution lies in proving crystallization rather than assuming it. We note, however, that our assumptions on the local geometry of the crystalline lattice are comparatively more stringent. Stratification technique In the two-dimensional case, finite-crystallization for hard spheres has often been obtained by adapting the induction method over bond-graph layers by Heitman & Radin []. This approach determines the fine geometry of the minimizers by inductively considering the relative effect of boundary vs. bulk particles in the configuration. A second elegant technique to prove finite-cristallizaton for hard spheres has been introduced by De Luca & Friesecke [], based on discrete geometry. Here, a notion of discrete combinatorial curvature is associated to the natural bond graph of the configuration, and a discrete Gauß-Bonnet-like theorem is applied. However, neither of these techniques appears to extend naturally to the case of particle interactions with a substrate. This is particularly evident for the induction method over bond-graph layers [], relying on the idea that adding (or removing) a boundary layer to a minimizer should maintain minimality – a property which is simply false in the presence of a substrate. A key technical difference of the present work compared to earlier ones is our use of an alternative argument based on stratification. This method, first introduced by the first two authors in [], provided an alternative proof of the finite-crystallization result for the square lattice of []. We now give a heuristic overview of the stratification technique, postponing details to Sect. below. Using purely variational arguments, we first show that minimizers are regular: every bond between two neighboring particles has lenght approximately 1, and the angle formed by two adjacent bonds is approximately a multiple of π/2. To each such regular configuration we associate its strata, namely all bond paths that are approximately straight. The crystallization result follows by characterizing the topology of these strata. In particular, due to the presence of the substrate, we can distinguish between those strata that are interacting or noninteracting with it. We prove that two interacting strata, or two noninteracting ones, cannot intersect, and that each interacting stratum crosses all noninteracting ones. This structural property allows us to reconstruct the global topology of the bond graph of minimizers, ultimately proving Theorem . The stratification approach may be applied to other lattices and higher dimensions, as well. It would also be interesting to consider the case of a nonflat substrate. With respect to our setting, these generalization seem however to imply some relevant technical complications. Discrepancy theory Previous results on fluctuation estimates have relied either on direct manipulations of minimizing configurations [, , , ] or on quantitative discrete isoperimetric inequalities [, , , ] (or both). Extending these techniques to the case of a substrate is not straightforward. On the one hand, the substrate restricts the possibility of directly manipulating minimizers. On the other hand, discrete isoperimetric inequalities in the presence of a substrate are currently not available. In this paper, we are hence forced to follow a different path, essentially based on discrepancy theory for sequences []. This plays a crucial role to treat the case β∈(0,1) irrational but algebraic. In particular, it is used to prove the following fact: for all N0 large enough, one can find at most Cβ,δN02/3+δ and at least cβ,δN02/3-δ minimizers with N particles for N∈[N0,2N0] taking the form of an exact rectangle (i.e., arrangements of particles in Z2 with equal rows and equal columns). Here, Cδ→∞ and cδ→0 as δ→0. Moreover, we prove that any two such optimal rectangles differ in the number of particles at least by cβ,δN01/3-δ and at most by Cβ,δN01/3+δ. As a consequence, all minimizers with N∈[N0,2N0] are at most Cβ,δN01/3+δ far from an optimal rectangle. At the same time, one finds N∈[N0,2N0] such that the closest optimal rectangle has distance at least cβ,δN01/3-δ and uses such a rectangle to the same lower bound on the fluctuation. The case of a nonalgebraic irrational β is not covered by our approach with discrepancy theory, and is hence left open at present. Structure of the paper The model is introduced in Sect. , where we also state the main results. In particular, Theorem and Theorem contain the crystallization and the fluctuation results, respectively. Section presents the stratification technique used in the analysis and applies it to derive both local and global properties of minimizers. The proof of a technical lemma is postponed to the Appendix. Finally, Sect. contains the proof of the main results. Specifically, Theorem is proved in Sect. and Theorem is proved in Sect. . Setting and Main Results We consider particle systems in two dimensions, and model their interaction by classical potentials in the frame of Molecular Mechanics [, ]. Let β>0 be given and define the substrate by L-=Z2∩{x2≤0}. Indicating a configuration of particles by CN={x1,...,xN}⊂R2∩{x2>0}, we define its energy by where v2 and v3 denote two-body and three-body interaction potentials, respectively, which are specified below and depicted in Fig. . The third sum runs over triples (x,y,z)∈(CN∪L-)3 such that |x-y|,|y-z|0 and N∈N, it will be convenient to consider the normalized energy 2(Fβ(CN)+2N) which essentially counts the number of missing bonds at the free surface of the configuration and the number of bonds between configuration and substrate (plus the angle part, if CN⊄Z2). In this regard, we further define the normalized minimal energy and the largest optimal height by We now state the main results of the paper. Theorem 2.1 (Crystallization). For all β>0 there exists Nβ∈N such that for each CN∈(R2)N with N≥Nβ it holds that In case of equality, we have CN⊂Z2∩{x2>0} and #CN∩{x2=1}≥N/h∗(β,N). A configuration of minimal energy is depicted in Fig. . Remark 2.2 (Finer characterization of minimizers). Given N≥Nβ, minimizers CN of Fβ satisfy some more geometric properties: Remark 2.3 (β≥1 : Wetting regime). If β≥1, it is elementary to check that h∗(β,N)=1. Thus, Theorem and Remark show that in this case minimizers are horizontal integer shifts of the single chain of particles {(i,1):i=1,...,N}, and the minimal energy is given by -(N-1)-βN. In order to state the next theorem, we introduce a reference rectangle Theorem 2.4 (Fluctuations). Let β∈(0,1) and CN⊂Z2∩{x2>0} be a minimizer of Fβ. Then the following upper bounds hold true: Moreover, the following lower bounds hold true: The result provides a fluctuation law of N3/4 in the case of rational β and a law of N1/3 for irrational, algebraic β. The case of a transcendental interaction parameter is not addressed here. Some configurations satisfying the optimal fluctuation estimates are depicted in Fig. . A direct consequence of the sharp fluctuation estimates of Theorem is the characterization of the Winterbottom shape emerging in the large-particle limit N→∞. Arguing along the lines of [] and using also Lemma below, one may prove the following. Corollary 2.5 (Winterbottom shape). Let β∈(0,1). The optimal height h∗(β,N) defined in () satisfies N-1/2h∗(β,N)→1-β as N→∞. Letting CN={x1N,⋯,xNN}⊂Z2∩{x2>0} be a minimizer of Fβ with min{x1j:j=1,⋯,N}=1, and defining the sequence of rescaled empirical measures as N→∞, the measures μN converge weakly* to the restriction of the Lebesgue in R2 to the rectangle Stratification After a short preliminary on graph theory, this section is devoted to detail the main technique of this paper, namely a modification of bond graphs, called stratification. Bond graph Let G=(V,E) be a graph, where V indicates the set of vertices and E⊂{{x,y}:x,y∈Vandx≠y} is the set of edges. For our purposes, we shall always assume that graphs are embedded in R2. For x∈V, we denote the (graph) neighborhood with respect to G and the (graph) neighborhood with respect to L- by for r0>0 as given in (ii2). For all such graphs G=(V,E) we define where is the bond energy and is the excess energy. For V′⊂V, let G[V′] be the (vertex) induced subgraph of V′ in G, that is G[V′]=(V′,E′) with E′={{x,y}∈E:x,y∈V′}. We will identify each configuration CN⊂R2 with its natural bond graph Gnat=(V,Enat), where V=CN and the natural edges are given by This definition is motivated by the relation to the energy Fβ defined in (), namely In Sect. below, we will successively modify Enat to a smaller set of edges E⊂Enat, according to a set of given rules. Definition 3.1 We say that G=(V,E) is ε-regular for ε>0 small if the following two conditions hold: - (i)For x,y∈V∪L- with x≠y it holds that |x-y|≥1-ε; - (ii)For each bond angle θ=θx,y,z with x,y,z∈V∪L- it holds that θ∈[π/2-ε,π/2+ε]∪[π-ε,π+ε]∪[3π/2-ε,3π/2+ε]. Note that, if Gnat=(V,Enat) is ε-regular, then G=(V,E) is ε-regular for all E⊂Enat. Lemma 3.2 There exists ε0>0 such that the following holds true: if v2, v3 satisfy (i2)–(iii2) and (i3)–(iv3) for some 0<ε<ε0 and if CN is a minimizer of (), then its natural bond graph Gnat=(V,Enat) is ε-regular. Moreover, it holds that #N(x,Enat)+#NL-(x)≤4 and #NL-(x)≤1 for all x∈V. The proof of this statement is similar to the one of [, Lemma 3.2]. We include it in for convenience of the reader. For the remainder of this paper, we assume that ε0>0 is chosen small enough such that Lemma holds true and that v2 and v3 satisfy (i2)–(iii2) and (i3)–(iv3) for some 0<ε<ε0. Stratified bond graph Given G=(V,E), we say that γ=(x1,...,xn) with xi∈V for all i=1,...,n is a straight path if n≥2 and the following holds: - (i){xi,xi+1}∈E for all i=1,...,n-1; - (ii)θi∈[π-ε,π+ε] for all i∈2,...,n-1, where θi=θxi+1,xi,xi-1; - (iii){xi,xi+1}≠{xj,xj+1} for all i,j=1,...,n-1, j≠i. (If n=2, (ii) and (iii) are empty.) The set of straight paths is denoted by We drop G and simply write Γ if no confusion arises. If γ∈Γ and x1=xn, we say that γ is closed and otherwise that γ is open. We define Note that in the second definition one could equally use the angle θx2,x,x1 as θx2,x,x1=2π-θx1,x,x2. We define the set of strata by Here, we set s(x)={(x),(x)} for x∈V0 and s(x)={(x)} for x∈V1∪V2π. Adding the degenerate stratum (x) twice for V0 and once for V1∪V2π has no geometrical interpretation but is merely convenient to relate the number of strata to Fbond,β, see Lemma below. In particular, it will ensure that each particle is contained in exactly two strata, see Lemma (iii). We say that s∈SΓ, s=(x1,...,xn), is an interacting stratum (with the substrate) if there is z0∈NL-(x1) such that θz0,x1,x2∈[π-ε,π+ε] and/or zn+1∈NL-(xn) such that θxn-1,xn,zn+1∈[π-ε,π+ε]. The set of interacting strata is defined by Also here the addition of degenerate strata is convenient to relate the number of strata to Fbond,β. The set of noninteracting strata is given by Sno(G):=S(G)\Sint(G). We say that s∈Sint(G), s=(x1,...,xn), is not double touching if either NL-(x1)≠∅ or NL-(xn)≠∅. We drop G and write S, Sint, and Sno if no confusion arises. Different possibilities of interacting strata are illustrated in Fig. . Definition 3.3 (Length, orthogonal strata, and span). Let s∈S. By l(s):=#s we denote its length. We define the set of orthogonal strata to s by Given s0∈S we define A stratum s∈S and its orthogonal strata are illustrated in Fig. . We proceed with the definition of the angle excess for straight paths and a corresponding lemma. Definition 3.4 (Angle excess). Given γ=(x1,...,xn)∈Γ, we define the angle excess by Lemma 3.5 (Small angle excess). Let G=(V,E) be an ε-regular graph. The following implications hold true: Proof Statements (i)–(iii) follow as in [, Proof of Lemma 3.6]. We first prove (iv). Assume that s=(x1,...,xn) is interacting, say at x1 and z1∈NL-(x1), and assume that s′=(y1,...,ym) satisfies s′∈S⊥(s), i.e., yj=xl for some j∈{1,...,m} and some l∈{1,...,n}. Assume by contradiction that also s′ is interacting, say at ym and zk∈NL-(ym). Consider a straight path γ1=(z1,...,zk) connecting z1 and zk in L-, as well as γ2=(x1,...,xl) and γ3=(yj,...,ym). The closed path (in the sense of graph theory) γ1∘γ2∘γ3 forms a (k+l+m-j)-gon. The sum of its interior angles is (k+l+m-j-2)π. In particular, the interior angles at z2,...,zk-1 are π, the interior angles at z1 and zk are in [π/2-ε,π/2+ε], the interior angle at xl is at least π/2-ε, and the interior angles at x1 and ym are in [π-ε,π+ε]. Thus, as θex(γ1)=0, this implies that θex(γ2)+θex(γ3)≥π2-5ε. This yields a contradiction since the angle excess of both straight paths is smaller than π4-52ε by assumption. We now prove (v). Assume by contradiction that γ=(x1,...,xn) is double touching, say in z1,zk∈L-. Consider a straight path (z1,...,zk) connecting z1 and zk in L-. The path (x1,...,xn,zk,zk-1...,z1) (in the sense of graph theory) forms an (n+k)-gon. The sum of its interior angles is (n+k-2)π. The interior angles at z2,...,zk-1 are π, the interior angles at z1 and zk are in [π/2-ε,π/2+ε], and the interior angles at x1 and xn are in [π-ε,π+ε]. This implies that θex(γ)≥π-4ε, a contradiction. Next, we prove (vi) and refer to Fig. for an illustration. Assume by contradiction that s^1=s^2=:s^. We denote by γ1⊂s1 the path connecting the point s1∩s′ and the point s1∩s2. Similarly, we denote γ2⊂s2 the path connecting the point s1∩s2 and the point s′′∩s2, γ′⊂s′ the path connecting the point s′∩s^ and the point s′∩s1, γ′′⊂s′′ the path connecting the point s′′∩s2 and the point s′′∩s^, and γ^⊂s^ the path connecting the point s′′∩s^ and the point s′∩s^. Note that the path γ′∘γ1∘γ2∘γ′′∘γ^ is a closed and thus, taking into account the fact that all bond angles belong to [kπ/2-ε,kπ/2+ε], k=1,2,3, we obtain that θex(γ1)+θex(γ2)+θex(γ′)+θex(γ′′)+θex(γ^)≥π/2-5ε. This yields a contradiction since the angle excess of all five straight paths is smaller than π10-ε by assumption. Lastly, the proof of (vii) follows like the one of (vi) by taking as s^⊂L- to be the path connecting s′ and s′′, see Fig. , observing additionally that there are two particles bonded to the substrate where the interior angles of the closed path are in [π-ε,π+ε], cf. the proof of (v).□ Remark 3.6 In the following, we assume that ε0>0 is small enough such that π/10-7ε/5≤π/6-4ε for all 0<ε<ε0. We state a consequence of Lemma . (Actually, we only use properties (i), (iv), and (v).) Lemma 3.7 (Graphs with open paths, no double touching). Let G=(V,E) be an ε-regular graph such that maxγ∈Γθex(γ)<π10-75ε. Then, Proof Statement (i) follows exactly as in [, Proof of Lemma 3.3], so we only need to prove (ii) and (iii). We first prove (ii). To this end, recalling () it suffices to prove the claims The proof of (a) can be found in [, Proof of Lemma 3.3]. We prove (b) by induction over m:=#{x∈V:NL-(x)≠∅}. The statement is clearly true for m=0. Let now m≥1 and fix x∈V such that NL-(x)≠∅. We have that #NL-(x)=1 by Lemma . Considering G^=(V\{x},E\{{x,y}:y∈N(x,E)})=(V^,E^) we get #{z∈V^:NL-(z)≠∅}=m-1. We can apply the induction hypothesis to obtain It thus remains to show that #Sint(G^)=#Sint(G)-1. By definition we clearly have Sint(G^)⊂Sint(G). Consider the unique s∈Sint(G) with x∈s (uniqueness follows from Lemma (iv)). It suffices to show that s\{x}∉Sint(G^). If s=(x), this is clear. On the other hand, if s=(x1,...,xn), for n≥2 with x1=x, then, as s is not double touching, we have that (x2,...,xn)∉Sint(G^). Therefore #Sint(G^)=#Sint(G)-1. We now turn to the proof of (iii). Note that, if x∈V3∪V4∪(V2\V2π), then there exist two straight paths passing through x and thus there exist s1,s2∈SΓ that contain x. (Note that all γ are open so the same stratum cannot pass through x twice). If x∈V2π∪V1, there exists a straight path containing x and thus there exists s∈SΓ containing x. By construction, we added the degenerate stratum (x) containing x. Therefore, also in this case there are two strata s1,s2∈S containing x. Lastly, for x∈V0 we added twice the degenerate stratum (x) which contains x. This concludes the proof.□ We now come to the stratification of bond graphs. The following lemma allows to reduce the problem of crystallization to a purely geometric problem of minimizing the number of strata in graphs containing only open strata with small angle excess. Recall ()–(). Lemma 3.8 (Construction of a graph with small angle excess). Let G=(V,E) be ε-regular. Then, there exists Go=(V,Eo) with Eo⊂E such that Proof The proof can be found in [, Proof of Lemma 3.7], replacing π6 by π10. (Compare the assumption in (iii3) with the one in [].)□ We now come to an estimate on Fbond,β for graphs satisfying Lemma (i). This will be the key ingredient for the proof of Theorem in Sect. . Recall Definition . Lemma 3.9 (Estimate on Fbond,β, β∈(0,1)). Let 0<ε<ε0 and β∈(0,1). Let G=(V,E) be an ε-regular graph such that Let s1∈S and s2∈S⊥(s1). Then, the following holds true: Proof By () and Remark we observe that all properties in Lemma hold. Therefore, also the properties in Lemma are satisfied. Proof of (i): First of all, by Lemma (i) all s∈Γ (and thus all s∈S) are open. By Lemma (iii) we have S⊥(s1)∩S⊥(s2)=∅. Moreover, note that, if there is s0∈S⊥(s1) (or s0∈S⊥(s2), respectively) such that s0∈Sint, then, due to Lemma (vii), all s^∈S⊥(s2) (or s^∈S⊥(s1), respectively) satisfy s^∈Sno. Together with Lemma (ii) and Lemma (ii), this implies This concludes the proof of (i). Proof of (ii): If () holds, then we can find x∈V\span(s1). There are two cases to consider: - x∉span(s2); - x∈span(s2). Case (a): In this case, there are two strata s^1,s^2∈S\S⊥(s1)∪S⊥(s2) such that x∈s^1∪s^2 and thus s^1∈S⊥(s^2). By Lemma (iv) at most one of the two strata s^1,s^2 belongs to Sint. Now, as in the proof of (i), we obtain This concludes (ii) in case (a). Case (b): As span(s2)⊊V, there exists z∈V\span(s2). We can assume that z∈span(s1) as otherwise the estimate follows by repeating the argument in (a). Now, due to Lemma (iii), for s1⊥,s2⊥∈S⊥(s2) we have that s1⊥∩s2⊥=∅, and thus there exists exactly one stratum s′′∈S⊥(s2) such that s′′∩{x}≠∅. In the same fashion, there exists exactly one s′∈S⊥(s1) such that s′∩{z}≠∅. We also note that s∩{x}=∅ for all s∈S⊥(s1) and s∩{z}=∅ for all s∈S⊥(s2) since x∈V\span(s1) and z∈V\span(s2). By Lemma (iii), for x∈V there exist two strata s^2,s′′ such that x∈s^2,s′′ and therefore (by definition) s^2∈S⊥(s′′). Similarly, there exist two strata s^1,s′ such that z∈s^1,s′′ and therefore s^1∈S⊥(s′). We get Indeed, as x∈s′′ with s′′∈S⊥(s2), by Lemma (iii) we have that s^2∉S⊥(s2) and, by the same reasoning, s^1∉S⊥(s1). As x∉span(s1), we clearly have that s^2∉S⊥(s1) and analogously s^1∉S⊥(s2). Moreover, by Lemma (vi) we have that s^1≠s^2, and, if both strata s^1,s^2 belong to Sint, we necessarily have s′,s′′∈Sno by Lemma (iv). Arguing as in the proof of (i) and taking () into account, we therefore obtain This concludes the proof.□ Lemma 3.10 (Estimate on Fbond,β, β≥1). Let 0<ε<ε0 and β≥1. Let G=(V,E) be an ε-regular graph satisfying (). Then, Fbond,β(G)≥2(1-β)N+2 with equality if and only if #Sint=N. Proof Lemma (iii) and Lemma (iv) clearly imply #Sint≤N and #Sno≥1. This along with the identity Fbond,β(G)=2(1-β)#Sint+2#Sno from Lemma (ii) yields the statement.□ Proof of the Main Results This section is devoted to the proofs of Theorem Theorem and Theorem . Crystallization For the proof of Theorem Theorem , our strategy is to show that the minimum of Fβ is given by mβ(N), and that it is attained by subsets of Z2 that touch the substrate. In view of (), this will show the result. Recall the definition of Gnat in (). We first state the following upper bound. Lemma 4.1 (Upper bound). Let CN be a minimizer of (). Then, Gnat satisfies Proof Fix N∈N and let h∈{1,...,N} be such that 2h+2(1-β)⌈Nh⌉=mβ(N). We have two cases: - (i)Nh∉N. In this case, N=h⌊Nh⌋+k for some k∈{1,...,h-1}, and we define C ̄N=1,...,⌈Nh⌉-1×{1,...,h}∪⌈Nh⌉×{1,...,k}. - (ii)Nh∈N. Here, one chooses C ̄N={1,...Nh}×{1,...,h}. In view of (), by directly checking that Fβ(G ̄nat)=Fbond,β(G ̄nat)=2h+2(1-β)⌈Nh⌉=mβ(N), where G ̄nat denotes the natural bond graph related to C ̄N, one concludes the proof.□ Given the upper bound of Lemma , the proof of Theorem follows by checking the corresponding lower bound. Proof of Theorem 2.1 Let CN be a minimizer of (). Then, its natural bond graph Gnat is ε-regular by Lemma . We denote by Go=(V,Eo) the graph obtained in Lemma . The graph Go is also ε-regular and satisfies i.e., all properties derived in Sect. hold for the graph Go. The main part of the proof consists in verifying Once () is proven, we conclude as follows. First, () holds due to Lemma (ii) and (). To characterize the equality case, we get from Lemma that G=Go, that all bond lengths are 1 (including the ones between V and L-), and all bond angles belong to {π/2,π,3/2π}. This shows that each connected component (in the sense of graph theory) of G lies in a rotated and shifted version of Z2. If there was more than one connected component, one could obtain a modified configuration with an additional bond. This contradicts minimality. Furthermore, if Sint=∅, one could obtain a modified configuration with strictly less energy than the original one by placing a rotated and shifted version of G in such a way that V⊂Z2 and NL-(x)≠∅ for some x∈V. As a consequence, we get that V⊂Z2∩{x2>0} with length l=#{k:({k}×Z)∩V≠∅} and height h=#{k:(Z×{k})∩V≠∅}, where the properties stated in Remark (i),(ii) follow by repeating the argument in [, Proposition 6.3]. This also shows Remark (iii) and eventually the remaining property #(V∩{x2=1})≥N/h∗(β,N). We now show (). In the following, for simplicity we write S, Sint, and Sno in place of S(Go), Sint(Go), and Sno(Go). The case β≥1 directly follows from Lemma , so we focus on the case β∈(0,1). Recalling () we distinguish two cases: - (i)There exists s∈S such that 4.2span(s)=V; - (ii)For all s∈S we have 4.3span(s)⊊V. Case (i): Select s∈S such that () holds true. By () and Lemma (ii),(iii), choosing some s^∈argmaxs′∈S⊥(s)l(s′) we get l(s)·l(s^)≥∑s′∈S⊥(s)l(s′)=N. Assume without loss of generality that l(s^)≥l(s). Then, l(s^)≥⌈N/l(s)⌉. Using Lemma (i), we obtain This concludes Case (i). Case (ii): Our goal is to find Nβ such that for all N≥Nβ the estimate holds which shows that this case can never occur for a minimizer. Define p:S→{12,12(1-β)} by We prove that there exists x0∈V such that where s(x0),s⊥(x0)∈S are the two strata such that x0∈s(x0)∩s⊥(x0), see Lemma (iii). To this end, define μ(s)=1p(s). By Lemma (ii) we have Now, by Jensen’s inequality together with Lemma (i) and () we obtain On the other hand, letting s(x),s⊥(x)∈S be the two strata such that x∈s(x)∩s⊥(x), exchanging the order of summation, and using Lemma (iii), we obtain Inequality () follows from () and () by selecting x0∈V such that p(s(x0))l(s(x0))+p(s⊥(x0))l(s⊥(x0))≥p(s(x))l(s(x))+p(s⊥(x))l(s⊥(x)) for all x∈V. Next, using () and Lemma (ii) for s(x0) and s⊥(x0), we obtain Now, by () and by multiplying () with 4(1-β) we obtain By solving the quadratic equation t2-(4-2β)t-16(1-β)N≥0 for t≥0, we obtain Let h ̄=⌈((1-β)N)1/2⌉, hc=((1-β)N)1/2, and fN(h)=2h+2(1-β)Nh. We then have |hc-h ̄|≤1 and using that minhfN(h)=fN(hc), by Taylor expansion, together with the fact that h ̄≥hc and thus fN′′(hc)≥fN′′(h) for all h∈[hc,h ̄], we get Therefore, using (), we obtain For N≥Nβ:=β-24(1-β)-1 this shows () and concludes the proof.□ Fluctuation estimates The goal of this section is to prove Theorem . Notice that we assume β∈(0,1) throughout this section. Definition 4.2 Given N∈N we define Hβ,N:{1,...,N}→R and Lβ,N:{1,...,N}→R by and Eβ:N×N→R by Moreover, we set We observe that and Thus, as the next lemma shows, h∗ and l∗ provide good reference values for minimizers of Hβ,N and Lβ,N, respectively. Lemma 4.3 There exists a constant Cβ>0 such that for all minimizers h∈N and l∈N of Hβ,N and Lβ,N, respectively, we have where h∗ and l∗ are given by (). Furthermore, if β∈R\Q, then the minimizers of Hβ,N and Lβ,N are unique. Proof Note that the proofs for minimizers of Hβ,N and Lβ,N are analogous. We therefore only prove the statements for minimizers of Hβ,N. We first show that for β∈R\Q the minimizer of Hβ,N is unique. To this end, assume by contradiction that there exist two minimizers h1,h2∈{1,...,N}. Then, which is equivalent to 2(h1-h2)=2(1-β)(⌈Nh2⌉-⌈Nh1⌉). However, this last equation cannot hold as 2(1-β)∈R\Q. Next, we prove (). Let h∈N be a minimizer of Hβ,N and define h∗=⌈h∗⌉. Then Multiplying by h, we obtain The left-hand side is quadratic in h with zeros at h±=2-β2+h∗±14(4-2β)2+16(2-β)h∗12 and thus h-≤h≤h+. Noting that h∗≤N12 by (), this implies By choosing Cβ:=2-β2+143212(2-β)12 we conclude the proof.□ We now first prove Theorem for β∈Q. Proof of Theorem , upper bound (U1). Let β∈Q, N≥Nβ, with Nβ from Theorem , and let CN be a minimizer of Fβ. By Theorem we have that CN⊂Z2. By Lemma it suffices to prove that there exists a horizontal translation in Z×{0} and Cβ>0 such that where Rmax(β,N)={1,...,lmax}×{1,...,hmax} with Without loss of generality we can assume that the term on the left-hand side of () is minimized for τ=0. By Lemma , Remark , and Definition , we have that CN⊂Rmax(β,N). As #CN=N, it suffices to show that Now, by Lemma we have that |hmax-h∗|≤CβN1/4 and |lmax-l∗|≤CβN1/4 with l∗,h∗ given in (). As l∗,h∗≤CβN1/2 and l∗·h∗=N, this shows (U1). □ Proof of Theorem , lower bound (L1). Let again β∈Q. Let p ̄,q ̄∈N, p ̄0 depending on β, for all τ∈Z×{0} we have This shows that (L1) must be satisfied either for CNk1 or CNk2. It remains to prove (). Testing () with h=pk2, we obtain mβ(Nk)≤2k2p+2(1-β)(k2q+1). We now need to prove mβ(Nk)≥2k2p+2(1-β)(k2q+1). To this end, observe first that for Nk0:=pqk4 we get Indeed, we have h∗=pk2,l∗=qk2∈N (defined in () for Nk0 in place of N) and therefore This shows (). By the convexity of h↦Nk0h we further have For any minimizer h∈{1,...,Nk} it holds that |h-h∗|≤CβNk1/4 by Lemma . Along with ()–(), for k large enough this shows Here, we used that h∗h≥1-δ for k large enough such that CβNk1/4≤12h∗=12(1-β)1/2Nk1/2 and 2Cβ(1-β)-1/2Nk-1/4≤δ, recalling () and Lemma . In view of (), in order to conclude the proof of (), it suffices to show 2(h-h∗)+2(1-β)⌈11-β(h∗-h)-2δ⌉≥0. Recalling that 1-β=p ̄q ̄ and setting q ̄(h-h∗)=p ̄l+r with l∈Z, r∈{0,...,p ̄-1}, we have because -rp ̄-2δ>-p ̄-1p ̄-212p ̄=-1 as δ<12p ̄. This shows () and concludes the proof of (L1). □ We now consider the case that β∈R\Q is algebraic. Denoting by N the number of particles of a configuration and by h its height, and by l its length, it is not restrictive to reduce the problem to considering configurations where (h-1)-rows consist of l particles and the upmost row consists of at most l particles, cf. Remark . Such configurations will be called (h, l)-configurations in the following. If also the upmost row consists of l particles, we call such a configuration an (h, l)-rectangle. In the latter case, we have N=hl. Given an (h, l)-configuration, the corresponding energy is given by Hβ,N(h)=2h+2(1-β)⌈Nh⌉, as defined in (). We recall by Lemma that, due to the fact that β is irrational, the minimizer h of Hβ,N is unique. Moreover, the energy of different (h, l)-rectangles is necessarily different. We fix N0∈N sufficiently large. Our strategy consists in characterizing the optimal (h, l)-rectangles with particle number N with N∈[N0,2N0]. By Lemma we know that the unique optimal height corresponding to N, denoted by h∗(β,N) (see ()), is given by This shows that for particle numbers N∈[N0,2N0] it suffices to consider heights between cβN0 and CβN0 for some 0N denotes the smallest number bigger than N such that there exists an optimal rectangle with N′ particles, and cβ>0 is a constant depending only on β. The above preliminary observation yields minHβ,N′=minHβ,N. Consider an optimal (h, l)-rectangle consisting of N′ particles, and consider the corresponding (h, l)-configuration consisting of N particles by removing N′-N particles from the upmost row. Due to minHβ,N′=minHβ,N, this configuration is also optimal. As N′-N≥cβN01/3-δ, this configuration allows for a fluctuation of order ∼N01/3-δ. Since this holds for all N0≥Nβ,δ, the lower bound (L2) follows for a constant cβ,δ>0 depending on β and δ. Proof of the upper bound (U2): Fix N∈[N0,2N0] and consider an optimal (h, l)-configuration consisting of N particles. By the above preliminary observation and by Lemma we find that the (h, l)-rectangle with particle number N′=hl is optimal and it holds that N′-N≤N01/3+δ. (Indeed, if N′-N>N01/3+δ, Lemma (ii) (applied for N′) would imply that there exists N0 depending on β and δ. □ We now come to the proofs of Lemma and Lemma . Proof of Lemma 4.4 Let 0<δ<16. Along the proof we will progressively increase the lower bound Nβ,δ, depending on δ and β. Fix a (h0,l)-rectangle, and set N=h0l. Given h∈N, we first rewrite the energy Hβ,N defined in () by easily computing where for shorthand we have set We observe that Proof of (i): Consider l∈N with |l-l(h0)|≤N01/6-δ. In view of (), for fixed h, it suffices to show that ⌈t⌉≥⌈11-β(h0-h)⌉. Since β is algebraic, by Roth’s theorem, see [], we find c=c(β,δ)>0 such that for all p∈Z and q∈N. For the choice q=|h0-h| this yields |11-β|h0-h|-p|≥c|h-h0|-1-δ for all p∈Z. Thus, writing x=h0-h, we get from () By using () and the fact that |l-l(h0)|≤N01/6-δ we further get It is not restrictive to assume that h∈[12cβN01/2,2CβN01/2] as otherwise h cannot be a minimizer, see (). We have h0∈[12cβN01/2,2CβN01/2] by assumption and therefore also l≥12cβN01/2, cf. (). This shows If |x|>32(Cβcβ-1)2N01/6-δ, the sum of the last two terms on the right-hand side is positive. If |x|<(cβc4N01/3+δ)1/(2+δ), the sum of the first two terms on the right-hand side is positive. Since 32(Cβcβ-1)2N01/6-δ<(cβc4N01/3+δ)1/(2+δ) for N0≥Nβ,δ for some Nβ,δ depending on δ and β, we conclude This shows ⌈t⌉≥⌈11-β(h-h0)⌉ and concludes the proof of (i). Proof of (ii): Consider l∈N with |l-l(h0)|≥N01/6+δ. Without restriction we treat the case l0 as otherwise the configuration is clearly not optimal. Using also h,h0≥12cβN01/2, cf. below (), we estimate Using x>0, lh0-11-β≤(-N01/6+δ+1)h0-1 (see () and use the choice of l), as well as h0≤2CβN01/2 we get provided that Nβ,δ is chosen sufficiently large. Recalling that N0α(1-3δ)≤x≤N0α and 1/3-δ2-4δ<α<1/6+δ1+3δ, for N0≥Nβ,δ and Nβ,δ sufficiently large, we find by an elementary computation that This yields ⌈t⌉≤⌈11-β(h0-h)⌉-1 and concludes the proof. In the case l>l(h0), we use Lemma (ii) below to find x∈N with N0α(1-3δ)≤x≤N0α and 11-β(-x)-⌊11-β(-x)⌋≤N0-α(1-δ), and define h=h0+x. Then, the statement follows by repeating the estimates above.□ We check the following result which has been used in the previous proof and will be also instrumental in the proof of Lemma below. Lemma 4.6 Let 0<δ<16 and 0<α<1. There exist Nβ,δ∈N (depending on δ, α, and β) and c>0 (depending on δ and β) such that for N0≥Nβ,δ - (i)we find x∈N with N0α(1-3δ)≤x≤N0α such that 4.23cN0-α(1+δ)≤11-βx-11-βx≤N0-α(1-δ); - (ii)we find x∈N with N0α(1-3δ)≤x≤N0α such that cN0-α(1+δ)≤⌈11-βx⌉-11-βx≤N0-α(1-δ). Proof First we show (i). We define the m-discrepancy of the sequence (j11-βmod1)j∈N with respect to the interval [0, s], s>0, by As β is algebraic, by [, Thm. 3.2 and Ex. 3.1, pp. 123-124], for δ>0 there exists Cδ,β>0 depending on δ and β such that With s=N0-α(1-δ) and m=⌊N0α⌋ we get provided that N0≥Nβ,δ for some Nβ,δ depending on δ, α, and β. This shows that there exists x∈N0 such that x≤N0α and 11-βx-⌊11-βx⌋≤s=N0-α(1-δ), i.e., the upper bound in () holds. By Roth’s Theorem [] there exists c>0 depending on δ and β such that |q11-β-p|>cq1+δ for all p∈Z and q∈N. Thus, |11-βx-⌊11-βx⌋|>cx1+δ≥cN0-α(1+δ) which yields the lower bound in (). Furthermore, as this also implies x1+δc≥N0α(1-δ) by the upper bound in (), we obtain x≥N0α(1-3δ) for N0≥Nβ,δ, provided that Nβ,δ is chosen sufficiently large. The proof of (ii) follows along the same lines by using the m-discrepancy of the sequence (j11-βmod1)j∈N with respect to the interval [1-s,1], s>0, namely Using again [] we can control φ^m in the same way as φm above. □ We prove now Lemma . Proof of Lemma 4.5 Let h,l∈N, N0=hl, and suppose that the (h, l)-rectangle is optimal. We only prove (i) as the proof of (ii) is completely analogous and we briefly indicate the necessary adaptations at the end of the proof. Let σl=1 if l-l(h)≥0 and σl=-1 else. We apply Lemma ((i) or (ii), depending on the sign of σl) for α=16-3δ to choose p,q∈N with such that We claim that Indeed, by (), (), (), and the fact that |l-l(h)|≤N01/6+δ (see Lemma (ii)), we obtain where the constant depends on δ and β. Here, we used that h≥cβN01/2. On the other hand, using again h≥cβN01/2, This shows (). Recalling that |l-l(h)|≤N01/6+δ and p≤Cq≤CN01/6-3δ, we can further choose k∈N, k≤CN01/6+δ/q, such that for some universal C>0. We now define l+=l-σlkp, h+=h+σlkq, and N+=l+h+. We claim that provided that Nβ,δ is sufficiently large. Indeed, By () we find Using that cβN012≤h≤CβN01/2, and that kp,kq≤CN01/6+δ we get where the first inequality holds for N0≥Nβ,δ for some Nβ,δ sufficiently large depending on δ and β. This concludes the proof of (). We observe that the desired estimate follows from () by replacing δ with δ/8. It remains to check that also the (h+,l+)-rectangle is optimal. First note that cβN01/2≤h≤CβN01/2 implies cβ2N01/2≤h+≤2CβN01/2 for N0≥Nβ,δ and Nβ,δ sufficiently large as kq≤CN01/6+δ. Thus, in view of Lemma (i), it suffices to check that as then, by using (), we get where the last inequality holds for Nβ,δ sufficiently large depending on δ and β. Let us finally show (). By () and (), we calculate for N0≥Nβ,δ and Nβ,δ sufficiently large depending on δ. Here, we used that σl=1 if l-l(h)≥0 and σl=-1 else. This shows () and concludes the proof. Finally, the proof of (ii) follows along the same lines by choosing p and q such that () is replaced by cN0-(16-3δ)(1+δ)≤σl(-q11-β+p)≤N0-(16-3δ)(1-δ). □ Proof of Lemma Proof of Lemma 3.2 Let Cn be a minimizer of (). For simplicity, we write G=(V,E) instead of Gnat=(V,Enat) for the associated natural bond graph. Step 1. We define Vsub=V∪⋃x∈VNL-(x) and Esub=E∪{{x,y}:x∈V,y∈NL-(x)}. In this step, we show We define where Br(x) denotes the closed ball with radius r>0 centered at x and we write Br in the following whenever x=0. Claim () will follow by showing M=1. Let x0∈R2 be a maximizer in (). After translation of Vsub, it is not restrictive to assume that x0=0. As #{{x,y}∈Esub:x,y∈B12(1-ε)}≥M(M-1)2, by assumption (iii2), letting λ=min{1,2β}, we have Consider the annulus Aε:=B12(1-ε)+2\B12(1-ε)⊂B12+2. By covering B12+2 with discs of radius 14 we get that there exists K∈N and {zi}i=1K⊂R2 such that for all 0<ε≤12 Note that K is independent of ε (provided ε<12). Recalling (), we have By (), the definition of M, and (i2), letting Λ=max{1,β} we have We write V∩B12(1-ε)={xi}i=1M ̄, M ̄≤M, and consider a competitor V^ (with associated natural bond graph G^) given by where τi∈R2 are chosen such that By (), (), (ii2), and the optimality of G we have Now, using (), (), and (), we obtain For ε>0 small enough (ε<λ5ΛK suffices), this inequality can only be true for M=1. This yields () and concludes Step 1. Step 2. In this step we prove that all bond angles satisfy In particular, by choosing ε<110π, this will also imply that #N(x,E)≤4 and #NL-(x)≤1 for all x∈V, where we recall that in () the energy contributions is due to triples (x,y,z)∈CN∪L-3. To see (), we first claim that Indeed, by Step 1, (ii2), and by the fact that B12(1-ε)(y)⊂B2+12(x) for all y∈N(x,E)∪NL-(x) we have i.e., () holds. We now show (). In fact, assuming by contradiction that x has a bond angle θy,x,z that does not satisfy (), we could define V^=(V\{x})∪{x+τ} for some τ∈R2 such that dist(x+τ,V^\{x+τ})≥2. Then, by (i2), (iv3), and () we obtain a contradiction to the minimality of G, namely Summarizing, with choosing ε0:=min{110π,λ8KΛ}, the statement holds. □ Footnotes