Pinning of Fermionic Occupation Numbers
Christian Schilling,1 David Gross,2 and Matthias Christandl1 1Institute for Theoretical Physics, ETH Zurich,¨ Wolfgang–Pauli–Strasse 27, CH-8093 Zurich,¨ Switzerland 2Institute for Physics, University of Freiburg, Rheinstrasse 10, D-79104 Freiburg, Germany (Dated: October 10, 2018) The Pauli exclusion principle is a constraint on the natural occupation numbers of fermionic states. It has been suspected since at least the 1970’s, and only proved very recently, that there is a multitude of further constraints on these numbers, generalizing the Pauli principle. Here, we provide the first analytic analysis of the physical relevance of these constraints. We compute the natural occupation numbers for the ground states of a family of interacting fermions in a harmonic potential. Intriguingly, we find that the occupation numbers are almost, but not exactly, pinned to the boundary of the allowed region (quasi-pinned). The result suggests that the physics behind the phenomenon is richer than previously appreciated. In particular, it shows that for some models, the generalized Pauli constraints play a role for the ground state, even though they do not limit the ground-state energy. Our findings suggest a generalization of the Hartree-Fock approximation.
PACS numbers: 03.67.-a, 05.30.Fk, 31.15.-p
Introduction.— In 1925 the study of atomic transitions led In a ground-breaking work building on recent progress in to Pauli’s exclusion principle [1]. It states that for identical invariant theory and representation theory, Klyachko exhib- fermions the occupation number for any quantum state cannot ited an algorithm for computing all such Pauli-like constraints exceed the value 1. By 1926, Dirac [2] and Heisenberg [3] had [4, 8]. In fact, his work is part of a more general effort in identified the exclusion principle as a consequence of a much quantum information theory addressing the quantum marginal deeper statement: the anti-symmetry of the many-fermion problem which asks when a given set of single-site reduced wave function. While anti-symmetry allows one to find the density operators (marginals) is compatible in the sense that correct solutions to the full many-fermion Schrodinger¨ equa- they arise from a common pure global state (see also [9– tion, it does not render the exclusion principle obsolete: In 12]). The global state may be subject to certain symme- many situations, the latter is sufficient to predict the qualita- try constraints—one obtains the fermionic case (commonly tive behavior of fermionic systems without the need to resort known as the N−representability problem [13, 14]) by requir- to (often computationally intractable) ab initio methods. The ing total anti-symmetry under particle exchange. Klyachko Aufbau principle for elements in the periodic table serves as a showed that for fixed particle number N and dimension d of prime example. the 1−particle Hilbert space, the generalized Pauli constraints This observation motivates the study of generalizations of amount to affine inequalities of the form the exclusion principle, which, maybe surprisingly, exist and κ + κ λ + ... + κ λ ≥ 0. exhibit an extremely rich structure [4]. To set the scene, note 0 1 1 d d (4) that the Pauli constraint can be stated succinctly as Geometrically, these constraints define a convex polytope d PN,d ⊂ R of possible spectra (for more details see Appendix 0 ≤ λi ≤ 1 ∀i (1) A). In general, if a spectral inequality such as (1) or (4) is (approximately) saturated we say that the corresponding spec- in terms of the natural occupation numbers λi, which are trum is (quasi-) pinned to its extremum. the eigenvalues of the 1-particle reduced density operator The natural question arises as to whether ground states of (1−RDO) ρ1, normalized to the particle number N. The util- relevant many-body models saturate some of those inequal- ity of the exclusion principle is grounded in the fact that in the ities. Strong numerical evidence supporting this conjecture ground states of many-fermion systems, one often observes has been presented in [15]. The problem is challenging to ad- arXiv:1210.5531v1 [quant-ph] 19 Oct 2012 λi ≈ 0 or λi ≈ 1, which is equivalent to stating that the dress analytically, as one has to not only compute the ground Hartree-Fock approximation works fairly well in these sys- state, but also determine and diagonalize the corresponding tems. 1−RDO. It had been observed in the 1970s that there are further lin- Here, we present for the first time an analytic analysis. For ear inequalities respected by the natural occupation numbers the ground state of a model of interacting fermions in a har- as a result of global anti-symmetry [5–7]. One particular ex- monic potential, the natural occupation numbers are calcu- 3 ample is the so-called Borland-Dennis setting ∧ [H6] of three lated. We obtain several results. We confirm that for this fermions and a six dimensional 1−particle Hilbert space H6 very natural model, the natural occupation numbers lie indeed [7]. Here, the set of constraints is given by close to the boundary of set of allowed ones. The analytic analysis enables us to track the “trajectory” of eigenvalues as λ1 + λ6 = λ2 + λ5 = λ3 + λ4 = 1 (2) a function of the interaction strength between the fermions. (6) D := λ5 + λ6 − λ4 ≥ 0 (3) What is conceptually also important, is the fact that the eigen- values never lie exactly on the boundary. To see why one on the ordered eigenvalues λi ≥ λi+1. could expect the opposite, note that the ground state energy of 2
P (i,j) (i,j) N 2 a Hamiltonian H = i,j h with two-particle terms h Hilbert space ∧ [H∞], with H∞ = L (R), i.e. we are treat- can be represented as a constrained optimization problem ing the N particles as fermions (without spin). In [16] H has been diagonalized and the ground state reads in spatial repre- N X (i,j) (i,j) sentation (~x = (x1, . . . , xN )) Emin = min tr[h ρ ] (i,j) 2 ρ2 i,j=1 Y (i,j) ΨN (~x) = const × (xi − xj) (7) where the ρ2 are 2-particle density operators that are com- 1≤i
The solution for N = 3 reads: which is positive for δ small enough. Hence the full spectrum 0 0 40 1390 λ also fails to be pinned. The case d = ∞ works in the same 1 − λ = δ6 − δ8 + O(δ10) 1 729 59049 way, up to some mild assumptions (see Appendix B). 2 232 3926 We will now apply these considerations to our model. First, 1 − λ = δ4 − δ6 + δ8 + O(δ10) 2 9 729 10935 we truncate to 6 dimensions, which has the advantage that the 3 2 64 81902 spectral polytope corresponding to ∧ [H6] is 3-dimensional 1 − λ = δ4 − δ6 + δ8 + O(δ10) 3 9 243 295245 and can thus be visualized. In a second step, we take a seventh 2 64 73802 eigenvalue into account. This setting turns out to be strong λ = δ4 − δ6 + δ8 + O(δ10) 4 9 243 295245 enough to establish all statements we have mentioned above – 2 232 3976 λ = δ4 − δ6 + δ8 + O(δ10) namely that the total spectrum is not exactly pinned, but does 5 9 729 10935 lie close to the boundary (quasi-pinned). 40 2200 λ = δ6 − δ8 + O(δ10) 6 729 59049 (d) 80 λ = δ8 + O(δ10) (a) 7 2187 10 λ8 = O(δ ) . . . . (10) (b) Similar results follow for N = 2. Note the non-trivial hierar- chy of the eigenvalues, (c)
2k−6 2k−4 λk = ck δ + O(δ ) , (11) FIG. 1: Spectral “trajectory” v(δ) (thick line, partially covered by facet, schematic) up to correction of order δ8 and small part of for all k ≥ 5. Moreover, the spectrum λ for δ not too large the polytope P around vertex v(a) obtained by cutting P along the is very close to the one of a single Slater determinant. For dashed lines instance, λi, i = 1, 2, 3 deviate from 1 and λj, j ≥ 4 from 0 3 only by at most 1 percent if |δ| ≤ 0.5. This emphasizes the The simplest non-trivial setting ∧ [H6] becomes an appro- priate description if λ , λ ,... ≈ 0. By (11), this condition relevance of the Pauli constraints (1). 7 8 is fulfilled if δ is small enough that contributions of order δ8 Quasi-Pinning by Generalized Pauli Constraints.— can be neglected. Choosing λ4, λ5 and λ6 as free parameters Equipped with the explicit solution (10), we can proceed according to (2), the corresponding polytope P3,6 effectively to analyze whether the generalized Pauli constraints play a 3 reduces [15] to a 3-dimensional polytope P ⊂ R with ver- role for the ground state. While the underlying 1−particle tices, Hilbert space H∞ is infinite-dimensional, the scaling (11) 1 1 implies that the spectrum is strongly concentrated on a v(a) = (0, 0, 0) , v(b) = , 0 low-dimensional subspace, at least for small δ. One can use 2 2 1 1 1 1 1 1 this fact to deduce statements about the position of the total v(c) = , , , v(d) = , , . (14) eigenvalues from truncated information alone. 2 4 4 2 2 2 This can be understood from simple geometric consid- Hence the vertex v(a) corresponds to single Slater determi- erations. Let d < d0 < ∞. Because a d-dimensional nants and the 2-facet spanned by v(a), v(b) and v(c) is defined 1−particle Hilbert space can be imbedded into any (larger) by D(6) = 0, which is here the one of interest and represents d0-dimensional one, one sees that the convex polytope P is N,d exact pinning by constraint (3). We first illustrate schemati- nothing but the intersection between P 0 and the set of spec- N,d cally our result (10) in Fig. 1. There, the spectral “trajectory”, tra with only d non-zero eigenvalues (see also Appendix B). Hence any facet of PN,d arises from the intersection of some v(δ) = (λ4(δ), λ5(δ), λ6(δ)), P 0 facet of N,d with the subspace of said spectra. Formally, a 8 0 is shown as a thick line (neglecting effects of order δ and F P 0 facet of N,d consists of points saturating a generalized (a) Pauli constraint higher). It starts at the vertex v which corresponds to the non-interacting situation δ = 0. When increasing the fermion- d d0 fermion interaction, v(δ) leaves the vertex v(a) and moves 0 X X (a) (b) (a) D (λ) = κ0 + κiλi + κiλi ≥ 0. (12) along the edge (v , v ), the distance to v growing as 4 6 i=1 i=d+1 δ . On the finer scale δ , v(δ) also moves away from the edge but is still pinned to the boundary of the polytope, lying on Denote the first two summands by D(λtr), where λtr = the 2-facet spanned by v(a), v(b) and v(c). This is the bottom d tr (λi)i=1 is the truncated spectrum. Clearly, D(λ ) = 0 de- area in Fig. 1, corresponding to the constraint (3). scribes the restriction of the facet to the d-dimensional setting. The pinning seems to disappear if we consider higher or- Now assume the truncated spectrum λtr(δ) is not pinned, i.e. ders. From (10), we can infer that the distance to the 2-facet D(λtr(δ)) > 0, then the hierarchical scaling (11) implies (v(a), v(b), v(c)) increases as δ8, D0 λ(δ) = D λtr(δ) + O(δ2d−4), (13) D(6)(δ) = ζ(6) δ8 + O(δ10) (15) 4
(6) 4510 with ζ = 59049 . However, this calculation is inconclusive, a new entanglement measure has been suggested, which, for as the distance to the boundary is of the same order, δ8, as the the Borland-Dennis setting, naturally separates exactly pinned truncation error (recall (13)). and non-pinned states. We believe that these findings open To resolve the issue, we take another eigenvalue, λ7, into up a potentially significant avenue for investigating the struc- 3 account. We thus work in the setting ∧ [H7] with four con- ture of fermionic ground states via their natural occupation straints D(7) ≥ 0 for i = 1,..., 4 [4]. This setting is valid as numbers— generalizing a program that has long been carried i out for the Hartree-Fock case [18]. long as λ8, λ9,... ≈ 0 or in other words we neglect terms of 10 3 order δ or higher (but in contrast to the setting ∧ [H6] we We close by speculating that these insights could give rise include δ8−terms). Since the polytope is now 6−dimensional to improved numerical procedures. The idea is to replace the we cannot present our results graphically anymore. The re- ground state ansatz of one single Slater determinant by the sults (10) lead to (i = 1, 2, 3, 4) states corresponding to the points lying on the (quasi-)pinning polytope facet. In contrast to the configuration interaction (CI) (7) (7) 8 10 Di = ζi δ + O(δ ) , (16) methods in quantum chemistry which improve the Hartree- Fock approximation by adding several arbitrary Slater deter- (7) 20 (7) 10 (7) 50 (7) 2890 minants to the Hartree-Fock state our method would add only with ζ1 = 2187 , ζ2 = 243 , ζ3 = 2187 , ζ4 = 59049 . 3 a few but carefully chosen additional Slater determinants. Here in the ∧ [H7]-analysis, the new result is that all four dis- (7) 8 Conclusions.— For a natural model of interacting tances Di are non-zero to a smaller order, δ , than the error of spectral truncation, δ10. Together with the comments at fermions in a harmonic trap we analytically calculated the the beginning of this section, this shows that the absence of leading orders of the eigenvalues of the 1−RDO correspond- pinned spectra is genuine, rather than an artifact of the trunca- ing to the fermionic ground state as function of δ, a measure tion. Given this, the quasi-pinning found here is surprisingly for the fermion-fermion interaction strength. The investiga- strong. In particular it exceeds by four additional orders the tion of the generalized Pauli constraints has shown that none (quasi-)pinning by Pauli’s exclusion principle constraints (1), of them is completely saturated, which might be a generic property of all continuous models of interacting fermions. In 2 particular, the findings show that it is likely extremely chal- 0 ≤ 1−λ (δ), 1−λ (δ), λ (δ), λ (δ) = δ4 +O(δ6). (17) 2 3 4 5 9 lenging to use numerical methods to distinguish between gen- uinely pinned and mere quasi-pinned states. This underscores Generalizing Hartree-Fock.— In this section, we dis- the need for analytical analyses, first provided here. On the cuss what conclusions can be drawn about the N−fermion other hand the pinning up to corrections of order δ8 we found state |Ψi itself, given information just about the position here is surprisingly strong. In particular it exceed the one by of the eigenvalues of the corresponding 1-RDO relative to the Pauli exclusion principle constraints (1), which are pin- the boundary of the spectral polytope. In this way, quasi- ning up to corrections of order δ4 only. pinned spectra are endowed with a physical significance. Acknowledgements.— We thank F.Verstraete for helpful To this end, recall the basic fact that the spectrum λ = Sl discussions. CS and MC acknowledge support from the (1,..., 1, 0,..., 0) can only arise from a Slater determinant Swiss National Science Foundation (grants PP00P2-128455 |Ψi = |1,...,Ni. It is well-known that this statement is and 20CH21- 138799), the National Centre of Competence in stable under small deviations: if λ ≈ λ , then |Ψi is well- Sl Research ‘Quantum Science and Technology’ and the German approximated by a Slater determinant (see [18] or Appendix Science Foundation (grant CH 843/2-1). DG’s research is sup- D). ported by the Excellence Initiative of the German Federal and For exactly pinned spectra, there is a simple generalization State Governments (grant ZUK 43). of these observations. In [15], it is stated that constraint (3) can be saturated only by states of the form
|Ψi = α|1, 2, 3i + β|1, 4, 5i + γ|2, 4, 6i, Appendix a fact is dubbed “selection rule” for Slater determinants (see also Appendix C). The general statement reads: if D(λ) ≥ 0 This appendix is split into four sections. The first one in- is a generalized Pauli constraint, then D(λ) = 0 can only troduces the notation and repeats the solution of the fermionic be achieved by states |Ψi which are superpositions of those quantum marginal problem. In the second section we explain Slater determinants whose (unordered) spectra also saturate how to simplify the pinning analysis by truncating the spec- D. trum. This amounts to the proof of statement (13), a relation What is more important, a stable version of this state- connecting polytope distances of the correct and truncated ment applying to quasi-pinned states can be found—at least marginal setting. The third section introduces a selection rule, for specific situations. In the Appendix D, we show that which explains how the structure of a N−fermion state sim- for the Borland-Dennis setting, spectra in the vicinity of the plifies if its natural occupation numbers are exactly pinned to facet corresponding to constraint (3) are approximately of the some Pauli facet and applies it to the Borland-Dennis setting. form above. In particular, quasi-pinned states are close to In the last section we present a modification of this selection states containing fairly low amounts of multi-partite entan- rule for the case of only approximate pinning. This then justi- glement as quantified by the Schmidt number [19]. In [20] fies our Hartree-Fock generalization. 5
A. Notation and Fermionic Quantum Marginal Problem.— and their utility for applications. These natural orbitals induced by a fixed state |Ψi ∈ N 0 d0 The problem of determining all spectra ∧ [Hd0 ], d ∈ N ∪ {∞}, define a basis B1 := {|ki}k=1 for the 1−particle Hilbert space Hd0 . For ease of notation d0 λ = (λi)i=1 , 1 ≥ λ1 ≥ λ2 ≥ ... ≥ λd0 ≥ 0 (18) we skip the argument Ψ of |i(Ψ)i. Basis B1 then induces the N basis BN for ∧ [Hd0 ] of corresponding Slater determinants 0 of 1−particle reduced density operators (1−RDO) ρ1 arising (1 ≤ i1 < . . . < iN ≤ d ) N from some pure N−fermion state |Ψi ∈ ∧ [Hd0 ], |ii ≡ |i1, . . . , iN i ≡ AN [|i1i ⊗ ... ⊗ |iN i], (25) ρ1 ≡ N trN−1[|ΨihΨ|] (19)
where AN is the anti-symmetrizing operator on the by tracing out N −1 particles, is known as the fermionic quan- ⊗N N 0 N−particle Hilbert space H 0 . By expanding |Ψi w.r.t. to tum marginal problem of the setting ∧ [Hd0 ]. Here d ∈ d 0 BN , N ∪ {∞}, Hd0 is the d −dimensional separable 1−particle Hilbert space and we use the trace normalization convention, X |Ψi = ci |ii (26) tr[ρ1] = λ1 + ... + λd0 = N, (20) i common in quantum chemistry. the natural occupation numbers are given by For d0 finite, the family of possible spectra (we call them N 0 X 2 compatible w.r.t ∧ [Hd ]), is described by finitely many in- λk = |ci| . (27) dependent conditions {Ci}, the generalized Pauli constraints. i, k∈i Each of them has the form To compare marginal settings of different dimensions, d, d0 Ci : Di(λ) = κ0 + κ1λ1 + . . . κd0 λd0 ≥ 0, (21) 0 with d < d ≤ ∞ we imbed Hd into Hd0 , d0 κ0, . . . , κd0 ∈ Z and describes a half-space Vi of R . These d d0 span{|ii} ≡ H ≤ H 0 ≡ span{|ii} , (28) constraints together with the trivial conditions (18) and (20) i=1 d d i=1 d0 define the polytope P 0 ⊂ of possible spectra. In that N,d R where the closure is only relevant for the case d0 infinite. In sense every constraint (21) gives rise to a facet F of this poly- i the same way, tope, N N ∧ [H ] ≤ ∧ [H 0 ]. (29) Fi = {λ ∈ PN,d0 | Di(λ) = 0}. (22) d d
Note that besides these Pauli facets there are also further Indeed, according to (26), we find that every state facets, those corresponding to the trivial constraints (18), but X N they will not be of interest in our work. Moreover, the quantity |Ψi = ci |ii ∈ ∧ [Hd] (30) D (·), which is only defined up to a positive factors, defines i 1≤i1<...
B. Truncation of the Spectrum.— Proof. The direction “⇒” was already explained at the end of Section A. To prove “⇐” we show that a state |Ψ0i expanded In our work we have determined the “trajectory” of spectra according to (26),
∞ 0 X N λ(δ) = (λi(δ))i=1 ∈ P3,∞, (33) |Ψ i = ci |ii ∈ ∧ [Hd0 ] , (39) 1≤i <... d : 0 = λk = |ci| . (40) is quasi-pinned to the facet Fi. Since P3,∞ is not explicitly i, k∈i known and quite involved (it is described by infinitely many constraints on infinitely many eigenvalues), we have truncated Hence ci = 0 if iN > d. the spectrum and simplified the pinning analysis by consider- ing only the largest d eigenvalues, What does Lemma 1 imply for the relation between the families of generalized Pauli constraints of two settings? Let tr 0 0 λ = (λ1, . . . , λd), (34) us consider two settings with d, d finite, d < d . Every con- 0 N straint Dj for the setting ∧ [Hd0 ] is linear and hence its re- and analyzed the saturation of the constraints corresponding striction 3 to the setting ∧ [Hd]. The following fact justifies this ap- 0 ˆ 0 0 proach: For d < d every Pauli facet F of PN,d is contained Dj(λ1, . . . , λd) ≡ Dj(λ1, . . . , λd, 0,...) ≥ 0 (41) 0 in some Pauli facet F of PN,d0 , i.e. F is the intersection of 0 F with the hyperplane of spectra with only d non-zero eigen- to the hyperplane defined by 0 = λd+1, λd+2,... is also a lin- tr values. Then, for small λd+1, λd+2,..., small distance of λ ear constraint in the remaining coordinates λ1, . . . , λd. How 0 d to F translates to small distances of λ to F modulo an error is the half space Vj ⊂ R corresponding to (41) related to the of order of the largest neglected eigenvalue, λd+1. To illus- polytope PN,d? Lemma 1 states that 3 trate this, we present the example ∧ [H6], which is one of the two settings studied in our work. There one generalized Pauli PN,d ⊂ Vj (42) constraint reads [4–7] and (6) D (λ) := 2 − (λ1 + λ2 + λ4) ≥ 0 . (35) PN,d = ∩jVj|∗, (43) 3 For the setting ∧ [H∞] the known constraint [4] where the star ∗ denotes here the restriction to spectra, i.e. (∞) D (λ) = 2 − (λ1 + λ2 + λ4 + λ7 + λ11 + λ16 + ...) ≥ 0 , ordered and normalized vectors. There are two possible re- (36) lations between Vj (or Vj|∗) and PN,d. They are illustrated coincides with constraint (35) up to a linear combination of in Figure 2 in form of a simplified 2−dimensional picture: eigenvalues λ7, λ11, λ16,..., which where neglected in the There, we consider two half spaces V1 and V2 correspond- truncated setting. A first important step in proving the universality of this re- lation between polytope distances of correct and truncated set- ting is the next lemma: S S 2 ~ 1 Lemma 1. Consider the quantum marginal problems of the S2 N N 0 two settings ∧ [Hd] and ∧ [Hd0 ], d < d ∈ N ∪ {∞} and let λ = (λ1, . . . , λd) be a spectrum. Then, PN,d N (λ1, . . . , λd) compatible w.r.t. ∧ [Hd] ⇔ N (λ1, . . . , λd, 0,..., 0) compatible w.r.t. ∧ [Hd0 ] (37). | {z } d0−d FIG. 2: Polytope PN,d and two restricted generalized Pauli con- straints Dˆ 1, Dˆ 2 ≥ 0 with boundaries S1,S2 arising from two con- For the corresponding polytopes this means straints D1,D2 ≥ 0 belonging to a higher dimensional marginal N settings ∧ [Hd0 ]. 0 PN,d = PN,d |λd+1,λd+2,...=0, (38) ˆ 0 ˆ 0 the polytope PN,d0 intersected with the hyperplane given by ing to the “restricted” constraints D1 ≥ 0 and D2 ≥ 0 with λd+1, λd+2,... = 0 coincides with PN,d. boundaries S1 and S2 and orientation indicated by stripes. 7
Such hyperplanes can either contain a facet of maximal (ex- stated, which |Ψi in that case satisfies: ample S1) or lower dimension of PN,d or they lie outside of Dˆ|Ψi ≡ (κ Id + κ N + . . . κ N ) |Ψi = 0. (48) PN,d (example S2). The third case of a proper intersection 0 1 1 d d 0 is not possible due to Lemma 1. Every constraint D with Applying this condition to the expansion of |Ψi in Slater de- ˆ 0 boundary S of its restriction D lying outside of PN,d is a terminants induced by the natural orbitals, constraint, which is irrelevant for the pinning analysis since it X has the form |Ψi = ci |ii (49) i D0(λ) = c + D˜(λtr) + O(λ ) , (44) d+1 it implies Klyachko’s selection rule, which states that when- tr N ever where D˜(λ ) ≥ 0 is a constraint of the setting ∧ [Hd] with ˜ a boundary shown in Figure 2 as hyperplane S2 and c > 0 Dˆ|ii= 6 0, (50) is some offset. Hence if the spectrum decays sufficiently fast, constraint D0 is not saturated at all due to the offset c and the corresponding coefficient ci vanishes. To show the thus irrelevant. Moreover, for every Pauli facet of PN,d corre- strength of this selection rule we study states in the Borland- 3 sponding to some constraint D > 0, Lemma 1 guarantees the Dennis setting. The corresponding Hilbert space ∧ [H6] has 0 6 existence of a constraint D > 0 in the larger setting whose dimension 3 = 20 and the generalized Pauli constraints read projection Dˆ0 coincides with D. We summarize these insights [4–7] by stating λ1 + λ6, λ2 + λ5, λ3 + λ4 ≤ 1 (51) N (6) Lemma 2. Given two marginal settings ∧ [Hd] and D := 2 − (λ1 + λ2 + λ4) ≥ 0. (52) N 0 ∧ [Hd0 ] with d < d ∈ N. Every generalized Pauli con- 0 N The normalization together with the non-negativity of the straint D ≥ 0 of the setting ∧ [H 0 ] relevant for the pinning d eigenvalues leads to analysis is given by a linear modification of some generalized N Pauli constraint D ≥ 0 of the setting ∧ [Hd], λ1 + λ6 = λ2 + λ5 = λ3 + λ4 = 1. (53) 0 tr D (λ) = D(λ ) + O(λd+1). (45) Hence the constraints in (51) are always saturated and this implies according to (48) Finally, we remark that for the important case d0 infinite ef- fectively the same results holds but one has to deal with one (Id − N1 − N6) |Ψi = 0 subtlety. Since PN,∞ is described by infinitely many con- (Id − N2 − N5) |Ψi = 0 straints Lemma 1 guarantees for every constraint D ≥ 0 of (Id − N − N ) |Ψi = 0. (54) N 3 4 the setting ∧ [Hd], only the existence of a sequence of con- 0 ˆ 0 Klyachko’s selection rule applied to (54) implies that every straints Dj ≥ 0 whose restrictions Dj ≥ 0 converge to the constraint D ≥ 0. This means that condition (45) in Lemma Slater determinant showing up in the ansatz (49) for |Ψi is 2 holds up to a small error ε, built up by natural orbitals with one index from each set {1, 6}, {2, 5} and {3, 4}. Those are the 8 states |1, 2, 3i, 0 tr Dε(λ) = ε + D(λ ) + O(λd+1), (46) |1, 2, 4i, |1, 3, 5i, |1, 4, 5i, |2, 3, 6i, |2, 4, 6i, |3, 5, 6i and |4, 5, 6i. If the constraint (52) is also saturated the selection which can be made arbitrarily small by choosing appropriate rule restricts this family of Slater determinants to the three 0 constraints Dε. Hence, to minimize the technical effort we states |1, 2, 3i, |1, 4, 5i and |2, 4, 6i and in that case we find assume in our work that Lemma 2 holds in its original form also for the case d0 infinite. |Ψ3i = α|1, 2, 3i + β|1, 4, 5i + γ|2, 4, 6i. (55)
C. Selection Rule.— D. Quasi-Pinning and modified Selection Rule.—
In this section we state a selection rule which explains how In this section we show for the Borland-Dennis setting that N 3 the structure of the N−fermion state |Ψi ∈ ∧ [Hd] simpli- any state |Ψi ∈ ∧ [H6] whose natural occupation numbers fies if the spectrum of the corresponding 1−RDO is pinned are approximately saturating the corresponding generalized to some Pauli facet of PN,d. Moreover, we apply it for the Pauli constraint (52) also fulfill approximately condition (48). 3 setting ∧ [H6]. We also quantify this relation. This result then guarantees that Let’s consider a state |Ψi with natural occupation numbers our Hartree-Fock extension will work for systems exposing d strong pinning. λ = (λi)i=1 saturating some generalized Pauli constraint As a warm-up and since we will need the result we first D(λ) = κ0 + κ1λ1 + . . . κdλd ≥ 0. (47) study a simpler question. It is a basic fact that the spectrum λSl = (1,..., 1, 0,..., 0) can only arise from a Slater deter- In [15], by introducing the creation and annihilation operator minant |Ψi = |1,...,Ni. Is this statement stable under small † ak, ak of a fermion in the natural orbital |ki and the parti- deviations, i.e. λ ≈ λSl ⇒ |Ψi ≈ |1,...,Ni? Yes, it is true † cle number operators Nk ≡ akak, an important condition is according to 8
N d Lemma 3. Consider a state |Ψi ∈ ∧ [Hd], let {|ki}k=1 be Now, we come back to the original question. We first state its natural orbitals and denote the projection operator onto the mathematical result and present the proof afterwards. the space spanned by |1,...,Ni by P . Then, Sl 3 Theorem 4. Given a state |Ψi ∈ ∧ [H6] with natu- 1 6 2 ral occupation numbers (λk)k=1. Let P be the projec- 1 − δ ≤ kP Ψk 2 ≤ 1 − δ, (56) Sl L N tion operator onto the subspace spanned by the states |1, 2, 3i, |1, 4, 5i, |2, 4, 6i, which corresponds to exact pinning where (6) of D = λ5 + λ6 − λ4 ≥ 0 (recall (55)). Then as long as 0 ≤ N − (λ + ... + λ ) =: δ. (57) 1 N 1 δ ≡ 3 − λ − λ − λ ≤ (66) Proof. We expand the state |Ψi in Slater determinants induced 1 2 3 4 by natural orbitals (recall Section A), (which means nothing else but being not too far away from X the spectrum λSl = (1, 1, 1, 0, 0, 0) of a single Slater determi- |Ψi = ci |ii. (58) nant) we find i
We define the operator (6) 2 1 (6) 1 − χδD ≤ kP Ψk2 ≤ 1 − D , (67) 2 Sˆ = N Id − a†a + ... + a† a . (59) 1 1 N N with Since all operators a†a , i = 1, . . . , d commute it is clear that 1 + 2δ i i χ ≡ . (68) Sˆ has the spectrum {0, 1,...,N} with eigenstates |ii. The δ 1 − 4δ eigenvalue corresponding to |ii is the number of indices k ∈ i not belonging to the set {1,...,N}. We denote the set of Proof. In Section C we concluded that |Ψi has the form indices leading to the eigenvalue k by J and find k |Ψi = α|1, 2, 3i + β|1, 2, 4i + γ|1, 3, 5i δ ≡ N − (λ1 + ... + λd) + δ|2, 3, 6i + ν|1, 4, 5i + µ|2, 4, 6i † † + ξ|3, 5, 6i + ζ|4, 5, 6i , (69) = hΨ|N Id − a1a1 + ... + adad |Ψi ˆ 6 ≡ hΨ|S|Ψi with natural orbitals {|ki}k=1. Since the corresponding 6 X ∗ ˆ 1−RDO is diagonal w.r.t. {|ki}k=1, = cj ci hj|S|ii i,j∈J ∪...∪J 0 N hk|ρ1|li = δkl λk, (70) X 2 ˆ = |ci| hi|S|ii. (60) we find (recall (27)) i∈J0∪...∪JN 2 2 2 2 Since for i ∈ Jk, λ4 = |β| + |ν| + |µ| + |ζ| (71) λ = |γ|2 + |ν|2 + |ξ|2 + |ζ|2 (72) hi|Sˆ|ii = k (61) 5 2 2 2 2 λ6 = |δ| + |µ| + |ξ| + |ζ| (73) we find The goal is now to show that the coefficients β, γ, δ, ξ and ζ N N X X 2 X X 2 are small, i.e. δ = |ci| k ≥ |ci| (62)
k=0 i∈Jk k=1 i∈Jk 2 2 2 2 kP ΨkL2 = |α| + |µ| + |ν| and alternatively also = 1 − |β|2 + |γ|2 + |δ|2 + |ξ|2 + |ζ|2 (74)
N X X 2 is close to 1, whenever (52), which here reads δ ≤ N |ci| . (63) (6) 2 2 2 2 2 k=1 i∈Jk D = −|β| + |γ| + |δ| + 2|ξ| + |ζ| , (75) P 2 The normalization of |Ψi, i |ci| = 1 yields (J0 = is approximately saturated. First we observe {(1, 2,...,N)} contains only one element) 2 1 2 2 2 2 2 δ 2 kP ΨkL2 ≤ 1 − |β| + |γ| + |δ| + 2|ξ| + |ζ| ≤ 1 − |c(1,...,N)| ≤ δ. (64) 2 N 1 ≤ 1 − −|β|2 + |γ|2 + |δ|2 + 2|ξ|2 + |ζ|2 and thus 2 1 1 (6) 1 − δ ≤ |c |2 ≤ 1 − δ. (65) = 1 − D , (76) (1,...,N) N 2 2 which is the upper bound for kP ΨkL2 in Theorem 4. 9
To derive the lower bound note the essential difference in By choosing (75) and (76), the sign of the term |β|2. To get rid of this 2 2 2 we write |β| = −χ |β| + (1 + χ)|β| , χ > 0 and estimate 2 2 2 2 2 2 2|α| (1 + χ)|β| in terms of |γ| , |δ| , |ξ| , |ζ| . For this observe r = (80) that (70) in particular implies 4|α|2 − 3 4(1 − |α|2) s = (81) 0 = h4|ρ1|3i = αβ + γν + δµ + ξζ , (77) 4|α|2 − 3 which leads by the triangle inequality, the identity (A + B + 2 2 2 2 2 2 2 2 2 C) ≤ 3 (A + B + C ) and |µ| , |ν| , |ξ| , |ζ| ≤ 1 − |α| the last expression in (79) coincides with D(6) up to a global to factor χ. Both parameters r, s are non-negative as long as 2 3 2 |α| ≥ 4 . Finally, this leads to 2 1 |β| = (γν + δµ + ξζ) α 2 2 2 2 2 2 1 kP Ψk 2 = 1 − (|β| + |γ| + |δ| + |ξ| + |ζ| ) ≤ (|γ| |ν| + |δ| |µ| + |ξ| |ζ|)2 L |α|2 ≥ 1 − (r − 1)D(6) 3 2 2 2 2 2 2 (6) ≤ |γ| |ν| + |δ| |µ| + |ξ| |ζ| ≡ 1 − χ1−|α|2 D , (82) |α|2 3(1 − |α|2) 1 ≤ |γ|2 + |δ|2 + (2|ξ|2 + |ζ|2) (78). with |α|2 3
Now, for all s, r ≥ 0 we find by using (78) 3 − 2|α|2 χ 2 ≡ r − 1 = 1−|α| 4|α|2 − 3 |β|2 + |γ|2 + |δ|2 + |ξ|2 + |ζ|2 1 + 2(1 − |α|2) ≤ (1 − r)|β|2 + |γ|2 + |δ|2 + (1 + s)(2|ξ|2 + |ζ|2) + r|β|2 = . (83) 1 − 4(1 − |α|2) ≤ (1 − r)|β|2 + |γ|2 + |δ|2 + (1 + s)(2|ξ|2 + |ζ|2) 3r(1 − |α|2) 1 + |γ|2 + |δ|2 + (2|ξ|2 + |ζ|2) 2 |α|2 3 Lemma 3 states |α| ≥ 1 − δ and since χ is monotonously increasing, χ1−|α|2 ≤ χδ, which finishes the proof. 3r(1 − |α|2) = (1 − r)|β|2 + 1 + |γ|2 + |δ|2 |α|2 r(1 − |α|2) + 1 + s + 2|ξ|2 + |ζ|2 . (79) |α|2
[1] W. Pauli, Z. Phys. A Hadrons and Nuclei 31, 765 (1925). [12] M. Christandl, B. Doran, S. Kousidis, and M. Walter, [2] P. A. M. Dirac, Proc R. Soc. Lond. A 112, 661 (1926). ArXiv:1204.0741 (2012). [3] W. Heisenberg, Z. Phys. A Hadrons and Nuclei 38, 411 (1926). [13] A. J. Coleman and V. I. Yukalov, Reduced Density Matrices: [4] M. Altunbulak and A. Klyachko, Commun. Math. Phys. 282, Coulsons Challenge (Springer, New York, 2000). 287 (2008). [14] A. J. Coleman, Rev. Mod. Phys. 35, 668 (1963). [5] M. B. Ruskai, Phys. Rev. A 5, 1336 (1972). [15] A. Klyachko, ArXiv:0904.2009 (2009). [6] M. B. Ruskai, Phys. Rev. A 40 (2007). [16] Z. Wang, A. Wang, Y. Yang, and X. Li, ArXiv:1108.1607 [7] R. E. Borland and K. Dennis, J. Phys. B 5, 7 (1972). (2012). [8] A. Klyachko, J. Phys.: Conf. Ser. 36, 72 (2006). [17] R. B. Laughlin, Phys. Rev. Lett. 50, 1395 (1983). [9] A. Klyachko, ArXiv:0409113 (2004). [18] V. Bach, Commun. Math. Phys. 147, 527 (1992). [10] M. Christandl and G. Mitchison, Commun. Math. Phys. 261, [19] J. Eisert and H. Briegel, Phys. Rev. A 64, 022306 (2001). 789 (2006). [20] P. Levay´ and P. Vrana, Phys. Rev. A 78, 022329 (2008). [11] S. Daftuar and P. Hayden, Ann. Phys. 315, 80 (2005).