---
title: Real-Space Cluster Perturbation Theory (rCPT)
url: https://www.emergentmind.com/topics/real-space-cluster-perturbation-theory-rcpt
type: topic
---

# Real-Space Cluster Perturbation Theory (rCPT)

Searching arXiv for recent and foundational papers on real-space cluster perturbation theory and closely related CPT formulations.
Real-space Cluster Perturbation Theory (rCPT) is a non-self-consistent cluster method in which a lattice is partitioned into finite real-space clusters, local and short-range correlations are treated exactly within each cluster, and the infinite lattice is reconstructed by treating inter-cluster hopping perturbatively at the level of Green’s functions. In the formulations surveyed across Hubbard, multi-orbital, topological, inhomogeneous, and non-equilibrium settings, rCPT is used to obtain momentum-resolved spectra, local spectral functions, and, in extended variants, spin and charge susceptibilities. Its characteristic approximation is the reuse of a finite-cluster self-energy, or finite-cluster irreducible vertex, as an embedding object for the infinite system [1607.08861], [2507.22269].

## 1. Conceptual definition and position within quantum cluster methods

rCPT is the real-space formulation of CPT: the interacting lattice problem is solved exactly on a finite cluster, while inter-cluster hopping is reintroduced through a Dyson-like embedding equation. In this sense it sits between simple perturbative schemes and self-consistent cluster methods. Compared to single-site RPA or FLEX, it treats local and short-range correlations non-perturbatively within the cluster. Compared to DMFT, CDMFT, or DCA, it is not self-consistent: the cluster self-energy is that of an isolated cluster, and there is no feedback from the embedding lattice onto the cluster problem. Compared to VCA, it is the “bare” version without variational optimization or Weiss fields unless such fields are explicitly added in an extension [1607.08861], [0806.2690], [1404.1287].

Several standard properties recur across the literature. CPT is exact in the non-interacting limit and in the atomic limit, and its accuracy is improved by increasing cluster size rather than by summing higher-order diagrams in a controlled expansion [0806.2690], [2309.14035]. A corresponding limitation is that long-range correlations are not incorporated into the self-energy beyond the cluster. This suggests that rCPT is most reliable for short-range correlation physics, broad spectral features, and regimes in which the relevant energy scales exceed the cluster level spacing [2309.14035].

Cluster geometry is not a secondary implementation detail but a methodological constraint. For the Kane–Mele–Hubbard model on the honeycomb lattice, the 6-site cluster preserves the point-group symmetry, whereas an 8-site cluster introduces anisotropic and unphysical band structures. The associated conclusion is explicit: only clusters that preserve the point-group symmetries of the lattice should be used for topological properties and quasiparticle dispersions [1404.1287].

## 2. Real-space formulation of the single-particle problem

The formal starting point is a decomposition of the Hamiltonian into intra-cluster and inter-cluster parts,
$$
H = \sum_g H_g^{\mathrm{intra}} + H^{\mathrm{inter}},
$$
or equivalently \(H = H' + V\), where \(H^{\mathrm{inter}}\) or \(V\) contains only inter-cluster one-body hopping, while local Hubbard interactions remain entirely inside the cluster Hamiltonian [1607.08861], [0806.2690]. In multi-orbital realizations, this intra-cluster part may include intra-orbital and inter-orbital Coulomb repulsion, Hund’s exchange, and pair hopping, as in the two-orbital Fe-based model with
$$
U' = U - 2J, \qquad J' = J
$$
under spin-rotational invariance [1607.08861].

The exact cluster Green’s function \(G^{(c)}(\omega)\) is computed in the cluster site-orbital basis, typically by Lanczos exact diagonalization or a Lehmann representation. Inter-cluster hopping enters as a matrix \(V(K)\) or \(V(k)\) in the same basis. The central CPT equation is the matrix Dyson form
$$
\hat G^{-1}(K,\omega)=\hat G^{(c)\,-1}(\omega)-\hat V(K),
$$
equivalently
$$
\hat G(K,\omega)=\left[\hat G^{(c)\,-1}(\omega)-\hat V(K)\right]^{-1}.
$$
In self-energy language, the cluster self-energy \(\Sigma^{(c)}(\omega)\) is used as an approximate lattice self-energy, independent of the superlattice momentum [1607.08861], [0806.2690].

Restoration of approximate lattice translational invariance is achieved by periodization. In the two-orbital Fe-based construction this is written as
$$
G_{lm}(k,\omega)=\frac{1}{N_c}\sum_{i,j=1}^{N_c} G_{il,jm}(k,\omega)\,e^{-ik\cdot(r_i-r_j)},
$$
which reduces the cluster-site/orbital Green’s function to the lattice orbital Green’s function [1607.08861]. The spectral function then follows from
$$
A_{lm}(k,\omega)=-\frac{1}{\pi}\,\mathrm{Im}\,G_{lm}(k,\omega+i\delta),
$$
with the total spectral intensity obtained by summing diagonal orbital components [1607.08861].

When translational symmetry is absent, the same logic is retained in real space rather than momentum space. For the 1D Hubbard model with a parabolic potential, the full Green’s function is constructed directly as
$$
\hat G(z)\approx \left[\bigoplus_{m=1}^{N_c}\left(\hat G^{(m)}(z)\right)^{-1}-\hat t^{(c-c)}\right]^{-1},
$$
and the main observable is the local Green’s function \(G_{ii}(z)\) and the local density of states \(D_i(\omega)\). Because naive real-space CPT generates an artificial \(L\)-periodic “patchwork pattern”, rCPT averages over shifted cluster boundaries,
$$
\bar G_{ii}(z)=\frac{1}{L}\sum_{s=-L/2}^{L/2-1}\Bigl[G_{i+s,i+s}(z)\Bigr]_{v_i\mapsto v_{i+s}},
$$
which suppresses cluster-boundary artifacts without using translational symmetry [2507.22269].

## 3. Cluster solvers, numerical realizations, and algorithmic variants

The canonical solver in rCPT is exact diagonalization. In the tutorial literature, the cluster Green’s function is constructed from Lanczos continued fractions or a Lehmann/Q-matrix representation, with Hilbert-space truncation controlled by symmetries and particle-number sectors [0806.2690]. In realistic multi-orbital settings, such as MnO, the interacting part may be restricted to a correlated subcluster, solved by band Lanczos, and then embedded into a larger orbital cluster before the CPT step [1306.5231].

A major extension replaces ED by DMRG or tDMRG. For the 2D Hubbard model, a \(2\times L\) ladder is solved in real space and time, with \(L\) up to 80 at half filling and 40 at 10% hole doping. The cluster Green’s function is then embedded by a \(2\times2\) CPT equation in the transverse direction, yielding high-resolution 2D spectral functions in good agreement with QMC. In that construction, the cluster is defined in real space, the cluster Green’s function is computed in real space and time, and the CPT step reconstructs the full 2D dispersion from inter-ladder hopping [1512.06102].

Another extension uses DQMC as a cluster solver. In CPT+DQMC, the finite-temperature cluster Green’s function \(G^c_{rr'}(\tau)\) is measured on open clusters, Fourier transformed to Matsubara frequencies, inserted into the standard CPT matrix equation, and analytically continued by Maximum Entropy. The stated advantages over CPT+ED are larger clusters, access to temperature dependence, and applicability to models beyond the ED limit; the paper reports square clusters up to \(8\times8\) sites [2107.11284].

A distinct but structurally related implementation appears in an Anderson impurity solver. There, a large real-frequency bath is optimized by unitary transformations so as to minimize the Frobenius norm of the cluster–environment coupling,
$$
E=\frac{1}{N_T}\sum_{i,j}|T_{ij}|^2,
$$
after which the interacting cluster and non-interacting environment are recombined by CPT on the real-frequency axis,
$$
G^{-1}(\omega)=
\begin{pmatrix}
G_C(\omega)&0\\
0&G_R(\omega)
\end{pmatrix}^{-1}
-V.
$$
This is a one-cluster CPT embedding rather than a lattice tiling, but it uses the same structural approximation: the full self-energy is taken to be the cluster self-energy in the correlated block [1707.08841].

Non-equilibrium real-space CPT extends the formalism to contour-ordered Green’s functions. For the transverse-field Ising model mapped to hard-core bosons, the lattice is partitioned into finite clusters with open boundaries, the Kadanoff–Baym contour Green’s functions are computed on each cluster, and the inter-cluster coupling is inserted through the contour Dyson equation
$$
\hat G(q)=\hat G_0+\hat G_0\bullet \hat V(q)\bullet \hat G(q).
$$
The resulting Volterra structure permits direct time propagation after a quench, and the accuracy window increases with cluster size [1605.00472].

## 4. Extensions to two-particle response functions

The simplest route to two-particle observables is CPT+RPA. In the two-orbital Fe-based model, the “bare” multi-orbital susceptibility tensor is built not from non-interacting propagators, but from CPT-dressed Green’s functions via the spectral representation
$$
G_{lm}(k,i\omega_n)=\int d\omega'\,\frac{A_{lm}(k,\omega')}{i\omega_n-\omega'},
$$
leading to the CPT-renormalized bubble
$$
\chi^{\mathrm{cluster}}_{ll',mm'}(q,\Omega)
=
-\sum_p\iint d\omega' d\omega''\,
A_{ml}(p,\omega')A_{l'm'}(p+q,\omega'')
\frac{f(\omega')-f(\omega'')}{\omega'-\omega''+\Omega+i0^+}.
$$
Spin and charge susceptibilities then follow from multiorbital RPA matrix equations,
$$
\hat\chi^s=\left[\hat1-\hat U^s\hat\chi^{\mathrm{cluster}}\right]^{-1}\hat\chi^{\mathrm{cluster}},\qquad
\hat\chi^c=\left[\hat1+\hat U^c\hat\chi^{\mathrm{cluster}}\right]^{-1}\hat\chi^{\mathrm{cluster}}.
$$
In the weak-coupling regime \(U\le 2\) eV, \(J=U/4\), the resulting self-energy corrections are small, and the calculations support the statement that the rigid band approximation and conventional band-RPA are adequate for this model at small \(U\) [1607.08861].

A more direct two-particle construction appears in cuprates. In spin-CPT and charge-CPT, the cluster transverse spin susceptibility and cluster charge susceptibility are computed directly by ED in Lehmann form and then periodized,
$$
\chi_{\alpha\beta}(\mathbf{k},\Omega)=
\frac{1}{N_c}\sum_{i,j}e^{-i(\mathbf r_i-\mathbf r_j)\cdot \mathbf k}\,
\chi^c_{\alpha\beta,ij}(\Omega),
$$
with an analogous expression for the charge channel. The same work also formulates CPT-RPA for a two-band effective Hubbard model derived from the Emery model, using CPT spectral functions inside the bubble and a multiorbital interaction matrix in the ladder resummation [2305.01882].

A third route approximates the Bethe–Salpeter equation itself. For the Hubbard spin susceptibility, the irreducible vertex is approximated as frequency-local and then by its cluster counterpart, which leads to the two-particle CPT equation
$$
\bm{\chi}_{\mathrm{CPT}}^{-1}(\tilde{\mathbf q},i\omega)
=
\bm{\chi}_{0,\mathrm{CPT}}^{-1}(\tilde{\mathbf q},i\omega)
+
\bm{\chi}_c^{-1}(i\omega)
-
\bm{\chi}_{0,c}^{-1}(i\omega).
$$
This extends the usual one-particle CPT logic from the self-energy to the irreducible vertex. In the 1D half-filled Hubbard benchmarks, the method reproduces the main structure of the spin response, although a small finite-size gap at \(q=\pi\) remains [1908.10361].

## 5. Representative applications

The range of applications is unusually broad: weakly and moderately correlated multi-orbital metals, Mott systems, topological insulators, trapped cold-atom systems, impurity models, and non-equilibrium spin systems have all been cast into an rCPT language. A compact summary is given below.

| System | rCPT role | Representative conclusion |
|---|---|---|
| Two-orbital Fe-based model | CPT self-energy + RPA vertex | For \(U\le 2\) eV, self-energy effects are modest and rigid-band+RPA remains adequate [1607.08861] |
| 1D Hubbard model in a parabolic trap | Real-space local Green’s functions with boundary-shift averaging | The LDOS at each site mirrors that of a homogeneous system with the same local filling [2507.22269] |
| Kane–Mele–Hubbard model | CPT Green’s functions and topological Hamiltonian | The \(U\)–\(t_{KM}\) phase diagram separates QSH and trivial insulating regions [1404.1287] |
| 2D Hubbard model with ladder clusters | tDMRG cluster solver + CPT embedding | Waterfalls, kinks, and pseudogap features are traced to spin–charge scattering [1512.06102] |
| Effective cuprate two-band model | One- and two-particle CPT, spin-CPT, charge-CPT, CPT-RPA | Low-energy incommensurate spin response appears in overdoped regimes [2305.01882] |
| MnO | Multi-orbital realistic CPT | Hubbard splitting of Mn \(d\) states opens an insulating gap [1306.5231] |

In Fe-based materials, rCPT has been used to compute the interacting band structure, spectral functions, Fermi surface, and spin and charge susceptibilities of a two-orbital model with on-site multiorbital Hubbard interactions. For \(U\le 2\) eV the Fermi surface remains close to the non-interacting one, with one hole pocket at \(\Gamma\) and two electron pockets at \(X=(\pi,0)\) and \(Y=(0,\pi)\), while spin susceptibility grows and charge susceptibility decreases with \(U\) [1607.08861].

In trapped 1D Hubbard systems, rCPT was adapted to inhomogeneous lattices by computing local Green’s functions and averaging over multiple cluster boundaries. The principal finding is that the local density of states at site \(i\) in the trap is, after a suitable energy shift, almost identical to the DOS of a homogeneous Hubbard chain at the same local filling. This underlies the “scan calculation” or “one-shot spectroscopy” interpretation of a harmonic trap [2507.22269].

In interacting topological systems, CPT Green’s functions have been combined with the topological Hamiltonian
$$
h_{\mathrm{topo}}(k)\equiv -G^{-1}(k,\omega=0),
$$
and the \(\mathbb Z_2\) invariant has been evaluated either from the sewing-matrix Pfaffian formula or, with inversion symmetry, from the Fu–Kane parity criterion applied to \(h_{\mathrm{topo}}\). For the Kane–Mele–Hubbard model, the QSH region broadens as \(t_{KM}\) increases, and the corresponding ribbon spectra display or lose gapless edge states as the bulk invariant changes [1404.1287].

In the 2D Hubbard model, ladder-based rCPT with a tDMRG cluster solver reconstructs the 2D spectral function from large \(2\times L\) real-space clusters. At half filling, the low-energy dispersion is well described by a mean-field antiferromagnetic form, while at finite doping the waterfall evolves into a kink and a pseudogap-like depletion emerges near the \(X\) point. The interpretation advanced there is that these anomalies are connected to scattering between spin and charge degrees of freedom [1512.06102].

## 6. Accuracy, artifacts, and known limitations

The dominant limitations of rCPT are consistent across its different realizations. First, there is no self-consistency: the cluster self-energy does not adapt to the lattice environment. Second, longer-range correlations are absent from the self-energy beyond the cluster size. Third, spontaneous symmetry breaking is not generated in plain CPT; ordered phases appear only through precursors in susceptibilities or by adding explicit variational or Weiss-field terms in an extension [1607.08861], [0806.2690], [1605.00472].

Finite-size effects are not merely quantitative. In the analysis of Hubbard models, the finite-size level spacing of the cluster is identified as the limiting resolution scale of CPT spectra. The rule of thumb
$$
\eta \approx \frac{4t}{M_C}
$$
is used to smear discrete cluster levels into a smooth spectrum, but this broadening then masks narrow low-energy structures. The stated conclusion is that CPT cannot resolve asymptotic low-energy properties of the metal-insulator transition or the supposed pseudogap regime of the half-filled 2D square lattice at intermediate \(U\le 3t\) with currently tractable cluster sizes [2309.14035].

Broken lattice symmetry at the cluster level can generate qualitatively wrong results. In the honeycomb topological case, clusters that do not preserve the full point-group symmetry produce anisotropic dispersions and inconsistent gap closings at symmetry-related Dirac points. This issue is presented as general and not confined to topological observables [1404.1287].

Loss of translational invariance also creates artifacts in explicitly inhomogeneous calculations. In naive real-space CPT for a smooth trap, local density and LDOS develop an artificial \(L\)-periodic “patchwork pattern” tied to cluster boundaries. Averaging over shifted boundaries is introduced precisely to suppress this artifact, and in the homogeneous 1D benchmark the resulting density profile becomes flat to numerical accuracy [2507.22269].

In non-equilibrium settings, accuracy is limited to a cluster-size-dependent time window. For quenches of the transverse-field Ising model, the characteristic time \(\tau\) increases with cluster size, is quite large for quenches within the ordered or disordered phases, and is smaller for quenches across the equilibrium phase transition [1605.00472].

These limitations do not negate the method’s utility; they delimit it. A plausible implication is that rCPT is best regarded as a short-range, high-structure embedding scheme: it is especially effective when the principal physics is encoded in local multiplets, short-range magnetic or orbital correlations, broad Hubbard-band rearrangements, or local filling variations, and less effective when long-wavelength collective phenomena or asymptotically small gaps dominate [2309.14035], [1607.08861].

Source: https://www.emergentmind.com/topics/real-space-cluster-perturbation-theory-rcpt