Magnetic order in the repulsive Fermi-Hubbard model in three-dimensions and the crossover to two-dimensions Jie Xu, Simone Chiesa, Eric J. Walter, Shiwei Zhang Department of Physics, College of William and Mary, Williamsburg, VA 23187, USA Systemsoffermionsdescribedbythethree-dimensional(3D)repulsiveHubbardmodelonacubic lattice have recently attracted considerable attention due to their possible experimental realiza- tion via cold atoms in an optical lattice. Because analytical and numerical results are limited away from half-filling, we study the ground state of the doped system from weak to intermediate interaction strengths within the generalized Hartree-Fock approximation. The exact solution to the self-consistent-field equations in the thermodynamic limit is obtained and the ground state is shown to exhibit antiferromagnetic order and incommensurate spin-density waves (SDW). At low 3 interaction strengths, the SDW has unidirectional character with a leading wave-vector along the 1 (cid:104)100(cid:105)-direction, and the system is metallic. As the interaction increases, the system undergoes a 0 simultaneous structural and metal-to-insulator transition to a unidirectional SDW state along the 2 (cid:104)111(cid:105)-direction,withadifferentwavelength. Wesystematicallydeterminethereal-andmomentum- space properties of these states. The crossover from 3D to two-dimensions (2D) is then studied n by varying the inter-plane hopping amplitude, which can be experimentally realized by tuning the a J distancebetweenastackofsquare-latticelayers. Detailedcomparisonsaremadebetweentheexact numerical results and predictions from the pairing model, a variational ansatz based on the pair- 5 ing of spins in the vicinity of the Fermi surface. Most of the numerical results can be understood 2 quantitatively from the ansatz, which provides a simple picture for the nature of the SDW states. ] l PACSnumbers: 75.30.Fv,71.15.Ap,71.45.Lr,71.10.Fd,75.10.Lp,75.50.Ee e - r t s I. INTRODUCTION larly important and timely to understand such phases in . the Hubbard model. Somewhat surprisingly, apart from t a half-filling (one particle per site), which displays anti- m Over the past several years, optical lattices have be- ferromagnetic (AFM) order, the nature of the magnetic come an increasingly powerful tool for emulating many d- systems in condensed matter physics1–4. An optical lat- properties in the 3D Hubbard model has not been char- acterized, even at the mean-field level. In this work, we n tice can provide exceptionally clean access to a vari- study the magnetic properties in the ground states of o ety of model many-body Hamiltonians in which param- c eters can be systematically tuned and controlled. Thus, the 3D Hubbard model and in the crossover regime, us- [ ing generalized Hartree-Fock (HF) theory. It is shown they make possible quantitative experimental study of that the system has a tendency to form unidirectional 1 thepropertiesofinteractingelectronmodels, whichhave v proven extremely challenging for analytic and numerical spin-density wave (SDW) states with AFM order and a 6 approaches alone. The combination of these approaches modulating wave along either the (cid:104)100(cid:105)- (at low U/t) 8 or the 111 -direction (at higher U/t). We examine the 1 presents unprecedented opportunities for improving our evolutio(cid:104)n of(cid:105)the SDW wavelength in the full mean-field understanding of interacting electron systems, by test- 6 solution as U, density and t vary and characterize the 1. ing theoretical concepts and increasing the accuracy and ground state by its propert⊥ies in real and momentum predictivepowerofnumericalapproachesviacomparison 0 space. with experiment. 3 1 The one-band Hubbard model is one of the most fun- Despite the simple nature of the mean-field approach, : damental models in condensed matter physics. It has thedeterminationofthecorrectequilibriumpropertiesin v been widely studied in two dimensions (2D)5–20 as the these models is not straightforward20,22. The main chal- i X simplest model for the Cu-O plane in cuprate supercon- lenge lies in finding an unbiased strategy to determine r ductors. Forthethree-dimensional(3D)Hubbardmodel, the leading wave-vector(s) characterizing the spatial de- a however,considerablylessisknownfromboththeoretical pendence of the order parameter. Calculations are per- and experimental sides. Optical lattices play, in this re- formed in a real space simulation cell and most choices spect, aparticularlyfundamentalroleastheyallowfora of the cell will return solutions that are biased by finite- clean experimental realization of the 3D model and offer size effects. This is often further complicated by shell theinterestingpossibilityoftuningthehoppingparame- effects and sensitivity of the solution to the topology of teralongonedirection,t ,therebyallowingasystematic the Fermi surface, which often lead to local minimum studyoftheevolutionof⊥propertiesasthesystemcrosses solutions. over from 3D to 2D. To overcome the difficulties, it is necessary to move to Thankstoadvancesintheabilitytodirectlycoolatoms larger and larger cells and gain insights from the evolu- in optical lattices21, experiments are nearing the realiza- tion of the corresponding solutions. This line of attack tion of phases with magnetic order. It is thus particu- hasbecomeincreasinglypossiblebecauseofthedramatic 2 increase in computing power and continuous algorithmic the Hubbard Hamiltonian reads progress. In the present work, we combine such an ap- proach with more targeted searches to obtain the global (cid:88) (cid:16) (cid:17) minimumsolutionoftheself-consistent-field(SCF)equa- H = −t c†rzσcr(cid:48)zσ+c†r(cid:48)zσcrzσ tionsinthethermodynamiclimit. Furthermore,weshow rr(cid:48) ,z,σ (cid:104) (cid:105) (cid:88) (cid:16) (cid:17) htioownatlhaennsautmzebraicsaeldroensutlhtsecpaanirbineguonfdseprisntsooindtbhyeavivcainriitay- −t⊥ c†rzσcrz(cid:48)σ+c†rz(cid:48)σcrzσ r, zz(cid:48) ,σ of the Fermi surface. Detailed comparisons are made (cid:104) (cid:105) (cid:88) between the direct numerical solutions and the pairing +U nrz nrz , (1) ↑ ↓ ansatzpredictions. Theexcellentagreementhelpstopro- r,z vide a simple, predictive picture for the properties of the SDW states. wheretheoperatorc (c )creates(annihilates)apar- †rzσ rzσ The mean-field approach is often the starting point ticlewithspinσ(σ = , )atsite(r,z)andn isthecor- rzσ in the study of strongly interacting systems such as the ↑ ↓ respondingnumberoperator. Thehoppingamplitudetis Hubbard model. Although the approximations involved between nearest neighbor sites within a plane (denoted can lead to significant errors, mean-field theory often by rr in the summation), t is the inter-plane hop- (cid:48) provides insights into qualitative aspects of the behav- ping(cid:104)am(cid:105)plitudebetweenneares⊥tneighborsitesbelonging ior of many-body systems. Moreover, comparisons with to different planes (denoted by zz in the summation), (cid:48) quantum Monte Carlo results19 have shown20 that, in (cid:104) (cid:105) and U > 0 is the on-site interacting strength. Through- 2D, the mean-field solution captures the basic physics of out this work, energy is quoted in units of t and we set SDW states at intermediate interaction strengths, and t=1. TheHamiltonianinEq.(1)describesthe3Dcubic providesagoodqualitative(orevenquantitativeinsome Hubbard model when t =1, the crossover between the aspects) description of the magnetic correlations in the squareandcubiclattice⊥swhen0<t <1andastackof true ground state. Because it is reasonable to expect a decoupled 2D Hubbard planes when⊥t =0. Only unpo- similar level of accuracy for the models presently consid- larizedsystemsareconsidered, i.e., th⊥eaveragedensities ered,weexpectthatourfindingswillprovideguidanceto of the two spin species are kept equal: n = n . The many-body approaches and experimental studies alike. nature of the ground state is thus character↑ized b↓y three WehavelimitedourstudytoU (cid:46)6t,wherethemean- parameters: the inter-plane hopping amplitude t , the fieldapproachcanbeexpectedtoofferusefulinsight. Be- on-site repulsion U and the doping (hole density)⊥ low we will discuss the mean-field predictions and their implications (and the caveats) on the true many-body states drawing from the comparison in 2D19,20. Clearly, h 1 (n +n ). (2) ≡ − ↑ ↓ the form of generalized mean-field theory considered in thisworkwillnotcaptureexoticinstabilities,suchasun- Theparticle-holetransformation,c ( 1)x+y+zc , conventionalpairingorder. Indeed,asU increases,itwill maps the h < 0 sector into the h >†rzσ0→one−, regardlesrszσof become increasingly inadequate for magnetic properties the value of t or U, and we therefore confine our study as well. to h>0. ⊥ The remainder of the paper is organized as follows. In Athalf-filling,h=0,thenon-interactingFermisurface Sec.II,weintroducetheHamiltonian,andbrieflyoutline is given by 2(cosk +cosk +t cosk ) = 0. Despite some of its basic properties to facilitate the ensuing dis- thelackofs−ymmetryxbetweenythe⊥z-andzther-directions cussion. In Sec. III, we summarize the strategies used to for any t = 1, perfect nesting via Q ( π, π, π) solvethemean-fieldequations. Resultsforthe3Dmodel remains t⊥hr(cid:54)oughout the whole surface,≡an±d ca±uses±an are presented in Sec. IV; numerical results for the 100 - (cid:104) (cid:105) AFM instability for any t = 0 and arbitrary small U and the (cid:104)111(cid:105)-SDW are followed by a discussion where values. The evolution of ⊥the(cid:54) non-interacting half-filled the pairing ansatz is first introduced and then used to Fermi surface as t varies is shown in Fig. 1. In the first helpunderstandthenumericalfindings. Thedimensional column, representi⊥ng the 2D limit, the Fermi surface has crossover results are then presented in Sec. V, followed no dependence on k and any wave-vector of the form again by a discussion based on insights from the pairing z ( π, π,q)isperfectlynestedonit. Thearbitrarinessof ansatz. We conclude in Sec. VI. ± ± q is reflected in the complete lack of correlation between the r-planes. The large nesting degeneracy is abruptly interrupted as soon as t = 0, and Q remains the only II. BACKGROUND nesting vector as the sy⊥ste(cid:54)m evolves from the 2D limit toward 3D. The middle and bottom rows illustrate pro- Given our goal to study both the 3D case and the jections of the Fermi surface along the 100 - and 111 - (cid:104) (cid:105) (cid:104) (cid:105) crossover from 3D to 2D, it is most convenient to define directions; as we shall see in Sec.’s IV and V, the pro- the3DHubbardHamiltonianasastackofsquare-lattice jected area of the Fermi surface plays a central role in planes. We will use r (x,y) to denote in-plane coor- determining the character of the SDW in the proximity ≡ dinates and z to label the planes. With this convention of h=0. 3 are N N matrices with elements × [H (k)] = t (k)+δ (UD µ), σ βγ βγ βγ βσ − − (5) [S±(k)]βγ =UδβγSβ±, (cid:80) where tβγ(k)= Lexp(ik·L)tβ,γ+L, and Dβσ, Sβ± and µaredeterminedbytherequirementthatthefreeenergy F = H TS is a minimum for the targeted average 0 0 (cid:104) (cid:105) − density n=n +n . This amounts to the following SCF (gap) equation↑s ↓ (cid:90) 1 Dβ,−σ =(2π)3 dk(cid:104)c†βσ(k)cβσ(k)(cid:105)0 (cid:90) 1 Sβ± =− (2π)3 dk(cid:104)c†β,±σ(k)cβ,∓σ(k)(cid:105)0 (6) (cid:90) 1 (cid:88) n=N(2π)3 dk(cid:104)c†βσ(k)cβσ(k)(cid:105). β,σ To locate the ground state we proceed with two com- plementary approaches. In the first approach we select the L ’s so that they span a large supercell containing i (5000) sites. A twisted boundary condition23,24 is ap- O plied, namely, using a single randomly selected k-point FIG. 1: (Color online) Non-interacting half-filled Fermi sur- in place of the integrals in Eq. (6). The iterative process face from different view angles: 3D (top), along [010] (mid- is started with various initial states, including random dle),along[11¯1](bottom). Fromlefttoright,thecolumnsare ones, and multiple annealing cycles are performed. In fort =0,0.5and1,respectively. Notethatonlythesurface ⊥ each cycle a random perturbation (whose strength can inoneoctantisshowninthebottomrow. Perfectnestingvia be controlled) is applied to a converged solution and the Q=(±π,±π,±π) holds for any t at half-filling. ⊥ self-consistent process is repeated. Separate calculations for different k-points are done to check for consistency. Once an understanding of the character of the ground III. METHOD state is gained, we use a second approach to target the specific family of states compatible with the results of The following mean-field formalism in real space is the random search. For instance, suppose the random used in this work. A simulation cell of N sites is defined searchfindsaunidirectionalSDWatsmallU valueswith by three vectors, L1, L2 and L3, whose components are wave-vector along the 100 -direction. We then choose a (cid:104) (cid:105) integers. Bloch states are then introduced as cluster of L = (L,0,0), L = (0,1,0) and L = (0,0,1) 1 2 3 with G=π(( 1)L+2l,1,1), where l is the number of os- (cid:88) (cid:2) (cid:3) − c (k) c exp ik L , (3) cillationsoftheorderparameter, chosentobeaninteger β β+L ∝ · orhalfaninteger. Foragivensetofthethreeparameters L (t , U, h), Lisfinelyscanned(withLontheorderof50 whereLarevectorsoftheformL=n L +n L +n L , an⊥d step size of 1) until the energy minimum is found. 1 1 2 2 3 3 k is a reciprocal lattice vector that is free to vary within A large number of k-points is used (on the order of 100 the first Brillouin zone (BZ) defined by the L ’s, and β in the two short directions and a few in the other) so i labels sites inside the simulation cell. Using these states, that the character and properties of the targeted states the mean-field Hamiltonian can be decoupled into a sum can be accurately determined. This approach allows us (cid:80) of k-dependent pieces, H = H (k), with each piece to study different forms of SDW and long wavelength 0 k 0 of the form modes without increasing the computational cost. In our study, we mix the two numerical approaches H0(k)=[c†↑c†↓](cid:20)HS↑(+k) H↓(kS−−G)(cid:21)[c↑c↓]T, (4) aexsanmepedlee,dcoamndpaursiseonthoefmeninergcioemspalmemonegntsaervyerwaalyfas.miFlioesr of SDW is made with the second approach. To confirm where c (c ) represents a row of operators c (k) thecorrectnessofthegroundstate,thesolutionsarethen β (c (k ↑G))w↓ithindexβ runningthroughtheN site↑sof checkedagainstdifferentinitialstatesandannealingpro- β the↓cel−l. A non-zero value of G causes the spin densities cedures using the first approach on a supercell commen- at β and β +L to be related via a rotation by G L surate with the optimal wavelength. i i · around the z-axis. Charge and spin densities along z- Various observables are computed to characterize the direction obey periodic boundary conditions. H and S converged solutions. The local charge density ρ and the ± 4 16 16 16 14 12 12 12 z 8 z 8 z 8 z 7 4 4 4 16 12y8 4 4 8x 12 16 16 12y8 4 4 8x 12 16 16 12y8 4 4 8x 12 16 16 12y8 4 4 8x 12 16 0.896 0.898 0.9 0.902 −0.1 0 0.1 −0.2 0 0.2 −0.2 0 0.2 16 16 FIG. 3: (Color online) Spatial dependence of the order 12 12 parameter m for h = 13/128, same as in Fig. 2, but with U =2.9,ona16×16×16supercell(left)anda16×16×14 z 8 z 8 supercell (right). Uniform AFM order in the xy-plane disap- pears for L = 16, but linear SDW along z-direction is seen z 4 4 again on the right with L =14. z 0.894 0.897 0.9 0.903 −0.12 −0.06 0 0.06 0.12 ρ m with c (cid:80) exp( ik R)c . We use n to compare kσ ∝ R − · Rσ k theconvergedmean-fieldsolutionwiththepairingansatz FIG.2: (Coloronline)Chargedensityρ(left)andorderpa- predictionandAktocharacterizetheFermisurfaceofthe rameterm(right)ofthesolutionforthe3DHubbardmodel. ordered phase. Shownisa16×16×16supercell,withh=13/128atU =2.5. Alinearwaveisseenalongthez-direction,withuniformAFM order in the xy-plane. The bottom panel shows a line cut IV. 3D RESULTS along z-direction, with dots the actual data and the line a sinusoidal fit. A. SDW correlation in the (cid:104)100(cid:105)-direction local order parameter, identified as the local staggered At half-filling, the existence of perfect nesting allows magnetization m, are defined as an AFM solution for any U >0. Away from half-filling, perfect nesting ceases to exist and a finite critical value ρ(R) n + n , (7) rz rz of the interaction is needed to cause the onset of order. ≡ (cid:104) ↑(cid:105) (cid:104) ↓(cid:105) m(R) ( 1)x+y+z( n n ), (8) The critical value U depends on h. Using the first ap- rz rz c ≡ − (cid:104) ↑(cid:105)−(cid:104) ↓(cid:105) proach described in Sec. III, we have determined that, and used to characterize the state in real space (here justaboveU , thegroundstateofthesystemisanSDW c R (r,z)). Since all the minimum energy solutions withmodulationalongthe 100 -direction. Figure2illus- we≡find are unidirectional spin/charge density waves tratesthespatialdependen(cid:104)ceof(cid:105)ρandmina16 16 16 (SDW/CDW),itisnaturaltocharacterizethembytheir supercell at h = 13/128 0.10 and U = 2.5. T×he S×DW modulation wavelength along a relevant Cartesian axis, is characterized by a sin(cid:39)gle wave-vector and λ = 8. 100 λSDW/CDW, defined by the leading component of the TheamplitudeoftheSDWis 0.1,roughlythi(cid:104)rty(cid:105)times Fourier transform of ρ and m, respectively. The min- larger than that of the CDW(cid:39). The simple form of the ima in the CDW are found to coincide with nodes in order found for m(R) is indicative of the proximity of U the SDW, thus 2λCDW = λSDW. Below we will some- to Uc. All of the observations above are consistent with times discuss our results in terms of a single wavelength the pairing model, as discussed below in Sec. IVC. λ λCDW, which can also be identified as the distance We next examine the evolution of λ as the inter- ≡ 100 between two consecutive nodes of the order parameter. action strength changes. Keeping h an(cid:104)d t(cid:105)he simulation Whenwerefertothedirectionofthemodulation,wewill cellunchangedandincreasingU from2.5to2.9,theran- use 100 todenotesymmetry-equivalent[100]-directions, dom search returns the state displayed on the left panel (cid:104) (cid:105) and similarly for [110] and [111]. of Fig. 3, which suggests that a more complicated type To characterize the system in momentum space, we of order is seemingly settling in. We apply our second use the momentum distribution nk and the momentum- approach,usingadensek-pointgridanda1 1 Lsim- resolved single-particle spectral function Ak(ω), defined ulationcell,tosearchfortheoptimalwavelen×gth.×Weuse as large L (containing about 8 nodes in the cell) and vary its value until an energy minimum is found. Figure 4 nkσ =(cid:104)c†kσckσ(cid:105), (9) shows the result of such minimization in terms of λ 100 ; Akσ(ω)= π1Im(cid:104)ckσ(ω−H +E0−iη)−1c†kσ(cid:105) (10) theminimumoccurswhenλ(cid:104)100(cid:105) =7,indicatingtha(cid:104)tth(cid:105)e 5 −1.403 1 1 L h=1/32 16× 1×6 16 −1.4031 16×16×14 0.85 h=1/16 × × h=2/25 h=3/32 h=13/128 −1.4032 0.8 h=3/25 e h=1/8 t i si −1.4033 00 h=3/20 E/ α1h0.75 h=1/5 −1.4034 0.7 −1.4035 −1.4036 0.65 6.4 6.6 6.8 7 7.2 7.4 7.6 7.8 8 8.2 2 2.5 3 3.5 4 4.5 5 λ 100 U h i FIG.4: (Coloronline)Energyof(cid:104)100(cid:105)-SDW(blue)vs. λ FIG. 5: (Color online) Characteristic wavelength as a func- (cid:104)100(cid:105) for a system of h=13/128 and U =2.9. Horizontal lines are tionofU atvariousdoping. α givesthemodulationwave- (cid:104)100(cid:105) theenergiesofthecalculationsshowninFig.3. Theminimum length, λ , in units of 1/h. As U is increased, the value (cid:104)100(cid:105) of(cid:104)100(cid:105)-SDWisreachedwhenλ =7. Thestateintheleft of α converges to approximately 2/3 at small h (slightly (cid:104)100(cid:105) (cid:104)100(cid:105) panel of Fig. 3 leads to an energy higher than the minimum larger at larger h). but lower than the energy with λ =8. (cid:104)100(cid:105) in Fig. 5, with α defined as 100 (cid:104) (cid:105) 16 16 16supercellisnotcommensuratewiththewave- α =hλ . (11) len×gth×of the minimum energy SDW state and that the (cid:104)100(cid:105) (cid:104)100(cid:105) patternintheleftpanelofFig.3isaresultoffrustration Whenthedopingissmall,thewavelengthofthemodula- from an incommensurate supercell size. We next return tion is proportional to 1/h, with α almost indepen- 100 to our first approach, and perform a new mean-field cal- dent of U and roughly equal to 2/3(cid:104). Fo(cid:105)r larger h, α 100 culation, with random initial guess and annealing, on a converges to a slightly larger value. There is a gen(cid:104)era(cid:105)l 16 16 14supercell,asizewhichiscommensuratewith trend of an increase of α as U approaches U from 100 c the×wav×elengthoftheminimumenergysolution. Andin- above. We will be able to(cid:104)rat(cid:105)ionalize these trends within deedwefindthepredictedstatewithλ =7correctly the pairing model in Sec. IVC. 100 reproduced (right panel in Fig. 3). (cid:104) (cid:105) The evolution of the properties of the 3D SDW state with interaction U is similar to what is observed in 2D. In Fig. 4, we report the energies of the two large su- Figure 6 shows 1D cuts of m and ρ in the z-direction, at percell calculations of Fig. 3, to verify that the energy h=0.05 and U values such that α has saturated to odbutcaeidnebdyinthtehe1161×16L×c1lu4stseurpesercaerlclhiswciothrreλctly re=pro7-. ∼2/3. The figure illustrates the ex(cid:104)i1s0t0e(cid:105)nce of Uc and the The energy of th×e 16× 16 16 calculation, on(cid:104)t1h00e(cid:105)other increase of the SDW and CDW amplitudes with U. It × × also shows the crossover from a regime where the order hand, falls between those of λ =7 and 8. This gives 100 parameter has a smooth sinusoidal modulation and the a clear illustration of the chara(cid:104)cte(cid:105)ristics of the two types holes are delocalized, to one where it is characterized by of approaches. The supercell being incommensurate pre- domainwallsorstripes, withholeslocalizedinthenodal ventsthesolutionfromcollapsingontothelowestenergy regions. SDW state of λ = 7. The self-consistent solution in 100 There exists an important difference in the physics of a large supercel(cid:104)l th(cid:105)en finds a different pattern that cor- the mean-field ground state of the 3D system and its 2D responds to the true ground state compatible with the counterpart. While, inthelatter, thesystemremainsin- imposed constraint. The energy of this state, computed sulating when lightly doped, the 3D model immediately by fixing the density and converging the value using a turns metallic. The difference is a consequence of the dense k-point mesh, is higher than the global minimum, differentbehaviorsofthemodulatingwavelength. In2D, but lower than that of the SDW state with an imposed α is unity, independent of h and U, while in 3D it varies wavelength λ =8. 100 with parameters and has a non-integer value. To illus- (cid:104) (cid:105) We proceed to determine the exact dependence of the trate this in a simple case, consider a value of doping h wavelengths on h and U by explicit solutions of the SCF such that λ ( α/h) is an integer. For such a system to ≡ equationsin1 1 Lclusters. Verificationsoftheresults beaninsulator,thenumberofparticlesina1 1 λcell, × × × × are done on large supercells whose sizes are commensu- (1 h)λ=λ α,willhavetobeaninteger. However,be- − − rate with the wavelengths. Our results are summarized cause α 2/3 in the limit of small doping, the condition ∼ 6 1 0.96 22 0.92 ρ0.88 16 0.84 12 0.25 U =2.0 0.2 U =2.5 z11 z 8 0.15 UU ==33..05 4 0.1 0.05 16 12 16 m 0 16 8 12 −−0.00.51 1 2y8 4 4 8x 12 16 y 4 4 8x −0.15 −0.2 −0.5 0 0.5 −0.5 0 0.5 −0.25 1 14 27 FIG. 7: (Color online) Order parameter m at h = 1/8 and z U =5.0 in a 16×16×22 supercell (left) and a 16×16×16 supercell (right). Though the supercell of 16×16×22 is FIG.6: (Coloronline)Chargedensityρ(top)andorderpa- commensurate with the optimal (cid:104)100(cid:105)-wavelength, the ran- rameterm(bottom)vs.U. Thesystemhasdopingh=0.05. dom search produces a lower energy solution. The minimum Eachcurveisa1Dcutalongz-direction, thedirectionofthe energy solution is an SDW along the (cid:104)111(cid:105)-direction, which modulation. Beyond Uc, the (cid:104)100(cid:105)-SDW/CDW amplitudes is correctly reproduced by a random search in a 16×16×16 increase with U and the solution evolves from a sinusoidal supercell as shown on the right. wave to domain walls. The CDW amplitude is much weaker than that of the SDW. U = 4.0 U = 4.5 −1.231 −1.158 cannot be satisfied, and the system is necessarily metal- e −1.232 −1.159 t lic. A related way to see this is to consider the case of si −1.233 −1.16 / domainwallstates,forexampleU =3.5inFig.6. Inside E −1.234 −1.161 eachdomainwall(nodalregion)arelocalizedholeswhose −1.235 integrated (along the direction of the modulation) den- 5 6 7 8 9 5 6 7 8 sity is α. Thus the domain wall as a whole will act as a U = 5.0 U = 5.5 quasi-2Dliquidofholeswithnon-integerdensity. Wewill −1.034 discuss the corresponding momentum space signature in −1.094 the next section. te −1.036 i s / −1.096 −1.038 100 E h110i B. SDW modulation along the (cid:104)111(cid:105)-direction −1.04 h111i −1.098 h i 5 6 7 8 5 6 7 8 AsshownaboveandfurtherdiscussedinSec.IVC,an λ λ orientation of the SDW other than 100 is not the so- lution in the proximity of U . Howev(cid:104)er, w(cid:105) hen the inter- FIG. 8: (Color online) Energies per site of SDW in (cid:104)100(cid:105)-, c (cid:104)110(cid:105)- and (cid:104)111(cid:105)-directions vs. λ for h=1/8. At and above action grows larger, other Fermi liquid instabilities be- U =4.5,thelowestenergystateismodulatedalongthe(cid:104)111(cid:105)- come possible. This fact is clearly displayed when cal- instead of (cid:104)100(cid:105)-direction. culations on supercells commensurate with the optimal 100 -wavelength for a given U do not yield a state with (cid:104) (cid:105) 100 -SDWorder. Figure7showstheoccurrenceofsuch (cid:104)a cas(cid:105)e in a calculation with h = 1/8 and U = 5.0, for lowest energy state is given by a 111 -SDW, instead of (cid:104) (cid:105) which λ = 5.5. The 16 16 22 supercell should the 100 -order atlowerU. For U =5, theminimum en- have pre(cid:104)c1i0s0e(cid:105)ly accommodated×4 no×dal planes of the or- ergy(cid:104)solu(cid:105)tioniscorrectlyreproducedbyarandomsearch der parameter, but rather than doing so, the random in a 16 16 16 supercell as shown in the right panel of × × search produces the lower energy solution shown in the Fig. 7. In our searches, 110 -direction SDW’s are never (cid:104) (cid:105) left panel of the figure. found to be the global ground state. To search for the solution at higher U, we investigate By repeating the same procedure, we construct the unidirectionalSDW’swithmodulationlyingalongeither equation of states for U = 3 and U = 4 contained in the 110 - or 111 -direction. An example is given in Fig. 9. At U = 3, there is no density regime where the (cid:104) (cid:105) (cid:104) (cid:105) Fig. 8. The energies from constrained searches using the 111 -SDW is the global ground state. In contrast, for (cid:104) (cid:105) second approach are shown as a function of λ for a scan U = 4, a discontinuous transition from 100 to 111 (cid:104) (cid:105) (cid:104) (cid:105) of U values. It is seen that, at and above U = 4.5, the occurs around n = 0.9 with a small coexistence region. 7 5 1.0 themodelhelpstoexplainthenumericalresultsandpro- 0.02 vides a simple conceptual framework that captures the 4 [111] 0.01 [100] essential physics of the SDW states in 3D and in the 3 0 [110] 0.5 crossover regime discussed in the next section. We first -3 t ]10 2 0.99 0.995 1 soufm1m00ar-iSzeDWtheaftormmoadleissmt U, a,nfodlltohweendabpypltyheit 1to11th-SeDcaWse. y [ 0.0 A(cid:104)t lo(cid:105)w U and small h, the pairing model i(cid:104)s defi(cid:105)ned by g 1 ner spin-orbitals of the form e 0 -0.5 -1 φ†kσ =ukc†kσ+σvkc†k+qkσ. (12) UUU === 333 UU == 44 The construction requires excitations to outside the -2 -1.0 Fermi sea. It is reminiscent of the ansatz used to con- 0.88 0.92 0.96 0.88 0.92 0.96 1.00 struct the BCS pairing states for attractive interactions, density density except that here the tendency for small excitations is FIG. 9: (Color online) Ground-state energy per site from perhaps more “natural”, because of the repulsive inter- constrainedsearchof(cid:104)100(cid:105)-,(cid:104)110(cid:105)-and(cid:104)111(cid:105)-SDWatU =3 action. A collection of spin pairs in orbitals given by (left) and U = 4 (right). A linear common shift has been Eq. (12) leads to a uniform charge density, ρ(R) = n, applied to the energies to highlight the convexity and the and a spin density of the form different trends. At U = 3, (cid:104)111(cid:105)-SDW is not the global groundstate. (Theinsetshowsazoomoftheenergydifference 4 (cid:88) s(R)= a cos(q R) (13) between (cid:104)111(cid:105)- and (cid:104)100(cid:105)-SDW states as n→1.) At U =4, N k k· the ground state is (cid:104)111(cid:105)-SDW for n(cid:38)0.92. k ∈R with a = u v . The region over which k is summed k k k R will be closely related to the non-interacting Fermi sea, In both cases the low-doping ground state is character- andpreservingthevolumeof4π3n,butwillingeneralbe ized by a linear energy-density dispersion. This bears slightlymodifiedfromthevariationaloptimization,aswe two important consequences. First, contrary to what is further discuss below. To ensure orthogonality amongst observedusingvariationalstateswithauniformspiralor- the spin-orbitals, q must be such that k+q and k k derparameter25,thereisnosignofphaseseparationinto k+qk = k(cid:48)+qk(cid:48). The corresponding potentia(cid:54)∈l eRnergy a half-filled, AFM region and a hole-rich region. Second, per site(cid:54)is then given by the effective interaction between domain walls is short- rangedandtheirpreciselocationinthehole-dilutedlimit U (cid:88) V =Un2 s2(R). (14) is therefore irrelevant as long as they stay sufficiently far − 4N R apart. We find α =1 at any density for which the 111 - The potential energy lowering relative to the paramag- 111 SDW is the(cid:104)gro(cid:105)und state. The 111 -SDW state(cid:104)s ar(cid:105)e netic (PM) solution is thus: (cid:104) (cid:105) fullygapped,owingtotheintegervalueofα ,incon- 111 trast to the metallic behavior of the 100 s(cid:104)tate(cid:105)s. Upon U (cid:88) increase of U at a constant h, the st(cid:104)ruct(cid:105)ural transition ∆V = −N2 akak(cid:48)[δ(qk+qk(cid:48))+δ(qk−qk(cid:48))], k,k(cid:48) is therefore always accompanied by a metal-to-insulator ∈R (15) transition. We have verified, for selected cases, that ran- dom searches on larger supercells with sizes commensu- where the Kronecker δ is intended as periodic on the rate to the optimal wave-vector always return unidirec- reciprocal lattice, i.e., modulo 2π in any direction. tional SDW’s with the same predicted wavelength and Equation(15)makesitclearthatthemaximumreduc- orientation. This provides a strong indication that the tioninV isachievedbyhavingasmanypairsaspossible character of intermediate U instabilities remains that of withq ’swhichareparalleloranti-paralleltoeachother. k a unidirectional SDW. Thus, as we increase U at con- NotingthatthevectorQisperfectlynestedwhenh=0, stant density, the system is always expected to undergo let us consider the following explicit construction for : a discontinuous transition from a 100 - to a 111 -SDW R displacethehalf-filledFermisurfaceineachoctantofthe (cid:104) (cid:105) (cid:104) (cid:105) ground state. first BZ by ∆q/2, choosing the direction that shrinks ± the Fermi sea and with a length ∆q such that the en- closed volume is reduced to 4π3n. This construction is C. A variational pairing ansatz illustrated in the top panel in Fig. 10 for the case of the 100 -SDW (discussed next in Sec. IVC1). The sur- (cid:104) (cid:105) The pairing model is a variational ansatz that has face of can now anchor spin pairs with one common R proved extremely helpful in rationalizing the properties pairing vector, by making u less than 1 in a small layer k of SDW’s in the mean-field treatment of the electron immediatelyinsidethesurfaceof ,andcorrespondingly (cid:112) R gas22,26 and the 2D Hubbard model20. Similarly here, v = 1 u 2 >0. k k −| | 8 √2π 0 √2π estimate of α (Eq. (11)) −π π 1 (cid:90) ∆q 1 2 h100i 0.9 α∆q = 4π2 e e (cid:98)e∆q·dS. (17) (cid:98)∆q (cid:98) · ∆q S 100 h i Q 0.8 where (cid:98)e is a relevant Cartesian unit vector. z0 0 Different directions of ∆q lead to different reconstruc- k tions of the non-interacting doped Fermi surface. The q h100i 0.7 kinetic energy cost of the pairing ansatz can therefore be thought of as the combination of two contributions: the reconstruction energy due to using rather than 0.6 R the true non-interacting Fermi sea, and the kinetic en- π π ergychangeduetomovingparticlesfromk(inside )to −√2π 0 √2−π R − k[110] 0.5 k+qk (outside), i.e., the non-zero vk’s. It is easy to see that,similarto2D20,atsufficientlylargeU thepotential π 0 π π− π energy lowering will overtake the kinetic energy increase 0.4 for the constructions discussed here. The correct state is determined by maximizing the gain in the potential energy from pairing (larger areas near the Fermi surface 0.3 participating with parallel ∆q) while minimizing the ki- netic energy cost. y0 0 k 0.2 Theansatzgivesaclearpicturefortheonsetofthein- stability. First, by its form, the model captures how the energetically costly CDW can be suppressed compared 0.1 to the SDW. Second, it indicates that amongst different possiblechoicesofq ,theoneinvolvingonlyparalleland k anti-parallel vectors are optimal. Third, the direction of π π − π 0 π− ∆qmustbesuchastoleadtotheminimalpossiblere- − ± kx construction of the doped Fermi surface. These findings are in qualitative agreement with the numerical results FIG. 10: (Color online) Illustration of the pairing model for obtained by the solution of the SCF equations: 1) the the(cid:104)100(cid:105)-SDWstatein3D.Theschematicdiagramisdrawn SDW is much stronger than the accompanying CDW, on the actual momentum distribution from the exact numer- 2) the Fermi surface reconstructs in a way to enhance ical solution for h = 1/8 and U = 4.0. The top panel shows pairing and the SDW order tends to be unidirectional as thepairingconstructiononthecontourplotofthe(1¯10)cut. a result, 3) more drastic reconstructions are only possi- Thewhitedashedlinesrepresentthehalf-filledsurface,across ble with larger U. Much quantitative information can which is the nesting vector Q. The reconstructed surface, be obtained with straightforward calculations using this shown as magenta solid lines, is obtained by displacing the half-filled one along the z-direction by a distance ∆q /2, model, as we discuss next for the 100 - and 111 -SDW (cid:104)100(cid:105) (cid:104) (cid:105) (cid:104) (cid:105) as given by Eq. (16). The nesting vector across the shifted states, respectively. surface,q ,isshownbythelongsolidlinewitharrow. The (cid:104)100(cid:105) bottom panel shows n(k) of the k = π plane. This is in a z regionwheretheFermisurfacesurvivestheonsetoforderand 1. Analysis of the (cid:104)100(cid:105)-SDW where it differs more severely from the pairing construction. The actual Fermi surface, seen distinctly inside the recon- Among all directions, only a ∆q along the 100 - structed surface, differs little from the non-interacting Fermi (cid:104) (cid:105) direction causes the Fermi surface in each octant to surface(blacksolidline). n dropssharply,andnopairingis k shrink equally and this, it can be shown, leads to the present in this region. minimum kinetic energy increase at small h. An SDW with 100 modulation is thus the lowest energy solution (cid:104) (cid:105) For small h, we can determine ∆q directly from the at low h and just above Uc, consistent with the results construction: fromexplicitsolutionsoftheSCFequationsinSec.IVA. (cid:90) We numerically calculate the projected area along the Ω ∆q e dS=h BZ, (16) 100 -direction(shownintherightmiddlepanelinFig.1) (cid:98)∆q· 2 (cid:104)and(cid:105)obtain α 0.63, in very good agreement with 100 S the exact resu(cid:104)lts(cid:105)fr(cid:39)om direct solutions shown in Fig. 5, where S is the half-filled surface, e is the direction whereα 0.66. Thattheestimatedvalueisslightly (cid:98)∆q 100 of ∆q, and Ω = (2π)3 is the volume of the first BZ. smalleri(cid:104)sco(cid:105)n(cid:39)sistentwiththepresenceofsurvivingFermi BZ Equation(16)impliesalinearrelationshipbetweenhand surface inside the reconstructed doped Fermi surface as ∆q and, using λ = π/(∆q e), provides the following seen in Fig. 10. (cid:98) · 9 √2π 0 √2π −π π 30 π π 2 2 25 z 0 0 k 20 π π −2 −2 15 FIG. 12: (Color online) (a) Half-filled Fermi surface for the π π 3D Hubbard model in one of the octants (left) and (b) aver- −√2π 0 √2−π − k[110] aged∆q(cid:104)100(cid:105) overdifferentareasvs.doping(right). Theblue π kz = 0 kz = π/2 kz = π 10 stholeidblluineediont(ibn)(ias)c.aTlchueladteadshaetd/(−doπt/t2ed,−liπn/e2s,iπn/(2b)), sahroewanvears- aged over varying areas as indicated by the circles with the samecolor/stylein(a). Theblacklinein(b)showstheaver- 5 y0 agedvalueovertheentiresurface,whichleadstotheestimate k of α ∼0.63 discussed in the text. (cid:104)100(cid:105) π − π 0 π π 0 π π 0 π − k − k − k part of the Fermi surface survives the onset of order; the x x x parts that do not survive are in areas around the “hot FIG. 11: (Color online) Contour plots of the single-particle spots” k ( π/2, π/2, π/2), where the pairing con- spectralfunctionevaluatedattheFermienergy;onthe(1¯10) struction(cid:39)and±the d±oped F±ermi surface are most similar, cut (top) and on the kz =0,π/2,π planes (bottom from left andthechangein ∂(cid:15)k isataminimumimplyingacloser toright)forasystemwithU =2.7andh=1/8. Alargepart proximity to a perf∂ekczt common paring vector. oftheFermisurfacesurvives,exceptforareasaroundthehot These understandings allow quantitative explanations spots where it is most energetically favorable for pairing. of all the features of the data on α shown in Fig. 5. 100 For example, Fig. 12 shows that t(cid:104)he(cid:105)distance along z- directionbetweenthedopedandhalf-filledFermisurfaces A direct comparison between the pairing model and is at a minimum around the hot spots. Given that such the exact SCF solution can also be made in momentum distance equals ∆q/2 in the pairing construction, one space. We will identify the Fermi surface in the numeri- finds that the local ∆q value is smaller at the hot spots calsolutionfrommean-fieldtheoryasthelocusofpoints thanwhencomputedastheaveragedistanceovertheen- wheren =0.5. Dependingonthesystemandthevalue kσ tiresurface. Hence,whenonlythehotspotsareinvolved ofU,themomentumdistributionoftheexactmean-field in pairing, a larger α results, as seen at lower values ground state can maintain a true Fermi surface, char- ofU justaboveU . O(cid:104)b10v0i(cid:105)ously,thepairingreconstruction acterized by a discontinuity in n , or have it smeared c k becomesincreasingly accurateas h approaches 0. In this out by large pairing amplitudes (i.e., u close to 1/√2 k limit, one therefore finds the increasingly smaller U and neartheboundaryof ). Thetwoscenarioscanoccurin c R the faster convergence (in U) to the saturated value of the same system at different k values. The identification α 2/3 shown in Fig. 5. using nkσ =0.5 is consistent with both. (cid:104)100(cid:105) ∼ Figure 10 shows, in particular, that the portion of the Fermi surface where pairing takes place is in very good 2. Application to the (cid:104)111(cid:105)-SDW agreement with the construction based on the pairing model,whichindicatesthattheansatzcapturesthedom- inant ingredient of the physics of the SDW state. The Bytakingadisplacement∆qalongthe 111 -direction, (cid:104) (cid:105) figure also provides a direct explanation for the survival we can apply the pairing construction straightforwardly of the Fermi surface around k = π as it is there that to the diagonal-modulated SDW. Analogously to the z thepairingconstructionshowslargediscrepancywiththe 100 case, Fig. 13 shows a remarkable agreement be- (cid:104) (cid:105) true Fermi surface. This, in turn, implies that pairing in tween the shifted half-filled surface and the calculated that region would be associated with too large a kinetic Fermi surface of the SDW. The two cuts in the figure energycosttobefavorable. Theabsenceofanygapfrom clearly show broken cubic symmetry, where the Fermi pairing at the Fermi surface in the k = π plane is con- surface in one pair of the octants is further away from z sistent with the picture discussed earlier of a quasi-2D the half-filled one so as to share the common modula- liquid within each domain wall. These effects are ampli- tionwave-vector∆q withtheotherthreepairs. The 111 fied at smaller U’s as shown in Fig. 11, where a larger 111 -SDW state off(cid:104)ers,(cid:105) in this respect, a particularly (cid:104) (cid:105) 10 √2π 0 √2π this is not energetically too costly. When doing so the −π π system benefits from both pairing and band-insulating ∆q h111i mechanisms to gap the entire Fermi surface and further q′111 0.9 lower the ground-state energy. h i 0.8 z0 Q 0 V. DIMENSIONAL CROSSOVER RESULTS k q h111i 1 0.7 A. Results from full numerical HF solutions ∆q Q 2 h111i ∆qh111i 0.6 Themean-fieldgroundstateofthedoped2DHubbard model shares many similarities with its 3D counterpart. π π −√2π 0 √2−π JustaboveUc,the2DsystemdevelopsasinusoidalSDW − k 0.5 with a modulating wave along the 10 -direction and a [110] (cid:104) (cid:105) √2π 0 √2π much weaker accompanying CDW. As U is increased, −π π the SDW increases its amplitude before the SDW state 0.4 eventuallychangesintoacollectionofweaklyinteracting domainwalls. AboveacertainU,thereisadiscontinuous transition to a phase where the modulation is along the 0.3 11 -direction. ThecrossoverfromSDWtodomainwalls (cid:104) (cid:105) occurs before the 10 to 11 transition at small h, but kz0 0 0.2 after at larger h20(cid:104). A(cid:105)pec(cid:104)ulia(cid:105)rity of the 2D case, due to the special topology of the 2D half-filled surface, is that α=1andthesystemisaninsulatorregardlessofdoping, 0.1 U or direction of the modulation wave-vector apart from a region close to U . c π π Bycontrollingthedistancebetweensquarelatticelay- −√2π 0 √2−π ers,opticallatticeexperimentsallowthestudyoftheevo- − k[1¯10] lutionofthesystemasitcrossesoverfrom2Dto3D.This situationistheoreticallydescribedbyanincreaseoft in FIG.13: (Coloronline)Schematicillustrationofthepairing Hamiltonian(1)andthequestion,withinmean-field⊥the- modelforthe(cid:104)111(cid:105)-SDWstate,shownonthecontourplotsof ory, concerns the ensuing evolution of the ground-state the (1¯10) (top) and the (110) (bottom) cuts of n for a sys- k↑ properties. The pairing model and the arguments de- tem of h=1/8 and U =5.0. The white dashed lines are the scribedinSec.IV.Cremainvalidinthecrossoverregime. half-filled surface, across which is the nesting vector Q. The WethusrestrictourinvestigationtounidirectionalSDW magenta solid lines show the pairing construction, obtained by shifting the half-filled surface along [¯1¯11]-direction by a ground states, although we did use the first approach distance of ∆q /2. In the upper panel, q and q(cid:48) to carry out some searches, finding no additional struc- (cid:104)111(cid:105) (cid:104)111(cid:105) (cid:104)111(cid:105) give two equivalent representations (differing by a reciprocal tures. As in 3D, we verify that the SDW solution with lattice vector) of the pairing vector across the reconstructed minimum energy, identified using the second approach, Fermisurface. Notetheasymmetrybetweenthetwodiagonal can be obtained by the first approach in a large super- directions in the upper panel. cell that is commensurate with the optimal wavelengths, evenwhenstartingfromrandominitialguesses. SDWin directions different from 100 or 111 are not found to (cid:104) (cid:105) (cid:104) (cid:105) clearexamplewhereFermisurfacereconstructioncanbe be the global ground state for any value of t . observed. It also shows how an accurate experimental Results are summarized in the t -U mean⊥-field phase characterization of the momentum distribution in opti- diagram of Fig. 14. An overall incr⊥ease in the critical U callatticescanbeusedtocharacterizethebandstructure valuesisseenasdopingincreases, asaresultofagreater andpairingattheFermisurfacewhich, inturn, provides deformation from the perfectly nested half-filled Fermi momentum space evidence onthe real space character of surface and the need for more excitations to achieve re- the SDW. construction of the Fermi surface for pairing. As before, Using Eq. (17) we have estimated the wavelength of numerical calculations focus on small doping (h (cid:54) 0.2) the 111 -SDW,andfoundα =0.93. TheexactSCF and low to intermediate interactions (U (cid:54) 5.5), where 111 calc(cid:104)ulati(cid:105)ons in Sec. IVB sho(cid:104)wed(cid:105), instead, that α is mean-field theory can be expected to be more accurate. 111 precisely pinned at 1 in a fairly large regime of U(cid:104). T(cid:105)his Upon increase of U, and similarly to 3D, the system un- somewhat large discrepancy is a consequence of the nat- dergoes a first transition to a 100 -SDW state followed (cid:104) (cid:105) uraltendencyofthesystemto“lock”theintegratedden- by a second, discontinuous transition to a 111 -SDW (cid:104) (cid:105) sityofholespernodalregionto1wheneverthetopology state. The absence of cubic symmetry away from t =1 of the non-interacting doped Fermi surface is such that causes the modulation wave-vector for 100 -SDW⊥to lie (cid:104) (cid:105)