Introduction and main results
The Yukawa potential describes a screened Coulomb interaction and has been studied extensively since 1935 [39 ] , in which Yukawa combined Heisenberg’s idea of a short-range force interaction and Fermi’s idea of an exchange particle in order to resolve the problem of the neutron-proton interaction. Its screening parameter introduces a competition between two cubic lattice structures: numerical phase diagrams and lattice calculations predict BCC at weak screening and FCC at strong screening [31 , 16 , 17 , 19 ] . The transition has also been studied through applied-strain free-energy calculations [18 ] . The purpose of this paper is to determine the global minimizing Bravais lattices for every positive screening parameter at zero temperature and to prove that there is exactly one transition.
For κ > 0 \kappa>0 κ > 0 , the potential and the energy per particle are
v κ ( r ) = e − κ r r , E κ ( L ) = 1 2 ∑ v ∈ L ∖ { 0 } e − κ ∣ v ∣ ∣ v ∣ . (1) v_\kappa(r)=\frac{e^{-\kappa r}}{r},\qquad E_\kappa(L)=\frac{1}{2}\sum_{v\in L\setminus\{0\}}\frac{e^{-\kappa|v|}}{|v|}. \tag*{(1)} v κ ( r ) = r e − κ r , E κ ( L ) = 2 1 v ∈ L ∖ { 0 } ∑ ∣ v ∣ e − κ ∣ v ∣ . (1)
The history of this potential involves some physical settings. Debye and Hückel obtained the screened Coulomb form in 1923 in their theory of electrolyte solutions [13 ] . Mobile ions rearrange around a charged particle, forming an ionic atmosphere that reduces its electrostatic influence at large distances.
Mathematically, the normalized kernel (Yukawa potential)
G κ ( x ) = e − κ ∣ x ∣ 4 π ∣ x ∣ G_\kappa(x)=\frac{e^{-\kappa|x|}}{4\pi|x|} G κ ( x ) = 4 π ∣ x ∣ e − κ ∣ x ∣
is the fundamental solution of the modified Helmholtz operator:
( − Δ + κ 2 ) G κ = δ 0 in R 3 . (-\Delta+\kappa^2)G_\kappa=\delta_0\quad\text{in }\mathbb{R}^3. ( − Δ + κ 2 ) G κ = δ 0 in R 3 .
2020 Mathematics Subject Classification . 52C17, 82B20, 11H31.
Key words and phrases. Yukawa potential, lattice energy, body-centred cubic lattice, face-centred cubic lattice, phase transition.
This equation explains the combination of a Coulomb singularity and exponential spatial decay. The factor 1 / 2 1/2 1/2 in (1) counts each pair once. For a lattice L = g Z 3 L = g\mathbb{Z}^{3} L = g Z 3 , write ∣ L ∣ = ∣ det g ∣ |L| = |\det g| ∣ L ∣ = ∣ det g ∣ for its covolume. We first impose ∣ L ∣ = 1 |L| = 1 ∣ L ∣ = 1 and identify lattices that differ by an orthogonal transformation. Write L ∗ = g − T Z 3 L^{*} = g^{-T}\mathbb{Z}^{3} L ∗ = g − T Z 3 and λ 1 ( L ) = min v ∈ L ∖ { 0 } ∣ v ∣ \lambda_{1}(L) = \min_{v \in L\setminus\{0\}} |v| λ 1 ( L ) = min v ∈ L ∖ { 0 } ∣ v ∣ . The unit-covolume face-centred and body-centred cubic lattices are
D 3 = { m ∈ Z 3 : m 1 + m 2 + m 3 ∈ 2 Z } , L F = 2 − 1 / 3 D 3 , L B = L F ∗ . (2) D_{3} = \{m \in\mathbb{Z}^{3} : m_{1} + m_{2} + m_{3} \in2\mathbb{Z}\}, \qquad L_{\mathrm{F}} = 2^{-1/3}D_{3}, \qquad L_{\mathrm{B}} = L_{\mathrm{F}}^{*}. \tag*{(2)} D 3 = { m ∈ Z 3 : m 1 + m 2 + m 3 ∈ 2 Z } , L F = 2 − 1/3 D 3 , L B = L F ∗ . (2)
Here L F L_{\mathrm{F}} L F is the face-centred cubic (FCC) lattice and L B L_{\mathrm{B}} L B is the body-centred cubic (BCC) lattice, with ∣ L F ∣ = ∣ L B ∣ = 1 |L_{\mathrm{F}}| = |L_{\mathrm{B}}| = 1 ∣ L F ∣ = ∣ L B ∣ = 1 . We shall use a = λ 1 ( L F ) = 2 1 / 6 a = \lambda_{1}(L_{\mathrm{F}}) = 2^{1/6} a = λ 1 ( L F ) = 2 1/6 . Standard facts about these lattices can be found in [12 ] .
For t > 0 t > 0 t > 0 , define
Θ ( t , L ) = ∑ v ∈ L e − π t ∣ v ∣ 2 . (3) \Theta(t,L) = \sum_{v \in L}e^{-\pi t|v|^{2}}. \tag*{(3)} Θ ( t , L ) = v ∈ L ∑ e − π t ∣ v ∣ 2 . (3)
Theorem 1.1 (Minima of three-dimensional Yukawa lattice energy). Let L ⊂ R 3 L \subset\mathbb{R}^{3} L ⊂ R 3 be a lattice with ∣ L ∣ = 1 |L| = 1 ∣ L ∣ = 1 , and let κ > 0 \kappa> 0 κ > 0 . There is a unique threshold κ ∗ ∈ ( 0 , π ) \kappa_{*} \in(0,\sqrt{\pi}) κ ∗ ∈ ( 0 , π ) , characterized by
E κ ∗ ( L B ) = E κ ∗ ( L F ) . E_{\kappa_{*}}(L_{\mathrm{B}}) = E_{\kappa_{*}}(L_{\mathrm{F}}). E κ ∗ ( L B ) = E κ ∗ ( L F ) . More explicitly, κ ∗ \kappa_{*} κ ∗ is the unique positive solution of
∫ 1 ∞ ( Θ ( t , L B ) − Θ ( t , L F ) ) [ t − 1 / 2 e − κ ∗ 2 / ( 4 π t ) − e − κ ∗ 2 t / ( 4 π ) ] d t = 0. (4) \int_{1}^{\infty}\left(\Theta(t,L_{\mathrm{B}})-\Theta(t,L_{\mathrm{F}})\right)\left[t^{-1/2}e^{-\kappa_{*}^{2}/(4\pi t)}-e^{-\kappa_{*}^{2}t/(4\pi)}\right]\,dt = 0. \tag*{(4)} ∫ 1 ∞ ( Θ ( t , L B ) − Θ ( t , L F ) ) [ t − 1/2 e − κ ∗ 2 / ( 4 π t ) − e − κ ∗ 2 t / ( 4 π ) ] d t = 0. (4) Numerically,
κ ∗ ≈ 1.71792186846990. \kappa_{*} \approx1.71792186846990. κ ∗ ≈ 1.71792186846990. The following statements hold.
(1) If κ > κ ∗ \kappa> \kappa_{*} κ > κ ∗ , then E κ ( L ) ≥ E κ ( L F ) E_{\kappa}(L) \ge E_{\kappa}(L_{\mathrm{F}}) E κ ( L ) ≥ E κ ( L F ) , with equality if and only if L L L is orthogonally equivalent to L F L_{\mathrm{F}} L F , corresponding to FCC lattice.
(2) If 0 < κ < κ ∗ 0 < \kappa< \kappa_{*} 0 < κ < κ ∗ , then E κ ( L ) ≥ E κ ( L B ) E_{\kappa}(L) \ge E_{\kappa}(L_{\mathrm{B}}) E κ ( L ) ≥ E κ ( L B ) , with equality if and only if L L L is orthogonally equivalent to L B L_{\mathrm{B}} L B , corresponding to BCC lattice.
(3) If κ = κ ∗ \kappa= \kappa_{*} κ = κ ∗ , then E κ ( L ) ≥ E κ ( L F ) = E κ ( L B ) E_{\kappa}(L) \ge E_{\kappa}(L_{\mathrm{F}}) = E_{\kappa}(L_{\mathrm{B}}) E κ ( L ) ≥ E κ ( L F ) = E κ ( L B ) , with equality if and only if L L L is orthogonally equivalent to L F L_{\mathrm{F}} L F or L B L_{\mathrm{B}} L B , corresponding to FCC or BCC lattice.
Equivalently, up to orthogonal transformations,
arg min ∣ L ∣ = 1 E κ ( L ) = { { L B } , 0 < κ < κ ∗ , { L B , L F } , κ = κ ∗ , { L F } , κ > κ ∗ . (5) \underset{|L|=1}{\operatorname{arg\,min}}\ E_{\kappa}(L) =
\begin{cases}
\{L_{\mathrm{B}}\}, & 0 < \kappa< \kappa_{*},\\
\{L_{\mathrm{B}},L_{\mathrm{F}}\}, & \kappa= \kappa_{*},\\
\{L_{\mathrm{F}}\}, & \kappa> \kappa_{*}.
\end{cases} \tag*{(5)} ∣ L ∣ = 1 arg min E κ ( L ) = ⎩ ⎨ ⎧ { L B } , { L B , L F } , { L F } , 0 < κ < κ ∗ , κ = κ ∗ , κ > κ ∗ . (5) At density ϱ > 0 \varrho> 0 ϱ > 0 , the minimizing lattices are the corresponding FCC and BCC lattices multiplied by ϱ − 1 / 3 \varrho^{-1/3} ϱ − 1/3 , and the transition occurs at κ = ϱ 1 / 3 κ ∗ \kappa= \varrho^{1/3}\kappa_{*} κ = ϱ 1/3 κ ∗ .
The density dependence follows from
E κ ( ϱ − 1 / 3 L ) = ϱ 1 / 3 E κ ϱ − 1 / 3 ( L ) . (6) E_{\kappa}(\varrho^{-1/3}L) = \varrho^{1/3}E_{\kappa\varrho^{-1/3}}(L). \tag*{(6)} E κ ( ϱ − 1/3 L ) = ϱ 1/3 E κ ϱ − 1/3 ( L ) . (6)
Throughout the paper, minimization is over Bravais lattices of fixed density, with one particle in each primitive cell.
The comparison underlying Theorem 1.1 also includes zero screening. It therefore yields the following Coulomb consequence. Here Z ( L , s ) Z(L,s) Z ( L , s ) denotes the meromorphic continuation of ∑ v ∈ L ∖ { 0 } ∣ v ∣ − 2 s \sum_{v\in L\setminus\{0\}}|v|^{-2s} ∑ v ∈ L ∖ { 0 } ∣ v ∣ − 2 s , initially defined for ℜ s > 3 / 2 \Re s > 3/2 ℜ s > 3/2 .
Corollary 1.1 (Coulomb minimizers). For every lattice L L L with ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 ,
lim κ ↓ 0 ( 2 E κ ( L ) − 4 π κ 2 ) = Z ( L , 1 / 2 ) ≥ Z ( L B , 1 / 2 ) , (7) \lim_{\kappa\downarrow0}\left(2E_{\kappa}(L)-\frac{4\pi}{\kappa^{2}}\right)=Z(L,1/2)\ge Z(L_{\mathrm{B}},1/2), \tag*{(7)} κ ↓ 0 lim ( 2 E κ ( L ) − κ 2 4 π ) = Z ( L , 1/2 ) ≥ Z ( L B , 1/2 ) , (7) with equality in the inequality if and only if L L L is orthogonally equivalent to L B L_{\mathrm{B}} L B . Equivalently, Z ( M , 1 ) ≥ Z ( L F , 1 ) Z(M,1)\ge Z(L_{\mathrm{F}},1) Z ( M , 1 ) ≥ Z ( L F , 1 ) for every lattice M M M with ∣ M ∣ = 1 |M|=1 ∣ M ∣ = 1 , with equality if and only if M M M is orthogonally equivalent to L F L_{\mathrm{F}} L F .
Muratov and Novaga [26 Theorem 1] proved that BCC uniquely minimizes the renormalized three-dimensional Coulomb energy among Bravais lattices of fixed covolume. Their energy is exactly the limit in (7) . Thus Corollary 1.1 recovers their theorem, including its equality case, as the zero-screening consequence of our uniform Yukawa comparison. In their Madelung normalization,
K L = 1 2 ( 3 4 π ) 1 / 3 Z ( L , 1 / 2 ) ( ∣ L ∣ = 1 ) . K_L=\frac{1}{2}\left(\frac{3}{4\pi}\right)^{1/3}Z(L,1/2)\qquad(|L|=1). K L = 2 1 ( 4 π 3 ) 1/3 Z ( L , 1/2 ) ( ∣ L ∣ = 1 ) .
The proof below obtains uniqueness from the strict comparison at zero screening and the strict FCC–BCC energy gap.
The lattice minimization problem belongs to the broader theory of crystallization; see the reviews of Blanc–Lewin [5 ] and Lewin [21 ] .
Fourier analysis provides another approach to energy minimization. Cohn–Elkies [8 ] developed linear programming bounds for the density of sphere packings. Cohn–Kumar [9 ] proved universal optimality for sharp configurations on spheres. Cohn, Kumar, Miller, Radchenko and Viazovska [10 ] subsequently proved universal optimality of the E 8 E_8 E 8 and Leech lattices at fixed density in dimensions 8 and 24, respectively, for completely monotone functions of squared distance. The Yukawa potential belongs to this class: its positive Gaussian representation (9) shows that it is completely monotone as a function of squared distance.
The Gaussian representation also connects the present problem with the Gaussian core model. Cohn, Kumar and Schürmann [11 ] numerically investigated its ground states and identified formal duality relations. Cohn–de Courcy-Ireland [7 ] established asymptotically sharp energy bounds for Gaussian interactions in a range of parameters, at fixed density as the dimension tends to infinity.
Renormalized energy connects these problems with the microscopic theory of Coulomb and Riesz gases. Sandier–Serfaty [33 , 34 ] develop this approach for planar vortices and logarithmic gases. Rougerie–Serfaty [32 ] and Petrache–Serfaty [28 ] identify microscopic renormalized energies in next-order Coulomb and Riesz asymptotics. Leblé–Serfaty [20 ] prove an empirical-field large-deviation principle, while Armstrong–Serfaty [1 ] obtain local laws and rigidity estimates. Petrache–Serfaty [29 ] connect universal optimality with renormalized Coulomb and Riesz crystallization, giving unconditional minimization results in dimensions 8 and 24. Serfaty’s monographs [36 , 37 ] provide systematic accounts of these developments.
For three-dimensional lattice energies, the work of Sarnak–Strömbergsson [35 ] on Epstein zeta functions and heights of flat tori provides a fundamental analytic framework. Bétermin–Petrache [3 ] develop dimension-reduction methods for lattice theta functions. Related Fourier-duality identities for soft pair potentials are obtained by Torquato–Stillinger [38 ] . Classical local optimality results are due to Ennola [14 ] ; Bétermin [2 ] treats local variations for general interaction potentials. Three-dimensional Riesz and Lennard–Jones lattice energies are studied in [4 ] . Ren–Wei [30 ] establish BCC selection for a different long-range interaction.
Our argument follows the local–global method developed in our previous work [25 ] , extending the framework of Sarnak–Strömbergsson [35 ] . Our analysis of theta differences [22 ] includes the hexagonal Yukawa minimum in two dimensions. Sharp lattice transitions for competing Epstein zeta energies and for theta derivatives are studied in [23 , 24 ] , respectively. The present problem requires a comparison in all five shape variables of a three-dimensional Bravais lattice, uniformly in the screening parameter. The unique Yukawa transition is determined by integrating the FCC–BCC theta difference, rather than by a pointwise theta comparison.
The global comparison reduces minimization to FCC and BCC for every screening parameter. We then prove that their energies cross exactly once. Put c = κ 2 4 π c=\frac{\kappa^2}{4\pi} c = 4 π κ 2 . Ewald splitting [15 ] gives a positive functional P ( L , c ) P(L,c) P ( L , c ) , defined in (2.4), with the same lattice differences as E κ E_\kappa E κ . We prove
P ( L , c ) ≥ min { P ( L B , c ) , P ( L F , c ) } ( c ≥ 0 ) , (8) P(L,c)\ge\min\{P(L_{\mathrm{B}},c),P(L_{\mathrm{F}},c)\}\qquad(c\ge0), \tag*{(8)} P ( L , c ) ≥ min { P ( L B , c ) , P ( L F , c )} ( c ≥ 0 ) , (8)
with equality only when L L L is orthogonally equivalent to L B L_{\mathrm{B}} L B or L F L_{\mathrm{F}} L F . A shortest-vector estimate first excludes the cusp. An elementary lattice reduction then places every remaining competitor in a fixed compact rectangle in five coordinates. Near each reference, cubic symmetry reduces the Hessian and third variation to finitely many scalar coefficients. A uniform fourth-variation estimate gives a fixed neighbourhood on which the appropriate local comparison is strict.
Outside these neighbourhoods, convexity of the positive Ewald kernels reduces the shape comparison to the 32 corners of each rectangle. We subtract the reference before expanding in c c c and bound the resulting polynomial by its Bernstein coefficients. This preserves the cancellation near the transition. For large screening, the normalized energy 2 a e κ a E κ 2ae^{\kappa a}E_\kappa 2 a e κa E κ admits a convexity argument that extends a finite comparison to the whole remaining half-line. Finally, Poisson duality and the scalar FCC–BCC theta ordering prove that the two reference energies cross exactly once.
The proof is computer assisted. Appendix A collects the independent numerical comparisons; the complete interval computations and exact partitions accompany the manuscript.
Section 2 establishes the integral representations, duality identities, and Coulomb limit; Section 3 gives the compact reduction. Sections 4 and 5 prove the local and global comparisons. Section 6 determines the transition and completes the main proofs. Appendix A contains the numerical lemmas.
Integral representation and duality
We express the Yukawa energy through positive Gaussian kernels and separate its shape-dependent part. The same representation yields an exact Coulomb limit and compactness uniform in weak screening.
Theorem 2.1. For every κ > 0 \kappa>0 κ > 0 and every lattice L L L with ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 ,
2 E κ ( L ) = ∫ 0 ∞ t − 1 / 2 e − κ 2 / ( 4 π t ) ( Θ ( t , L ) − 1 ) d t . (9) 2E_\kappa(L)=\int_{0}^{\infty}t^{-1/2}e^{-\kappa^2/(4\pi t)}\bigl(\Theta(t,L)-1\bigr)\,\mathrm{d}t. \tag*{(9)} 2 E κ ( L ) = ∫ 0 ∞ t − 1/2 e − κ 2 / ( 4 π t ) ( Θ ( t , L ) − 1 ) d t . (9) The energy is continuous, its sublevel sets are compact, and its minimum is attained.
The Gaussian identity permits Poisson summation, while the singularity of the potential excludes degenerating lattices.
Proof. For A , B > 0 A,B>0 A , B > 0 ,
∫ 0 ∞ t − 1 / 2 e − A t − B / t d t = π A e − 2 A B . \int_{0}^{\infty}t^{-1/2}e^{-At-B/t}\,\mathrm{d}t=\sqrt{\frac{\pi}{A}}e^{-2\sqrt{AB}}. ∫ 0 ∞ t − 1/2 e − A t − B / t d t = A π e − 2 A B .
Take A = π r 2 A=\pi r^2 A = π r 2 and B = κ 2 / ( 4 π ) B=\kappa^2/(4\pi) B = κ 2 / ( 4 π ) , and apply Tonelli’s theorem. In a lattice chart L = g Z 3 L=g\mathbb{Z}^3 L = g Z 3 on which g , g − 1 g,g^{-1} g , g − 1 are bounded, the summands of E κ E_\kappa E κ are bounded by a constant multiple of e − κ ∣ m ∣ / ∣ m ∣ e^{-\kappa\lvert m\rvert}/\lvert m\rvert e − κ ∣ m ∣ / ∣ m ∣ for some ϵ > 0 \epsilon>0 ϵ > 0 . Normal convergence gives continuity. A shortest opposite pair gives
E κ ( L ) ≥ e − κ λ 1 ( L ) λ 1 ( L ) . (10) E_\kappa(L)\ge\frac{e^{-\kappa\lambda_1(L)}}{\lambda_1(L)}. \tag*{(10)} E κ ( L ) ≥ λ 1 ( L ) e − κ λ 1 ( L ) . (10)
Thus λ 1 ( L ) \lambda_1(L) λ 1 ( L ) is bounded away from zero on every sublevel set. Mahler’s compactness theorem [6 ] makes each sublevel relatively compact, and continuity makes it closed. Hence every minimizing sequence has a convergent subsequence whose limit attains the minimum. ▢
Proposition 2.1. Put c = κ 2 / ( 4 π ) c=\kappa^2/(4\pi) c = κ 2 / ( 4 π ) . For lattices L , R L,R L , R with ∣ L ∣ = ∣ R ∣ = 1 |L|=|R|=1 ∣ L ∣ = ∣ R ∣ = 1 ,
2 ( E κ ( L ) − E κ ( R ) ) = ∫ 1 ∞ t − 1 / 2 e − c / t ( Θ ( t , L ) − Θ ( t , R ) ) d t + ∫ 1 ∞ e − c t ( Θ ( t , L ∗ ) − Θ ( t , R ∗ ) ) d t . (11) 2\left(E_\kappa(L)-E_\kappa(R)\right)=\int_1^\infty t^{-1/2}e^{-c/t}\left(\Theta(t,L)-\Theta(t,R)\right)\,\mathrm{d}t+\int_1^\infty e^{-ct}\left(\Theta(t,L^*)-\Theta(t,R^*)\right)\,\mathrm{d}t. \tag*{(11)} 2 ( E κ ( L ) − E κ ( R ) ) = ∫ 1 ∞ t − 1/2 e − c / t ( Θ ( t , L ) − Θ ( t , R ) ) d t + ∫ 1 ∞ e − c t ( Θ ( t , L ∗ ) − Θ ( t , R ∗ ) ) d t . (11) Both integrals converge absolutely.
Proof. Split (2.1) at one. On ( 0 , 1 ) (0,1) ( 0 , 1 ) use Θ ( t , L ) = t − 3 / 2 Θ ( 1 / t , L ∗ ) \Theta(t,L)=t^{-3/2}\Theta(1/t,L^*) Θ ( t , L ) = t − 3/2 Θ ( 1/ t , L ∗ ) and substitute t ↦ 1 / t t\mapsto1/t t ↦ 1/ t . The zero terms cancel in the difference, and Gaussian decay proves absolute convergence. □
For q > 0 q>0 q > 0 and c ≥ 0 c\geq0 c ≥ 0 , define
I ( q , c ) = ∫ 1 ∞ t − 1 / 2 e − π q t − c / t d t , J ( q , c ) = ∫ 1 ∞ e − ( π q + c ) t d t = e − π q − c π q + c . I(q,c)=\int_1^\infty t^{-1/2}e^{-\pi qt-c/t}\,\mathrm{d}t,\qquad J(q,c)=\int_1^\infty e^{-(\pi q+c)t}\,\mathrm{d}t=\frac{e^{-\pi q-c}}{\pi q+c}. I ( q , c ) = ∫ 1 ∞ t − 1/2 e − π q t − c / t d t , J ( q , c ) = ∫ 1 ∞ e − ( π q + c ) t d t = π q + c e − π q − c .
For a determinant-one Gram matrix Y = g T g Y=g^{T}g Y = g T g , write Y [ m ] = m T Y m Y[m]=m^{T}Ym Y [ m ] = m T Y m and put
P ( Y , c ) = 1 2 ∑ m ∈ Z 3 ∖ { 0 } { I ( Y [ m ] , c ) + J ( Y − 1 [ m ] , c ) } . (12) P(Y,c)=\frac{1}{2}\sum_{m\in\mathbb{Z}^{3}\setminus\{0\}}\left\{I(Y[m],c)+J(Y^{-1}[m],c)\right\}. \tag*{(12)} P ( Y , c ) = 2 1 m ∈ Z 3 ∖ { 0 } ∑ { I ( Y [ m ] , c ) + J ( Y − 1 [ m ] , c ) } . (12)
We also write P ( L , c ) P(L,c) P ( L , c ) for this basis-independent functional. Every summand is positive, and both kernels are completely monotone in q q q . Their sums and any fixed number of shape derivatives converge normally on compact sets of nonsingular lattices and bounded screening intervals. These assertions follow by differentiation under the positive integrals and Gaussian domination of the resulting polynomial factors.
Let Z ( L , s ) Z(L,s) Z ( L , s ) denote the meromorphic continuation of ∑ v ∈ L ∖ { 0 } ∣ v ∣ − 2 s \sum_{v\in L\setminus\{0\}}|v|^{-2s} ∑ v ∈ L ∖ { 0 } ∣ v ∣ − 2 s from ℜ s > 3 / 2 \Re s>3/2 ℜ s > 3/2 , and define
W ( L ) = 1 2 Z ( L , 1 / 2 ) = 1 2 π Z ( L ∗ , 1 ) , F κ ( L ) = E κ ( L ) − 2 π κ 2 − κ 2 , F 0 = W . (13) W(L)=\frac{1}{2}Z(L,1/2)=\frac{1}{2\pi}Z(L^*,1),\qquad F_\kappa(L)=E_\kappa(L)-\frac{2\pi}{\kappa^2}-\frac{\kappa}{2},\qquad F_0=W. \tag*{(13)} W ( L ) = 2 1 Z ( L , 1/2 ) = 2 π 1 Z ( L ∗ , 1 ) , F κ ( L ) = E κ ( L ) − κ 2 2 π − 2 κ , F 0 = W . (13)
The identity for W W W is the Epstein functional equation; the displayed zeta values are continued values, not divergent lattice sums.
Proposition 2.2. For κ > 0 \kappa>0 κ > 0 and c = κ 2 / ( 4 π ) c=\kappa^2/(4\pi) c = κ 2 / ( 4 π ) ,
2 F κ ( L ) = 2 P ( L , c ) + e − c − 1 c − 2 e − c − κ erf ( c ) . (14) 2F_\kappa(L)=2P(L,c)+\frac{e^{-c}-1}{c}-2e^{-c}-\kappa\operatorname{erf}(\sqrt{c}). \tag*{(14)} 2 F κ ( L ) = 2 P ( L , c ) + c e − c − 1 − 2 e − c − κ erf ( c ) . (14) The last three terms tend to − 3 -3 − 3 as κ ↓ 0 \kappa\downarrow0 κ ↓ 0 , and
P ( L , c ) − P ( R , c ) = F κ ( L ) − F κ ( R ) , W ( L ) = P ( L , 0 ) − 3 2 . (15) P(L,c)-P(R,c)=F_\kappa(L)-F_\kappa(R),\qquad W(L)=P(L,0)-\frac{3}{2}. \tag*{(15)} P ( L , c ) − P ( R , c ) = F κ ( L ) − F κ ( R ) , W ( L ) = P ( L , 0 ) − 2 3 . (15) Thus P ( ⋅ , c ) P(\cdot,c) P ( ⋅ , c ) and E κ E_\kappa E κ have the same minimizers when κ > 0 \kappa>0 κ > 0 .
Proof. Splitting (2.1) and using Poisson summation gives
2 E κ ( L ) = 2 P ( L , c ) + e − c c − ∫ 0 1 t − 1 / 2 e − c / t d t . 2E_\kappa(L)=2P(L,c)+\frac{e^{-c}}{c}-\int_0^1 t^{-1/2}e^{-c/t}\,\mathrm{d}t. 2 E κ ( L ) = 2 P ( L , c ) + c e − c − ∫ 0 1 t − 1/2 e − c / t d t . With u = c / t u=\sqrt{c/t} u = c / t , integration by parts gives
∫ 0 1 t − 1 / 2 e − c / t d t = 2 c ∫ c ∞ e − u 2 u 2 d u = 2 e − c − κ erfc ( c ) . \int_0^1 t^{-1/2}e^{-c/t}\,\mathrm{d}t=2\sqrt{c}\int_{\sqrt{c}}^\infty\frac{e^{-u^2}}{u^2}\,\mathrm{d}u=2e^{-c}-\kappa\operatorname{erfc}(\sqrt{c}). ∫ 0 1 t − 1/2 e − c / t d t = 2 c ∫ c ∞ u 2 e − u 2 d u = 2 e − c − κ erfc ( c ) . Since 2 F κ = 2 E κ − 1 / c − κ 2F_\kappa=2E_\kappa-1/c-\kappa 2 F κ = 2 E κ − 1/ c − κ and erfc ( c ) = 1 − erf ( c ) \operatorname{erfc}(\sqrt{c})=1-\operatorname{erf}(\sqrt{c}) erfc ( c ) = 1 − erf ( c ) , subtraction gives (2.6). The regularized Mellin formula at s = 1 / 2 s=1/2 s = 1/2 is
Z ( L , 1 / 2 ) = ∫ 1 ∞ t − 1 / 2 ( Θ ( t , L ) − 1 ) d t + ∫ 1 ∞ ( Θ ( t , L ∗ ) − 1 ) d t − 3. Z(L,1/2)=\int_1^\infty t^{-1/2}\left(\Theta(t,L)-1\right)\,\mathrm{d}t+\int_1^\infty\left(\Theta(t,L^*)-1\right)\,\mathrm{d}t-3. Z ( L , 1/2 ) = ∫ 1 ∞ t − 1/2 ( Θ ( t , L ) − 1 ) d t + ∫ 1 ∞ ( Θ ( t , L ∗ ) − 1 ) d t − 3. It proves the assertion at zero screening and the difference identity. □
The direct kernel also has the elementary form
I ( q , c ) = e − κ q erfc ( π q − c ) + e κ q erfc ( π q + c ) 2 q , κ = 4 π c . I(q,c)=\frac{e^{-\kappa\sqrt{q}}\operatorname{erfc}(\sqrt{\pi q}-\sqrt{c})+e^{\kappa\sqrt{q}}\operatorname{erfc}(\sqrt{\pi q}+\sqrt{c})}{2\sqrt{q}},\qquad\kappa=\sqrt{4\pi c}. I ( q , c ) = 2 q e − κ q erfc ( π q − c ) + e κ q erfc ( π q + c ) , κ = 4 π c .
A shortest opposite pair contributes I ( λ 1 ( L ) 2 , c ) I(\lambda_1(L)^2,c) I ( λ 1 ( L ) 2 , c ) to P ( L , c ) P(L,c) P ( L , c ) . Since e − c / t ≥ e − c e^{-c/t}\ge e^{-c} e − c / t ≥ e − c for t ≥ 1 t\ge1 t ≥ 1 ,
P ( L , c ) ≥ I ( λ 1 ( L ) 2 , c ) ≥ e − c I ( λ 1 ( L ) 2 , 0 ) = e − c erfc ( π λ 1 ( L ) ) λ 1 ( L ) . P(L,c)\ge I(\lambda_1(L)^2,c)\ge e^{-c}I(\lambda_1(L)^2,0)=\frac{e^{-c}\operatorname{erfc}(\sqrt{\pi}\lambda_1(L))}{\lambda_1(L)}. P ( L , c ) ≥ I ( λ 1 ( L ) 2 , c ) ≥ e − c I ( λ 1 ( L ) 2 , 0 ) = λ 1 ( L ) e − c erfc ( π λ 1 ( L )) .
Substitution in (14) gives
F κ ( L ) ≥ e − c erfc ( π λ 1 ( L ) ) λ 1 ( L ) + e − c − 1 2 c − e − c − κ 2 erf ( c ) . (16) F_\kappa(L)\ge\frac{e^{-c}\operatorname{erfc}(\sqrt{\pi}\lambda_1(L))}{\lambda_1(L)}+\frac{e^{-c}-1}{2c}-e^{-c}-\frac{\kappa}{2}\operatorname{erf}(\sqrt{c}). \tag*{(16)} F κ ( L ) ≥ λ 1 ( L ) e − c erfc ( π λ 1 ( L )) + 2 c e − c − 1 − e − c − 2 κ erf ( c ) . (16)
The right side extends continuously to κ = 0 \kappa=0 κ = 0 and tends to infinity as λ 1 ( L ) → 0 \lambda_1(L)\to0 λ 1 ( L ) → 0 , uniformly on bounded screening intervals. This uniform exclusion of degenerating lattices permits passage from the Coulomb limit of each energy to the limit of the global minima.
Proposition 2.3 (The Coulomb limit). For every lattice L L L with ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 and every κ > 0 \kappa>0 κ > 0 ,
F κ ( L ) = W ( L ) − κ 2 8 π 3 ∑ w ∈ L ∗ ∖ { 0 } 1 ∣ w ∣ 2 ( ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ) . (17) F_\kappa(L)=W(L)-\frac{\kappa^2}{8\pi^3}\sum_{w\in L^*\setminus\{0\}}\frac{1}{|w|^2\left(|w|^2+\kappa^2/(4\pi^2)\right)}. \tag*{(17)} F κ ( L ) = W ( L ) − 8 π 3 κ 2 w ∈ L ∗ ∖ { 0 } ∑ ∣ w ∣ 2 ( ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ) 1 . (17) The correction converges absolutely, and
0 ≤ F κ ( L ) − W ( L ) + κ 2 8 π 3 Z ( L ∗ , 2 ) ≤ κ 4 32 π 5 Z ( L ∗ , 3 ) . (18) 0\le F_\kappa(L)-W(L)+\frac{\kappa^2}{8\pi^3}Z(L^*,2)\le\frac{\kappa^4}{32\pi^5}Z(L^*,3). \tag*{(18)} 0 ≤ F κ ( L ) − W ( L ) + 8 π 3 κ 2 Z ( L ∗ , 2 ) ≤ 32 π 5 κ 4 Z ( L ∗ , 3 ) . (18) In particular F κ → W F_\kappa\to W F κ → W locally uniformly as κ ↓ 0 \kappa\downarrow0 κ ↓ 0 . The functional W W W attains its minimum, and
lim κ ↓ 0 { min ∣ L ∣ = 1 E κ ( L ) − 2 π κ 2 − κ 2 } = min ∣ L ∣ = 1 W ( L ) . (19) \lim_{\kappa\downarrow0}\left\{\min_{|L|=1}E_\kappa(L)-\frac{2\pi}{\kappa^2}-\frac{\kappa}{2}\right\}=\min_{|L|=1}W(L). \tag*{(19)} κ ↓ 0 lim { ∣ L ∣ = 1 min E κ ( L ) − κ 2 2 π − 2 κ } = ∣ L ∣ = 1 min W ( L ) . (19) Every sequence of Yukawa minimizers with κ ↓ 0 \kappa\downarrow0 κ ↓ 0 has a convergent subsequence, and every such limit minimizes W W W .
Proof. Put Q L ( t ) = Θ ( t , L ) − 1 − t − 3 / 2 Q_L(t)=\Theta(t,L)-1-t^{-3/2} Q L ( t ) = Θ ( t , L ) − 1 − t − 3/2 . Poisson summation shows that Q L ( t ) = − 1 + O ( t − 3 / 2 e − d / t ) Q_L(t)=-1+O(t^{-3/2}e^{-d/t}) Q L ( t ) = − 1 + O ( t − 3/2 e − d / t ) at zero and Q L ( t ) = − t − 3 / 2 + O ( e − d t ) Q_L(t)=-t^{-3/2}+O(e^{-dt}) Q L ( t ) = − t − 3/2 + O ( e − d t ) at infinity, for some d > 0 d>0 d > 0 . Hence t − 1 / 2 Q L ( t ) t^{-1/2}Q_L(t) t − 1/2 Q L ( t ) is integrable, and the regularized Mellin formula reads Z ( L , 1 / 2 ) = ∫ 0 ∞ t − 1 / 2 Q L ( t ) d t Z(L,1/2)=\int_0^\infty t^{-1/2}Q_L(t)\,dt Z ( L , 1/2 ) = ∫ 0 ∞ t − 1/2 Q L ( t ) d t . Since ∫ 0 ∞ t − 2 e − c / t d t = 1 / c \int_0^\infty t^{-2}e^{-c/t}\,dt=1/c ∫ 0 ∞ t − 2 e − c / t d t = 1/ c , we also have
2 E κ ( L ) − 1 c = ∫ 0 ∞ t − 1 / 2 e − c / t Q L ( t ) d t . (20) 2E_\kappa(L)-\frac{1}{c}=\int_0^\infty t^{-1/2}e^{-c/t}Q_L(t)\,dt. \tag*{(20)} 2 E κ ( L ) − c 1 = ∫ 0 ∞ t − 1/2 e − c / t Q L ( t ) d t . (20)
Poisson summation gives the explicit identity
u − 3 / 2 Q L ( 1 / u ) = Θ ( u , L ∗ ) − 1 − u − 3 / 2 . u^{-3/2}Q_L(1/u)=\Theta(u,L^*)-1-u^{-3/2}. u − 3/2 Q L ( 1/ u ) = Θ ( u , L ∗ ) − 1 − u − 3/2 .
Subtract the Mellin identity from (20) and put t = 1 / u t=1/u t = 1/ u . The weighted measure transforms as t − 1 / 2 d t = − u − 3 / 2 d u t^{-1/2}dt=-u^{-3/2}du t − 1/2 d t = − u − 3/2 d u , so
2 E κ ( L ) − 1 c − Z ( L , 1 / 2 ) = ∫ 0 ∞ ( e − c u − 1 ) ( Θ ( u , L ∗ ) − 1 − u − 3 / 2 ) d u = ∑ w ∈ L ∗ ∖ { 0 } { 1 π ∣ w ∣ 2 + c − 1 π ∣ w ∣ 2 } + κ . \begin{aligned}
2E_\kappa(L)-\frac{1}{c}-Z(L,1/2)=\int_0^\infty(e^{-cu}-1)(\Theta(u,L^*)-1-u^{-3/2})\,du \\
&=\sum_{w\in L^*\setminus\{0\}}\left\{\frac{1}{\pi|w|^2+c}-\frac{1}{\pi|w|^2}\right\}+\kappa.
\end{aligned} 2 E κ ( L ) − c 1 − Z ( L , 1/2 ) = ∫ 0 ∞ ( e − c u − 1 ) ( Θ ( u , L ∗ ) − 1 − u − 3/2 ) d u = w ∈ L ∗ ∖ { 0 } ∑ { π ∣ w ∣ 2 + c 1 − π ∣ w ∣ 2 1 } + κ .
After separating u − 3 / 2 u^{-3/2} u − 3/2 , the theta integral is absolutely convergent: its integrand is O ( u − 1 / 2 ) O(u^{-1/2}) O ( u − 1/2 ) at zero and decays exponentially at infinity. Summation and integration therefore commute. The remaining term is ∫ 0 ∞ ( 1 − e − c u ) u − 3 / 2 d u = 2 π c = κ \int_0^\infty(1-e^{-cu})u^{-3/2}\,du=2\sqrt{\pi c}=\kappa ∫ 0 ∞ ( 1 − e − c u ) u − 3/2 d u = 2 π c = κ . Thus F κ ( L ) − W ( L ) F_\kappa(L)-W(L) F κ ( L ) − W ( L ) is one half of the reciprocal sum above. The identity
1 2 ( 1 π ∣ w ∣ 2 + c − 1 π ∣ w ∣ 2 ) = − c 2 π 2 ∣ w ∣ 2 ( ∣ w ∣ 2 + c / π ) , \frac{1}{2}\left(\frac{1}{\pi|w|^2+c}-\frac{1}{\pi|w|^2}\right)=-\frac{c}{2\pi^2|w|^2(|w|^2+c/\pi)}, 2 1 ( π ∣ w ∣ 2 + c 1 − π ∣ w ∣ 2 1 ) = − 2 π 2 ∣ w ∣ 2 ( ∣ w ∣ 2 + c / π ) c ,
together with c / π = κ 2 / ( 4 π 2 ) c/\pi=\kappa^2/(4\pi^2) c / π = κ 2 / ( 4 π 2 ) and c / ( 2 π 2 ) = κ 2 / ( 8 π 3 ) c/(2\pi^2)=\kappa^2/(8\pi^3) c / ( 2 π 2 ) = κ 2 / ( 8 π 3 ) , proves (2.9).
The identity
1 q ( q + x ) = 1 q 2 − x q 2 ( q + x ) ( q , x > 0 ) \frac{1}{q(q+x)}=\frac{1}{q^2}-\frac{x}{q^2(q+x)}\qquad(q,x>0) q ( q + x ) 1 = q 2 1 − q 2 ( q + x ) x ( q , x > 0 )
gives
F κ ( L ) − W ( L ) + κ 2 8 π 3 Z ( L ∗ , 2 ) = κ 4 32 π 5 ∑ w ∈ L ∗ ∖ { 0 } 1 ∣ w ∣ 4 ( ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ) . (21) F_\kappa(L)-W(L)+\frac{\kappa^2}{8\pi^3}Z(L^*,2)=\frac{\kappa^4}{32\pi^5}\sum_{w\in L^*\setminus\{0\}}\frac{1}{|w|^4\left(|w|^2+\kappa^2/(4\pi^2)\right)}. \tag*{(21)} F κ ( L ) − W ( L ) + 8 π 3 κ 2 Z ( L ∗ , 2 ) = 32 π 5 κ 4 w ∈ L ∗ ∖ { 0 } ∑ ∣ w ∣ 4 ( ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ) 1 . (21)
Since ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ≥ ∣ w ∣ 2 |w|^2+\kappa^2/(4\pi^2)\ge|w|^2 ∣ w ∣ 2 + κ 2 / ( 4 π 2 ) ≥ ∣ w ∣ 2 , this proves (2.10). Normal convergence bounds Z ( L ∗ , 2 ) Z(L^*,2) Z ( L ∗ , 2 ) and Z ( L ∗ , 3 ) Z(L^*,3) Z ( L ∗ , 3 ) on compact sets of lattice classes, proving locally uniform convergence.
For any fixed κ 0 > 0 \kappa_0>0 κ 0 > 0 , the right side of (2.8) tends to infinity as λ 1 ( L ) → 0 \lambda_1(L)\to0 λ 1 ( L ) → 0 , uniformly for 0 ≤ κ ≤ κ 0 0\le\kappa\le\kappa_0 0 ≤ κ ≤ κ 0 : the lattice-independent terms in (2.8) extend continuously to c = 0 c=0 c = 0 , and e − c e^{-c} e − c is bounded below by a positive constant. Thus W W W has compact sublevels by Mahler’s theorem and attains a minimum. Moreover, a minimizer L κ L_\kappa L κ of F κ F_\kappa F κ satisfies
F κ ( L κ ) ≤ F κ ( L B ) ≤ W ( L B ) , F_\kappa(L_\kappa)\le F_\kappa(L_B)\le W(L_B), F κ ( L κ ) ≤ F κ ( L B ) ≤ W ( L B ) ,
where the last inequality follows from (2.9). Consequently all these minimizers, for 0 < κ ≤ κ 0 0<\kappa\le\kappa_0 0 < κ ≤ κ 0 , lie in one compact set.
If κ j ↓ 0 \kappa_j\downarrow0 κ j ↓ 0 , choose a convergent subsequence of minimizers with limit L 0 L_0 L 0 . Locally uniform convergence gives F κ j ( L κ j ) → W ( L 0 ) F_{\kappa_j}(L_{\kappa_j})\to W(L_0) F κ j ( L κ j ) → W ( L 0 ) . For any fixed competitor R R R , F κ j ( L κ j ) ≤ F κ j ( R ) F_{\kappa_j}(L_{\kappa_j})\le F_{\kappa_j}(R) F κ j ( L κ j ) ≤ F κ j ( R ) ; passing to the limit shows that L 0 L_0 L 0 minimizes W W W . Every sequence admits such a subsequence, and its limiting minimum value is always min W \min W min W . This proves (2.11) and the assertion about minimizers. □
An elementary compact reduction
We reduce the global problem to an explicit compact set of lattice shapes. The reduction is uniform in the screening parameter, including the continued Coulomb endpoint. Throughout this section c ≥ 0 c\ge0 c ≥ 0 denotes the screening variable in the positive Ewald functional P ( L , c ) P(L,c) P ( L , c ) .
Proposition 3.1 (Compact reduction). For every c ≥ 0 c\ge0 c ≥ 0 , the functional P ( ⋅ , c ) P(\cdot,c) P ( ⋅ , c ) attains its minimum among lattices L L L with ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 . Every global minimizer satisfies λ 1 ( L ) > 4 / 5 \lambda_1(L)>4/5 λ 1 ( L ) > 4/5 and has a Gram matrix of the form
Y = N T D N , N = ( 1 x 1 x 2 0 1 x 3 0 0 1 ) , D = diag ( d 1 , d 2 , d 3 ) , (22) Y=N^{T}DN,\qquad N=\begin{pmatrix}1&x_1&x_2\\0&1&x_3\\0&0&1\end{pmatrix},\qquad D=\operatorname{diag}(d_1,d_2,d_3), \tag*{(22)} Y = N T D N , N = 1 0 0 x 1 1 0 x 2 x 3 1 , D = diag ( d 1 , d 2 , d 3 ) , (22) where, on putting y i = e ℓ i y_i=e^{\ell_i} y i = e ℓ i ,
d 1 = y 1 2 / 3 y 2 1 / 3 , d 2 = y 1 − 1 / 3 y 2 1 / 3 , d 3 = y 1 − 1 / 3 y 2 − 2 / 3 , d_1=y_1^{2/3}y_2^{1/3},\qquad d_2=y_1^{-1/3}y_2^{1/3},\qquad d_3=y_1^{-1/3}y_2^{-2/3}, d 1 = y 1 2/3 y 2 1/3 , d 2 = y 1 − 1/3 y 2 1/3 , d 3 = y 1 − 1/3 y 2 − 2/3 , and
( y 1 , y 2 , x 1 , x 2 , x 3 ) ∈ [ 11 25 , 4 3 ] × [ 147 1000 , 4 3 ] × [ 0 , 1 2 ] × [ 0 , 1 ] × [ 0 , 1 2 ] . (23) (y_1,y_2,x_1,x_2,x_3)\in\left[\frac{11}{25},\frac{4}{3}\right]\times\left[\frac{147}{1000},\frac{4}{3}\right]\times\left[0,\frac{1}{2}\right]\times[0,1]\times\left[0,\frac{1}{2}\right]. \tag*{(23)} ( y 1 , y 2 , x 1 , x 2 , x 3 ) ∈ [ 25 11 , 3 4 ] × [ 1000 147 , 3 4 ] × [ 0 , 2 1 ] × [ 0 , 1 ] × [ 0 , 2 1 ] . (23) In the basis constructed below, d 1 = λ 1 ( L ) 2 d_1=\lambda_1(L)^2 d 1 = λ 1 ( L ) 2 .
We first choose a shortest-vector basis and reduce its two-dimensional orthogonal projection. A comparison of the Ewald kernel ratios then excludes the cusp for every c ≥ 0 c\ge0 c ≥ 0 . The resulting rectangle retains all five shape parameters; it does not restrict the problem to a symmetric family of lattices. We also record the separate convexity which will control every shape rectangle by its corners.
Lemma 3.1 (Elementary covering reduction). Every lattice L L L with ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 has a basis with Gram matrix (3.1) satisfying 0 ≤ x 1 , x 3 ≤ 1 / 2 0 \le x_1,x_3 \le1/2 0 ≤ x 1 , x 3 ≤ 1/2 , 0 ≤ x 2 ≤ 1 0 \le x_2 \le1 0 ≤ x 2 ≤ 1 , and
1 ≤ ( 1 + x 1 − x 2 ) 2 + y 1 − 1 ( ( 1 − x 3 ) 2 + y 2 − 1 ) , 1 ≤ ( x 1 − x 2 ) 2 + y 1 − 1 ( ( 1 − x 3 ) 2 + y 2 − 1 ) , 1 ≤ x 1 2 + y 1 − 1 , 1 ≤ x 2 2 + y 1 − 1 ( x 3 2 + y 2 − 1 ) , 1 ≤ ( x 2 − 1 ) 2 + y 1 − 1 ( x 3 2 + y 2 − 1 ) , 1 ≤ x 3 2 + y 2 − 1 . (24) \begin{aligned}
1 &\le(1+x_1-x_2)^2+y_1^{-1}\left((1-x_3)^2+y_2^{-1}\right),\\
1 &\le(x_1-x_2)^2+y_1^{-1}\left((1-x_3)^2+y_2^{-1}\right),\\
1 &\le x_1^2+y_1^{-1},\\
1 &\le x_2^2+y_1^{-1}\left(x_3^2+y_2^{-1}\right),\\
1 &\le(x_2-1)^2+y_1^{-1}\left(x_3^2+y_2^{-1}\right),\\
1 &\le x_3^2+y_2^{-1}. \tag*{(24)}
\end{aligned} 1 1 1 1 1 1 ≤ ( 1 + x 1 − x 2 ) 2 + y 1 − 1 ( ( 1 − x 3 ) 2 + y 2 − 1 ) , ≤ ( x 1 − x 2 ) 2 + y 1 − 1 ( ( 1 − x 3 ) 2 + y 2 − 1 ) , ≤ x 1 2 + y 1 − 1 , ≤ x 2 2 + y 1 − 1 ( x 3 2 + y 2 − 1 ) , ≤ ( x 2 − 1 ) 2 + y 1 − 1 ( x 3 2 + y 2 − 1 ) , ≤ x 3 2 + y 2 − 1 . (24) Its first basis vector is shortest, so d 1 = λ 1 ( L ) 2 d_1=\lambda_1(L)^2 d 1 = λ 1 ( L ) 2 .
Proof. Choose a shortest nonzero vector v 1 ∈ L v_1 \in L v 1 ∈ L . It is primitive and hence extends to an integral basis. If p p p is orthogonal projection onto v 1 ⊥ v_1^\perp v 1 ⊥ , then p ( L ) p(L) p ( L ) is a rank-two lattice and ker ( p ∣ L ) = Z v 1 \ker(p|_L)=\mathbb{Z}v_1 ker ( p ∣ L ) = Z v 1 . Indeed, projection of the other two vectors of any basis beginning with v 1 v_1 v 1 gives two linearly independent vectors generating p ( L ) p(L) p ( L ) . Choose a shortest nonzero vector w 2 ∈ p ( L ) w_2 \in p(L) w 2 ∈ p ( L ) . It is primitive, so it extends to a basis ( w 2 , w 3 ) (w_2,w_3) ( w 2 , w 3 ) of p ( L ) p(L) p ( L ) . Choose lifts v 2 , v 3 ∈ L v_2,v_3 \in L v 2 , v 3 ∈ L . Then ( v 1 , v 2 , v 3 ) (v_1,v_2,v_3) ( v 1 , v 2 , v 3 ) is a basis of L L L : projecting an arbitrary vector first expresses its image in terms of w 2 , w 3 w_2,w_3 w 2 , w 3 , and the remaining vector belongs to Z v 1 \mathbb{Z}v_1 Z v 1 .
Gram–Schmidt gives an orthonormal frame ( e 1 , e 2 , e 3 ) (e_1,e_2,e_3) ( e 1 , e 2 , e 3 ) and positive numbers d i d_i d i such that
v 1 = d 1 e 1 , v 2 = x 1 d 1 e 1 + d 2 e 2 , v 3 = x 2 d 1 e 1 + x 3 d 2 e 2 + d 3 e 3 . v_1=\sqrt{d_1}e_1,\qquad v_2=x_1\sqrt{d_1}e_1+\sqrt{d_2}e_2,\qquad v_3=x_2\sqrt{d_1}e_1+x_3\sqrt{d_2}e_2+\sqrt{d_3}e_3. v 1 = d 1 e 1 , v 2 = x 1 d 1 e 1 + d 2 e 2 , v 3 = x 2 d 1 e 1 + x 3 d 2 e 2 + d 3 e 3 .
Add an integral multiple of v 1 v_1 v 1 to v 2 v_2 v 2 to arrange ∣ x 1 ∣ ≤ 1 / 2 |x_1| \le1/2 ∣ x 1 ∣ ≤ 1/2 . Add an integral multiple of v 2 v_2 v 2 to v 3 v_3 v 3 to arrange ∣ x 3 ∣ ≤ 1 / 2 |x_3| \le1/2 ∣ x 3 ∣ ≤ 1/2 , and then an integral multiple of v 1 v_1 v 1 to arrange ∣ x 2 ∣ ≤ 1 / 2 |x_2| \le1/2 ∣ x 2 ∣ ≤ 1/2 . These changes preserve the basis and the shortest projected vector p ( v 2 ) p(v_2) p ( v 2 ) . Changing the signs of the basis vectors sends the shears to ( s 1 s 2 x 1 , s 1 s 3 x 2 , s 2 s 3 x 3 ) (s_1s_2x_1,s_1s_3x_2,s_2s_3x_3) ( s 1 s 2 x 1 , s 1 s 3 x 2 , s 2 s 3 x 3 ) , with s i ∈ { ± 1 } s_i \in\{\pm1\} s i ∈ { ± 1 } . Take s 2 = 1 s_2=1 s 2 = 1 and choose s 1 , s 3 s_1,s_3 s 1 , s 3 to make x 1 , x 3 x_1,x_3 x 1 , x 3 nonnegative. If the resulting x 2 x_2 x 2 is negative, replace v 3 v_3 v 3 by v 3 + v 1 v_3+v_1 v 3 + v 1 . Then 0 ≤ x 2 ≤ 1 0 \le x_2 \le1 0 ≤ x 2 ≤ 1 , with x 1 , x 3 x_1,x_3 x 1 , x 3 unchanged.
The five nonzero vectors v 1 + v 2 − v 3 v_1+v_2-v_3 v 1 + v 2 − v 3 , v 2 − v 3 v_2-v_3 v 2 − v 3 , v 2 v_2 v 2 , v 3 v_3 v 3 , and v 3 − v 1 v_3-v_1 v 3 − v 1 have squared length at least d 1 d_1 d 1 . Computing these lengths and dividing by d 1 d_1 d 1 gives the first five inequalities in (3.3). The shortest-vector property of p ( v 2 ) p(v_2) p ( v 2 ) gives x 3 2 d 2 + d 3 ≥ d 2 x_3^2d_2+d_3 \ge d_2 x 3 2 d 2 + d 3 ≥ d 2 , which is the sixth inequality. Finally, d 1 d 2 d 3 = ∣ L ∣ 2 = 1 d_1d_2d_3=|L|^2=1 d 1 d 2 d 3 = ∣ L ∣ 2 = 1 ; the displayed formulas for d i d_i d i follow from y 1 = d 1 / d 2 y_1=d_1/d_2 y 1 = d 1 / d 2 and y 2 = d 2 / d 3 y_2=d_2/d_3 y 2 = d 2 / d 3 . □
Lemma 3.2 (Uniform cusp exclusion). For every c ≥ 0 c \ge0 c ≥ 0 , if λ 1 ( L ) ≤ 4 / 5 \lambda_1(L) \le4/5 λ 1 ( L ) ≤ 4/5 , then
P ( L , c ) > P ( L F , c ) . P(L,c)>P(L_F,c). P ( L , c ) > P ( L F , c ) . Proof. For q ≥ 16 / 25 q \ge16/25 q ≥ 16/25 , we claim that
I ( q , c ) I ( 16 / 25 , c ) ≤ I ( q , 0 ) I ( 16 / 25 , 0 ) , J ( q , c ) I ( 16 / 25 , c ) ≤ J ( q , 0 ) I ( 16 / 25 , 0 ) . (25) \frac{I(q,c)}{I(16/25,c)}\le\frac{I(q,0)}{I(16/25,0)},\qquad
\frac{J(q,c)}{I(16/25,c)}\le\frac{J(q,0)}{I(16/25,0)}. \tag*{(25)} I ( 16/25 , c ) I ( q , c ) ≤ I ( 16/25 , 0 ) I ( q , 0 ) , I ( 16/25 , c ) J ( q , c ) ≤ I ( 16/25 , 0 ) J ( q , 0 ) . (25) The functions t ↦ e − c / t t\mapsto e^{-c/t} t ↦ e − c / t and t ↦ e − π ( q − 16 / 25 ) t t\mapsto e^{-\pi(q-16/25)t} t ↦ e − π ( q − 16/25 ) t are respectively increasing and decreasing on [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) . Hence, for t , u ≥ 1 t,u \ge1 t , u ≥ 1 ,
( e − c / t − e − c / u ) ( e − π ( q − 16 / 25 ) t − e − π ( q − 16 / 25 ) u ) ≤ 0. \left(e^{-c/t}-e^{-c/u}\right)\left(e^{-\pi(q-16/25)t}-e^{-\pi(q-16/25)u}\right)\le0. ( e − c / t − e − c / u ) ( e − π ( q − 16/25 ) t − e − π ( q − 16/25 ) u ) ≤ 0. Multiply by ( t u ) − 1 / 2 e − ( 16 π / 25 ) ( t + u ) (tu)^{-1/2}e^{-(16\pi/25)(t+u)} ( t u ) − 1/2 e − ( 16 π /25 ) ( t + u ) , integrate in t , u t,u t , u , and expand. The resulting inequality is
I ( q , c ) I ( 16 / 25 , 0 ) ≤ I ( q , 0 ) I ( 16 / 25 , c ) , I(q,c)I(16/25,0)\le I(q,0)I(16/25,c), I ( q , c ) I ( 16/25 , 0 ) ≤ I ( q , 0 ) I ( 16/25 , c ) , which proves the first ratio bound. For the second, use J ( q , c ) ≤ e − c J ( q , 0 ) J(q,c)\le e^{-c}J(q,0) J ( q , c ) ≤ e − c J ( q , 0 ) and I ( 16 / 25 , c ) ≥ e − c I ( 16 / 25 , 0 ) I(16/25,c)\ge e^{-c}I(16/25,0) I ( 16/25 , c ) ≥ e − c I ( 16/25 , 0 ) . This argument proves the second ratio bound for every q > 0 q>0 q > 0 .
All nonzero FCC vectors have squared length at least 2 1 / 3 > 16 / 25 2^{1/3}>16/25 2 1/3 > 16/25 . Apply the first ratio bound to the direct FCC sum and the second, which holds for every q > 0 q>0 q > 0 , to the reciprocal BCC sum. Summing gives
P ( L F , c ) I ( 16 / 25 , c ) ≤ P ( L F , 0 ) I ( 16 / 25 , 0 ) < 1. \frac{P(L_F,c)}{I(16/25,c)}\le\frac{P(L_F,0)}{I(16/25,0)}<1. I ( 16/25 , c ) P ( L F , c ) ≤ I ( 16/25 , 0 ) P ( L F , 0 ) < 1.
The strict scalar inequality is proved in Lemma A.1 of the appendix. A shortest opposite pair in L L L contributes I ( λ 1 ( L ) 2 , c ) I(\lambda_1(L)^2,c) I ( λ 1 ( L ) 2 , c ) to P ( L , c ) P(L,c) P ( L , c ) . Since I ( q , c ) I(q,c) I ( q , c ) decreases in q q q ,
P ( L , c ) ≥ I ( λ 1 ( L ) 2 , c ) ≥ I ( 16 / 25 , c ) > P ( L F , c ) , P(L,c) \ge I(\lambda_1(L)^2,c) \ge I(16/25,c) > P(L_F,c), P ( L , c ) ≥ I ( λ 1 ( L ) 2 , c ) ≥ I ( 16/25 , c ) > P ( L F , c ) ,
as required. □
Proof of Proposition 3.1. It suffices to minimize over lattices with energy at most P ( L F , c ) P(L_F,c) P ( L F , c ) . Lemma 3.2 implies that every such lattice has λ 1 ( L ) > 4 / 5 \lambda_1(L)>4/5 λ 1 ( L ) > 4/5 . Choose the basis in Lemma 3.1. Since 0 ≤ x 1 , x 3 ≤ 1 / 2 0\le x_1,x_3\le1/2 0 ≤ x 1 , x 3 ≤ 1/2 , the third and sixth inequalities in (24) give
y 1 − 1 ≥ 1 − x 1 2 ≥ 3 4 , y 2 − 1 ≥ 1 − x 3 2 ≥ 3 4 . y_1^{-1}\ge1-x_1^2\ge\frac{3}{4},\qquad y_2^{-1}\ge1-x_3^2\ge\frac{3}{4}. y 1 − 1 ≥ 1 − x 1 2 ≥ 4 3 , y 2 − 1 ≥ 1 − x 3 2 ≥ 4 3 .
Thus y 1 , y 2 ≤ 4 / 3 y_1,y_2\le4/3 y 1 , y 2 ≤ 4/3 . Also d 1 3 = y 1 2 y 2 > ( 16 / 25 ) 3 d_1^3=y_1^2y_2>(16/25)^3 d 1 3 = y 1 2 y 2 > ( 16/25 ) 3 . Hence
y 1 2 > 3 4 ( 16 25 ) 3 = 3072 15625 > ( 11 25 ) 2 , y 2 > ( 3 4 ) 2 ( 16 25 ) 3 = 2304 15625 > 147 1000 . y_1^2>\frac{3}{4}\left(\frac{16}{25}\right)^3=\frac{3072}{15625}>\left(\frac{11}{25}\right)^2,\qquad
y_2>\left(\frac{3}{4}\right)^2\left(\frac{16}{25}\right)^3=\frac{2304}{15625}>\frac{147}{1000}. y 1 2 > 4 3 ( 25 16 ) 3 = 15625 3072 > ( 25 11 ) 2 , y 2 > ( 4 3 ) 2 ( 25 16 ) 3 = 15625 2304 > 1000 147 .
Thus all relevant lattices have representatives in (23) . The corresponding closed rectangle of logarithmic coordinates is compact, and its continuous Gram image consists of positive definite determinant-one matrices. The Ewald sums converge normally on this image, so P ( ⋅ , c ) P(\cdot,c) P ( ⋅ , c ) is continuous there and attains a minimum. The rectangle contains a representative of L F L_F L F , and every lattice outside its image either has energy larger than P ( L F , c ) P(L_F,c) P ( L F , c ) or has another representative inside. Its minimum is therefore the global minimum. The cusp comparison proves the asserted strict bound on its shortest vector. □
Lemma 3.3 (Separate matrix convexity). Let ξ = ( ℓ 1 , ℓ 2 , x 1 , x 2 , x 3 ) \xi=(\ell_1,\ell_2,x_1,x_2,x_3) ξ = ( ℓ 1 , ℓ 2 , x 1 , x 2 , x 3 ) . Both Y ( ξ ) Y(\xi) Y ( ξ ) and Y ( ξ ) − 1 Y(\xi)^{-1} Y ( ξ ) − 1 are convex in the Loewner order in each individual coordinate, with the other four fixed. Consequently, for S , T ⪰ 0 S,T\succeq0 S , T ⪰ 0 and a coordinate rectangle B B B ,
tr { S Y ( ξ ) + T Y ( ξ ) − 1 } ≤ max ζ ∈ C ( B ) tr { S Y ( ζ ) + T Y ( ζ ) − 1 } ( ξ ∈ B ) . (26) \operatorname{tr}\{SY(\xi)+TY(\xi)^{-1}\}\le\max_{\zeta\in\mathcal{C}(B)}\operatorname{tr}\{SY(\zeta)+TY(\zeta)^{-1}\}\qquad(\xi\in B). \tag*{(26)} tr { S Y ( ξ ) + T Y ( ξ ) − 1 } ≤ ζ ∈ C ( B ) max tr { S Y ( ζ ) + T Y ( ζ ) − 1 } ( ξ ∈ B ) . (26)
Proof. The matrix Y = N T D N Y=N^{T}DN Y = N T D N is a sum of positive d i d_i d i times the outer products of the rows of N N N . Each row is affine in an individual shear variable, so its outer product has a positive semidefinite second derivative in that variable. Each d i d_i d i is the exponential of an affine function of ( ℓ 1 , ℓ 2 ) (\ell_1,\ell_2) ( ℓ 1 , ℓ 2 ) , with the outer product fixed. The same argument applies to Y − 1 = N − 1 D − 1 N − T Y^{-1}=N^{-1}D^{-1}N^{-T} Y − 1 = N − 1 D − 1 N − T , using
N − 1 = ( 1 − x 1 x 1 x 3 − x 2 0 1 − x 3 0 0 1 ) . N^{-1}=
\begin{pmatrix}
1 & -x_1 & x_1x_3-x_2\\
0 & 1 & -x_3\\
0 & 0 & 1
\end{pmatrix}. N − 1 = 1 0 0 − x 1 1 0 x 1 x 3 − x 2 − x 3 1 .
Each column of N − 1 N^{-1} N − 1 is affine in each shear separately. Taking the trace against positive semidefinite matrices preserves convexity. Successive one-dimensional convex interpolation then proves (26) . The same interpolation proves the corresponding matrix inequalities in the Loewner order. □
The local comparison near FCC and BCC
We prove local comparisons on neighborhoods whose radii are independent of the screening parameter. On bounded screening intervals we estimate the quadratic and cubic variations and bound the fourth-order remainder uniformly. For large screening, the twelve FCC contact vectors give the required estimate directly. This is the local part of the method developed in our previous work [25 ] .
Throughout this section, c = κ 2 / ( 4 π ) c=\kappa^2/(4\pi) c = κ 2 / ( 4 π ) and F 0 = W F_0=W F 0 = W . For κ > 0 \kappa>0 κ > 0 , replacing F κ F_\kappa F κ by E κ E_\kappa E κ leaves every energy difference and shape derivative unchanged.
Theorem 4.1. Let X = X T X=X^{T} X = X T , tr X = 0 \operatorname{tr}X=0 tr X = 0 , and ∥ X ∥ F ≤ 1 / 40 \lVert X\rVert_{F}\le1/40 ∥ X ∥ F ≤ 1/40 . Then
F κ ( e X / 2 L B ) − F κ ( L B ) ≥ 1 1000 ∥ X ∥ F 2 ( 0 ≤ c ≤ 1 / 4 ) , (27) F_{\kappa}(e^{X/2}L_{B})-F_{\kappa}(L_{B})\ge\frac{1}{1000}\lVert X\rVert_{F}^{2}\qquad(0\le c\le1/4), \tag*{(27)} F κ ( e X /2 L B ) − F κ ( L B ) ≥ 1000 1 ∥ X ∥ F 2 ( 0 ≤ c ≤ 1/4 ) , (27) and, for every κ ≥ 0 \kappa\ge0 κ ≥ 0 ,
F κ ( e X / 2 L F ) − F κ ( L F ) ≥ ∥ X ∥ F 2 { 10 − 6 , c ≤ 10 , ( κ a + 1 ) e − κ a / ( 200 a ) , c > 10. (28) F_{\kappa}(e^{X/2}L_{F})-F_{\kappa}(L_{F})\ge\lVert X\rVert_{F}^{2}
\begin{cases}
10^{-6}, & c\le10,\\
(\kappa a+1)e^{-\kappa a/(200a)}, & c>10.
\end{cases} \tag*{(28)} F κ ( e X /2 L F ) − F κ ( L F ) ≥ ∥ X ∥ F 2 { 1 0 − 6 , ( κa + 1 ) e − κa / ( 200 a ) , c ≤ 10 , c > 10. (28) Moreover, if A = A T > 0 A=A^{T}>0 A = A T > 0 , det A = 1 \det A=1 det A = 1 , and ∥ A − I ∥ op ≤ 3 / 100 \lVert A-I\rVert_{\operatorname{op}}\le3/100 ∥ A − I ∥ op ≤ 3/100 , then
E κ ( A 1 / 2 L F ) − E κ ( L F ) ≥ s + 1 160 a e − s ∥ A − I ∥ F 2 , s = κ a ≥ 12. (29) E_{\kappa}(A^{1/2}L_{F})-E_{\kappa}(L_{F})\ge\frac{s+1}{160a}e^{-s}\lVert A-I\rVert_{F}^{2},\qquad s=\kappa a\ge12. \tag*{(29)} E κ ( A 1/2 L F ) − E κ ( L F ) ≥ 160 a s + 1 e − s ∥ A − I ∥ F 2 , s = κa ≥ 12. (29) Equality of the two energies in any of these comparisons forces the deformation to be the identity.
We first give the elementary estimates needed for (29) .
Lemma 4.1. Suppose A = I + T = A T > 0 A=I+T=A^{T}>0 A = I + T = A T > 0 , det A = 1 \det A=1 det A = 1 , and ∥ T ∥ op ≤ d < 2 / 3 \lVert T\rVert_{\operatorname{op}}\le d<2/3 ∥ T ∥ op ≤ d < 2/3 . Write T = η I + S T=\eta I+S T = η I + S , where tr S = 0 \operatorname{tr}S=0 tr S = 0 .
∥ T ∥ F 2 6 ( 1 + d ) ≤ η ≤ ∥ T ∥ F 2 6 ( 1 − d ) , η ≤ ∥ S ∥ F 2 6 − 9 d , ∥ S ∥ F ≤ 3 d . (30) \frac{\lVert T\rVert_{F}^{2}}{6(1+d)}\le\eta\le\frac{\lVert T\rVert_{F}^{2}}{6(1-d)},\qquad\eta\le\frac{\lVert S\rVert_{F}^{2}}{6-9d},\qquad\lVert S\rVert_{F}\le\sqrt{3}d. \tag*{(30)} 6 ( 1 + d ) ∥ T ∥ F 2 ≤ η ≤ 6 ( 1 − d ) ∥ T ∥ F 2 , η ≤ 6 − 9 d ∥ S ∥ F 2 , ∥ S ∥ F ≤ 3 d . (30) Proof. For ∣ t ∣ ≤ d \lvert t\rvert\le d ∣ t ∣ ≤ d ,
t 2 2 ( 1 + d ) ≤ t − log ( 1 + t ) ≤ t 2 2 ( 1 − d ) . \frac{t^{2}}{2(1+d)}\le t-\log(1+t)\le\frac{t^{2}}{2(1-d)}. 2 ( 1 + d ) t 2 ≤ t − log ( 1 + t ) ≤ 2 ( 1 − d ) t 2 . Sum over the eigenvalues of T T T , using log det A = 0 \log\det A=0 log det A = 0 , to get
∥ T ∥ F 2 2 ( 1 + d ) ≤ tr T ≤ ∥ T ∥ F 2 2 ( 1 − d ) . (31) \frac{\lVert T\rVert_{F}^{2}}{2(1+d)}\le\operatorname{tr}T\le\frac{\lVert T\rVert_{F}^{2}}{2(1-d)}. \tag*{(31)} 2 ( 1 + d ) ∥ T ∥ F 2 ≤ tr T ≤ 2 ( 1 − d ) ∥ T ∥ F 2 . (31) Thus 0 ≤ η ≤ d 0\le\eta\le d 0 ≤ η ≤ d . Since tr T = 3 η \operatorname{tr}T=3\eta tr T = 3 η and ∥ T ∥ F 2 = ∥ S ∥ F 2 + 3 η 2 \lVert T\rVert_{F}^{2}=\lVert S\rVert_{F}^{2}+3\eta^{2} ∥ T ∥ F 2 = ∥ S ∥ F 2 + 3 η 2 , the upper bound gives 6 ( 1 − d ) η ≤ ∥ S ∥ F 2 + 3 η 2 ≤ ∥ S ∥ F 2 + 3 d η 6(1-d)\eta\le\lVert S\rVert_{F}^{2}+3\eta^{2}\le\lVert S\rVert_{F}^{2}+3d\eta 6 ( 1 − d ) η ≤ ∥ S ∥ F 2 + 3 η 2 ≤ ∥ S ∥ F 2 + 3 d η . Rearranging yields ( 6 − 9 d ) η ≤ ∥ S ∥ F 2 (6-9d)\eta\le\lVert S\rVert_{F}^{2} ( 6 − 9 d ) η ≤ ∥ S ∥ F 2 . Finally, ∥ S ∥ F ≤ ∥ T ∥ F ≤ 3 d \lVert S\rVert_{F}\le\lVert T\rVert_{F}\le\sqrt{3}d ∥ S ∥ F ≤ ∥ T ∥ F ≤ 3 d . □
Lemma 4.2. If x 1 , … , x 6 ∈ R x_{1},\ldots,x_{6}\in\mathbb{R} x 1 , … , x 6 ∈ R , ∑ j x j = 0 \sum_{j}x_{j}=0 ∑ j x j = 0 , and σ 2 = 1 6 ∑ j x j 2 \sigma^{2}=\frac{1}{6}\sum_{j}x_{j}^{2} σ 2 = 6 1 ∑ j x j 2 , then
log ( 1 6 ∑ j = 1 6 e x j ) ≥ 1 13 min { σ 2 , σ } . (32) \log\left(\frac{1}{6}\sum_{j=1}^{6}e^{x_{j}}\right)\ge\frac{1}{13}\min\{\sigma^{2},\sigma\}. \tag*{(32)} log ( 6 1 j = 1 ∑ 6 e x j ) ≥ 13 1 min { σ 2 , σ } . (32) Proof. There is nothing to prove when σ = 0 \sigma=0 σ = 0 . Since the entries sum to zero and at most five are positive,
∑ x j < 0 x j 2 ≤ ( ∑ x j > 0 x j ) 2 ≤ 5 ∑ x j > 0 x j 2 , \sum_{x_{j}<0}x_{j}^{2}\le\left(\sum_{x_{j}>0}x_{j}\right)^{2}\le5\sum_{x_{j}>0}x_{j}^{2}, x j < 0 ∑ x j 2 ≤ x j > 0 ∑ x j 2 ≤ 5 x j > 0 ∑ x j 2 , and hence ∑ x j > 0 x j 2 ≥ σ 2 \sum_{x_{j}>0}x_{j}^{2}\ge\sigma^{2} ∑ x j > 0 x j 2 ≥ σ 2 . Apply e x ≥ 1 + x e^{x}\ge1+x e x ≥ 1 + x to negative entries and e x ≥ 1 + x + x 2 / 2 e^{x}\ge1+x+x^{2}/2 e x ≥ 1 + x + x 2 /2 to positive entries. This yields 1 6 ∑ e x j ≥ 1 + σ 2 / 12 \frac{1}{6}\sum e^{x_{j}}\ge1+\sigma^{2}/12 6 1 ∑ e x j ≥ 1 + σ 2 /12 , so its logarithm is at least σ 2 / ( 12 + σ 2 ) \sigma^{2}/(12+\sigma^{2}) σ 2 / ( 12 + σ 2 ) . This proves the assertion for 0 ≤ σ ≤ 8 0\le\sigma\le8 0 ≤ σ ≤ 8 .
If σ ≥ 8 \sigma\ge8 σ ≥ 8 , put m = max j x j m=\max_{j}x_{j} m = max j x j . Every entry lies in [ − 5 m , m ] [-5m,m] [ − 5 m , m ] . Averaging ( m − x j ) ( x j + 5 m ) ≥ 0 (m-x_{j})(x_{j}+5m)\ge0 ( m − x j ) ( x j + 5 m ) ≥ 0 gives σ 2 ≤ 5 m 2 \sigma^{2}\le5m^{2} σ 2 ≤ 5 m 2 ; therefore
log ( 1 6 ∑ e x j ) ≥ σ 5 − log 6 > 4 σ 9 − 2 ≥ σ 13 . \log\left(\frac{1}{6}\sum e^{x_{j}}\right)\ge\frac{\sigma}{\sqrt{5}}-\log6>\frac{4\sigma}{9}-2\ge\frac{\sigma}{13}. log ( 6 1 ∑ e x j ) ≥ 5 σ − log 6 > 9 4 σ − 2 ≥ 13 σ .
Proposition 4.1. Let N n N_n N n denote the number of F C C FCC F C C vectors of squared length n a 2 na^2 n a 2 . Then N 1 = 12 N_1=12 N 1 = 12 , N 2 = 6 N_2=6 N 2 = 6 , N n < 30 n N_n<30n N n < 30 n for every n ≥ 1 n\ge1 n ≥ 1 , and
∑ n ≥ 2 N n e − s ( n − 1 ) < 3 , E κ ( L F ) < 15 2 a e − s ( s = κ a ≥ 10 ) . (33) \sum_{n\ge2}N_n e^{-s(\sqrt{n}-1)}<3,\qquad E_\kappa(L_F)<\frac{15}{2a}e^{-s}\qquad(s=\kappa a\ge10). \tag*{(33)} n ≥ 2 ∑ N n e − s ( n − 1 ) < 3 , E κ ( L F ) < 2 a 15 e − s ( s = κa ≥ 10 ) . (33)
Proof. The n n n th shell is represented by integer triples of squared norm 2 n 2n 2 n and even coordinate sum. The first two counts are immediate. Fixing the first two coordinates leaves at most two choices for the third, so
N n ≤ 2 ( 2 ⌊ 2 n ⌋ + 1 ) 2 ≤ 16 n + 8 2 n + 2 < 30 n . N_n\le2(2\lfloor\sqrt{2n}\rfloor+1)^2\le16n+8\sqrt{2n}+2<30n. N n ≤ 2 ( 2 ⌊ 2 n ⌋ + 1 ) 2 ≤ 16 n + 8 2 n + 2 < 30 n .
For s ≥ 10 s\ge10 s ≥ 10 , the function x e − s ( x − 1 ) xe^{-s(\sqrt{x}-1)} x e − s ( x − 1 ) decreases on [ 2 , ∞ ) [2,\infty) [ 2 , ∞ ) . Sum–integral comparison gives
∑ n ≥ 2 N n e − s ( n − 1 ) < 30 e − s ( 2 − 1 ) [ 2 + 2 ( 2 2 s + 6 s 2 + 6 2 s 3 + 6 s 4 ) ] < 90 e − 4 < 3. \begin{aligned}
\sum_{n\ge2}N_ne^{-s(\sqrt{n}-1)}
<30e^{-s(\sqrt{2}-1)}
\left[2+2\left(\frac{2\sqrt{2}}{s}+\frac{6}{s^2}+\frac{6\sqrt{2}}{s^3}+\frac{6}{s^4}\right)\right] \\
&<90e^{-4}<3.
\end{aligned} n ≥ 2 ∑ N n e − s ( n − 1 ) < 30 e − s ( 2 − 1 ) [ 2 + 2 ( s 2 2 + s 2 6 + s 3 6 2 + s 4 6 ) ] < 90 e − 4 < 3.
Here the bracket is less than 2 + 6 / 10 + 12 / 10 2 + 18 / 10 3 + 12 / 10 4 < 3 2+6/10+12/10^2+18/10^3+12/10^4<3 2 + 6/10 + 12/1 0 2 + 18/1 0 3 + 12/1 0 4 < 3 , and 2 − 1 > 2 / 5 \sqrt{2}-1>2/5 2 − 1 > 2/5 . Summing the energy by shells now proves its stated upper bound. □
For the bounded parameter range, consider the determinant-preserving deformations e X / 2 C e^{X/2}\mathcal{C} e X /2 C , where C ∈ { L F , L B } C\in\{L_F,L_B\} C ∈ { L F , L B } , X = X T X=X^T X = X T , and tr X = 0 \operatorname{tr}X=0 tr X = 0 .
For A > 0 A>0 A > 0 and j ∈ Z j\in\mathbb{Z} j ∈ Z , define
K j ( A , c ) = ∫ 1 ∞ t j − 1 / 2 e − A t − c / t d t , T j ( A , c ) = ∫ 1 ∞ t j e − ( A + c ) t d t , K_j(A,c)=\int_1^\infty t^{j-1/2}e^{-At-c/t}\,dt,\qquad T_j(A,c)=\int_1^\infty t^j e^{-(A+c)t}\,dt, K j ( A , c ) = ∫ 1 ∞ t j − 1/2 e − A t − c / t d t , T j ( A , c ) = ∫ 1 ∞ t j e − ( A + c ) t d t ,
and write K j ( A ) = K j ( A , 0 ) K_j(A)=K_j(A,0) K j ( A ) = K j ( A , 0 ) . These kernels satisfy K 0 ( π q , c ) = I ( q , c ) K_0(\pi q,c)=I(q,c) K 0 ( π q , c ) = I ( q , c ) and T 0 ( π q , c ) = J ( q , c ) T_0(\pi q,c)=J(q,c) T 0 ( π q , c ) = J ( q , c ) , so (2.4) gives the derivatives of X ↦ F κ ( e X / 2 C ) X\mapsto F_\kappa(e^{X/2}C) X ↦ F κ ( e X /2 C ) . Normal convergence, as established in Section 2, justifies the following differentiations, also at c = 0 c=0 c = 0 .
For a cubic shell S S S set
M ( S ) = ∑ v ∈ S v 1 2 , A D ( S ) = ∑ v ∈ S ( v 1 4 − v 1 2 v 2 2 ) , A O ( S ) = 2 ∑ v ∈ S v 1 2 v 2 2 . M(S)=\sum_{v\in S}v_1^2,\qquad A_D(S)=\sum_{v\in S}(v_1^4-v_1^2v_2^2),\qquad A_O(S)=2\sum_{v\in S}v_1^2v_2^2. M ( S ) = v ∈ S ∑ v 1 2 , A D ( S ) = v ∈ S ∑ ( v 1 4 − v 1 2 v 2 2 ) , A O ( S ) = 2 v ∈ S ∑ v 1 2 v 2 2 .
For a traceless symmetric matrix H H H put
I 1 ( H ) = ∑ i H i i 3 , I 2 ( H ) = H 11 H 23 2 + H 22 H 13 2 + H 33 H 12 2 , I 3 ( H ) = H 12 H 13 H 23 . I_1(H)=\sum_i H_{ii}^3,\qquad I_2(H)=H_{11}H_{23}^2+H_{22}H_{13}^2+H_{33}H_{12}^2,\qquad I_3(H)=H_{12}H_{13}H_{23}. I 1 ( H ) = i ∑ H ii 3 , I 2 ( H ) = H 11 H 23 2 + H 22 H 13 2 + H 33 H 12 2 , I 3 ( H ) = H 12 H 13 H 23 .
Lemma 4.3 (Variations at the cubic lattices). The first variation at C ∈ { L F , L B } C\in\{L_F,L_B\} C ∈ { L F , L B } vanishes on traceless symmetric directions. Its Hessian is scalar on the diagonal traceless and off-diagonal subspaces, with Frobenius eigenvalues
μ τ ( c ) = 1 2 ∑ S ⊂ C { π 2 A τ ( S ) K 2 ( π q S , c ) − π M ( S ) K 1 ( π q S , c ) } \mu_\tau(c)=\frac{1}{2}\sum_{S\subset C}\left\{\pi^2A_\tau(S)K_2(\pi q_S,c)-\pi M(S)K_1(\pi q_S,c)\right\} μ τ ( c ) = 2 1 S ⊂ C ∑ { π 2 A τ ( S ) K 2 ( π q S , c ) − π M ( S ) K 1 ( π q S , c ) } + 1 2 ∑ S ⊂ C ∗ { π 2 A τ ( S ) T 2 ( π q S , c ) − π M ( S ) T 1 ( π q S , c ) } , τ ∈ { D , O } . (34) +\frac{1}{2}\sum_{S\subset C^*}\left\{\pi^2A_\tau(S)T_2(\pi q_S,c)-\pi M(S)T_1(\pi q_S,c)\right\},\qquad\tau\in\{D,O\}. \tag*{(34)} + 2 1 S ⊂ C ∗ ∑ { π 2 A τ ( S ) T 2 ( π q S , c ) − π M ( S ) T 1 ( π q S , c ) } , τ ∈ { D , O } . (34) where q S = ∣ v ∣ 2 q_S=|v|^2 q S = ∣ v ∣ 2 on the shell S S S . The third variation satisfies
d 3 d t 3 F κ ( e t H / 2 C ) ∣ t = 0 = α ( c ) I 1 ( H ) + β ( c ) I 2 ( H ) + γ ( c ) I 3 ( H ) . (35) \left.\frac{d^3}{dt^3}F_\kappa(e^{tH/2}C)\right|_{t=0}
=\alpha(c)I_1(H)+\beta(c)I_2(H)+\gamma(c)I_3(H). \tag*{(35)} d t 3 d 3 F κ ( e t H /2 C ) t = 0 = α ( c ) I 1 ( H ) + β ( c ) I 2 ( H ) + γ ( c ) I 3 ( H ) . (35) Here H = H T H=H^T H = H T and tr H = 0 \operatorname{tr}H=0 tr H = 0 . For ∥ H ∥ F = 1 \lVert H\rVert_F=1 ∥ H ∥ F = 1 ,
∣ I 1 ( H ) ∣ ≤ 1 6 < 5 12 , ∣ I 2 ( H ) ∣ ≤ 2 9 < 1 6 , ∣ I 3 ( H ) ∣ ≤ 1 6 6 < 5 72 . (36) |I_1(H)|\le\frac{1}{\sqrt{6}}<\frac{5}{12},\qquad
|I_2(H)|\le\frac{\sqrt{2}}{9}<\frac{1}{6},\qquad
|I_3(H)|\le\frac{1}{6\sqrt{6}}<\frac{5}{72}. \tag*{(36)} ∣ I 1 ( H ) ∣ ≤ 6 1 < 12 5 , ∣ I 2 ( H ) ∣ ≤ 9 2 < 6 1 , ∣ I 3 ( H ) ∣ ≤ 6 6 1 < 72 5 . (36)
Proof. Sign changes and permutations make the first derivative zero and split the Hessian into the stated eigenspaces. The first two derivatives of v T e t H v v^{T}e^{tH}v v T e t H v at zero are v T H v v^{T}Hv v T H v and v T H 2 v v^{T}H^{2}v v T H 2 v . For a unit Frobenius direction in the diagonal traceless or off-diagonal subspace, shell symmetry gives, respectively,
∑ v ∈ S ( v T H v ) 2 = A τ ( S ) , ∑ v ∈ S v T H 2 v = M ( S ) ( τ = D , O ) . \sum_{v\in S}(v^{T}Hv)^{2}=A_{\tau}(S),\qquad\sum_{v\in S}v^{T}H^{2}v=M(S)\qquad(\tau=D,O). v ∈ S ∑ ( v T H v ) 2 = A τ ( S ) , v ∈ S ∑ v T H 2 v = M ( S ) ( τ = D , O ) .
Differentiation under the integral gives I q ( q , c ) = − π K 1 ( π q , c ) I_q(q,c)=-\pi K_1(\pi q,c) I q ( q , c ) = − π K 1 ( π q , c ) and I q q ( q , c ) = π 2 K 2 ( π q , c ) I_{qq}(q,c)=\pi^2K_2(\pi q,c) I q q ( q , c ) = π 2 K 2 ( π q , c ) . The chain rule now gives the direct part of (4.8); the dual deformation e − t H / 2 e^{-tH/2} e − t H /2 gives the same second-variation formula with T j T_j T j in place of K j K_j K j . The fourth moments are nonnegative, since A D ( S ) = 1 2 ∑ S ( v 1 2 − v 2 2 ) 2 A_D(S)=\frac{1}{2}\sum_S(v_1^2-v_2^2)^2 A D ( S ) = 2 1 ∑ S ( v 1 2 − v 2 2 ) 2 and A O ( S ) ≥ 0 A_O(S)\ge0 A O ( S ) ≥ 0 .
A cubic invariant under coordinate sign changes and permutations is a linear combination of I 1 , I 2 , I 3 I_1,I_2,I_3 I 1 , I 2 , I 3 . A surviving monomial contains three diagonal entries, one diagonal entry and a squared off-diagonal entry, or the three different off-diagonal entries; tr H = 0 \operatorname{tr}H=0 tr H = 0 gives the stated reduction. For the middle bound in (4.10), put d 2 = ∑ H i i 2 d^2=\sum H_{ii}^2 d 2 = ∑ H ii 2 . Since max i ∣ H i i ∣ ≤ 2 3 d \max_i|H_{ii}|\le\sqrt{\frac{2}{3}}d max i ∣ H ii ∣ ≤ 3 2 d and ∑ i < j H i j 2 = ( 1 − d 2 ) / 2 \sum_{i<j}H_{ij}^2=(1-d^2)/2 ∑ i < j H ij 2 = ( 1 − d 2 ) /2 ,
∣ I 2 ( H ) ∣ ≤ max i ∣ H i i ∣ ∑ i < j H i j 2 ≤ 2 / 3 2 d ( 1 − d 2 ) ≤ 2 9 . |I_2(H)|\le\max_i|H_{ii}|\sum_{i<j}H_{ij}^2\le\frac{\sqrt{2/3}}{2}d(1-d^2)\le\frac{\sqrt{2}}{9}. ∣ I 2 ( H ) ∣ ≤ i max ∣ H ii ∣ i < j ∑ H ij 2 ≤ 2 2/3 d ( 1 − d 2 ) ≤ 9 2 .
The last step uses max 0 ≤ d ≤ 1 d ( 1 − d 2 ) = 2 / ( 3 3 ) \max_{0\le d\le1}d(1-d^2)=2/(3\sqrt{3}) max 0 ≤ d ≤ 1 d ( 1 − d 2 ) = 2/ ( 3 3 ) . The other two bounds follow from the traceless diagonal constraint and the arithmetic–geometric mean inequality.
For the coefficients in (4.9), use the matrices
H D = ( 1 0 0 0 1 0 0 0 − 2 ) , H T = ( 1 0 0 0 − 1 1 0 1 0 ) , H O = ( 0 1 1 1 0 1 1 1 0 ) . H_D=\begin{pmatrix}1&0&0\\0&1&0\\0&0&-2\end{pmatrix},\qquad
H_T=\begin{pmatrix}1&0&0\\0&-1&1\\0&1&0\end{pmatrix},\qquad
H_O=\begin{pmatrix}0&1&1\\1&0&1\\1&1&0\end{pmatrix}. H D = 1 0 0 0 1 0 0 0 − 2 , H T = 1 0 0 0 − 1 1 0 1 0 , H O = 0 1 1 1 0 1 1 1 0 .
Their cubic invariants are ( − 6 , 0 , 0 ) (-6,0,0) ( − 6 , 0 , 0 ) , ( 0 , 1 , 0 ) (0,1,0) ( 0 , 1 , 0 ) , and ( 0 , 0 , 1 ) (0,0,1) ( 0 , 0 , 1 ) . For a shell represented by integer vectors m m m with physical vector v = s m v=\sqrt{s}\,m v = s m , define
U 3 ( H ) = ∑ m ( m T H m ) 3 , U 12 ( H ) = ∑ m ( m T H m ) ( m T H 2 m ) , U 1 ( H ) = ∑ m m T H 3 m . U_3(H)=\sum_m(m^{T}Hm)^3,\qquad U_{12}(H)=\sum_m(m^{T}Hm)(m^{T}H^2m),\qquad U_1(H)=\sum_m m^{T}H^3m. U 3 ( H ) = m ∑ ( m T H m ) 3 , U 12 ( H ) = m ∑ ( m T H m ) ( m T H 2 m ) , U 1 ( H ) = m ∑ m T H 3 m .
Its direct contribution to the directional third derivative is
1 2 { − π 3 s 3 U 3 ( H ) K 3 + 3 π 2 s 2 U 12 ( H ) K 2 − π s U 1 ( H ) K 1 } . (37) \frac{1}{2}\{-\pi^3s^3U_3(H)K_3+3\pi^2s^2U_{12}(H)K_2-\pi sU_1(H)K_1\}. \tag*{(37)} 2 1 { − π 3 s 3 U 3 ( H ) K 3 + 3 π 2 s 2 U 12 ( H ) K 2 − π s U 1 ( H ) K 1 } . (37)
The arguments are ( π s ∣ m ∣ 2 , c ) (\pi s|m|^2,c) ( π s ∣ m ∣ 2 , c ) . The dual contribution is the negative of the same expression with T j T_j T j in place of K j K_j K j , because the dual deformation is e − X / 2 e^{-X/2} e − X /2 . Sum these expressions and evaluate on H D , H T , H O H_D,H_T,H_O H D , H T , H O to obtain − 6 α , β , γ -6\alpha,\beta,\gamma − 6 α , β , γ . □
Lemma 4.4 (Uniform derivative tails). Let δ = 1 / 40 \delta=1/40 δ = 1/40 . After retaining FCC shells n ≤ 9 n\le9 n ≤ 9 and BCC shells ℓ ≤ 30 \ell\le30 ℓ ≤ 30 , the omitted directional second and third variations at the origin are less in absolute value than 10 − 10 10^{-10} 1 0 − 10 and 10 − 8 10^{-8} 1 0 − 8 , respectively, for ∥ H ∥ F = 1 \lVert H\rVert_F=1 ∥ H ∥ F = 1 . The omitted fourth variation on ∥ X ∥ F ≤ δ \lVert X\rVert_F\le\delta ∥ X ∥ F ≤ δ is less than 10 − 6 10^{-6} 1 0 − 6 under the same normalization. These bounds hold for both references and every c ≥ 0 c\ge0 c ≥ 0 .
Proof. Set δ = 1 / 40 \delta=1/40 δ = 1/40 . If ∥ X ∥ F ≤ δ \lVert X\rVert_F\le\delta ∥ X ∥ F ≤ δ and ∥ H ∥ F = 1 \lVert H\rVert_F=1 ∥ H ∥ F = 1 , then all derivatives of v T e X + u H v v^{T}e^{X+uH}v v T e X + u H v of order j ≤ 4 j\le4 j ≤ 4 at u = 0 u=0 u = 0 have absolute value at most e δ ∣ v ∣ 2 e^\delta|v|^2 e δ ∣ v ∣ 2 . This follows either by the iterated integral formula for derivatives of the matrix exponential or, along the radial Taylor path, by diagonalizing H H H . Also v T e X v ≥ e − δ ∣ v ∣ 2 v^{T}e^Xv\ge e^{-\delta}|v|^2 v T e X v ≥ e − δ ∣ v ∣ 2 . Put A = π e − δ ∣ v ∣ 2 A=\pi e^{-\delta}|v|^2 A = π e − δ ∣ v ∣ 2 and U = π e δ ∣ v ∣ 2 U=\pi e^\delta|v|^2 U = π e δ ∣ v ∣ 2 . The fourth derivative of a summand is therefore bounded, after integration, by
U 4 K 4 ( A , c ) + 6 U 3 K 3 ( A , c ) + 7 U 2 K 2 ( A , c ) + U K 1 ( A , c ) . (38) U^4K_4(A,c)+6U^3K_3(A,c)+7U^2K_2(A,c)+UK_1(A,c). \tag*{(38)} U 4 K 4 ( A , c ) + 6 U 3 K 3 ( A , c ) + 7 U 2 K 2 ( A , c ) + U K 1 ( A , c ) . (38)
and the same formula with T j T_j T j for a dual summand. The coefficients 1 , 6 , 7 , 1 1,6,7,1 1 , 6 , 7 , 1 come from the fourth derivative of an exponential of a function. The factor 1 / 2 1/2 1/2 from (2.4) is retained when summing this estimate.
We spell out all omitted-shell estimates. FCC shells have ∣ v ∣ 2 = 2 1 / 3 n |v|^2 = 2^{1/3}n ∣ v ∣ 2 = 2 1/3 n , with N n < 30 n N_n < 30n N n < 30 n . BCC can be written
L B = { 2 − 2 / 3 m : m ∈ Z 3 , m 1 ≡ m 2 ≡ m 3 ( m o d 2 ) } . L_{\mathrm{B}} = \left\{2^{-2/3}m : m \in\mathbb{Z}^3,\ m_1 \equiv m_2 \equiv m_3 \pmod{2}\right\}. L B = { 2 − 2/3 m : m ∈ Z 3 , m 1 ≡ m 2 ≡ m 3 ( mod 2 ) } .
The number of integer triples of squared norm ℓ \ell ℓ is less than 30 ℓ 30\ell 30 ℓ , by fixing two coordinates as in Proposition 4.1. For every omitted vector, A = π e − δ ∣ v ∣ 2 > 3 A = \pi e^{-\delta}|v|^2 > 3 A = π e − δ ∣ v ∣ 2 > 3 , and
2 1 / 3 π < 4 , 2 − 4 / 3 π < 5 4 , e 8 δ < 5 4 , 2^{1/3}\pi< 4,\qquad2^{-4/3}\pi< \frac{5}{4},\qquad e^{8\delta} < \frac{5}{4}, 2 1/3 π < 4 , 2 − 4/3 π < 4 5 , e 8 δ < 4 5 ,
e − 39 ⋅ 2 1 / 3 π / 40 < 1 47 , e − 39 ⋅ 2 − 4 / 3 π / 40 < 3 10 . (39) e^{-39\cdot2^{1/3}\pi/40} < \frac{1}{47},\qquad e^{-39\cdot2^{-4/3}\pi/40} < \frac{3}{10}. \tag*{(39)} e − 39 ⋅ 2 1/3 π /40 < 47 1 , e − 39 ⋅ 2 − 4/3 π /40 < 10 3 . (39)
Since e − δ > 39 / 40 e^{-\delta} > 39/40 e − δ > 39/40 , the two exponential bounds also hold with 39 / 40 39/40 39/40 replaced by e − δ e^{-\delta} e − δ . Dropping screening and using t − 1 / 2 ≤ 1 t^{-1/2} \le1 t − 1/2 ≤ 1 shows that an omitted summand in (38) , including its factor 1 / 2 1/2 1/2 , is at most
5 8 e − A ( A 3 + 10 A 2 + 37 A + 75 + 75 / A ) ≤ 5 8 e − A ( A 3 + 10 A 2 + 37 A + 100 ) . \frac{5}{8}e^{-A}(A^3+10A^2+37A+75+75/A) \le\frac{5}{8}e^{-A}(A^3+10A^2+37A+100). 8 5 e − A ( A 3 + 10 A 2 + 37 A + 75 + 75/ A ) ≤ 8 5 e − A ( A 3 + 10 A 2 + 37 A + 100 ) .
Thus, after retaining FCC shells n ≤ 9 n \le9 n ≤ 9 and BCC shells ℓ ≤ 30 \ell\le30 ℓ ≤ 30 , the omitted fourth derivative norm is less than
R 4 = 75 4 ∑ n = 10 ∞ n ( 64 n 3 + 160 n 2 + 148 n + 100 ) 47 n + 75 4 ∑ ℓ = 31 ∞ ℓ ( 3 10 ) ℓ ( 125 64 ℓ 3 + 125 8 ℓ 2 + 185 4 ℓ + 100 ) < 10 − 6 . (40) R_4 = \frac{75}{4}\sum_{n=10}^{\infty}\frac{n(64n^3+160n^2+148n+100)}{47^n}
+\frac{75}{4}\sum_{\ell=31}^{\infty}\ell\left(\frac{3}{10}\right)^\ell\left(\frac{125}{64}\ell^3+\frac{125}{8}\ell^2+\frac{185}{4}\ell+100\right)<10^{-6}. \tag*{(40)} R 4 = 4 75 n = 10 ∑ ∞ 4 7 n n ( 64 n 3 + 160 n 2 + 148 n + 100 ) + 4 75 ℓ = 31 ∑ ∞ ℓ ( 10 3 ) ℓ ( 64 125 ℓ 3 + 8 125 ℓ 2 + 4 185 ℓ + 100 ) < 1 0 − 6 . (40)
The fourth derivative bound M 4 M_4 M 4 used below is the finite sum in (38) , plus 10 − 6 10^{-6} 1 0 − 6 .
At the origin the corresponding absolute Hessian tail is bounded by
R 2 = 30 ∑ n ≥ 10 n ( 4 n + 20 ) 49 n + 30 ∑ ℓ ≥ 31 ℓ ( 3 10 ) ℓ ( 5 ℓ / 4 + 20 ) < 10 − 10 . (41) R_2 = 30\sum_{n\ge10}\frac{n(4n+20)}{49^n}+30\sum_{\ell\ge31}\ell\left(\frac{3}{10}\right)^\ell(5\ell/4+20)<10^{-10}. \tag*{(41)} R 2 = 30 n ≥ 10 ∑ 4 9 n n ( 4 n + 20 ) + 30 ℓ ≥ 31 ∑ ℓ ( 10 3 ) ℓ ( 5 ℓ /4 + 20 ) < 1 0 − 10 . (41)
Indeed, a unit Frobenius direction has ∣ v T H v ∣ ≤ ∣ v ∣ 2 |v^T Hv| \le|v|^2 ∣ v T H v ∣ ≤ ∣ v ∣ 2 and ∣ v T H 2 v ∣ ≤ ∣ v ∣ 2 |v^T H^2v| \le|v|^2 ∣ v T H 2 v ∣ ≤ ∣ v ∣ 2 . After dropping screening, the contribution of a shell S S S in absolute value is at most 1 2 # S e − A ( A + 3 + 3 / A ) ≤ # S e − A ( A + 20 ) \frac{1}{2}\#S e^{-A}(A+3+3/A) \le\#S e^{-A}(A+20) 2 1 # S e − A ( A + 3 + 3/ A ) ≤ # S e − A ( A + 20 ) , where A = π ∣ v ∣ 2 > 3 A=\pi|v|^2>3 A = π ∣ v ∣ 2 > 3 ; this proves the displayed tail bound. For completeness, integrating the third derivative polynomial ( A t ) 3 + 3 ( A t ) 2 + A t (At)^3+3(At)^2+At ( A t ) 3 + 3 ( A t ) 2 + A t bounds the omitted third derivative norm by
R 3 = 18 ∑ n ≥ 10 n ( 16 n 2 + 24 n + 26 ) 49 n + 18 ∑ ℓ ≥ 31 ℓ ( 3 10 ) ℓ ( 25 ℓ 2 / 16 + 15 ℓ / 2 + 26 ) < 10 − 8 . (42) R_3 = 18\sum_{n\ge10}\frac{n(16n^2+24n+26)}{49^n}+18\sum_{\ell\ge31}\ell\left(\frac{3}{10}\right)^\ell(25\ell^2/16+15\ell/2+26)<10^{-8}. \tag*{(42)} R 3 = 18 n ≥ 10 ∑ 4 9 n n ( 16 n 2 + 24 n + 26 ) + 18 ℓ ≥ 31 ∑ ℓ ( 10 3 ) ℓ ( 25 ℓ 2 /16 + 15 ℓ /2 + 26 ) < 1 0 − 8 . (42)
For example, the polynomial before its final simplification is e − A ( A 2 + 6 A + 13 + 13 / A ) e^{-A}(A^2+6A+13+13/A) e − A ( A 2 + 6 A + 13 + 13/ A ) , and the prefactor 3 / 5 3/5 3/5 used here exceeds the required 1 / 2 1/2 1/2 . All inequalities in (40) –(42) are rational: for their evaluation one can use
∑ n = N ∞ n k r n = ( r d d r ) k r N 1 − r , 0 < r < 1. \sum_{n=N}^{\infty}n^k r^n=\left(r\frac{d}{dr}\right)^k\frac{r^N}{1-r},\qquad0<r<1. n = N ∑ ∞ n k r n = ( r d r d ) k 1 − r r N , 0 < r < 1.
For the truncated FCC sum, retain the integer vectors satisfying
z ∈ Z 3 , 0 < ∣ z ∣ 2 ≤ 18 , z 1 + z 2 + z 3 ∈ 2 Z ; z\in\mathbb{Z}^3,\qquad0<|z|^2\le18,\qquad z_1+z_2+z_3\in2\mathbb{Z}; z ∈ Z 3 , 0 < ∣ z ∣ 2 ≤ 18 , z 1 + z 2 + z 3 ∈ 2 Z ;
for the BCC sum, retain those satisfying
z ∈ Z 3 , 0 < ∣ z ∣ 2 ≤ 30 , z 1 ≡ z 2 ≡ z 3 ( m o d 2 ) . z\in\mathbb{Z}^3,\qquad0<|z|^2\le30,\qquad z_1\equiv z_2\equiv z_3\pmod{2}. z ∈ Z 3 , 0 < ∣ z ∣ 2 ≤ 30 , z 1 ≡ z 2 ≡ z 3 ( mod 2 ) .
Group the vectors by m = ∣ z ∣ 2 m=|z|^2 m = ∣ z ∣ 2 . Every coordinate lies in [ − 6 , 6 ] [-6,6] [ − 6 , 6 ] . The physical squared scales are 2 − 2 / 3 2^{-2/3} 2 − 2/3 for FCC and 2 − 4 / 3 2^{-4/3} 2 − 4/3 for BCC; s s s denotes the applicable scale below. For C = L F C=L_F C = L F the direct sum uses the first family and the dual sum the second; for C = L B C=L_{\mathrm{B}} C = L B interchange the two families. For each shell form the integer moments
m 2 = ∑ z z 1 2 , a D = ∑ z ( z 1 4 − z 1 2 z 2 2 ) , a O = 2 ∑ z z 1 2 z 2 2 . m_2=\sum_z z_1^2,\qquad a_{\mathrm{D}}=\sum_z(z_1^4-z_1^2z_2^2),\qquad a_{\mathrm{O}}=2\sum_z z_1^2z_2^2. m 2 = z ∑ z 1 2 , a D = z ∑ ( z 1 4 − z 1 2 z 2 2 ) , a O = 2 z ∑ z 1 2 z 2 2 .
and the three sets of integer moments in (37) . The physical moments in (34) are M = s m 2 M=sm_2 M = s m 2 and A τ = s 2 a τ A_\tau=s^2a_\tau A τ = s 2 a τ . Thus every retained shell moment is a finite integer sum times a power of 2 − 2 / 3 2^{-2/3} 2 − 2/3 or 2 − 4 / 3 2^{-4/3} 2 − 4/3 .
For 1 ≤ j ≤ 4 1\le j\le4 1 ≤ j ≤ 4 put
L j ( A , c ) = ∑ n = 0 61 ( − c ) n n ! K j − n ( A ) , U j ( A , c ) = ∑ n = 0 60 ( − c ) n n ! K j − n ( A ) . L_j(A,c)=\sum_{n=0}^{61}\frac{(-c)^n}{n!}K_{j-n}(A),\qquad U_j(A,c)=\sum_{n=0}^{60}\frac{(-c)^n}{n!}K_{j-n}(A). L j ( A , c ) = n = 0 ∑ 61 n ! ( − c ) n K j − n ( A ) , U j ( A , c ) = n = 0 ∑ 60 n ! ( − c ) n K j − n ( A ) .
Only the indices down to − 60 -60 − 60 are needed in the expressions below. Taylor’s formula with its signed integral remainder gives, for every c ≥ 0 c\ge0 c ≥ 0 ,
L j ( A , c ) ≤ K j ( A , c ) ≤ U j ( A , c ) . (43) L_j(A,c)\le K_j(A,c)\le U_j(A,c). \tag*{(43)} L j ( A , c ) ≤ K j ( A , c ) ≤ U j ( A , c ) . (43)
No assumption about decreasing Taylor terms is needed. Both K j ( A , c ) K_j(A,c) K j ( A , c ) and T j ( A , c ) T_j(A,c) T j ( A , c ) decrease in c c c .
On a closed parameter interval [ u , v ] [u,v] [ u , v ] , define h τ h_\tau h τ by the finite sum
h τ = 1 2 ∑ direct shells { π 2 s 2 a τ L 2 ( π s m , v ) − π s m 2 U 1 ( π s m , u ) } + 1 2 ∑ dual shells { π 2 s 2 a τ T 2 ( π s m , v ) − π s m 2 T 1 ( π s m , u ) } − 10 − 10 . (44) h_\tau=\frac{1}{2}\sum_{\text{direct shells}}\left\{\pi^2s^2a_\tau L_2(\pi sm,v)-\pi sm_2U_1(\pi sm,u)\right\}
+\frac{1}{2}\sum_{\text{dual shells}}\left\{\pi^2s^2a_\tau T_2(\pi sm,v)-\pi sm_2T_1(\pi sm,u)\right\}-10^{-10}. \tag*{(44)} h τ = 2 1 direct shells ∑ { π 2 s 2 a τ L 2 ( π s m , v ) − π s m 2 U 1 ( π s m , u ) } + 2 1 dual shells ∑ { π 2 s 2 a τ T 2 ( π s m , v ) − π s m 2 T 1 ( π s m , u ) } − 1 0 − 10 . (44)
Thus μ τ ( c ) ≥ h τ \mu_\tau(c)\ge h_\tau μ τ ( c ) ≥ h τ throughout this interval. Set h = min ( h D , h O ) h=\min(h_{\mathrm{D}},h_{\mathrm{O}}) h = min ( h D , h O ) . To bound the cubic variation, enclose the three cubic coefficients using (37) and, in a direct summand,
K j ( π s m , c ) ∈ [ L j ( π s m , v ) , U j ( π s m , u ) ] , K_j(\pi sm,c)\in[L_j(\pi sm,v),U_j(\pi sm,u)], K j ( π s m , c ) ∈ [ L j ( π s m , v ) , U j ( π s m , u )] ,
and in a dual summand use [ T j ( π s m , v ) , T j ( π s m , u ) ] [T_j(\pi sm,v),T_j(\pi sm,u)] [ T j ( π s m , v ) , T j ( π s m , u )] . Summing the signed moment enclosures and adding [ − 10 − 6 , 10 − 6 ] [-10^{-6},10^{-6}] [ − 1 0 − 6 , 1 0 − 6 ] encloses each full cubic coefficient. The added interval covers the omitted shells: each test matrix has Frobenius norm cubed less than 15 15 15 , and 15 R 3 < 10 − 6 15R_3<10^{-6} 15 R 3 < 1 0 − 6 . Let b α , b β , b γ b_\alpha,b_\beta,b_\gamma b α , b β , b γ be the corresponding absolute upper bounds, and put
C 3 = 5 12 b α + 1 6 b β + 5 72 b γ . (45) C_3=\frac{5}{12}b_\alpha+\frac{1}{6}b_\beta+\frac{5}{72}b_\gamma. \tag*{(45)} C 3 = 12 5 b α + 6 1 b β + 72 5 b γ . (45)
Finally, define M 4 M_4 M 4 by the finite fourth derivative sum (38) , taking c = u c=u c = u , replacing each direct K j K_j K j by U j U_j U j , retaining each dual T j T_j T j exactly, and adding 10 − 6 10^{-6} 1 0 − 6 . The resulting bound holds on the whole shape ball and parameter interval.
Lemma 4.5. For C = L F C=L_{\mathrm{F}} C = L F , let 0 ≤ j ≤ 999 0\le j\le999 0 ≤ j ≤ 999 ; for C = L B C=L_{\mathrm{B}} C = L B , let 0 ≤ j ≤ 24 0\le j\le24 0 ≤ j ≤ 24 . There are bounds h h h , C 3 C_3 C 3 , and M 4 M_4 M 4 , uniform for c ∈ [ j / 100 , ( j + 1 ) / 100 ] c\in[j/100,(j+1)/100] c ∈ [ j /100 , ( j + 1 ) /100 ] , such that, for every symmetric traceless matrix H H H ,
d d t F κ ( e t H / 2 C ) ∣ t = 0 = 0 , \left.\frac{d}{dt}F_\kappa(e^{tH/2C})\right|_{t=0}=0, d t d F κ ( e t H /2 C ) t = 0 = 0 , d 2 d t 2 F κ ( e t H / 2 C ) ∣ t = 0 ≥ h ∥ H ∥ F 2 , \left.\frac{d^2}{dt^2}F_\kappa(e^{tH/2C})\right|_{t=0}\ge h\lVert H\rVert_{\mathrm{F}}^2, d t 2 d 2 F κ ( e t H /2 C ) t = 0 ≥ h ∥ H ∥ F 2 , ∣ d 3 d t 3 F κ ( e t H / 2 C ) ∣ t = 0 ∣ ≤ C 3 ∥ H ∥ F 3 , \left|\left.\frac{d^3}{dt^3}F_\kappa(e^{tH/2C})\right|_{t=0}\right|\le C_3\lVert H\rVert_{\mathrm{F}}^3, d t 3 d 3 F κ ( e t H /2 C ) t = 0 ≤ C 3 ∥ H ∥ F 3 , ∣ d 4 d t 4 F κ ( e ( X + t H ) / 2 C ) ∣ t = 0 ∣ ≤ M 4 ∥ H ∥ F 4 ( ∥ X ∥ F ≤ 1 / 40 ) . \left|\left.\frac{d^4}{dt^4}F_\kappa(e^{(X+tH)/2C})\right|_{t=0}\right|\le M_4\lVert H\rVert_{\mathrm{F}}^4\qquad(\lVert X\rVert_{\mathrm{F}}\le1/40). d t 4 d 4 F κ ( e ( X + t H ) /2 C ) t = 0 ≤ M 4 ∥ H ∥ F 4 (∥ X ∥ F ≤ 1/40 ) .
They satisfy
h 2 − C 3 240 − M 4 38400 > { 10 − 6 , C = L F , 10 − 3 , C = L B . (46) \frac{h}{2}-\frac{C_3}{240}-\frac{M_4}{38400}>\begin{cases}10^{-6},&C=L_F,\\10^{-3},&C=L_B.\end{cases} \tag*{(46)} 2 h − 240 C 3 − 38400 M 4 > { 1 0 − 6 , 1 0 − 3 , C = L F , C = L B . (46)
Proof. The preceding derivative formulas and truncation bounds give the four variation estimates, uniformly on each closed parameter interval and on the whole shape ball. The finite endpoint inequalities in Lemma A.2 of Appendix A are precisely (46) . □
Proof of Theorem 4.1. Taylor’s theorem and Lemma 4.5 give
F κ ( e X / 2 C ) − F κ ( C ) ≥ h 2 ∥ X ∥ F 2 − C 3 6 ∥ X ∥ F 3 − M 4 24 ∥ X ∥ F 4 ≥ ( h 2 − C 3 240 − M 4 38400 ) ∥ X ∥ F 2 . \begin{aligned}
F_\kappa(e^{X/2}C)-F_\kappa(C)\geq\frac{h}{2}\lVert X\rVert_F^2-\frac{C_3}{6}\lVert X\rVert_F^3-\frac{M_4}{24}\lVert X\rVert_F^4 \\
&\geq\left(\frac{h}{2}-\frac{C_3}{240}-\frac{M_4}{38400}\right)\lVert X\rVert_F^2.
\end{aligned} F κ ( e X /2 C ) − F κ ( C ) ≥ 2 h ∥ X ∥ F 2 − 6 C 3 ∥ X ∥ F 3 − 24 M 4 ∥ X ∥ F 4 ≥ ( 2 h − 240 C 3 − 38400 M 4 ) ∥ X ∥ F 2 .
This proves (27) and the part c ≤ 10 c\leq10 c ≤ 10 of (28) .
We next prove (29) . Put d = 3 / 100 d=3/100 d = 3/100 , s = κ a s=\kappa a s = κa , and write A − I = T = η I + S A-I=T=\eta I+S A − I = T = η I + S as in Lemma 4.1. In particular,
η ≤ 9 50 ∥ S ∥ F 2 , ∥ S ∥ F < 13 250 . (47) \eta\leq\frac{9}{50}\lVert S\rVert_F^2,\qquad\lVert S\rVert_F<\frac{13}{250}. \tag*{(47)} η ≤ 50 9 ∥ S ∥ F 2 , ∥ S ∥ F < 250 13 . (47)
Set f ( q ) = q − 1 / 2 e − κ q f(q)=q^{-1/2}e^{-\kappa\sqrt{q}} f ( q ) = q − 1/2 e − κ q and β = ( s + 1 ) / 2 \beta=(s+1)/2 β = ( s + 1 ) /2 . The identity ( log f ) ′ ′ ( q ) = κ / ( 4 q 3 / 2 ) + 1 / ( 2 q 2 ) > 0 (\log f)''(q)=\kappa/(4q^{3/2})+1/(2q^2)>0 ( log f ) ′′ ( q ) = κ / ( 4 q 3/2 ) + 1/ ( 2 q 2 ) > 0 gives
f ( a 2 ( 1 + b ) ) ≥ f ( a 2 ) e − β b ( b > − 1 ) . (48) f(a^2(1+b))\geq f(a^2)e^{-\beta b}\qquad(b>-1). \tag*{(48)} f ( a 2 ( 1 + b )) ≥ f ( a 2 ) e − β b ( b > − 1 ) . (48)
We first bound the possible loss from all higher shells. On a shell of squared radius q = a 2 n q=a^{2n} q = a 2 n , sign changes and permutations imply ∑ v v T = ( N n q / 3 ) I \sum vv^T=(N_nq/3)I ∑ v v T = ( N n q /3 ) I . Since f f f is convex,
∑ v i n s h e l l { f ( v T A v ) − f ( q ) } ≥ f ′ ( q ) ∑ v v T T v = η N n q f ′ ( q ) . \sum_{v\ \mathrm{in\ shell}}\{f(v^TAv)-f(q)\}\geq f'(q)\sum_v v^T Tv=\eta N_nqf'(q). v in shell ∑ { f ( v T A v ) − f ( q )} ≥ f ′ ( q ) v ∑ v T T v = η N n q f ′ ( q ) .
Moreover,
− q f ′ ( q ) = κ q + 1 2 f ( q ) ≤ β f ( a 2 ) e − s ( n − 1 ) , -qf'(q)=\frac{\kappa\sqrt{q}+1}{2}f(q)\leq\beta f(a^2)e^{-s(\sqrt{n}-1)}, − q f ′ ( q ) = 2 κ q + 1 f ( q ) ≤ β f ( a 2 ) e − s ( n − 1 ) ,
because ( s n + 1 ) / n ≤ s + 1 (s\sqrt{n}+1)/\sqrt{n}\leq s+1 ( s n + 1 ) / n ≤ s + 1 . Proposition 4.1 therefore bounds the total higher-shell loss, before the pair factor, by 3 β η f ( a 2 ) 3\beta\eta f(a^2) 3 β η f ( a 2 ) .
For the twelve unit contact directions w = v / a w=v/a w = v / a ,
∣ w T T w ∣ ≤ d , ∑ w w T T w = 12 η , |w^TTw|\leq d,\qquad\sum_w w^TTw=12\eta, ∣ w T T w ∣ ≤ d , w ∑ w T T w = 12 η ,
∑ w ( w T T w ) 2 = ∑ i T i i 2 + ( tr T ) 2 + 4 ∑ i < j T i j 2 ≥ ∥ T ∥ F 2 . \sum_w(w^TTw)^2=\sum_iT_{ii}^2+(\operatorname{tr}T)^2+4\sum_{i<j}T_{ij}^2\geq\lVert T\rVert_F^2. w ∑ ( w T T w ) 2 = i ∑ T ii 2 + ( tr T ) 2 + 4 i < j ∑ T ij 2 ≥ ∥ T ∥ F 2 .
Using e − β b ≥ 1 − β b + 1 2 β 2 e − β d b 2 e^{-\beta b}\geq1-\beta b+\frac{1}{2}\beta^2e^{-\beta d}b^2 e − β b ≥ 1 − β b + 2 1 β 2 e − β d b 2 in (48) gives the first-shell increase before the pair factor:
∑ w { f ( a 2 ( 1 + w T T w ) ) − f ( a 2 ) } ≥ f ( a 2 ) ( β 2 2 e − β d ∥ T ∥ F 2 − 12 β η ) . \sum_w\{f(a^2(1+w^TTw))-f(a^2)\}\geq f(a^2)\left(\frac{\beta^2}{2}e^{-\beta d}\lVert T\rVert_F^2-12\beta\eta\right). w ∑ { f ( a 2 ( 1 + w T T w )) − f ( a 2 )} ≥ f ( a 2 ) ( 2 β 2 e − β d ∥ T ∥ F 2 − 12 β η ) .
The higher shells can decrease the energy by at most 3 β η f ( a 2 ) 3\beta\eta f(a^2) 3 β η f ( a 2 ) . By (30) ,
15 η ≤ 6 ( 1 − d ) ∥ T ∥ F 2 = 250 97 ∥ T ∥ F 2 . 15\eta\leq\frac{6(1-d)}{ } \lVert T\rVert_F^2=\frac{250}{97}\lVert T\rVert_F^2. 15 η ≤ 6 ( 1 − d ) ∥ T ∥ F 2 = 97 250 ∥ T ∥ F 2 .
Thus the unhalved energy increase is bounded below by
β f ( a 2 ) ( β 2 e − 3 β / 100 − 250 97 ) ∥ T ∥ F 2 . (49) \beta f(a^2)\left(\frac{\beta}{2}e^{-3\beta/100}-\frac{250}{97}\right)\lVert T\rVert_F^2. \tag*{(49)} β f ( a 2 ) ( 2 β e − 3 β /100 − 97 250 ) ∥ T ∥ F 2 . (49)
For 13 / 2 ≤ β ≤ 64 13/2\leq\beta\leq64 13/2 ≤ β ≤ 64 , the function β e − 3 β / 100 \beta e^{-3\beta/100} β e − 3 β /100 attains its minimum at an endpoint. Its value at the left endpoint is at least 13 2 ( 1 − 39 / 200 ) = 2093 / 400 \frac{13}{2}(1-39/200)=2093/400 2 13 ( 1 − 39/200 ) = 2093/400 , while at the right endpoint it exceeds 8 8 8 , since e 48 / 25 < 8 e^{48/25}<8 e 48/25 < 8 . Thus
β 2 e − 3 β / 100 − 250 97 ≥ 2093 800 − 250 97 = 3021 77600 > 1 40 . \frac{\beta}{2}e^{-3\beta/100}-\frac{250}{97}\geq\frac{2093}{800}-\frac{250}{97}=\frac{3021}{77600}>\frac{1}{40}. 2 β e − 3 β /100 − 97 250 ≥ 800 2093 − 97 250 = 77600 3021 > 40 1 .
Restoring the pair factor proves (4.3) for this range of β \beta β .
For β ≥ 64 \beta\ge64 β ≥ 64 , the twelve contact values are η − S k k / 2 ± S i j \eta-S_{kk}/2\pm S_{ij} η − S k k /2 ± S ij , each repeated twice, with ( k , i , j ) = ( 1 , 2 , 3 ) , ( 2 , 1 , 3 ) , ( 3 , 1 , 2 ) (k,i,j)=(1,2,3),(2,1,3),(3,1,2) ( k , i , j ) = ( 1 , 2 , 3 ) , ( 2 , 1 , 3 ) , ( 3 , 1 , 2 ) . Hence their deformed energy is at least 12 f ( a 2 ) e − β η M 12f(a^2)e^{-\beta\eta M} 12 f ( a 2 ) e − β η M , where
M = 1 6 ∑ ( k , i , j ) ( e β S k k / 2 + β S i j + e β S k k / 2 − β S i j ) . M=\frac{1}{6}\sum_{(k,i,j)}\left(e^{\beta S_{kk}/2+\beta S_{ij}}+e^{\beta S_{kk}/2-\beta S_{ij}}\right). M = 6 1 ( k , i , j ) ∑ ( e β S k k /2 + β S ij + e β S k k /2 − β S ij ) .
The six exponents have mean zero and variance
σ 2 = β 2 12 ( ∑ i S i i 2 + 4 ∑ i < j S i j 2 ) ≥ β 2 12 ∥ S ∥ F 2 . (50) \sigma^2=\frac{\beta^2}{12}\left(\sum_i S_{ii}^2+4\sum_{i<j}S_{ij}^2\right)\ge\frac{\beta^2}{12}\lVert S\rVert_F^2. \tag*{(50)} σ 2 = 12 β 2 ( i ∑ S ii 2 + 4 i < j ∑ S ij 2 ) ≥ 12 β 2 ∥ S ∥ F 2 . (50)
Lemma 4.2 and (4.24) give
σ ≤ 1 : log M ≥ σ 2 13 ≥ β 2 156 ∥ S ∥ F 2 ≥ 9 25 β ∥ S ∥ F 2 ≥ 2 β η , \sigma\le1:\quad\log M\ge\frac{\sigma^2}{13}\ge\frac{\beta^2}{156}\lVert S\rVert_F^2\ge\frac{9}{25}\beta\lVert S\rVert_F^2\ge2\beta\eta, σ ≤ 1 : log M ≥ 13 σ 2 ≥ 156 β 2 ∥ S ∥ F 2 ≥ 25 9 β ∥ S ∥ F 2 ≥ 2 β η ,
σ ≥ 1 : log M ≥ σ 13 ≥ β 13 12 ∥ S ∥ F ≥ 9 25 β ∥ S ∥ F 2 ≥ 2 β η . \sigma\ge1:\quad\log M\ge\frac{\sigma}{13}\ge\frac{\beta}{13\sqrt{12}}\lVert S\rVert_F\ge\frac{9}{25}\beta\lVert S\rVert_F^2\ge2\beta\eta. σ ≥ 1 : log M ≥ 13 σ ≥ 13 12 β ∥ S ∥ F ≥ 25 9 β ∥ S ∥ F 2 ≥ 2 β η .
Here the first line uses β ≥ 64 \beta\ge64 β ≥ 64 and 64 / 156 > 9 / 25 64/156>9/25 64/156 > 9/25 ; the second uses ∥ S ∥ F < 13 / 250 \lVert S\rVert_F<13/250 ∥ S ∥ F < 13/250 and 1 / ( 13 12 ) > 1 / 52 > ( 9 / 25 ) ( 13 / 250 ) 1/(13\sqrt{12})>1/52>(9/25)(13/250) 1/ ( 13 12 ) > 1/52 > ( 9/25 ) ( 13/250 ) . The final inequalities are (4.21). The first-shell gain is therefore at least 12 f ( a 2 ) ( e β η − 1 ) ≥ 12 β η f ( a 2 ) 12f(a^2)(e^{\beta\eta}-1)\ge12\beta\eta f(a^2) 12 f ( a 2 ) ( e β η − 1 ) ≥ 12 β η f ( a 2 ) . After the higher-shell loss and the pair factor,
E κ ( A 1 / 2 L F ) − E κ ( L F ) ≥ 9 2 β η f ( a 2 ) ≥ 9 12 ( 1 + d ) β f ( a 2 ) ∥ T ∥ F 2 , E_\kappa(A^{1/2}L_F)-E_\kappa(L_F)\ge\frac{9}{2}\beta\eta f(a^2)\ge\frac{9}{12(1+d)}\beta f(a^2)\lVert T\rVert_F^2, E κ ( A 1/2 L F ) − E κ ( L F ) ≥ 2 9 β η f ( a 2 ) ≥ 12 ( 1 + d ) 9 β f ( a 2 ) ∥ T ∥ F 2 ,
which is stronger than (4.3). Absolute convergence justifies summing the shellwise tangent inequalities.
It remains to prove (4.2) when c > 10 c>10 c > 10 . Then s 2 = 2 7 / 3 π c > 156 > 144 s^2=2^{7/3}\pi c>156>144 s 2 = 2 7/3 π c > 156 > 144 . For traceless symmetric X X X , ∥ X ∥ o p ≤ 2 / 3 ∥ X ∥ F < 1 / 48 \lVert X\rVert_{\mathrm{op}}\le\sqrt{2/3}\lVert X\rVert_F<1/48 ∥ X ∥ op ≤ 2/3 ∥ X ∥ F < 1/48 , so
∥ e X − I ∥ o p < e 1 / 48 − 1 < 1 / 47 < 3 / 100. \lVert e^X-I\rVert_{\mathrm{op}}<e^{1/48}-1<1/47<3/100. ∥ e X − I ∥ op < e 1/48 − 1 < 1/47 < 3/100.
Comparison of the exponential and geometric series proves the middle inequality. Also, ∥ e X − I ∥ F 2 ≥ e − 1 / 20 ∥ X ∥ F 2 ≥ ( 4 / 5 ) ∥ X ∥ F 2 \lVert e^X-I\rVert_F^2\ge e^{-1/20}\lVert X\rVert_F^2\ge(4/5)\lVert X\rVert_F^2 ∥ e X − I ∥ F 2 ≥ e − 1/20 ∥ X ∥ F 2 ≥ ( 4/5 ) ∥ X ∥ F 2 , by applying the scalar mean value theorem to each eigenvalue. Thus (4.3), with A = e X A=e^X A = e X , gives
E κ ( e X / 2 L F ) − E κ ( L F ) ≥ s + 1 160 a e − s ∥ e X − I ∥ F 2 ≥ s + 1 200 a e − s ∥ X ∥ F 2 . \begin{aligned}
E_\kappa(e^{X/2}L_F)-E_\kappa(L_F)\ge\frac{s+1}{160a}e^{-s}\lVert e^X-I\rVert_F^2 \\
&\ge\frac{s+1}{200a}e^{-s}\lVert X\rVert_F^2.
\end{aligned} E κ ( e X /2 L F ) − E κ ( L F ) ≥ 160 a s + 1 e − s ∥ e X − I ∥ F 2 ≥ 200 a s + 1 e − s ∥ X ∥ F 2 .
Since s = κ a s=\kappa a s = κa , this is the remaining part of (4.2). All coefficients are positive, which also proves the asserted equality cases. □
Corollary 4.1. The comparisons (4.1) and (4.2) apply when A = A T > 0 A=A^T>0 A = A T > 0 , det A = 1 \det A=1 det A = 1 , and ∥ A − I ∥ o p ≤ 7 / 500 \lVert A-I\rVert_{\mathrm{op}}\le7/500 ∥ A − I ∥ op ≤ 7/500 , with deformation A 1 / 2 A^{1/2} A 1/2 of the corresponding reference lattice and the same parameter ranges.
Proof. If det A = 1 \det A=1 det A = 1 and ∥ A − I ∥ o p ≤ 7 / 500 \lVert A-I\rVert_{\mathrm{op}}\le7/500 ∥ A − I ∥ op ≤ 7/500 , then X = log A X=\log A X = log A is symmetric and traceless, and
∥ X ∥ F 2 ≤ 3 ( 7 / 493 ) 2 = 147 / 243049 < 1 / 1600. \lVert X\rVert_F^2\le3(7/493)^2=147/243049<1/1600. ∥ X ∥ F 2 ≤ 3 ( 7/493 ) 2 = 147/243049 < 1/1600. Here ∣ log ( 1 + t ) ∣ ≤ ∣ t ∣ / ( 1 − 7 / 500 ) \lvert\log(1+t)\rvert\le\lvert t\rvert/(1-7/500) ∣ log ( 1 + t )∣ ≤ ∣ t ∣ / ( 1 − 7/500 ) on the stated interval. □
Global comparison
We now compare all lattice shapes uniformly in the screening parameter. The following theorem is the global estimate used to prove Theorem 1.1.
Theorem 5.1. For every c ≥ 0 c \ge0 c ≥ 0 and every lattice L L L with ∣ L ∣ = 1 |L| = 1 ∣ L ∣ = 1 ,
P ( L , c ) ≥ min { P ( L F , c ) , P ( L B , c ) } . P(L,c) \ge\min\{P(L_{\mathrm{F}},c),P(L_{\mathrm{B}},c)\}. P ( L , c ) ≥ min { P ( L F , c ) , P ( L B , c )} . Equality is possible only if L L L is orthogonally equivalent to L F L_{\mathrm{F}} L F or L B L_{\mathrm{B}} L B .
The cusp estimate and the elementary reduction restrict the comparison to a fixed rectangle in five lattice coordinates. For 0 ≤ c ≤ 32 0 \le c \le32 0 ≤ c ≤ 32 , we subtract a cubic reference before expanding in the screening parameter; this retains the cancellation in the energy difference. For κ ≥ 20 \kappa\ge20 κ ≥ 20 , convexity extends a comparison at one screening parameter to the entire half-line. These two parameter ranges overlap.
Let Q Q Q be a rectangle in the reduced coordinates ξ = ( ℓ 1 , ℓ 2 , x 1 , x 2 , x 3 ) \xi= (\ell_1,\ell_2,x_1,x_2,x_3) ξ = ( ℓ 1 , ℓ 2 , x 1 , x 2 , x 3 ) , and denote its 32 corners by C ( Q ) C(Q) C ( Q ) . Choose ξ 0 ∈ Q \xi_0 \in Q ξ 0 ∈ Q and put Q = Y ( ξ 0 ) Q = Y(\xi_0) Q = Y ( ξ 0 ) and Y z = Y ( z ) Y_z = Y(z) Y z = Y ( z ) . For finite symmetric sets M I , M J ⊂ Z 3 ∖ { 0 } \mathcal{M}_{I},\mathcal{M}_{J} \subset\mathbb{Z}^3 \setminus\{0\} M I , M J ⊂ Z 3 ∖ { 0 } , define
p ( c ) = 1 2 ∑ m ∈ M I I ( Q [ m ] , c ) + 1 2 ∑ m ∈ M J J ( Q − 1 [ m ] , c ) , p(c) = \frac{1}{2}\sum_{m\in\mathcal{M}_{I}} I(Q[m],c) + \frac{1}{2}\sum_{m\in\mathcal{M}_{J}} J(Q^{-1}[m],c), p ( c ) = 2 1 m ∈ M I ∑ I ( Q [ m ] , c ) + 2 1 m ∈ M J ∑ J ( Q − 1 [ m ] , c ) ,
S ( c ) = − 1 2 ∑ m ∈ M I I q ( Q [ m ] , c ) m m T , T ( c ) = − 1 2 ∑ m ∈ M J J q ( Q − 1 [ m ] , c ) m m T . S(c) = -\frac{1}{2}\sum_{m\in\mathcal{M}_{I}} I_q(Q[m],c)mm^{T}, \qquad T(c) = -\frac{1}{2}\sum_{m\in\mathcal{M}_{J}} J_q(Q^{-1}[m],c)mm^{T}. S ( c ) = − 2 1 m ∈ M I ∑ I q ( Q [ m ] , c ) m m T , T ( c ) = − 2 1 m ∈ M J ∑ J q ( Q − 1 [ m ] , c ) m m T .
Both S ( c ) S(c) S ( c ) and T ( c ) T(c) T ( c ) are positive semidefinite. Let R R R be a reference lattice, and let p R ( c ) p_R(c) p R ( c ) be a finite positive Ewald sum satisfying
0 ≤ P ( R , c ) − p R ( c ) ≤ τ R ( u ≤ c ≤ u + C ) , (51) 0 \le P(R,c)-p_R(c) \le\tau_R \qquad(u \le c \le u+C), \tag*{(51)} 0 ≤ P ( R , c ) − p R ( c ) ≤ τ R ( u ≤ c ≤ u + C ) , (51)
where u ≥ 0 u \ge0 u ≥ 0 and C > 0 C > 0 C > 0 . At each corner set
F R , z ( c ) = p ( c ) − p R ( c ) + tr { S ( c ) ( Q − Y z ) + T ( c ) ( Q − 1 − Y z − 1 ) } . F_{R,z}(c) = p(c)-p_R(c)+\operatorname{tr}\{S(c)(Q-Y_z)+T(c)(Q^{-1}-Y_z^{-1})\}. F R , z ( c ) = p ( c ) − p R ( c ) + tr { S ( c ) ( Q − Y z ) + T ( c ) ( Q − 1 − Y z − 1 )} .
Lemma 5.1 (Comparison of screening profiles). Suppose that, for 0 ≤ s ≤ C 0 \le s \le C 0 ≤ s ≤ C and every z ∈ C ( Q ) z \in C(Q) z ∈ C ( Q ) ,
∣ F R , z ( u + s ) − ∑ n = 0 N a z , n s n ∣ ≤ r z s N + 1 . \left|F_{R,z}(u+s)-\sum_{n=0}^{N}a_{z,n}s^n\right| \le r_zs^{N+1}. F R , z ( u + s ) − n = 0 ∑ N a z , n s n ≤ r z s N + 1 . Define
b z , k = ∑ n = 0 k ( k n ) ( N n ) C n a z , n , 0 ≤ k ≤ N . b_{z,k}=\sum_{n=0}^{k}\frac{\binom{k}{n}}{\binom{N}{n}}C^n a_{z,n}, \qquad0 \le k \le N. b z , k = n = 0 ∑ k ( n N ) ( n k ) C n a z , n , 0 ≤ k ≤ N . Then, for ξ ∈ Q \xi\in Q ξ ∈ Q and u ≤ c ≤ u + C u \le c \le u+C u ≤ c ≤ u + C ,
P ( Y ( ξ ) , c ) − P ( R , c ) ≥ min z ∈ C ( Q ) 0 ≤ k ≤ N { b z , k − C N + 1 r z } − τ R . (52) P(Y(\xi),c)-P(R,c)\ge\min_{\substack{z\in C(Q)\\0\le k\le N}}\{b_{z,k}-C^{N+1}r_z\}-\tau_R. \tag*{(52)} P ( Y ( ξ ) , c ) − P ( R , c ) ≥ z ∈ C ( Q ) 0 ≤ k ≤ N min { b z , k − C N + 1 r z } − τ R . (52) When N = 0 N=0 N = 0 , the coefficient formula means b z , 0 = a z , 0 b_{z,0}=a_{z,0} b z , 0 = a z , 0 .
Proof. Convexity of the kernels in squared distance and positivity of the omitted competitor terms give
P ( Y , c ) ≥ p ( c ) + tr { S ( c ) ( Q − Y ) + T ( c ) ( Q − 1 − Y − 1 ) } . P(Y,c) \ge p(c)+\operatorname{tr}\{S(c)(Q-Y)+T(c)(Q^{-1}-Y^{-1})\}. P ( Y , c ) ≥ p ( c ) + tr { S ( c ) ( Q − Y ) + T ( c ) ( Q − 1 − Y − 1 )} .
By Lemma 3.3 , tr { S ( c ) Y + T ( c ) Y − 1 } \operatorname{tr}\{S(c)Y+T(c)Y^{-1}\} tr { S ( c ) Y + T ( c ) Y − 1 } is bounded above by its largest corner value. Hence (51) implies
P ( Y , c ) − P ( R , c ) ≥ min z F R , z ( c ) − τ R . P(Y,c)-P(R,c)\ge\min_{z}F_{R,z}(c)-\tau_R. P ( Y , c ) − P ( R , c ) ≥ z min F R , z ( c ) − τ R .
For 0 ≤ n ≤ N 0 \le n \le N 0 ≤ n ≤ N , the binomial theorem gives
t n = ∑ k = n N ( k n ) ( N n ) ( N k ) t k ( 1 − t ) N − k . t^n=\sum_{k=n}^{N}\frac{\binom{k}{n}}{\binom{N}{n}}\binom{N}{k}t^k(1-t)^{N-k}. t n = k = n ∑ N ( n N ) ( n k ) ( k N ) t k ( 1 − t ) N − k .
Consequently, with t = s / C ∈ [ 0 , 1 ] t=s/C \in[0,1] t = s / C ∈ [ 0 , 1 ] , the polynomial at corner z z z has the Bernstein representation
∑ n = 0 N a z , n s n = ∑ k = 0 N b z , k ( N k ) t k ( 1 − t ) N − k . \sum_{n=0}^{N} a_{z,n}s^n=\sum_{k=0}^{N}b_{z,k}\binom{N}{k}t^k(1-t)^{N-k}. n = 0 ∑ N a z , n s n = k = 0 ∑ N b z , k ( k N ) t k ( 1 − t ) N − k .
The basis functions are nonnegative and sum to one. The remainder bound therefore gives (52) . □
The Taylor expansion is applied to the difference F R , z F_{R,z} F R , z , rather than to separate upper and lower bounds for the two energies. Thus cancellation between competitor and reference derivatives is retained. The next lemma gives the coefficients and a uniform remainder on any finite screening interval.
Lemma 5.2 (Coefficients and remainders for the screening comparison). For q , α > 0 q,\alpha>0 q , α > 0 , u ≥ 0 u\geq0 u ≥ 0 , and integers j ≥ − 1 j\geq-1 j ≥ − 1 , n ≥ 0 n\geq0 n ≥ 0 , set
H j ( q , u ) = ∫ 1 ∞ t − j − 1 / 2 e − π q t − u / t d t , A n ( α ) = 1 n ! ∫ 1 ∞ t n e − α t d t . H_j(q,u)=\int_1^\infty t^{-j-1/2}e^{-\pi qt-u/t}\,\mathrm{d}t,\qquad A_n(\alpha)=\frac{1}{n!}\int_1^\infty t^n e^{-\alpha t}\,\mathrm{d}t. H j ( q , u ) = ∫ 1 ∞ t − j − 1/2 e − π q t − u / t d t , A n ( α ) = n ! 1 ∫ 1 ∞ t n e − α t d t . The Taylor coefficients in Lemma 5.1 are
a z , n = − p R ( n ) ( u ) n ! + ( − 1 ) n 2 n ! ∑ m ∈ M I [ H n ( Q [ m ] , u ) + π H n − 1 ( Q [ m ] , u ) m T ( Q − Y z ) m ] + ( − 1 ) n 2 ∑ m ∈ M J [ A n ( π Q − 1 [ m ] + u ) + π ( n + 1 ) A n + 1 ( π Q − 1 [ m ] + u ) m T ( Q − 1 − Y z − 1 ) m ] . \begin{aligned}
a_{z,n}={}&-\frac{p_R^{(n)}(u)}{n!}+\frac{(-1)^n}{2n!}\sum_{m\in\mathcal{M}_I}\left[H_n(Q[m],u)+\pi H_{n-1}(Q[m],u)m^T(Q-Y_z)m\right]\\
&+\frac{(-1)^n}{2}\sum_{m\in\mathcal{M}_J}\left[A_n(\pi Q^{-1}[m]+u)\right.\\
&\qquad\left.+\pi(n+1)A_{n+1}(\pi Q^{-1}[m]+u)m^T(Q^{-1}-Y_z^{-1})m\right].
\end{aligned} a z , n = − n ! p R ( n ) ( u ) + 2 n ! ( − 1 ) n m ∈ M I ∑ [ H n ( Q [ m ] , u ) + π H n − 1 ( Q [ m ] , u ) m T ( Q − Y z ) m ] + 2 ( − 1 ) n m ∈ M J ∑ [ A n ( π Q − 1 [ m ] + u ) + π ( n + 1 ) A n + 1 ( π Q − 1 [ m ] + u ) m T ( Q − 1 − Y z − 1 ) m ] . A remainder constant valid on every finite interval 0 ≤ s ≤ C 0\leq s\leq C 0 ≤ s ≤ C is
r z = ( − 1 ) N + 1 p R ( N + 1 ) ( u ) ( N + 1 ) ! + 1 2 ( N + 1 ) ! ∑ m ∈ M I H N + 1 ( Q [ m ] , u ) + 1 2 ∑ m ∈ M J A N + 1 ( π Q − 1 [ m ] + u ) + π 2 ( N + 1 ) ! ∑ i , j = 1 3 ∑ m ∈ M I H N ( Q [ m ] , u ) ∣ m i m j ( Q − Y z ) i j ∣ + π ( N + 2 ) 2 ∑ i , j = 1 3 ∑ m ∈ M J A N + 2 ( π Q − 1 [ m ] + u ) ∣ m i m j ( Q − 1 − Y z − 1 ) i j ∣ . (53) \begin{aligned}
r_z={}&\frac{(-1)^{N+1}p_R^{(N+1)}(u)}{(N+1)!}+\frac{1}{2(N+1)!}\sum_{m\in\mathcal{M}_I}H_{N+1}(Q[m],u)\\
&+\frac{1}{2}\sum_{m\in\mathcal{M}_J}A_{N+1}(\pi Q^{-1}[m]+u)\\
&+\frac{\pi}{2(N+1)!}\sum_{i,j=1}^{3}\sum_{m\in\mathcal{M}_I}H_N(Q[m],u)\lvert m_i m_j(Q-Y_z)_{ij}\rvert\\
&+\frac{\pi(N+2)}{2}\sum_{i,j=1}^{3}\sum_{m\in\mathcal{M}_J}A_{N+2}(\pi Q^{-1}[m]+u)\lvert m_i m_j(Q^{-1}-Y_z^{-1})_{ij}\rvert. \tag*{(53)}
\end{aligned} r z = ( N + 1 )! ( − 1 ) N + 1 p R ( N + 1 ) ( u ) + 2 ( N + 1 )! 1 m ∈ M I ∑ H N + 1 ( Q [ m ] , u ) + 2 1 m ∈ M J ∑ A N + 1 ( π Q − 1 [ m ] + u ) + 2 ( N + 1 )! π i , j = 1 ∑ 3 m ∈ M I ∑ H N ( Q [ m ] , u ) ∣ m i m j ( Q − Y z ) ij ∣ + 2 π ( N + 2 ) i , j = 1 ∑ 3 m ∈ M J ∑ A N + 2 ( π Q − 1 [ m ] + u ) ∣ m i m j ( Q − 1 − Y z − 1 ) ij ∣ . (53)
Proof. Differentiation under the defining integrals gives, in particular,
− ∂ c n I q ( q , u ) n ! = π ( − 1 ) n n ! H n − 1 ( q , u ) , -\frac{\partial_c^n I_q(q,u)}{n!}=\frac{\pi(-1)^n}{n!}H_{n-1}(q,u), − n ! ∂ c n I q ( q , u ) = n ! π ( − 1 ) n H n − 1 ( q , u ) ,
− ∂ c n J q ( q , u ) n ! = π ( − 1 ) n ( n + 1 ) A n + 1 ( π q + u ) . -\frac{\partial_c^n J_q(q,u)}{n!}=\pi(-1)^n(n+1)A_{n+1}(\pi q+u). − n ! ∂ c n J q ( q , u ) = π ( − 1 ) n ( n + 1 ) A n + 1 ( π q + u ) .
The factor n + 1 n+1 n + 1 in the reciprocal tangent term is the factorial ratio ( n + 1 ) ! / n ! (n+1)!/n! ( n + 1 )! / n ! . The coefficient of s n s^n s n in e − s v e^{-sv} e − s v is ( − v ) n / n ! (-v)^n/n! ( − v ) n / n ! , and, for s , v ≥ 0 s,v\geq0 s , v ≥ 0 , its absolute remainder after degree N N N is at most s N + 1 v N + 1 / ( N + 1 ) ! s^{N+1}v^{N+1}/(N+1)! s N + 1 v N + 1 / ( N + 1 )! . Apply this with v = 1 / t v=1/t v = 1/ t in the direct kernel and v = t v=t v = t in the reciprocal kernel, including their squared-length derivatives. Substitution into F R , z F_{R,z} F R , z gives the displayed coefficients and the first three terms of the remainder bound. The reference contribution is nonnegative because p R p_R p R is completely monotone in the screening variable.
For a symmetric matrix D D D and nonnegative weights w m w_m w m ,
∑ m w m ∣ m T D m ∣ ≤ ∑ i , j = 1 3 ∑ m w m ∣ m i m j D i j ∣ . \sum_m w_m\lvert m^T Dm\rvert\leq\sum_{i,j=1}^{3}\sum_m w_m\lvert m_i m_jD_{ij}\rvert. m ∑ w m ∣ m T D m ∣ ≤ i , j = 1 ∑ 3 m ∑ w m ∣ m i m j D ij ∣ .
Apply this to D = Q − Y z D=Q-Y_z D = Q − Y z and D = Q − 1 − Y z − 1 D=Q^{-1}-Y_z^{-1} D = Q − 1 − Y z − 1 , using the corresponding derivative remainders as weights. This gives the last two terms of (5.3). Absolute values are taken before summing the rank-one contributions, so no cancellation of matrix entries is used to bound a remainder. □
The moments can be evaluated by integration by parts. At u = 0 u=0 u = 0 ,
H 0 ( q , 0 ) = erfc ( π q ) q , H − 1 ( q , 0 ) = e − π q + H 0 ( q , 0 ) / 2 π q , H_0(q,0)=\frac{\operatorname{erfc}(\sqrt{\pi q})}{\sqrt{q}},
\qquad
H_{-1}(q,0)=\frac{e^{-\pi q}+H_0(q,0)/2}{\pi q}, H 0 ( q , 0 ) = q erfc ( π q ) , H − 1 ( q , 0 ) = π q e − π q + H 0 ( q , 0 ) /2 ,
H n ( q , 0 ) = e − π q − π q H n − 1 ( q , 0 ) n − 1 / 2 ( n ≥ 1 ) . H_n(q,0)=\frac{e^{-\pi q}-\pi qH_{n-1}(q,0)}{n-1/2}
\qquad(n\ge1). H n ( q , 0 ) = n − 1/2 e − π q − π q H n − 1 ( q , 0 ) ( n ≥ 1 ) .
For u > 0 u>0 u > 0 , use H 0 ( q , u ) = I ( q , u ) H_0(q,u)=I(q,u) H 0 ( q , u ) = I ( q , u ) and
H 1 ( q , u ) = π 2 u [ e − 2 π q u erfc ( π q − u ) − e 2 π q u erfc ( π q + u ) ] . H_1(q,u)=\frac{\sqrt{\pi}}{2\sqrt{u}}\left[e^{-2\sqrt{\pi q u}}\operatorname{erfc}(\sqrt{\pi q}-\sqrt{u})-e^{2\sqrt{\pi q u}}\operatorname{erfc}(\sqrt{\pi q}+\sqrt{u})\right]. H 1 ( q , u ) = 2 u π [ e − 2 π q u erfc ( π q − u ) − e 2 π q u erfc ( π q + u ) ] .
The remaining moments satisfy
H − 1 ( q , u ) = e − π q − u + H 0 ( q , u ) / 2 + u H 1 ( q , u ) π q , H_{-1}(q,u)=\frac{e^{-\pi q-u}+H_0(q,u)/2+uH_1(q,u)}{\pi q}, H − 1 ( q , u ) = π q e − π q − u + H 0 ( q , u ) /2 + u H 1 ( q , u ) ,
H j + 1 ( q , u ) = π q H j − 1 ( q , u ) + ( j − 1 / 2 ) H j ( q , u ) − e − π q − u u ( j ≥ 1 ) . H_{j+1}(q,u)=\frac{\pi qH_{j-1}(q,u)+(j-1/2)H_j(q,u)-e^{-\pi q-u}}{u}
\qquad(j\ge1). H j + 1 ( q , u ) = u π q H j − 1 ( q , u ) + ( j − 1/2 ) H j ( q , u ) − e − π q − u ( j ≥ 1 ) .
For α > 0 \alpha>0 α > 0 , the reciprocal moments satisfy
A 0 ( α ) = e − α α , A n ( α ) = A n − 1 ( α ) + e − α / n ! α ( n ≥ 1 ) . A_0(\alpha)=\frac{e^{-\alpha}}{\alpha},
\qquad
A_n(\alpha)=\frac{A_{n-1}(\alpha)+e^{-\alpha}/n!}{\alpha}
\qquad(n\ge1). A 0 ( α ) = α e − α , A n ( α ) = α A n − 1 ( α ) + e − α / n ! ( n ≥ 1 ) .
For the reference sum, the FCC squared lengths are 2 1 / 3 n 2^{1/3}n 2 1/3 n , where N n = # { m ∈ Z 3 : ∣ m ∣ 2 = 2 n } N_n=\#\{m\in\mathbb{Z}^3:|m|^2=2n\} N n = # { m ∈ Z 3 : ∣ m ∣ 2 = 2 n } , and the BCC squared lengths are 2 − 4 / 3 ℓ 2^{-4/3}\ell 2 − 4/3 ℓ , where M ℓ M_\ell M ℓ counts integer triples of squared norm ℓ \ell ℓ whose coordinates have the same parity. We use the shells n ≤ 48 n\le48 n ≤ 48 and ℓ ≤ 192 \ell\le192 ℓ ≤ 192 in the finite reference sum p R p_R p R . They are completely enumerated in [ − 14 , 14 ] 3 [-14,14]^3 [ − 14 , 14 ] 3 . A uniform bound for the omitted reference shells is
τ = 15 e − 49 π 2 1 / 3 π 2 1 / 3 ( 1 − e − π 2 1 / 3 ) + 15 e − 193 π 2 − 4 / 3 π 2 − 4 / 3 ( 1 − e − π 2 − 4 / 3 ) . (54) \tau=\frac{15e^{-49\pi2^{1/3}}}{\pi2^{1/3}(1-e^{-\pi2^{1/3}})}
+\frac{15e^{-193\pi2^{-4/3}}}{\pi2^{-4/3}(1-e^{-\pi2^{-4/3}})}. \tag*{(54)} τ = π 2 1/3 ( 1 − e − π 2 1/3 ) 15 e − 49 π 2 1/3 + π 2 − 4/3 ( 1 − e − π 2 − 4/3 ) 15 e − 193 π 2 − 4/3 . (54)
For c ≥ 0 c\ge0 c ≥ 0 , both kernels satisfy I ( q , c ) , J ( q , c ) ≤ e − π q / ( π q ) I(q,c),J(q,c)\le e^{-\pi q}/(\pi q) I ( q , c ) , J ( q , c ) ≤ e − π q / ( π q ) . The shell bound N n < 30 n N_n<30n N n < 30 n was proved in Proposition 4.1; likewise, fixing two coordinates gives M ℓ ≤ 2 ( 2 ℓ + 1 ) 2 ≤ 18 ℓ < 30 ℓ M_\ell\le2(2\sqrt{\ell}+1)^2\le18\ell<30\ell M ℓ ≤ 2 ( 2 ℓ + 1 ) 2 ≤ 18 ℓ < 30 ℓ . Since R R R and R ∗ R^* R ∗ are one FCC and one BCC lattice, the factor 1 / 2 1/2 1/2 in each Ewald sum and geometric summation give 0 ≤ P ( R , c ) − p R ( c ) ≤ τ 0\le P(R,c)-p_R(c)\le\tau 0 ≤ P ( R , c ) − p R ( c ) ≤ τ , uniformly for R ∈ { L F , L B } R\in\{L_F,L_B\} R ∈ { L F , L B } and c ≥ 0 c\ge0 c ≥ 0 . Thus (5.1) holds with τ R = τ \tau_R=\tau τ R = τ .
For a shape rectangle Q Q Q and J = [ u , u + C ] J=[u,u+C] J = [ u , u + C ] , the explicit sufficient condition for a strict comparison on Q × J Q\times J Q × J is
min z ∈ C ( Q ) 0 ≤ k ≤ N { b z , k − C N + 1 r z } > τ . (55) \min_{\substack{z\in\mathcal{C}(Q)\\0\le k\le N}}\{b_{z,k}-C^{N+1}r_z\}>\tau. \tag*{(55)} z ∈ C ( Q ) 0 ≤ k ≤ N min { b z , k − C N + 1 r z } > τ . (55)
The finite competitor sets are fixed on each rectangle, and positivity allows the omitted terms to be discarded. For the direct and reciprocal sums, the sets are chosen independently from { − 3 , − 2 , … , 3 } 3 ∖ { 0 } \{-3,-2,\ldots,3\}^3\setminus\{0\} { − 3 , − 2 , … , 3 } 3 ∖ { 0 } . A vector is retained only when its squared length q q q , evaluated at the centre of the rectangle, satisfies π q < 60 \pi q<60 π q < 60 . One representative of each pair { m , − m } \{m,-m\} { m , − m } may be used if the factor 1 / 2 1/2 1/2 is removed.
For the local comparisons we use the reduced reference matrices
Y F = 2 1 / 3 ( 1 1 / 2 1 / 2 1 / 2 1 1 / 2 1 / 2 1 / 2 1 ) , Y B = 2 − 4 / 3 ( 3 1 2 1 3 2 2 2 4 ) . (56) Y_F=2^{1/3}
\begin{pmatrix}
1 & 1/2 & 1/2\\
1/2 & 1 & 1/2\\
1/2 & 1/2 & 1
\end{pmatrix},
\qquad
Y_B=2^{-4/3}
\begin{pmatrix}
3 & 1 & 2\\
1 & 3 & 2\\
2 & 2 & 4
\end{pmatrix}. \tag*{(56)} Y F = 2 1/3 1 1/2 1/2 1/2 1 1/2 1/2 1/2 1 , Y B = 2 − 4/3 3 1 2 1 3 2 2 2 4 . (56)
The reference may be chosen separately on each rectangle. Indeed, P ( L , c ) > P ( R , c ) P(L,c)>P(R,c) P ( L , c ) > P ( R , c ) for either R ∈ { L F , L B } R\in\{L_F,L_B\} R ∈ { L F , L B } implies P ( L , c ) > min { P ( L F , c ) , P ( L B , c ) } P(L,c)>\min\{P(L_F,c),P(L_B,c)\} P ( L , c ) > min { P ( L F , c ) , P ( L B , c )} . The same observation applies to each local comparison on its stated parameter range.
Lemma 5.3 (Comparison on the bounded screening range). For 0 ≤ c ≤ 32 0 \le c \le32 0 ≤ c ≤ 32 and every lattice L L L with ∣ L ∣ = 1 |L| = 1 ∣ L ∣ = 1 ,
P ( L , c ) ≥ min { P ( L F , c ) , P ( L B , c ) } , P(L,c) \ge\min\{P(L_{\mathrm{F}},c),P(L_{\mathrm{B}},c)\}, P ( L , c ) ≥ min { P ( L F , c ) , P ( L B , c )} , and equality implies that L L L is orthogonally equivalent to L F L_{\mathrm{F}} L F or L B L_{\mathrm{B}} L B .
Proof. Lemma 3.2 treats the cusp. Every remaining reduced representative has logarithmic coordinates in D \mathcal{D} D . Fix c c c and choose an interval J j J_j J j containing it and a rectangle Q \mathcal{Q} Q containing the reduced coordinates. Apply Lemma A.3.
Each right-hand side of (24) is separately convex. Thus alternative (i), which concerns one fixed right-hand side at every corner, contradicts that reduction inequality throughout Q \mathcal{Q} Q . Alternative (ii) lies in the excluded cusp, since d 1 = λ 1 ( L ) 2 d_1 = \lambda_1(L)^2 d 1 = λ 1 ( L ) 2 for a reduced representative. Alternative (iii) gives a strict comparison with R R R by Lemma 5.1, and hence with the smaller reference energy.
In alternative (iv), separate matrix convexity extends (A.5) to the whole rectangle. Choose g R g_R g R with g R T g R = Y R g_R^{\mathsf T}g_R = Y_R g R T g R = Y R and put A = g R − T Y g R − 1 A = g_R^{-\mathsf T}Yg_R^{-1} A = g R − T Y g R − 1 . The basis A 1 / 2 g R A^{1/2}g_R A 1/2 g R has Gram matrix Y Y Y , so A 1 / 2 R A^{1/2}R A 1/2 R is orthogonally equivalent to L L L . Since A − 1 = g R Y − 1 g R T A^{-1} = g_RY^{-1}g_R^{\mathsf T} A − 1 = g R Y − 1 g R T , congruence transforms the direct and inverse matrix bounds into
A ≺ ( 1 + δ ) I , A − 1 ≺ ( 1 − δ ) − 1 I . A \prec(1+\delta)I,\qquad A^{-1} \prec(1-\delta)^{-1}I. A ≺ ( 1 + δ ) I , A − 1 ≺ ( 1 − δ ) − 1 I .
Inverting the second inequality yields ( 1 − δ ) I ≺ A (1-\delta)I \prec A ( 1 − δ ) I ≺ A . Thus every eigenvalue λ \lambda λ of A A A satisfies ∣ log λ ∣ ≤ ∣ λ − 1 ∣ / ( 1 − δ ) |\log\lambda| \le|\lambda-1|/(1-\delta) ∣ log λ ∣ ≤ ∣ λ − 1∣/ ( 1 − δ ) . Since det A = det Y / det Y R = 1 \det A = \det Y/\det Y_R = 1 det A = det Y / det Y R = 1 , we obtain
det A = 1 , ∥ A − I ∥ o p < δ , ∥ log A ∥ F ≤ 3 δ 1 − δ . \det A = 1,\qquad\|A-I\|_{\mathrm{op}} < \delta,\qquad\|\log A\|_{\mathrm{F}} \le\frac{\sqrt{3}\delta}{1-\delta}. det A = 1 , ∥ A − I ∥ op < δ , ∥ log A ∥ F ≤ 1 − δ 3 δ .
If δ ≤ 7 / 500 \delta\le7/500 δ ≤ 7/500 , the last bound is less than 1 / 40 1/40 1/40 , since 147 / 243049 < 1 / 1600 147/243049 < 1/1600 147/243049 < 1/1600 . Theorem 4.1 applies. All BCC comparisons are restricted to 0 ≤ c ≤ 1 / 4 0 \le c \le1/4 0 ≤ c ≤ 1/4 . For j = 8 , 9 , 10 j = 8,9,10 j = 8 , 9 , 10 , c ≥ 12 c \ge12 c ≥ 12 implies κ a > 12 \kappa_a > 12 κ a > 12 , so the FCC operator radius 3 / 100 3/100 3/100 follows from the same theorem. Every lattice in Q \mathcal{Q} Q different from its local reference has strictly greater energy than that reference. These alternatives prove the comparison and its equality statement. □
We turn to the unbounded screening range. Normalize the physical energy by
Φ κ ( L ) = 2 a e κ a E κ ( L ) = ∑ v ∈ L ∖ { 0 } a ∣ v ∣ e κ ( a − ∣ v ∣ ) , a = 2 1 / 6 . \Phi_\kappa(L) = 2ae^{\kappa a}E_\kappa(L) = \sum_{v\in L\setminus\{0\}}\frac{a}{|v|}e^{\kappa(a-|v|)},\qquad a=2^{1/6}. Φ κ ( L ) = 2 a e κa E κ ( L ) = v ∈ L ∖ { 0 } ∑ ∣ v ∣ a e κ ( a − ∣ v ∣ ) , a = 2 1/6 .
For a finite set M ⊂ Z 3 ∖ { 0 } \mathcal{M} \subset\mathbb{Z}^3 \setminus\{0\} M ⊂ Z 3 ∖ { 0 } and a positive definite Gram matrix Y Y Y , put
V κ ( Y ) = ∑ m ∈ M a Y [ m ] e κ ( a − Y [ m ] ) . V_\kappa(Y) = \sum_{m\in\mathcal{M}}\frac{a}{\sqrt{Y[m]}}e^{\kappa(a-\sqrt{Y[m]})}. V κ ( Y ) = m ∈ M ∑ Y [ m ] a e κ ( a − Y [ m ] ) .
Lemma 5.4 (Continuation along a screening half-line). Let K > 0 K>0 K > 0 . If
V K ( Y ) > Φ K ( L F ) , ∂ κ V κ ( Y ) ∣ κ = K ≥ 0 , (57) V_K(Y) > \Phi_K(L_{\mathrm{F}}),\qquad\left.\partial_\kappa V_\kappa(Y)\right|_{\kappa=K} \ge0, \tag*{(57)} V K ( Y ) > Φ K ( L F ) , ∂ κ V κ ( Y ) ∣ κ = K ≥ 0 , (57) then E κ ( L ) > E κ ( L F ) E_\kappa(L) > E_\kappa(L_{\mathrm{F}}) E κ ( L ) > E κ ( L F ) for every κ ≥ K \kappa\ge K κ ≥ K , where Y Y Y is a Gram matrix of L L L .
Proof. Each summand of V κ ( Y ) V_\kappa(Y) V κ ( Y ) is convex in κ \kappa κ , so the second inequality makes V κ ( Y ) V_\kappa(Y) V κ ( Y ) nondecreasing for κ ≥ K \kappa\ge K κ ≥ K . Every nonzero FCC vector has length at least a a a , and therefore Φ κ ( L F ) \Phi_\kappa(L_{\mathrm{F}}) Φ κ ( L F ) is nonincreasing. Positivity of the omitted terms now gives
Φ κ ( L ) ≥ V κ ( Y ) ≥ V K ( Y ) > Φ K ( L F ) ≥ Φ κ ( L F ) . \Phi_\kappa(L) \ge V_\kappa(Y) \ge V_K(Y) > \Phi_K(L_{\mathrm{F}}) \ge\Phi_\kappa(L_{\mathrm{F}}). Φ κ ( L ) ≥ V κ ( Y ) ≥ V K ( Y ) > Φ K ( L F ) ≥ Φ κ ( L F ) .
Division by 2 a e κ a > 0 2ae^{\kappa a}>0 2 a e κa > 0 proves the claim. □
We next derive the uniform bounds used to apply this principle. Fix K = 20 K=20 K = 20 and a shape rectangle Q \mathcal{Q} Q . For each nonzero integer vector m m m , write q m ( ξ ) = Y ( ξ ) [ m ] q_m(\xi)=Y(\xi)[m] q m ( ξ ) = Y ( ξ ) [ m ] . In the coordinates of (22) ,
q m = d 1 ( m 1 + x 1 m 2 + x 2 m 3 ) 2 + d 2 ( m 2 + x 3 m 3 ) 2 + d 3 m 3 2 . q_m=d_1(m_1+x_1m_2+x_2m_3)^2+d_2(m_2+x_3m_3)^2+d_3m_3^2. q m = d 1 ( m 1 + x 1 m 2 + x 2 m 3 ) 2 + d 2 ( m 2 + x 3 m 3 ) 2 + d 3 m 3 2 .
Let [ x i ] [x_i] [ x i ] be the shear intervals of Q \mathcal{Q} Q , and put d i − = min z ∈ C ( Q ) d i ( z ) d_i^-=\min_{z\in C(\mathcal{Q})}d_i(z) d i − = min z ∈ C ( Q ) d i ( z ) . Then
q ‾ m = d 1 − 1 min x 1 ∈ [ x 1 ] x 2 ∈ [ x 2 ] ( m 1 + x 1 m 2 + x 2 m 3 ) 2 + d 2 − 1 min x 3 ∈ [ x 3 ] ( m 2 + x 3 m 3 ) 2 + d 3 − 1 m 3 2 , (58) \overline{q}_m=d_1^{-1}\min_{\substack{x_1\in[x_1]\\x_2\in[x_2]}}(m_1+x_1m_2+x_2m_3)^2+d_2^{-1}\min_{x_3\in[x_3]}(m_2+x_3m_3)^2+d_3^{-1}m_3^2, \tag*{(58)} q m = d 1 − 1 x 1 ∈ [ x 1 ] x 2 ∈ [ x 2 ] min ( m 1 + x 1 m 2 + x 2 m 3 ) 2 + d 2 − 1 x 3 ∈ [ x 3 ] min ( m 2 + x 3 m 3 ) 2 + d 3 − 1 m 3 2 , (58)
0 < q m ≤ q m ( ξ ) ≤ q ‾ m : = max z ∈ C ( Q ) q m ( z ) . 0<q_m\leq q_m(\xi)\leq\overline{q}_m:=\max_{z\in C(\mathcal{Q})}q_m(z). 0 < q m ≤ q m ( ξ ) ≤ q m := z ∈ C ( Q ) max q m ( z ) .
The lower bound follows term by term and is positive because the last nonzero coordinate of m m m contributes a positive constant square. The extrema of the d i d_i d i occur at corners. The upper bound follows from separate convexity, which also gives
tr ( S Y ( ξ ) ) ≤ max z ∈ C ( Q ) tr ( S Y z ) ( S ⪰ 0 , ξ ∈ Q ) . (59) \operatorname{tr}(SY(\xi))\leq\max_{z\in C(\mathcal{Q})}\operatorname{tr}(SY_z)\qquad(S\succeq0,\ \xi\in\mathcal{Q}). \tag*{(59)} tr ( S Y ( ξ )) ≤ z ∈ C ( Q ) max tr ( S Y z ) ( S ⪰ 0 , ξ ∈ Q ) . (59)
Set
f ( q ) = a q e K ( a − q ) , h ( q ) = ( a − q ) a q e K ( a − q ) . f(q)=\frac{a}{\sqrt{q}}e^{K(a-\sqrt{q})},\qquad h(q)=(a-\sqrt{q})\frac{a}{\sqrt{q}}e^{K(a-\sqrt{q})}. f ( q ) = q a e K ( a − q ) , h ( q ) = ( a − q ) q a e K ( a − q ) .
A direct differentiation gives
− f ′ ( q ) = f ( q ) ( K q + 1 ) 2 q > 0 , f ′ ′ ( q ) = f ( q ) 4 q 2 ( K 2 q + 3 K q + 3 ) > 0. -f'(q)=\frac{f(q)(K\sqrt{q}+1)}{2q}>0,\qquad f''(q)=\frac{f(q)}{4q^2}(K^2q+3K\sqrt{q}+3)>0. − f ′ ( q ) = 2 q f ( q ) ( K q + 1 ) > 0 , f ′′ ( q ) = 4 q 2 f ( q ) ( K 2 q + 3 K q + 3 ) > 0.
At the midpoint ξ 0 \xi_0 ξ 0 of Q \mathcal{Q} Q , put Q = Y ( ξ 0 ) Q=Y(\xi_0) Q = Y ( ξ 0 ) and S f = ∑ m ∈ M ( − f ′ ( Q [ m ] ) ) m m T ⪰ 0 S_f=\sum_{m\in\mathcal{M}}(-f'(Q[m]))mm^T\succeq0 S f = ∑ m ∈ M ( − f ′ ( Q [ m ])) m m T ⪰ 0 . The tangent inequality and (59) yield
V K ( Y ( ξ ) ) ≥ ∑ m ∈ M f ( Q [ m ] ) + tr ( S f Q ) − max z ∈ C ( Q ) tr ( S f Y z ) . (60) V_K(Y(\xi))\geq\sum_{m\in\mathcal{M}}f(Q[m])+\operatorname{tr}(S_fQ)-\max_{z\in C(\mathcal{Q})}\operatorname{tr}(S_fY_z). \tag*{(60)} V K ( Y ( ξ )) ≥ m ∈ M ∑ f ( Q [ m ]) + tr ( S f Q ) − z ∈ C ( Q ) max tr ( S f Y z ) . (60)
There are three useful lower bounds for the screening derivative. First, put
q ∗ = 1 4 ( a + a 2 + 4 a / K ) 2 . q_*=\frac{1}{4}\left(a+\sqrt{a^2+4a/K}\right)^2. q ∗ = 4 1 ( a + a 2 + 4 a / K ) 2 .
Since
d d r ( ( a − r ) a r e K ( a − r ) ) = a e K ( a − r ) r 2 ( K r 2 − K a r − a ) , \frac{d}{dr}\left((a-r)\frac{a}{r}e^{K(a-r)}\right)=\frac{ae^{K(a-r)}}{r^2}(Kr^2-Kar-a), d r d ( ( a − r ) r a e K ( a − r ) ) = r 2 a e K ( a − r ) ( K r 2 − K a r − a ) ,
h h h decreases up to q ∗ q_* q ∗ and increases thereafter. Thus
∂ κ V κ ( Y ( ξ ) ) ∣ κ = K ≥ ∑ m ∈ M min q ∈ [ q ‾ m , q ‾ m ] h ( q ) . (61) \left.\partial_\kappa V_\kappa(Y(\xi))\right|_{\kappa=K}\geq\sum_{m\in\mathcal{M}}\min_{q\in[\underline{q}_m,\overline{q}_m]}h(q). \tag*{(61)} ∂ κ V κ ( Y ( ξ )) ∣ κ = K ≥ m ∈ M ∑ q ∈ [ q m , q m ] min h ( q ) . (61)
Each minimum is attained at an endpoint or at q ∗ q_* q ∗ when the interval contains it. Second, if q ‾ m < a 2 \overline{q}_m<a^2 q m < a 2 for every retained vector, then every derivative term is positive.
Third, suppose q ‾ m < q ∗ \overline{q}_m<q_* q m < q ∗ for every retained vector. On ( 0 , q ∗ ) (0,q_*) ( 0 , q ∗ ) , h h h is decreasing and convex. Indeed,
− h ′ ( q ) = f ( q ) ( a K q + a − K q ) 2 q > 0 , -h'(q)=\frac{f(q)(aK\sqrt{q}+a-Kq)}{2q}>0, − h ′ ( q ) = 2 q f ( q ) ( a K q + a − K q ) > 0 ,
and, writing r = q r=\sqrt{q} r = q ,
4 q 2 h ′ ′ ( q ) f ( q ) = ( a − r ) ( K 2 r 2 + 3 K r + 3 ) + 2 K r 2 + 3 r . \frac{4q^2h''(q)}{f(q)}=(a-r)(K^2r^2+3Kr+3)+2Kr^2+3r. f ( q ) 4 q 2 h ′′ ( q ) = ( a − r ) ( K 2 r 2 + 3 K r + 3 ) + 2 K r 2 + 3 r .
This is positive for r ≤ a r\leq a r ≤ a . For a ≤ r ≤ q ∗ a\leq r\leq\sqrt{q_*} a ≤ r ≤ q ∗ , K q ∗ ( q ∗ − a ) = a K\sqrt{q_*}(\sqrt{q_*}-a)=a K q ∗ ( q ∗ − a ) = a gives r − a ≤ a / ( K r ) r-a\leq a/(Kr) r − a ≤ a / ( K r ) , and hence
4 q 2 h ′ ′ ( q ) f ( q ) ≥ K r ( 2 r − a ) + 3 ( r − a ) − 3 a K r ≥ K a 2 − 3 K > 0. \frac{4q^2h''(q)}{f(q)}\geq Kr(2r-a)+3(r-a)-\frac{3a}{Kr}\geq Ka^2-\frac{3}{K}>0. f ( q ) 4 q 2 h ′′ ( q ) ≥ K r ( 2 r − a ) + 3 ( r − a ) − K r 3 a ≥ K a 2 − K 3 > 0.
Consequently, with S h = ∑ m ( − h ′ ( Q [ m ] ) ) m m T ⪰ 0 S_h=\sum_m(-h'(Q[m]))mm^T\succeq0 S h = ∑ m ( − h ′ ( Q [ m ])) m m T ⪰ 0 ,
∂ κ V κ ( Y ( ξ ) ) ∣ κ = K ≥ ∑ m ∈ M h ( Q [ m ] ) + tr ( S h Q ) − max z ∈ C ( Q ) tr ( S h Y z ) . (62) \left.\partial_\kappa V_\kappa(Y(\xi))\right|_{\kappa=K}\geq\sum_{m\in\mathcal{M}}h(Q[m])+\operatorname{tr}(S_hQ)-\max_{z\in C(\mathcal{Q})}\operatorname{tr}(S_hY_z). \tag*{(62)} ∂ κ V κ ( Y ( ξ )) ∣ κ = K ≥ m ∈ M ∑ h ( Q [ m ]) + tr ( S h Q ) − z ∈ C ( Q ) max tr ( S h Y z ) . (62)
In all three cases, the value and derivative bounds concern the same finite sum V κ V_\kappa V κ , with the same fixed set M \mathcal{M} M . No sign is imposed on derivatives of omitted terms: positivity is used only to compare the full energy with V κ V_\kappa V κ .
It remains to bound the full FCC reference in the value comparison. Its shell expansion is
Φ K ( L F ) = ∑ n ≥ 1 N n n e − K a ( n − 1 ) . \Phi_K(L_F)=\sum_{n\ge1}\frac{N_n}{\sqrt{n}}e^{-K a(\sqrt{n}-1)}. Φ K ( L F ) = n ≥ 1 ∑ n N n e − K a ( n − 1 ) .
All shells through n = 64 n=64 n = 64 are enumerated in [ − 12 , 12 ] 3 [-12,12]^3 [ − 12 , 12 ] 3 . The bound N n < 30 n N_n<30n N n < 30 n and monotonicity of x e − K a ( x − 1 ) \sqrt{x}e^{-K a(\sqrt{x}-1)} x e − K a ( x − 1 ) on [ 64 , ∞ ) [64,\infty) [ 64 , ∞ ) imply
Φ K ( L F ) ≤ ∑ n = 1 64 N n n e − K a ( n − 1 ) + 60 e − 7 K a ( 64 K a + 16 ( K a ) 2 + 2 ( K a ) 3 ) . (63) \Phi_K(L_F)\le\sum_{n=1}^{64}\frac{N_n}{\sqrt{n}}e^{-K a(\sqrt{n}-1)}+60e^{-7Ka}\left(\frac{64}{Ka}+\frac{16}{(Ka)^2}+\frac{2}{(Ka)^3}\right). \tag*{(63)} Φ K ( L F ) ≤ n = 1 ∑ 64 n N n e − K a ( n − 1 ) + 60 e − 7 K a ( K a 64 + ( K a ) 2 16 + ( K a ) 3 2 ) . (63)
The final term is the explicit integral 30 ∫ 64 ∞ x e − K a ( x − 1 ) d x 30\int_{64}^{\infty}\sqrt{x}e^{-Ka(\sqrt{x}-1)}\,dx 30 ∫ 64 ∞ x e − K a ( x − 1 ) d x ; it is retained in every comparison.
Theorem 5.2. For every κ ≥ 20 \kappa\ge20 κ ≥ 20 , the FCC lattice L F L_F L F uniquely minimizes E κ ( L ) E_\kappa(L) E κ ( L ) under ∣ L ∣ = 1 |L|=1 ∣ L ∣ = 1 , up to orthogonal transformations.
Proof. Lemma 3.2 gives the strict comparison when λ 1 ( L ) ≤ 4 / 5 \lambda_1(L)\le4/5 λ 1 ( L ) ≤ 4/5 . Otherwise the reduced coordinates lie in (23) . The finite inequalities of Lemma A.4 cover that rectangle. A rectangle which violates a reduction inequality contains no relevant representative. On every remaining rectangle outside the FCC local neighbourhood, (57) holds at K = 20 K=20 K = 20 , and Lemma 5.4 gives the strict comparison for the entire half-line. On the local rectangles the strict estimate of Theorem 4.1 applies, since κ a ≥ 20 a > 12 \kappa a\ge20a>12 κa ≥ 20 a > 12 . Equality is therefore possible only when L L L is orthogonally equivalent to L F L_F L F . □
Proof of Theorem 5.1. For 0 ≤ c ≤ 32 0\le c\le32 0 ≤ c ≤ 32 , the assertion, including its equality statement, is Lemma 5.3 . If c > 32 c>32 c > 32 , then κ = 4 π c > 20 \kappa=\sqrt{4\pi c}>20 κ = 4 π c > 20 , because 128 π > 400 128\pi>400 128 π > 400 . Theorem 5.2 and (15) show that P ( L F , c ) P(L_F,c) P ( L F , c ) is the minimum and that equality forces L L L to be orthogonally equivalent to L F L_F L F . These two ranges exhaust c ≥ 0 c\ge0 c ≥ 0 . □
The transition and the main theorems
We now compare the two cubic references and locate the transition. Throughout this section put D ( t ) = Θ ( t , L B ) − Θ ( t , L F ) D(t)=\Theta(t,L_B)-\Theta(t,L_F) D ( t ) = Θ ( t , L B ) − Θ ( t , L F ) . We prove the scalar theta ordering below, using the finite estimates in Lemma A.5 .
Theorem 6.1 (FCC–BCC classification). There is a unique κ ∗ ∈ ( 0 , π ) \kappa_*\in(0,\sqrt{\pi}) κ ∗ ∈ ( 0 , π ) such that
arg min L ∈ { L F , L B } E κ ( L ) = { { L B } , 0 < κ < κ ∗ , { L F , L B } , κ = κ ∗ , { L F } , κ > κ ∗ . (64) \underset{L\in\{L_F,L_B\}}{\arg\min}\ E_\kappa(L)=
\begin{cases}
\{L_B\}, & 0<\kappa<\kappa_*,\\
\{L_F,L_B\}, & \kappa=\kappa_*,\\
\{L_F\}, & \kappa>\kappa_*.
\end{cases} \tag*{(64)} L ∈ { L F , L B } arg min E κ ( L ) = ⎩ ⎨ ⎧ { L B } , { L F , L B } , { L F } , 0 < κ < κ ∗ , κ = κ ∗ , κ > κ ∗ . (64) The threshold is the unique positive root of (4) , with the enclosure (A.9); in particular, κ ∗ ≈ 1.71792186846990 \kappa_*\approx1.71792186846990 κ ∗ ≈ 1.71792186846990 .
The proof uses the scalar FCC–BCC theta ordering and a strictly increasing transform of their Yukawa energy difference.
Lemma 6.1. The function D ( t ) = Θ ( t , L B ) − Θ ( t , L F ) D(t)=\Theta(t,L_B)-\Theta(t,L_F) D ( t ) = Θ ( t , L B ) − Θ ( t , L F ) satisfies
D ( 1 ) = 0 , D ( t ) > 0 ( t > 1 ) . (65) D(1)=0,\qquad D(t)>0\quad(t>1). \tag*{(65)} D ( 1 ) = 0 , D ( t ) > 0 ( t > 1 ) . (65)
Proof. The equality follows from Poisson summation and L B = L F ∗ L_B=L_F^* L B = L F ∗ . For t ≥ 1 t\ge1 t ≥ 1 , the shell bounds N n < 30 n N_n<30n N n < 30 n , M ℓ < 30 ℓ M_\ell<30\ell M ℓ < 30 ℓ , together with e − π a 2 < 1 / 49 e^{-\pi a^2}<1/49 e − π a 2 < 1/49 , e − π 2 − 4 / 3 < 3 / 10 e^{-\pi2^{-4/3}}<3/10 e − π 2 − 4/3 < 3/10 and π 2 − 4 / 3 < 5 / 4 \pi2^{-4/3}<5/4 π 2 − 4/3 < 5/4 , give
∑ n ≥ 13 N n e − π a 2 n t < 30 ∑ n ≥ 13 n 49 n < 10 − 18 , \sum_{n\ge13}N_ne^{-\pi a^2nt}<30\sum_{n\ge13}\frac{n}{49^n}<10^{-18}, n ≥ 13 ∑ N n e − π a 2 n t < 30 n ≥ 13 ∑ 4 9 n n < 1 0 − 18 ,
π 2 − 4 / 3 ∑ ℓ ≥ 49 ℓ M ℓ e − π 2 − 4 / 3 ℓ t < 75 2 ∑ ℓ ≥ 49 ℓ 2 ( 3 / 10 ) ℓ < 10 − 18 . \pi2^{-4/3}\sum_{\ell\ge49}\ell M_\ell e^{-\pi2^{-4/3}\ell t}
<\frac{75}{2}\sum_{\ell\ge49}\ell^2(3/10)^\ell<10^{-18}. π 2 − 4/3 ℓ ≥ 49 ∑ ℓ M ℓ e − π 2 − 4/3 ℓ t < 2 75 ℓ ≥ 49 ∑ ℓ 2 ( 3/10 ) ℓ < 1 0 − 18 .
These are the only negative omitted tails in the lower bounds for D D D and D ′ D' D ′ , respectively. On a cell [ u , v ] ⊂ [ 1 , 11 / 10 ] [u,v] \subset[1,11/10] [ u , v ] ⊂ [ 1 , 11/10 ] , monotonicity of the exponentials gives
D ′ ( t ) ≥ π a 2 ∑ n = 1 12 n N n e − π a 2 n v − π 2 − 4 / 3 ∑ ℓ = 1 48 ℓ M ℓ e − π 2 − 4 / 3 ℓ u − 10 − 18 ( u ≤ t ≤ v ) . D'(t) \ge\pi a^2 \sum_{n=1}^{12} nN_n e^{-\pi a^2nv} - \pi2^{-4/3} \sum_{\ell=1}^{48} \ell M_\ell e^{-\pi2^{-4/3}\ell u} - 10^{-18} \qquad(u \le t \le v). D ′ ( t ) ≥ π a 2 n = 1 ∑ 12 n N n e − π a 2 n v − π 2 − 4/3 ℓ = 1 ∑ 48 ℓ M ℓ e − π 2 − 4/3 ℓ u − 1 0 − 18 ( u ≤ t ≤ v ) .
Take u = 1 + j / 10000 u=1+j/10000 u = 1 + j /10000 and v = u + 1 / 10000 v=u+1/10000 v = u + 1/10000 and apply (A.6). For D D D itself, the positive BCC terms are bounded at the right endpoint and the negative FCC terms at the left endpoint; the first tail bound and (A.7) apply on the corresponding cells. We obtain
D ′ ( t ) > 3 / 1000 ( 1 ≤ t ≤ 11 / 10 ) , D ( t ) > 1 / 10000 ( 11 / 10 ≤ t ≤ 21 / 10 ) . D'(t) > 3/1000 \qquad(1 \le t \le11/10), \qquad D(t) > 1/10000 \qquad(11/10 \le t \le21/10). D ′ ( t ) > 3/1000 ( 1 ≤ t ≤ 11/10 ) , D ( t ) > 1/10000 ( 11/10 ≤ t ≤ 21/10 ) .
For t ≥ 21 / 10 t \ge21/10 t ≥ 21/10 , we have e − π a 2 t < 1 / 128 e^{-\pi a^2t} < 1/128 e − π a 2 t < 1/128 , so Proposition 4.1 gives
Θ ( t , L F ) − 1 < e − π a 2 t [ 12 + 30 ( 1 ( 1 − e − π a 2 t ) 2 − 1 ) ] < 25 2 e − π a 2 t . \Theta(t,L_F)-1 < e^{-\pi a^2t}\left[12+30\left(\frac{1}{(1-e^{-\pi a^2t})^2}-1\right)\right] < \frac{25}{2}e^{-\pi a^2t}. Θ ( t , L F ) − 1 < e − π a 2 t [ 12 + 30 ( ( 1 − e − π a 2 t ) 2 1 − 1 ) ] < 2 25 e − π a 2 t .
The first BCC shell gives Θ ( t , L B ) − 1 ≥ 8 e − 3 π 2 − 4 / 3 t \Theta(t,L_B)-1 \ge8e^{-3\pi2^{-4/3}t} Θ ( t , L B ) − 1 ≥ 8 e − 3 π 2 − 4/3 t . The last inequality of Lemma A.5 states 8 e ( π a 2 − 3 π 2 − 4 / 3 ) 21 / 10 > 25 / 2 8e^{(\pi a^2-3\pi2^{-4/3})21/10}>25/2 8 e ( π a 2 − 3 π 2 − 4/3 ) 21/10 > 25/2 . As a 2 > 3 ⋅ 2 − 4 / 3 a^2>3\cdot2^{-4/3} a 2 > 3 ⋅ 2 − 4/3 , this inequality remains strict for every t ≥ 21 / 10 t\ge21/10 t ≥ 21/10 . Thus D ( t ) > 0 D(t)>0 D ( t ) > 0 also in this range. □ \square □
Proof of Theorem 6.1. Since L B = L F ∗ L_B=L_F^* L B = L F ∗ , Poisson summation gives D ( 1 / t ) = − t 3 / 2 D ( t ) D(1/t)=-t^{3/2}D(t) D ( 1/ t ) = − t 3/2 D ( t ) and D ( 1 ) = 0 D(1)=0 D ( 1 ) = 0 . Writing c = κ 2 / ( 4 π ) c=\kappa^2/(4\pi) c = κ 2 / ( 4 π ) , Proposition 2.1 gives
2 ( E κ ( L B ) − E κ ( L F ) ) = ∫ 1 ∞ D ( t ) [ t − 1 / 2 e − c / t − e − c t ] d t . (66) 2\bigl(E_\kappa(L_B)-E_\kappa(L_F)\bigr)=\int_1^\infty D(t)\left[t^{-1/2}e^{-c/t}-e^{-ct}\right]\,dt. \tag*{(66)} 2 ( E κ ( L B ) − E κ ( L F ) ) = ∫ 1 ∞ D ( t ) [ t − 1/2 e − c / t − e − c t ] d t . (66)
This expression has a continuous limit at c = 0 c=0 c = 0 , where it is strictly negative by (65) . At c = 1 / 4 c=1/4 c = 1/4 the bracket is strictly positive for every t > 1 t>1 t > 1 : indeed,
t − 1 t > 2 log t ( t > 1 ) , t-\frac{1}{t}>2\log t \qquad(t>1), t − t 1 > 2 log t ( t > 1 ) ,
as follows by differentiating the difference. Thus a crossing exists with 0 < κ ∗ < π 0<\kappa_*<\sqrt{\pi} 0 < κ ∗ < π .
For uniqueness, substitute u = 1 / t u=1/t u = 1/ t in the unsplit Gaussian expression and use D ( 1 / u ) = − u 3 / 2 D ( u ) D(1/u)=-u^{3/2}D(u) D ( 1/ u ) = − u 3/2 D ( u ) . This gives
2 ( E κ ( L B ) − E κ ( L F ) ) = ∫ 0 ∞ u − 3 / 2 e − c u D ( 1 / u ) d u = − ∫ 0 ∞ e − c u D ( u ) d u . \begin{aligned}
2\bigl(E_\kappa(L_B)-E_\kappa(L_F)\bigr)=\int_0^\infty u^{-3/2}e^{-cu}D(1/u)\,du \\
&=-\int_0^\infty e^{-cu}D(u)\,du.
\end{aligned} 2 ( E κ ( L B ) − E κ ( L F ) ) = ∫ 0 ∞ u − 3/2 e − c u D ( 1/ u ) d u = − ∫ 0 ∞ e − c u D ( u ) d u .
Both integrals converge absolutely. Multiply by e c e^c e c and differentiate with respect to c c c . The derivative is
∫ 0 ∞ ( u − 1 ) e c ( 1 − u ) D ( u ) d u > 0. \int_0^\infty(u-1)e^{c(1-u)}D(u)\,du>0. ∫ 0 ∞ ( u − 1 ) e c ( 1 − u ) D ( u ) d u > 0.
Both factors u − 1 u-1 u − 1 and D ( u ) D(u) D ( u ) have the same sign on either side of 1. The exponential decay of D D D at infinity and its Poisson decay at zero justify differentiation, including at c = 0 c=0 c = 0 . The rescaled energy difference is therefore strictly increasing. Together with the opposite signs at c = 0 c=0 c = 0 and c = 1 / 4 c=1/4 c = 1/4 , this proves that it has exactly one zero. Equations (64) and (4) follow, and Lemma A.6 gives the stated enclosure. □ \square □
Proof of Theorem 1.1. Existence follows from Theorem 2.1. Theorem 5.1 and (15) give
E κ ( L ) ≥ min { E κ ( L F ) , E κ ( L B ) } , E_\kappa(L)\ge\min\{E_\kappa(L_F),E_\kappa(L_B)\}, E κ ( L ) ≥ min { E κ ( L F ) , E κ ( L B )} ,
and equality restricts the lattice class to the two references. Theorem 6.1 determines which reference has smaller energy, including the equality case at the unique transition. Finally, (6) gives the statement at arbitrary density. □ \square □
Proof of Corollary 1.1. Proposition 2.3 gives the limit in (7) :
2 E κ ( L ) − 4 π κ 2 = 2 F κ ( L ) + κ → 2 W ( L ) = Z ( L , 1 / 2 ) . 2E_{\kappa}(L)-\frac{4\pi}{\kappa^2}=2F_{\kappa}(L)+\kappa\to2W(L)=Z(L,1/2). 2 E κ ( L ) − κ 2 4 π = 2 F κ ( L ) + κ → 2 W ( L ) = Z ( L , 1/2 ) .
Apply Theorem 5.1 at c = 0 c=0 c = 0 . The identity (66) at zero and D ( t ) > 0 D(t)>0 D ( t ) > 0 for t > 1 t>1 t > 1 yield W ( L B ) < W ( L F ) W(L_{\mathrm{B}})<W(L_{\mathrm{F}}) W ( L B ) < W ( L F ) . Since W ( L ) = P ( L , 0 ) − 3 / 2 W(L)=P(L,0)-3/2 W ( L ) = P ( L , 0 ) − 3/2 , BCC is the unique minimizing class. The Epstein functional equation in (13) , with M = L ∗ M=L^* M = L ∗ , gives the equivalent statement for the continued value Z ( M , 1 ) Z(M,1) Z ( M , 1 ) . □
Appendix A. Numerical comparisons
We collect the finite numerical comparisons used in the preceding proofs. Each lemma states inequalities on explicit parameter regions; the accompanying interval computations include all indicated tails and rounding errors.
Lemma A.1 (Scalar cusp comparison). For the unit-covolume FCC lattice,
P ( L F , 0 ) < 55995 10 6 < 7 125 < I ( 16 25 , 0 ) . P(L_{\mathrm{F}},0)<\frac{55995}{10^6}<\frac{7}{125}<I\left(\frac{16}{25},0\right). P ( L F , 0 ) < 1 0 6 55995 < 125 7 < I ( 25 16 , 0 ) . Proof. The FCC and BCC squared lengths are 2 1 / 3 n 2^{1/3}n 2 1/3 n and 2 − 4 / 3 ℓ 2^{-4/3}\ell 2 − 4/3 ℓ , with shell multiplicities N n N_n N n and M ℓ M_\ell M ℓ . Fixing two integer coordinates gives N n < 30 n N_n<30n N n < 30 n and M ℓ < 30 ℓ M_\ell<30\ell M ℓ < 30 ℓ . Two integrations by parts yield
I ( q , 0 ) ≤ e − π q ( π q + 1 ) π q ( π q + 3 / 2 ) ≤ e − π q π q . (67) I(q,0)\leq\frac{e^{-\pi q}(\pi q+1)}{\pi q(\pi q+3/2)}\leq\frac{e^{-\pi q}}{\pi q}. \tag*{(67)} I ( q , 0 ) ≤ π q ( π q + 3/2 ) e − π q ( π q + 1 ) ≤ π q e − π q . (67) The first two FCC multiplicities are 12 and 6. Consequently,
1 2 ∑ v ∈ L F ∖ { 0 } I ( ∣ v ∣ 2 , 0 ) ≤ 6 I ( 2 1 / 3 , 0 ) + 3 I ( 2 4 / 3 , 0 ) + 15 e − 3 π 2 1 / 3 π 2 1 / 3 ( 1 − e − π 2 1 / 3 ) < 26459 10 6 . \frac{1}{2}\sum_{v\in L_{\mathrm{F}}\setminus\{0\}}I(|v|^2,0)
\leq6I(2^{1/3},0)+3I(2^{4/3},0)+\frac{15e^{-3\pi2^{1/3}}}{\pi2^{1/3}(1-e^{-\pi2^{1/3}})}
<\frac{26459}{10^6}. 2 1 v ∈ L F ∖ { 0 } ∑ I ( ∣ v ∣ 2 , 0 ) ≤ 6 I ( 2 1/3 , 0 ) + 3 I ( 2 4/3 , 0 ) + π 2 1/3 ( 1 − e − π 2 1/3 ) 15 e − 3 π 2 1/3 < 1 0 6 26459 . For BCC, the nonzero integer squared norms through twelve are 3, 4, 8, 11, 12, with multiplicities 8, 6, 12, 24, 8. Thus
1 2 ∑ v ∈ L B ∖ { 0 } J ( ∣ v ∣ 2 , 0 ) ≤ 1 2 ∑ ℓ ∈ { 3 , 4 , 8 , 11 , 12 } M ℓ e − π 2 − 4 / 3 ℓ π 2 − 4 / 3 ℓ + 15 e − 13 π 2 − 4 / 3 π 2 − 4 / 3 ( 1 − e − π 2 − 4 / 3 ) < 29536 10 6 . \frac{1}{2}\sum_{v\in L_{\mathrm{B}}\setminus\{0\}}J(|v|^2,0)
\leq\frac{1}{2}\sum_{\ell\in\{3,4,8,11,12\}}M_\ell\frac{e^{-\pi2^{-4/3}\ell}}{\pi2^{-4/3}\ell}
+\frac{15e^{-13\pi2^{-4/3}}}{\pi2^{-4/3}(1-e^{-\pi2^{-4/3}})}
<\frac{29536}{10^6}. 2 1 v ∈ L B ∖ { 0 } ∑ J ( ∣ v ∣ 2 , 0 ) ≤ 2 1 ℓ ∈ { 3 , 4 , 8 , 11 , 12 } ∑ M ℓ π 2 − 4/3 ℓ e − π 2 − 4/3 ℓ + π 2 − 4/3 ( 1 − e − π 2 − 4/3 ) 15 e − 13 π 2 − 4/3 < 1 0 6 29536 . The finite expressions, including the displayed geometric tails, give these strict bounds by outward interval evaluation. The same evaluation of I ( 16 / 25 , 0 ) = 5 4 erfc ( 4 π / 5 ) I(16/25,0)=\frac{5}{4}\operatorname{erfc}(4\sqrt{\pi}/5) I ( 16/25 , 0 ) = 4 5 erfc ( 4 π /5 ) gives I ( 16 / 25 , 0 ) > 7 / 125 I(16/25,0)>7/125 I ( 16/25 , 0 ) > 7/125 . Adding the two upper bounds proves the lemma. □
Lemma A.2 (Numerical bounds for the local expansion). Let h h h , C 3 C_3 C 3 , and M 4 M_4 M 4 be the endpoint bounds defined by (44) , (45) , and (38) , including the tail enlargements in § 4 . On each of the following closed parameter intervals they satisfy:
Proof. The retained FCC and BCC integer shells are those of § 4 , with squared integer norm at most 18 and 30, respectively. Their integer moments determine all terms in (44) , (45) , and (38) . Evaluate these endpoint expressions with outward rational interval enclosures with 256 binary fractional digits, taking lower endpoints for h h h and upper endpoints for C 3 C_3 C 3 , M 4 M_4 M 4 . The companion interval computation encloses all scalar kernels and elementary functions rigorously; its arithmetic specification is supplied with the certificate.
For every one of the 1000 FCC intervals and 25 BCC intervals, the lower endpoint of h / 2 − C 3 / 240 − M 4 / 38400 h/2-C_3/240-M_4/38400 h /2 − C 3 /240 − M 4 /38400 exceeds the stated rational bound. The shell allowances 10 − 10 10^{-10} 1 0 − 10 and 10 − 6 10^{-6} 1 0 − 6 are already included in those expressions. The denominators are 6 / ( 1 / 40 ) = 240 6/(1/40)=240 6/ ( 1/40 ) = 240 and 24 / ( 1 / 40 ) 2 = 38400 24/(1/40)^2=38400 24/ ( 1/40 ) 2 = 38400 . Monotonicity in c c c and the signed Taylor bounds in (43) make these comparisons valid on the entire closed intervals. This proves the lemma. □
For the global comparisons, use the exact rectangle
D = [ α 1 , β ] × [ α 2 , β ] × [ 0 , 1 / 2 ] × [ 0 , 1 ] × [ 0 , 1 / 2 ] , \mathcal{D}=[\alpha_1,\beta]\times[\alpha_2,\beta]\times[0,1/2]\times[0,1]\times[0,1/2], D = [ α 1 , β ] × [ α 2 , β ] × [ 0 , 1/2 ] × [ 0 , 1 ] × [ 0 , 1/2 ] ,
α 1 = − 7394735416760227 2 53 , α 2 = − 2158713440539059 2 50 , β = 5182419497180053 2 54 . (68) \alpha_1=-\frac{7394735416760227}{2^{53}},\qquad\alpha_2=-\frac{2158713440539059}{2^{50}},\qquad\beta=\frac{5182419497180053}{2^{54}}. \tag*{(68)} α 1 = − 2 53 7394735416760227 , α 2 = − 2 50 2158713440539059 , β = 2 54 5182419497180053 . (68)
In the reduction inequalities (24) , take y i = e ℓ i y_i=e^{\ell_i} y i = e ℓ i . All corners below are the 32 corners in the five reduced coordinates.
Lemma A.3 (Numerical rectangular comparisons). The endpoints in (A.3) satisfy
e α 1 ≤ 11 25 , e α 2 ≤ 147 1000 , e β ≥ 4 3 . (69) e^{\alpha_1}\leq\frac{11}{25},\qquad e^{\alpha_2}\leq\frac{147}{1000},\qquad e^\beta\geq\frac{4}{3}. \tag*{(69)} e α 1 ≤ 25 11 , e α 2 ≤ 1000 147 , e β ≥ 3 4 . (69) Put
( c 0 , … , c 10 ) = ( 0 , 1 / 4 , 1 / 2 , 1 , 2 , 4 , 8 , 12 , 16 , 24 , 32 ) , J j = [ c j − 1 , c j ] . (c_0,\ldots,c_{10})=(0,1/4,1/2,1,2,4,8,12,16,24,32),\qquad J_j=[c_{j-1},c_j]. ( c 0 , … , c 10 ) = ( 0 , 1/4 , 1/2 , 1 , 2 , 4 , 8 , 12 , 16 , 24 , 32 ) , J j = [ c j − 1 , c j ] . For 1 ≤ j ≤ 10 1\leq j\leq10 1 ≤ j ≤ 10 , the rectangle D \mathcal{D} D has a partition into n j n_j n j closed dyadic rectangles with disjoint interiors, where
( n 1 , … , n 5 ) = ( 20037 , 15481 , 10757 , 8852 , 7032 ) , (n_1,\ldots,n_5)=(20037,15481,10757,8852,7032), ( n 1 , … , n 5 ) = ( 20037 , 15481 , 10757 , 8852 , 7032 ) , ( n 6 , … , n 10 ) = ( 5523 , 4295 , 3409 , 4154 , 4503 ) . (n_6,\ldots,n_{10})=(5523,4295,3409,4154,4503). ( n 6 , … , n 10 ) = ( 5523 , 4295 , 3409 , 4154 , 4503 ) . Use the degrees and FCC radii
( N 1 , … , N 10 ) = ( 12 , 14 , 18 , 24 , 30 , 40 , 40 , 40 , 40 , 40 ) , δ j = { 17 / 2000 , j = 1 , 7 / 500 , 2 ≤ j ≤ 7 , 3 / 100 , 8 ≤ j ≤ 10. (N_1,\ldots,N_{10})=(12,14,18,24,30,40,40,40,40,40),\qquad
\delta_j=\begin{cases}
17/2000,&j=1,\\
7/500,&2\leq j\leq7,\\
3/100,&8\leq j\leq10.
\end{cases} ( N 1 , … , N 10 ) = ( 12 , 14 , 18 , 24 , 30 , 40 , 40 , 40 , 40 , 40 ) , δ j = ⎩ ⎨ ⎧ 17/2000 , 7/500 , 3/100 , j = 1 , 2 ≤ j ≤ 7 , 8 ≤ j ≤ 10. For every rectangle Q Q Q in the j j j th partition, at least one of the following holds:
(i) One fixed right-hand side of (24) is strictly less than 1 at every corner of Q Q Q .
*(ii) max z ∈ C ( Q ) d 1 ( z ) ≤ 16 / 25 \max_{z\in C(Q)}d_1(z)\leq16/25 max z ∈ C ( Q ) d 1 ( z ) ≤ 16/25 .
(iii) For at least one R ∈ { L F , L B } R\in\{L_F,L_B\} R ∈ { L F , L B } , inequality (55) holds with J = J j J=J_j J = J j and N = N j N=N_j N = N j .
(iv) For R = L F R=L_F R = L F and δ = δ j \delta=\delta_j δ = δ j , or for R = L B R=L_B R = L B , j = 1 j=1 j = 1 and δ = 7 / 500 \delta=7/500 δ = 7/500 , every corner satisfies
( 1 + δ ) Y R − Y z ≻ 0 , ( 1 − δ ) − 1 Y R − 1 − Y z − 1 ≻ 0. (70) (1+\delta)Y_R-Y_z\succ0,\qquad(1-\delta)^{-1}Y_R^{-1}-Y_z^{-1}\succ0. \tag*{(70)} ( 1 + δ ) Y R − Y z ≻ 0 , ( 1 − δ ) − 1 Y R − 1 − Y z − 1 ≻ 0. (70) with Y R Y_R Y R as in (56) .
There are 84,043 rectangles in total. The BCC local inequalities are used only for 0 ≤ c ≤ 1 / 4 0\leq c\leq1/4 0 ≤ c ≤ 1/4 .
Proof. The accompanying partition data give exact dyadic subdivisions of (A.3). At each subdivision, the two closed rectangles have disjoint interiors and their union is the original rectangle. These rational identities verify coverage of D \mathcal{D} D and give the ten stated counts.
At the midpoint of each rectangle, form the coefficients and remainders of Lemma 5.2, using the fixed competitor sets and reference shells specified in Section 5. Substitution in (55) , the reduction expressions, and the three leading principal minors of the two matrices in (A.5) establishes one of (i)–(iv). These finite evaluations use outward interval bounds at precision 128 bits, including the exponential and complementary error functions [27 ] . Every asserted strict inequality has an enclosed endpoint of the required strict sign; the nonstrict bounds in (ii) are checked with their upper endpoints. The same interval calculation gives (A.4). The reference tails are included through (54) , and the screening remainders through (53) ; hence each comparison covers the entire screening interval and shape rectangle.
Lemma A.4 (Finite inequalities for the screening half-line). At K = 20 K=20 K = 20 , D \mathcal{D} D has a dyadic partition into 7,225 rectangles. Of these, 1,118 satisfy alternative (i) of Lemma A.3, 193 satisfy (ii), and 86 satisfy the FCC inequalities (A.5) with δ = 3 / 100 \delta=3/100 δ = 3/100 . On each of the remaining 5,828 rectangles there is a fixed set M ⊂ { − 3 , − 2 , … , 3 } 3 ∖ { 0 } \mathcal{M}\subset\{-3,-2,\ldots,3\}^3\setminus\{0\} M ⊂ { − 3 , − 2 , … , 3 } 3 ∖ { 0 } for which the right-hand side of (60) strictly exceeds the right-hand side of (63) , and at least one of the following holds:
(a) the right-hand side of (61) is nonnegative;
(b) q ‾ m < a 2 \overline{q}_m<a^2 q m < a 2 for every m ∈ M m\in\mathcal{M} m ∈ M ;
(c) q ‾ m < q ∗ \overline{q}_m < q_* q m < q ∗ for every m ∈ M m \in\mathcal{M} m ∈ M and the right-hand side of (62) is nonnegative. All bounds are evaluated at K = 20 K = 20 K = 20 , with the same fixed set M \mathcal{M} M on each rectangle. Consequently (5.7} holds throughout each of these 5,828 rectangles.
Proof. The partition data give 7,224 exact dyadic subdivisions. At each subdivision the two closed rectangles have disjoint interiors and union equal to the original rectangle; hence the resulting 7,225 rectangles cover D \mathcal{D} D . Evaluate (58) at each retained vector, then the value bound and the applicable derivative bound above. Outward interval enclosures with 512-bit endpoints give the asserted signs on all 5,828 rectangles. The full reference upper bound in (63) , including its integral tail, is retained. The same enclosures verify the reduction, cusp and local inequalities on the other rectangles; positive definiteness is checked through the three leading principal minors. Equations (60) –(63) extend these finite comparisons to the whole rectangles, and (a)–(c) each imply the derivative inequality in (5.7). □
Lemma A.5 (Finite theta comparisons). Let N n N_n N n count integer triples of squared norm 2 n 2n 2 n , and let M ℓ M_\ell M ℓ count integer triples of squared norm ℓ \ell ℓ whose coordinates all have the same parity. For every integer 0 ≤ j < 1000 0 \le j < 1000 0 ≤ j < 1000 ,
π a 2 ∑ n = 1 12 N n e − π a 2 n ( 1 + ( j + 1 ) / 10000 ) − π 2 − 4 / 3 ∑ ℓ = 1 48 M ℓ e − π 2 − 4 / 3 ℓ ( 1 + j / 10000 ) − 10 − 18 > 3 1000 . (71) \pi a^2 \sum_{n=1}^{12} N_n e^{-\pi a^2 n(1+(j+1)/10000)}-\pi2^{-4/3}\sum_{\ell=1}^{48}M_\ell e^{-\pi2^{-4/3}\ell(1+j/10000)}-10^{-18}>\frac{3}{1000}. \tag*{(71)} π a 2 n = 1 ∑ 12 N n e − π a 2 n ( 1 + ( j + 1 ) /10000 ) − π 2 − 4/3 ℓ = 1 ∑ 48 M ℓ e − π 2 − 4/3 ℓ ( 1 + j /10000 ) − 1 0 − 18 > 1000 3 . (71) For every integer 0 ≤ j < 2000 0 \le j < 2000 0 ≤ j < 2000 ,
∑ ℓ = 1 48 M ℓ e − π 2 − 4 / 3 ℓ ( 11 / 10 + ( j + 1 ) / 2000 ) − ∑ n = 1 12 N n e − π a 2 n ( 11 / 10 + j / 2000 ) − 10 − 18 > 1 10000 . (72) \sum_{\ell=1}^{48}M_\ell e^{-\pi2^{-4/3}\ell(11/10+(j+1)/2000)}-\sum_{n=1}^{12}N_n e^{-\pi a^2n(11/10+j/2000)}-10^{-18}>\frac{1}{10000}. \tag*{(72)} ℓ = 1 ∑ 48 M ℓ e − π 2 − 4/3 ℓ ( 11/10 + ( j + 1 ) /2000 ) − n = 1 ∑ 12 N n e − π a 2 n ( 11/10 + j /2000 ) − 1 0 − 18 > 10000 1 . (72) Finally,
8 e ( π a 2 − 3 π 2 − 4 / 3 ) 21 / 10 > 25 2 . (73) 8e^{(\pi a^2-3\pi2^{-4/3})21/10}>\frac{25}{2}. \tag*{(73)} 8 e ( π a 2 − 3 π 2 − 4/3 ) 21/10 > 2 25 . (73)
Proof. Integer enumeration gives
( N 1 , … , N 12 ) = ( 12 , 6 , 24 , 12 , 24 , 8 , 48 , 6 , 36 , 24 , 24 , 24 ) . (N_1,\ldots,N_{12})=(12,6,24,12,24,8,48,6,36,24,24,24). ( N 1 , … , N 12 ) = ( 12 , 6 , 24 , 12 , 24 , 8 , 48 , 6 , 36 , 24 , 24 , 24 ) .
The nonzero BCC indices and their multiplicities are
ℓ = ( 3 , 4 , 8 , 11 , 12 , 16 , 19 , 20 , 24 , 27 , 32 , 35 , 36 , 40 , 43 , 44 , 48 ) , \ell=(3,4,8,11,12,16,19,20,24,27,32,35,36,40,43,44,48), ℓ = ( 3 , 4 , 8 , 11 , 12 , 16 , 19 , 20 , 24 , 27 , 32 , 35 , 36 , 40 , 43 , 44 , 48 ) ,
M ℓ = ( 8 , 6 , 12 , 24 , 8 , 6 , 24 , 24 , 24 , 32 , 12 , 48 , 30 , 24 , 24 , 24 , 8 ) . M_\ell=(8,6,12,24,8,6,24,24,24,32,12,48,30,24,24,24,8). M ℓ = ( 8 , 6 , 12 , 24 , 8 , 6 , 24 , 24 , 24 , 32 , 12 , 48 , 30 , 24 , 24 , 24 , 8 ) .
Substitution in the finite exponential sums, with outward interval enclosures on the stated rational cells, gives (A.6) and (A.7) for every indicated index, as well as (A.8). The margins in the statement include the omitted-shell bounds used in the proof of Lemma 6.1. □
Lemma A.6 (Exact transition enclosure). At the two rational screening parameters,
2 ( E κ ( L B ) − E κ ( L F ) ) < − 10 − 19 if κ = 1717921868469903 10 15 , 2(E_\kappa(L_{\mathrm{B}})-E_\kappa(L_{\mathrm{F}}))<-10^{-19}
\qquad\text{if }\kappa=\frac{1717921868469903}{10^{15}}, 2 ( E κ ( L B ) − E κ ( L F )) < − 1 0 − 19 if κ = 1 0 15 1717921868469903 , 2 ( E κ ( L B ) − E κ ( L F ) ) > 10 − 20 if κ = 1717921868469904 10 15 . 2(E_\kappa(L_{\mathrm{B}})-E_\kappa(L_{\mathrm{F}}))>10^{-20}
\qquad\text{if }\kappa=\frac{1717921868469904}{10^{15}}. 2 ( E κ ( L B ) − E κ ( L F )) > 1 0 − 20 if κ = 1 0 15 1717921868469904 . Consequently, the unique root in Theorem 6.1 satisfies
1.717921868469903 < κ ∗ < 1.717921868469904. (74) 1.717921868469903<\kappa_*<1.717921868469904. \tag*{(74)} 1.717921868469903 < κ ∗ < 1.717921868469904. (74)
Proof. Retain FCC shells n ≤ 12 n \le12 n ≤ 12 and BCC shells ℓ ≤ 48 \ell\le48 ℓ ≤ 48 in the Ewald difference, with the multiplicities in Lemma A.5. Since I ( q , c ) , J ( q , c ) ≤ e − π q / ( π q ) I(q,c),J(q,c)\le e^{-\pi q/(\pi q)} I ( q , c ) , J ( q , c ) ≤ e − π q / ( π q ) , the absolute omitted tail in 2 ( E κ ( L B ) − E κ ( L F ) ) 2(E_\kappa(L_{\mathrm{B}})-E_\kappa(L_{\mathrm{F}})) 2 ( E κ ( L B ) − E κ ( L F )) is at most
60 ( 1 / 49 ) 13 1 − 1 / 49 + 60 ( 3 / 10 ) 49 1 − 3 / 10 < 7 ⋅ 10 − 21 . \frac{60(1/49)^{13}}{1-1/49}+\frac{60(3/10)^{49}}{1-3/10}<7\cdot10^{-21}. 1 − 1/49 60 ( 1/49 ) 13 + 1 − 3/10 60 ( 3/10 ) 49 < 7 ⋅ 1 0 − 21 .
Outward interval evaluation of the finite sums at the two rational parameters, including this tail, gives the two signed bounds in the statement. The uniqueness proved in Theorem 6.1 then gives (A.9). □
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. The research of J. Wei is partially supported by the General Research Fund (GRF) of Hong Kong ”New frontiers in singularity formation of nonlinear partial differential equations”. The authors used AI to assist the long but finite computations in the Appendix.
(S. Luo) School of Mathematics and Statistics, Jiangxi Normal University, Nanchang, 330022, China
(J. Wei) Department of Mathematics, Chinese University of Hong Kong, Shatin, NT, Hong Kong
Email address , S. Luo: luosp1989@163.com
Email address , J. Wei: wei@math.cuhk.edu.hk