arXiv:cond-mat/0701025v2 [cond-mat.mtrl-sci] 22 Feb 2007 cie h olm euso ewe the between repulsion Coulomb the scribes of hc stedniyfntoa o h xhneenergy. exchange the for functional Generally, density the is which in htaeascae ihtesm ie electron given same the with associated density are that Ψ tions Aoi nt r sdtruhu hswr. Even- antisymmetric work.) all this of throughout out used tually, are units (Atomic inlo h ate nryadteoperator the and energy Hartree the of tional htteparameter the that as defined is integrand The tegho opig“constant”. coupling or strength iiu xetto fteoperator the of expectation minimum T Here, neatepeso o hsfntoa stecoupling- the is functional this integral, for constant expression exact An uniyta utb prxmtdi rciei the is practice in approximated be functional must that solids. quantity and molecules many for accuracy cost computational reasonable low and of combination unique be- chemistry, its quantum of cause and physics matter calculations condensed structure in electronic for method popular most ˆ At napyn F oagvneeto ytm h only the system, electron given a to DFT applying In the now by is (DFT) [1–3] theory functional Density = α ic h lcrn ntesaeΨ state the in electrons the since , aoaor eCii herqe NS Universit´e Pierr Th´eorique, CNRS, Chimie de Laboratoire − α U 2 1 0, = ρ [ P ρ = = ] W W i N E E =1 opr u eut ihpeiu prxmt ouin an solutions u c approximate explicitly allows previous theory. is with This functional which results solution, general minimum. our a compare degenerate write a to and with limit, problem classical a W ρ α xc α x ( [ [ [ ∇ ρ erfruaetesrn-neato ii felectronic of limit strong-interaction the reformulate We r α V ρ [ ρ 2 1 ˆ ρ ,Ψ ), = ] nttt fTertclPyis nvriyo Regensbur of University Physics, Theoretical of Institute = ] [ .INTRODUCTION I. samntnclydcesn function decreasing monotonically a is ] ee E i 2 R eea omlto ihapiain oshrcldensi spherical to applications with formulation general A ρ o h xhnecreainenergy. exchange-correlation the for ] trsotwt h value the with out starts ] xc stekntceeg prtr Notice operator. kinetic-energy the is d = titycreae lcrn ndniyfntoa theor functional density in electrons correlated Strictly r α [ α ρ h h R [ Ψ = ] Ψ ρ ok sa dutbeinteraction adjustable an as works N X i d eoe h n htyed the yields that one the denotes ] =1 α 0 − r [ [ 1 ρ ρ ′ ρ ] ] j Z | | ( = X V V 0 r ˆ ˆ N i 1 ee ee ) +1 ρ αW dα | | ( Ψ Ψ r | r ′ 0 α N ) i [ [ / ρ ρ − α eeto aefunc- wave -electron 1 | ] ] r − i [ − i ρ r − ] al oiGog n nra Savin Andreas and Gori-Giorgi Paola j . T α | ˆ U . U r [ + ρ ′ [ | [ increasingly ] ρ ρ N α (4) ] stefunc- the is ] . V ˆ Dtd uut1,2018) 13, August (Dated: electrons, ee Here, . V ˆ ee ihe Seidl Michael de- (3) (1) (2) tMreCre lc use,F722Prs France Paris, F-75252 Jussieu, Place 4 Curie, Marie et e otx,w ar u ealdsuyo h limit the of study detailed a out carry we context, ute eeoe nteDTfaeoki es and at 5 information Refs. the in [4], which framework Wigner in DFT of 6, the idea in early developed the further follows instead, methods. work, sophisticated more This or theory perturbation e.g., where α terpolating edt vi ahohri pc sterplinstrength repulsion the as space α in other each avoid to tend a rpsd[,8], [5, with model proposed case (PC) was general point-charge-plus-continuum density, the the spherical For a electrons, in arguments. electrons physical two only of using found case was special here presented the solution in previous the in investigated 7–9], was [5, (5) work Eq. of limit the Although ain,wsue ocntuta neplto for interpolation an construct to used was lations, rxmto,tgte ihasmlroefrtenext the for one similar a term, leading with together proximation, ic h lcrn r ocdt tywti h fixed the within stay to forced are density indefinitely, electrons decrease the electrons, cannot since the it However, between distances decrease. inverse must average the for raht lcrncreainotncnit ntrying in consists ( often correlation to electron to proach xlsvl h w functionals two the exclusively in- work. can of DFT idea strong-interaction 3.4 in the general limits of and the weak- error the that between absolute showing terpolating at- 10], mean accurate [6, a predicts kcal/mole) (with ISI energies [6]. omization model (ISI) interpolation between α rw.Tu,teepcainof expectation the Thus, grows. → ceaial,tetaiinlqatmceityap- chemistry quantum traditional the Schematically, tsol eepaie htIIue singredients as uses ISI that emphasized be should It )fo h o-neatn ii ( limit non-interacting the from 1) = extrapolate W ,adifiieysrn interaction, strong infinitely and 0, luae o peia este.W then We densities. spherical for alculated ∞ PC A ρ est ucinlter ntrsof terms in theory functional density ,D900Rgnbr,Germany Regensburg, D-93040 g, ics h mlctosfrdensity for implications the discuss d α ( [ = ρ r = ] ). and 0 = ewe h w iiso ekinteraction, weak of limits two the between − ocaiymn set fthis of aspects many clarify to s W 10 9 h nomto ntepyia system physical the on information the Z ∞ ′ ( W 4 [ 3 π d ρ ∞ r ,dsrbn eopitmto oscil- motion zero-point describing ], ) α 1 [ ρ / 3 lim = ] ρ A = and ∞ ( r α alditrcinstrength interaction called , ) →∞ 4 B / 3 α = W E + sotie by obtained is 1 = V y: ˆ ties x α ee 350 [ 3 B [ ρ hc sameasure a is which , ρ and ] ( ] |∇ . ρ 4 3 α π ( α ρ r ) )b using, by 0) = ( ) 1 ∞ → r 4 / / ) 3 E 3 | hsap- This . 2 c GL2 nthis In . . [ W ρ (of ] α (6) (5) in- [ N ρ ] 2 the second-order correlation energy in G¨orling-Levy per- order of O(α). Therefore, Ψα[ρ] Tˆ Ψα[ρ] has the order turbation theory [11]) from the relatively simple non- of O(√α) and we may writeh in Eq.| | (7) i interacting limit α 0 (with single-particle orbitals), → 1 plus the two functionals W [ρ] and W ′ [ρ] from the op- W [ρ]+ U[ρ] = lim Ψα[ρ] Tˆ + Vˆee Ψα[ρ] ∞ ∞ ∞ α posite α limit of infinitely strong repulsion, which →∞ α are, so far,→ not ∞ known exactly (as said, except for W [ρ] D 1 E ∞ = lim min Ψ Tˆ + Vˆee Ψ . (10) in the special case of two electrons in a spherical density α Ψ ρ α →∞ → D E [7]). In this work we reformulate the α limit of DFT In the second step we have applied the definition of the → ∞ wave function Ψ [ρ]; the constraint “Ψ ρ” addresses in terms of a classical problem with a degenerate mini- α → mum. This allows us to construct the general solution all those wave functions Ψ that are associated with the same given density ρ. Provided that the limit α for W [ρ], and to clarify many aspects of this limit. As → ∞ α> 0∞ grows beyond its real-world value α = 1, the con- can be applied directly to the operators in Eq. (10) (there cept of single-particle orbitals becomes completely mean- is no rigorous proof for this reasonable conjecture – see ingless. When α , however, an entirely new type also the Appendix), it is simplified to of simplicity with so-called→ ∞ “co-motion” functions arises, W [ρ]+ U[ρ] = min Ψ Vˆee Ψ . (11) as we shall see in Secs. II-IV. The “co-motion” func- ∞ Ψ ρh | | i tions, which entirely determine W [ρ], can be directly → constructed from the one-electron density∞ ρ(r). In this case, the expectation of Vˆee alone is to be mini- The paper is organized as follows. In Sec. II we first mized, regardless of the kinetic-energy operator Tˆ. This give a general overview of the problem, anticipating the apparently purely classical problem corresponds to the solution on intuitive physical grounds, and leaving the quantum-mechanical limit of infinitely large masses. mathematical details in Secs. III and IV. We then use If Eq. (9) holds, that is if the density ρ(r) is both our formalism to explicitly evaluate the limit of Eq. (5) in N- and v-representable also in the very α limit, the case of spherical densities, with applications to atoms the minimizing Ψ in Eq. (11) is the ground→ state ∞ of the (Sec. V). In Sec. VI we compare our solution with the ap- pure multiplicative operator Vˆee +Vˆ . The local one-body proximation of Eq. (6), and we discuss the implications potential Vˆ is the Lagrange multiplier for the constraint for the ISI functional. The last Sec. VII is devoted to “Ψ ρ”, and can be found via the Legendre-transform conclusions and perspectives. In the Appendix we also formulation→ of Eq. (11) [12], consider the simple case of harmonic forces, in order to analyze how the nature of the electron-electron interac- ˆ ˆ tion affects the solution. W [ρ]+ U[ρ] = max min Ψ Vee + V Ψ ρ v . ∞ v Ψ h | | i− Z (12) To start to address the problem of N-representability II. SMOOTH DENSITIES FROM A CLASSICAL PROBLEM in the α limit, we write the minimizing wavefunc- tion Ψ in→ Eq. ∞ (12) as the product r For a given N-electron density ρ = ρ( ) we generally Ψ= ψ(r1, ..., rN )χ(σ1, ..., σN ), (13) wish to find the limit (5) or, equivalently, where ψ is a spatial wavefunction and χ is a chosen eigen- W [ρ]+ U[ρ] = lim Ψα[ρ] Vˆee Ψα[ρ] . (7) ˆ2 ˆ ∞ α h | | i state of the total spin S and its projection Sz. InSec. III →∞ we analyze the N-representability problem in more de- If the density ρ is both N- and v-representable for ev- tail, showing that, in the special α case, we can al- ery α, there exists an α-dependent external potential ways construct an antisymmetric Ψ→ of ∞ the form (13), and ˆ α N α r Vext[ρ]= i=1 vext([ρ], i) such that Ψα[ρ] is the ground that the choice of the spin eigenfunction χ does not af- state of the Hamiltonian fect the square of the spatial wavefunction ψ(r , ..., r ) 2 P 1 2 and thus, as shown by Eq. (15) below, the| energy. For| Hˆ α[ρ]= Tˆ + αVˆ + Vˆ α [ρ]. (8) ee ext this reason, in what follows we only consider the spatial α r wavefunction ψ(r , ..., r ). As α , the binding external potential vext([ρ], ) in 1 N the Hamiltonian→ ∞ (8) must compensate the strong repul- As said, sion O(α) between the electrons. Therefore, we expect Vˆee + Vˆ Epot([v]; r1, ..., rN ) (14) vα ([ρ], r) ≡ lim ext = v([ρ], r), (9) α α is a pure multiplicative operator, so that, after integrat- →∞ ing out the spin variables, the unconstrained minimiza- r r with v([ρ], ) a smooth function of . Thus, for large tion in the brackets of Eq. (12) reads α 1, the kinetic energy in the state Ψ [ρ] is mainly ≫ α due to zero-point oscillations of strongly repulsive elec- 2 min dr1...drN ψ(r1, ..., rN ) Epot([v]; r1, ..., rN ). (15) trons, bound by a strong attractive force that has the ψ | | Z 3
We see that the minimum in Eq. (15) is reached when The requirement “Ψ ρ” can of course be fulfilled 2 → the square of the spatial wavefunction ψ(r1, ..., rN ) is also if the degeneracy of the absolute minimum of Epot a distribution that is zero everywhere except| for values| is higher, e.g., if M is a 6D subspace described by (r1, ..., rN ) M, where (r, r′, f3(r, r′), ..., fN (r, r′)) : r, r′ P , which cor- ∈ responds{ to the physical state{ in which} ∈ the} position of
M (r1, ..., rN ): Epot([v]; r1, ..., rN ) = min (16) two electrons fixes the positions of all the others. How- ≡ ever, we have to keep in mind that we can affect M only n o by varying the one-body potential v(r). Physically, this is the set of all configurations (r , ..., r ) for which the 1 N means that v(r) must compensate the repulsive forces of 3N-dimensional function E of Eq. (14) assumes its pot the other N 1 electrons acting on the electron in r. If absolute (global) minimum. Notice that the set M is the position of− the other N 1 electrons only depends on purely determined by the choice of v(r), M = M[v]. We r [as in Eq. (18)], then it may− be possible to find such a shall see in the next Sec. III that to such a minimizing v(r). If, instead, we require the degeneracy of M to be distribution ψ(r , ..., r ) 2 we can indeed associate an 1 N higher, finding such a v(r) seems a daunting task. antisymmetric| wavefunction| of the form (13). We shall go through these physical considerations more For a given reasonable (attractive) potential v(r), we in detail in Sec. IV, where we analyze the solution (18) may expect that M will comprise only one single config- and write down an explicit equation for v(r). Before do- uration (plus its permutations) or a small finite number ing so, we first show that we can consistently construct of configurations: the set M corresponds, in fact, to the an antisymmetric wavefunction Ψ of the form (13) such solution of the classical electrostatic equilibrium prob- that ψ(r , ..., r ) 2 is a distribution that is zero every- lem for N identical point charges in the external poten- 1 2 where| except on the| minimizing set M. tial v(r). The density associated with the corresponding spatial wave function ψ will then be a finite sum of delta functions, III. ANTISYMMETRY OF THE K WAVEFUNCTION ρ(r) δ(r r ), (17) ∝ − k k=1 We want to construct an antisymmetric wavefunction X of the form (13) such that its spatial part ψ(r , ..., r ) 2 r 1 N where the k are the K N different position vec- is a distribution that is zero everywhere| except in the| tors forming the configurations≥ in M. When v(r) has set M where Epot has its global minimum. Given a some symmetry (e.g., spherical), M can also be a low- configuration (r1, ..., rN ) of M, we first notice that all dimensional continuum (e.g., the surface of a sphere). its electronic positions ri must be different from each For a quantum-mechanical system (atom, molecule, other: without the kinetic energy operator Tˆ, two or solid), the electronic density is typically smooth, so that more electrons on-top of each other would make Epot Eq. (17) seems in contradiction with the final require- infinite because of the singular nature of Coulomb repul- ment “Ψ ρ” [or the maximization (12)]. In other sion at short distances. Moreover, for any configuration words, it seems→ like there is a v-representability prob- M, also all its N! permutations are M, since Epot lem for smooth densities in the very α limit. The ∈ ∈ r → ∞ is invariant under permutation of its variables i. Then, solution to this apparent problem is that the potential given the value of the wavefunction for one configuration v(r) must be such that M[v] is a continuum with at least (r1, ..., rN ) M, with spins (σ1, ..., σN ), its value for the three dimensions. That is, the absolute minimum of the permutations∈ π of this configuration must be simply de- 3N-dimensional function Epot must be degenerate over fined as (at least) a 3-dimensional subspace of R3N , which we write as ψ(rπ(1), ..., rπ(N))χ(σπ(1), ..., σπ(N))= π ( 1) ψ(r1, ..., rN )χ(σ1, ..., σN ) (19) M = (r, f (r), ..., f (r)) : r P , (18) − 2 N ∈ n o where ( 1)π denotes the sign of the permutation π. Be- 3 − where P R is the region of space where ρ(r) =0. As cause the ri in a given configuration are all different from we shall see⊆ in Sec. IV, the ansatz (18) can be consistent6 each other, and ψ is zero everywhere except on M, the with the requirement that Epot is minimum over M only N electrons are always localized in different regions of when the functions fn(r) satisfy special properties. space that have exactly zero overlap. This means that From the physical point of view, the corresponding spa- when we consider the square of the spatial wavefunction 2 tial wavefunction ψ (which is zero everywhere except on ψ(r1, ..., rN ) all the cross terms are exactly zero, and M), describes a state in which the position of one of the |thus ψ 2 is completely| independent of the choice of the | | electrons can be freely chosen in P , r1 = r, but it then spin eigenstate χ. fixes the positions of all the other electrons through the Similarly, for repulsive bosons (with any spin), Ψ can functions fn(r), r2 = f2(r), ..., rN = fN (r). This is what be symmetrized. Both the symmetric and the antisym- we call [5, 7] “strictly correlated electrons” (SCE). metric choice give the same spatial ψ(r , ..., r ) 2, and | 1 N | 4 thus the same expectation for Vˆee: when Coulomb re- properties pulsion between the particles becomes dominant with re- spect to the kinetic energy, particle-particle overlap is f2(fn(r)), ..., fN (fn(r)) = f1(r), ..., fN (r) fn(r) , { } { }\{ } suppressed, so that the particles no longer know wether (22) f x they are fermions or bosons. i.e., the set of N 1 functions i( ) with i = 2, ..., N, −x f r The same construction (19) applies to the SCE state in when evaluated in = n( ) must yield any permuta- f r which M has the form (18), since all the considerations tion of the set of N 1 i( ) with i =1, ..., N and i = n f r − r 6 made so far are obviously still valid for any configuration (thus including 1( )= ). This means that applying one f r in M characterized by a given r P . co-motion function k to one position n of a given SCE ∈ configuration C M must always yield a position of the ∈ same C again. If the functions fi satisfy this property, IV. CO-MOTION FUNCTIONS then the fulfillment of the first of the Eqs. (21) automat- ically implies the fufillment of the other N 1 equations, and thus the stationarity of the solution (18)− for any r. r We have seen that the potential v( ) of Eq. (9) for the One has then to verify that such a stationary solution α limit of DFT must be such that the minimum → ∞ is the global minimum of the 3N-dimensional function of the 3N-dimensional function Epot of Eq. (14) is de- Epot([v], r1, ..., rN ). generate over the 3-dimensional subspace M of Eq. (18). We shall now analyze how to construct such a v(r), and f r thus which properties the functions n( ) must satisfy, B. The SCE external potential and how they can be determined from the density ρ(r). Since the solution (18) corresponds to a state in which Once we have found some functions f (r) that satisfy the position of one electron dictates the postions of all the i the properties (22), Eqs. (21) provide N equivalent equa- others, we call the functions f (r) co-motion functions. n tions for the potential v(r),
N r f (r) A. Properties of the co-motion functions r i v( )= − 3 . (23) ∇ r fi(r) i=1 X6 | − | If we assume that the v(r) we are looking for is a smooth potential with a continuous gradient v(r), then Equation (23) has the clear physical meaning anticipated ∇ in Sec. II: the potential v(r) must compensate the net Epot([v]; r1, ..., rN ) is a continuous and smooth function r force acting on the electron in r, resulting from the re- of its variables n, except for configurations with two or f r more electrons on top of each other, which, as explained pulsion of the other N 1 electrons at positions i( ). We start now to see− that the α limit of DFT is in Sec. III, cannot belong to the minimizing set M. In → ∞ entirely characterized by the co-motion functions fi(r), this case, a minimizing configuration (r1, ..., rN ) M must satisfy the stationarity conditions for the function∈ which also give, as a byproduct via Eq. (23), the external potential v(r). In the next Subsec. IV C we shall see how Epot([v]; r1, ..., rN ), i.e., to determine these functions from the density ρ(r).
N r1 ri r 3 v( 1)= i=1 r1 −ri ∇ N6 | r2− ri| r 3 v( 2)= i=2 r2 −ri (20) C. Co-motion functions for a given density ∇ P 6 | − | . . P The SCE problem for a given density ρ(r) reduces then to the construction of the appropriate co-motion func- If we insert the solution (18) into Eqs. (20) we obtain tions fn(r). To do this, we simply use the quantum me-
N r fi(r) chanical meaning of the electronic density. Since in the v(r)= − 3 i=1 r fi(r) SCE state the position of the first electron determines ∇ 6 | − N | f2(r) fi(r) v(x) x f r = − 3 the positions of all the others, the probability of finding =P2 ( ) i=2 f2(r) fi(r) (21) ∇ | 6 | − | the first electron in the volume element dr around the . . P position r must be the same of finding the nth electron in the volume element df (r) around the position f (r). f r r n n where we have defined 1( ) . Now, suppose that we This means that all the co-motion functions f (r) must f ≡r r n have found N 1 functions i( ) and a potential v( ) such satisfy the differential equation that the first of− the Eqs. (21) is satisfied for all r R3. If r f s ∈ we now evaluate this first equation for = n( ) and then ρ(fn(r))dfn(r)= ρ(r)dr, n =2, ..., N. (24) put s = r, we see that its left-hand-side coincides with the left-hand-side of the nth equation. It is then easy to In order to construct the co-motion functions we thus verify that, if we want also the right-hand-sides to be the have to find the initial conditions for the integration same, the functions fi(r) must satisfy the transformation of (24) that (i) satisfy the properties (22), (ii) yield a 5 smooth potential v(r) via Eq. (23), and (iii) give the min- of lenghts r1, ..., rN having the same point-charge q glued imum expectation of Vˆee. In Sec. V we solve this problem at one end, and the other end fixed in the origin, in such explicitly for spherically symmetric N-electron densities. a way that they are free to rotate in the 3D space. Equations (24) can also be proven by explicitly We now insert the function Ω(r1, ..., rN ) in the station- constructing the square of the spatial wavefunction arity equations for the radial variables ri, ψ(r , ..., r ) 2 that is zero everywhere except on M of | 1 N | Eq. (18), and by imposing that the expectation value of v (r )= ∂ V (r , r , ..., r , Ω) ′ 1 ∂r1 ee 1 2 N 1 N N 2 − Ω(r ,...,r ) the operatorρ ˆ(r) = δ(r r ) on ψ(r1, ..., r ) ∂ i=1 i N v′(r2)= Vee(r1, r2, ..., rN , Ω) r − | | − ∂r2 Ω(r1,...,rN ) (27) is ρ( ). In this way, we also see that the value of . ψ(r, f (r)..., f (r)) 2 inP a given SCE configuration M . 2 N . must| be set equal to| 1 ρ(r). ∈ N As explained in the previous Sec. IV, in order to have a smooth (e.g., atomic) radial density ρ(r) we need to D. The value of W∞[ρ] choose a v(r) that makes these equations fulfilled by a set of radial distances (r, f2(r), ..., fN (r)) in which r can take After all, maximising with respect to the potential v any value in the domain in which ρ(r) = 0. These radial 6 in Eq. (12) seems to be equivalent to constructing the co-motion functions fn(r) must have the same transfor- f r co-motion functions n( ). In terms of these functions, mation properties of Eq. (22), as it can be verified by the expectation of the operator Vˆee in the SCE state [see inspection of Eqs. (27). also Eqs. (2) and (5)] reads To construct the radial co-motion functions fn(r) for a given density ρ(r) we proceed as follows. Our starting N 1 N 1 − ρ(r) point is Eq. (24) that becomes, in the case of spherical W [ρ]+ U[ρ] = dr N , (25) ∞ f (r) f (r) symmetry, i=1 j=i+1 i j X X Z | − | 2 2 where, again, we have used the convention f (r) = r. 4πfn(r) ρ(fn(r)) fn′ (r) dr =4π r ρ(r) dr. (28) 1 | | Equation (25) comes from the fact that, as said in the previous Subsec. IVC, for a given r, the configuration We then define the function Ne(r) that gives the expected r f r f r 1 r ( , 2( ), ..., N ( )) in M has a weight equal to N ρ( ), number of electrons between 0 and r, and that the only possible electron-electron distances are f r f r r i( ) j ( ) , with i = j, and i, j =1, ..., N. 2 | − | 6 Ne(r)= 4π x ρ(x) dx. (29) Z0 V. THE SCE SOLUTION FOR N ELECTRONS Ne(r) is a monotonously increasing function of r, and its IN A SPHERICAL DENSITY 1 inverse Ne− (y) is well defined for y [0,N). By inte- grating both sides of Eq. (28) we see∈ that each function As an example, we consider here N electrons in a fn(r) is of the form spherical density, and we explictly calculate the functions fn(r) and the value W [ρ] for a few atoms. 1 ∞ Ne− (Ne(fn(0)) Ne(r)) 0 r a1 Given the symmetry of the problem, we choose a 1 ± ≤ ≤ r f (r)= Ne− (Ne(fn(a1)) Ne(a1) Ne(r)) a1 r a2 spherical-coordinate reference system i = (ri,θi, φi) in n . ∓ ± ≤ ≤ which the position of the first electron defines the z- . axis (θ1 = 0, φ1 = 0) and the second electron is on (30) the xz plane (φ2 = 0). The 2N 3 relative angles where the upper (lower) sign corresponds to the case − θ2,θ3, φ3, ..., θN , φN are globally denoted by Ω. The fn′ 0 (fn′ < 0), and [0,a1], [a1,a2], ..., [ak 1,ak] are { } ≥ − total classical energy we want to minimize then reads the intervals in which (i) fn′ (r) does not change sign and 1 (ii) Ne(fn(ai)) Ne(ai) Ne(r) [0,N], so that N − is N e well defined. ∓ ± ∈ E ([v], r , ..., r , Ω) = v(r )+ V (r , ..., r , Ω). pot 1 N i ee 1 N The SCE problem for spherical densities thus consists i=1 X in finding the signs, the intervals [ai 1,ai], and the ini- (26) − Since the external potential does not depend on the rela- tial conditions fn(0), ..., fn(ak) in Eqs. (30) that (i) make tive angles Ω, we can decouple the stationarity equations the fn(r) satisfy the transformation properties (22), (ii) yield, via Eq. (23), a v(r) with a continuous first deriva- (20) into an angular part, which only involves Vee, and ˆ a radial part that also involves v(r). By solving the an- tive, and (iii) give the minimum expectation value of Vee gular part for any given r1, ..., rN we can define a func- when inserted in Eq. (25). tion Ω(r1, ..., rN ) for the minimizing angles, valid for any We shall now see how to proceed in the case of atomic spherically-symmetric v. These angles are the solution of densities. We first treat the case of the He atom, and we the electrostatic equilibrium problem for N neutral sticks then generalize the solution to N > 2. 6
A. N = 2 (He atom) is the true minimum of Epot([v], r1, r2,θ2). Besides the mathematical arguments, the choice (35) Two electrons in a smooth spherical density in the is the most “physical” one: it makes the two electrons α limit have been already considered in Ref. 7, always be in two different spherical shells, each one con- where→ ∞ the SCE state was proposed on the basis of physi- taining, on average in the quantum mechanical problem, cal considerations. Here, we restate the problem in terms one electron [7]. This is exactly what we expect for two of the formalism just derived, and we show that the so- electrons that repel each other infinitely strong, but that lution proposed in Ref. 7 was indeed the correct one. have to fulfill the constraint of yielding a given smooth, When N = 2 we have only one relative angle, θ2. By atomic-like density ρ(r). solving the stationarity equations for the angular part we immediatly obtain θ2 =0 or π. It is easy to verify that B. N = 3 (Li atom) θ2 = 0 is a maximum and θ2 = π is a minimum for any value r1 and r2. By inserting the minimizing angle θ2 = π in the stationarity equations for the radial variables, we When N = 3 we have 3 angles, θ2,θ3, φ3, but we im- obtain mediately obtain φ3 = 0: the three electrons must be on 2 the same plane, containing the nucleus, to achieve com- v′(r1) = (r1 + r2)− 2 (31) pensation of the forces (see Sec. IV B). We then find nu- v′(r2) = (r1 + r2)− merically, for any given r1, r2, r3, the minimizing angles These equations admit a solution of the kind (r, f(r)) if θ2 and θ3. and only if f(r) satisfies the property f(f(r)) = r, in To construct the radial co-motion functions f2(r) and agreement with Eq. (22). This means that, among all f3(r) we should, in principle, try out all the possible fn(r) the possible f(r) of the form (30), we can only choose of the form (30) that satisfy the properties (22), and se- 1 the ones for which f − = f. One can verify that the lect the ones that yield a continous v′(r) and the lowest only possible choices are then expectation for Vˆee. Instead, we use the physical idea that
1 was behind the solution for two electrons of Eq. (35): we f(r)= N − (N (r)) = r (32) e e expect that the correct f2 and f3 are the ones that make 1 Ne− (Ne(f(0)) Ne(r)) r
6 2 ∞ Li atom 1.5 a) 5 1 f (r) 4 3 0.5
3 0
-0.5 2 -1 f2(r) 1 -1.5
0 -2 0 1 2 3 4 5 6 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 r 2
FIG. 1: The radial co-motion functions f2(r) and f3(r) for 1.5 b) the Li atom density. Hartree atomic units are used. 1
0.5 electron 3. We clearly see from this figure, that in any 0 SCE configuration the space is divided in three spheri- cal shells, which never contain more than one electron. -0.5 These three spherical shells are the same that, in the -1 quantum mechanical problem, contain on average one electron: the SCE state makes them contain always ex- -1.5 actly ∞ one electron, thus suppressing any accidental clus- -2 tering (“correlation suppresses fluctuations” [14, 15]). -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 Because of the properties of the co-motion functions 2 (which are a consequence of the fact that the electrons are ∞ indistinguishable), the three panels are equivalent. Our 1.5 c) physical problem is in fact invariant for a rigid rotation 1 of any of the configurations corresponding to a given r. This means that in panel b) we could, for each value 0.5 of r, rigidly rotate the position of the three electrons in 0 such a way that the electron in the inner shell moves on a straight line. In this way, we would reobtain a rigid -0.5 rotation of panel a). The same can be done with panel -1 c). Thus, to calculate the expectation of Vˆee we only need to consider the case 0 r a . -1.5 ≤ ≤ 1 The solution of Eqs. (37)-(38) is thus very reasonable -2 from the physical point of view. Moreover, it can be -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 easily verified that it satisfies the properties required by stationarity of Eq. (22), and that it corresponds to a FIG. 2: The positions of the 3 electrons in the SCE state for smooth external potential v(r), which can be computed the Li atom density. The two big circles are the projections of via the spheres containing, on average in the quantum mechanical problem, 1 electron (radius a1) and 2 electrons (radius a2). ∂ The position r of the first electron is varied along the vertical v′(r)= V (r , r , r ,θ ,θ ) , 1 1 2 2 ∞ ee 1 2 3 2 3 r,f2(r),f3(r),θ2(r),θ3(r) z axis a) from 0 to a , b) from a to a , and c) from a to . −∂r1 (39) The other two electrons have distances from the center given, respectively, by f2(r) and f3(r) of Eqs. (37)-(38), and angu- where θ (r) is a shortened notation for θ (r, f (r),f (r)). i i 2 3 lar positions given by the minimizing angles θ2(r, f2(r), f3(r)) The potential obtained by integrating Eq. (39) (again and θ3(r, f2(r), f3(r)). In each panel, the starting position of using the accurate density for the Li atom of Ref. [13]) the 3 electrons is represented by a full circle (•) and their final is reported in Fig. 3. Notice that v(r) is finite, with a position by an empty circle (◦). The dashed curves represent zero first derivative, at r = 0: the electron-nucleus cusp the “trajectories” of the 3 electrons, and the arrows their di- in the density ρ(r) is compensated by the kinetic energy rection. The three panels are actually equivalent (see text). for any value of α (including the very limit α ), as Distances are in atomic units. discussed in Ref. 16. For large r, we have v(r→ ∞ ) = N 1 → ∞ − , a necessary condition to include as a stationary − r 8
0 state in a smooth density have one configuration in which they are both at the same distance from the nucleus, but -0.5 for three electrons this does not happen. One can also easiliy check that, again, the “breathing” solution, with its potential v(r)= 1 , is not a minimum for the cor- α → ∞ -1 − √3r responding E ([v], r , r , r ,θ ,θ , φ ).
(r) | pot 1 2 3 2 3 3
α ext We can at this point clearly see the difference between
v -1.5 Li atom 1
− the SCE state for a smooth atomic density (Fig. 2) and α − 2/r -2 the more familiar Wigner-crystal-like state (for a case with three electrons see, e.g., Ref. 17), in which the three electrons are localized at the vertices of a triangle at a -2.5 0 2 4 6 8 10 certain distance from the nucleus, say r0. In this latter r case, the density becomes very peaked around r0, loosing any resemblence with an atomic density, and becoming FIG. 3: The external potential v(r) of Eq. (9) for the SCE more and more similar to Eq. (17). state in the case of the Li atom. The asymptotic expansion for → ∞ − N−1 large r, v(r ) = r , is also reported. All quantities are in Hartree atomic units. C. N = 4 (Be atom)
We have now 5 relative angles to consider. As in the configuration the one in which one of the N electrons is N = 3 case, we compute the minimizing angular function at infinity. The large-r asympotic expansion 2/r is also Ω(r , r , r , r ) numerically. shown in the same Fig. 3. − 1 2 3 4 We then construct the radial co-motion functions, With this potential v(r) we have also verified that the f (r), f (r), f (r), following the same ideas used for the stationary solution of Eqs. (37)-(38) is the absolute min- 2 3 4 case N = 2 and 3: we divide the space in four spher- imum of E ([v], r , r , r ,θ ,θ , φ ). All these checks pot 1 2 3 2 3 3 ical shells, each containing, on average in the quantum make us believe that Eqs. (37)-(38) are the correct SCE mechanical problem, one electron. The radial co-motion solution for the Li atom: it is difficult to imagine a differ- functions that make the four electrons always be in four ent solution that yields a lower value for the expectation distinct shells are then of Vˆee, since our solution is the one that makes the 3 1 electrons be always as far as possible from each other, Ne− (2 Ne(r)) r a2 f2(r)= 1 − ≤ (41) without violating the constraint of yielding the smooth N − (Ne(r) 2) r>a2 e − density ρ(r). 1 Ne− (2 + Ne(r)) r a2 We can thus finally obtain the value W [ρ] = f3(r)= 1 ≤ (42) ∞ N − (6 Ne(r)) r>a2 2.6030 Hartree for the Li atom, via e − − 1 f4(r)= Ne− (4 Ne(r)), (43) W [ρ]+ U[ρ]= − ∞ 1 a1 where, again, ai = N − (i). These functions satisfy the 2 e dr 4π r ρ(r) Vee(r, f2(r),f3(r),θ2(r),θ3(r)), (40) transformation properties (22), and are reported in Fig. 4 Z0 for the case of an accurate correlated density [18] of the where we have used the fact that, because the three elec- Be atom. trons are indistinguishable, integrating from 0 to is We can then obtain, using the analogue of Eq. (40) for ∞ the same as integrating three times from 0 to a1. N = 4, the value W [ρ] = 4.0212 Hartree for the Be As a final remark, we comment briefly on the fact atom. ∞ − that one never obtains the configuration in which the To check our result, we also computed the correspond- three electrons are at the same distance from the nu- ing external potential v(r), which, as shown in Fig. 5, is cleus, at the vertices of an equilateral triangle. The rea- continuous and smooth. We also verified that our sta- son is that the only f2(r) and f3(r) compatible with tionary solution is the absolute minimum for the corre- this configuration (even for only one single value of r, sponding Epot([v], r1, r2, r3, r4, Ω). say r0) that also satisfy the required properties (22) Interestingly, we found that, for the case of the Be are the ones corresponding to the “breathing” solution, atomic density, the four electrons always lie on the same f2(r) = f3(r) = r, as it can be easily verified by inte- plane, containing the nucleus. This effect seems to be due grating both sides of Eq. (28) with the initial condition to the shell structure of the atomic density: by repeating r = r0, fn(r0) = r0. This gives the two possibilites the calculation using a simple exponential density, ρ(r)= 1 3 βr f (r)= r or f (r)= N − (2N (r ) N (r)): any choice β e− /2π, we found that the four electrons are often A B e e 0 − e for f2(r) and f3(r) which contains the function fB(r) non-complanar. Instead, by using other test densities does not satisfy the properties (22). This is a difference with a shell structure, we found again the four electrons with the N = 2 case in which, instead, the solution is always on the same plane. The shell structure seems to f(r)= fB(r) with Ne(r0) = 1: two electrons in the SCE force the four electrons to have distances from the nucleus 9
6 4 B atom Be atom 3.5 5 3 f3(r) f5(r) 4 f (r) 2.5 3
3 2
1.5 2 f4(r) f4(r) 1 f (r) f2(r) 1 2 0.5
0 0 0 1 2 3 4 5 6 0 0.5 1 1.5 2 2.5 3 3.5 4 r r
FIG. 4: The radial co-motion functions f2(r), f3(r) and f4(r) FIG. 6: The radial co-motion functions for the sphericalized for the Be atom density. Hartree atomic units are used. B atom density. Hartree atomic units are used.
0 3.5 C atom -0.5 3 -1 f (r) f3(r) -1.5 Be atom 2.5 5 α → ∞
(r) | -2 2 α ext
v -2.5 − 3/r 1 1.5 − α -3 f (r) f (r) 1 f4(r) 6 2 -3.5 -4 0.5 0 2 4 6 8 10 0 r 0 0.5 1 1.5 2 2.5 3 3.5 r FIG. 5: The external potential v(r) of Eq. (9) for the SCE state in the case of the Be atom. The asymptotic expansion FIG. 7: The radial co-motion functions for the sphericalized → ∞ − N−1 for large r, v(r ) = r , is also reported. All quanti- C atom density. Hartree atomic units are used. ties are in Hartree atomic units.
1 where ai = Ne− (i). For odd N, these equations give all that are distributed in a highly nonuniform way, so that the needed N 1 radial co-motion functions, while for forces compensation (see Sec. IVB) can be achieved only even N we have− to add the last function, if they are on the same plane. 1 f (r)= N − (N N (r)). (46) N e − e
D. The general N-electron solution Using these radial co-motion functions, which satisfy the properties (22) for any N, we calculated the SCE value W [ρ] for accurate sphericalized densities of the B This way of constructing the solution can be general- and C atoms∞ [19], and for the Ne atom density [18, 20]. ized to any number of electrons N. The radial co-motion The resulting radial co-motion functions are displayed in functions that make the N electrons always occupy N Figs. 6-8. The results for W [ρ] are reported in Table I, different shells containing, on average in the quantum and the corresponding potentials∞ v(r) are shown in Fig. 9. mechanical problem, one electron, are as follows. Define In all cases we find the N electrons non-complanar. an integer index k running for odd N from 1 to (N 1)/2, and for even N from 1 to (N 2)/2. Then − − VI. COMPARISON WITH THE PC MODEL 1 Ne− (2k Ne(r)) r a2k f2k(r)= 1 − ≤ (44) N − (Ne(r) 2k) r>a2k e − In the absence of spherical symmetry there seems to be 1 Ne− (Ne(r)+2k) r aN 2k no general strategy for finding the initial conditions to in- f2k+1(r)= 1 ≤ − (45) Ne− (2N 2k Ne(r)) r>aN 2k, tegrate Eqs. (24), although we are still investigating this − − − 10
SCE PC W∞ [ρ] (H) W∞ [ρ] (H) error (mH) Ne atom 2 H −0.3125 −0.3128 0.3 f3 He −1.500 −1.463 37 f 9 f Li −2.6030 −2.5559 47 f7 5 1.5 Be −4.0212 −3.9608 60 B −5.7063 −5.6502 56 − . − . f2 C 7 7817 7 7192 63 1 Ne −20.035 −20.000 35
f4 ∞ f TABLE I: Comparison of the values W [ρ] in Hartree atomic 0.5 6 units obtained with the SCE construction, and with the PC model [8]. The absolute errors of the PC model are also re- f10 f8 ported. 0 0 0.5 1 1.5 2 r weak-interaction limit α 0. In this way, properties → FIG. 8: The radial co-motion functions for the Ne atom den- of real atoms and molecules, including those with weak sity. Hartree atomic units are used. correlations, can be predicted accurately [6] without deal- ing explicitly with the complicated wave function at the realistic interaction strength α = 1. Both the extreme 0 limits α 0 and α are mathematically much sim- → → ∞ B pler than the realistic situation at α = 1. Consequently, C more relevant than the error of the quantity W [ρ] it- -5 self is the error it causes in the ISI correlation energy∞ [6] Ne ISI ISI
α → ∞ E [ρ] = E [ρ] E [ρ]. Since the ISI model of Ref. 6 c xc − x
(r) | -10 uses, along with W [ρ], also the coefficient W ′ [ρ] which ∞ ∞ α ext is not known exactly, we consider here the earlier ISI v 1
− version SPL of Ref. 5, α -15 SPL Ex[ρ] W [ρ] Wα [ρ] = W [ρ]+ − ∞ , (47) ∞ 1+2Q[ρ] α -20 0 1 2 3 4 5 6 7 with p r 2 EGL2[ρ] Q[ρ]= | c | . (48) FIG. 9: The external potential v(r) of Eq. (9) for the SCE Ex[ρ] W [ρ] state in the case of the Ne atom and of the sphericalized B − ∞ 1 and C atoms. All quantities are in Hartree atomic units. Analytical integration of dαW SPL[ρ] E [ρ] yields 0 α − x R SPL 1+2Q[ρ] 1 Ec [ρ]= Ex[ρ] W [ρ] − 1 . problem. So far, such cases can only be treated approx- − ∞ " Q[ρ] − # p imately, using instead of the exact expression (25) the (49) PC model (6) for the functional W [ρ]. The PC model SPL ∞ The correlation energies Ec [ρ], evaluated with the val- has been tested successfully [8] against the meta-GGA ues W [ρ] of Table I, are reported in the columns “(PC)” functional of Ref. 21 that is expected to be accurate in and “(SCE)”∞ of Table II. While the SPL prediction the limit α . The present exact (SCE) results, in → ∞ comes quite close to the exact correlation energy of the contrast, provide a rigorous test for the PC model. He atom, the case of the Be and Ne atoms is less satisfy- Table I documents reasonable values from the PC ing. Nevertheless, the improvement beyond the second- model for all the atoms considered here. A severe test order correlation energy is remarkable, since the only ad- is the trivial case of the hydrogen atom H, with the one- ditional information used here is the coefficient W [ρ] 1 2r ∞ electron density ρ(r) = e− , where Vˆ = 0 for any π h eei (but not W ′ [ρ]). value of α and thus W [ρ]= 5 Hartree. For the other For the He,∞ Be, and Ne atoms, the error introduced ∞ − 16 atoms in Table I, the PC error is less than 63 mH 39 by the PC model in ESPL[ρ] is always less than 1 mH, ≈ c kcal/mole. showing that in the case of neutral atoms the SPL-ISI The quantity W [ρ] refers to the limit of infinitely correlation functional is not too sensitive to the exact strong repulsion between∞ electrons. Being the sub- value of W [ρ]. However, we have to keep in mind that ject of the present paper, this extreme limit is not di- neutral atoms∞ are systems that resemble much more to rectly relevant for real electron systems such as atoms the noninteracting KS system (α = 0) than to the SCE or molecules. Rather, it is studied here in order to be state (α = ). We expect that for more correlated sys- combined with additional information from the opposite tems (e.g.,∞ streched bonds) the ISI correlation energy is 11
GL2 exact Ex[ρ] Ec [ρ] (PC) (SCE) Ec [ρ] in this paper gives the α limit of DFT. We thus He −1.0246 −0.0503 −0.0413 −0.0418 −0.042 consider the exactly solvable→ 2-electron ∞ hamiltonian [25] Be −2.674 −0.125 −0.1054 −0.1061 −0.096 − − − − − 1 k α Ne 12.084 0.469 0.4205 0.4207 0.394 Hˆ α = 2 + 2 + α r2 + r2 r r 2, (A1) −2 ∇1 ∇2 2 1 2 − 2 | 1 − 2| TABLE II: The exchange energy Ex[ρ], the second order GL2 and we analyze the case in which α and the density G¨orling-Levy correlation energy Ec [ρ] [22], and the esti- is kept fixed, equal to the one at α→= ∞ 1, by means of a mate of the correlation energy Ec[ρ] from the SPL ISI model SCE PC suitable k . [5] using the values W∞ [ρ] (SCE) and W∞ [ρ] (PC), com- α pared with “exact” values. All energies are in Hartree atomic The ground-state one-electron density of the hamilto- α units. nian Hˆ is given by
3/2 2 2β βr 2 kα(kα 2α) ρ(r)= e− , β = − , (A2) much more sensitive to the exact value of W [ρ]. The 3/2 π √pkα 2α + √kα investigation of such cases will be the object∞ of future − work. so that kα must keep β independent of α and thus satisfies the equation
k (k 2α) k (k 2) VII. CONCLUSIONS AND PERSPECTIVES α α − = 1 1 − , (A3) √k 2α + √k √k 2+ √k pα − α p1 − 1 We have reformulated the strong-interaction limit of where k1 > 2 in order to have a bound system. We easily density functional theory in terms of a classical problem find with a degenerate minimum, obtaining a consistent solu- 2 β 1/2 tion for this limit. Even if an antisymmetric wavefunc- kα =2α + + O(α− ). (A4) tion can be explicitly constructed, the strong-interaction →∞ 4 limit of DFT is entirely characterized by much simpler The square of the spatial wavefunction is equal to quantities, the so called “co-motion” functions, which are related to the electronic density ρ(r) via the differential 3/2 k (k 2α) equation (24). The results of this work can be useful r r 2 α α ψα( 1, 2) = 2 − to construct interpolations between the weak- and the | | " p π # × 1 2 2 strong-interaction limits of DFT, and to test other ap- (√kα+√kα 2α)(r1 +r2 ) (√kα √kα 2α)r1 r2 e− 2 − − − − · . (A5) proximate functionals in the α limit [23]. → ∞ Future work will be mainly devoted to the calculation Now, we want to compare the exact result with the of this limit for non-spherical densities, and to study the SCE construction. We thus follow the same steps done generalization to the next leading term, W ′ [ρ]. The cal- for the case of the He atom in Subsec. V A. Again, we find culation of the intracule density (the probability∞ distribu- that the minimizing angle is θ = π and that the radial co- tion for the electron-electron distance) will be also carried motion function must satisfy the property f(f(r)) = r, out in connection with the correlation energy functional so that the only possible choices are the ones of Eqs. (32)- constructed in Ref. 24. (34). In this case, however, the “breathing” solution of Eq. (32), f(r) = r, is the one that yields the minimum expectation of “Vˆee”, and that is the global minimum of Acknowledgments the corresponding Epot. The very different nature of har- monic forces with respect to Coulomb repulsion makes We thank Prof. C. Bunge for the fully correlated den- the solution completely different also from a qualitative sity of the Li atom, Prof. C.J Umrigar for the VMC den- point of view. The “breathing” solution, that was a max- sities of the Ne and Be atoms, and Dr. J. Toulouse for imum for Coulomb forces, becomes the true minimum in the sphericalized densities of the B and C atoms. One of the case of harmonic forces. In this latter case, in fact, the authors (P.G.-G.) gratefully acknowledges Prof. E.J. the interaction is not singular at zero electron-electron Baerends for the warm hospitality in his group, where distance so that the two electrons can get on top of each part of this work was done. other (and they do, at the “nucleus”). This means that in this case even in the α limit the spin state can still play a role. The SCE one-body→ ∞ potential is obtained APPENDIX A: HARMONIC INTERACTIONS by
v′(r)= r + f(r)=2r, (A6) It is instructive to see how the SCE solution changes if we replace the Coulomb repulsion with an harmonic so that v(r) = r2 agrees with the leading term (O(α)) interaction. Since in this case we have the exact solution, of Eq. (A4) (remember that in the SCE construction we we can also clarify how the SCE construction presented directly consider all the quantities divided by α). 12
In order to compare the SCE solution with the ex- When α , the first gaussian in Eq. (A7) with its → ∞ act wavefunction of Eq. (A5) in the α limit, we factor in front tends to the distribution δ(r1 + r2). The have first to keep in mind that Eq. (A5)→ is ∞ the solution second gaussian gives 1 ρ(r ). Thus, as α we have 2 1 → ∞ of the hamiltonian Hˆ α, so that it includes also the ki- netic energy operator Tˆ. Equation (A4) shows that the ˆ α external potential Vext that keeps the density fixed has ˆ α ˆ ˆ 1/2 the large-α expansion Vext = αV + V0 + O(α− ). The ∞ SCE construction only gives Vˆ , but the exact solution of ∞ Eq. (A5) always includes also the next term, Vˆ0, because r it is of the same order of Tˆ. The comparison with the 2 ρ( 1) lim ψα(r1, r2) = δ(r1 + r2), (A8) α SCE solution can be done in the following way. Replace →∞ | | 2 0 in Eq. (A5) the large-α expansion of kα up to orders α , 2 2 k 2α + β . We can rearrange the result, with c = β , α → 4 4 3/2 1 2 2 √2α + c √c (√2α+c √c) r1+r2 ψ (r , r ) − e− 2 − | | | α 1 2 | → 2π 3/2 2 2 which is exactly the SCE solution, since in the case of 1 2 c(2α + c) √c(r1 +r2 ) e− (A7). harmonic forces f(r)= r. ×π3/2 √2α + c √c − p − !
[1] W. Kohn, Rev. Mod. Phys. 71, 1253 (1999). [16] A.´ Nagy and Zs. J´anosfalvi, Philos. Mag. 86, 2101 (2006). [2] A.E. Mattsson, Science 298, 759 (2002). [17] J. Cioslowski and K. Pernal, J. Chem. Phys. 125, 064106 [3] C. Fiolhais, F. Nogueira, and M. Marques (eds.), A (2006). Primer in Density Functional Theory (Springer-Verlag, [18] C. Filippi, X. Gonze, and C.J. Umrigar, in Recent de- Berlin, 2003). velopments and applications in modern DFT, edited by [4] E.P. Wigner, Phys. Rev. 46, 1002 (1934); Trans. Faraday J.M. Seminario, (Elsevier, 1996). Soc. 34, 678 (1938). [19] J. Toulouse, private communication. The sphericalized [5] M. Seidl, J. P. Perdew, M. Levy, Phys. Rev. A 59, 51 densities of the B and C atoms were obtained from vari- (1999). ational Monte Carlo using accurate optimized wavefunc- [6] M. Seidl, J. P. Perdew, S. Kurth, Phys. Rev. Lett. 84, tions (see J. Toulouse and C.J. Umrigar, submitted to J. 5070 (2000). Chem. Phys.; C. J. Umrigar, J. Toulouse, C. Filippi, S. [7] M. Seidl, Phys. Rev. A 60, 4387 (1999). Sorella, and R.G. Hennig, cond-mat/0611094). [8] M. Seidl, J. P. Perdew, S. Kurth, Phys. Rev. A 62, [20] A. I. Al-Sharif, R. Resta, and C. J. Umrigar, Phys. Rev. 012502 (2000); 72, 029904 (2005) (E). A 57, 2466 (1998). [9] R.J. Magyar, W. Terilla, and K. Burke, J. Chem. Phys. [21] J. P. Perdew, S. Kurth, A. Zupan, and P. Blaha, Phys. 119, 696 (2003). Rev. Lett. 82, 2544 (1999); 82, 5179 (1999) (E). [10] T. Gimon, A. G¨orling, M. Seidl, and J. P. Perdew, in [22] M. Ernzerhof, Chem. Phys. Lett. 263, 499 (1996); private preparation. communication. [11] A. G¨orling and M. Levy, Phys. Rev. B 47, 13 105 (1993); [23] V. N. Staroverov, G. E. Scuseria, J. Tao, and J. P. Phys. Rev. A 52, 4493 (1995). Perdew, Phys. Rev. B 69, 075102 (2004); J. Jung, P. [12] E. Lieb, Int. J. Quantum Chem. 24, 243 (1983). Garcia-Gonzalez, J. E. Alvarellos, and R. W. Godby, [13] C. Bunge, private communication. The full CI density of Phys. Rev. A 69, 052501 (2004). the Li atom has been computed with a very large basis [24] P. Gori-Giorgi and A. Savin, Phys. Rev. A 71, 032513 set with 8 s functions and up to k functions. (2005); Philos. Mag. 86, 2643 (2006); cond-mat/0611324. [14] see, e.g., P. Fulde, Electron Correlation in Molecules and [25] see, e.g., E.R. Davidson, Reduced Density Matrices in Solids (Springer, Berlin, 1995). Quantum Chemistry (Academic Press, New York, 1976). [15] P. Ziesche, J. Tao, M. Seidl, and J. P. Perdew, Int. J. Quantum Chem. 77, 819 (2000).