Constrained Random Phase Approximation (cRPA)

• Constrained Random Phase Approximation (cRPA) is a method that downfolds electron–electron interactions by separating target and screening polarization to yield effective Hubbard parameters.
• It decomposes the full polarization into contributions from a correlated subspace and the remainder, projecting the resulting interactions onto localized Wannier functions for low-energy models.
• Its accuracy and practical utility depend on subspace definition, treatment of band entanglement, and the dynamic screening effects inherent in complex materials.

Searching arXiv for recent and foundational cRPA papers to support the article. Constrained random phase approximation (cRPA) is a first-principles downfolding framework for computing partially screened electron–electron interactions for a chosen correlated subspace in solids. Its central purpose is to determine Hubbard-model parameters—most prominently the on-site Hubbard interaction $U$ and Hund’s exchange $J$—from the full electronic structure while excluding screening processes that are intended to be treated explicitly in the low-energy many-body model. In its standard formulation, cRPA decomposes the total polarization into contributions internal to the target subspace and the remainder, and defines the effective interaction by screening the bare Coulomb interaction only by the latter (Sasioglu et al., 2011). This construction has become a standard ingredient in DFT+$U$, DFT+DMFT, and related low-energy effective Hamiltonian approaches, while also remaining the subject of ongoing methodological refinement and criticism, particularly in systems with strong band entanglement, small gaps, or strong correlations (Carta et al., 6 May 2025).

1. Definition and formal structure

The defining idea of cRPA is the distinction between the fully screened interaction $W$ of the solid and the partially screened interaction $U$ appropriate for a reduced low-energy model (Sasioglu et al., 2011). In ordinary RPA, screening is described by the polarization $P(\omega)$, and the fully screened interaction is

$W(\omega) = [1 - v P(\omega)]^{-1} v,$

where $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$ is the bare Coulomb interaction (Sasioglu et al., 2011).

For downfolding, the one-particle Hilbert space is partitioned into a correlated subspace and the remainder. In the transition-metal setting emphasized by Şaşıoğlu, Friedrich, and Blügel, the correlated subspace is formed by localized $d$-like Wannier orbitals, while the rest contains the remaining Kohn–Sham states (Sasioglu et al., 2011). The polarization is then decomposed as

$P(\omega) = P_d(\omega) + P_r(\omega),$

where $J$0 contains screening processes internal to the correlated subspace and $J$1 contains all other screening channels (Sasioglu et al., 2011). The cRPA interaction is defined by

$J$2

and the fully screened interaction is recovered by reintroducing the omitted low-energy screening,

$J$3

(Sasioglu et al., 2011).

This separation addresses the double-counting problem that would arise if one used the fully screened interaction directly in a low-energy Hubbard or LDA+DMFT Hamiltonian. The correlated electrons would then screen each other both in the ab initio interaction construction and again in the many-body solution of the reduced model (Sasioglu et al., 2011). The cRPA interaction is therefore not the bare Coulomb matrix element of the localized orbitals, nor the fully screened interaction of the full solid, but an intermediate, model-dependent quantity.

In practical implementations, $J$4 is projected onto a localized basis, typically Wannier functions. For a set of Wannier orbitals $J$5, the matrix elements are

$J$6

(Sasioglu et al., 2011). From these one defines orbital-averaged on-site interactions such as the intra-orbital $J$7, the exchange $J$8, and, under rotational invariance, $J$9 (Sasioglu et al., 2011).

2. Correlated subspaces, Wannierization, and model dependence

The practical content of cRPA depends decisively on how the correlated subspace is defined. This is not a secondary implementation detail but part of the definition of the effective model itself (Reddy et al., 14 Mar 2025). In systems with an isolated set of correlated bands, the target subspace may be identified directly by band index and cRPA is comparatively straightforward. In entangled systems, however, the target orbitals must be defined through projections or Wannierization, and different prescriptions can lead to materially different $U$0 values (Reddy et al., 14 Mar 2025).

A common choice is a maximally localized Wannier function basis. In the transition-metal study of (Sasioglu et al., 2011), the correlated subspace is built from maximally localized Wannier functions constructed from six bands per transition-metal atom, namely five $U$1 bands plus one itinerant $U$2 band, in order to capture both $U$3 character and $U$4–$U$5 hybridization. This permits a parameter-free definition of the correlated subspace even in the presence of strong sp–d entanglement (Sasioglu et al., 2011). In later comparative work, different Wannier windows were shown to generate “frontier orbitals” that are relatively delocalized and “atomic-like orbitals” that are more localized, with correspondingly different bare and screened interactions (Reddy et al., 14 Mar 2025).

For entangled bands, the central technical problem is the construction of $U$6. The MLWF-based prescription introduced in (Sasioglu et al., 2011) assigns to each Bloch state $U$7 a probability

$U$8

that it belongs to the correlated subspace, and weights a transition $U$9 by $W$0 when building $W$1 (Sasioglu et al., 2011). This avoids the arbitrariness of fixed energy-window or band-index cuts and remains well defined under strong hybridization.

More recent comparative analyses distinguish several strategies for defining the correlated polarization in entangled systems: a band method, a projector method, a weighted method, and disentanglement constructions (Reddy et al., 14 Mar 2025). Their numerical differences can be large. In isolated $W$2-manifold systems such as LiMO$W$3, the band method gives systematically larger static screened $W$4 than the projector method because it neglects some hybridization-induced screening channels (Reddy et al., 14 Mar 2025). In strongly entangled systems such as SrMO$W$5, the choice between frontier and atomic-like Wannier functions can even invert qualitative trends across a chemical series (Reddy et al., 14 Mar 2025). This suggests that cRPA does not return a unique material constant; it returns an interaction for a specified low-energy Hamiltonian and orbital representation.

The same issue appears in the comparison between cRPA and linear-response theory. When the interacting subspace is an isolated set of bands and the same maximally localized Wannier functions are used in both methods, cRPA and linear response show good quantitative agreement; by contrast, strong hybridization makes the cRPA construction ambiguous and can lead to unrealistically small $W$6 values (Carta et al., 6 May 2025). A plausible implication is that the consistency of projector definitions across DFT+$W$7, DFT+DMFT, and cRPA calculations is indispensable for transferability.

3. Implementations and representative material results

A foundational implementation for elemental transition metals was given within full-potential LAPW using FLEUR, SPEX, and Wannier90 (Sasioglu et al., 2011). In that work, the effective on-site Coulomb interaction for 3$W$8, 4$W$9, and 5$U$0 transition metals was found to lie between 1.5 and 5.7 eV and to depend on crystal structure, spin polarization, $U$1-electron number, and orbital filling (Sasioglu et al., 2011). The bare direct interaction $U$2 and bare exchange $U$3 increase monotonically across each transition-metal series because of orbital contraction, whereas the screened $U$4 displays a non-monotonic dependence governed by the efficiency of residual screening (Sasioglu et al., 2011).

Crystal structure strongly affects the effective interaction. For example, in bcc V, Nb, and Ta, the orbital-resolved interaction satisfies $U$5 about 0.3 eV larger than $U$6, while in close-packed fcc Ni, Pd, and Pt the difference is only about 0.1 eV (Sasioglu et al., 2011). Spin polarization also matters: for Cr, Fe, Co, and Ni the magnetic state has larger $U$7 than the nonmagnetic state, whereas Mn shows the opposite tendency due to changes in the screening associated with the density of states near the Fermi level (Sasioglu et al., 2011).

The LAPW cRPA implementation was extended to transition-metal oxides in (Vaugier et al., 2012), which emphasized that interaction trends depend qualitatively on the low-energy model. In 3$U$8 perovskites SrMO$U$9 with $P(\omega)$0 V, Cr, Mn, a $P(\omega)$1-only Hamiltonian yields decreasing effective interactions along the series because screening becomes more efficient, whereas a more localized “d–dp Hamiltonian” recovers the generally expected increase with atomic number (Vaugier et al., 2012). The same work found that 4$P(\omega)$2 perovskites SrMO$P(\omega)$3 with $P(\omega)$4 Nb, Mo, Tc exhibit weaker screening than their 3$P(\omega)$5 analogues, leading in a $P(\omega)$6 model to effectively larger $P(\omega)$7 values on 4$P(\omega)$8 shells than on 3$P(\omega)$9 shells (Vaugier et al., 2012). This is a particularly clear example of the competition between orbital localization and screening emphasized in cRPA studies.

A PAW-based cRPA implementation for uranium dioxide and cerium was reported in (Amadon et al., 2014). It allows the interaction to be computed in the same Wannier basis used in DFT+$W(\omega) = [1 - v P(\omega)]^{-1} v,$0 and DFT+DMFT. That work argues that self-consistent determination of the static screened interaction together with a consistent Wannier basis is mandatory for $W(\omega) = [1 - v P(\omega)]^{-1} v,$1-Ce and UO$W(\omega) = [1 - v P(\omega)]^{-1} v,$2, because LDA starting points can place correlated states too close to the Fermi level and thereby artificially enhance screening (Amadon et al., 2014). For $W(\omega) = [1 - v P(\omega)]^{-1} v,$3-Ce, by contrast, the static approximation is described as too drastic because $W(\omega) = [1 - v P(\omega)]^{-1} v,$4 exhibits strong frequency dependence (Amadon et al., 2014). This underscores that cRPA is intrinsically a theory of dynamic interactions even when its output is later reduced to a static scalar.

4. Frequency dependence, dynamic screening, and self-consistency

The cRPA interaction is a function of frequency, $W(\omega) = [1 - v P(\omega)]^{-1} v,$5, not merely a number (Sasioglu et al., 2011). Its low-frequency limit provides the static Hubbard parameter used in DFT+$W(\omega) = [1 - v P(\omega)]^{-1} v,$6 and in many DFT+DMFT applications, but the full frequency structure encodes plasmonic and interband screening processes and can be essential for quantitatively accurate many-body treatments (Amadon et al., 2014).

In the transition-metal study of (Sasioglu et al., 2011), $W(\omega) = [1 - v P(\omega)]^{-1} v,$7 was computed on the real axis. For certain bcc metals such as V, Nb, and Ta, the interaction varies strongly at low frequencies, indicating that a purely static $W(\omega) = [1 - v P(\omega)]^{-1} v,$8 is questionable. For close-packed fcc or hcp metals such as Ni, Pd, and Pt, the low-frequency behavior is much smoother, making a static approximation more plausible (Sasioglu et al., 2011). At energies above the plasmon peak, screening becomes ineffective and $W(\omega) = [1 - v P(\omega)]^{-1} v,$9 approaches the bare interaction $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$0, whereas the exchange $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$1 is only weakly frequency dependent (Sasioglu et al., 2011).

In UO$$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$2, (Amadon et al., 2014) identified subplasmon features in $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$3 associated with $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$4–$$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$5 and O-$$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$6–U-$$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$7 transitions, and showed that opening a gap within DFT+$$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$8 strongly suppresses the low-energy screening, thereby making the static limit more reliable. In $$v(\mathbf{r},\mathbf{r}') = e^2/|\mathbf{r}-\mathbf{r}'|$$9- and $d$0-Ce, the same study found qualitatively different behavior: in $d$1-Ce the low-frequency interaction is about 2 eV while the higher-frequency plateau is about 6 eV, so a single static interaction cannot simultaneously describe quasiparticles and Hubbard bands (Amadon et al., 2014). This motivates DMFT or related solvers with retarded interactions.

The relation between cRPA and self-consistency is methodologically important. If the Kohn–Sham starting point badly represents the correlated subspace, then the cRPA polarization also inherits that deficiency. In UO$d$2 and $d$3-Ce, (Amadon et al., 2014) argues that a self-consistent DFT+$d$4+cRPA loop is necessary to obtain physically meaningful $d$5 values. This suggests that cRPA is not fully decoupled from the one-particle theory on which it is built; rather, it is sensitive to the fidelity of the reference band structure.

5. Accuracy, limitations, and criticisms

The standard cRPA approximation retains only RPA-type direct particle–hole screening and neglects vertex corrections and other interaction channels. This limitation has been scrutinized in controlled model studies. In one-dimensional two- and three-orbital Hubbard models with a target band and one or two screening bands, exact quantum Monte Carlo and constrained functional renormalization group analyses show that cRPA can predict substantial screening where the exact low-energy effective interaction is only weakly screened or even antiscreened (Honerkamp et al., 2018). The discrepancy is traced to cancellation of the cRPA contribution by other one-loop diagrams, especially vertex corrections in the direct particle–hole channel and contributions from crossed particle–hole and particle–particle channels (Honerkamp et al., 2018).

A related DMFT benchmark on multi-orbital Hubbard models concluded that cRPA systematically overestimates screening of the Hubbard $d$6 relevant for DMFT impurity solvers (Han et al., 2018). In those models, a one-band effective Hamiltonian using the bare projected interaction reproduced the full multi-orbital spectra far better than the cRPA-screened model, while the cRPA model remained too metallic near the Mott transition (Han et al., 2018). The same work introduced a local-polarization-based diagnostic and argued that RPA misses local screening signatures that are important in strongly correlated systems (Han et al., 2018).

Another source of inaccuracy is identified in the analysis of downfolding based on cRPA for a three-orbital lattice model with one target band and two screening bands (Shinaoka et al., 2014). There, the violation of the Pauli principle in the standard spin-independent cRPA treatment leads to overscreening effects when the inter-orbital interaction is small (Shinaoka et al., 2014). A spin-dependent variant that restores the Pauli principle removes the overscreening in that limit, although it still does not capture the full dependence on inter-orbital coupling (Shinaoka et al., 2014). This criticism is more specific than the generic statement that vertex corrections are missing; it points to a concrete structural problem in the usual RPA resummation for local Hubbard interactions.

The Emery-model comparison between cRPA and constrained functional RG reaches a similar conclusion in a more cuprate-relevant setting (Han et al., 2020). There, cRPA overscreens the onsite interaction relative to cfRG, the direct particle–hole RPA contribution is nearly canceled by vertex corrections, and cRPA misses the antiscreening that appears in the hole-doped regime (Han et al., 2020). Nevertheless, when the full frequency dependence of the effective interaction is retained in DMFT, both cRPA- and cfRG-based models show the Mott transition at similar values of the charge-transfer energy $d$7, implying that static $d$8 values alone can be misleading (Han et al., 2020).

These criticisms do not imply that cRPA is universally unreliable. They indicate that its accuracy is parameter- and system-dependent, especially in the presence of strong hybridization, small band separations, or strong local correlations. A balanced interpretation is that cRPA supplies a controlled partial resummation of screening diagrams, but not a systematically complete low-energy interaction.

6. Regimes of validity, methodological developments, and contemporary directions

A theoretical justification for cRPA in gapped systems was developed in (Loon et al., 2021). That work argues that when the screening bands are separated from the Fermi level by a sizeable gap, the corresponding Green’s functions are short-ranged in real space and nonlocal vertex corrections are suppressed relative to the RPA bubble series. In that regime, RPA is asymptotically justified for the nonlocal screening part, in a manner described as a real-space analogue of Migdal’s theorem (Loon et al., 2021). The same analysis further argues that when cRPA is built on Kohn–Sham bands in insulators, the leading local excitonic vertex correction is effectively incorporated through the Kohn–Sham gap (Loon et al., 2021). This provides a counterpoint to the model-based critiques: cRPA is on firmer footing when the eliminated subspace is genuinely high-energy and gapped.

Recent work has focused on the practical ambiguities of cRPA and on alternatives or extensions. A Wannier-based comparison between cRPA and linear-response theory showed good agreement when the interacting subspace corresponds to an isolated set of bands, but found that strong hybridization renders cRPA ambiguous and can produce unrealistically small $d$9 values, while linear response remains better behaved (Carta et al., 6 May 2025). This strengthens the view that cRPA is most robust in downfolding problems with a clear energetic separation between target and screening sectors.

A systematic comparison of cRPA methodologies for isolated and entangled $P(\omega) = P_d(\omega) + P_r(\omega),$0-electron systems similarly concluded that the definition of the correlated space and the method used to compute the correlated polarization critically determine both the magnitude and the trend of the resulting Hubbard parameters (Reddy et al., 14 Mar 2025). A plausible implication is that reported discrepancies between published $P(\omega) = P_d(\omega) + P_r(\omega),$1 values often arise less from intrinsic material differences than from differences in model definition.

Methodological innovation has also targeted known numerical pathologies. A 2025 proposal termed spectral cRPA (s-cRPA) introduces a spectral selection of correlated Bloch states based on leverage scores and reports systematically larger $P(\omega) = P_d(\omega) + P_r(\omega),$2 values than weighted or projector cRPA, together with improved numerical stability and elimination of negative interaction values for filled $P(\omega) = P_d(\omega) + P_r(\omega),$3 shells (Kaltak et al., 21 Aug 2025). Applied to CaFeO$P(\omega) = P_d(\omega) + P_r(\omega),$4, s-cRPA yields interaction parameters closer to those needed in DFT+$P(\omega) = P_d(\omega) + P_r(\omega),$5 to obtain the experimentally observed insulating state than standard cRPA does (Kaltak et al., 21 Aug 2025). This suggests that some of the standard underestimation of $P(\omega) = P_d(\omega) + P_r(\omega),$6 may be partly technical, arising from the handling of band entanglement and the long-wavelength limit, rather than purely conceptual.

Across these developments, a common theme is that cRPA remains central because it provides a principled definition of partial screening and frequency-dependent interactions. At the same time, its quantitative reliability depends on band separation, subspace definition, and the importance of beyond-RPA channels. In practice, cRPA is best understood as a framework whose outputs require interpretation in the context of the chosen low-energy model, the basis construction, and, where relevant, comparison with beyond-RPA approaches or alternative response theories (Carta et al., 6 May 2025).