Research · White paper · Long-range order

Ground states of the 2D dipolar Ising model and of dipolar spin ice

We prove which phase is a ground state of the Ising model on Z2\mathbb Z^2 with nearest-neighbour ferromagnetic coupling JJ and an antiferromagnetic 1/r31/r^3 tail, among all configurations, for every J∈[0,3.95]J\in[0,3.95] except one window of width 2.8×10−62.8\times10^{-6}; inside each interval it is the only translation-invariant one. The phases are the Néel state, checkerboards and stripes; ten couplings where two phases tie are certified to at least 17 decimals. We prove that the Melko–den Hertog–Gingras (MHG) state is the unique translation-invariant ground state of tin-foil Ewald-summed dipolar spin ice with nearest-neighbour exchange only, for J1/D≤9/2J_1/D\le9/2. Numerical studies had located these phases 1, 3, 8, 10, 12, 18; we know of no earlier proof, apart from the Néel state at J=0J=0.

The question

The 2D model has spins sx=±1s_x=\pm1 on Z2\mathbb Z^2, a ferromagnetic nearest-neighbour coupling and an antiferromagnetic tail ∣r∣−p|r|^{-p}. For p=3p=3, the dipolar case, it describes an ultrathin magnetic film with out-of-plane anisotropy 18. The energy is

H=∑x<yW(x−y) sxsy,W(r)=∣r∣−3−J [ ∣r∣=1 ],H=\sum_{x<y}W(x-y)\,s_xs_y,\qquad W(r)=|r|^{-3}-J\,[\,|r|=1\,],

where J≥0J\ge0 is the ratio δ\delta of Fonseca et al. 2, whose dipolar unit sums over all distinct pairs. Giuliani and Seiringer proved stripes just below the ferromagnetic transition for p>4p>4 and wrote in 2015 that the dipolar case had "eluded any rigorous treatment so far" 4. Fermi and Giuliani proved the stripe conjecture for J≥J0J\ge J_0, with J0J_0 unspecified, among periodic states with straight domain walls 5.

Dipolar spin ice, the model used for Dy2_2Ti2_2O7_7 and Ho2_2Ti2_2O7_7 3, 19, has Ising spins σi=±1\sigma_i=\pm1 on the pyrochlore lattice along local ⟨111⟩\langle111\rangle axes. With Si=σiz^i\mathbf S_i=\sigma_i\hat z_i, z^i\hat z_i the axis, and rnnr_{nn} the nearest-neighbour distance, the energy is

H=D rnn3∑i<j[Si ⁣⋅ ⁣Sjrij3−3(Si ⁣⋅ ⁣rij)(Sj ⁣⋅ ⁣rij)rij5]+∑i<jJij Si ⁣⋅ ⁣Sj.H=D\,r_{nn}^3\sum_{i<j}\Big[\frac{\mathbf S_i\!\cdot\!\mathbf S_j}{r_{ij}^3}-\frac{3(\mathbf S_i\!\cdot\!\mathbf r_{ij})(\mathbf S_j\!\cdot\!\mathbf r_{ij})}{r_{ij}^5}\Big]+\sum_{i<j}J_{ij}\,\mathbf S_i\!\cdot\!\mathbf S_j .

Here DD is the dipolar strength, JijJ_{ij} is J1J_1 or J2J_2 for first or second neighbours, and for third neighbours J3tJ_{3t} if the midpoint of the pair is a lattice site and J3hJ_{3h} if not. J1>0J_1>0 favours all-in-all-out (AIAO). The nearest-neighbour couplings JnnJ_{nn} and DnnD_{nn} of 3 obey Dnn=5D/3D_{nn}=5D/3 and Jnn/Dnn=−J1/(5D)J_{nn}/D_{nn}=-J_1/(5D); for Dy2_2Ti2_2O7_7, Jnn=−1.24J_{nn}=-1.24 K and Dnn=2.35D_{nn}=2.35 K give D=1.41D=1.41 K and J1/D=2.638≈66/25J_1/D=2.638\approx66/25. Energies are per spin, the dipolar sum being taken by Ewald's method with metallic (tin-foil) boundary. Monte Carlo found the MHG state, a zero-moment state with ordering vector (0,0,2π/a)(0,0,2\pi/a) 3; that it is the ground state had rested on numerics 10, 12.

What we found

A periodic state is a ground state if no configuration of Z2\mathbb Z^2 has lower energy per site ee along any sequence of boxes. Images are translates, lattice rotations and reflections, and spin flips. The checkerboard (k×1)(k\times1) is s(x,y)=sqk(x)(−1)ys(x,y)=\mathrm{sq}_k(x)(-1)^y, sqk\mathrm{sq}_k the square wave of period 2k2k, so (1×1)(1\times1) is the Néel state; stripe hh is sqh(y)\mathrm{sq}_h(y).

NamePhasesJ∗J^*Images
c1c_1Néel and (2×1)(2\times1)0.415203055875182819010
c2c_2(2×1)(2\times1) and (3×1)(3\times1)0.4394199028659450445320
c3c_3(3×1)(3\times1) and (4×1)(4\times1)0.4402790703331051159728
c4c_4(4×1)(4\times1) and (5×1)(5\times1)0.4403109600340542596336
c5c_5(5×1)(5\times1) and (6×1)(6\times1)0.4403121858138043068644, SAT
s1s_1stripes 1 and 21.258538518261292358312
s2s_2stripes 2 and 32.17245449537855555820
s3s_3stripes 3 and 42.8144951381707532128
s4s_4stripes 4 and 53.304241945342747495236
s5s_5stripes 5 and 63.6990247013763192689544, SAT
Couplings where two phases tie, with the images of both (SAT: exhaustive SAT enumeration without proof log); every digit is certified.

Theorem 1 (Ground states of the 2D model). Let c0=0c_0=0 and s6=3.95s_6=3.95. (a) For k=1,…,5k=1,\dots,5 the checkerboard (k×1)(k\times1) is a ground state for ck−1≤J≤ckc_{k-1}\le J\le c_k, and (6×1)(6\times1) is one at J=c5J=c_5. (b) Stripe 1 is a ground state for 0.440315≤J≤s10.440315\le J\le s_1, and stripe hh for sh−1≤J≤shs_{h-1}\le J\le s_h, h=2,…,6h=2,\dots,6. (c) In the interior of each interval of (a) and (b), the translation-invariant ground states are exactly the mixtures of the images of that phase. (d) At c1,…,c5c_1,\dots,c_5 and s1,…,s5s_1,\dots,s_5 both neighbouring phases are ground states, JJ is the crossing of their energy lines, and the translation-invariant ground states are the mixtures of their images; for c5c_5 and s5s_5 the last assertion rests on the SAT enumeration.

Theorem 2 (MHG ground state). For J2=J3=0J_2=J_3=0 and every J1/D≤9/2J_1/D\le9/2, the MHG state is the unique translation-invariant ground state, up to its 12 images. The same holds for J1/D≤9/2J_1/D\le9/2 with −1/25≤J2/D≤1/125-1/25\le J_2/D\le1/125, ∣J3t−J3h∣≤0.002D|J_{3t}-J_{3h}|\le0.002D and 0≤J3t+J3h≤0.06D0\le J_{3t}+J_{3h}\le0.06D. At J2=J3=0J_2=J_3=0 the energy is e0=12(aX+23J1)e_0=\tfrac12(a_X+\tfrac23J_1) with aX=−3.89513176378013880505 Da_X=-3.89513176378013880505\,D, enclosed to 10−2610^{-26}, the dipolar part of the eigenvalue λX=aX+23J1\lambda_X=a_X+\tfrac23J_1 defined below. At J1/D=66/25J_1/D=66/25, the rounded Dy2_2Ti2_2O7_7 ratio of the nearest-neighbour model, e0=−1.0675658818900694025 D=−1.5053e_0=-1.0675658818900694025\,D=-1.5053 K.

Proposition 3 (AIAO crossing). For J2=J3=0J_2=J_3=0, eMHG−eAIAO=43(J1−J1∗)e_{\rm MHG}-e_{\rm AIAO}=\tfrac43(J_1-J_1^*) with J1∗=4.52708544914207102 DJ_1^*=4.52708544914207102\,D, that is Jnn/Dnn=−0.905417089828J_{nn}/D_{nn}=-0.905417089828. For J1>J1∗J_1>J_1^* the AIAO state is lower than every MHG state.

Proposition 4 (Double-chain comparison state). Let S be one fixed ice-rule configuration of 64 spins on a 2×2×12\times2\times1 block of cubic cells, with Fourier amplitudes at 2π(12,12,0)/a2\pi(\tfrac12,\tfrac12,0)/a and its images; it is not published. Then eS−eMHG=c+23J2+2J3te_S-e_{\rm MHG}=c+\tfrac23J_2+2J_{3t} in units of DD, with c=0.0413995194175±9×10−14c=0.0413995194175\pm9\times10^{-14}, independent of J1J_1 and J3hJ_{3h}. Where J3t+J2/3+0.0207D<0J_{3t}+J_2/3+0.0207D<0, S is below MHG, so MHG is not a ground state there.

Result 5 (Energy window). At (J1,J2,J3t,J3h)/D=(2.579,−0.106,0.0189,0.0189)(J_1,J_2,J_{3t},J_{3h})/D=(2.579,-0.106,0.0189,0.0189), rounded from Yavors'kii et al. with D=1.3224D=1.3224 K 14 (not 1.41 K), the lowest energy per spin emin⁡e_{\min} of a translation-invariant state obeys −1.0617825485567 D≤emin⁡≤−1.0550325485567 D-1.0617825485567\,D\le e_{\min}\le-1.0550325485567\,D to 10−1310^{-13}. The width is 0.00675 D0.00675\,D, or 8.9 mK per spin; the upper end is the MHG energy. The ground state there is undecided: S is 11.3 mK per spin above MHG, and no ice-rule state in a cell of up to 32 spins is lower by more than 10−9D10^{-9}D (Exhaustive computation).

Why it is true

Dual. Translation-invariant (TI) states suffice, since ∑r∣W(r)∣<∞\sum_r|W(r)|<\infty. A TI state has correlations C(r)=∫eik⋅rdσ(k)C(r)=\int e^{ik\cdot r}d\sigma(k) with σ≥0\sigma\ge0 of mass 1 on the torus. For real aoa_o, one per class oo of displacements under the lattice symmetries, set f=W^−∑oaoSof=\hat W-\sum_oa_oS_o, with W^\hat W and SoS_o the cosine sums of WW and of oo. Then e=12∫f dσ+12∑oNoaoCoe=\tfrac12\int f\,d\sigma+\tfrac12\sum_oN_oa_oC_o, with No=∣o∣N_o=|o| and CoC_o the correlation on oo. A slack Λ≥0\Lambda\ge0 on all sign patterns of a small window gives ∑oNoaoCo≥y0\sum_oN_oa_oC_o\ge y_0. If f≥λf\ge\lambda then e≥(λ+y0)/2e\ge(\lambda+y_0)/2 for every TI state, with equality for a periodic state when Λ=0\Lambda=0 on the windows of its images and f=λf=\lambda at its stars, the atoms of its spectrum.

Crossing and uniqueness. Two phases share one dual when f=λf=\lambda with zero gradient at all their stars: a square linear system in the dual parameters and JJ. Both energies then equal (λ+y0)/2(\lambda+y_0)/2 at the solved JJ, the crossing of their energy lines. The cover proves f>λf>\lambda off the stars, so equality forces σ\sigma onto the stars and Λ=0\Lambda=0 on all windows of the images, and a finite enumeration returns the images. Energies are affine in JJ, so a ground state at two couplings is one between them, and the only TI one inside if the enumeration holds at one end.

Spin ice. The energy per spin is 18∫tr⁡[M(k)Σ~(dk)]\tfrac18\int\operatorname{tr}[M(k)\tilde\Sigma(dk)] with Σ~⪰0\tilde\Sigma\succeq0 of trace 4 and MM a real symmetric 4×44\times4 Bloch matrix. Hence the Luttinger–Tisza (LT) bound 11: if M(k)−λ⪰0M(k)-\lambda\succeq0 for all kk then e≥λ/2e\ge\lambda/2, with equality only on the kernels. Let λX\lambda_X be the eigenvalue at X=(2π,0,0)X=(2\pi,0,0) on E=span{e0−e1,e2−e3}E=\mathrm{span}\{e_0-e_1,e_2-e_3\}. Of the 2162^{16} Ising states of the 16-site cubic cell, 12 have Fourier amplitudes in the kernels at XX and its images and zero at the origin; they obey the ice rule (two spins in, two out of every tetrahedron) and attain λX/2\lambda_X/2. So the bound is tight, as 10 observed numerically, and an atomic spectrum forces periodicity and uniqueness. The computer-assisted step is M(k)−λXI⪰0M(k)-\lambda_XI\succeq0 on the whole Brillouin zone: Ewald sums in ball arithmetic with explicit tails, a cover by boxes closed by a second-order matrix lemma (the box lemma), and a separate lemma at XX, where the minimum is degenerate. Lowering J1J_1 raises M−λXIM-\lambda_XI, since 23I−S1(k)⪰0\tfrac23I-S_1(k)\succeq0 for the exchange matrix S1S_1, with equality on EE at XX; so J1=9/2J_1=9/2 covers all smaller J1J_1, and affinity makes the prism cost its eight vertices.

How it was checked

Theorems 1 and 2 and the lower end of Result 5 are computer-assisted proofs, Propositions 3 and 4 certified computations, the enumerations exhaustive computations. The 2D checker evaluates the dual slack on all sign patterns of the small window, exactly on the windows of the two phases and in ball arithmetic elsewhere (for the larger windows after a floating-point screen with a rigorous error bound), and the star system and the cover of f−λf-\lambda in ball arithmetic. A second program, written from the proof without the checker's code, checked the slack, the star system and positivity with its own cover, which for the four staircase couplings uses floating point with an explicit margin on one region. Bessel and Hurwitz-zeta sums for the energy lines reproduced all ten couplings. A second reader re-derived the lemmas without finding a gap; the second program shares the statement of the method. Tampered certificates, and certificates re-used past a crossing, were rejected, except modifications that are valid alternative duals. The enumerations were repeated with CP-SAT and, at c2c_2 to c5c_5 and s5s_5, a second SAT solver. Shared trust is python-flint, mpmath and the solvers.

For spin ice the cover was re-run in full for J2=J3=0J_2=J_3=0, the prism and the window, with code written from the proof. A 45-digit Ewald evaluation reproduced aXa_X and J1∗J_1^* to every printed digit, and a damped direct sum without Ewald splitting reproduced aXa_X to 6×10−76\times10^{-7}. Separately written floating-point code reproduced cc to 13 digits. Both covers follow the proof's route, so a flaw in the box lemma would be shared. The proofs have not been peer reviewed.

What it does not show

The 2D results are for classical Ising spins at zero temperature and field, not for films. Uniqueness is among TI states only. Open are the window c5<J<0.440315c_5<J<0.440315 and J>3.95J>3.95, except that stripe 7 is certified as a ground state on [4.15,4.2][4.15,4.2] and stripe 8 at J=4.45J=4.45.

For spin ice the energy is the tin-foil Ewald energy of TI states; samples of definite shape are not treated for states with net moment, which MHG and AIAO lack. Temperature, quantum effects and real materials are not addressed. The interval 9/2<J1/D≤J1∗9/2<J_1/D\le J_1^* has no certificate. The curvature of the lowest LT band at XX vanishes at J2/D≈−0.0563J_2/D\approx-0.0563 and +0.0226+0.0226 for J1/D=66/25J_1/D=66/25 and J3t=J3hJ_{3t}=J_{3h} (numerical); the proof stops near −0.055-0.055 on the negative side. S is one configuration, identified with the double-chain state of 13 by propagation vector only, and not shown to be a ground state anywhere; identifying J3tJ_{3t} with J3aJ_{3a} of 13 is an inference.

Earlier and parallel work

2D model. At J=0J=0 the Néel state is known from reflection positivity 6, 7. Giuliani, Lebowitz and Lieb 6 proved matching large-JJ energy asymptotics for 2<p≤32<p\le3, without the ground state. MacIsaac et al. 8 found stripe ground states numerically; Cannas et al. 9 quote the transitions between stripes 1 and 2, 2 and 3, 3 and 4 near 1.26, 2.2 and 2.8, in the convention of 2, and a Néel-to-stripe threshold near 0.425, inside our (2×1)(2\times1) interval; Pighin and Cannas 18 repeat it and mark the stripe crossings in a figure without printed values. Rastelli, Regina and Tassi 1 reported the checkerboards ⟨1,n⟩\langle1,n\rangle, our (n×1)(n\times1). Fonseca, Rizzi and Alves 2 give the ranges 0<δ<0.41520<\delta<0.4152 (Néel), 0.4152<δ<0.44030.4152<\delta<0.4403, 0.4403<δ<1.25850.4403<\delta<1.2585 (stripe 1) and 1.2585<δ<2.17241.2585<\delta<2.1724 (stripe 2) with data from 1; 0.4403 appears earlier in 20. Our c1c_1, s1s_1 and s2s_2 agree with the digits printed in 2. We found no printed value for c2,…,c5c_2,\dots,c_5, s4s_4 or s5s_5, but read only the abstract of 1 and could not read 8, 17, the 1995 Booth et al. letter cited in 9 or the 2006 Rastelli et al. paper cited in 2; they may have some. We add the proof that each phase is a ground state among all configurations, the intervals of Theorem 1 and the certified digits.

Spin ice. Den Hertog and Gingras 19 put the boundary to AIAO at about Jnn/Dnn=−0.91J_{nn}/D_{nn}=-0.91 by Monte Carlo, and Melko et al. 12 at −0.905-0.905 by mean-field theory and Ewald energies. Gingras and den Hertog 21 had located a weak maximum of the mean-field spectrum at (2π/a)(1,0,0)(2\pi/a)(1,0,0) numerically; 12 also has the soft mode with its two-in two-out eigenvectors and the mean-field order at that vector for J1/D<J1∗J_1/D<J_1^*. McClarty et al. 10 showed numerically, for J1=J3=0J_1=J_3=0 and J2<0J_2<0, that the lowest LT eigenvectors at (0,0,2π/a)(0,0,2\pi/a) are Ising-realisable, so MHG is a ground state for J2/D>−0.057(1)J_2/D>-0.057(1). They also have the MHG and double-chain energies, the boundary J2/D=−0.0621J_2/D=-0.0621 and −0.0622-0.0622 on 128 and 1024 sites, and Dy2_2Ti2_2O7_7 with the parameters of 14 in the MHG phase. Henelius et al. 13 gave the approximate boundary J3a+J2/3+0.02D=0J_{3a}+J_2/3+0.02D=0. We add the proof, for J1/D≤9/2J_1/D\le9/2 and on the prism including positive J2J_2, the energy and crossing to 17 or more digits, the 13-digit coefficients of Proposition 4 with J3tJ_{3t} separated from J3hJ_{3h}, and the window. Paul and Refael 16 give a numerically solved semidefinite hierarchy of bounds that contains the LT bound; Shah et al. 15 used the LT bound with an Ising-realisable minimiser for a pyrochlore Rydberg-array model with a 1/r61/r^6 tail, without interval arithmetic.

Searches of arXiv, Zenodo, Crossref and the web on 9 and 15 November 2026 found no earlier proof or certified bound.

References

  1. 1.E. Rastelli, S. Regina and A. Tassi, "Phase diagram of a square Ising model with exchange and dipole interactions: Monte Carlo simulations", Phys. Rev. B 76, 054438 (2007). doi.org/10.1103/PhysRevB.76.054438
  2. 2.J. S. M. Fonseca, L. G. Rizzi and N. A. Alves, "Stripe-tetragonal phase transition in the two-dimensional Ising model with dipole interactions: Partition function zeros approach", Phys. Rev. E 86, 011103 (2012), arXiv:1206.4501. arxiv.org/abs/1206.4501
  3. 3.R. G. Melko, B. C. den Hertog and M. J. P. Gingras, "Long-range order at low temperatures in dipolar spin ice", Phys. Rev. Lett. 87, 067203 (2001), arXiv:cond-mat/0009225. arxiv.org/abs/cond-mat/0009225
  4. 4.A. Giuliani and R. Seiringer, "Periodic striped ground states in Ising models with competing interactions", Commun. Math. Phys. 347, 983 (2016), arXiv:1509.00057. arxiv.org/abs/1509.00057
  5. 5.D. Fermi and A. Giuliani, "Periodic striped states in Ising models with dipolar interactions", in The Physics and Mathematics of Elliott Lieb, EMS Press (2022), pp. 269–293, arXiv:2203.01249. arxiv.org/abs/2203.01249
  6. 6.A. Giuliani, J. L. Lebowitz and E. H. Lieb, "Ising models with long-range antiferromagnetic and short-range ferromagnetic interactions", Phys. Rev. B 74, 064420 (2006), arXiv:cond-mat/0604668. arxiv.org/abs/cond-mat/0604668
  7. 7.A. Giuliani, J. L. Lebowitz and E. H. Lieb, "Checkerboards, stripes, and corner energies in spin models with competing interactions", Phys. Rev. B 84, 064205 (2011), arXiv:1106.0922. arxiv.org/abs/1106.0922
  8. 8.A. B. MacIsaac, J. P. Whitehead, M. C. Robinson and K. De'Bell, "Striped phases in two-dimensional dipolar ferromagnets", Phys. Rev. B 51, 16033 (1995). doi.org/10.1103/PhysRevB.51.16033
  9. 9.S. A. Cannas, P. M. Gleiser and F. A. Tamarit, "Two dimensional Ising model with long-range competing interactions", Recent Research Developments in Physics 5, 751 (2004), arXiv:cond-mat/0502403. arxiv.org/abs/cond-mat/0502403
  10. 10.P. A. McClarty, O. Sikora, R. Moessner, K. Penc, F. Pollmann and N. Shannon, "Chain-based order and quantum spin liquids in dipolar spin ice", Phys. Rev. B 92, 094418 (2015), arXiv:1410.0451. arxiv.org/abs/1410.0451
  11. 11.J. M. Luttinger and L. Tisza, "Theory of dipole interaction in crystals", Phys. Rev. 70, 954 (1946). doi.org/10.1103/PhysRev.70.954
  12. 12.R. G. Melko, M. Enjalran, B. C. den Hertog and M. J. P. Gingras, "Thermodynamic properties of the dipolar spin ice model", arXiv:cond-mat/0308282 (2003). The arXiv listing gives J. Phys.: Condens. Matter 16, R1277 (2004), whose Crossref record is R. G. Melko and M. J. P. Gingras, "Monte Carlo studies of the dipolar spin ice model"; we read the arXiv version only. arxiv.org/abs/cond-mat/0308282
  13. 13.P. Henelius, T. Lin, M. Enjalran, Z. Hao, J. G. Rau, J. Altosaar, F. Flicker, T. Yavors'kii and M. J. P. Gingras, "Refrustration and competing orders in the prototypical Dy2Ti2O7 spin ice material", Phys. Rev. B 93, 024402 (2016), arXiv:1512.05361. arxiv.org/abs/1512.05361
  14. 14.T. Yavors'kii, T. Fennell, M. J. P. Gingras and S. T. Bramwell, "Dy2Ti2O7 spin ice: a test case for emergent clusters in a frustrated magnet", Phys. Rev. Lett. 101, 037204 (2008), arXiv:0707.3477. arxiv.org/abs/0707.3477
  15. 15.J. Shah, G. Nambiar, A. V. Gorshkov and V. Galitski, "Quantum spin ice in three-dimensional Rydberg atom arrays", Phys. Rev. X 15, 011025 (2025), arXiv:2301.04657. arxiv.org/abs/2301.04657
  16. 16.N. Paul and G. Refael, "Bootstrapping ground state properties of classical frustrated magnets", arXiv:2605.06784 (2026). arxiv.org/abs/2605.06784
  17. 17.K. De'Bell, A. B. MacIsaac and J. P. Whitehead, "Dipolar effects in magnetic thin films and quasi-two-dimensional systems", Rev. Mod. Phys. 72, 225 (2000). doi.org/10.1103/RevModPhys.72.225
  18. 18.S. A. Pighín and S. A. Cannas, "Phase diagram of an Ising model for ultrathin magnetic films: comparing mean field and Monte Carlo predictions", Phys. Rev. B 75, 224433 (2007), arXiv:cond-mat/0611260. arxiv.org/abs/cond-mat/0611260
  19. 19.B. C. den Hertog and M. J. P. Gingras, "Dipolar interactions and origin of spin ice in Ising pyrochlore magnets", Phys. Rev. Lett. 84, 3430 (2000), arXiv:cond-mat/0001369. arxiv.org/abs/cond-mat/0001369
  20. 20.L. G. Rizzi and N. A. Alves, "Phase transitions and autocorrelation times in two-dimensional Ising model with dipole interactions", arXiv:0903.4084 (2009). arxiv.org/abs/0903.4084
  21. 21.M. J. P. Gingras and B. C. den Hertog, "Origin of spin-ice behavior in Ising pyrochlore magnets with long-range dipole interactions: an insight from mean-field theory", Can. J. Phys. 79, 1339 (2001), arXiv:cond-mat/0012275. arxiv.org/abs/cond-mat/0012275