betaFeedback

WorksMathematics

How to arrange four points in the plane

Abstract

Arranging points in the plane appears to be an elementary geometric problem, yet it is deeply rooted in statistical physics. As a toy model, we consider the configuration of four points. We completely classify the minimal configurations of four points interaction energy in R2\R^2 under the general Lennard-Jones potentials V(r)=r−p−br−qV(r)=r^{-p}-br^{-q}, where p>q>0p>q>0 and b∈Rb\in\R. We find that there is a threshold curve Dp,q=2−(p−q)/2(1+p−qq+2(1+2−(q+2)/2)) D_{p,q}=2^{-(p-q)/2}\Bigl(1+\frac{p-q}{q+2}\bigl(1+2^{-(q+2)/2}\bigr)\Bigr) to determine the minimal shapes. The main results can be stated concisely as follows: consider min⁡∑1≤i<j≤4(∣xi−xj∣−p−b∣xi−xj∣−q)\min\sum_{1\le i<j\le4}\bigl(|x_i-x_j|^{-p}-b|x_i-x_j|^{-q}\bigr). For b≤0b\le0, the minimizer does not exist. For b>0b>0, the minimizer is unique up to rotation and translation. Furthermore, the minimizer is {a square,if    Dp,q≥1,a rhombus with angle θ∈(π/3,π/2),if    Dp,q<1. \begin{cases} \text{a square},&\hbox{if}\;\; D_{p,q}\ge1, \text{a rhombus with angle }θ\in(π/3,π/2),&\hbox{if}\;\; D_{p,q}<1. \end{cases}

Introduction and statement of main results

At low temperatures, atoms arrange themselves into periodic structures and regular shapes. How interatomic forces the periodic structures and regular shapes is a fundamental problem in statistical and solid mechanics [11, 25, 7]. To model the physical problems from the first principle and simplify the statements, we formulate the problems in a purely mathematical form, and we call that how to arrange points minimizing certain functionals.

Let X=(x1,…,xN)∈(R2)NX=(x_1,\ldots,x_N)\in(\mathbb{R}^{2})^N consist of N≥1N\geq1 distinct points. For a radial two-body potential VV, the NN points interaction energy is

EV(N)(X)=∑1≤i<j≤NV(∣xi−xj∣)=12∑1≤i,j≤Ni≠jV(∣xi−xj∣).(1)E_V^{(N)}(X)=\sum_{1\leq i<j\leq N}V(|x_i-x_j|)=\frac{1}{2}\sum_{\substack{1\leq i,j\leq N\\i\ne j}}V(|x_i-x_j|). \tag*{(1)}

The NN points interaction energy arises in lots of physical and mathematical problems, for which we refer to the recent Lecture notes by Serfaty [27], and the reviews by Radin [25] and Blanc-Lewin [7].

There are freedom on particle numbers NN and the potentials VV. We start from a simple yet useful potential. The Heitmann–Radin potential is defined as follows:

VHR(r)={+∞,0<r<1,−1,r=1,0,r>1.(2)V_{\mathrm{HR}}(r)= \begin{cases} +\infty, & 0<r<1,\\ -1, & r=1,\\ 0, & r>1. \end{cases} \tag*{(2)}

Correspondingly, the Heitmann–Radin energy [15] is NN points interaction energy (1.1) EVHR(N)(X)E_{V_{\mathrm{HR}}}^{(N)}(X) under the potential (1.2).

1991 Mathematics Subject Classification. Primary 49Q10, 82B21; Secondary 52C15, 70C20, 74G65.

Key words and phrases. Lennard–Jones potential, Mie potential, four-point energy, square, rhombus, phase transition.

In 1980, Heitmann–Radin proves the following celebrated result.

Theorem A (Heitmann–Radin [15]). Every minimizer XX of EVHR(N)(X)E_{\mathrm{VHR}}^{(N)}(X) is subset of triangular lattice. That is,

{x1,…,xN}⊂y+Q(Z(1,0)+Z(12,32))\{x_1,\ldots,x_N\}\subset y+Q\left(\mathbb{Z}(1,0)+\mathbb{Z}\left(\frac{1}{2},\frac{\sqrt{3}}{2}\right)\right)

for some y∈R2y\in\mathbb{R}^2 and Q∈SO(2)Q\in\mathrm{SO}(2).

A new proof of Theorem A can be found in De Luca-Friesecke [12]. Theorem A does not include any uniqueness of the minimizers up to translation and rotation. This question was completely solved by De Luca-Friesecke [11] as follows:

Theorem B (De Luca-Friesecke [11]). Let N∈NN\in\mathbb{N}. The minimizers of EVHR(N)(X)E_{\mathrm{VHR}}^{(N)}(X) are unique up to translation and rotation if and only if either

(a) N=3s2+3s+1N=3s^2+3s+1, for some s∈N∪{0}s\in\mathbb{N}\cup\{0\}

(b) N=3s2+3s+1+(s+1)k+sN=3s^2+3s+1+(s+1)k+s, for some s∈N∪{0}s\in\mathbb{N}\cup\{0\} and some k∈{0,1,2,3,4}k\in\{0,1,2,3,4\}.

Theorems A and B together provide a complete classification including existence and uniqueness on the minimizers of the NN points interaction energy under Heitmann–Radin potential.

The Heitmann–Radin potential (1.2) is useful to model solid collision among particles, while physically there is a widely used Lennard-Jones potential as follows:

VLJ(r):=r−2p−2r−p,p>0.(3)V_{\mathrm{LJ}}(r):=r^{-2p}-2r^{-p},\quad p>0. \tag*{(3)}

See, for example, in [12]. The Lennard-Jones has long history and receives extensive studies ([21, 16, 17, 19, 7]). As noted in De Luca-Friesecke [11], the Lennard-Jones potential converges to Heitmann–Radin potential as p→+∞p\to+\infty. This reveals a connection between these two potentials. To fully understand particles interaction under Lennard-Jones potentials, it is natural to ask the following problem, in parallel to Theorems A and B:

Problem A. Classify the minimizers of NN-points interaction energy (1.1) under the Lennard-Jones potential (1.3).

There is a related interesting problem concerning Theorem A. The triangular and hexagonal shapes are prevail and ubiquitous in nature. Understanding the triangular and hexagonal shapes from mathematical perspective from the first principle is important for both in physics and applied and pure mathematics. There are many interesting works; see Montgomery [23] for the two-dimensional theta functions, Béterin-Faulhuber-Steinerberger [2] for shifted Gaussian lattice sum with wide applications, Chen-Oshita [8] for copolymers, Sandier-Serfaty [26] for Ginzburg-Landau superconductivity, Theil [28] and E-Li [14] for triangular crystallization and hexagonal crystallization, respectively.

While, we shall raise the following problem from the opposite direction.

Problem B. For what situations, the minimizers of NN-points interaction energy (1.1) are not the triangular or regular shapes (or their subsets).

For N=4N=4, Theorem A indicates that the rhombus with angle π/3\pi/3 (two equilateral triangles) minimizes the four points interaction energy under Heitmann–Radin potential.

To answer Problems A and B, we focus on N=4N=4 as a toy model and consider the most general form of Lennard-Jones potentials V(r)=r−p−br−qV(r)=r^{-p}-br^{-q} with p>q>0p>q>0 and b∈Rb\in\mathbb{R}. Then the four points interaction energy has the form as follows:

Ep,q,b(X):=∑1≤i<j≤4(∣xi−xj∣−p−b∣xi−xj∣−q).(4)E_{p,q,b}(X):=\sum_{1\leq i<j\leq4}\left(|x_i-x_j|^{-p}-b|x_i-x_j|^{-q}\right). \tag*{(4)}

We provide a complete classification on the minimizers of this four points interaction energy for all p>q>0p>q>0 and b∈Rb\in\mathbb{R}. It is stated as follows:

Theorem 1.1 (Minimal configurations). Let p>q>0p>q>0 and b∈Rb\in\mathbb{R}. Consider the four points interaction energy under the general Lennard-Jones potentials as follows:

∑1≤i<j≤4(∣xi−xj∣−p−b∣xi−xj∣−q).\sum_{1\leq i<j\leq4}\left(\lvert x_i-x_j\rvert^{-p}-b\lvert x_i-x_j\rvert^{-q}\right).

Then, there is a threshold curve defined as follows:

Dp,q:=2−(p−q)/2(1+p−qq+2(1+2−(q+2)/2)).(5)D_{p,q}:=2^{-(p-q)/2}\left(1+\frac{p-q}{q+2}\left(1+2^{-(q+2)/2}\right)\right). \tag*{(5)}

The following hold.

(i) For b≤0b\leq0, the minimizer does not exist.

(ii) For b>0b>0, the minimizer is a rhombus. Up to rotation and translation, it is

Xp,q,b∗=ab(cos⁡(θp,q/2)0−cos⁡(θp,q/2)00sin⁡(θp,q/2)0−sin⁡(θp,q/2)),ab>0.X_{p,q,b}^{*}=a_b\begin{pmatrix} \cos(\theta_{p,q}/2)&0&-\cos(\theta_{p,q}/2)&0\\ 0&\sin(\theta_{p,q}/2)&0&-\sin(\theta_{p,q}/2) \end{pmatrix},\qquad a_b>0.

(iii) For b>0b>0, the angle θp,q\theta_{p,q} of the rhombus satisfies

{θp,q=π/2,Dp,q≥1,π/3<θp,q<π/2,Dp,q<1.\begin{cases} \theta_{p,q}=\pi/2,&D_{p,q}\geq1,\\ \pi/3<\theta_{p,q}<\pi/2,&D_{p,q}<1. \end{cases}

In particular, the minimizers are rhombi whose angles can never be π/3\pi/3.

(iv) For b>0b>0, the following hold.

(a) The map (p,q)↦θp,q(p,q)\mapsto\theta_{p,q} is real analytic on {(p,q):p>q>0, Dp,q<1}\{(p,q):p>q>0,\ D_{p,q}<1\}. Here, ∂pθp,q,∂qθp,q<0\partial_p\theta_{p,q},\partial_q\theta_{p,q}<0.

(b) For each fixed q>0q>0,

lim⁡p→∞(p+2)(θp,q−π/3)=23log⁡1+3−(q+2)/21−3−(q+2)/2.\lim_{p\to\infty}(p+2)(\theta_{p,q}-\pi/3)=\frac{2}{\sqrt{3}}\log\frac{1+3^{-(q+2)/2}}{1-3^{-(q+2)/2}}.

(c) For p=2qp=2q. As q→∞q\to\infty,

θ2q,q=π3+4⋅3−q/233 q+O(3−q/q).\theta_{2q,q}=\frac{\pi}{3}+\frac{4\cdot3^{-q/2}}{3\sqrt{3}\,q}+O(3^{-q}/q).

(d) If p∗>q∗>0p^{*}>q^{*}>0 and Dp∗,q∗=1D_{p^{*},q^{*}}=1, then

lim⁡(p,q)→(p∗,q∗)Dp,q<1π/2−θp,q1−Dp,q=26(p∗−q∗)[p∗+q∗+10+(4−2p∗+q∗)2−(p∗+2)/2].\lim_{\substack{(p,q)\to(p^{*},q^{*})\\D_{p,q}<1}} \frac{\pi/2-\theta_{p,q}}{\sqrt{1-D_{p,q}}} = \frac{2\sqrt{6}}{\sqrt{(p^{*}-q^{*})[p^{*}+q^{*}+10+(4-2p^{*}+q^{*})2^{-(p^{*}+2)/2}]}}.

Remark 1.2. For p=2qp=2q and b=2b=2, the minimizer in (ii) converges to a rhombus with angle π/3\pi/3. More precisely,

X2q,q,2∗⟶(3/20−3/2001/20−1/2)(q→∞).X_{2q,q,2}^{*}\longrightarrow \begin{pmatrix} \sqrt{3}/2&0&-\sqrt{3}/2&0\\ 0&1/2&0&-1/2 \end{pmatrix} \qquad(q\to\infty).

The limit gives the minimizer of EVHR(4)(X)E_{\mathrm{VHR}}^{(4)}(X).

The square shape is not the first time being proved as the minimizer; see Béterin-De Luca-Petrache [1] on square crystallization. Recently, Del Nin-De Luca proved square crystallization for supremum-norm sticky disks [13]. The square lattice shape appear also in triblock copolymer (Luo-Ren-Wei [18]) and Bose-Einstein condensates (Mueller-Ho [22] and Luo-Wei [20]).

The minimal shapes: Square, $D_{p,q} \ge 1$; Non-square rhombus, $D_{p,q} < 1$.

Figure 1. The minimal shapes.

Considering thermodynamic of interaction energy (1.1) per particle, and assuming the particles arrange themselves into a lattice, we obtain the ff−potential lattice energy

Ef(Λ):=∑P∈Λ∖{0}f(∣P∣2),∣⋅∣ is the Euclidean norm on R2.(6)E_f(\Lambda) := \sum_{P\in\Lambda\setminus\{0\}} f(|P|^2),\quad|\cdot| \text{ is the Euclidean norm on } \mathbb{R}^2. \tag*{(6)}

Here Λ\Lambda denotes a lattice. This is another related interesting topic on lattices. We refer to Bétermin-Petrache [3] for a detailed review on lattices and Blanc-Lewin [7] for the broad review on crystallization including lattice minimizations and special functions as a section. Cohn-Kumar-Miller-Radchenko-Viazovska [9] prove the universality of E8E_8 and Leech lattices in dimensions 8 and 24, respectively. Interestingly, Petrache-Serfaty [24] then prove the crystallization results in these two dimensions building on this remarkable theorem ([9]). For the ff−potential lattice energy (1.6) under the classical Lennard-Jones potential in dimension two, we refer to Bétermin and his collaborators [5, 4, 6] for pioneering work and Luo-Wei [19] for a complete resolution; while this problem is open in dimensions higher than two.

The paper is organized as follows: Section 2 reduces the minimizers to rhombi. It rules out nonconvex minimizers by comparison with squares. For a convex minimizer, Connelly’s rank-one equilibrium stress [10] shows that the diagonals are perpendicular and bisect each other. We consider the minimization over rhombi in Section 3. We consider the classical Lennard-Jones case in Section 4.

Reduction to rhombi

For b≤0b \le0, dilation gives a zero infimum that is not attained. For b>0b > 0, (9) and Lemmas 2.1–2.3 give existence and reduce the problem to rhombi.

Write

ep,q,b:=inf⁡XEp,q,b(X),rij:=∣xi−xj∣.e_{p,q,b}:=\inf_X E_{p,q,b}(X), \qquad r_{ij}:=|x_i-x_j|.

For b≤0b \le0,

0<Ep,q,b(X),lim⁡t→∞Ep,q,b(tX)=0,(min⁡1≤i<j≤4rij)−p≤Ep,q,b(X).0 < E_{p,q,b}(X), \qquad\lim_{t\to\infty} E_{p,q,b}(tX)=0, \qquad\left(\min_{1\le i<j\le4}r_{ij}\right)^{-p}\le E_{p,q,b}(X).

The energy tends to zero exactly when min⁡1≤i<j≤4rij→∞\min_{1\le i<j\le4}r_{ij}\to\infty. Differentiation under dilation gives

ddtEp,q,b(tX)∣t=1=−p∑1≤i<j≤4∣xi−xj∣−p+bq∑1≤i<j≤4∣xi−xj∣−q<0,\left.\frac{d}{dt}E_{p,q,b}(tX)\right|_{t=1}=-p\sum_{1\le i<j\le4}|x_i-x_j|^{-p}+bq\sum_{1\le i<j\le4}|x_i-x_j|^{-q}<0,

so there are no critical configurations. This proves part (i) of Theorem 1.1.

For b>0b > 0, the potential r−p−br−qr^{-p}-br^{-q} has a unique minimum at ℓb\ell_b, with well depth εb\varepsilon_b, where

ℓb=(pbq)1/(p−q),εb=−(ℓb−p−bℓb−q)=p−qq(bqp)p/(p−q).(7)\ell_b=\left(\frac{p}{bq}\right)^{1/(p-q)}, \qquad\varepsilon_b=-\left(\ell_b^{-p}-b\ell_b^{-q}\right)=\frac{p-q}{q}\left(\frac{bq}{p}\right)^{p/(p-q)}. \tag*{(7)}

Define the normalized potential and energy by

W(r)=qr−p−pr−qp−q,Ep,q(X)=∑1≤i<j≤4W(rij),(8)W(r)=\frac{qr^{-p}-pr^{-q}}{p-q}, \qquad E_{p,q}(X)=\sum_{1\le i<j\le4}W(r_{ij}), \tag*{(8)}

so that

Ep,q,b(ℓbX)=εbEp,q(X).(9)E_{p,q,b}(\ell_bX)=\varepsilon_bE_{p,q}(X). \tag*{(9)}

Differentiating (8) gives

W′(r)=pqp−qr−p−1(rp−q−1).(10)W'(r)=\frac{pq}{p-q}r^{-p-1}(r^{p-q}-1). \tag*{(10)}

Then

W(r)≥W(1)=−1.(11)W(r)\ge W(1)=-1. \tag*{(11)}

W(r)<0W(r)<0 exactly when r>(q/p)1/(p−q)r>(q/p)^{1/(p-q)}.

Lemma 2.1 (Existence and compactness). Let p>q>0p>q>0 and mp,q=inf⁡XEp,q(X)m_{p,q}=\inf_X E_{p,q}(X). A minimizer exists, and

mp,q<−5+W(3)<−5.(12)m_{p,q}<-5+W(\sqrt{3})<-5. \tag*{(12)}

Every distance in a minimizer lies in ((q/p)1/(p−q),3)\left((q/p)^{1/(p-q)},\sqrt{3}\right).

Proof. A rhombus of side length aa and acute angle π/3\pi/3 has diagonals aa and 3a\sqrt{3}a, and hence energy

5W(a)+W(3a).5W(a)+W(\sqrt{3}a).

By (10),

dda[5W(a)+W(3a)]∣a=1=3W′(3)>0.\left.\frac{d}{da}\left[5W(a)+W(\sqrt{3}a)\right]\right|_{a=1}=\sqrt{3}W'(\sqrt{3})>0.

For a<1a<1 sufficiently close to 11, the energy is less than −5+W(3)-5+W(\sqrt{3}), its value at a=1a=1. Since (8) gives W(3)<0W(\sqrt{3})<0, this proves (12).

Fix 1≤i<j≤41 \le i < j \le4. Each of the other five pair terms is at least −1-1 by (2.5), so

Ep,q(X)=W(rij)+∑1≤k<ℓ≤4(k,ℓ)≠(i,j)W(rkℓ)≥W(rij)−5.\mathcal{E}_{p,q}(X)=W(r_{ij})+\sum_{\substack{1\le k<\ell\le4\\(k,\ell)\ne(i,j)}}W(r_{k\ell})\ge W(r_{ij})-5.

If Ep,q(X)<−5+W(3)\mathcal{E}_{p,q}(X)<-5+W(\sqrt{3}), the preceding estimate gives

W(rij)<W(3)<0(1≤i<j≤4).W(r_{ij})<W(\sqrt{3})<0 \qquad(1\le i<j\le4).

By (2.2), W(rij)<0W(r_{ij})<0 implies rij>(q/p)1/(p−q)r_{ij}>(q/p)^{1/(p-q)}. Since (2.4) makes WW strictly increasing on [1,∞)[1,\infty), W(rij)<W(3)W(r_{ij})<W(\sqrt{3}) also implies rij<3r_{ij}<\sqrt{3}.

By (2.5), mp,q≥−6m_{p,q}\ge-6. Translation invariance of Ep,q\mathcal{E}_{p,q} allows us to impose x1(n)=0x_1^{(n)}=0 on a minimizing sequence X(n)X^{(n)}. By (2.6) and the preceding distance bounds, ∣xi(n)∣<3|x_i^{(n)}|<\sqrt{3} for 2≤i≤42\le i\le4 and all sufficiently large nn. A subsequence therefore converges to some X∗=(x1∗,…,x4∗)X^*=(x_1^*,\ldots,x_4^*). Passing to the limit in the lower bound gives

∣xi∗−xj∗∣≥(q/p)1/(p−q)>0(1≤i<j≤4),(13)|x_i^*-x_j^*|\ge(q/p)^{1/(p-q)}>0 \qquad(1\le i<j\le4), \tag*{(13)}

so continuity gives Ep,q(X∗)=mp,q\mathcal{E}_{p,q}(X^*)=m_{p,q}. By (2.6), every minimizer lies in the strict sublevel considered above and hence satisfies both strict distance bounds. □

Lemma 2.2 (Comparison with a square). Let p>q>0p>q>0. Suppose one point of XX lies in the convex hull of the other three. Then there is a square YY such that Ep,q(Y)<Ep,q(X)\mathcal{E}_{p,q}(Y)<\mathcal{E}_{p,q}(X).

Proof. Set

F(t)=1+W(et),Fλ(t)=∫0tue−λu du(λ>0).F(t)=1+W(e^t),\qquad F_\lambda(t)=\int_0^t u e^{-\lambda u}\,du\quad(\lambda>0).

F(0)=0F(0)=0, and (2.4) gives

F′(t)=pqp−q(e−qt−e−pt)=pqp−q∫qpte−λt dλ,F'(t)=\frac{pq}{p-q}\left(e^{-qt}-e^{-pt}\right)=\frac{pq}{p-q}\int_q^p t e^{-\lambda t}\,d\lambda,

and integration gives

F(t)=pqp−q∫qpFλ(t) dλ.(14)F(t)=\frac{pq}{p-q}\int_q^p F_\lambda(t)\,d\lambda. \tag*{(14)}

Thus F≥0F\ge0, FF increases on [0,∞)[0,\infty), and F′′>0F''>0 on (−∞,0](-\infty,0]. For t>0t>0, Fλ(−t)=∫0tueλu du>Fλ(t)F_\lambda(-t)=\int_0^t u e^{\lambda u}\,du>F_\lambda(t).

Step 1. A lower bound for 6+Ep,q(X)6+\mathcal{E}_{p,q}(X). Translate the point in the convex hull to the origin and write ui=∣xi∣u_i=|x_i| and ti=max⁡{0,−log⁡ui}t_i=\max\{0,-\log u_i\} for the remaining three points. If one of the three angles around the origin exceeded π\pi, the three points would lie in an open half-plane bounded by a line through the origin, contradicting the choice of that point. Ordering the gaps as ϕ1≥ϕ2≥ϕ3\phi_1\ge\phi_2\ge\phi_3, we therefore have

ϕ1+ϕ2+ϕ3=2π,2π3≤ϕ1≤π,2ϕ2≥2π−ϕ1≥π.\phi_1+\phi_2+\phi_3=2\pi,\qquad\frac{2\pi}{3}\le\phi_1\le\pi,\qquad2\phi_2\ge2\pi-\phi_1\ge\pi.

In the collinear case the gaps are π,π,0\pi,\pi,0. Let (ik,jk)(i_k,j_k) denote the pair that subtends the angle ϕk\phi_k, for k=1,2k=1,2, and set sk=(tik+tjk)/2s_k=(t_{i_k}+t_{j_k})/2 and s=max⁡{s1,s2}s=\max\{s_1,s_2\}. Because uikujk≥e−2sk≥e−2su_{i_k}u_{j_k}\ge e^{-2s_k}\ge e^{-2s}, the law of cosines yields

rikjk2=(uik−ujk)2+2uikujk(1−cos⁡ϕk)≥2e−2s(1−cos⁡ϕk),k=1,2,r_{i_kj_k}^2=(u_{i_k}-u_{j_k})^2+2u_{i_k}u_{j_k}(1-\cos\phi_k)\ge2e^{-2s}(1-\cos\phi_k),\qquad k=1,2,

and hence ri1j1≥3e−sr_{i_1j_1}\ge\sqrt{3}e^{-s} and ri2j2≥2e−sr_{i_2j_2}\ge\sqrt{2}e^{-s}. For the pair with sk=ss_k=s, convexity gives

F(log⁡uik)+F(log⁡ujk)≥F(−tik)+F(−tjk)≥2F(−s).F(\log u_{i_k})+F(\log u_{j_k})\ge F(-t_{i_k})+F(-t_{j_k})\ge2F(-s).

Each of the two outer pairs contributes F(log⁡rikjk)≥F((log⁡rikjk)+)F(\log r_{i_kj_k}) \ge F((\log r_{i_kj_k})_+), where t+=max⁡{t,0}t_+ = \max\{t, 0\}. The four pairs are distinct, and F≥0F \ge0 bounds the remaining pair terms. Since FF increases on [0,∞)[0,\infty),

6+Ep,q(X)≥2F(−s)+F((log⁡2−s)+)+F((log⁡3−s)+).(15)6+\mathcal{E}_{p,q}(X) \ge2F(-s)+F((\log\sqrt{2}-s)_+)+F((\log\sqrt{3}-s)_+). \tag*{(15)}

Step 2. The comparison with a square. Let YY be a unit square. Its four sides contribute 4F(0)=04F(0)=0, while the two diagonals give 6+Ep,q(Y)=2F(log⁡2)6+\mathcal{E}_{p,q}(Y)=2F(\log\sqrt{2}). If s=0s=0, the right-hand side of (2.9) reads F(log⁡2)+F(log⁡3)>2F(log⁡2)F(\log\sqrt{2})+F(\log\sqrt{3})>2F(\log\sqrt{2}). If s≥log⁡2s\ge\log\sqrt{2}, it is at least 2F(−s)>2F(s)≥2F(log⁡2)2F(-s)>2F(s)\ge2F(\log\sqrt{2}). In either case, (2.9) gives Ep,q(X)>Ep,q(Y)\mathcal{E}_{p,q}(X)>\mathcal{E}_{p,q}(Y).

For 0<s<log⁡20<s<\log\sqrt{2}, let YY be the square of side e−s/2e^{-s/2}. Then

6+Ep,q(Y)=4F(−s/2)+2F(log⁡2−s/2).6+\mathcal{E}_{p,q}(Y)=4F(-s/2)+2F(\log\sqrt{2}-s/2).

For every λ>0\lambda>0, 4Fλ(−s/2)=∫0sueλu/2 du<Fλ(−s)4F_\lambda(-s/2)=\int_0^s u e^{\lambda u/2}\,du<F_\lambda(-s), whence

2Fλ(−s)−4Fλ(−s/2)>Fλ(−s)>s22.(16)2F_\lambda(-s)-4F_\lambda(-s/2)>F_\lambda(-s)>\frac{s^2}{2}. \tag*{(16)}

We next estimate the difference between the diagonal terms. First suppose 0<s≤log⁡3/20<s\le\log\sqrt{3}/2. Since Fλ′′(t)=(1−λt)e−λt≥−e−2F_\lambda''(t)=(1-\lambda t)e^{-\lambda t}\ge-e^{-2} for t≥0t\ge0,

2Fλ(log⁡2−s/2)−Fλ(log⁡2−s)−Fλ(log⁡3−s)≤∫log⁡2−slog⁡2−s/2[Fλ′(t)−Fλ′(t+s/2)] dt≤e−2s24<s22.\begin{aligned} 2F_\lambda(\log\sqrt{2}-s/2)-F_\lambda(\log\sqrt{2}-s)-F_\lambda(\log\sqrt{3}-s) &\le\int_{\log\sqrt{2}-s}^{\log\sqrt{2}-s/2} \left[F_\lambda'(t)-F_\lambda'(t+s/2)\right]\,dt \\ &\le\frac{e^{-2}s^2}{4}<\frac{s^2}{2}. \end{aligned}

For log⁡3/2<s<log⁡2\log\sqrt{3}/2<s<\log\sqrt{2}, we have log⁡3−s>log⁡2−s/2\log\sqrt{3}-s>\log\sqrt{2}-s/2 and s>47log⁡2s>\frac{4}{7}\log\sqrt{2}. Using Fλ′(t)≤tF_\lambda'(t)\le t for t≥0t\ge0,

2Fλ(log⁡2−s/2)−Fλ(log⁡2−s)−Fλ(log⁡3−s)≤Fλ(log⁡2−s/2)−Fλ(log⁡2−s)≤∫log⁡2−slog⁡2−s/2t dt=slog⁡2−34s22<s22.\begin{aligned} 2F_\lambda(\log\sqrt{2}-s/2)-F_\lambda(\log\sqrt{2}-s)-F_\lambda(\log\sqrt{3}-s) &\le F_\lambda(\log\sqrt{2}-s/2)-F_\lambda(\log\sqrt{2}-s) \\ &\le\int_{\log\sqrt{2}-s}^{\log\sqrt{2}-s/2}t\,dt =\frac{s\log\sqrt{2}-\frac{3}{4}s^2}{2}<\frac{s^2}{2}. \end{aligned}

For 0<s<log⁡20<s<\log\sqrt{2}, the difference between the diagonal terms is less than s2/2s^2/2, while the side terms increase by more than s2/2s^2/2 by (2.10). Integrating in λ\lambda and using (2.8) and (2.9) gives

6+Ep,q(X)≥2F(−s)+F(log⁡2−s)+F(log⁡3−s)>4F(−s/2)+2F(log⁡2−s/2)=6+Ep,q(Y),\begin{aligned} 6+\mathcal{E}_{p,q}(X)&\ge2F(-s)+F(\log\sqrt{2}-s)+F(\log\sqrt{3}-s)\\ &>4F(-s/2)+2F(\log\sqrt{2}-s/2)=6+\mathcal{E}_{p,q}(Y), \end{aligned}

which proves the comparison with the square of side e−s/2e^{-s/2}.

Lemma 2.3 (Rhombic symmetry). Let p>q>0p>q>0. Every strictly convex critical configuration of Ep,q\mathcal{E}_{p,q} is a rhombus. In particular, every global minimizer is a rhombus.

Proof. Let XX be a strictly convex critical configuration, and label its vertices in cyclic order.

Step 1. The signs of the forces. Set wij=W′(rij)/rijw_{ij}=W'(r_{ij})/r_{ij} and Ωij=−wij\Omega_{ij}=-w_{ij} for i≠ji\ne j, while the diagonal entries are defined by

Ωii=∑1≤j≤4j≠iwij,1≤i≤4.\Omega_{ii}=\sum_{\substack{1\le j\le4\\j\ne i}}w_{ij},\qquad1\le i\le4.

Write xj=(xj,1,xj,2)x_j=(x_{j,1},x_{j,2}). The definition of Ωii\Omega_{ii} and the criticality of XX for (8) give, for 1≤i≤41\leq i\leq4,

∑j=14Ωij=Ωii−∑1≤j≤4j≠iwij=0,∑j=14Ωijxj,k=∑1≤j≤4j≠iwij(xi,k−xj,k)=0,k=1,2.\sum_{j=1}^{4}\Omega_{ij}=\Omega_{ii}-\sum_{\substack{1\leq j\leq4\\j\neq i}}w_{ij}=0,\qquad \sum_{j=1}^{4}\Omega_{ij}x_{j,k}=\sum_{\substack{1\leq j\leq4\\j\neq i}}w_{ij}(x_{i,k}-x_{j,k})=0,\quad k=1,2.

Since the four points are not collinear, the vectors (1,1,1,1)(1,1,1,1), (x1,1,…,x4,1)(x_{1,1},\ldots,x_{4,1}) and (x1,2,…,x4,2)(x_{1,2},\ldots,x_{4,2}) are linearly independent. Every row of Ω\Omega lies in their one-dimensional orthogonal complement.

If all off-diagonal entries of Ω\Omega vanish, (10) gives rij=1r_{ij}=1 for every i<ji<j. After translating x4x_4 to the origin, the Gram determinant of x1,x2,x3x_1,x_2,x_3 would satisfy

det⁡(11/21/21/211/21/21/21)=12,\det\begin{pmatrix} 1 & 1/2 & 1/2\\ 1/2 & 1 & 1/2\\ 1/2 & 1/2 & 1 \end{pmatrix}=\frac{1}{2},

contradicting their linear dependence in R2\mathbb{R}^2. Hence some off-diagonal entry is nonzero.

Let the diagonals meet at

o=λx1+(1−λ)x3=μx2+(1−μ)x4,0<λ,μ<1,o=\lambda x_1+(1-\lambda)x_3=\mu x_2+(1-\mu)x_4,\qquad0<\lambda,\mu<1,

and set (z1,z2,z3,z4)=(λ,−μ,1−λ,μ−1)(z_1,z_2,z_3,z_4)=(\lambda,-\mu,1-\lambda,\mu-1). Then

∑j=14zj=0,∑j=14zjxj,k=0(k=1,2).\sum_{j=1}^{4}z_j=0,\qquad\sum_{j=1}^{4}z_jx_{j,k}=0\quad(k=1,2).

The orthogonal complement is spanned by zz, so Ωij=cizj\Omega_{ij}=c_i z_j for some ci∈Rc_i\in\mathbb{R}. By symmetry, cizj=cjzic_i z_j=c_j z_i. Since every ziz_i is nonzero, the ratios ci/zic_i/z_i have a common value κ≠0\kappa\neq0. Thus

Ωij=κzizj,1≤i,j≤4,z1,z3>0,z2,z4<0.(17)\Omega_{ij}=\kappa z_i z_j,\qquad1\leq i,j\leq4,\qquad z_1,z_3>0,\quad z_2,z_4<0. \tag*{(17)}

If κ>0\kappa>0, then (17) and (10) give r13,r24<1r_{13},r_{24}<1 together with rij>1r_{ij}>1 for adjacent pairs. Choose on each diagonal the endpoint closest to oo, say xix_i and xjx_j. These vertices are adjacent, and

1<rij≤∣xi−o∣+∣xj−o∣≤r13+r242<1,1<r_{ij}\leq|x_i-o|+|x_j-o|\leq\frac{r_{13}+r_{24}}{2}<1,

a contradiction. Thus κ<0\kappa<0, and (17) and (10) give

r12,r23,r34,r14<1<r13,r24.(18)r_{12},r_{23},r_{34},r_{14}<1<r_{13},r_{24}. \tag*{(18)}

Perpendicular diagonals, one of them bisected

For 0<r<10<r<1 set

f(r)=−W′(r)r=pqp−qr−p−2(1−rp−q),0<r<1.(19)f(r)=-\frac{W'(r)}{r}=\frac{pq}{p-q}r^{-p-2}(1-r^{p-q}),\qquad0<r<1. \tag*{(19)}

Differentiating (19) shows that ff and r2fr^2f are strictly decreasing on this interval.

Place the intersection of the diagonals at the origin and write

x1=ae,x3=−de,x2=αe′,x4=−βe′,∣e∣=∣e′∣=1,a,d,α,β>0,x_1=ae,\qquad x_3=-de,\qquad x_2=\alpha e',\qquad x_4=-\beta e',\qquad|e|=|e'|=1,\qquad a,d,\alpha,\beta>0,

so that

(z1,z2,z3,z4)=(da+d,−βα+β,aa+d,−αα+β).(z_1,z_2,z_3,z_4)=\left(\frac{d}{a+d},-\frac{\beta}{\alpha+\beta},\frac{a}{a+d},-\frac{\alpha}{\alpha+\beta}\right).

Set K=−κ/((a+d)(α+β))>0K=-\kappa/((a+d)(\alpha+\beta))>0. The side lengths lie in (0,1)(0,1) by (18), so (17) and (19) give

f(r12)=Kdβ,f(r23)=Kaβ,f(r34)=Kaα,f(r14)=Kdα.(20)f(r_{12})=Kd\beta,\qquad f(r_{23})=Ka\beta,\qquad f(r_{34})=Ka\alpha,\qquad f(r_{14})=Kd\alpha. \tag*{(20)}

Assume the intersection bisects neither diagonal, and relabel so that a>da>d and α>β\alpha>\beta. Equation (20), together with the monotonicity of ff and of r2fr^2f, gives

r12>r23,r14>r34,dr122<ar232,dr142<ar342,r_{12}>r_{23},\qquad r_{14}>r_{34},\qquad dr_{12}^{2}<ar_{23}^{2},\qquad dr_{14}^{2}<ar_{34}^{2},
r12>r14,r23>r34,βr122<αr142,βr232<αr342.r_{12}>r_{14},\qquad r_{23}>r_{34},\qquad\beta r_{12}^{2}<\alpha r_{14}^{2},\qquad\beta r_{23}^{2}<\alpha r_{34}^{2}.

For example, the law of cosines gives

dr122−ar232=(a−d)(ad−α2)−4adα e⋅e′,dr_{12}^{2}-ar_{23}^{2}=(a-d)(ad-\alpha^{2})-4ad\alpha\,e\cdot e',

and the mixed terms cancel in the sums

0>β(dr122−ar232)+α(dr142−ar342)=(a−d)(α+β)(ad−αβ),0>\beta\left(dr_{12}^{2}-ar_{23}^{2}\right)+\alpha\left(dr_{14}^{2}-ar_{34}^{2}\right)=(a-d)(\alpha+\beta)(ad-\alpha\beta),
0>d(βr122−αr142)+a(βr232−αr342)=(α−β)(a+d)(αβ−ad).0>d\left(\beta r_{12}^{2}-\alpha r_{14}^{2}\right)+a\left(\beta r_{23}^{2}-\alpha r_{34}^{2}\right)=(\alpha-\beta)(a+d)(\alpha\beta-ad).

Since a>da>d and α>β\alpha>\beta, the first identity gives ad<αβad<\alpha\beta and the second gives αβ<ad\alpha\beta<ad, a contradiction. The intersection bisects at least one diagonal. Relabel so that a=da=d. Equation (20) and the strict decrease of ff give r12=r23r_{12}=r_{23}. Since r232−r122=4aα e⋅e′r_{23}^{2}-r_{12}^{2}=4a\alpha\,e\cdot e', we obtain e⋅e′=0e\cdot e'=0.

Step 3. The second diagonal is bisected as well. Suppose α≠β\alpha\ne\beta. Interchanging x2x_2 and x4x_4 if necessary, assume α>β\alpha>\beta. Write c=a2c=a^2, u=c+α2=r122u=c+\alpha^2=r_{12}^{2} and v=c+β2=r142v=c+\beta^2=r_{14}^{2}. By (18) we have 0<v<u<10<v<u<1, and α+β>1\alpha+\beta>1. Then

βu(1−v)−αv(1−u)=(α−β)(uv+αβ−c),\beta u(1-v)-\alpha v(1-u)=(\alpha-\beta)(uv+\alpha\beta-c),

with

uv+αβ−c=(c−αβ)2+c[(α+β)2−1]+αβ>0,uv+\alpha\beta-c=(c-\alpha\beta)^2+c\left[(\alpha+\beta)^2-1\right]+\alpha\beta>0,

so that α(1−u)/u<β(1−v)/v\alpha(1-u)/u<\beta(1-v)/v. For s>0s>0, both t−s/2t^{-s/2} and

−log⁡t1−t=∫01dτ1−τ+τt\frac{-\log t}{1-t}=\int_{0}^{1}\frac{d\tau}{1-\tau+\tau t}

strictly decrease on (0,1)(0,1). Multiplying the preceding inequality by these positive factors and using u>vu>v gives

α(−log⁡u)u−(s+2)/2<β(−log⁡v)v−(s+2)/2.\alpha(-\log u)u^{-(s+2)/2}<\beta(-\log v)v^{-(s+2)/2}.

For 0<r<10<r<1, (19) also implies

f(r)=pqp−q∫qp(−log⁡r)r−s−2 ds.f(r)=\frac{pq}{p-q}\int_{q}^{p}(-\log r)r^{-s-2}\,ds.

Substituting u=r122u=r_{12}^{2} and v=r142v=r_{14}^{2} into the previous inequality and integrating over s∈(q,p)s\in(q,p) yields αf(r12)<βf(r14)\alpha f(r_{12})<\beta f(r_{14}), which contradicts αf(r12)=Kdαβ=βf(r14)\alpha f(r_{12})=K d\alpha\beta=\beta f(r_{14}) from (20). Hence α=β\alpha=\beta.

Together with Step 2, this gives a=da=d and e⋅e′=0e\cdot e'=0, and therefore

r122=r232=r342=r142=a2+α2,r_{12}^{2}=r_{23}^{2}=r_{34}^{2}=r_{14}^{2}=a^{2}+\alpha^{2},

so XX is a rhombus.

Every global minimizer is strictly convex by Lemma 2.2 and is critical because Ep,q\mathcal{E}_{p,q} is smooth on the open set of distinct configurations. The preceding argument therefore applies. □\square

The square and rhombic phases

By Lemma 2.3, every minimizer is a rhombus for b>0b > 0. We prove uniqueness, locate the square region Dp,q≥1D_{p,q} \ge1, and study the angle as a function of (p,q)(p,q), including the four-point Heitmann–Radin limit.

A rhombus of side length a>0a > 0 and angle 0<θ≤π/20 < \theta\le\pi/2 has diagonals 2asin⁡(θ/2)2a\sin(\theta/2) and 2acos⁡(θ/2)2a\cos(\theta/2). For s>0s > 0, set

Ss(θ)=4+(2sin⁡(θ/2))−s+(2cos⁡(θ/2))−s.(21)S_s(\theta) = 4 + (2\sin(\theta/2))^{-s} + (2\cos(\theta/2))^{-s}. \tag*{(21)}

Proposition 3.1 (Minimization among rhombi). For p>q>0p > q > 0 and b>0b > 0, the minimizing rhombus is unique up to congruence and relabelling, and its side length is uniquely determined. It is a square if Dp,q≥1D_{p,q} \ge1, where Dp,qD_{p,q} is defined in (5). If Dp,q<1D_{p,q} < 1, its acute angle is θp,q=2arctan⁡(e−Lp,q)∈(π/3,π/2)\theta_{p,q} = 2\arctan(e^{-L_{p,q}}) \in(\pi/3,\pi/2), where Lp,qL_{p,q} is the unique solution in (0,12log⁡3)(0,\frac{1}{2}\log3) of

1−e−(p+2)L1−e−(q+2)L=1+(41+e2L)(p+2)/21+(41+e2L)(q+2)/2.(22)\frac{1-e^{-(p+2)L}}{1-e^{-(q+2)L}} = \frac{1+\left(\frac{4}{1+e^{2L}}\right)^{(p+2)/2}}{1+\left(\frac{4}{1+e^{2L}}\right)^{(q+2)/2}}. \tag*{(22)}

To prove Proposition 3.1, fix θ\theta and write Ss=Ss(θ)S_s = S_s(\theta). By (4) and (21), the rhombus has energy Spa−p−bSqa−qS_p a^{-p} - bS_q a^{-q}. Differentiating with respect to aa gives

adda(Spa−p−bSqa−q)=a−p(bqSqap−q−pSp).a\frac{d}{da}\left(S_p a^{-p}-bS_q a^{-q}\right)=a^{-p}\left(bqS_q a^{p-q}-pS_p\right).

The factor in parentheses increases strictly from −pSp-pS_p to +∞+\infty. Hence

min⁡a>0{Spa−p−bSqa−q}=−εbSqp/(p−q)Spq/(p−q),ap−q=pSpbqSq.(23)\min_{a>0}\{S_p a^{-p}-bS_q a^{-q}\} =-\varepsilon_b\frac{S_q^{p/(p-q)}}{S_p^{q/(p-q)}}, \qquad a^{p-q}=\frac{pS_p}{bqS_q}. \tag*{(23)}

Thus the minimizing angle maximizes Sqp/SpqS_q^p/S_p^q. By (9) and Lemmas 2.1 and 2.3, a rhombus attains the global minimum. Scaling this minimizer by ℓb−1\ell_b^{-1} and applying (18) gives

1<2asin⁡(θ/2)ℓb<2sin⁡(θ/2)⟹θ>π/3.1<\frac{2a\sin(\theta/2)}{\ell_b}<2\sin(\theta/2) \quad\Longrightarrow\quad \theta>\pi/3.

Set L=log⁡cot⁡(θ/2)=arctanh⁡(cos⁡θ)L=\log\cot(\theta/2)=\operatorname{arctanh}(\cos\theta). Then

Ss(θ)=4+2(cosh⁡L2)s/2cosh⁡(sL/2),dSsdL=scosh⁡L(cosh⁡L2)s/2sinh⁡((s+2)L/2).S_s(\theta)=4+2\left(\frac{\cosh L}{2}\right)^{s/2}\cosh(sL/2), \qquad \frac{dS_s}{dL}=\frac{s}{\cosh L}\left(\frac{\cosh L}{2}\right)^{s/2}\sinh((s+2)L/2).

Define R(L)=R1(L)+R2(L)R(L)=R_1(L)+R_2(L) for L>0L>0, where

R1(L)=(cosh⁡L2)(p−q)/2sinh⁡((p+2)L/2)sinh⁡((q+2)L/2),R2(L)=(cosh⁡L2)(p+2)/2sinh⁡((p−q)L/2)sinh⁡((q+2)L/2).(24)R_1(L)=\left(\frac{\cosh L}{2}\right)^{(p-q)/2} \frac{\sinh((p+2)L/2)}{\sinh((q+2)L/2)}, \qquad R_2(L)=\left(\frac{\cosh L}{2}\right)^{(p+2)/2} \frac{\sinh((p-q)L/2)}{\sinh((q+2)L/2)}. \tag*{(24)}

The addition formulas for sinh give

sinh⁡((q+2)L/2)cosh⁡(pL/2)−sinh⁡((p+2)L/2)cosh⁡(qL/2)=−cosh⁡Lsinh⁡((p−q)L/2),\sinh((q+2)L/2)\cosh(pL/2)-\sinh((p+2)L/2)\cosh(qL/2) =-\cosh L\sinh((p-q)L/2),

and therefore

ddLlog⁡SqpSpq=4pqcosh⁡L SpSq(cosh⁡L2)q/2sinh⁡((q+2)L2)[1−R(L)].(25)\frac{d}{dL}\log\frac{S_q^p}{S_p^q} = \frac{4pq}{\cosh L\,S_pS_q} \left(\frac{\cosh L}{2}\right)^{q/2} \sinh\left(\frac{(q+2)L}{2}\right)[1-R(L)]. \tag*{(25)}

Lemma 3.2 (Monotonicity of R(L)R(L)). The function RR is strictly increasing on (0,12log⁡3)(0,\frac{1}{2}\log3). Its endpoint values satisfy

R(0)=Dp,q,R(12log⁡3)>1,R(0)=D_{p,q}, \qquad R\left(\frac{1}{2}\log3\right)>1,

where R(0)R(0) is defined by continuity.

Proof. Set z=sL>0z=sL>0. Since sinh⁡zcosh⁡z>z\sinh z\cosh z>z and zcoth⁡z>1z\coth z>1,

0<sinh⁡zcosh⁡z−zsinh⁡2z=∂∂s[scoth⁡(sL)]=z−coth⁡z(zcoth⁡z−1)<z.0<\frac{\sinh z\cosh z-z}{\sinh^2z}=\frac{\partial}{\partial s}[s\coth(sL)]=z-\coth z(z\coth z-1)<z.

Taking logarithmic derivatives in (24) gives

R1′R1=p−q2tanh⁡L+∫(q+2)/2(p+2)/2∂∂s[scoth⁡(sL)] ds>p−q2tanh⁡L,\frac{R_1'}{R_1}=\frac{p-q}{2}\tanh L+\int_{(q+2)/2}^{(p+2)/2}\frac{\partial}{\partial s}[s\coth(sL)]\,ds>\frac{p-q}{2}\tanh L,
R2′R2=p+22tanh⁡L+∫(q+2)/2(p−q)/2∂∂s[scoth⁡(sL)] ds.\frac{R_2'}{R_2}=\frac{p+2}{2}\tanh L+\int_{(q+2)/2}^{(p-q)/2}\frac{\partial}{\partial s}[s\coth(sL)]\,ds.

If p−q≥q+2p-q\ge q+2, both logarithmic derivatives are positive, so R′=R1′+R2′>0R'=R_1'+R_2'>0.

If p−q<q+2p-q<q+2,

R2′R2>−∫(p−q)/2(q+2)/2sL ds=−(2q+2−p)(p+2)L8,\frac{R_2'}{R_2}>-\int_{(p-q)/2}^{(q+2)/2}sL\,ds=-\frac{(2q+2-p)(p+2)L}{8},
R2R1≤(cosh⁡L2)(q+2)/2p−qp+2,\frac{R_2}{R_1}\le\left(\frac{\cosh L}{2}\right)^{(q+2)/2}\frac{p-q}{p+2},

where the second inequality follows because sinh⁡z/z=∫01cosh⁡(tz) dt\sinh z/z=\int_0^1\cosh(tz)\,dt increases on (0,∞)(0,\infty). For 0<L<12log⁡30<L<\frac{1}{2}\log3,

12cosh⁡L<13,tanh⁡L=∫0Lsech⁡2t dt≥3L4.\frac{1}{2}\cosh L<\frac{1}{\sqrt{3}},\qquad\tanh L=\int_0^L\operatorname{sech}^2t\,dt\ge\frac{3L}{4}.

Hence, whenever p−q<q+2p-q<q+2,

2q+2−p2(cosh⁡L2)(q+2)/2<q+223−(q+2)/4≤sup⁡x>0x3−x/2=2elog⁡3<1,\frac{2q+2-p}{2}\left(\frac{\cosh L}{2}\right)^{(q+2)/2}<\frac{q+2}{2}3^{-(q+2)/4}\le\sup_{x>0}x3^{-x/2}=\frac{2}{e\log3}<1,

and combining these bounds for R1′/R1R_1'/R_1, R2′/R2R_2'/R_2 and R2/R1R_2/R_1 gives

R′R1=R1′R1+R2R2′R1R2>p−q2[tanh⁡L−(2q+2−p)L4(cosh⁡L2)(q+2)/2]>(p−q)L8.(26)\frac{R'}{R_1}=\frac{R_1'}{R_1}+\frac{R_2R_2'}{R_1R_2} >\frac{p-q}{2}\left[\tanh L-\frac{(2q+2-p)L}{4}\left(\frac{\cosh L}{2}\right)^{(q+2)/2}\right]>\frac{(p-q)L}{8}. \tag*{(26)}

Since R1(L)>0R_1(L)>0, this proves R′(L)>0R'(L)>0 in the second case.

Letting L↓0L\downarrow0 in (24) and using sinh⁡t/t→1\sinh t/t\to1, we obtain

R(0)=2−(p−q)/2[p+2q+2+2−(q+2)/2p−qq+2]=Dp,q,R(0)=2^{-(p-q)/2}\left[\frac{p+2}{q+2}+2^{-(q+2)/2}\frac{p-q}{q+2}\right]=D_{p,q},

where the last equality follows from (5). For the other endpoint, set x=cos⁡θ=tanh⁡Lx=\cos\theta=\tanh L. Then

Ss(arccos⁡x)=4+[2(1−x)]−s/2+[2(1+x)]−s/2.S_s(\arccos x)=4+[2(1-x)]^{-s/2}+[2(1+x)]^{-s/2}.

At L=12log⁡3L=\frac{1}{2}\log3, we have x=1/2x=1/2, so

1sddxlog⁡Ss(arccos⁡x)∣x=1/2=1−3−s/2−15+3−s/2,\left.\frac{1}{s}\frac{d}{dx}\log S_s(\arccos x)\right|_{x=1/2}=\frac{1-3^{-s/2-1}}{5+3^{-s/2}},

and

dds1−3−s/2−15+3−s/2=4log⁡333−s/2(5+3−s/2)2>0.\frac{d}{ds}\frac{1-3^{-s/2-1}}{5+3^{-s/2}}=\frac{4\log3}{3}\frac{3^{-s/2}}{(5+3^{-s/2})^2}>0.

At L=12log⁡3L=\frac{1}{2}\log3, since p>qp>q,

ddLlog⁡SqpSpq=pq(1−x2)[1−3−q/2−15+3−q/2−1−3−p/2−15+3−p/2]<0.\frac{d}{dL}\log\frac{S_q^p}{S_p^q}=pq(1-x^2)\left[\frac{1-3^{-q/2-1}}{5+3^{-q/2}}-\frac{1-3^{-p/2-1}}{5+3^{-p/2}}\right]<0.

The prefactor in (25) is positive for L>0L>0, so the derivative has the sign of 1−R(L)1-R(L). The endpoint inequality therefore gives R(12log⁡3)>1R(\frac{1}{2}\log3)>1.

Proof of Proposition 3.1. By (25), the derivative of log⁡(Sqp/Spq)\log(S_q^p/S_p^q) has the sign of 1−R(L)1-R(L). Lemma 3.2 therefore gives a unique maximizing value of LL,

L={0,Dp,q≥1,Lp,q∈(0,12log⁡3),R(Lp,q)=1,Dp,q<1.L=\begin{cases} 0, & D_{p,q}\ge1,\\ L_{p,q}\in(0,\frac{1}{2}\log3),\quad R(L_{p,q})=1, & D_{p,q}<1. \end{cases}

For s,L>0s,L>0, set

Ks(L)=log⁡(1−e−sL)−log⁡(1+(41+e2L)s/2).K_s(L)=\log(1-e^{-sL})-\log\left(1+\left(\frac{4}{1+e^{2L}}\right)^{s/2}\right).

The identity 1−e−sL=2e−sL/2sinh⁡(sL/2)1-e^{-sL}=2e^{-sL/2}\sinh(sL/2) gives

eKp+2(L)−Kq+2(L)−1=R(L)−11+(1+e2L4)(p+2)/2.(27)e^{K_{p+2}(L)-K_{q+2}(L)}-1=\frac{R(L)-1}{1+\left(\frac{1+e^{2L}}{4}\right)^{(p+2)/2}}. \tag*{(27)}

Thus R(L)=1R(L)=1 is equivalent to (3.2). Differentiating (27) at the root and using R′(L)>0R'(L)>0 from Lemma 3.2 gives

∂L(Kp+2−Kq+2)=R′(L)1+(1+e2L4)(p+2)/2>0.\partial_L(K_{p+2}-K_{q+2})=\frac{R'(L)}{1+\left(\frac{1+e^{2L}}{4}\right)^{(p+2)/2}}>0.

The map L↦2arctan⁡(e−L)L\mapsto2\arctan(e^{-L}) is one-to-one, so the maximizing value above determines a unique minimizing angle, equal to π/2\pi/2 exactly when Dp,q≥1D_{p,q}\ge1. Equation (23) gives a unique minimizing side length. The rhombus is therefore unique up to congruence and relabelling.

Lemmas 2.1 and 2.3 and Proposition 3.1 prove parts (ii)–(iii) of Theorem 1.1. Equation (23) gives aba_b and ep,q,be_{p,q,b} at θ=θp,q\theta=\theta_{p,q}.

To locate the curve Dp,q=1D_{p,q}=1, define

H(s)=log⁡s−log⁡(1+2s/2),s>0.(28)H(s)=\log s-\log(1+2^{s/2}),\qquad s>0. \tag*{(28)}

Lemma 3.3 (The square region). There is a unique q0∈(1,2)q_0\in(1,2) with H′(q0+2)=0H'(q_0+2)=0. For 0<q<q00<q<q_0, the equation Dp,q=1D_{p,q}=1 has a unique solution p=pc(q)>qp=p_c(q)>q. The function pcp_c is real analytic and strictly decreasing, and its limits are

lim⁡q↓0pc(q)=4,lim⁡q↑q0pc(q)=q0.\lim_{q\downarrow0}p_c(q)=4,\qquad\lim_{q\uparrow q_0}p_c(q)=q_0.

For p>q>0p > q > 0 and b>0b > 0, the square region in Figure 2 is

The phase diagram for $p > q > 0$, showing the regions Square, $D_{p,q} > 1$, and Non-square rhombus, $D_{p,q} < 1$, separated by $D_{p,q}=1$.

Figure 2. The phase diagram for p>q>0p > q > 0. The endpoints of Dp,q=1D_{p,q}=1 are (4,0)(4,0) and (q0,q0)(q_0,q_0), where q0≈1.68887q_0 \approx1.68887.

θp,q=π/2⟺0<q<q0 and q<p≤pc(q).\theta_{p,q} = \pi/2 \quad\Longleftrightarrow\quad0 < q < q_0 \text{ and } q < p \le p_c(q).

Proof. Equation (5) gives

(p+2)(1+2(q+2)/2)−(q+2)(1+2(p+2)/2)=(q+2)2(p+2)/2(Dp,q−1),(p + 2)(1 + 2^{(q+2)/2}) - (q + 2)(1 + 2^{(p+2)/2}) = (q + 2)2^{(p+2)/2}(D_{p,q} - 1),

so (3.8) gives sgn⁡(H(p+2)−H(q+2))=sgn⁡(Dp,q−1)\operatorname{sgn}(H(p + 2) - H(q + 2)) = \operatorname{sgn}(D_{p,q} - 1). Moreover,

H′(s)=1s−log⁡222s/21+2s/2,H′′(s)=−1s2−(log⁡2)242s/2(1+2s/2)2<0.H'(s) = \frac{1}{s} - \frac{\log2}{2}\frac{2^{s/2}}{1 + 2^{s/2}}, \qquad H''(s) = -\frac{1}{s^2} - \frac{(\log2)^2}{4}\frac{2^{s/2}}{(1 + 2^{s/2})^2} < 0.

Since H′(3)>0>H′(4)H'(3) > 0 > H'(4) and H′′<0H'' < 0, there is a unique solution s=q0+2∈(3,4)s = q_0 + 2 \in(3,4) of H′(s)=0H'(s) = 0. The function HH increases strictly up to q0+2q_0 + 2 and then decreases strictly to −∞-\infty. For 0<q<q00 < q < q_0, we have H(q+2)<H(q0+2)H(q + 2) < H(q_0 + 2). The decreasing branch therefore contains exactly one point pc(q)+2p_c(q) + 2 with H(pc(q)+2)=H(q+2)H(p_c(q) + 2) = H(q + 2), and the increasing branch gives no solution with q<p≤q0q < p \le q_0. The sign identity gives a unique root of Dp,q=1D_{p,q} = 1 with p>qp > q.

Since HH is real analytic and H′(pc(q)+2)<0H'(p_c(q) + 2) < 0, the implicit function theorem shows that pcp_c is real analytic and

pc′(q)=H′(q+2)H′(pc(q)+2)<0.p_c'(q) = \frac{H'(q + 2)}{H'(p_c(q) + 2)} < 0.

As q↓0q \downarrow0, we have H(q+2)→H(2)=H(6)H(q + 2) \to H(2) = H(6). Continuity of the inverse of the decreasing branch gives pc(q)+2→6p_c(q) + 2 \to6. As q↑q0q \uparrow q_0, we have H(q+2)→H(q0+2)H(q + 2) \to H(q_0 + 2), and the same inverse gives pc(q)+2→q0+2p_c(q) + 2 \to q_0 + 2.

For q≥q0q \ge q_0 and p>qp > q, strict decrease of HH gives H(p+2)<H(q+2)H(p + 2) < H(q + 2). For 0<q<q00 < q < q_0, strict increase on [q+2,q0+2][q + 2, q_0 + 2] and strict decrease on [q0+2,∞)[q_0 + 2, \infty) give H(p+2)≥H(q+2)H(p + 2) \ge H(q + 2) exactly when q<p≤pc(q)q < p \le p_c(q). Proposition 3.1 and the sign identity therefore give

θp,q=π/2⟺Dp,q≥1⟺H(p+2)≥H(q+2)⟺0<q<q0 and q<p≤pc(q).\theta_{p,q} = \pi/2 \quad\Longleftrightarrow\quad D_{p,q} \ge1 \quad\Longleftrightarrow\quad H(p + 2) \ge H(q + 2) \quad\Longleftrightarrow\quad0 < q < q_0 \text{ and } q < p \le p_c(q).

By Lemma 3.3, the strict square region above q=1q = 1 is

{(p,q):p>q>1, Dp,q>1}={(p,q):1<q<q0, q<p<pc(q)}.\{(p,q) : p > q > 1,\ D_{p,q} > 1\} = \{(p,q) : 1 < q < q_0,\ q < p < p_c(q)\}.

Equation (5) gives D2,4/3>1>D5/2,1D_{2,4/3} > 1 > D_{5/2,1}. Hence Lemma 3.3 gives pc(1)<5/2p_c(1) < 5/2. Since pcp_c decreases, Dp,q<1D_{p,q} < 1 for p>q>1p > q > 1 with p≥5/2p \ge5/2.

Proof of part (iv) of Theorem 1.1. (a) Fix 0<L<12log⁡30 < L < \frac{1}{2}\log3 and set t=12log⁡(4/(1+e2L))t = \frac{1}{2}\log(4/(1 + e^{2L})). Then

∂s2Ks(L)=−L2esL(esL−1)2−t2est(1+est)2<0.\partial_s^2 K_s(L) = -\frac{L^2 e^{sL}}{(e^{sL} - 1)^2} - \frac{t^2 e^{st}}{(1 + e^{st})^2} < 0.

By Proposition 3.1 and (3.7), Kp+2(Lp,q)=Kq+2(Lp,q)K_{p+2}(L_{p,q}) = K_{q+2}(L_{p,q}). Strict concavity in ss then implies

∂sKs∣s=q+2>Kp+2−Kq+2p−q=0>∂sKs∣s=p+2.\left.\partial_s K_s\right|_{s=q+2} > \frac{K_{p+2} - K_{q+2}}{p - q} = 0 > \left.\partial_s K_s\right|_{s=p+2}.

The positive LL-derivative of Kp+2−Kq+2K_{p+2} - K_{q+2} at the root, proved in Proposition 3.1, allows us to apply the implicit function theorem. Thus Lp,qL_{p,q} is real analytic, and differentiating the defining equation gives

∂pLp,q=−∂sKs∣s=p+2∂L(Kp+2−Kq+2)>0,∂qLp,q=∂sKs∣s=q+2∂L(Kp+2−Kq+2)>0.\partial_p L_{p,q} = -\frac{\left.\partial_s K_s\right|_{s=p+2}}{\partial_L(K_{p+2} - K_{q+2})} > 0,\qquad\partial_q L_{p,q} = \frac{\left.\partial_s K_s\right|_{s=q+2}}{\partial_L(K_{p+2} - K_{q+2})} > 0.

By Proposition 3.1, θp,q=2arctan⁡(e−Lp,q)\theta_{p,q}=2\arctan(e^{-L_{p,q}}). Since ∂L[2arctan⁡(e−L)]=−sech⁡L\partial_L[2\arctan(e^{-L})]=-\operatorname{sech}L, both angle derivatives are negative.

(b) Fix q>0q>0. By (5), Dp,q→0D_{p,q}\to0 as p→∞p\to\infty; choose p1>qp_1>q so that Dp,q<1D_{p,q}<1 for p≥p1p\ge p_1. Part (a) then gives Lp,q≥Lp1,q>0L_{p,q}\ge L_{p_1,q}>0 on this range. Equation (3.2) reads

(41+e2Lp,q)(p+2)/2=1−e−(p+2)Lp,q1−e−(q+2)Lp,q[1+(41+e2Lp,q)(q+2)/2]−1=O(1).(29)\left(\frac{4}{1+e^{2L_{p,q}}}\right)^{(p+2)/2} = \frac{1-e^{-(p+2)L_{p,q}}}{1-e^{-(q+2)L_{p,q}}} \left[1+\left(\frac{4}{1+e^{2L_{p,q}}}\right)^{(q+2)/2}\right]-1=O(1). \tag*{(29)}

Since 4/(1+e2L)4/(1+e^{2L}) exceeds 1 for L<12log⁡3L<\frac{1}{2}\log3, boundedness of the left-hand side gives Lp,q→12log⁡3L_{p,q}\to\frac{1}{2}\log3. Passing to the limit on the right-hand side gives

(41+e2Lp,q)(p+2)/2⟶1+3−(q+2)/21−3−(q+2)/2.\left(\frac{4}{1+e^{2L_{p,q}}}\right)^{(p+2)/2}\longrightarrow\frac{1+3^{-(q+2)/2}}{1-3^{-(q+2)/2}}.

At L=12log⁡3L=\frac{1}{2}\log3 the derivatives of 2arctan⁡(e−L)2\arctan(e^{-L}) and 12log⁡(4/(1+e2L))\frac{1}{2}\log(4/(1+e^{2L})) are −32-\frac{\sqrt{3}}{2} and −34-\frac{3}{4}, hence

lim⁡L↑12log⁡32arctan⁡(e−L)−π/312log⁡(4/(1+e2L))=23.\lim_{L\mathrel{\uparrow}\frac{1}{2}\log3} \frac{2\arctan(e^{-L})-\pi/3}{\frac{1}{2}\log(4/(1+e^{2L}))} =\frac{2}{\sqrt{3}}.

Multiplying this limit by the logarithm of the preceding limit proves part (iv)(b).

(c) Take p=2qp=2q with q≥2q\ge2; Lemma 3.3 gives D2q,q<1D_{2q,q}<1, and part (a) gives L2q,q≥L4,2>0L_{2q,q}\ge L_{4,2}>0. Write L=L2q,qL=L_{2q,q}, L0=12log⁡3L_0=\frac{1}{2}\log3 and T=41+e2L>1T=\frac{4}{1+e^{2L}}>1. Equation (3.2) becomes

Tq/2−11+T−(q+2)/2=e−(q+2)L(1−e−qL)1−e−(q+2)L.\frac{T^{q/2}-1}{1+T^{-(q+2)/2}} = \frac{e^{-(q+2)L}(1-e^{-qL})}{1-e^{-(q+2)L}}.

Because log⁡T=∫LL02e2t1+e2t dt≥L0−L\log T=\int_L^{L_0}\frac{2e^{2t}}{1+e^{2t}}\,dt\ge L_0-L,

0<q(L0−L)≤qlog⁡T≤2(Tq/2−1)=O(e−(q+2)L)=o(1).0<q(L_0-L)\le q\log T\le2(T^{q/2}-1)=O(e^{-(q+2)L})=o(1).

Thus e−(q+2)L=3−(q+2)/2e(q+2)(L0−L)=O(3−q/2)e^{-(q+2)L}=3^{-(q+2)/2}e^{(q+2)(L_0-L)}=O(3^{-q/2}), so q(L0−L)=O(3−q/2)q(L_0-L)=O(3^{-q/2}). Expanding both sides of (3.2) gives

Tq/2−11+T−(q+2)/2=q4log⁡T+O(3−q)=3q8(L0−L)+O(3−q),\frac{T^{q/2}-1}{1+T^{-(q+2)/2}} =\frac{q}{4}\log T+O(3^{-q}) =\frac{3q}{8}(L_0-L)+O(3^{-q}),
e−(q+2)L(1−e−qL)1−e−(q+2)L=3−(q+2)/2+O(3−q),\frac{e^{-(q+2)L}(1-e^{-qL})}{1-e^{-(q+2)L}} =3^{-(q+2)/2}+O(3^{-q}),

Hence L0−L=89q3−q/2+O(3−q/q)L_0-L=\frac{8}{9q}3^{-q/2}+O(3^{-q}/q). The angle formula in Proposition 3.1 gives

θ2q,q−π3=32(L0−L)+O((L0−L)2)=433q−13−q/2+O(3−q/q).\theta_{2q,q}-\frac{\pi}{3} =\frac{\sqrt{3}}{2}(L_0-L)+O((L_0-L)^2) =\frac{4}{3\sqrt{3}}q^{-1}3^{-q/2}+O(3^{-q}/q).

At θ=θ2q,q\theta=\theta_{2q,q}, Taylor expansion gives

qlog⁡(2sin⁡(θ/2))=233−q/2+O(3−q),qlog⁡(2cos⁡(θ/2)/3)=O(3−q/2),q\log(2\sin(\theta/2))=\frac{2}{3}3^{-q/2}+O(3^{-q}),\qquad q\log(2\cos(\theta/2)/\sqrt{3})=O(3^{-q/2}),

and substitution into (21) gives

Sq(θ)=5+133−q/2+O(3−q),S2q(θ)=5−433−q/2+O(3−q).S_q(\theta)=5+\frac{1}{3}3^{-q/2}+O(3^{-q}),\qquad S_{2q}(\theta)=5-\frac{4}{3}3^{-q/2}+O(3^{-q}).

For b=2b=2, equation (23) then reads

qlog⁡a2=log⁡S2q(θ)Sq(θ)=−133−q/2+O(3−q),q\log a_2=\log\frac{S_{2q}(\theta)}{S_q(\theta)} =-\frac{1}{3}3^{-q/2}+O(3^{-q}),

and in particular a2→1a_2 \to1.

(d) Write Rp,qR_{p,q} for RR in (24). Dividing the numerator and denominator of each hyperbolic quotient by LL gives a joint analytic extension near (p∗,q∗,0)(p^*,q^*,0), even in LL. The expansion

(cosh⁡L)γsinh⁡(uL)sinh⁡(vL)=uv[1+(γ2+u2−v26)L2+O(L4)](\cosh L)^\gamma\frac{\sinh(uL)}{\sinh(vL)}=\frac{u}{v}\left[1+\left(\frac{\gamma}{2}+\frac{u^2-v^2}{6}\right)L^2+O(L^4)\right]

gives

Rp,q′′(0)=(p−q)(p+q+10)12Rp,q(0)+(q+2)(4−2p+q)12R2(0)=p−q12[(p+q+10)Dp,q+(4−2p+q)2−(p+2)/2].\begin{aligned} R_{p,q}''(0)=\frac{(p-q)(p+q+10)}{12}R_{p,q}(0)+\frac{(q+2)(4-2p+q)}{12}R_2(0) \\ &=\frac{p-q}{12}\left[(p+q+10)D_{p,q}+(4-2p+q)2^{-(p+2)/2}\right]. \end{aligned}

On the curve Dp,q=1D_{p,q}=1 the bracket equals p(1−2−p/2)+q+10+(q+4)2−(p+2)/2>0p(1-2^{-p/2})+q+10+(q+4)2^{-(p+2)/2}>0.

Let (p,q)→(p∗,q∗)(p,q)\to(p^*,q^*) through Dp,q<1D_{p,q}<1. Proposition 3.1 and (27) give Rp,q(Lp,q)=1R_{p,q}(L_{p,q})=1. If a subsequence satisfied Lp,q→L∗>0L_{p,q}\to L^*>0, Lemma 3.2 and continuity would give 1=Rp∗,q∗(L∗)>Rp∗,q∗(0)=11=R_{p^*,q^*}(L^*)>R_{p^*,q^*}(0)=1, a contradiction. Thus Lp,q→0L_{p,q}\to0. The even Taylor expansion is uniform near (p∗,q∗)(p^*,q^*) and gives

1−Dp,q=12Rp,q′′(0)Lp,q2+O(Lp,q4),π2−2arctan⁡(e−Lp,q)=Lp,q+O(Lp,q3).1-D_{p,q}=\frac{1}{2}R_{p,q}''(0)L_{p,q}^2+O(L_{p,q}^4),\qquad\frac{\pi}{2}-2\arctan(e^{-L_{p,q}})=L_{p,q}+O(L_{p,q}^3).

Therefore

π/2−θp,q1−Dp,q=1+O(Lp,q2)12Rp,q′′(0)+O(Lp,q2)⟶2Rp∗,q∗′′(0).\frac{\pi/2-\theta_{p,q}}{\sqrt{1-D_{p,q}}}=\frac{1+O(L_{p,q}^2)}{\sqrt{\frac{1}{2}R_{p,q}''(0)+O(L_{p,q}^2)}}\longrightarrow\sqrt{\frac{2}{R_{p^*,q^*}''(0)}}.

and substitution of the formula for R′′(0)R''(0) gives the stated constant. □

Proof of Remark 1.2. Part (iv)(c) of the proof of Theorem 1.1 gives a2→1a_2\to1 and θ2q,q→π/3\theta_{2q,q}\to\pi/3. Substitution in part (ii) gives the stated matrix. Its pair distances are

r12=r23=r34=r14=r24=1,r13=3.r_{12}=r_{23}=r_{34}=r_{14}=r_{24}=1,\qquad r_{13}=\sqrt{3}.

The Gram determinant argument in the proof of Lemma 2.3 rules out six unit pair distances in the plane. Thus every four-point configuration with finite Heitmann–Radin energy has at most five unit bonds, and the displayed limit is a minimizer. □

The classical Lennard-Jones case

Theorem 1.1 gives a sharp inequality for the six pair distances. For (p,q)=(12,6)(p,q)=(12,6), we determine the minimizing angle, side length, and energy explicitly.

Corollary 4.1 (Sharp scale-free inequality). For every configuration XX and p>q>0p>q>0,

(∑1≤i<j≤4∣xi−xj∣−q)p(∑1≤i<j≤4∣xi−xj∣−p)q≥Sq(θp,q)pSp(θp,q)q.(30)\frac{\left(\sum_{1\le i<j\le4}|x_i-x_j|^{-q}\right)^p}{\left(\sum_{1\le i<j\le4}|x_i-x_j|^{-p}\right)^q}\ge\frac{S_q(\theta_{p,q})^p}{S_p(\theta_{p,q})^q}. \tag*{(30)}

Equality holds if and only if XX is similar to the minimizing rhombus in Theorem 1.1.

Proof. Fix b>0b>0. By Theorem 1.1(ii) and the scale minimization (23),

min⁡t>0Ep,q,b(tX)=−εb[(∑1≤i<j≤4∣xi−xj∣−q)p(∑1≤i<j≤4∣xi−xj∣−p)q]1/(p−q)≥ep,q,b=−εb[Sq(θp,q)pSp(θp,q)q]1/(p−q).\begin{aligned} \min_{t>0}E_{p,q,b}(tX)=-\varepsilon_b\left[\frac{\left(\sum_{1\le i<j\le4}|x_i-x_j|^{-q}\right)^p}{\left(\sum_{1\le i<j\le4}|x_i-x_j|^{-p}\right)^q}\right]^{1/(p-q)} \\ &\ge e_{p,q,b}=-\varepsilon_b\left[\frac{S_q(\theta_{p,q})^p}{S_p(\theta_{p,q})^q}\right]^{1/(p-q)}. \end{aligned}

Since εb>0\varepsilon_b > 0, this proves (30). Lemma 2.3 and Proposition 3.1 give

min⁡t>0Ep,q,b(tX)=ep,q,b⟺X is similar to Xp,q,b∗.\min_{t>0} E_{p,q,b}(tX) = e_{p,q,b} \quad\Longleftrightarrow\quad X\text{ is similar to }X^*_{p,q,b}.

Proposition 4.2 (Classical Lennard–Jones potential). Let b>0b > 0. The minimizing angle is θ12,6=arccos⁡u∗\theta_{12,6} = \arccos\sqrt{u^*}, where u∗u^* is the unique root in (0,1/4)(0,1/4) of

397−1585u−336u2+1008u3−512u4=0.(31)397 - 1585u - 336u^2 + 1008u^3 - 512u^4 = 0. \tag*{(31)}

The minimizing side length and energy satisfy

ab6=128(1−u∗)6+1+15u∗+15u∗2+u∗34b(1−u∗)3[16(1−u∗)3+1+3u∗],(32)a_b^6 = \frac{128(1-u^*)^6 + 1 + 15u^* + 15u^{*2} + u^{*3}}{4b(1-u^*)^3[16(1-u^*)^3 + 1 + 3u^*]}, \tag*{(32)}
e12,6,b=−b2[16(1−u∗)3+1+3u∗]22[128(1−u∗)6+1+15u∗+15u∗2+u∗3].e_{12,6,b} = -\frac{b^2[16(1-u^*)^3 + 1 + 3u^*]^2}{2[128(1-u^*)^6 + 1 + 15u^* + 15u^{*2} + u^{*3}]}.

Proof. By (5), D12,6=115/512<1D_{12,6} = 115/512 < 1. Theorem 1.1(iii) therefore gives a minimizing angle in (π/3,π/2)(\pi/3,\pi/2). Set u=cos⁡2θ∈(0,1/4)u = \cos^2\theta\in(0,1/4). Then

4sin⁡2(θ/2)=2(1−u),4cos⁡2(θ/2)=2(1+u),4\sin^2(\theta/2) = 2(1-\sqrt{u}), \qquad4\cos^2(\theta/2) = 2(1+\sqrt{u}),

and expansion of (21) gives

S6=16(1−u)3+1+3u4(1−u)3,S12=128(1−u)6+1+15u+15u2+u332(1−u)6.S_6 = \frac{16(1-u)^3 + 1 + 3u}{4(1-u)^3}, \qquad S_{12} = \frac{128(1-u)^6 + 1 + 15u + 15u^2 + u^3}{32(1-u)^6}.

Differentiating, we obtain

ddulog⁡S62S12=3(1−u)2(397−1585u−336u2+1008u3−512u4)[16(1−u)3+1+3u][128(1−u)6+1+15u+15u2+u3].\frac{d}{du}\log\frac{S_6^2}{S_{12}} = \frac{3(1-u)^2(397 - 1585u - 336u^2 + 1008u^3 - 512u^4)} {[16(1-u)^3 + 1 + 3u][128(1-u)^6 + 1 + 15u + 15u^2 + u^3]}.

By (23), the minimizing angle maximizes S62/S12S_6^2/S_{12}. Thus u=cos⁡2θ12,6u = \cos^2\theta_{12,6} satisfies (31). On [0,1/4][0,1/4], the derivative of the polynomial is at most −1585+3024/16<0-1585 + 3024/16 < 0, so this root is unique. By (23),

ab6=2S12bS6,e12,6,b=−b2S624S122,a_b^6 = \frac{2S_{12}}{bS_6}, \qquad e_{12,6,b} = -\frac{b^2S_6^2}{4S_{12}^2},

and substitution of S6S_6 and S12S_{12} at u=u∗u = u^* gives (32). □

For b=2b = 2,

u∗≈0.2459296640551374,a2≈0.9980123196944918,e12,6,2≈−5.073420858462792.u^* \approx0.2459296640551374,\qquad a_2 \approx0.9980123196944918,\qquad e_{12,6,2} \approx-5.073420858462792.

Acknowledgements. The research of S. Luo is partially supported by the National Natural Science Foundation of China (NSFC) under Grant Nos. 12261045 and 12001253, and by the Jiangxi Jieqing Fund under Grant No. 20242BAB23001.

Statements and Declarations: there is no conflict of interest.

Data availability: the manuscript has no associated data.

School of Mathematics and Statistics, Jiangxi Normal University, Nanchang 330022, China

Email address: [email protected]

References

  1. [1]Béterin, L., De Luca, L., Petrache, M.: Crystallization to the square lattice for a two-body potential, Arch. Ration. Mech. Anal. 240, 987–1053 (2021).
  2. [2]Béterin, L., Faulhuber, M., Steinerberger, S.: A variational principle for Gaussian lattice sums, Forum Math. Pi 14, article no. e22 (2026).
  3. [3]Béterin, L., Petrache, M.: Optimal and non-optimal lattices for non-completely monotone interaction potentials, Anal. Math. Phys. 9, 2033–2073 (2019).
  4. [4]Béterin, L., Zhang, P.: Minimization of energy per particle among Bravais lattices in R²: Lennard–Jones and Thomas–Fermi cases, Commun. Contemp. Math. 17, article no. 1450049 (2015).
  5. [5]Béterin, L.: Local variational study of 2d lattice energies and application to Lennard-Jones type interactions, Nonlinearity 31(9) (2018), 3973–4005.
  6. [6]Béterin, L.: Optimality of the triangular lattice for Lennard-Jones type lattice energies: a computer-assisted method, J. Phys. A 56 (2023), no. 14, Paper No. 145204, 19 pp.
  7. [7]Blanc, X., Lewin, M.: The crystallization conjecture: A review, EMS Surv. Math. Sci. 2, 255–306 (2015).arxiv.org/abs/1504.01153
  8. [8]Chen, X., Oshita, Y.: An application of the modular function in nonlocal variational problems, Arch. Ration. Mech. Anal. 186, 109–132 (2007).
  9. [9]Cohn, H., Kumar, A., Miller, S. D., Radchenko, D., Viazovska, M.: Universal optimality of the E₈ and Leech lattices and interpolation formulas, Ann. of Math. (2) 196, 983–1082 (2022).
  10. [10]Connelly, R.: Rigidity and energy, Invent. Math. 66, 11–33 (1982).DOI
  11. [11]De Luca, L., Friesecke, G.: Classification of particle numbers with unique Heitmann–Radin minimizer, J. Stat. Phys. 167, 1586–1592 (2017).
  12. [12]De Luca, L., Friesecke, G.: Crystallization in two dimensions and a discrete Gauss–Bonnet theorem, J. Nonlinear Sci. 28, 69–90 (2018).
  13. [13]Del Nin, G., De Luca, L.: The square sticky disk: crystallization and Gamma-convergence to the octagonal anisotropic perimeter, SIAM J. Math. Anal. 58, 3238–3266 (2026).
  14. [14]E, W. and Li, D.: On the Crystallization of 2D Hexagonal Lattices, Commun. Math. Phys. 286 (2009), no. 3, 1099–1140.DOI
  15. [15]Heitmann, R. C., Radin, C.: The ground state for sticky disks, J. Stat. Phys. 22, 281–287 (1980).DOI
  16. [18]Luo, S., Ren, X., Wei, J.: Nonhexagonal lattices from a two species interacting system, SIAM J. Math. Anal. 52, 1903–1942 (2020).arxiv.org/abs/1902.09611
  17. [19]Luo, S., Wei, J.: On minima of differences of Epstein zeta functions and exact solutions to Lennard–Jones lattice energy, J. Eur. Math. Soc., published online first (2025).
  18. [20]Luo, S., Wei, J.: On minima of sum of theta functions and application to Mueller–Ho conjecture, Arch. Ration. Mech. Anal. 243, 139–199 (2022).
  19. [22]Mueller, E. J. and Ho, T. L.: Two-component Bose-Einstein condensates with a large number of vortices, Physical review letters, 88 (2002), 180403.
  20. [23]Montgomery, H. L.: Minimal theta functions, Glasgow Math. J. 30, 75–85 (1988).DOI
  21. [24]Petrache, M., Serfaty, S.: Crystallization for Coulomb and Riesz interactions as a consequence of the Cohn–Kumar conjecture, Proc. Amer. Math. Soc. 148, 3047–3057 (2020).
  22. [25]Radin, C.: Low temperature and the origin of crystallization symmetry, International Journal of Modern Physics B, Vol. 01, No. 05-06, pp. 1157–1191 (1987).
  23. [26]Sandier, E., Serfaty, S.: From the Ginzburg–Landau model to vortex lattice problems, Comm. Math. Phys. 313, 635–743 (2012).
  24. [27]Serfaty, S.: Lectures on Coulomb and Riesz gases, to appear in the AMS Colloquium series.
  25. [28]Theil, F.: A proof of crystallization in two dimensions, Comm. Math. Phys. 262, 209–236 (2006).

Paper details

vlnc:W_cp83cng7
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents