Ab initio study of the magnetic interactions in the spin-ladder compound SrCu 2 O 3

A wide range of experimental, semiempirical, and theoretical values have been reported in the literature for the magnetic coupling parameters of the two-leg ladder compound SrCu 2O3. We apply quantum chemical and density functional techniques to calculate accurate Nlectron wave functions or densities for two different Cu2O7 clusters that represent the leg ( Ji) and rung (J') of the ladder. Our data indicate that Ji is slightly larger thanJ' (J' /Ji'0.9) with J'52139 meV~21670 K! andJi52156 meV~21870 K!. Recent experimental data indicate a more strongly anisotropic ratio, J' /Ji'0.5. The origin of the difference is unclear, as our ab initio estimates of J' andJi seem to be converged with respect to the size of the basis set, the level of electron correlation, and the size of the cluster. However, we also find a surprisingly strong ferromagnetic interladder interaction which may play a role in resolving this discrepancy. @S0163-1829 ~99!01126-1#


I. INTRODUCTION
Over the last ten years, condensed-matter chemists and physicists have been exploring the surprising richness of cupric oxides or cuprates.A wide variety of compounds have been synthesized and investigated.One of the most important classes of cupric oxides is formed by the high-T c superconductors and their undoped parent compounds.These compounds are characterized by ͑virtually͒ isolated CuO 2 planes formed by corner-sharing CuO 4 squares in which all copper ions are connected to four other copper ions by a linear Cu-O-Cu bond.These bonds give rise to strong antiferromagnetic interactions and, hence, these compounds can be classified as two-dimensional antiferromagnets.Moreover, the magnetic spin moments of the undoped materials show long-range order at low temperatures.Another interesting class of cuprates is formed by the so-called antiferromagnetic copper oxide chains such as Sr 2 CuO 3 , which have the characteristic feature of corner-sharing CuO 4 squares in one direction only.In contrast to the two-dimensional antiferromagnet, the ground state of the one-dimensional spin-1/2 system can be solved exactly within the Heisenberg model Hamiltonian. 1 It is well-known that this exact ground state does not show true long-range antiferromagnetic order.
The discovery in the early 1990s of the Sr nϪ1 Cu n O 2nϪ1 ͑with nу2) series by Hiroi et al. 2 opened a different direction in the field.The compounds in this series are built from n one-dimensional chains in which the copper ions in adjacent chains are coupled by oxygen centers to form so-called n-leg spin ladders.The magnetic interaction between the ladders is rather small since ladders are connected to each other by a Cu-O-Cu bond of ϳ90°͑see Fig. 1͒.In principle, the spin ladders interpolate between the one-dimensional and the two-dimensional case: two interacting chains for nϭ2, while for very large n values a two-dimensional CuO 2 plane appears.However, intensive theoretical and experimental investigations showed that the progression from one to two dimensions is far from smooth.Even-leg ladders show a spin gap, [3][4][5][6][7][8][9][10][11] neither observed in the copper oxide chains nor in two-dimensional antiferromagnets.5][16] The spin ladders also attracted much attention because of the possible appearance of superconductivity upon doping the ladders with holes. 4,5,17,18Uehara et al. have actually been able to demonstrate superconductivity in a spin-ladder compound recently. 19,20Comprehensive reviews of all interesting phenomena of the spin ladders, not only those of the Sr nϪ1 Cu n O 2nϪ1 series but also of other spin-ladder compounds such as LaCuO 2.5 and Sr 14 Cu 24 O 41 , are given by Dagotto and Rice, 10 Rice, 11 and Maekawa. 18n important parameter in theoretical models of the electronic structure of cupric oxides, either two-dimensional antiferromagnets, spin chains, or ladder systems, is the effective magnetic coupling constant J, which parameterizes the strength of the magnetic interaction between the spin moments on the Cu 2ϩ ions.The magnitude of J is well established for the two-dimensional antiferromagnets, e.g., the magnetic interactions in the parent compound La 2 CuO 4 are generally considered to be characterized accurately by a J of Ϫ130 meV.However, for the lower-dimensional magnetic systems the situation is less clear.The best experimental realizations of one-dimensional spin-1/2 chains are the compounds Sr 2 CuO 3 and Ca 2 CuO 3 .For these compounds J values for the intrachain coupling have been reported ranging from Ϫ140 meV to Ϫ260 meV. 21n extra complication arises in the spin-ladder systems.In principle, there are now two different J's, one for the interaction along the legs and one along the rungs of the ladder.Although both the legs and the rungs are built from similar linear Cu-O-Cu bonds, the possibility cannot be dismissed that J ʈ ͑along the legs͒ differs from J Ќ ͑along the rungs͒.Given the similarity in Cu-O bond distances along legs and rungs, one would certainly expect a nearly isotropic situation (J Ќ /J ʈ ϭ1).3][24][25][26][27] If the magnetic interactions in the spin-ladder compounds are indeed best described by a parameter set with J ʈ ӷJ Ќ , the intuitive picture of the ground-state wave function as rung singlets with weak antiferromagnetic interaction along the legs 11 may need to be reconsidered.
The first experimental estimates of J were derived from measurements of the spin gap ⌬. 7,9 Earlier theoretical studies had established the relation between the spin gap and the magnetic coupling parameters.For isotropic coupling, it was demonstrated that ⌬ϷϪ0.5J. 3,4,6,15The nuclear-spin-lattice relaxation rate measurements of Ishida et al. 7 and Azuma et al. 9 lead to a value of the spin gap of 680 K ͑56 meV͒.Assuming isotropic coupling, the authors arrived at an estimate of J of Ϫ113 meV.In addition, Azuma et al. also performed measurements of the magnetic susceptibility as a function of the temperature.This results in a ⌬ of 420 K ͑35 meV͒ and, hence, a J of Ϫ70 meV.Azzouz et al. 24 were able to fit the data of Azuma et al. with a unique set of parameters.The assumption of isotropic coupling was given up and the following J's were reported: J ʈ ϭϪ71 meV and J Ќ ϭ Ϫ51 meV, which corresponds to a ratio J Ќ /J ʈ ϭ0.72.Sandvik et al. 23 calculated the magnetic susceptibility, spin-echo decay rate, and spin-lattice relaxation rate in good agreement with experimental data and extracted from these functions values for J ʈ and J Ќ .The best agreement was obtained for a ratio of 0.8 with J ʈ ϭϪ91 meV and J Ќ ϭϪ73 meV.It must be noted that these parameters give rise to an anomalous g factor of Ϸ1.5.In the fit of the magnetic susceptibility pre-sented by Johnston, 22 no such anomalous g factor appears, however, his results show rather strong anisotropy ͑a ratio of 0.5͒ and also a somewhat large value for J ʈ ͑158 meV͒ compared to previous estimates.Nevertheless, this interpretation of the magnetic interactions in SrCu 2 O 3 has recently been confirmed by two different groups.Eccleston et al. 26 report neutron-scattering experiments on the spin-ladder compound Sr 14 Cu 24 O 41 which is composed of layers of spin ladders equivalent to the ones found in SrCu 2 O 3 intermediated by layers composed of CuO 2 chains.These measurements give a ratio of the two J values of 0.55 and absolute values are estimated as Ϫ130 and Ϫ72 meV for J ʈ and J Ќ , respectively.Imai et al. 27 have published 63 Cu and 17 O NMR studies of the same compound and they reach the conclusion that J Ќ /J ʈ ϭ0.5, with J Ќ ϭϪ79Ϯ25 meV.Table I collects the experimental, semiempirical, and theoretical values of J ʈ and J Ќ mentioned in this paragraph.
So there now appears to be a consensus forming for a strong anisotropic ratio.A possible origin for this effect might be the different Cu-Cu distances in SrCu 2 O 3 ͑3.934Å for the leg and 3.858 Å for the rung͒.Note, however, that the shorter distance, and hence presumably larger magnetic interaction, is along the rung.This contradicts the experimental determination J Ќ /J ʈ Ϸ0.5.Another possibility might be differences in the Madelung potential for the oxygen centers mediating the superexchange interaction.In SrCu 2 O 3 , there is a difference of nearly 1 eV, the Madelung potential is 22.44 eV for the leg and 21.35 eV for the rung, assuming formal charges of ϩ2, ϩ2, and Ϫ2 for Sr, Cu, and O, respectively.Figure 1 illustrates the fact that the oxygen on the leg has five near neighbors ͑three Cu 2ϩ and two Sr 2ϩ ions͒, while that on the rung has six ͑two Cu 2ϩ and four Sr 2ϩ ).This local coordination difference can affect not only the Madelung potential at the oxygen site, but also the polarization of the Cu-O-Cu linkages.However, the magnitudes of the oxygen Madelung potentials implies that the O-2p ˜Cu-3d charge transfer energy is smaller along the rung.This suggests a larger J along the rung and J Ќ /J ʈ Ͼ1, again in contradiction with experiment.
There are thus two unexpected results here.The first is the large anisotropy, and the second is that the ratio is opposite to what we would expect in terms of bond lengths and Madelung potentials.In this paper we present the results of an ab initio quantum chemical study of the magnetic coupling parameters in SrCu 2 O 3 .These parameters are obtained by mapping spin eigenfunctions of the exact ͑nonrelativistic͒ Hamiltonian of a given material model onto the Heisenberg Hamiltonian.This approach does not involve the fitting of any data to any kind of analytical function and only depends on the quality of the approximated eigenfunction obtained from the calculation and the appropriateness of the material model applied.
Typically, the materials of interest are represented by a cluster of 5-20 all-electron atoms embedded in an electrostatic background that accounts for the rest of the crystal.Although this seems a rather poor description for a periodic structure, several reasons can be given to justify the embedded cluster approach to investigate the magnetic interactions in ionic solids.First, we mention that empirical estimates of the magnetic coupling constants are usually extracted from experimental data with the help of the Heisenberg Hamiltonian.This operator normally contains two-body operators only, which implies that a cluster with two magnetic centers is sufficient.Furthermore, it has been demonstrated for several ionic compounds ͑e.g., KNiF 3 , K 2 NiF 4 , NiO, La 2 CuO 4 ) that the magnetic coupling parameter J does not hide any possible interactions between more than two magnetic moments. 28In addition, it has been found that the magnetic interactions are completely additive in these compounds, [28][29][30] and hence, magnetic coupling parameters can be deduced from an appropriately chosen cluster model embedded in an accurate background.The question remains of the quality of the calculated wave function.It is well-known that it is essential to account for the very large electron correlation effects in the cuprates in order to obtain meaningful estimates of the magnetic coupling parameter.Modern quantum chemical techniques combined with extensive computational resources permit accurate approximations to the exact eigenfunction or to the differences in energy eigenvalues of the exact eigenfunctions.Starting from a mean-field approximation, i.e., the Hartree-Fock wave function, different computational schemes can be applied to calculate wave functions that have included the major part of the electron correlation effects.][33][34][35][36][37][38] In the remaining part of this paper, we present the application of ab initio quantum chemical techniques to the problem of the magnitude of the magnetic coupling parameters in the ladder compound SrCu 2 O 3 .In the next section, we give a short explanation of how one obtains estimates of the magnetic coupling parameters from ab initio cluster calculations, and moreover, a description of the computational methods and material model used.After that, we carefully analyze our results, in which attention is focused on the validity of the applied material model and the quality of the approximated eigenfunctions of the cluster Hamiltonian.The last section contains a summary and a further discussion of the results.

II. MATERIAL MODEL
The spin-ladder SrCu 2 O 3 is modeled by two different clusters, as illustrated in Fig. 1.One Cu 2 O 7 cluster, which has D 2h local symmetry, is used to extract a J value for the rung and another Cu 2 O 7 cluster, which has C 2v local symmetry, models the interaction along the leg.Some test calcu-lations are performed for a Cu 4 O 10 cluster from which J values for leg and rung can be extracted at the same time, and finally, we use a Cu 2 O 6 cluster to calculate the interladder magnetic interaction.All these clusters are embedded in an electrostatic background represented by optimized point charges that reproduce the Madelung potential in the whole cluster region with an accuracy better than 1 meV.To avoid an artificial polarization of the electrons of the cluster towards the point charges, we also include the short-range electrostatic repulsion between the cluster atoms and their near neighbors in the Cu 2 O 7 clusters.Because of the quantum-mechanical nature of this short-range interaction, the most desirable approach would be to extend the size of the cluster treated ab initio.Unfortunately, this is computationally too demanding and therefore, we apply a more approximate representation of the interaction by representing the ions in the direct environment of the cluster with total ion potentials ͑TIP's͒. 39We use the crystal structure as determined experimentally by Hiroi et al. 2 The Cu 2ϩ cations in the SrCu 2 O 3 crystal are characterized electronically by their local electronic ground state 2 D, arising from the d 9 open-shell configuration.Two interacting copper cations can be coupled to form either a singlet or a triplet function.Under the assumption of a common orbital part for these two functions, the eigenvalues of the Heisenberg Hamiltonian (Ϫ1/4J for the triplet and 3/4J for the singlet function͒ are directly related to the energy expectation values of the full electronic Hamiltonian.This allows us to extract estimates of the magnetic coupling parameters from our spin-restricted calculations by the relation E s ϪE t ϭJ.From this relation it is clear that a negative J arises when the antiferromagnetic ordering is the preferred one.For the cluster model with four copper cations, the values of J are derived in a similar, although slightly more complicated way. 40On the other hand, for the spin-unrestricted approaches ͓unrestricted Hartree-Fock, density-functional theory ͑UHF, DFT͔͒ no such relationship exists.Noodleman and Davidson 41 have derived the relationship between the eigenvalues of the spin-unrestricted cluster wave functions and the true singlet-triplet energy eigenvalues.Under the assumption of zero overlap between the open-shell orbitals of the broken symmetry solution representing the antiferromagnetic ͑AF͒ state, this relation reads: E AF ϪE F ϭ1/2J, as proposed by Caballol et al. 42 Again a negative J indicates that antiferromagnetic ordering is the most stable one.We refer to previous work 38,43 for a more comprehensive description of how to obtain J values from ab initio calculations.

III. COMPUTATIONAL METHODS
The original ideas of superexchange formulated by Anderson and Nesbet 44,45 can be followed by performing a CASCI ͑complete active space CI͒ in which the active orbitals are the open-shell orbitals on the Cu 2ϩ ions.This choice of the N-particle basis normally reproduces the correct sign of J, although the value computed is usually only 20-30 % of the experimental.Several approaches exist-and have been applied successfully over the last few years-to improve the CASCI approximation.We use methods based on the understanding that the large majority of the determinants outside the CAS space contribute equally to the energy ex-pectation value of the singlet and the triplet wave functions. 46,47By skipping these determinants, a relatively short list of determinants can be constructed that directly contribute to the energy difference of the two spin states.Under the assumption of a reference space with equal expectation values for triplet and singlet, the list of contributing determinants only contains singly and doubly excited determinants with the restriction of at most two holes in the inactive orbitals, one hole in the inactive plus one particle in the secondary, or two particles in the secondary orbitals.The effect of the determinants in this list can be evaluated either with second-order perturbation theory-here referred to as MP2-2-or by a complete diagonalization of the interaction matrix, i.e., by configuration interaction, here referred to as DDCI2. 47,48This approximation already gives rather good results, but it has been found recently [49][50][51] that adding some well-defined determinants to the CI space yields a significant improvement of the results.The extra determinants treated in this approach, labeled as DDCI3, 47,48 are characterized by at most one hole in the inactive orbitals plus two particles in the secondary, or two holes in the inactive and one particle in the secondary orbitals.These determinants are precisely the ones that cause a relaxation of the configurations connected with a charge transfer excitation from the bridging ligand to a copper cation, a contribution which has been found extremely important by van Oosten et al. 32,33 For a discussion of the way in which these correlations modify the charge-transfer energy, see Ref. 31.
The other computational schemes applied are less well grounded but are important in the context of solid-state physics.These methods are all so-called unrestricted, or spinpolarized formalisms, i.e., the wave functions ͑or densities͒ are no longer necessarily eigenfunctions of S ˆ2.Within the UHF ͑unrestricted Hartree-Fock͒ calculations, the unrestricted equivalent of CASCI, wave functions are constructed for the ferromagnetic and antiferromagnetic states in the cluster model.The wave function for the ferromagnetic state approaches the maximum m s component of the spinrestricted triplet wave function.However, the wave function of the AF state corresponds to a broken symmetry solution, a mixture of the true m s ϭ0 singlet and triplet wave functions. 41,42,52Again, the estimate of J computed with this method has the right sign, but usually only a small fraction of the experimental value is obtained.Density-functional theory ͑DFT͒ seems to be a very promising method to tackle this shortcoming.The method offers a possibility to improve the J value computed in the Anderson model at a lower computational cost than the traditional restricted quantum chemical methods.The simplest and most widely used approxima-tion for the unknown part of the density functional, usually referred to as the exchange-correlation potential, is the localdensity approximation ͑LDA͒, in which the functional is derived for a noninteracting electron gas.However, it is well known that LDA is not able to reproduce the insulating character of many transition-metal compounds, moreover it has been shown lately that the magnetic interaction parameter is poorly estimated by this method. 38Improvement on the correlation potential does not have a large effect.A gradient corrected exchange potential does improve the situation, although the computed values are still not comparable to experimental values.Illas and Martin 37,38 showed that with hybrid methods accurate estimates of J can be calculated.4][55] This functional combines in equal parts the exact Fock exchange functional with Becke's 1988 gradient corrected exchange functional.The Lee-Yang-Parr gradient-corrected functional 56 is used for the correlation part.In addition, we examine the popular B3:LYP hybrid functional. 57,58he calculations reported in this article have been performed with the PSHF-CIPSI-CASDI chain of programs, 59 with GAUSSIAN 94 ͑Ref.60͒ and with MOLCAS 4.0. 61

IV. RESULTS
Within the material models for rung and leg as described above, N-electron wave functions for the singlet and triplet states or for the ferromagnetic and antiferromagnetic states are expanded in an identical set of one-particle functions.Table II presents a set of calculations in which the oneparticle space is systematically improved.We have only investigated the UHF and B-F:LYP approximations in this way because the spin-restricted methods, especially DDCI3, are computationally much more demanding.Basis set A in the Table uses the 6-3111ϩG basis [62][63][64] for Cu (19s,13p,7d)/͓7s,6p,5d͔ and the 6-31G* basis (10s,4p,1d)/͓3s,2p,1d͔ for all oxygen atoms in the cluster. 65Table II shows that both spin-unrestricted methods do reproduce the experimental observation that J Ќ /J ʈ Ͻ1, although the ratio is only about 0.9 ͑It should be noted that preliminary calculations utilizing effective core potentials for both Cu and O atoms give an inverted ratio, that is J Ќ /J ʈ Ͼ1).The difference in the absolute magnitudes of the coupling parameters in the UHF and hybrid DFT ͑B-F:LYP͒ is expected.The UHF approximation is well-known to underestimate J because it places the O-2p˜Cu-3d chargetransfer energy too high.In previous research, we found that the B-F:LYP approximation yields ϳ95% of the experimen- tal coupling constant.The B3:LYP functional yields values that are clearly outside the experimental range, more than a factor of 2 larger than the corresponding B-F:LYP calculation.This has been observed before for other compounds as well, 38 and therefore, we do not consider this functional in any further calculations.
Next we introduce an f-type Gaussian function centered on Cu in the one-electron basis set ͑basis B͒ to permit polarization of the charge distribution of the Cu ions.This augmentation of the basis set does not give rise to significant changes in the calculated magnetic interaction parameters.The ratio remains constant, while the absolute values are lowered by not more than 0.6 meV.A larger change in the calculated values can be observed when the TIP description of the Sr 2ϩ ions around the Cu 2 O 7 clusters is replaced by an explicit description with basis functions of the valence electrons, albeit by a minimal, LAN2MB basis set 66 ͑basis C͒.Both J Ќ and J ʈ decreases by approximately 10% and the ratio of both parameters is slightly smaller now.These changes are mainly explained by a size effect: The TIP used to represent Sr 2ϩ was derived for neutral Sr and because neutral Sr is significantly larger than Sr 2ϩ , the short-range repulsion between cluster and environment, which is primarily due to Pauli repulsion, is overestimated by the TIP.With the explicit representation of the Sr 2ϩ valence electrons, we introduce extra ions into the cluster model with a smaller radial extent and hence, a more realistic description of the short-range repulsion.
As stated in the Introduction, it has been shown in several studies that for a variety of ionic solids the magnetic interactions are genuinely local and entangle two-body interactions only.For SrCu 2 O 3 we can test this finding by defining the cluster such that it includes both legs and rungs.For this purpose, a Cu 4 O 10 cluster is constructed that contains two subsequent rungs along the legs of the ladder.The unpaired electrons on the four copper ions give rise to one quintet spin function, three different triplet spin functions and two singlet spin functions.The energy eigenvalues are determined in a CASCI, i.e., the Anderson model, and we apply the following all-electron one-particle basis: [67][68][69] Cu (21s,15p,10d)/͓5s,4p,3d͔; bridging O (14s,9p,4d)/ ͓4s,3p,1d͔; and all other O (10s,6p)/͓3s,2p͔.From the energy differences of all the spin eigenfunctions J ʈ and J Ќ are calculated as Ϫ21.7 and Ϫ23.9 meV, respectively.The ratio J Ќ /J ʈ is 0.92.The J values obtained in the CASCI calculation from separate clusters for rung and leg, given in Table III, deviate less than 1% from the values obtained for the large cluster, in agreement with the earlier findings for other compounds that J does not have included collective interactions.
We now report the results of spin-restricted calculations in which all the above-mentioned considerations are taken into account in order to obtain reliable ab initio estimates for J Ќ and J ʈ .The one-particle space is equal to the one used for the Cu 4 O 10 cluster ͑vide infra͒, the short-range repulsion is included in the description of the electronic structure by replacing the infinitely small point charges nearest to the bridging oxygen by frozen charge distributions.These charge distributions are obtained in Hartree-Fock calculations on a ͓Sr 4 ͔ 8ϩ ͑for the rung͒ or a ͓Sr 2 Cu͔ 6ϩ ͑for the leg͒ fragment embedded in the point charges used in all calculations and the ions of the Cu 2 O 7 cluster are replaced by their formal ionic charges.Subsequently, these charge distributions are combined with their respective cluster charge distribution by a Gramm-Schmidt orthogonalization.In the calculations of the spin states the ͓Sr 4 ͔ 8ϩ and ͓Sr 2 Cu͔ 6ϩ are kept frozen.The open-shell character of the Cu ion would mean that a partially occupied orbital, the Cu-3d(x 2 -y 2 ), has to be kept frozen, a feature which is not yet implemented in our programs.Instead calculations are done with a frozen charge distribution of a Mg 2ϩ ion.The divalent magnesium ion has an effective ionic radius that is virtually equal to that of Cu 2ϩ :0.86 Å vs 0.87 Å for a sixfold coordination and 0.71 Å for both in case of a fourfold coordination. 70The wave function of the frozen fragments is expressed in a minimal basis set for the ions, i.e., (13s,10p,4d)/͓4s,3p,1d͔ for Sr 2ϩ ͑Ref.71͒ and (13s,8p)/͓2s,1p͔ for Mg 2ϩ . 69able III gives the results of the CASCI, MP2-2, DDCI2, and DDCI3 calculations for leg and rung.These numbers clearly show that the ratio between J Ќ and J ʈ does not show a very strong dependence of the computational scheme applied; all of them indicate a slightly smaller magnetic interaction for the rung than for the leg.However, as known from previous studies, the absolute values of the calculated J's differ strongly between one method and the other.We observe a relative small antiferromagnetic interaction in the CASCI calculations, which is greatly enhanced by the inclusion of external electron correlation ͑MP2-2, DDCI2, or DDCI3͒.The fact that MP2-2 and DDCI2 cause an enlargement of J by a factor of 3 has been observed before in many other compounds, e.g., La 2 CuO 4 and NiO, but the doubling of J by adding the determinants connected to a relaxation of the charge-transfer excitations to the wave function ͑DDCI3͒ is significantly larger than has been observed for La 2 CuO 4 , for which an increase of about 40% has been observed. 49his indicates that covalent interactions are relatively more important in SrCu 2 O 3 , which is also expressed in the values of J obtained with DDCI3, which are about 25% larger than in La 2 CuO 4 .Our final and most reliable ab initio estimates for J Ќ and J ʈ are Ϫ139.3 and Ϫ155.8 meV, and hence, the ratio of both parameters equals 0.9.
Finally, we consider a different type of magnetic interaction in SrCu 2 O 3 arising from the interaction of the spin moments on Cu ions located on the legs of different ladders.This interaction is assumed to be very weak because it involves an interaction in which the two O-2p orbitals which participate are orthogonal to each other.The explicit calculation of the magnitude of this interaction can be done in a Cu 2 O 6 cluster and indeed yields a very small and ferromag- netic interaction of 0.95 meV at the CASCI level; i.e., the computational details are those of the Cu 4 O 10 cluster and hence, external electron correlation is not accounted for in the present estimate.However, the inclusion of these external electron correlation effects through DDCI3 increases the calculated magnitude of J inter up to 12.5 meV ͑see Table III͒.This is an unexpected result, the interladder interaction is now ϳ10% of the magnetic interactions along the legs and rungs.This interaction may need to be considered in the fitting of experimental data, and the assumption of isolated two-leg ladders in SrCu 2 O 3 reconsidered.

V. SUMMARY AND CONCLUSIONS
We have calculated ab initio estimates of J Ќ , J ʈ , and J inter of the ladder compound SrCu 2 O 3 .The material has been represented by small clusters that contain the essential ions involved in the superexchange processes.The clusters are embedded in a background of point charges that account for the long-range electrostatic interactions ͑the Madelung po-tential͒ and in addition we have included the short-range Pauli repulsion to prevent an artificial polarization of the cluster charge distribution towards the point charges.Within this material model we obtain accurate approximations of the quantities determining the magnetic coupling constants with quantum chemical and density functional techniques.The simple CASCI and UHF approximations result in antiferromagnetic but relatively small magnetic coupling parameters for leg and rung.The values are greatly enlarged by incorporating external electron correlation effects in the N-electron wave function.While the absolute magnitudes of the coupling parameters change dramatically with electron correlation, the ratio is fairly stable of J Ќ /J ʈ Ϸ0.9.
Beside the dependency of J Ќ and J ʈ on the choice of the one-electron space and the details of the representation of the short-range repulsion, we have also investigated how the magnetic interactions along the rung and the leg interfere with one another.By comparing the J values extracted from a Cu 4 O 10 cluster which contains both superexchange paths along the leg and along the rung with those obtained from the two-center clusters, we can conclude that the two superexchange processes are completely independent.The parameters differ by less than 1% in the two-center and the fourcenter clusters.In addition we have calculated the interchain coupling by constructing a cluster with two Cu ions from different legs.As expected, this interaction turns out to be ferromagnetic, but its magnitude is significantly larger than generally assumed.
Our final and most reliable ab initio estimates of J Ќ and J ʈ are Ϫ139.3 and Ϫ155.8 meV respectively, as listed in Table III.These values lead to a ratio of 0.9.Although the values of J Ќ and J ʈ are very reasonable in comparison with the experimental and semiempirical estimates reported in the early literature, the ratio is in sharp contrast to the most recent determinations by Johnston et al., 22 Eccleston et al., 26 and Imai et al. 27 Their interpretation of the experiments suggest that J Ќ is only half of J ʈ .As put forward by Eccleston, this is a rather surprising finding since the Cu-O-Cu links for leg and rung are very similar.In the Introduction we have discussed the differences between these two bonds and none of these can be the origin of such a large difference in mag-netic coupling along the leg and the rung.In fact, the difference in bond length favors the magnetic interaction along the rung, since this bond is slightly shorter.The exact effect of the difference in local geometry for the two bonds is not easily quantized.The effect of the local geometry is threefold, first it contributes to the differences in the Madelung potential between the oxygen on the rung and the oxygen on the leg; secondly, it introduces a different Pauli repulsion between the cluster ions and the environment; and in the third place it polarizes the oxygen anions on the leg and the rung in a different manner.The second point is not essential since the ratio of 0.9 is also observed for an embedding with point charges only, i.e., the Pauli repulsion causes an approximately equal increase of J for leg and rung.Although J is susceptible to changes in V MAD , the dependency is rather weak 36 and actually, the difference affects the parameters in the opposite direction because J increases for a reduction of the Madelung potential. 36,72Therefore, the most likely origin of the ratio of 0.9 encountered in our calculations is the different polarization of the Cu-O-Cu bonds by the ions in the direct neighborhood of the Cu 2 O 7 clusters for leg and rung.Note that this polarization effect is also present in the clusters in which no Pauli repulsion ͑i.e., an embedding of point charges only͒ is included between the cluster ions and the surrounding centers, consistent with our findings that the ratio J Ќ /J ʈ does not depend on the exact details of the representation of the cluster surroundings.In addition, we have shown that the superexchange processes along rung and leg do not interfere and cannot cause J Ќ to be half of J ʈ .
Finally, we discuss our results in terms of the well-known t-J model Hamiltonian, which can be derived from the Hubbard model Hamiltonian by a perturbation expansion in t/U.In the t-J model, J is proportional to t 2 /U, where t is an effective Cu-Cu hopping integral and U is the effective onsite two-electron Coulomb repulsion integral.We see no reason to doubt that all Cu sites exhibit roughly equal U ͑within 10% or so͒.The origin of the anisotropy would then appear to lie exclusively in the influence of the oxygen centers on the effective hopping parameter t.The ratio would suggest that t along the rung should be smaller than t for the leg.However, the explicit calculation of t for leg and rung indicates that this parameter is ͑as might have been expected͒ larger for the shorter distance ͑i.e., for the rung͒.We conclude that the observed anisotropy cannot be explained within a simple t-J model.
We have also found that the assumption of a weak interladder coupling is not supported by our calculations, and hence, this magnetic interaction cannot be excluded as a possible source of large anisotropy within the ladder.We must point out that the DDCI3 calculated value for J ʈ is in excellent agreement with the experimental estimate whereas a fairly large discrepancy is found for J Ќ .However, the fact that a significant J inter coupling is found means that the measured J Ќ is in fact an effective magnetic coupling parameter for the two magnetic interactions.Notice that the appearance of J inter introduces spin frustrations in the systems that may play a role in the experimental determination of the J Ќ /J ʈ ratio.
In summary, the two-leg ladder material SrCu 2 O 3 represents a very challenging case.We have not found a simple explanation for the experimentally observed strong anisot-ropy.While the present ab initio cluster model calculations yield a qualitatively correct anisotropy, in the sense that J along the leg is larger than J for the rung, they predict a fairly small anisotropy and fail to predict the experimentally observed ratio of about 1/2.We examined a number of possible sources of this discrepancy and found a significant value for the interladder interaction that is not taken into account in the models used to fit experimental data to extract the magnetic coupling constants.This may possibly play a role in resolving the discrepancy between the ab initio calculated values and the experimental estimates.

TABLE I .
Overview of parameters reported in the literature for the strength of the magnetic coupling in SrCu 2 O 3 .

TABLE II .
Dependency of the magnetic coupling parameters on the basis set and the material model.Basis A consists of the 6-3111ϩG basis for Cu, and the 6-31G* basis for O. Basis B is as A augmented with an extra f-type function on Cu.In basis C the TIP's for Sr are replaced by the LAN2MB basis set.