Towards a formal definition of static and dynamic electronic correlations

Carlos L. Benavides-Riveros,1, ∗ Nektarios N. Lathiotakis,2 and Miguel A. L. Marques1 1Institut für Physik, Martin-Luther-Universität Halle-Wittenberg, 06120 Halle (Saale), Germany 2Theoretical and Physical Chemistry Institute, National Hellenic Research Foundation, GR-11635 Athens, Greece Some of the most spectacular failures of density-functional and Hartree-Fock theories are related to an incorrect description of the so-called static electron correlation. Motivated by recent progress on the N-representability problem of the one-body density matrix for pure states, we propose a way to quantify the static contribution to the electronic correlation. By studying several molecular systems we show that our proposal correlates well with our intuition of static and dynamic electron correlation. Our results bring out the paramount importance of the occupancy of the highest occupied natural spin-orbital in such quantification.

PACS numbers: 31.15.V-, 31.15.xr, 31.70.-f

I. INTRODUCTION (the orthogonally twisted ethylene C2H4 and the methy- lene CH2, for example) and bond stretching in H2O, C2 The concept of correlation and more precisely the idea and N2, are well described by such a method [6,7]. of correlation energy are central in quantum chemistry. In recent years a considerable effort has been devoted Indeed, the electron-correlation problem (or how the dy- to characterize the correlation of a quantum system in namics of each electron is affected by the others) is terms of more meaningful quantities, such as the Slater perhaps the single largest source of error in quantum- rank for two-electron systems [8,9], the entanglement chemical computations [1]. The success of Hartree-Fock classification for the three-fermion case [10], the squared theory in providing a workable upper bound for the Frobenius norm of the cumulant part of the two-particle ground-state energy is largely due to the fact that a reduced density matrix [11] or the comparison with un- single Slater determinant is usually the simplest wave correlated states [12]. Along with Christian Schilling, we function having the correct symmetry properties for a have recently stressed the importance of the energy gap system of fermions. Since the description of interacting in the understanding of the electronic correlations [13]. fermionic systems requires multi-determinantal reference Notwithstanding, these measures do not draw a distinc- wave functions, the correlation energy is commonly de- tion between qualitatively different kinds of electronic fined as the difference between the exact ground-state correlations. In quantum chemistry, for instance, it is and the Hartree-Fock energy [2,3]. Beyond Hartree-Fock customary to distinguish between static (or nondynamic) theory, numerous other methods (such as configuration and dynamic correlations. The former corresponds to interaction or coupled-cluster theory) aim at reconstruct- configurations which are nearly degenerate with respect ing the part of the energy missing from a description to the reference Slater determinant (if any), whilst the based on a single-determinantal wave function. Indeed, latter arises from the need of mixing the Hartree-Fock one common indicator of the accuracy of a model is, by state with higher-order excited states [14, 15]. Heuris- and large, the percentage of the correlation energy it is tically, one usually states that in systems with (strong) able to recover. static correlation the wavefunction differs qualitatively For small molecules, variational methods based on from the reference Slater determinant, while strong dy- configuration interaction techniques describe well elec- namic correlation implies a wavefunction including a tronic correlations. However, due to their extreme com- large number of excited determinants, all with compa- putational cost, configuration-interaction wavefunctions rable, small occupations. Some of the most spectacular are noteworthily difficult to evaluate for larger systems. failures of the Hartree-Fock theory and density functional Among other procedures at hand, the correlation can theory (with standard exchange-correlation functionals) be treated efficiently by applying a Jastrow correlation are related to an incorrect description of static correla- term to an antisymmetrized wave function (e.g. a sin- tion [16].

arXiv:1705.00238v3 [physics.chem-ph] 19 May 2017 gle Slater determinant or an antisymmetrized geminal It is commonly believed that, to a large extent, both power) within quantum Monte Carlo methods [4,5]. The static and dynamic contributions should be included in so-called Jastrow antisymmetric geminal ansatz accounts the global computation of the electronic correlation. Yet for inter-pair interactions and multiple resonance struc- there are few systems for which one can distinguish un- tures, maintaining a polynomial scaling cost, compara- ambiguously between these two types of correlations. For ble to that of the simpler Jastrow single determinant ap- instance, the ground state of helium has no excited elec- proach. Highly correlated systems, as diradical molecules tronic states nearby, leading therefore to the absence of static correlation. In the dissociation limit of H2 a state with fractional occupations arises [17] and the correla- tion is purely static. According to Hollett and Gill [18], ∗ [email protected] static correlation comes in two “flavors”: one that can be 2 captured by breaking the spin symmetry of the Hartree- vided a set of constraints on the natural occupation num- Fock wave function (like in stretched H2) and another bers, stronger than the Pauli principle [23]. Although that cannot. The measures of correlation proposed so far rudimentary schemes to construct such constraints were, purport to include both static and dynamic correlations, to some extent, routine in quantum-chemistry literature although in an uncontrolled manner [19]. [27], it was not only with the work of Klyachko that this For pure quantum states, global structural features of rich structure could be decrypted. The main goal of this the wave function can be abstracted from local infor- section is to review the physical consequences of such a mation alone. Multiparticle entanglement, for instance, generalization. N can be completely classified with the more accesible one- Given an N-fermion state |Ψi ∈ ∧ [H1], with H1 be- particle picture [20]. Such a characterization is addressed ing the one-particle Hilbert space, the natural occupation by a finite set of linear inequalities satisfied by the eigen- numbers are the eigenvalues {ni} and the natural spin- values of the single-particle states [21]. Furthermore, by orbitals are the eigenvectors |ϕii of the one-body reduced using the two-particle density matrix and its deviation density matrix, from idempotency, it is possible to propose a criterion X to distinguish static from dynamic correlation, which for ρˆ1 ≡ NTrN−1[|ΨihΨ|] = ni|ϕiihϕi|. (1) two-fermion systems only requires the occupancies of the i natural orbitals [22]. Needless to say, grasping global in- The natural occupation numbers, arranged in decreasing formation of a many-body quantum system by tackling order ni ≥ ni+1, fulfill the Pauli condition n1 ≤ 1. The only one-particle information is quite remarkable, mainly natural spin-orbitals define an orthonormal basis B1 for because in this way a linear number of degrees of freedom H1 and can also be used to generate an orthonormal basis is required. N BN for the N-fermion Hilbert space HN ≡ ∧ [H1], given Recent progress on the N-representability problem of by the Slater determinants |ϕi . . . ϕi i = |ϕi i ∧ · · · ∧ the one-body reduced density matrix for pure states pro- 1 N 1 |ϕi i. For practical purposes, the dimension of the one- vides an extension of the well-known Pauli exclusion prin- N particle Hilbert space H1 is usually finite. Henceforth, ciple [23]. This extension is important because it pro- HN,d denotes an antisymmetric N-particle Hilbert space vides stringent constraints beyond those from the Pauli with an underlying d-dimensional one-particle Hilbert principle, which can be used, among others, to improve d
space. In principle, the total dimension of HN,d is N , reduced-density-matrix functional theories [24–26]. Our but symmetries usually lower it. main aim in this paper is to employ the generalized Pauli It is by now known that the antisymmetry of N- exclusion principle to establish a general criterion to dis- fermion pure quantum states not only implies the well- tinguish static and dynamic contributions to the elec- known Pauli exclusion principle, which restricts the oc- tronic correlation in fermionic systems. cupation numbers according to [28] 0 ≤ ni ≤ 1, but The paper is organized as follows. For completeness, also entails a set of so-called generalized Pauli constraints SectionII summarizes the key aspects of the so-called [23, 29–31]. These take the form of independent linear generalized Pauli exclusion principle and its potential rel- inequalities evance for quantum chemistry. In SectionIII we discuss a Shull-Löwdin-type functional for three-fermion systems, d 0 X i which can be constructed by using the pertinent gen- Dj(~n) ≡ κj + κjni ≥ 0. (2) eralized Pauli constraints along with the spin symme- i=1 tries. Since this functional depends only on the occupa- i Here the coefficients κj ∈ Z and j = 1, 2, . . . , νN,d < ∞. tion numbers, it is possible to distinguish the correlation Accordingly, for pure states the spectrum of a physical degree of the so-called Borland-Dennis setting (with an fermionic one-body reduced density matrix must satisfy underlining six-dimensional one-particle Hilbert space) a set of independent linear inequalities of the type (2). by using one-particle information alone. In SectionIV The total number of independent inequalities νN,d de- we discuss a formal way to distinguish static from dy- pends on the number of fermions and the dimension of namic correlation. SectionV is devoted to investigate the underlying one-particle Hilbert space. For instance the static and dynamic electronic correlation in molec- [29], ν = 4, ν = 4, ν = 31, ν = 52, ν = 93, ular systems. We compare our results with the well- 3,6 3,7 3,8 3,9 3,10 ν4,8 = 14, ν4,9 = 60, ν4,10 = 125 and ν5,10 = 161. known von-Neumann entanglement entropy. The paper From a geometrical viewpoint, for each fixed pair N ends with a conclusion and an appendix. and d, the family of generalized Pauli constrains, to- gether with the normalization and the ordering condi- tion, forms a “Paulitope”[32]: a polytope PN,d of allowed II. THE GENERALIZATION OF THE PAULI d vectors ~n ≡ (ni)i=1. The physical relevance of this gener- PRINCIPLE alized Pauli exclusion principle has been already stressed, among others, in quantum chemistry [24, 33–43], in open In a groundbreaking work, aimed at solving the quan- quantum systems [44] or in condensed matter [45, 46]. tum marginal problem for pure states, Alexander Kly- The generalized Pauli exclusion principle is partic- achko generalized the Pauli exclusion principle and pro- ularly relevant whenever the natural occupation num- 3 bers of a given system saturate some of the general- III. THE BORLAND-DENNIS SETTING ized Pauli constraints [47]. This so-called “pinning” effect can potentially simplify the complexity of the wave func- A. A Löwdin-Shull functional for three-fermion tion [39]. In fact, whenever a constraint of the sort (2) is systems saturated or pinned (namely, Dj(~n) = 0), any compat- ible N-fermion state |Ψi (with occupation numbers ~n) The famous Borland-Dennis setting H3,6, the rank- belongs to the null eigenspace of the operator six approximation for the three-electron system, is com- pletely characterized by 4 constrains [53]: the equalities Dˆ = κ0 + κ1nˆ + ··· + κdnˆ , (3) j j j 1 j d n1 + n6 = n2 + n5 = n3 + n4 = 1 (5) and the inequality: where nˆi denotes the number operator of the natural or- bital |ϕii of |Ψi. This result not only connects the N- n1 + n2 + n4 ≤ 2. (6) and 1-particle descriptions, which is in itself striking, but provides an important selection rule for the determinants This latter inequality together with the decreasing order- 6 that can appear in the configuration interaction expan- ing rule defines a polytope in R , called here the Borland- sion of the wave function. Indeed, for a given wave func- Dennis Paulitope. Conditions (5) imply that, in the nat- tion |Ψi, whenever Dj(~n) = 0, the Slater determinants ural orbital basis, every Slater determinant, built up from ˆ three natural spin-orbitals, showing up in the configura- for which the relation Dj|ϕi1 . . . ϕiN i = 0 does not hold are not permitted in the configuration expansion of |Ψi. tion expansion (4), satisfies In this way, pinned wave functions undergo an extraor- |ϕ ϕ ϕ i = (ˆn +n ˆ )|ϕ ϕ ϕ i, (7) dinary structural simplification which suggests a natural i j k 7−s s i j k extension of the Hartree-Fock ansatz of the form: for s ∈ {1, 2, 3}. Therefore, each natural spin-orbital belongs to one of three different sets, say ϕi ∈ {ϕ1, ϕ6}, X ϕj ∈ {ϕ2, ϕ5} and ϕk ∈ {ϕ3, ϕ4}. Consequently, the |Ψi = c |ϕ . . . ϕ i. (4) i1,...,iN i1 iN dimension of the total Hilbert space is eight. {i ,...,i }∈I 1 N Dj In the symmetry-adapted description of this system, the spin of three natural orbitals points down, and the spin of the other three points up. The corresponding Here I stands for the family of configurations that may Dj one-body reduced density matrix (a 6 × 6 matrix) is a contribute to the wave function in case of pinning to a block-diagonal matrix that can be written as the direct given generalized Pauli constraint Dj [39]. These remark- sum of two (3 × 3) matrices (say, ρˆ↑ and ρˆ↓), one related able global implications of extremal local information are to the spin up and the other one related to the spin down. stable, i.e. they hold approximately for spectra close to For the doublet configuration, each acceptable Slater de- the boundary of the allowed region [48]. terminant contains two spin orbitals pointing up (for in- These structural simplifications can be used as a vari- stance) and one pointing down. It follows that ational ansatz, whose computational cost is cheaper than configuration interaction or other post-Hartree-Fock vari- Trˆρ↑ = 2 and Trˆρ↓ = 1. (8) ational methods [30, 39, 48]. For a given hamiltonian Hˆ , ˆ To meet the decreasing ordering of the natural occupa- the expectation value of the energy hΨ|H|Ψi is minimized tions as well as the representability conditions (5), two with respect to all states |Ψi of the form (4), i.e., with of the first three occupation numbers must belong to the natural occupation numbers saturating some specific gen- matrix whose trace is equal to two [37]. It is straight- eralized Pauli constraint. For the lithium atom, a wave forward to see that the only admisible set of occupation function with three Slater determinants chosen in this numbers are the ones lying in the hyperplane A : way accounts for more than 87% of the total correlation 1 energy [39]. For harmonium (a system of fermions inter- n1 + n2 + n4 = 2 (9) acting with an external harmonic potential and repelling each other by a Hooke-type force), this method accounts (equivalently, n3 + n5 + n6 = 1), which saturates the for more than 98% of the correlation energy for 3, 4 and generalized Pauli constraint (6), or in the hyperplane A2: 5 fermions [49]. n1 + n2 + n3 = 2 (10) In a nutshell, the main aim of the strategy is to se- lect the most important configurations popping up in an (equivalently, n4 + n5 + n6 = 1). Note that the two hy- 1 efficient configuration interaction computation. We ex- perplanes intersect on the line n3 = n4 = 2 . In Fig.1 pect that these are the first configurations to appear in hyperplanes A1 and A2 are shown within the Pauli hy- approaches whose attempt is also to choose (determinis- percube n1 ≤ 1. tically [50] or stochastically [51, 52]) the most important As stated above, pinning of natural occupation num- Slater determinants. bers undergoes a remarkable structural simplification of 4 the wave functions compatible with ~n. For the case of the Borland-Dennis setting, the equation (9) im- plies for the corresponding wave function the condition (ˆn1 +n ˆ2 +n ˆ4)|Ψi = 2|Ψi, while the constraint (10) im- plies (ˆn1 +n ˆ2 +n ˆ3)|Ψi = 2|Ψi. Consequently, a wave function (the so-called Borland-Dennis state) compatible with the hyperplane A1 can be written in the form [37]: √ √ √ |ΨBDi = n3 |ϕ1ϕ2ϕ3i + n5 |ϕ1ϕ4ϕ5i + n6 |ϕ2ϕ4ϕ6i, (11)

1 where n3 ≥ n5 + n6 and n3 ≥ 2 . A wave function com- patible with the hyperplane A2 reads: √ √ √ |Ψ2i = n4 |ϕ1ϕ2ϕ4i + n5 |ϕ1ϕ3ϕ5i + n6 |ϕ2ϕ3ϕ6i, (12)

1 where n4 ≤ n5 + n6 and n4 ≤ 2 . Notice that, just like in the famous Löwdin-Shull functional for two-fermion systems [54], the wave function is explicitly written in terms of both the natural occupation numbers and the FIG. 1: The hyperplanes 2 = n1 + n2 + n4 and natural orbitals. Likewise, any sign dilemma that may 2 = n1 + n2 + n3, subject to the conditions occur when writing the amplitudes of the states (11) and 1 ≥ n1 ≥ n2 ≥ n3 ≥ 0.5 and 2 ≥ n1 + n2 + n4. The blue (12) can be dodged by absorbing the phase into the spin- dot is the Hartree-Fock point 1, 1, 1
. The red ones are orbitals. Moreover, only doubly excited configurations 1 1
3 3 1
2 2 2
1, 2 , 2 , 4 , 4 , 2 and 3 , 3 , 3 . are permitted here. For |ΨBDi such double excitations are referred to the Slater determinant whose one-particle density matrix is the best idempotent approximation to 2 2 2
• The point Pa ≡ 3 , 3 , 3 , which corresponds to the the true one-particle density matrix. The state |Ψ2i is strongly (static) correlated state: orthogonal to the state |ϕ1ϕ2ϕ3i. Interestingly, a non- vanishing overlap of a wave function with this latter state 1 |Ψai = √ (|ϕ1ϕ2ϕ4i + |ϕ1ϕ3ϕ5i + |ϕ2ϕ3ϕ6i). (13) can only be guaranteed if the sum of the first three nat- 3 ural occupation numbers is larger than two [55]. Both 1 1
|ΨBDi and |Ψ2i lead to diagonal one-particle reduced • The point Pb ≡ 1, 2 , 2 . These occupation num- density matrices. bers correspond to the state For any given Slater determinant, the seniority num- |Ψ i = √1 (|ϕ ϕ ϕ i + |ϕ ϕ ϕ i). (14) ber is defined as the number of orbitals which are singly b 2 1 2 3 1 4 5 occupied. Such an important concept is used in nuclear and condensed matter physics to partition the Hilbert In quantum information theory, this state is said space and construct compact configuration-interaction to be biseparable because one of the particles is wave functions [56, 57]. Since the wave functions (11) and disentangled from the other ones [58]. (12) are eigenfunctions of the spin operators, each Slater 3 3 1
determinant showing up in these expansions is also an • The point Pc ≡ 4 , 4 , 2 , which correspond to the eigenfunction of such operators. The latter is only pos- (static) correlated state: sible if one orbital is doubly occupied and therefore the |Ψ i = √1 |ϕ ϕ ϕ i + 1 (|ϕ ϕ ϕ i + |ϕ ϕ ϕ i). (15) seniority number of each Slater determinant is 1. The c 2 1 2 3 2 1 4 5 2 4 6 seniority number of |ΨBDi and |Ψ2i is also 1. Points Pb and Pc lie in the intersection of A1 and 1 A2, namely, the degeneracy line n3 = n4 = 2 . B. Correlations Since n3 and n4 are identical, the choice of the highest occupied natural orbital |ϕ3i and the low- In Fig.1 we illustrate the discussion of the previous est unoccupied natural orbital |ϕ4i is not unique section, highlighting four special configurations, namely: anymore and the indices 3 and 4 can be swapped in (14) and (15) without changing the spectra.
• The “Hartree-Fock” point (n1, n2, n3) = 1, 1, 1 , which corresponds to the single Slater determinant These four points are important because they belong to |ϕ1ϕ2ϕ3i. Note that it does not coincide in general two different correlation regimes. On the one hand, the with the Hartree-Fock state, since it is described in states |Ψai and |Ψbi exhibit static correlation, as they the natural-orbital basis set. However, we call it so are equiponderant superpositions of Slater determinants. because its spectrum is ~nHF = (1, 1, 1, 0, 0, 0). On the other, the state |Ψci is the superposition of two 5 states (|ϕ1ϕ4ϕ5i and |ϕ2ϕ4ϕ6i) and the nearly degen- C. Borland-Dennis for three active electrons erate |ϕ1ϕ4ϕ5i and its correlation is also static. This is reminiscent of the zero-order description of the beryllium In the configuration interaction picture, the full wave- ground state, for which the 2s and 2p orbitals are nearly function is to be expressed in a given one-electron basis as degenerate and the state is a equiponderant superposi- a linear combination of all possible Slater determinants, tion of three Slater determinants plus a highly weighted save symmetries. In the basis of natural orbitals, it reads: reference state [1]. X |Ψi = ci1...iN |ϕi1 . . . ϕiN i (16)

1≤i1<···

in a similar fashion to the Hartree-Fock ansatz (4). It is well known that the expansion (16) contains a very large number of configurations that are superfluous or negligible for computing molecular electronic properties. In practice, the configurations considered effective are sparse if an arbitrary threshold for the value of the am- plitudes in (16) is enforced [59]. As such, one often in- troduces the notion of active space to select the most relevant configurations at the level of the one-particle pic- ture. A complete active space classifies the one-particle Hilbert space in core (fully occupied), active (partially occupied) and virtual (empty) spin-orbitals. The core spin-orbitals are pinned (completely populated) and are not treated as correlated. Adding active-space con- straints improves the estimate of the ground-state energy in the framework of reduced-density-matrix theory [60]. The generalized Pauli principle can shed some light on this important concept [42, 48]. In fact, for the case of r core (and consequently d − r active orbitals) the Hilbert space HN,d is isomorphic to the wedge product core active Hr,r ∧ HN−r,d−r. Hence, a wave function |Ψi ∈ HN,d can be written in the following way:

active |Ψi = |ϕ1 . . . ϕri ∧ |Ψ i, (17)

active active FIG. 2: Entanglement entropy of the hyperplanes A where |Ψ i ∈ HN−r,d−r. The first r natural occupa- 1 tion numbers are saturated to 1. The remaining d − r and A2. occupation numbers (nr+1, . . . , nd) satisfy a set of gener- alized Pauli constraints and lie therefore inside the poly- tope P . The space Hactive is called here the P N−r,d−r N−r,d−r In Fig.2 the entanglement entropy S = − i ni ln ni is “active Hilbert space”. For instance, for the “Hartree- plotted as a function of n2 and n3 for the hyperplanes A1 Fock” space HN,N , the corresponding zero dimensional and A2. The entanglement entropy of the state |ϕ1ϕ2ϕ3i active active Hilbert space is H0,0 . is zero since it is uncorrelated. The entropies of the states It is possible to characterize a hierarchy of active spaces |Ψbi and |Ψci are 1.3862 and 1.8178, respectively. As active by the effective dimension of HN−r,d−r and the number one might expect, the configurations present in A2 all of Slater determinants appearing in the configuration in- are strongly correlated. For the highest correlated state teraction expansion of |Ψactivei [42, 48]. For the “active” |Ψai, S = 1.9095. The particular structure of the states active Borland-Dennis setting H3,6 we can apply the same |Ψai, |Ψbi and |Ψci prompts us to say that the correlation considerations discussed in the last subsections: if the effects of the states lying in the hyperplane A2 are all due corresponding constraint (6) is saturated, the wave func- to static effects, while the states in A are due to both 1 tion fulfills (ˆnr+1 +ˆnr+2 +ˆnr+4)|Ψi = 2|Ψi, and the set of static and dynamic effects. possible Slater determinants reduces to just three, taking According to the particle-hole symmetry, when applied thus the form: to a three-electron system, the nonzero eigenvalues and active √ √ |Ψ i = nr+3|ϕr+1ϕr+2ϕr+3i + nr+5|ϕr+1ϕr+4ϕr+5i their multiplicities are the same for the one- and the two- √ body reduced matrices. Thus, the results in this section + nr+6|ϕr+2ϕr+4ϕr+6i, (18) based on one-particle information alone are also valid at 1 the level of the 2-particle picture. provided that nr+3 ≥ 2 and nr+3 ≥ nr+5 + nr+6. 6

IV. CORRELATIONS AND CORRELATION In Fig.3 we plot δHF(~n)/2 for the points in the hyper- MEASURES plane A1. As expected, δHF(~nHF) = 0. More interest- ing, the correlation increases monotonically with n3. All Even if the peculiar role played by electronic correla- the points on the degeneracy line n3 = n4 are at the tions in quantum mechanics were noticed from the onset, same l1-distance from the Hartree-Fock point. In effect, the problem of how to measure quantum correlations is δHF(~ns(η))/2 = 1, where still subject to an intense research [61, 62]. The degree 3 3 1 1 1 1 of entanglement D(Ψ) of an arbitrary vector |Ψi can be ~ns(η) ≡ ( 4 + η, 4 − η, 2 , 2 , 4 + η, 4 − η), (22) expressed by its projection onto the nearest normalized 1 unentangled (or uncorrelated) pure state [63, 64]: with 0 ≤ η ≤ 4 , is the set of points lying on the intersec- 1 tion line n3 = n4 = 2 . Moreover, δHF(A2)/2 = 1, for all 2 D(Ψ) = 1 − max |hΨ|Φi| , (19) the points lying on the hyperplane A2. Φ where the maximum is over all unentangled states, nor- malized so that hΦ|Φi = 1. Although this measures sound conventional, it has the merit of being zero when- ever |Ψi is uncorrelated. ˜ 2 This measure (and the minimum minΦ˜∈F ||Ψ − Φ|| , where F denotes the set of unnormalized unentangled (or uncorrelated) pure states[63]) is also important in the realm of quantum chemistry. In the Appendix we state and prove that the set of pure quantum systems with predetermined energy is connected: given a Hamiltonian Hˆ there are two wavefunctions |ψ1i and |ψ2i with ener- 2 gies E1 and E2 whose distance ||ψ1 − ψ2|| is bounded by a function of |E − E | (see Theorem2 in the Ap- 1 2 FIG. 3: l1-distance of the hyperplane A with respect pendix). Recently, we have shown that when |Ψi is the 1 to the Hartree-Fock point ~n . ground state of a given Hamiltonian the measure (19) HF is closely related to the concept of correlation energy as understood in quantum chemistry [13]. A key ingredient in such connection turns out to be the energy gap within the symmetry-adapted Hilbert subspace. B. Static correlation

Roughly speaking, the idea of static correlation is as- A. Dynamic correlation sociated with the presence of a wave function built up from an equiponderant superposition of more than one The N-particle description of a quantum system and Slater determinant, namely, its reduced one-fermion picture can be related in mean- 1 ingful ways. In effect, D(Ψ) can be bounded from above √ (|Φ1i + ··· + |Φmi), (23) and from below by the l1-distance of the natural occupa- m tion numbers. In fact, the distance between a wave func- For the Borland-Dennis setting H3,6, the hyperplane A2 tion |Ψi and any Slater determinant |ϕi1 . . . ϕiN i satisfies [30]: contains states with three configurations being almost equiponderant. After this long discussion, it is natural δ (~n) δ (~n) to define all the points lying on the hyperplane A2 as i ≤ 1 − |hϕ . . . ϕ |Ψi|2 ≤ i , (20) 2 min(N, d − N) i1 iN 2 statically correlated. The hyperplane A1 contains the uncorrelated Hartree-Fock state and the correlation of where d is the dimension of the underlying one-particle the rest of the states present is due to static as well as P dynamic effects. Hilbert space, as defined in Sec.II, and δi(~n) ≡ i∈i(1− P 1 The “static” states |Ψs(η)i that lead to the occupancies ni) + i/∈i ni is the l -distance between ~n (the natural occupation numbers of |Ψi) and the natural occupation ~ns(η) as defined in Eq. (22) read: numbers of the Slater determinant in display (here i ≡ q q √1 1 1 {i1, . . . , iN }). This result is also valid for Hartree-Fock or |Ψs(η)i ≡ |ϕ1ϕ2ϕ3i+ 4 + η|ϕ1ϕ4ϕ5i+ 4 − η|ϕ2ϕ4ϕ6i, 1 2 Brueckner orbitals [13, 65]. In particular, the l -distance (24) to the Hartree-Fock point is given by: 1 where 0 ≤ η ≤ 4 . As stated before, since n3 and n4 X X are identical, the choice of the highest occupied natural δHF(~n) = (1 − ni) + ni. (21) orbital and the lowest unoccupied natural orbital is not i≤N i>N unique and the indices 3 and 4 can be swapped without 7 changing the spectra. However, by doing so the resulting The distance to the Hartree-Fock point δHF(~n) can be state is orthogonal to |ϕ1ϕ2ϕ3i. viewed as a measure of the dynamic part of the electronic The L2-distance (Eq. (19)) between the Borland- correlation, for it quantifies how much a wave function Dennis state (11) and |Ψs(η)i is given by Js(η) ≡ 1 − differs from the uncorrelated Hartree-Fock state. The 2 |hΨBD|Ψs(η)i| . The minimum of this distance depends static distance δs(~n) can be viewed as a measure of the only on the value of the natural occupation number cor- static part of the correlation, as it quantifies how much responding to the highest occupied natural orbital. To a wave function differs from the set of static states. One see this notice that the minimum of Js(η) is attained expects for helium δHF(~nHe) ≈ 0 while for H2 at infinite ∗ n5−n6 when dJs(η)/dη = 0, which happens at η = . separation δs(~nH ) = 0. For convenience we renormal- 4(1−n3) 2,∞ The distance is therefore ize these two l1-distances by means of

√ 2   δHF(~n) δs(~n) ∗ n3 1 − n2 n2 − n3 P (~n) = and P (~n) = . Js(η ) = 1 − √ + + sta dyn 2 p2(1 − n ) p2(1 − n ) δHF(~n) + δs(~n) δHF(~n) + δs(~n) 3 3 (28) 1 √ √ 2 1 p = 1 − 2 ( n3 + 1 − n3) = 2 − n3(1 − n3), (25) Since the measures Pdyn(~n) and Psta(~n) are normal- ized (while δHF(~n) and δs(~n) are not), they are much which is zero when the correlation of the Borland-Dennis more useful to compare different systems. In this way, 1 state is completely static and is 2 when the state is the when the correlation is due to static effects Psta(~n) = 1 uncorrelated Hartree-Fock state. and Pdyn(~n) = 0, while the contrary occurs for a com- We can also investigate the l1-distance between any pletely dynamic state. For instance, one expects for state ~n ∈ A1 and the state ~ns(η) (22), namely: He Psta(~nHe) ≈ 0 while for H2 at infinite separation

Pdyn(~nH2,∞ ) = 0. These quantities have the merit of δs(~n) = min dist1(~n,~ns(η)). (26) η being zero or one when the correlation is completely dy- namic or completely static and hence they separate the Notice that the l1-distance reads correlation in two contributions. It is worth saying that our considerations do not only 3 3 1 dist1(~n,~ns(η)) = 2(|n1 −( 4 +η)|+|n2 −( 4 −η)|+|n3 − 2 |). apply for the Borland-Dennis setting but also for larger ones. Recall that for the settings HN,d and HN,d0 , such So, to minimize dist (~n,~n (η)) is the same as minimizing 0 1 s that d < d ∈ N, the corresponding polytopes satisfy: |n − 3 −η|+|n − 3 +η|. Since the l1-sphere with radius 1 4 2 4 0 PN,d = PN,d |nd+1=···=nd0 =0. It means that, intersected z centered at ~n is the convex hull of the vertices ((n1 ± with the hyperplane given by nd+1 = ··· = nd0 = 0, z, n2, n3), (n1, n2 ± z, n3), (n1, n2, n3 ± z)), the minimum the polytope PN,d0 coincides with PN,d [30]. Therefore, of the distance is: we have completely characterized the static states up to 3 1 1 six dimensional one-particle Hilbert spaces. By choosing δs(~n) = 2(|n1 − (−n2 + 2 )| + |n3 − 2 |) = 4(n3 − 2 ), (27) n1 = 1, we freeze one electron, we are effectively deal- ing with a two active-electron system and our measures which depends on n3 alone. Notice that for the Hartree- can also be used for characterizing the correlation of two- Fock point δs(~nHF)/2 = 1 and remember that by our electron systems. Note that for this latter case the nat- ∗ definition δs(~n) = 0 if ~n ∈ A2. Remarkably, η , the ural occupation numbers are evenly degenerated, a very minimizer of Js(η), is also a minimizer of dist1(~n,~ns(η)). well known representability condition for systems with This latter statement can be proved by noting that the a even number of electrons and time-reversal symmetry first occupation number of ~ns(η∗) satisfies [66].

∗ 3 n1 − n2 2 − 4n3 + 2n1 n1 n1(η ) = + = ≤ ≤ n1, 4 4(1 − n3) 4(1 − n3) 2(1 − n3) V. CORRELATION IN MOLECULAR SYSTEMS

1 ∗ since n3 ≥ 2 . Equivalently, n2(η ) ≤ n2. Therefore: To illustrate these concepts we plot in Fig.4 our mea- sures of static and dynamic correlation for the ground dist (~n,~n (η∗)) = 2[(n − 3 − η ) + (n − 3 + η∗) + (n − 1 )] 1 s 1 4 ∗ 2 4 3 2 states of the diatomic molecules H2 and Li2 as a func- 3 1 1 tion of the interatomic distance. In the same plot we can = 2[(n1 + n2 − 2 ) + (n3 − 2 )] = 4(n3 − 2 ), also see the von-Neumann entropy and the value of the which is the minimum (27). highest occupancy of the highest occupied natural spin- orbital. These numbers were obtained from CAS-SCF calculation using the code Gamess [67] and cc-pVTZ ba- C. Correlation measures sis sets with all electrons active and as large as possible active space of orbitals. The comparison of the distances (21) and (26) allows us The molecule H2, and in particular its dissociation to distinguish between dynamic and static correlations. limit, is the quintessential example of static correla- 8

(a) H2. (a) H3.

(b) Li2. (b) Three-fermion three-site Hubbard model.

FIG. 4: Correlation curves for H2 and Li2. Psta and FIG. 5: Correlation curves for H3 and the three-fermion Pdyn as well as the occupancy of the highest occupied three-site Hubbard model. Psta and Pdyn as well as the natural spin-orbital and the von Neumann entropy S occupancy of the highest occupied natural spin-orbital are plotted as functions of the interatomic distance (in and the von Neumann entropy S are plotted as Å). The equilibrium bonding length is 0.74 Å for H2 functions of the interatomic distance (in Å) or the and 2.67 Å for Li2. coupling strength.

imum. Beyond that value, the static correlation grows tion [68]. It is well known that the restricted Hartree- and the dynamic correlation decreases slowly. This be- Fock approach describes very well the equilibrium chem- haviour changes around 4 Å, where the static correlation ical bond, but fails dramatically as the molecule is speeds up. stretched. Around the equilibrium separation, Psta is In Fig.5 we plot the correlation measures for three- close to zero and Pdyn reaches its maximum. There is a electron systems: the ground state of the equilateral H3 change of regime around 1.5 Å because the static correla- and the three-site three-fermion Hubbard model, which tion begins to grow rapidly. Beyond this point, restricted is very well known for it is analytically solvable [13, 45]. Hartree-Fock theory is unable to predict a bound system The Hamiltonian (in second quantization) of the one- anymore. At the dissociation limit, the correlation is due dimensional r-site Hubbard model reads: to static effects only, as expected. Both measures allow t X X us to observe the smooth increasing of static effects when Hˆ = − (c† c + h.c.) + 2U nˆ nˆ , (29) 2 iσ (i+1)σ i↑ i↓ the molecule is elongated. A different situation can be i,σ i observed for the diatomic Li2. For lengths smaller than † the bond length the static correlation decreases as the i ∈ {1, 2, . . . , r}, where ciσ and ciσ are the fermionic cre- distance increases. The energy, the static correlation and ation and annihilation operators for a particle on the site † the von-Neumann entropy reach their minimum around i with spin σ ∈ {↑, ↓} and nˆiσ = ciσciσ. The first term 2.9 Å, very close to the equilibrium bonding length, while in Eq. (29) describes the hopping between two neighbor- the dynamic correlation as well as the occupancy of the ing sites while the second represents the on-site interac- highest occupied natural spin-orbital acquired their max- tion. Periodic boundary conditions for the case r = 3 9 are also assumed. Achieved experimentally very recently 0.25 with full control over the quantum state [69], this model |E | may be considered as a simplified tight-binding descrip- corr 0.20 Energy gap tion of the Hr molecule. For the case of H3 the correla- tion measures are plotted as a function of the interatomic distance (in Å) and for the Hubbard model as a function 0.15 of the coupling U/t. In both cases, as the molecule is elongated or the interaction in the Hubbard model is en- hanced, the energy gap (the energy difference between 0.10 the first-excited and the ground states) shortens and the Energy (Hartree) electronic correlation increases, leading to the appear- 0.05 ance of static effects [13]. While H3 exhibits a behaviour essentially similar to H2, the Hubbard model shows off 0.00 two different regimes of correlation. For positive values 0 10 20 30 40 50 60 70 80 90 of the relative coupling, the static correlation plays a Torsional Angle (Degrees) prominent role. In particular, beyond U/t = 3.2147, the system lies in the hyperplane A2 of the Borland-Dennis FIG. 6: Correlation energy and energy gap for the setting (10) and the correlation is completely static. For twisted ethylene C2H4 as a function of the torsion angle this strongly correlated regime, the ground state can be around the C=C double bond. written as a equiponderant superposition of three Slater determinants. For negative values of the coupling there is always a fraction of the correlation due to dynamic effects. In that limit the ground state is written as a superposition of two Slater determinants with different amplitudes. It is known that static and dynamic electronic corre- lations play a prominent role in the orthogonally twisted ethylene [6]. In fact, the energy gap between the ground and the first excited state shortens when the torsion an- gle around the C=C double bond is increased. While the ground state of the planar ethylene is very well de- scribed by a single Slater determinant, at ninety degrees at least two Slater determinants are needed, resembling the dihydrogen in the dissociation limit. In Fig.6 we plot the correlation energy and the energy gap of ethylene as a function of the torsion angle, using CAS-SCF(12,12) FIG. 7: Correlation curves for ethylene. Psta and Pdyn method and a cc-pVDZ basis set. In Fig.7 we plot as well as the occupancy of the highest occupied natural the correlation measures for ethylene along the torsional spin-orbital and the von Neumann entropy S are path. For the planar geometry the correlation is almost plotted as functions of the torsion angle around the completely dynamic and the situation remains in this way C=C double bond. until the torsional degree reaches 60o. From this angle on the static correlation shows up. At 80o the static and dynamic correlation are equally important in the to- paper, we have proposed a general criterion to distinguish tal electron correlation of ethylene. When orthogonally the static and dynamic parts of the electronic correla- twisted, the correlation of ethylene is 90% due to static ef- tion in fermionic systems, by tackling only one-particle fects and there is still an important part due to dynamic information. By doing so, we provided two kinds of l1- correlation. Remarkably, the rise of static correlation distances: (a) to the Hartree-Fock point, which can be around 60o coincides with the increase of correlation en- viewed as a measure of the dynamic part of the electronic ergy. From this perspective, the gain of total correlation correlation, and (b) to the static states, which can be is mainly due to static effects. viewed as a measure of the static part of the correlation. We gave some examples of physical systems and showed that these correlation measures correlate well with our VI. SUMMARY AND CONCLUSION intuition of static and dynamic correlation. Though we focused our attention on two and three Thanks to the generalization of the Pauli exclusion “active”-fermion systems, the results can in principle be principle, it is possible to relate equiponderant superpo- generalized to larger settings. So far, the complete set sitions of Slater determinants to certain sets of fermionic of generalized Pauli constraints is only known for small occupation numbers lying inside the Paulitope. In this systems with three, four and five particles. There is, how- 10 ever, an algorithm which provide in principle the repre- Theorem 2. Let Hˆ be a Hamiltonian on H with a unique sentability conditions for larger settings [23]. ground-state. Let SE be the set of pure states with ex- In this paper we have highlighted the paramount im- pected energy E: SE = {ψ ∈ H|hψ|Hˆ |ψi = E}. If portance of the occupancy of the highest occupied natu- |E1 − E2| < , then SE2 has an element close to an ele- ral spin-orbital in the understanding of the static correla- ment of SE1 . tion. The quantities we proposed in this paper can allow us to construct reliable ways to separate dynamic and Proof. Let us write the spectral decomposition of the ˆ P static correlations and, more important, to better under- Hamiltonian in the following way: H = i ei|φiihφi|, with e < e ≤ e ≤ · · · . A wavefunction |ψ i ∈ S can stand the qualitative nature of the correlation present in 0 1 2 1 E1 ˆ P real physical and chemical electronic systems. They can be written in the eigenbasis of H as |ψ1i = i ai|φii. also be a tool for analysing the failures of quantum many If E2 = E1 there is nothing to prove. Without loss of body theories (like density functional theory) [70]. Re- generality, let us assume that E2 < E1, take 0 ≤ a0 ≤ 1 cent progress in fermionic mode entanglement can also and choose a state |ψ2i in SE2 as a superposition of |ψ1i shed more light in these directions [71, 72]. and the ground state |φ0i, namely: |ψ2i = α|ψ1i + β|φ0i with α and β positive real numbers smaller than 1. Nor- 2 2 malization dictates that α + β + 2αβa0 = 1 and the 2 2 ACKNOWLEDGEMENT energy constraint reads α E1 + β e0 + 2αβa0e0 = E2. It is easy to see that both conditions translate into: α = p 2 2 2 1/2 We thank D. Gross, C. Schilling and M. Springborg (E2 − e0)/(E1 − e0) and β = [α a0 +(1−α )] −αa0. for helpful discussions. We acknowledge financial sup- Using the fact that 0 ≤ α, β, a0 ≤ 1 and Lemma1 one port from the GSRT of the Hellenic Ministry of Edu- obtains (for E1 ≤ e1): cation (ESPA), through “Advanced Materials and De- q q 2 2 2 2 2 2 2 vices” program (MIS:5002409) (N.N.L.) and the DFG a0β = a0 α a0 + (1 − α ) − αa0 ≥ a0 α a0 + (1 − α ) − αa0 through Projects No. SFB-762 and No. MA 6787/1-1 r r r ! (M.A.L.M.). e1 − E1 e1 − E2 E2 − e0 ≥ − ≡ f(e0, e1,E1,E2). e1 − e0 e1 − e0 E1 − e0 Appendix Now we can compare the states |ψ1i and |ψ2i: Lemma 1. Let Hˆ be a Hamiltonian on the Hilbert space r 2 E1 − E2 H with a unique ground state |φ0i with energy e0 and an ||ψ1 − ψ2|| ≤ 2 − 2θ(e1 − E1)g(e0, e1,E1,E2), E1 − e0 energy gap egap = e1 − e0, where e1 is the energy of the first excited state. Then, for any |ψi ∈ H with energy where the Heaviside function reads θ(x) = 0 if x < 0 and 2 Eψ = hψ|Hˆ |ψi we have [13]: |hφ0|ψi| ≥ (e1 − Eψ)/egap. θ(x) = 1 if x ≥ 0 and g(·) = max(0, f(·)).

[1] D. P. Tew, W. Klopper, and T. Helgaker, J. Comput. [14] A. D. Becke, J. Chem. Phys. 138, 074109 (2013). Chem. 28, 1307 (2007). [15] P. Ziesche, O. Gunnarsson, W. John, and H. Beck, Phys. [2] E. Wigner, Phys. Rev. 46, 1002 (1934). Rev. B 55, 10270 (1997). [3] P.-O. Löwdin, Phys. Rev. 97, 1509 (1955). [16] A. J. Cohen, P. Mori-Sánchez, and W. Yang, Science [4] M. Casula and S. Sorella, J. Chem. Phys. 119, 6500 321, 792 (2008). (2003). [17] A. J. Cohen, P. Mori-Sánchez, and W. Yang, J. Chem. [5] M. Casula, C. Attaccalite, and S. Sorella, J. Chem. Phys. Phys. 129, 121104 (2008). 121, 7110 (2004). [18] J. W. Hollett and P. M. W. Gill, J. Chem. Phys. 134, [6] A. Zen, E. Coccia, Y. Luo, S. Sorella, and L. Guidoni, 114111 (2011). J. Chem. Theory Comput. 10, 1048 (2014). [19] J. W. Hollett, H. Hosseini, and C. Menzies, J. Chem. [7] E. Neuscamman, J. Chem. Phys. 139, 194105 (2013). Phys. 145, 084106 (2016). [8] J. Schliemann, J. I. Cirac, M. Kuś, M. Lewenstein, and [20] M. Walter, B. Doran, D. Gross, and M. Christandl, Sci- D. Loss, Phys. Rev. A 64, 022303 (2001). ence 340, 1205 (2013). [9] A. R. Plastino, D. Manzano, and J. S. Dehesa, EPL 86, [21] A. Sawicki, M. Walter, and M. Kuś, J. Phys. A 46, 20005 (2009). 055304 (2013). [10] G. Sárosi and P. Lévay, Phys. Rev. A 89, 042310 (2014). [22] E. Ramos-Cordoba, P. Salvador, and E. Matito, Phys. [11] T. Juhász and D. A. Mazziotti, J. Chem. Phys. 125, Chem. Chem. Phys. 18, 24015 (2016). 174105 (2006). [23] A. Klyachko, J. Phys. 36, 72 (2006). [12] A. D. Gottlieb and N. J. Mauser, Phys. Rev. Lett. 95, [24] I. Theophilou, N. Lathiotakis, M. Marques, and N. Hel- 123003 (2005). big, J. Chem. Phys. 142, 154108 (2015). [13] C. L. Benavides-Riveros, N. N. Lathiotakis, C. Schilling, [25] D. A. Mazziotti, Phys. Rev. A 94, 032516 (2016). and M. A. L. Marques, Phys. Rev. A 95, 032507 (2017). [26] A. E. DePrince, J. Chem. Phys. 145, 164109 (2016). 11

[27] C. W. Müller, J. Phys. A 32, 4139 (1999). [50] D. Cleland, G. H. Booth, and A. Alavi, J. Chem. Phys. [28] A. J. Coleman, Rev. Mod. Phys. 35, 668 (1963). 132, 041103 (2010). [29] M. Altunbulak and A. Klyachko, Commun. Math. Phys. [51] E. Giner, A. Scemama, and M. Caffarel, Can. J. Chem. 282, 287 (2008). 91, 879 (2013). [30] C. Schilling, D. Gross, and M. Christandl, Phys. Rev. [52] M. Caffarel, T. Applencourt, E. Giner, and A. Sce- Lett. 110, 040404 (2013). mama, “Recent progress in quantum monte carlo,” in Us- [31] C. Schilling, in Mathematical Results in Quantum Me- ing CIPSI Nodes in Diffusion Monte Carlo (ACS, 2016) chanics, edited by P. Exner, W. König, and H. Neidhardt Chap. 2, pp. 15–46. (World Scientific, 2014) Chap. 10, pp. 165–176. [53] R. Borland and K. Dennis, J. Phys. B 5, 7 (1972). [32] Norbert Mauser coined the term “Paulitope” during the [54] P.-O. Löwdin and H. Shull, Phys. Rev. 101, 1730 (1956). Workshop Generalized Pauli Constraints and Fermion [55] W. Kutzelnigg and V. H. Smith, Int. J. Quant. Chem. 2, Correlation, celebrated at the Wolfgang Pauli Institute 531 (1968). in Vienna in August 2016. [56] L. Bytautas, T. M. Henderson, C. A. Jiménez-Hoyos, [33] C. Schilling, Phys. Rev. A 91, 022105 (2015). J. K. Ellis, and G. E. Scuseria, J. Chem. Phys. 135, [34] C. L. Benavides-Riveros, J. M. Gracia-Bondia, and 044119 (2011). M. Springborg, Phys. Rev. A 88, 022508 (2013). [57] L. Bytautas, G. E. Scuseria, and K. Ruedenberg,J. [35] R. Chakraborty and D. Mazziotti, Phys. Rev. A 89, Chem. Phys. 143, 094105 (2015). 042505 (2014). [58] P. Lévay and P. Vrana, Phys. Rev. A 78, 022329 (2008). [36] C. L. Benavides-Riveros, J. M. Gracia-Bondía, and [59] L. M. Mentel, R. van Meer, O. V. Gritsenko, and E. J. M. Springborg, arXiv:1409.6435 (2014). Baerends, J. Chem. Phys. 140, 214105 (2014). [37] C. L. Benavides-Riveros and M. Springborg, Phys. Rev. [60] N. Shenvi and A. F. Izmaylov, Phys. Rev. Lett. 105, A 92, 012512 (2015). 213003 (2010). [38] R. Chakraborty and D. A. Mazziotti, Int. J. Quantum [61] R. Horodecki, P. Horodecki, M. Horodecki, and Chem. 115, 1305 (2015). K. Horodecki, Rev. Mod. Phys. 81, 865 (2009). [39] C. L. Benavides-Riveros and C. Schilling, Z. Phys. Chem. [62] S. Szalay, Phys. Rev. A 92, 042329 (2015). 230, 703 (2016). [63] A. Shimony, Ann. N. Y. Acad. Sci. 755, 675 (1995). [40] F. Tennie, V. Vedral, and C. Schilling, Phys. Rev. A 94, [64] J. M. Myers and T. T. Wu, Quantum Inf. Process. 9, 239 012120 (2016). (2010). [41] F. Tennie, D. Ebler, V. Vedral, and C. Schilling, Phys. [65] J. M. Zhang and M. Kollar, Phys. Rev. A 89, 012504 Rev. A 93, 042126 (2016). (2014). [42] F. Tennie, V. Vedral, and C. Schilling, Phys. Rev. A 95, [66] D. W. Smith, Phys. Rev. 147, 896 (1966). 022336 (2017). [67] M. W. Schmidt, K. K. Baldridge, J. A. Boatz, S. T. El- [43] Y. Wang, J. Wang, and H. Lischka, Int. J. Quantum bert, M. S. Gordon, J. H. Jensen, S. Koseki, N. Mat- Chem. , e25376 (2017), e25376. sunaga, K. A. Nguyen, S. Su, T. L. Windus, M. Dupuis, [44] R. Chakraborty and D. Mazziotti, Phys. Rev. A 91, and J. A. Montgomery, J. Comput. Chem. 14, 1347 010101 (2015). (1993). [45] C. Schilling, Phys. Rev. B 92, 155149 (2015). [68] C. A. Coulson and I. Fischer, Philos. Mag. 40, 386 (1949). [46] C. Schilling, Quantum marginal problem and its physical [69] S. Murmann, A. Bergschneider, V. M. Klinkhamer, relevance, Ph.D. thesis, ETH-Zürich (2014). G. Zürn, T. Lompe, and S. Jochim, Phys. Rev. Lett. [47] A. Klyachko, arXiv:0904.2009 (2009). 114, 080402 (2015). [48] C. Schilling, C. L. Benavides-Riveros, and P. Vrana, [70] J.-D. Chai, J. Chem. Phys. 136, 154104 (2012). arXiv:1703.01612 (2017). [71] K. Boguslawski, P. Tecmer, G. Barcza, Ö. Legeza, and [49] C. L. Benavides-Riveros, Disentangling the marginal M. Reiher, J. Chem. Theory Comput. 9, 2959 (2013). problem in quantum chemistry, Ph.D. thesis, Universidad [72] N. Friis, New J. Phys. 18, 033014 (2016). de Zaragoza (2015).