betaFeedback

WorksPhysics and Astronomy

Energies and spectra of solids from the algorithmic inversion of dynamical Hubbard functionals

Abstract

Energy functionals of the Green's function can simultaneously provide spectral and thermodynamic properties of interacting electrons' systems. Though powerful in principle, these formulations need to deal with dynamical (frequency-dependent) quantities, increasing the algorithmic and numerical complexity and limiting applications. We first show that, when representing all frequency-dependent propagators as sums over poles, the typical operations of dynamical formulations become closed (i.e., all quantities are expressed as sums over poles) and analytical. Importantly, we map the Dyson equation into a nonlinear eigenvalue problem that can be solved exactly; this is achieved by introducing a fictitious non-interacting system with additional degrees of freedom which shares, upon projection, the same Green's function of the real system. Last, we introduce an approximation to the exchange-correlation part of the Klein functional adopting a localized GWGW approach; this is a generalization of the static Hubbard extension of density-functional theory with a dynamical screened potential U(ω)U(ω). We showcase the algorithmic efficiency of the methods, and the physical accuracy of the functional, by computing the spectral, thermodynamic, and vibrational properties of SrVO3_3, finding results in close agreement with experiments and state-of-the-art methods, at highly reduced computational costs and with a transparent physical interpretation.

Accurate and predictive first-principles calculations of materials properties have been and remain a challenging task for scientific discoveries and technological innovation [1]. Though density-functional theory (DFT) has provided a major step forward in the prediction of ground-state properties [2, 3], addressing spectroscopic quantities remains challenging [4, 5, 6, 7]. To overcome this limitation, dynamical methods, mostly based on Green’s functions, have been used; these include many-body perturbation theory (MBPT) approaches such as GWGW [5, 8], dynamical mean-field theory [7, 9] (DMFT), spectral functional theories [6, 10, 11] (SFT), and electron-boson interaction schemes [12, 13, 14]; in all of these approaches, dynamical (frequency-dependent) self-energies or potentials arise.

In Green’s function theories, the Luttinger-Ward (LW) and Klein functionals [15, 16, 17, 18, 19] are energy functionals of the Green’s function that are variational and yield conserving potentials that are dynamical. Though explicitly known diagrammatically, the exact exchange-correlation term Φxc\Phi_{\mathrm{xc}} of the functional is computationally inaccessible and needs to be approximated [20, 21]. The choice of the approximation for Φxc\Phi_{\mathrm{xc}} determines the physics accessible to the functional, ranging from long-range plasmonic effects, as in GWGW [5, 8], to strong correlations as in DMFT [7, 9]. The stationarity condition of the functional yields the Dyson equation involving the interacting propagator GG and the dynamical self-energy (as a derivative of Φxc\Phi_{\mathrm{xc}} with respect to GG) [18, 19]. Therefore, the functional and its derivative determine the thermodynamics and the spectral properties of a material, also allowing one to compute, at self-consistency, ground-state quantities such as forces and phonons through the Hellmann-Feynman theorem [22].

Due to the presence of dynamical quantities, applications where both spectroscopic and thermodynamic quantities are computed together are limited (see Ref. [23] and references therein). In the context of GWGW, such calculations have been performed for model systems, such as the homogeneous electron gas [24] or Hubbard chains [25, 26, 27], and later addressed solids, typically using an imaginary-axis formalism [28, 29, 30, 31]. This latter approach has been proven very effective for the prediction of ground-state properties but retains limited accuracy for spectral properties, due to the challenges of performing the analytic continuation to the real axis [32]. Similarly, in DMFT, it is possible to calculate different ground-state properties accurately [33, 34], while spectral properties still need to deal with analytic continuation.

In this work, we propose a framework to enable the calculation of accurate spectral and thermodynamic properties on the same footing, and apply it to study SrVO3_3. First, we generalize to the non-homogeneous case the algorithmic-inversion method on sum over poles (AIM-SOP) [23] to tackle dynamical formulations (e.g., MBPT or DMFT) for condensed-matter applications; here, the dynamical (frequency-dependent) propagators are represented as sum over poles (SOP) and defined on the entire complex plane, thus avoiding analytic continuation (for clarity, by SOP we imply a sum over first-order poles). Second, we introduce a dynamical Hubbard functional that generalizes the DFT+U energy functional of Dudarev et al. [35] to host a dynamical screened potential U(ω)U(\omega), rather than a static UU, and that can be applied to solids with localized dd or ff frontier electrons. Third, we use AIM-SOP to implement the dynamical Hubbard functional and obtain the spectral and thermodynamic properties of SrVO3_3, a paradigmatic correlated metal. We find very good agreement with experiments and state-of-the-art methods, such as GW+GW+DMFT, for the spectrum, bulk modulus, and phonons, at a greatly reduced computational cost compared to established approaches.

Algorithmic-inversion method on sum over poles. AIM-SOP provides a theoretical and computational framework aimed to deal with dynamical quantities, such as Green’s functions (GF) and self-energies (SE), and which is closed for common operations appearing in many-body perturbation theory (MBPT), i.e., all quantities are expressed as sums over poles. Starting from our previous work [23], here we generalize the framework to the operatorial case to address realistic materials. AIM-SOP is based on the representation of dynamical (frequency-dependent) propagators as sum over poles:

Σ(r,r′,ω)=∑i=1NΓi(r,r′)ω−Ωi+Σ0(r,r′),(1)\Sigma(\mathbf{r},\mathbf{r}',\omega)=\sum_{i=1}^{N}\frac{\Gamma_i(\mathbf{r},\mathbf{r}')}{\omega-\Omega_i}+\Sigma_0(\mathbf{r},\mathbf{r}'), \tag*{(1)}

where Σ\Sigma may be a generic propagator (not only a SE, but also e.g. the GF or the polarizability), Γi\Gamma_i are operatorial residues (we omit spacial and orbital indexes henceforth), Ωi\Omega_i are scalar poles, and Σ0\Sigma_0 is a static operator (for example the Hartree-Fock term). In general, all time-ordered (TO) operators can be expanded as SOP [36]. It is evident that the sum of two operators on SOP is a SOP, and the same holds for multiplication and convolutions [23]. In addition, and crucially, the Dyson equation G(ω)=[ωI−h0−Σ(ω)]−1G(\omega)=[\omega I-h_0-\Sigma(\omega)]^{-1} is closed on SOP, i.e., a self-energy on SOP yields a GF on SOP.

To demonstrate this, we start by observing that the frequency-wise inversion of the Dyson equation can be exactly mapped into the solution of the nonlinear eigenvalue problem:

[h0+Σ(ω)]∣ψ⟩=ω∣ψ⟩,(2)[h_0+\Sigma(\omega)]|\psi\rangle=\omega|\psi\rangle, \tag*{(2)}

where the nonlinear eigenvalues are the poles of the Green’s function and the eigenvectors determine its residues. In fact, as derived in the Supplemental Material [37] (and in Ref. [38] within the mathematical framework of NLEPs), if the self-energy can be represented as a sum over poles (i.e., it is rational on the whole complex plane), the resulting GF is also a SOP:

G(ω)=∑s∣ψsr⟩⟨ψsl∣ω−zs,(3)G(\omega)=\sum_s\frac{|\psi_s^r\rangle\langle\psi_s^l|}{\omega-z_s}, \tag*{(3)}

where (zs,ψsr)(z_s,\psi_s^r) are the nonlinear eigenvalues and (right) eigenvectors of Eq. (2). Of course, when dealing with the exact self-energy, the Lehmann representation of the Green’s function guarantees Eq. (3), but the extension to approximate self-energies requires the mathematical treatment of NLEPs [37, 38]. Moreover, in writing Eq. (3) we have assumed the completeness of the nonlinear left/right eigenvectors; we also mention that, in the more general case of a self-energy that is analytic in a connected set of the complex plane, one needs to add to the rhs of Eq. (3) an analytic remainder function – see Supplemental Material [37] for an in-dept discussion.

Although there are multiple ways to solve Eq. (2) (e.g., see Refs. [38] or [39]), here we exploit the knowledge of residues and poles of the SE in Eq. (1) to find the SOP for the GF. Introducing a factorization of the self-energy residues Γm=VmVˉm†\Gamma_m=V_m\bar{V}_m^\dagger in Eq. (1), the Dyson equation can be rewritten as:

G(ω)=[ωI−h0−∑m=1NVm(ω−Ωm)−1Vˉm†]−1.(4)G(\omega)=\left[\omega I-h_0-\sum_{m=1}^{N}V_m(\omega-\Omega_m)^{-1}\bar{V}_m^\dagger\right]^{-1}. \tag*{(4)}

This can be interpreted as embedding a non-interacting system h0h_0 with NN non-interacting fictitious degrees of freedom, representing the bath, each having a Hamiltonian Ωm\Omega_m [40] coupled to h0h_0 by VmV_m and Vˉm†\bar{V}_m^\dagger. As can be proven by a direct calculation, the Green’s function in Eq. (4) can be obtained as G(ω)=P0(ωItot−HAIM)−1P0G(\omega)=P_0(\omega I_{\mathrm{tot}}-H_{\mathrm{AIM}})^{-1}P_0 and

HAIM=(h0V1⋯VNVˉ1†Ω1I100⋮0⋱0VˉN†0⋯ΩNIN),(5)H_{\mathrm{AIM}}= \begin{pmatrix} h_0 & V_1 & \cdots& V_N\\ \bar{V}_1^\dagger& \Omega_1 I_1 & 0 & 0\\ \vdots& 0 & \ddots& 0\\ \bar{V}_N^\dagger& 0 & \cdots& \Omega_N I_N \end{pmatrix}, \tag*{(5)}

with P0P_0 projecting onto the h0h_0 Hilbert space. We carry out the proof by exploiting embedding techniques in the Supplemental Material [37]. To summarize, AIM-SOP exactly maps the nonlinear eigenvalue problem, Eq. (2), into a linear eigenvalue problem by building a fictitious non-interacting system, with additional degrees of freedom (DOFs), that possesses the same GF as the interacting system upon projection. As a side but important note, the freedom in the factorization of Γm\Gamma_m can be exploited to minimize the dimension of the bath. In fact, when using SVD to factorize Γm\Gamma_m, the number of columns (rows) of VmV_m (Vˉm†\bar{V}_m^\dagger) is reduced to the rank of Γm\Gamma_m, thus making dim⁡[Im]=rank⁡[Γm]\dim[I_m]=\operatorname{rank}[\Gamma_m]. Then, the dimension of HAIMH_{\mathrm{AIM}} is reduced to dim⁡[HAIM]=dim⁡[h0]+∑mrank⁡[Γm]\dim[H_{\mathrm{AIM}}]=\dim[h_0]+\sum_m\operatorname{rank}[\Gamma_m].

It is important to note that, at variance with our previous work [23], in Eq. (5) the entries are matrix blocks and not scalars, also making apparent the link to NLEPs. In mathematical terms this treatment can be seen as a special case of the “linearization” of NLEPs using rational functions [38], with the HAIMH_{\mathrm{AIM}} matrix serving as an ad-hoc alternative to the companion matrix or other approaches [39, 41]. In electronic-structure methods, a construction comparable to HAIMH_{\mathrm{AIM}} is used in the context of DMFT to efficiently invert the Dyson equation [7, 42].

Here, also owing to the connection to the NLEP framework, the case of fully non-Hermitian (diagonalizable) Hamiltonians is also included, e.g., with Vm≠V‾mV_m \ne\overline{V}_m and complex Ωm\Omega_m. In the Supplemental Material [37] we also discuss the case of non-diagonalizable Hamiltonians together with the appearance of higher-order poles in the solution of the Dyson equation. By linking this approach to the theory of embedding and unfolding, it becomes apparent that the methods in Refs. [43, 44, 45, 46], all involving the construction of a supermatrix (or upfolded matrix) to solve the Dyson equation for the self-energy or the Bethe-Salpeter equation, can be seen as specialized cases of the present framework. It important to stress that to maintain computational efficiency it is essential to have a SOP representation of the self-energy; so, the physical and algorithmic assumption here is to work with approximate but analytical propagators (as SOP), and to solve exactly the Dyson equation within this representation.

Dynamical Hubbard functional. Taking AIM-SOP as a formal and effective approach to solving electronic-structure problems subject to dynamical potentials, we introduce a functional formulation that extends the DFT+U energy functional of Dudarev et al. [35] to host a frequency-dependent screened potential U(ω)U(\omega). In particular, we define a “dynamical Hubbard” Klein energy functional as:

EdynH[G]=EH[ρ]+Exc[ρ]+ΦdynH[G]−Tr⁡ω[G0−1G]+Tr⁡ωlog⁡G0−1G+Tr⁡ω[h0G0].(6)E_{\mathrm{dynH}}[G] = E_{\mathrm{H}}[\rho] + E_{\mathrm{xc}}[\rho] + \Phi_{\mathrm{dynH}}[G] - \operatorname{Tr}_{\omega}\left[G_0^{-1}G\right] + \operatorname{Tr}_{\omega}\log G_0^{-1}G + \operatorname{Tr}_{\omega}[h_0G_0]. \tag*{(6)}

Here, Tr⁡ω[…]\operatorname{Tr}_{\omega}[\ldots] stands for ∫dω2πieiω0+Tr⁡[…]\int\frac{d\omega}{2\pi i}e^{i\omega0^+}\operatorname{Tr}[\ldots], GG represents the Green’s function (GF) of the system, G0=ωI−h0G_0=\omega I-h_0 is the Kohn-Sham Green’s function, ρ\rho is the density derived from GG, and G\mathbf{G} is the projection of GG onto a localized manifold (typically dd- or ff- states), commonly referred to as the Hubbard manifold. For a single site, this dynamical Hubbard energy correction reads

ΦdynH[G]=12∫dω2πidω′2πieiω0++iω′0+U(ω′)××Tr⁡{G(ω+ω′)[δ(ω−c)I−G(ω)]},(7)\begin{aligned} \Phi_{\mathrm{dynH}}[G] = \frac{1}{2}\int\frac{d\omega}{2\pi i}\frac{d\omega'}{2\pi i}e^{i\omega0^+ + i\omega'0^+}U(\omega')\times\\ &\times\operatorname{Tr}\left\{\mathbf{G}(\omega+\omega')\left[\delta(\omega-c)\mathbf{I}-\mathbf{G}(\omega)\right]\right\}, \tag*{(7)} \end{aligned}

where U(ω)U(\omega) represents the screened potential W(ω)W(\omega) of the system projected onto a localized (Hubbard) space and averaged over it, and cc is a constant to fix double counting (DC). The DC term is included because we are correcting an approximate DFT functional, and its purpose is to suppress the exchange-correlation interactions of the localized manifold included at the DFT level (for a comprehensive review, see e.g. Ref. [47]).

To justify the ansatz for ΦdynH[G]\Phi_{\mathrm{dynH}}[G], we first look at its derivative, i.e., the (time-ordered) self-energy:

ΣdynH(ω)=2πiδΦdynH[G]δG=−∫dω′2πieiω′0+U(ω′)G(ω+ω′)+12U(c)I,(8)\begin{aligned} \Sigma_{\mathrm{dynH}}(\omega)=2\pi i\frac{\delta\Phi_{\mathrm{dynH}}[G]}{\delta G} \\ &=-\int\frac{d\omega'}{2\pi i}e^{i\omega'0^+}U(\omega')G(\omega+\omega')+\frac{1}{2}U(c)\mathbf{I}, \tag*{(8)} \end{aligned}

which is composed of a localization of the GWGW SE [48, 49] (first term), and the double-counting term 12U(c)I\frac{1}{2}U(c)\mathbf{I}. In particular, as shown in § 1.5 of the Supplemental Material [37] and similarly in Ref. [49], the dynamical Hubbard self-energy can be derived from the GWGW self-energy by discarding the itinerant part of the Lehmann (or KS) amplitudes. Also, if a static screening is considered, ΦdynH\Phi_{\mathrm{dynH}} reduces to the (static, rotationally invariant) DFT+U Hubbard correction of Ref. [35]. Indeed, similarly to considering DFT+U a truncation of COHSEX [50], the dynamical Hubbard functional stems from a localization of GWGW [49]. Note that, at variance with Ref. [49], the SE of Eq. (8) preserves the correct time-ordering, as no approximation in the frequency convolutions was carried out. Additionally, ΦdynH\Phi_{\mathrm{dynH}} reduces to the GW0GW_0 functional (plus a double-counting term) in the limit of the Hubbard manifold becoming the entire space. The generalization to multiple sites consists of summing different Φ\Phi terms, one per site, and is not treated here for simplicity. Note that, as discussed in § 1.6 of the Supplemental Material [37], we fix the double counting parameter as c→∞c\to\infty, so that U(c)=U∞U(c)=U_\infty, i.e., the bare Coulomb potential localized and averaged on the manifold.

Being a Klein functional with the screened interaction set to U(ω)U(\omega), EdynH[G]E_{\mathrm{dynH}}[G] can be made stationary by solving self-consistently the corresponding Dyson equation with a static term h0=hKS(ρ)h_0=h_{\mathrm{KS}}(\rho) and a SE, ΣdynH=∑m,m′∣ϕm⟩ΣdynHmm′(ω)⟨ϕm′∣\Sigma_{\mathrm{dynH}}=\sum_{m,m'}|\phi_m\rangle\Sigma_{\mathrm{dynH}}^{mm'}(\omega)\langle\phi_{m'}|, where {ϕm}\{\phi_m\} span the localized Hubbard manifold. Here, the use of the algorithmic inversion method is crucial and provides a closed (all within SOPs) formulation. This means that given an initial GF expressed as SOP, e.g., G=GKSG=G_{\mathrm{KS}}, and a SOP for the screened interaction U(ω)U(\omega), the SE ΣdynH\Sigma_{\mathrm{dynH}} can also be written as a SOP. Indeed, as mentioned above, the convolution of two SOPs is a SOP [23], and the projections are just linear operations on the residues.

With a SE on SOP, the algorithmic inversion method can be used to find the SOP for the GF. The cycle can then be iterated until self-consistency. Furthermore, the SOP form of the Green’s function naturally allows for the accurate evaluation of the generalized Hubbard energy of Eq. (7), the chemical potential, and, in general, integrated thermodynamic quantities [23]. In Ref. [48] Miyake et al. observe that their version of localized GWGW, termed GdWG_dW — same as Eq. (8) but with a different choice of double counting, see Supplemental Material [37] — gives very similar spectral results when the full kk-dependent GWGW self-energy is localized onto the manifold. Here, similarly to GdWG_dW [48], we first localize and then calculate the self-energy, given the existence of an energy functional for this form and the link to DFT+U.

Application to SrVO3_3. As a case study, we apply the formalism to SrVO3_3. This material is prototypical since it possesses localized electrons around the Fermi energy with dd character that are not strongly correlated [52, 53], and can thus be described within a GWGW-like approximation. Furthermore, localized GWGW has shown to provide qualitatively good spectral results for this material [48] by improving the GWGW bandwidth [48, 52] and yielding results similar to GW+GW+EDMFT [53]. Here, we take a step further and compute not only the spectral properties of the material but also integrated quantities, such as the total energy and its derivatives (lattice parameter, bulk modulus, phonons), obtained from the functional EdynH[G]E_{\mathrm{dynH}}[G]. Note that, while we choose to study a metallic system, the framework can be readily applied to semiconductors and insulators. As mentioned, a choice of U(ω)U(\omega), accounting for the dynamical screening of the electrons in the Hubbard localized manifold is needed. We consider U(ω)U(\omega) to be the average over the Hubbard manifold of the screened interaction WW in the random-phase approximation (RPA); at variance with EDMFT, which uses a cRPA screening [53, 54, 55], here we retain screening originating from all the bands (see the Supplemental Material [37] and the references therein for a discussion). Though designed for self-consistency, here we limit the approach to a one-shot calculation — i.e., a single step in the stationarization of EdynHE_{\mathrm{dynH}} — using as a starting propagator the self-consistent Kohn-Sham Green’s function computed with a standard semi-local DFT exchange-correlation functional (PBEsol [56, 57]). By doing so, we need to evaluate the energy functional at EdynH[G=GKS]E_{\mathrm{dynH}}[G=G_{\mathrm{KS}}], and the same for the self-energy, ΣdynH[G=GKS]\Sigma_{\mathrm{dynH}}[G=G_{\mathrm{KS}}]. Then, exploiting the algorithmic inversion method, we find the resulting non-self-consistent Green’s function. At this level, the dynamical Hubbard Klein energy func-

MethodBWm∗/mPBEsolm^*/m_{\mathrm{PBEsol}}LSUS
PBEsol1.01
PBEsol + U0.921.1
GW [52]0.651.5−22.2; 3.5
GW+C [52]0.651.5−22.2
(qp)locGW [48]0.52
GW+DMFT [58]0.52−1.62
GW+DMFT [59]0.62−1.52.5
GW+EDMFT [53]−1.72.8
This work0.52−2.5 ± 12.6; 3.5 ± 1
exp. [60]0.71.8−1.5
exp. [61]0.442−1.5

Table 1. Occupied bandwidth (BW) of the t2gt_{2g} bands, the mass enhancement factor (m∗/mPBEsolm^*/m_{\mathrm{PBEsol}}), and energy of the lower (LS) and upper satellites (US) from experiments and different theoretical frameworks—with two numbers indicating two different satellites. Results from the generalized dynamical Hubbard functional introduced in this work are estimated from the spectral function in Figure 1. m∗/mPBEsolm^*/m_{\mathrm{PBEsol}} is estimated using the ratio of the full bandwidths (LDA and PBEsol coincide). All energies are in eV.

tional simplifies to

EdynH[GKS]=EDFT[ρ]+ΦdynH[GKS],(9)E_{\mathrm{dynH}}[G_{\mathrm{KS}}] = E_{\mathrm{DFT}}[\rho] + \Phi_{\mathrm{dynH}}[G_{\mathrm{KS}}], \tag*{(9)}

correcting the DFT energy with a dynamical energy term. In this work, we use this last equation to evaluate the energy, and give all the numerical details of the simulations in the Supplemental Material [37].

In Figure 1 we present the spectral function resulting from our calculations. In contrast to DFT (PBEsol) or DFT+U (PBEsol+U), which only shift the chemical potential, the spectral function derived from ΣdynH\Sigma_{\mathrm{dynH}} drives a reduction in the bandwidth, along with a renormalization of the full width of the t2gt_{2g} bands. Table I provides details on the bandwidth (BW), the mass enhancement factor (m∗/mPBEsolm^*/m_{\mathrm{PBEsol}}), and the positions of the lower/upper satellites (LS/US) plus/minus the broadening. While PBEsol and PBEsol+U overestimate both the occupied and full bandwidth, the present results align well with the data from Ref. [48], obtained through localized-GWGW calculations (referred to as qp-locGW in the Table, as only the quasiparticle band structure is computed), and also closely match experimental data [60, 61] and results from DMFT [53, 58, 59].

Spectral function of SrVO3

Figure 1. Spectral function [51] of SrVO3_3 from this work (color plot) compared to PBEsol (solid white line) and PBEsol+U (dashed white line). Only the t2gt_{2g} bands are displayed. The chemical potential is shifted to 0 in all three cases. The color map is logarithmic.

Compared to the GWGW calculations in Ref. [52], the present results provide a further reduction in the bandwidth, from approximately 0.65 eV to about 0.5 eV. An asymmetric renormalization of t2gt_{2g} with respect to the Fermi energy, absent in the GWGW calculations, is also observed. Naturally, these effects are due to the localized nature of the correction. Concerning the satellites, the LS position is slightly underestimated (i.e. deeper in energy) compared to DMFT and GWGW, but exhibits a significant broadening of around 1 eV, making it still consistent with the references. While the position of the first upper satellite (US1) agrees with GWGW+EDMFT and GWGW+C results, the appearance of a second upper satellite (US2) at about 3.5 eV (similarly to GWGW) differs. A more detailed discussion of the incoherent part of the spectrum, density of states, and self-energy can be found in the Supplemental Material [37]. Notably, the consistency of quasi-particle results across different choices for the double-counting terms in our work and in Ref. [48] is reassuring. Additionally, the use of ortho-atomic dd orbitals to define the localized manifold, instead of maximally-localized Wannier functions from the dpdp model as done in Ref. [48], confirms that results are robust against specific details of the localized dd-manifold for this material. However, this may not always be the case, as, for instance, in situations where Wannier functions hybridize or localize along bonds.

From the knowledge of the SOP of the Green’s function (here KS) and the functional form of ΦdynH\Phi_{\mathrm{dynH}}, one can straightforwardly calculate total energies and total-energy differences. In Figure 2 we compare the equation of state for SrVO3_3 obtained from PBEsol, PBEsol+U, and the present dynamical formulation. In the legend, we report the estimated values for the equilibrium volume VV and the bulk modulus BB, using a Birch-Murnaghan third-order function for the fitting [63]. It can be observed that the present approach correctly predicts the softening of the bulk modulus while overcorrecting the equilibrium lattice parameter. Furthermore, in the Supplemental Material [37] we link the softening of the bulk modulus to the change of the charge density from the different approaches. While these findings are encourag-

Figure 2. Equation of state for SrVO3_3 calculated using PBEsol (blue), PBEsol+U (orange), and the present approach (green). Data (crosses) are fitted using the Birch-Murnaghan curve (solid line). The values and errors obtained using the fit are displayed in the legend. For reference, the experimental volume is marked in dashed black, and the bulk modulus is reported in the legend. Experimental values are taken from [62].

(THz)PBEsolLDA+EDMFT [34]This work
ω1\omega_14.864.33.9
ω2\omega_210.39.09.4
ω3\omega_311.011.212.1
ω4\omega_417.318.919.4

Table 2. Zone center phonons for SrVO3_3 calculated with different approaches; the last column shows the results obtained using the generalized dynamical functional introduced in this work (on top of PBEsol). The data from Ref. [34] have been estimated from the plot. Frequencies are in terahertz.

ing, they are to be considered preliminary in view of the absence of self-consistency.

Finally, utilizing the relaxed structures and capitalizing on the cost-effectiveness of the present method, we compute the zone-center phonons for SrVO3_3. Data are reported in Tab. II. Since, given the lack of self-consistency, the Hellmann-Feynman theorem does not apply, we use finite-energy differences. Consistently with the softening of the bulk modulus, the first two frequencies are lower compared to PBEsol calculations. Additionally, similar to state-of-the-art methods such as LDA+EDMFT [34], the remaining optical modes are shifted to higher frequencies.

In conclusion, in this work, we provide a computationally straightforward Green’s function framework to address the electronic structure of materials exhibiting correlation in a localized manifold. First, we establish a connection between the solution of Dyson equation and nonlinear eigenvalue problems; then we solve the equation exactly by extending the algorithmic inversion method on sum over poles (AIM-SOP), introduced in Ref. [23], to the general non-homogeneous (matricial) case to treat realistic materials. Next, we present an approximation to the exchange-correlation Φxc\Phi_{\mathrm{xc}} part of the Klein energy functional, yielding a localized-GWGW self-energy, and providing a dynamical generalization of DFT+U. We combine this functional with the algorithmic inversion method to study the paradigmatic case of SrVO3_3. Our results closely agree with state-of-the-art computational approaches but come with a very modest computational cost. The method allows one to simultaneously access spectral and thermodynamic properties, including total energies and their differences. Moreover, the existence of a Klein functional guarantees the possibility of performing self-consistent calculations. This, in turn, will enable the use of the Hellmann-Feynman theorem to calculate forces and other derivatives within a Green’s function formalism. Additionally, besides self-consistency, the method is readily applicable to the study of insulating materials. For example, it can be used to correct erroneous metallic ground states predicted by semi-local DFT, as is often the case with DFT+U.

We acknowledge stimulating discussions with Mario Caserta and Marco Vanzini. This work was supported by the Swiss National Science Foundation (SNSF) through grant No. 200021-179138 (T.C.) and NCCR MARVEL (N.M.), a National Centre of Competence in Research through grant No. 205602, and from the EU Commission for the MaX Centre of Excellence on ‘Materials Design at the eXascale’ under grant No. 101093374 (A.F., N.M.).

References

Note that, due to the finite lifetimes, we estimate the occupied bandwidth and effective mass, i.e. the position of the first t2gt_{2g} quasiparticle peak (BW) and last peak (full), as the center of the peak of the DOS. In the figure we do not use an extra broadening for the plotting.

Supplemental Material: Energies and spectra of solids from the algorithmic inversion of dynamical Hubbard functionals

Tommaso Chiarotti; Andrea Ferretti; Nicola Marzari

The Dyson equation as a nonlinear eigenvalue problem

To complement the discussion in the main text, in this Section we prove that if Σ\Sigma is sufficiently regular [38] (see below), one can exactly map the solution to the Dyson equation into the nonlinear eigenvalue problem:

[h0+Σ(ω)]∣ψr⟩=ω∣ψr⟩,(10)[h_0+\Sigma(\omega)]|\psi^r\rangle=\omega|\psi^r\rangle, \tag*{(10)}

where (zs,ψsr)(z_s,\psi_s^r) are the corresponding (discrete) nonlinear eigenvalues and right eigenvectors, as in the main text. In particular, one can show that that the nonlinear eigenvalues zsz_s provide the poles of the resolvent G(ω)G(\omega), i.e., [ω−h0−Σ(ω)]−1[\omega-h_0-\Sigma(\omega)]^{-1}, and, when poles are first-order (not granted in general), the right/left nonlinear eigenvectors ∣ψsr⟩⟨ψsl∣|\psi_s^r\rangle\langle\psi_s^l| provide their residues. Note that the results of this Section may also be deduced from the mathematical literature of NLEPs [38].

Under the hypothesis of a self-energy analytic in a connected set of the complex plane (which includes the case of rational functions) [38], we prove the above statement by showing that:

[ω−h0−Σ(ω)]−1=∑s∣ψsr⟩⟨ψsl∣ω−zs+r(ω),(11)[\omega-h_0-\Sigma(\omega)]^{-1}=\sum_s\frac{|\psi_s^r\rangle\langle\psi_s^l|}{\omega-z_s}+r(\omega), \tag*{(11)}

where r(z)r(z) is a smooth function without poles, and ∣ψsr⟩|\psi_s^r\rangle satisfying

[zs−h0−Σ(zs)]∣ψsr⟩=0,(12)[z_s-h_0-\Sigma(z_s)]|\psi_s^r\rangle=0, \tag*{(12)}

and the corresponding left equation for ⟨ψsl∣\langle\psi_s^l|.

We start by computing:

(∑s′∣ψs′r⟩⟨ψs′l∣ω−zs′[ω−h0−Σ(ω)])∣ψsr⟩=∑s′∣ψs′r⟩⟨ψs′l∣ω−zs′[zs−h0−Σ(zs)+[1−Σ′(zs)](ω−zs)+o(ω−zs)]∣ψsr⟩.(13)\left(\sum_{s'}\frac{|\psi_{s'}^r\rangle\langle\psi_{s'}^l|}{\omega-z_{s'}}[\omega-h_0-\Sigma(\omega)]\right)|\psi_s^r\rangle= \sum_{s'}\frac{|\psi_{s'}^r\rangle\langle\psi_{s'}^l|}{\omega-z_{s'}}\left[z_s-h_0-\Sigma(z_s)+[1-\Sigma'(z_s)](\omega-z_s)+o(\omega-z_s)\right]|\psi_s^r\rangle. \tag*{(13)}

By using Eq. (S3) we arrive to:

(∑s′∣ψs′r⟩⟨ψs′l∣ω−zs′[ω−h0−Σ(ω)])∣ψsr⟩=∣ψsr⟩⟨ψsl∣[1−Σ′(zs)]∣ψsr⟩+Rs(ω)∣ψsr⟩,(14)\left(\sum_{s'}\frac{|\psi_{s'}^r\rangle\langle\psi_{s'}^l|}{\omega-z_{s'}}[\omega-h_0-\Sigma(\omega)]\right)|\psi_s^r\rangle=|\psi_s^r\rangle\langle\psi_s^l|[1-\Sigma'(z_s)]|\psi_s^r\rangle+R_s(\omega)|\psi_s^r\rangle, \tag*{(14)}

where we stress that, owing to the left hand side of the equation above, the rest Rs(ω)∣ψsr⟩R_s(\omega)|\psi_s^r\rangle does not have poles in zs′≠zsz_{s'}\ne z_s, and, in addition, Rs(ω)∣ψsr⟩=O(ω−zs)∣ψsr⟩R_s(\omega)|\psi_s^r\rangle=O(\omega-z_s)|\psi_s^r\rangle. Then, setting the normalization of the left and right eigenvectors as ⟨ψsl∣[1−Σ′(zs)]∣ψsr⟩=1\langle\psi_s^l|[1-\Sigma'(z_s)]|\psi_s^r\rangle=1, we can define a smooth function R~(ω)\widetilde{R}(\omega) as:

∑s′∣ψs′r⟩⟨ψs′l∣ω−zs′[ω−h0−Σ(ω)]=I+R~(ω),(15)\sum_{s'}\frac{|\psi_{s'}^r\rangle\langle\psi_{s'}^l|}{\omega-z_{s'}}[\omega-h_0-\Sigma(\omega)]=I+\widetilde{R}(\omega), \tag*{(15)}

which, by using Eq. (S5), satisfies R~(ω)∣ψsr⟩=Rs(ω)∣ψsr⟩\widetilde{R}(\omega)|\psi_s^r\rangle=R_s(\omega)|\psi_s^r\rangle, and thus is a smooth function (on the whole vectorial space if the invertibility of ∑s∣ψsr⟩⟨ψsl∣\sum_s|\psi_s^r\rangle\langle\psi_s^l| is assumed). Thus, if we define:

r(ω):=[ω−h0−Σ(ω)]−1−∑s∣ψsr⟩⟨ψsl∣ω−zs,(16)r(\omega):=[\omega-h_0-\Sigma(\omega)]^{-1}-\sum_s\frac{|\psi_s^r\rangle\langle\psi_s^l|}{\omega-z_s}, \tag*{(16)}

we have r(ω)[ω−h0−Σ(ω)]=−R~(ω)r(\omega)[\omega-h_0-\Sigma(\omega)]=-\widetilde{R}(\omega). To prove that r(ω)r(\omega) does not have poles at any zsz_s, one can suppose the absurd thesis that it has, and then find the contradiction:

r(ω)[ω−h0−Σ(ω)]∣ψsr⟩=[C+O(ω−zs)]∣ψsr⟩≠−R~(ω)∣ψsr⟩=O(ω−zs)∣ψsr⟩,(17)r(\omega)[\omega-h_0-\Sigma(\omega)]|\psi_s^r\rangle=[C+O(\omega-z_s)]|\psi_s^r\rangle\ne-\widetilde{R}(\omega)|\psi_s^r\rangle=O(\omega-z_s)|\psi_s^r\rangle, \tag*{(17)}

using again the expansion of [ω−h0−Σ(ω)][\omega-h_0-\Sigma(\omega)] around zsz_s. This shows the validity of Eq. (11), as extensively discussed in Ref. [38]. Note that in the case of a rational self-energy R(ω)=0R(\omega)=0, as can be seen from the AIM-SOP construction (see the main text and the next Section).

AIM-SOP: proof for the solution of the Dyson equation via HAIM

In this section, we provide the proof for the solution of the Dyson equation in the sum-over-poles formalism. Specifically, we prove that given the poles and residues of the SOP of self-energy Σ\Sigma, Eq. (1), the eigenvalues of the of the algorithmic inversion matrix in Eq. (5) are the poles of the Green’s function, and the (projection of the) eigenvectors its residues.

We start by observing that HAIMH_{\mathrm{AIM}} can be rewritten as:

HAIM=(h0VVˉ†Ω),withΩ=(Ω1I1000⋱00⋯ΩNIN)(18)H_{\mathrm{AIM}}=\begin{pmatrix}h_0 & V\\ \bar{V}^{\dagger} & \Omega\end{pmatrix},\quad\text{with}\quad\Omega=\begin{pmatrix}\Omega_1 I_1 & 0 & 0\\ 0 & \ddots& 0\\ 0 & \cdots& \Omega_N I_N\end{pmatrix} \tag*{(18)}

and V=(V1 ⋯ VN)V=(V_1\ \cdots\ V_N) — and similarly for Vˉ†\bar{V}^{\dagger} but as a column vector. Taking an embedding perspective [18], we rewrite the Dyson equation for the fictitious system (with Hamiltonian HAIMH_{\mathrm{AIM}}) making it explicit the degrees of freedom (DOFs) of h0h_0 (SS), and the additional DOFs of Ω\Omega for the bath (BB):

(ω−h0VVˉ†ω−Ω)(GSGSBGBSGB)=(IS00IB)(19)\begin{pmatrix}\omega-h_0 & V\\ \bar{V}^{\dagger} & \omega-\Omega\end{pmatrix}\begin{pmatrix}G_S & G_{SB}\\ G_{BS} & G_B\end{pmatrix}=\begin{pmatrix}I_S & 0\\ 0 & I_B \tag*{(19)}\end{pmatrix}

This is a system of four matricial equations and four variables, the Green’s functions GSG_S and GBG_B and the coupling terms GSBG_{SB} and GBSG_{BS}. Using the first and the third equations, one arrives at GS−1(ω)=ω−h0−V1ω−ΩVˉ†G_S^{-1}(\omega)=\omega-h_0-V\frac{1}{\omega-\Omega}\bar{V}^{\dagger}, which gives Eq. (4) of the main text with GS=GG_S=G. Furthermore, from Eq. (19), we have G(ω)=GS(ω)=P0(ωItot−HAIM)−1P0G(\omega)=G_S(\omega)=P_0(\omega I_{\mathrm{tot}}-H_{\mathrm{AIM}})^{-1}P_0, which shows that the eigenvalues of the static (fictitious) Hamiltonian HAIMH_{\mathrm{AIM}} are the poles of the real Green’s function GG, while the projection, by P0P_0, of the HAIMH_{\mathrm{AIM}} eigenvectors are the Dyson orbitals (or residues). In addition, as anticipated in the previous Section, this particular case of a self-energy as a SOP (plus a constant term), yields the remainder r(ω)r(\omega) of Eq. (11) equal to zero.

AIM-SOP: higher-order poles

In this Section we discuss, in the context of the AIM-SOP framework, (i) the possibility of handling self-energies displaying higher order poles in their SOP representation, and (ii) the case of the AIM matrix being non-diagonalizable. While originating from different features of the approximate self-energy in use, (i) and (ii) are tightly related mathematically, since non-diagonalizability involves Jordan blocking, and the resolvents of Jordan blocks involve higher order poles. For instance, a non-positive definite approximation of the Φxc\Phi_{\mathrm{xc}} functional [64] may lead to (ii).

Jordan blocks and higher order poles

As discussed in the main text, diagonalizable AIM matrices have resolvents (Green’s functions) with first order poles (called simple or semi-simple in the presence of degeneracies). Therefore, it is not surprising that resolvents with higher order poles originate from non-diagonalizable matrices, and Jordan blocks in particular. Let us consider the nilpotent matrix

Nn=(01⋯0101⋱)(20)N_n=\begin{pmatrix}0 & 1 & \cdots\\ & 0 & 1\\ & & 0 & 1\\ & & \ddots\end{pmatrix} \tag*{(20)}

which satisfies the identity (Nn)n=0(N_n)^n=0, since each multiplication by NnN_n shifts by 1 the position of the non-zero superdiagonal. This is also a Jordan block with zero diagonal entries, the more general case being:

Jn(Ω)=ΩIn+Nn(21)J_n(\Omega)=\Omega I_n+N_n \tag*{(21)}

By taking into account the nilpotency property of NnN_n, its resolvent can be written as:

rn(ω)=[ωIn−Nn]−1=1ω[I+Nnω+Nn2ω2+⋯+Nnn−1ωn−1](22)\begin{aligned} r_n(\omega)=\left[\omega I_n-N_n\right]^{-1} \\ &=\frac{1}{\omega}\left[I+\frac{N_n}{\omega}+\frac{N_n^2}{\omega^2}+\cdots+\frac{N_n^{n-1}}{\omega^{n-1}}\right] \tag*{(22)} \end{aligned}

which can be directly verified by multiplying the second line of Eq. (22) by [ωIn−Nn][\omega I_n-N_n]. In practice, one has:

rn(ω)=(1ω1ω2⋯1ωn01ω1ω2⋯001ω⋯⋱)(23)r_n(\omega)=\begin{pmatrix}\frac{1}{\omega} & \frac{1}{\omega^2} & \cdots& \frac{1}{\omega^n}\\ 0 & \frac{1}{\omega} & \frac{1}{\omega^2} & \cdots\\ 0 & 0 & \frac{1}{\omega} & \cdots\\ & & & \ddots\end{pmatrix} \tag*{(23)}

where the resolvent shows a order-nn pole.

Self-energies with higher order poles

We are now in the position to discuss the case of self-energies expressed in SOP form with higher order poles. Let us assume a self-energy of the form:

Σ(ω)=Σ0(ω)+r(ω−Ω)n=Σ0(ω)+ΔΣ(ω),(24)\Sigma(\omega)=\Sigma_0(\omega)+\frac{r}{(\omega-\Omega)^n}=\Sigma_0(\omega)+\Delta\Sigma(\omega), \tag*{(24)}

where the first term, Σ0(ω)\Sigma_0(\omega), contains first order poles (one, for simplicity, in the present discussion), and, for the moment, we assume r=∣v⟩⟨v∣r=|v\rangle\langle v|. While we can exploit the standard AIM-SOP treatment to write Σ0(ω)=V1[ωI−diag⁡(Ω1)]−1Vˉ1†\Sigma_0(\omega)=V_1[\omega I-\operatorname{diag}(\Omega_1)]^{-1}\bar{V}_1^{\dagger}, where Ω1\Omega_1 represents the set of first order poles of Σ0(ω)\Sigma_0(\omega), we have to resort to Eq. (23) to write the higher order term ΔΣ(ω)\Delta\Sigma(\omega). In fact, one can write:

ΔΣ(ω)=VLrn(ω−Ω)VˉR†(25)\Delta\Sigma(\omega)=V_L r_n(\omega-\Omega)\bar{V}^{\dagger}_R \tag*{(25)}

where

VL=(v0)VˉR†=(0vˉ†).(26)V_L=\begin{pmatrix} v & 0 \end{pmatrix} \qquad \bar{V}^{\dagger}_R=\begin{pmatrix} 0 \\ \bar{v}^{\dagger} \end{pmatrix}. \tag*{(26)}

In the above expressions, vv and vˉ†\bar{v}^{\dagger} are representations on a basis of ∣v⟩\lvert v\rangle and ⟨vˉ∣\langle\bar{v}\rvert, located in the first column of VLV_L and last row of VˉR†\bar{V}^{\dagger}_R, respectively. Here, the number of rows (columns) of VLV_L (VˉR†\bar{V}^{\dagger}_R) is equal to the degree of the pole nn. Therefore, in Eq. (25) the matrices VLV_L, VˉR†\bar{V}^{\dagger}_R couple only to the upper-right element of rnr_n, which is 1/(ω−Ω)n1/(\omega-\Omega)^n.

On the basis of the above results, the AIM matrix for Σ(ω)\Sigma(\omega) in Eq. (24) can be written as:

HAIM=(h0V1VLVˉ1†diag⁡(Ω1)0VˉR†0Jn(Ω)),(27)\mathcal{H}_{\mathrm{AIM}}= \begin{pmatrix} h_0 & V_1 & V_L\\ \bar{V}^{\dagger}_1 & \operatorname{diag}(\Omega_1) & 0\\ \bar{V}^{\dagger}_R & 0 & J_n(\Omega) \end{pmatrix}, \tag*{(27)}

where the first 2×22\times2 blocks correspond to the AIM matrix for Σ0(ω)\Sigma_0(\omega) and the remaining blocks generate the higher order term of the self-energy. In the more general case, when the factorization of rr gives rise to VV and Vˉ†\bar{V}^{\dagger} with ranks higher than 1, a Jordan block Jn(Ω)J_n(\Omega) of order nn needs to be added for each linearly independent coupling vector ∣vi⟩\lvert v_i\rangle.

Non-diagonalizable AIM matrices. Once we have provided a AIM matrix able to handle the case of higher order poles in the input self-energy, we briefly discuss the case of the AIM matrix being non-diagonalizable (independently of order of the poles present in the self-energy). Non diagonalizable matrices can be transformed to the Jordan form, that is a block-diagonal structure where some of the blocks are Jordan blocks. Computing the resolvent of the AIM matrix then leads to a Green’s function with higher order poles. In order to find the explicit SOP form of the output GF, one needs first to transform the AIM matrix to Jordan form, and then apply the formal expression for the Jordan block resolvent given in Eq. (23). Given the numerical difficulties in finding the Jordan form, a similar scheme based, e.g., on Schur decomposition and triangular matrices can be adopted, this being beyond the scope of this work.

From static U to dynamical U()

Similarly to choosing the screening parameter UU in DFT+UU, a choice of U(ω)U(\omega), accounting for the dynamical screening of the electrons in the Hubbard localized manifold, is required. In the spirit of localized GWGW and Ref. [49], we calculate the screening for the Hubbard site

II as the manifold average

U(ω)=1NI2∑m,m′WIm′,Im(ω),(28)U(\omega)=\frac{1}{N_I^2}\sum_{m,m'}W_{Im',Im}(\omega), \tag*{(28)}
WIm,Im′(ω)=∫drdr′∣ϕIm(r)∣2W(r,r′,ω)∣ϕIm′(r′)∣2W_{Im,Im'}(\omega)=\int d\mathbf{r}d\mathbf{r}'\lvert\phi_{Im}(\mathbf{r})\rvert^2W(\mathbf{r},\mathbf{r}',\omega)\lvert\phi_{Im'}(\mathbf{r}')\rvert^2

with WW calculated within the random-phase approximation, and mm being the orbital index of localized functions centered on the site II; e.g., atomic projectors or maximally localized Wannier functions [65]. For SrVO3_3 we choose the vanadium as the active Hubbard site, and ortho-atomic projectors as localized functions [66].

After being evaluated on grid, U(ω)U(\omega) is transformed to a SOP representation using a mixed linear and non-linear least square minimization of the cost function

C=∑j∣U(ωj)−∑lAlωj−gl∣2,(29)\mathcal{C}=\sum_j\left|U(\omega_j)-\sum_l\frac{A_l}{\omega_j-g_l}\right|^2, \tag*{(29)}

where, at each step of the minimization, the real coefficients AlA_l are found with a generalized linear least-square approach [67], and the complex (time-ordered) poles glg_l with a simulated annealing on C\mathcal{C} using the newly found amplitudes. Such an algorithm requires an initial guess of the real (position) and imaginary (broadening) parts of the poles, and, at each step, is required to not move these by at most 10% of their value. The algorithm has been shown to be robust with respect to changes of the initial values.

In Figure S1 we show the real (left) and imaginary (right) parts of the calculated RPA screened interaction (black lines) compared to the result of the fitting procedure (green lines), having chosen 10 poles for the fit. The agreement is excellent at lower energies (ω<25\omega< 25 eV) and slightly worsen afterwards. Still, the error in the two limits ω→0\omega\to0 and ω→∞\omega\to\infty (300 eV) is of the order of 10 meV, and thus quite low. Furthermore, the result is in agreement with the W0W_0 calculation of Ref. [53] for the low-energy part displayed in the reference. Specifically, while the dispersion of the imaginary part of U(ω)U(\omega) is qualitatively the same, the values differ. This is most probably due to a different choice for the broadening of the G0G_0 propagator. In particular, the features are sharper in Figure S1.

Figure S1. Real and imaginary part of the localized screened interaction U(ω)U(\omega) calculated on a frequency grid using the random-phase approximation (black line) compared to its representation in a sum-over-poles form (green line) obtained using a non-linear least squared minimization technique.

From GW to localized-GW

In this section we show that

ΣGWLoc=−∫dω′2πieiω′0+G(ω+ω′)U(ω′)(30)\boldsymbol{\Sigma}_{GW\mathrm{Loc}}=-\int\frac{\mathrm{d}\omega'}{2\pi i}e^{i\omega'0^+}\mathbf{G}(\omega+\omega')U(\omega') \tag*{(30)}

can be seen as a localized form of the GWGW self-energy,

Σ(r,r′,ω)=−∫dω′2πieiω′0+G(r,r′,ω+ω′)W(r,r′,ω).(31)\Sigma(\mathbf{r},\mathbf{r}',\omega)=-\int\frac{\mathrm{d}\omega'}{2\pi i}e^{i\omega'0^+}G(\mathbf{r},\mathbf{r}',\omega+\omega')W(\mathbf{r},\mathbf{r}',\omega). \tag*{(31)}

This approach has been adapted from Ref. [49] and is similar to Refs. [50, 68]. As discussed in the main text, bold symbols such as G\mathbf{G} or Σ\boldsymbol{\Sigma} refer to Green’s function and self-energies projected on the localized manifold (dd orbitals in our case).

In the following we make use of the Lehmann representation of the Green’s function,

G(r,r′,ω)=∑sψs(r)ψs∗(r′)ω−ϵs±i0+,(32)G(\mathbf{r},\mathbf{r}',\omega)=\sum_s\frac{\psi_s(\mathbf{r})\psi_s^*(\mathbf{r}')}{\omega-\epsilon_s\pm i0^+}, \tag*{(32)}

where ψs(r)\psi_s(\mathbf{r}) are the Dyson orbitals, ϵs\epsilon_s the charged excitations of the material, and ±i0+\pm i0^+ provide the correct time ordering of the propagator. On a localized basis around a site II (labelled here with Greek letters like α\alpha, β\beta, γ\gamma) the self energy reads

⟨α∣Σ(ω)∣β⟩=−∑s∫dω′2πieiω′0+∫dr dr′[ψs(r)α∗(r)W(r,r′,ω)β(r′)ψs∗(r′)ω+ω′−ϵs±i0+].(33)\langle\alpha|\Sigma(\omega)|\beta\rangle=-\sum_s\int\frac{\mathrm{d}\omega'}{2\pi i}e^{i\omega'0^+}\int\mathrm{d}\mathbf{r}\,\mathrm{d}\mathbf{r}'\left[\frac{\psi_s(\mathbf{r})\alpha^*(\mathbf{r})W(\mathbf{r},\mathbf{r}',\omega)\beta(\mathbf{r}')\psi_s^*(\mathbf{r}')}{\omega+\omega'-\epsilon_s\pm i0^+}\right]. \tag*{(33)}

Introducing the four-point notation for the screened potential [18]

Wαβss(ω)=∫dr dr′ α∗(r)ψs(r)W(r,r′,ω)β(r′)ψs∗(r′),(34)W_{\alpha\beta}^{ss}(\omega)=\int\mathrm{d}\mathbf{r}\,\mathrm{d}\mathbf{r}'\,\alpha^*(\mathbf{r})\psi_s(\mathbf{r})W(\mathbf{r},\mathbf{r}',\omega)\beta(\mathbf{r}')\psi_s^*(\mathbf{r}'), \tag*{(34)}

the self-energy can be rewritten as

⟨α∣Σ(ω)∣β⟩=−∑s∫dω′2πieiω′0+Wαβss(ω)ω+ω′−ϵs±0+.(35)\langle\alpha|\Sigma(\omega)|\beta\rangle=-\sum_s\int\frac{\mathrm{d}\omega'}{2\pi i}e^{i\omega'0^+}\frac{W_{\alpha\beta}^{ss}(\omega)}{\omega+\omega'-\epsilon_s\pm0^+}. \tag*{(35)}

Then, as in Refs. [49, 50, 68], we separate each Dyson orbital into an itinerant (delocalized) component and a localized one, ∣ψs⟩=∣s⟩=∣sit⟩+∑γ∣γ⟩⟨γ∣s⟩|\psi_s\rangle=|s\rangle=|s_{\mathrm{it}}\rangle+\sum_\gamma|\gamma\rangle\langle\gamma|s\rangle. The screened potential is separated in

Wαβss(ω)=Wαβsitsit(ω)+∑γ[⟨s∣γ⟩Wαβγsit(ω)+Wαβsitγ(ω)⟨γ∣s⟩]+∑γ,γ′⟨s∣γ⟩Wαβγγ′(ω)⟨γ′∣s⟩,(36)W_{\alpha\beta}^{ss}(\omega)=W_{\alpha\beta}^{s_{\mathrm{it}}s_{\mathrm{it}}}(\omega)+\sum_\gamma\left[\langle s|\gamma\rangle W_{\alpha\beta}^{\gamma s_{\mathrm{it}}}(\omega)+W_{\alpha\beta}^{s_{\mathrm{it}}\gamma}(\omega)\langle\gamma|s\rangle\right] +\sum_{\gamma,\gamma'}\langle s|\gamma\rangle W_{\alpha\beta}^{\gamma\gamma'}(\omega)\langle\gamma'|s\rangle, \tag*{(36)}

where, since we suppose the screened interaction to be short-range, we discard the first three terms. Owing to the localization of the basis of the site, we can approximate the screened interaction of the fourth term as [50]:

Wαβγγ′(ω)=∫dr dr′ α∗(r)γ(r)W(r,r′,ω)γ′∗(r′)β(r′)≈δαγ′δγβWαβαβ(ω).(37)W_{\alpha\beta}^{\gamma\gamma'}(\omega)=\int\mathrm{d}\mathbf{r}\,\mathrm{d}\mathbf{r}'\,\alpha^*(\mathbf{r})\gamma(\mathbf{r})W(\mathbf{r},\mathbf{r}',\omega)\gamma'^*(\mathbf{r}')\beta(\mathbf{r}') \approx\delta_{\alpha\gamma'}\delta_{\gamma\beta}W_{\alpha\beta}^{\alpha\beta}(\omega). \tag*{(37)}

Discarding the three terms in Eq. (36), and using Eq. (37), we can write the localized form of the self-energy from Eq. (35)

⟨α∣ΣGWLoc(ω)∣β⟩=−∫dω′2πieiω′0+∑s⟨α∣s⟩⟨s∣β⟩ω+ω′−ϵs±i0+Wαββα(ω).(38)\langle\alpha|\Sigma_{GW\mathrm{Loc}}(\omega)|\beta\rangle=-\int\frac{\mathrm{d}\omega'}{2\pi i}e^{i\omega'0^+}\sum_s\frac{\langle\alpha|s\rangle\langle s|\beta\rangle}{\omega+\omega'-\epsilon_s\pm i0^+}W_{\alpha\beta}^{\beta\alpha}(\omega). \tag*{(38)}

Calling ∣α⟩=∣I,m⟩|\alpha\rangle=|I,m\rangle and ∣β⟩=∣J,m′⟩|\beta\rangle=|J,m'\rangle, together with discarding off-site contributions—those terms would give a +V+V-like term [49]—, and averaging the on site terms m,m′m,m', as in Eq. (28), the localized-GWGW self-energy reads as in Eq. (30), for a single site.

An immediate consequence of the derivation above is the use of the RPA screening WW, at variance with the constrained RPA (WrW_r) used, e.g., in EDMFT [53, 54, 55]. A possible way to understand the difference is presented in Ref. [69], where it is shown that the screening potential WW at RPA level of an embedded (sub)system dd having particles interacting with an effective cRPA potential WrW_r—instead of the bare vv and coming from the rest of the system rr—is the RPA WW of the whole system. Then, the GWGW self-energy comes out to be the one of Eq. (8) with the RPA WW as screening potential. Physically, a heuristic argument to understand this is that if the model does not add other forms of screening (at variance, e.g., with EDMFT where the exact solution for the subspace provides the local part of the screening) the effective screening to be used is the fully screened potential WW. As a side note, we cannot apply this argument to DFT+UU calculations, just taking the RPA U(ω=0)U(\omega= 0) value for UU. Indeed for SrVO3_3, the RPA U(0)U(0) value is around 1.1 eV, where instead the linear response value [70] that is used for the DFT+U calculations is ∼6\sim6 eV and is much more similar to the cRPA value. This discussion traces back to the fact that DFT+U more than correcting correlations in the localized manifold, corrects the piece-wise linearity of the approximate DFT functional in that manifold [66, 71].

Double counting for the dynamical Hubbard functional

Since the dynamical Hubbard functional adds a correction to the exchange-correlation term of the DFT functional, it is necessary to estimate how much of this correlation is already present at the approximate DFT level in such a way to not double count it. This estimate can be done for the total energy or equivalently for the self-energy. Here, we focus on the latter and build the correction for the Klein functional consequently. Similarly to Ref. [49], since we want a frequency-independent mean-field-like double counting, we look at the mean-field (ω→∞\omega\to\infty) limit averaged over the manifold. Looking at the expression of the self-energy of Eq. (30), the average of the localized Fock term is (note that we change the notation from ΣGWLoc\Sigma_{\mathrm{GWLoc}} to ΣdynH\Sigma_{\mathrm{dynH}}):

⟨ΣdynH(ω)⟩→ω→∞⟨n⟩U∞.(39)\left\langle\Sigma_{\mathrm{dynH}}(\omega) \right\rangle\xrightarrow{\omega\to\infty} \langle n \rangle U_{\infty}. \tag*{(39)}

This term would yield an around mean-field double counting, where instead we want to a fully localized double counting. The easiest way to derive such double counting from the expression of the around mean-field is [47] to set the average occupation to 12\frac{1}{2}. Thus the expression of the self-energy accounting for the double counting is:

ΣdynH(ω)=2πiδΦdynH[G]δG=−∫dω′2πieiω′0+U(ω′)G(ω+ω′)+12U∞I,(40)\begin{aligned} \Sigma_{\mathrm{dynH}}(\omega) = 2\pi i \frac{\delta\Phi_{\mathrm{dynH}}[G]}{\delta G} \\ &= -\int\frac{d\omega'}{2\pi i}e^{i\omega'0^+}U(\omega')G(\omega+\omega')+\frac{1}{2}U_{\infty}I, \tag*{(40)} \end{aligned}

that coincides with Eq. (8) of the main text once the limit c→∞c \to\infty is taken. This is the expression used for the calculations. Equivalently, the generalized Hubbard functional with a fixed choice for the double counting reads:

ΦdynH[G]=−12∫dω2πidω′2πieiω0+eiω′0+U(ω′)G(ω+ω′)G(ω)+U∞2∫dω2πieiω0+G(ω),(41)\Phi_{\mathrm{dynH}}[G] = -\frac{1}{2}\int\frac{d\omega}{2\pi i}\frac{d\omega'}{2\pi i}e^{i\omega0^+}e^{i\omega'0^+}U(\omega')G(\omega+\omega')G(\omega) +\frac{U_{\infty}}{2}\int\frac{d\omega}{2\pi i}e^{i\omega0^+}G(\omega), \tag*{(41)}

where in the last term we kept the integral form to make the dependence on GG explicit. In passing we note that contrary to GW, the double-counting term is not dependent on the underlying DFT functional, but instead aims to remove the static averaged correlated part that is added, as is done in the context of DFT+U. As a future perspective, using the localized version of the GW double counting, i.e., by localizing vxcv_{xc}, could be an interesting alternative.

Numerical details

All DFT and DFT+U calculations are performed using the PWscf code of QUANTUM ESPRESSO distribution [72]. The energy, self-energy, Green’s function, and spectral function are computed with a PYTHON implementation of the generalized Hubbard approach described in the main text. The localized screened-potential U(ω)U(\omega) has been calculated using RESPACK [73] using maximally localized Wannier functions. The pseudopotentials used for all the calculations are optimized norm-conserving pseudopotentials [74] (PBEsol standard-precision, nc-sr-04) from the Pseudo Dojo library [75]. The phonon calculations have been performed with the help of PHONOPY [76, 77] to compute the energy differences. The sum-over-poles form for U(ω)U(\omega) is found using an in-house PYTHON implementation of the algorithm described in Sec. S4.

For the calculations in Figure 1 (main text) and S4 (Supp. Mater.), and Table I (main text), i.e., all but those for the equation of state (EOS), we use a simple cubic cell having a lattice parameter of a=3.824a = 3.824 Å. This is the extrapolated value at the zero-temperature limit of the experimental value a=3.841a = 3.841 Å from Refs. [62] and [78], considering an average linear thermal expansion coefficient αl=1.45×10−5 K−1\alpha_l = 1.45 \times10^{-5}\,\mathrm{K}^{-1} from Ref. [62]. The phonon calculations in Table II are performed at the relaxed lattice parameter both for the DFT (PBEsol) and using the dynamical Hubbard functional.

For all electronic structure calculations we use a Marzari-Vanderbilt [79] smearing of ∼0.27\sim0.27 eV. For the generalized Hubbard calculations, involving GKSG_{\mathrm{KS}}, we shift the poles above and below the real axis with an effective amount of iη±=±i0.1i\eta_{\pm} = \pm i0.1 eV, according to whether the pole is occupied (+) or empty (−). To implement the smearing in GKSG_{\mathrm{KS}}, we double each pole to be both above and below the real axis multiplying its occupation—1 in the Kohn and Sham basis—, by a factor f±(1−f±)f^{\pm}(1-f^{\pm}) where + is the sign of η±\eta_{\pm}. Here, f+f^+ is the Marzari-Vanderbilt smearing function, and f−=1−f+f^- = 1-f^+.

For all total energy and ground-state density calculations—DFT, DFT+U, and generalized Hubbard—we use a 12×12×1212 \times12 \times12 k-points Monkhorst-Pack grid. Specifically, this k-grid is used to compute EtotE_{\mathrm{tot}} in the lines “DFT, DFT+U, This work” of Table I, as well as the equation of state of Figure 2 and the charge-density differences of Figure S4. With the chosen values we assure a 10 meV (or better) convergence on the total-energy in DFT.

Charge-density difference $\rho^{\mathrm{DFT}}-\rho^{\mathrm{DFT+U}}$

(a)

Charge-density difference $\rho^{\mathrm{DFT}}-\rho^{\mathrm{dynH}}$

(b)

Charge-density difference $\rho^{\mathrm{DFT+U}}-\rho^{\mathrm{dynH}}$

(c)

Figure S4. Comparison between charge-density differences from different methods along the 100 plane on SrVO3_3. In the cell, vanadium atoms are on the corners, oxygen atoms are on the midpoint of each edge, and strontium is at the center of the cell and cannot be seen from the plot since in the center of the cell and we are cutting with a plane on the surface of the cell. All calculations are done at the experimental equilibrium (T=0T = 0 limit) lattice parameter of SrVO3_3, see the text for more details. The color-coding is red-green-blue, with red being positive, green zero, blue negative. Same color for different plots are the same value. For example, in panel (b) the zone where is red ρDFT\rho^{\mathrm{DFT}} is higher then ρdynH\rho^{\mathrm{dynH}}, where is blue is viceversa, and where is green the system has the same electronic charge density.

Due to the kk-point sampling, the localized-GW approach needs an additional “pole-condensation” procedure. The issue originates from the number of poles of the self-energy, which scales linearly with the dimension of the HAIMH_{\mathrm{AIM}} that we need to diagonalize. The self-energy poles are calculated from Eq. (8), i.e. using the SOP form for GG and UU. Specifically, the number of poles of Σ\Sigma is equal to the number of poles of GG times those of UU. While UU brings in a number of poles in the order of 10, the projected KS Green’s function has poles at all KS states with residues as projections over the Hubbard manifold,

Gmm′KS(ω)=∑k,n⟨m∣ψnk⟩⟨ψnk∣m′⟩ω−ϵnk+iη±,(42)G^{\mathrm{KS}}_{mm'}(\omega) = \sum_{\mathbf{k},n} \frac{\langle m \vert\psi_{n\mathbf{k}} \rangle\langle\psi_{n\mathbf{k}} \vert m' \rangle}{\omega- \epsilon_{n\mathbf{k}} + i\eta_{\pm}}, \tag*{(42)}

where ∣m⟩\vert m \rangle span the Hubbard manifold, and ψnk,ϵnk\psi_{n\mathbf{k}}, \epsilon_{n\mathbf{k}} are the Kohn-Sham orbitals and eigenvalues. The sum is performed on all the Kohn-Sham states. Thus, with a 12×12×1212 \times12 \times12 grid and 25 orbitals in the calculation, one gets more than 10000 poles (without enforcing point-group symmetries). As it happens for the density of states, when projected to the localized manifold most of these poles effectively overlap, giving a structure that can be then represented with much fewer poles.

In order to condense the poles we adopt a strategy inspired from the weighted-mean of NN random variables. Indeed, one can see the spectral function of a sum-over-poles representation of a propagator, i.e., the absolute value of the imaginary part, as a distribution function. In SOP, this must be a sum of Lorentzians [23] each centered at ϵnk\epsilon_{n\mathbf{k}} and broadened by η\eta. Then, if the overlap of two Lorentzians is greater than a threshold the two poles can be regarded effectively as one. In particular:

  • We sequentially browse the (ordered) pole array.

  • We compare the weighted-mean of the first two (real part of the) poles — having a weight equal to each broadening — and compare it with the first.

  • If the separation between the position (real part) of the two poles is less than a threshold, we replace them with a newly defined pole, with a position equal to the weighted mean, broadening as the weighted mean variance, and residue (or amplitude [23]) as the sum of the two.

  • We continue with the rest of the poles.

We do this procedure separately for the four quarters of the complex plane, defined by the real axis and the position of the chemical potential. Please note that this procedure preserves the normalization of the Green’s function — since we are summing the amplitudes —, and preserves the total number of particles, since we are doing it for occupied and empty poles separately. We check that the value chosen for the threshold changes the total energy for less than 10 meV. With this procedure, we are able to reduce the number of poles to a few hundreds (without symmetries), making the diagonalization of HAIMH_{\mathrm{AIM}} computationally much cheaper.

With 25 Kohn and Sham orbitals in the calculation, the dimension of HAIMH_{\mathrm{AIM}} would then be ∼25×100\sim25 \times100. Importantly, owing to the SVD factorization of the Γm\Gamma_m matrices, Γm=VmV~m†\Gamma_m = V_m \tilde{V}_m^\dagger, only non-zero singular values of Γm\Gamma_m count to define the rectangular shape of VmV_m and V~m\tilde{V}_m, and, in turn, the overall dimension of HAIMH_{\mathrm{AIM}}. As mentioned in the main text, this allows for the dimension of HAIMH_{\mathrm{AIM}} to be reduced to dim⁡[HAIM]=dim⁡[h0]+∑mrank⁡[Γm]\dim[H_{\mathrm{AIM}}] = \dim[h_0] + \sum_m \operatorname{rank}[\Gamma_m]. In the calculations for SrVO3_3 the rank of the Γm\Gamma_m’s are at maximum 5, in some cases 3, and in a few 0. This is because the correction on the non t2gt_{2g} and ege_g bands is basically negligible. Overall, the most expensive part of the calculation is the evaluation of the screening potential U(ω)U(\omega) within RPA. Indeed, while for the 12×12×1212 \times12 \times12 kk-mesh, the CPU time for this calculation takes approximately ∼100\sim100 CPU hours, our PYTHON implementation takes approximately ∼12\sim12 hours to deliver all the quantities (here, most of the computational time is spent in diagonalizing HAIMH_{\mathrm{AIM}}).

In addition to the total energy, within these settings we also calculate the density matrix, then used e.g. in Figure S4, using a chemical potential that solves:

N=∫−∞μdωTr⁡A(ω),(43)N = \int_{-\infty}^{\mu} \mathrm{d}\omega\operatorname{Tr} A(\omega), \tag*{(43)}

where A(ω)A(\omega) is the spectral function and the integral is done analytically thanks to the SOP form of GG. For the spectral properties — Figure 1, and the values in the lines “DFT, DFT+U, This work” of Table I — we use a denser Monkhorst-Pack grid of 16×16×1616 \times16 \times16 kk-points. For DFT and DFT+U we use the self-consistent density calculated with the coarser grid. For generalized Hubbard, according to [23], we solve the Dyson equation,

G−1(ω)=G0−1(ω)−Σ(ω)+μ,(44)G^{-1}(\omega) = G_0^{-1}(\omega) - \Sigma(\omega) + \mu, \tag*{(44)}

with μ\mu calculated from Eq. (43), on the kk-path of Figure 1, using again the algorithmic inversion. The bandwidth in Table I is calculated at the Γ\Gamma point, considering the peak (in frequency) of the spectral function, for the dynamical Hubbard approach. The numerical value of the mass enhancement factor is computed by dividing the full bandwidth of the method, by the DFT value.

Satellites, self-energy, and density of states from the dynamical Hubbard functional

In this section, we discuss the origin of the satellites in the spectral function with the help of the self-energy and the density of states (DOS). First, we compare the DOS of the t2gt_{2g} bands in Figure S2 to the self-energy of the local t2gt_{2g} manifold (orthogonalized dxyd_{xy}, dyzd_{yz}, and dxzd_{xz} orbitals) in Figure S3. It is important to note that in this material, we can directly link the (kk-spanned) t2gt_{2g} bands with the local manifold since they do not hybridize with anything else, and hKSh_{\mathrm{KS}} is essentially diagonal and degenerate for the 3d3d local orbitals. In practice, this makes the self-energy correction for the t2gt_{2g} bands diagonal and kk independent, allowing for a clear link between the incoherent features of the DOS in Figure S2 and the local self-energy in Figure S3. It is not surprising that the separation between LS, US1, and US2 is exactly reproduced in the three features of the self-energy. As mentioned in the main text, determining the center of the LS satellite is difficult since it possesses a large broadening. While the US1 satellite matches well with the EDMFT and GWGW+C results (see Table I in the main text), the US2 satellite corresponds to a maximum of Im⁡Σ\operatorname{Im}\Sigma and not to a minimum, as in GWGW calculations [52]. Therefore, it cannot be interpreted as a plasmaron. Since the erroneous plasmaron of GWGW is corrected by GWGW+C calculations [52], it would be of interest to apply GWGW+C to our localized self-energy to see if US2 merges with US1, as it does for GWGW+C.

Figure S2. Density of states of the t2gt_{2g} bands of SrVO3_3 from the calculations of this work. The lower satellite (LS), quasi-particle occupied bandwidth (BW), full bandwidth (full), upper satellite (US), and upper plasmonic satellite (USP) are marked with grey dashed lines [80]. The frequency axis is aligned with the Fermi energy resulting from the calculations. Please note we have placed the labels BW and full to match the peak of the DOS for the lowest and highest quasi-particle t2gt_{2g} bands.

Self-energy on the localized $t_{2g}$ manifold

Figure S3. Self-energy on the localized t2gt_{2g} manifold with satellites labelled as in Figure S2. The frequency axis is aligned with the Kohn-Sham Fermi energy. Note that the US2 vertical line has been moved of 0.1 eV with respect to Figure S2 to match the peak in Im⁡Σ\operatorname{Im}\Sigma

Charge rearrangement within dynamical Hubbard

To get a qualitative picture concerning the lowering of the bulk-modulus of SrVO3_3 when applying the localized GW approach, in Figure S4 we compare charge-density differences from different methods along the (100) plane. On the plane, vanadium atoms are on the corners, and oxygens on the midpoint of the cell. The color-code in the plots is red-green-blue, with red being positive, green zero, blue negative. In panel S4a, one can see the comparison between the DFT and DFT+U charge densities. Looking at the left corner, one can see that there is no qualitative change in the charge density at the middle of the bond between the vanadium and the oxygen atoms. Conversely, in the other two plots, the DFT (panel S4b) or the DFT+U (panel S4c) charge in the middle is pulled toward the atoms. Effectively, this can account for a lowering of the bulk-modulus as obtained by the localized-GW approach with respect to DFT and DFT+U. Also, since the Hubbard manifold is around the vanadium atom, no changes in the charge density on the strontium are observed.

References

  1. [1]N. Marzari, A. Ferretti, and C. Wolverton, Nature Materials 20, 736 (2021).
  2. [2]P. Hohenberg and W. Kohn, Physical Review 136, B864 (1964).
  3. [3]W. Kohn and L. J. Sham, Physical Review 140, A1133 (1965).
  4. [4]J. P. Perdew, W. Yang, K. Burke, Z. Yang, E. K. U. Gross, M. Scheffler, G. E. Scuseria, T. M. Henderson, I. Y. Zhang, A. Ruzsinszky, H. Peng, J. Sun, E. Trushin, and A. Görling, Proceedings of the National Academy of Sciences 114, 2801 (2017).
  5. [5]L. Reining, Wiley Interdisciplinary Reviews: Computational Molecular Science 8, e1344 (2018).
  6. [6]A. Ferretti, I. Dabo, M. Cococcioni, and N. Marzari, Physical Review B 89, 195134 (2014).
  7. [7]G. Kotliar, S. Y. Savrasov, K. Haule, V. S. Oudovenko, O. Parcollet, and C. A. Marianetti, Reviews of Modern Physics 78, 865 (2006).
  8. [8]L. Hedin and S. Lundqvist, in Solid State Physics, Vol. 23, edited by F. Seitz, D. Turnbull, and H. Ehrenreich (Academic Press, 1970) pp. 1–181.
  9. [9]A. Georges, G. Kotliar, W. Krauth, and M. J. Rozenberg, Reviews of Modern Physics 68, 13 (1996).
  10. [10]S. Y. Savrasov and G. Kotliar, Physical Review B 69, 245101 (2004).
  11. [11]M. Gatti, V. Olevano, L. Reining, and I. V. Tokatly, Physical Review Letters 99, 057401 (2007).
  12. [12]F. Aryasetiawan, L. Hedin, and K. Karlsson, Physical Review Letters 77, 2268 (1996).
  13. [13]F. Caruso, C. Verdi, S. Poncé, and F. Giustino, Physical Review B 97, 165113 (2018).
  14. [14]J. S. Zhou, L. Reining, A. Nicolaou, A. Bendounan, K. Ruotsalainen, M. Vanzini, J. J. Kas, J. J. Rehr, M. Muntwiler, V. N. Strocov, F. Sirotti, and M. Gatti, Proceedings of the National Academy of Sciences 117, 28596 (2020).
  15. [15]J. M. Luttinger and J. C. Ward, Physical Review 118, 1417 (1960).
  16. [16]G. Baym and L. P. Kadanoff, Physical Review 124, 287 (1961).
  17. [17]G. Baym, Physical Review 127, 1391 (1962).
  18. [18]R. M. Martin, L. Reining, and D. M. Ceperley, “Interacting Electrons: Theory and Computational Approaches,” (2016).DOI
  19. [19]G. Stefanucci and R. v. Leeuwen, “Nonequilibrium Many-Body Theory of Quantum Systems: A Modern Introduction,” (2013).
  20. [20]C.-O. Almbladh, U. V. Barth, and R. V. Leeuwen, International Journal of Modern Physics B 13, 535 (1999).
  21. [21]N. E. Dahlen and U. V. Barth, Physical Review B 69, 195102 (2004).
  22. [22]K. Haule and G. L. Pascut, Physical Review B 94, 195146 (2016).
  23. [23]T. Chiarotti, N. Marzari, and A. Ferretti, Physical Review Research 4, 013242 (2022).
  24. [24]B. Holm and U. von Barth, Physical Review B 57, 2108 (1998).
  25. [25]M. Puig von Friesen, C. Verdozzi, and C.-O. Almbladh, Phys. Rev. B 82, 155108 (2010).
  26. [26]S. Di Sabatino, P.-F. Loos, and P. Romaniello, Front. Chem. 9 (2021), 10.3389/fchem.2021.751054.
  27. [27]A. Honet, L. Henrard, and V. Meunier, AIP Advances 12, 053238 (2022).
  28. [28]A. Kute pov, S. Y. Savrasov, and G. Kotliar, Physical Review B 80, 041103 (2009).
  29. [29]A. Kutepov, K. Haule, S. Y. Savrasov, and G. Kotliar, Physical Review B 85, 155129 (2012).
  30. [30]M. Grumet, P. Liu, M. Kaltak, J. c. v. Klimeš, and G. Kresse, Phys. Rev. B 98, 155143 (2018).
  31. [31]C.-N. Yeh, S. Iskakov, D. Zgid, and E. Gull, Physical Review B 106, 235104 (2022).
  32. [32]A. L. Kutepov and G. Kotliar, Physical Review B 96, 035108 (2017).
  33. [33]K. Haule and T. Birol, Physical Review Letters 115, 256402 (2015).
  34. [34]C. P. Koçer, K. Haule, G. L. Pascut, and B. Monserrat, Physical Review B 102, 245104 (2020).
  35. [35]S. L. Dudarev, G. A. Botton, S. Y. Savrasov, C. J. Humphreys, and A. P. Sutton, Physical Review B 57, 1505 (1998).
  36. [36]G. E. Engel, B. Farid, C. M. M. Nex, and N. H. March, Physical Review B 44, 13356 (1991).
  37. [37]See Supplemental Material for computational details and further discussion of the theoretical approach adopted.
  38. [38]S. Güttel and F. Tisseur, Acta Numerica 26, 1 (2017).
  39. [39]Y. Su and Z. Bai, SIAM Journal on Matrix Analysis and Applications 32, 201 (2011).
  40. [40]Note that for simplicity of notation we dropped the identity on the m-th bath, the correct Hamiltonian is ΩₘIₘ.
  41. [41]E. Jarlebring, M. Benndrich, G. Mele, E. Ringh, and P. Upadhyaya, “NEP-PACK: A Julia package for nonlinear eigenproblems,” (2018), https://github.com/nep-pack, arXiv:1811.09592.
  42. [42]S. Y. Savrasov, K. Haule, and G. Kotliar, Physical Review Letters 96, 036404 (2006).
  43. [43]S. J. Bintrim and T. C. Berkelbach, The Journal of Chemical Physics 154, 041101 (2021).
  44. [44]O. J. Backhouse, A. Santana-Bonilla, and G. H. Booth, The Journal of Physical Chemistry Letters 12, 7650 (2021).
  45. [45]S. J. Bintrim and T. C. Berkelbach, The Journal of Chemical Physics 156, 044114 (2022).
  46. [46]C. J. C. Scott, O. J. Backhouse, and G. H. Booth, The Journal of Chemical Physics 158, 124102 (2023).
  47. [47]B. Himmetoglu, A. Floris, S. de Gironcoli, and M. Cococcioni, International Journal of Quantum Chemistry 114, 14 (2014).
  48. [48]T. Miyake, C. Martins, R. Sakuma, and F. Aryasetiawan, Physical Review B 87, 115110 (2013).
  49. [49]M. Vanzini and N. Marzari, “Towards a minimal description of dynamical correlation in metals,” (2023), arXiv:2309.12144 [cond-mat].arxiv.org/abs/2309.12144
  50. [50]H. Jiang, R. I. Gomez-Abal, P. Rinke, and M. Scheffler, Physical Review B 82, 045108 (2010).
  51. [51]The spectral function is calculated as \(A(\omega) = \operatorname{sgn}(\mu-\omega)\operatorname{Im}G(\omega)\).
  52. [52]M. Gatti and M. Guzzo, Physical Review B 87, 155147 (2013).
  53. [53]L. Boehnke, F. Aryasetiawan, and P. Werner, Physical Review B 94, 201106 (2016).
  54. [54]S. Biermann, F. Aryasetiawan, and A. Georges, Physical Review Letters 90, 086402 (2003).
  55. [55]T. Ayral, S. Biermann, and P. Werner, Physical Review B 87, 125149 (2013).
  56. [56]J. P. Perdew, K. Burke, and M. Ernzerhof, Physical Review Letters 77, 3865 (1996).
  57. [57]J. P. Perdew, A. Ruzsinszky, G. I. Csonka, O. A. Vydrov, G. E. Scuseria, L. A. Constantin, X. Zhou, and K. Burke, Physical Review Letters 100, 136406 (2008).
  58. [58]J. M. Tomczak, M. Casula, T. Miyake, and S. Biermann, Physical Review B 90, 165138 (2014).
  59. [59]R. Sakuma, P. Werner, and F. Aryasetiawan, Physical Review B 88, 235110 (2013).
  60. [60]T. Yoshida, K. Tanaka, H. Yagi, A. Ino, H. Eisaki, A. Fujimori, and Z.-X. Shen, Physical Review Letters 95, 146404 (2005).
  61. [61]M. Takizawa, M. Minohara, H. Kumigashira, D. Toyota, M. Oshima, H. Wadati, T. Yoshida, A. Fujimori, M. Lippmaa, M. Kawasaki, H. Koinuma, G. Sordi, and M. Rozenberg, Physical Review B 80, 235104 (2009).
  62. [62]T. Maekawa, K. Kurosaki, and S. Yamanaka, Journal of Alloys and Compounds 426, 46 (2006).
  63. [63]F. Birch, Physical Review 71, 809 (1947).
  64. [64]Y. Pavlyukh, A.-M. Uimonen, G. Stefanucci, and R. van Leeuwen, Physical Review Letters 117, 206402 (2016).
  65. [65]N. Marzari, A. A. Mostofi, J. R. Yates, I. Souza, and D. Vanderbilt, Reviews of Modern Physics 84, 1419 (2012).
  66. [66]I. Timrov, F. Aquilante, L. Binc i, M. Cococcioni, and N. Marzari, Physical Review B 102, 235159 (2020).
  67. [67]W. H. Press, S. A. Teukolsky, W. T. Vetterling, and B. P. Flannery, Numerical Recipes 3rd Edition: The Art of Scientific Computing (Cambridge University Press, 2007).
  68. [68]V. I. Anisimov, F. Aryasetiawan, and A. I. Lichtenstein, Journal of Physics: Condensed Matter 9, 767 (1997).
  69. [69]F. Aryasetiawan, M. Imada, A. Georges, G. Kotliar, S. Biermann, and A. I. Lichtenstein, Physical Review B 70, 195104 (2004).
  70. [70]M. Cococcioni and S. de Gironcoli, Physical Review B 71, 035105 (2005).
  71. [71]H. J. Kulik, M. Cococcioni, D. A. Scherlis, and N. Marzari, Physical Review Letters 97, 103001 (2006).
  72. [72]P. Giannozzi, S. Baroni, N. Bonini, M. Calandra, R. Car, C. Cavazzoni, D. Ceresoli, G. L. Chiarotti, M. Cococcioni, A. D. Corso, S. d. Gironcoli, S. Fabris, G. Fratesi, R. Gebauer, U. Gerstmann, C. Gougoussis, A. Kokalj, M. Lazzeri, L. Martin-Samos, N. Marzari, F. Mauri, R. Mazzarello, S. Paolini, A. Pasquarello, L. Paulatto, C. Sbraccia, S. Scandolo, G. Sclauzero, A. P. Seitsonen, A. Smogunov, P. Umari, and R. M. Wentzcovitch, Journal of Physics: Condensed Matter 21, 395502 (2009).
  73. [73]K. Nakamura, Y. Yoshimoto, Y. Nomura, T. Tadano, M. Kawamura, T. Kosugi, K. Yoshimi, T. Misawa, and Y. Motoyama, Computer Physics Communications 261, 107781 (2021).
  74. [74]D. R. Hamann, Physical Review B 88, 085117 (2013).
  75. [75]M. J. van Setten, M. Giantomassi, E. Bousquet, M. J. Verstraete, D. R. Hamann, X. Gonze, and G. M. Rignanese, Computer Physics Communications 226, 39 (2018).
  76. [76]A. Togo and I. Tanaka, Scripta Materialia 108, 1 (2015).
  77. [77]A. Togo, Journal of the Physical Society of Japan 92, 012001 (2023).
  78. [78]Y. C. Lan, X. L. Chen, and M. He, Journal of Alloys and Compounds 354, 95 (2003).
  79. [79]N. Marzari, D. Vanderbilt, A. De Vita, and M. C. Payne, Physical Review Letters 82, 3296 (1999).
  80. [80]Note that, due to the finite lifetimes, we estimate the occupied bandwidth and effective mass, i.e. the position of the first \(t_{2g}\) quasiparticle peak (BW) and last peak (full), as the center of the peak of the DOS. In the figure we do not use an extra broadening for the plotting.

Paper details

vlnc:W_77bk7dv8
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents