Cluster Hubbard Model

• Cluster Hubbard Model is a framework that reformulates the traditional Hubbard Hamiltonian using finite clusters or embedded clusters to capture short-range correlations and frustration.
• It applies advanced techniques like VCA and DCA to resolve magnetic orders, Mott transitions, and spectral properties through methods such as exact diagonalization and self-energy functional theory.
• The approach has practical applications in modeling frustrated lattices and molecular systems, providing insights into Mottness, competing phases, and finite-cluster effects.

Searching arXiv for recent and foundational papers on cluster Hubbard models and cluster methods. In the literature considered here, the expression cluster Hubbard model denotes two closely related constructions. In one usage, it is a finite Hubbard-type Hamiltonian defined on a cluster, such as an 8-site cube at quarter filling or a 4-site molecular plaquette, studied by exact diagonalization or related methods (Szałowski et al., 2015). In the other, it is a cluster representation of an extended lattice problem, where an infinite Hubbard system is tiled into identical clusters or mapped to an embedded cluster whose self-energy, cumulants, or effective vertices are then used to reconstruct lattice properties (Misumi et al., 2015). Both usages preserve the central Hubbard ingredients—kinetic hopping and local repulsion—while promoting the cluster to the fundamental object for describing short-range correlations, frustration, Mottness, and competing ordered states (Dang et al., 2014).

1. Conceptual scope and formal definitions

At its most elementary level, the Hubbard model is written as

$$H = -\sum_{\langle ij\rangle,\sigma} t_{ij}\, c^\dagger_{i\sigma} c_{j\sigma} + U\sum_i n_{i\uparrow} n_{i\downarrow},$$

with $c^\dagger_{i\sigma}$ and $c_{i\sigma}$ fermionic creation and annihilation operators, $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$, hopping amplitudes $t_{ij}$, and on-site repulsion $U$ (Dang et al., 2014). Cluster formulations preserve this structure but reorganize it around a finite subset of sites. Depending on the problem, additional one-body and interaction terms are retained: anisotropic next-nearest-neighbor hoppings $t_2,t_2'$ on frustrated square and triangular geometries (Misumi et al., 2015), intersite Coulomb terms $V$ and next-neighbor hopping $t_2$ on cubic clusters (Szałowski et al., 2016), Kondo exchange $J_\perp$ to localized moments in Hubbard–Kondo lattices (Faye et al., 2018), or distinct bandwidths and frustrations in multi-orbital settings (Lee et al., 2011).

The conceptual distinction between a finite-cluster Hamiltonian and a cluster embedding is methodologically important. In the finite-cluster sense, the cluster itself is the physical system, so thermodynamics and spectroscopy reflect a discrete many-body spectrum and do not imply thermodynamic phase transitions (Szałowski et al., 2015). In the embedding sense, the cluster is a reference problem, impurity, or seed whose exact or quasi-exact solution supplies the local or short-ranged building blocks—self-energies, cumulants, or interaction-irreducible vertices—from which an approximation to the infinite lattice is constructed (Lira et al., 2022). This suggests that “cluster Hubbard model” is less a single Hamiltonian than a family of cluster-centered formulations for strongly correlated fermions.

2. Cluster geometries and representative Hamiltonians

The range of cluster constructions in the literature is broad, but several geometries recur because they match characteristic ordering vectors, local motifs, or experimental realizations.

Setting	Cluster structure	Characteristic role
Frustrated square-lattice Hubbard model	12-site cluster	Accommodates Néel, collinear, and 120° spiral orders
Triangular-lattice DCA	$c^\dagger_{i\sigma}$0 clusters	Tracks frustration-dependent Mott transition
Cubic-cluster Hubbard model	8-site cube	Exact thermodynamics on a finite 3D cluster
Molecular extended Hubbard model	4-site diamond or clover	Real-space extended Hubbard realization

In the frustrated square-lattice model, the Hamiltonian includes isotropic nearest-neighbor hopping $c^\dagger_{i\sigma}$1, anisotropic next-nearest-neighbor hoppings $c^\dagger_{i\sigma}$2 and $c^\dagger_{i\sigma}$3, on-site repulsion $c^\dagger_{i\sigma}$4, and chemical potential $c^\dagger_{i\sigma}$5 chosen for half filling. By tuning $c^\dagger_{i\sigma}$6 and $c^\dagger_{i\sigma}$7, the model interpolates continuously among the unfrustrated square lattice, the crossed-square lattice, and the isotropic triangular lattice (Misumi et al., 2015). In the strong-coupling limit, it maps to an anisotropic frustrated spin-$c^\dagger_{i\sigma}$8 Heisenberg model with $c^\dagger_{i\sigma}$9, $c_{i\sigma}$0, and $c_{i\sigma}$1 (Misumi et al., 2015).

The anisotropic triangular-lattice Hubbard model uses a single-band Hamiltonian at half filling with anisotropic nearest-neighbor hoppings $c_{i\sigma}$2 and $c_{i\sigma}$3, where $c_{i\sigma}$4 controls the degree of geometrical frustration from square-lattice to isotropic triangular limits (Dang et al., 2014). In multi-orbital variants, one orbital may be unfrustrated and another frustrated, or the two orbitals may simply have unequal bandwidths, producing orbital differentiation already at the one-body level (Lee et al., 2011).

Finite-cluster realizations make the geometry itself part of the physics. The 8-site cubic cluster at quarter filling uses the standard or extended Hubbard Hamiltonian on the vertices of a cube, with 4 electrons in 16 spin-orbitals and Hilbert-space dimension $c_{i\sigma}$5 (Szałowski et al., 2015). The extended version adds second-nearest-neighbor hopping $c_{i\sigma}$6 and nearest-neighbor Coulomb repulsion $c_{i\sigma}$7 to the basic nearest-neighbor hopping $c_{i\sigma}$8 and on-site $c_{i\sigma}$9 (Szałowski et al., 2016). In molecular systems, four PTCDA anions on NaCl/Ag(111) realize a 4-site extended Hubbard model with site-dependent on-site potentials, asymmetric hoppings $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$0, and intersite Coulomb terms $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$1, with $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$2 and occupation asymmetry controlled by intersite terms and on-site potentials rather than by $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$3 itself (Tong et al., 6 Sep 2025).

3. Cluster formulations and computational frameworks

A first major class is built on self-energy functional theory. In the variational cluster approximation (VCA), the infinite lattice is tiled into identical clusters, the reference system consists of decoupled interacting clusters with the same $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$4, and the lattice grand potential is treated as a functional of the cluster self-energy,

$n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$5

Stationary points with respect to cluster parameters define the approximation (Misumi et al., 2015). Weiss fields permit explicit competition among Néel, collinear, and spiral states on the same 12-site cluster (Misumi et al., 2015). A finite-temperature extension combines VCA with thermal pure quantum states, replacing full diagonalization of thermally populated eigenstates by typicality-based evaluation of $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$6 and $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$7, thereby accessing finite-$n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$8 spectra, entropy, specific heat, and order parameters without Monte Carlo (Nishida et al., 2019). The nonequilibrium VCA extends the same self-energy-functional logic to the Keldysh contour and uses isolated Hubbard dimers as a reference system; in that setting, the time derivative of the Euler equation is numerically tractable and stable, whereas a direct solution of the Euler equation is not (Hofmann et al., 2015).

A second class consists of DMFT-derived cluster theories. In the dynamical cluster approximation (DCA), the Brillouin zone is partitioned into $n_{i\sigma}=c^\dagger_{i\sigma}c_{i\sigma}$9 patches, the lattice self-energy is approximated as piecewise constant on those patches, and a periodic cluster of size $t_{ij}$0 is embedded in a self-consistent bath (Dang et al., 2014). This framework captures short-range nonlocal correlations absent in single-site DMFT and reveals, for example, cluster-size evolution of the Mott transition and sector-selective behavior (Dang et al., 2014). In the anisotropic two-orbital model, four-site DCA yields momentum-dependent orbital-selective physics, while single-site DMFT with antiferromagnetic order emphasizes the consequences of frustration anisotropy for AF metallic states (Lee et al., 2011). DCA$t_{ij}$1 removes the piecewise-constant momentum structure of conventional DCA by constructing a continuous lattice self-energy and then extending that logic to two-particle vertices, which permits more controlled determination of superconducting transition temperatures in the 2D Hubbard model (Staar et al., 2014). eMBEX pushes this further by using the interaction-irreducible multi-boson vertex $t_{ij}$2 of a small embedded cluster as input to lattice Hedin equations, so that short-range correlations are supplied by the cluster while long-range fluctuations are reconstructed diagrammatically on the lattice (Kiese et al., 2024).

A third class works directly with cluster Hamiltonians or cluster wavefunctions rather than a bath self-consistency. The cumulant Green’s functions method diagonalizes a seed cluster of $t_{ij}$3 correlated sites and uses the resulting cumulants to reconstruct the lattice Green’s function without any self-consistent loop; in 1D, the method converges systematically with cluster size for the gap, ground-state energy, and double occupancy (Lira et al., 2022). The cluster slave-spin method factorizes the electron operator into a fermionic spinon and a slave spin, then solves 2-site and 4-site slave-spin clusters self-consistently to treat symmetry-broken phases and strong-coupling charge fluctuations beyond single-site Gutzwiller theory (Lee et al., 2017). A different wavefunction route is cluster mean-field, in which the lattice ground state is written as a tensor product of optimized cluster states and the single-particle basis is variationally rotated; inter-cluster correlations are then added by second-order perturbation theory in the basis of cluster excitations (Jiménez-Hoyos et al., 2015).

4. Correlated phases, Mottness, and competing orders

The frustrated square-lattice Hubbard model provides one of the clearest illustrations of cluster Hubbard phenomenology. At half filling and strong coupling, VCA on a 12-site cluster finds three magnetically ordered Mott insulators—Néel, collinear, and 120° spiral—located near the square, crossed-square, and triangular limits, respectively. Between these ordered regions lies a nonmagnetic insulating phase with finite single-particle gap and vanishing magnetic order parameters, identified as a Mott insulator generated by quantum fluctuations of frustrated spins (Misumi et al., 2015). The transitions from Néel or spiral order to the nonmagnetic insulator are continuous, whereas the collinear-to-nonmagnetic transition is first order; the critical interaction for the Mott transition increases with frustration because frustration suppresses nesting and antiferromagnetism (Misumi et al., 2015).

On the anisotropic triangular lattice, DCA quantum Monte Carlo yields a frustration-dependent first-order Mott metal–insulator transition with coexistence lines $t_{ij}$4 and $t_{ij}$5. As $t_{ij}$6 increases from the square to the isotropic triangular limit, the slope of the low-temperature phase boundary changes sign, and this is tied, through cluster eigenstate analysis, to a change in the nature of the insulating state from one dominated by an AFM-like singlet $t_{ij}$7 to one with substantial weight of a second singlet $t_{ij}$8 and increased low-energy degeneracy (Dang et al., 2014). The paper discusses this within a scenario of a quantum critical Mott transition and a possible quantum spin liquid insulating state in triangular organics (Dang et al., 2014).

Cluster formulations also clarify more elaborate forms of differentiation. In the anisotropic two-orbital Hubbard model, four-site DCA finds a progression from Fermi liquid to non-Fermi liquid to Mott insulator in momentum sector $t_{ij}$9, together with orbital-selective phase transitions. In the complementary single-site DMFT treatment with one frustrated and one unfrustrated orbital, there is an intermediate regime in which one orbital is insulating and the other metallic, and an antiferromagnetic metallic state with small ordered moment emerges as a consequence of frustration anisotropy (Lee et al., 2011). In the Hubbard–Kondo lattice, VCA identifies competition among the Néel antiferromagnetic phase, the Kondo singlet phase, and, away from half filling, a ferrimagnetic phase in which conduction and localized subsystems are both ferromagnetically ordered internally but aligned antiferromagnetically with respect to one another (Faye et al., 2018).

Finite clusters exhibit the same underlying interactions in a different guise. On the 8-site cube at quarter filling, there is no thermodynamic phase transition, but exact diagonalization reveals strong short-range magnetic structure, interaction-driven suppression of double occupancy, and low-temperature anomalies tied to low-lying multiplets (Szałowski et al., 2015). In the extended cubic model, the energy gap between the ground state and first excited state, and hence the temperature positions of specific-heat and susceptibility maxima, depend nontrivially on $U$0 and $U$1, with Schottky-model behavior valid only while the first excited state remains well isolated from higher states (Szałowski et al., 2016).

5. Thermodynamics, spectroscopy, and real-time dynamics

Finite-temperature cluster methods make it possible to separate Mott physics from magnetic ordering scales. In the typicality-based finite-$U$2 VCA for the half-filled square-lattice Hubbard model at $U$3, the spectral function retains a robust Mott gap both below and above the VCA Néel temperature $U$4, while the staggered magnetization vanishes continuously at $U$5 and the entropy and specific heat display the expected signatures of a second-order transition within the approximation (Nishida et al., 2019). In one dimension, the same TPQ+CPT machinery reproduces low-temperature spinon and holon branches in the half-filled Hubbard chain and their thermal evolution, providing a finite-temperature cluster route to spin–charge separation (Nishida et al., 2019).

For literal finite clusters, thermodynamics is governed by the discrete many-body spectrum. The 8-site cubic cluster at quarter filling has a 15-fold degenerate ground state at $U$6 and a twofold degenerate singlet ground state for any $U$7 in the studied range, with entropy flowing from $U$8 or $U$9 at $t_2,t_2'$0 to $t_2,t_2'$1 at $t_2,t_2'$2 (Szałowski et al., 2015). The specific heat exhibits a single broad maximum at $t_2,t_2'$3 and a two-peak structure for $t_2,t_2'$4, while the magnetic susceptibility has a single broad maximum and Curie–Weiss-like high-temperature behavior with positive Weiss temperature (Szałowski et al., 2015). Adding $t_2,t_2'$5 and nearest-neighbor $t_2,t_2'$6 shifts these maxima, can generate an intermediate specific-heat peak, and modifies the regime where a two-level Schottky description remains accurate (Szałowski et al., 2016).

Real-time cluster Hubbard dynamics has been addressed by both self-energy and Green-function techniques. In nonequilibrium VCA for the 1D Fermi–Hubbard model, isolated Hubbard dimers serve as the reference system for real-time dynamics after fast changes of hopping parameters, and the method is used to study the double occupancy (Hofmann et al., 2015). A self-consistent nonequilibrium Green functions approach with the generalized Kadanoff–Baym ansatz and Hartree–Fock propagators has likewise been applied to finite Hubbard clusters, where it cures artifacts of prior two-time simulations, substantially speeds up the calculations, and compares favorably with exact diagonalization results available for up to 13 particles (Hermanns et al., 2014). These developments indicate that cluster formulations are not restricted to equilibrium phase diagrams but extend to genuine nonequilibrium correlation dynamics.

6. Benchmarks, limitations, and experimental realizations

The chief strength of cluster Hubbard constructions is the exact or quasi-exact treatment of short-range correlations; the chief limitation is that the cluster geometry restricts what can be represented. In the frustrated square-lattice VCA, the 12-site cluster supports 2-sublattice and 3-sublattice orders on equal footing but cannot access longer-period or incommensurate orders, and the resulting phase boundaries slightly overestimate the stability of Néel order relative to other methods (Misumi et al., 2015). In DCA on frustrated lattices, the fermionic sign problem sharply limits accessible temperatures and cluster sizes, especially at strong frustration, although convergence is faster in the fully frustrated triangular case than in the unfrustrated square case because long-range AFM is suppressed (Dang et al., 2014). In finite-$t_2,t_2'$7 VCA for the 2D square lattice, a finite Néel temperature appears despite the Mermin–Wagner theorem because correlations beyond the finite cluster are treated only at mean-field level; the paper explicitly notes that $t_2,t_2'$8 decreases as cluster size increases (Nishida et al., 2019). In the cumulant Green’s functions method, a finite-cluster phase with partially filled band and negative magnetization under positive field survives up to $t_2,t_2'$9 but tends to disappear as the cluster size increases, and is therefore identified as a finite-cluster effect rather than a thermodynamic phase (Lira et al., 2022).

At the same time, the literature consistently benchmarks cluster methods against controlled reference data. VCA phase boundaries in the frustrated square-lattice model are compared to Heisenberg-model calculations in both square $V$0–$V$1 and anisotropic triangular limits (Misumi et al., 2015). Triangular-lattice DCA is checked against previous DCA-like studies, entropy and specific heat from finite-temperature Lanczos, and qualitative trends from other cluster and variational methods (Dang et al., 2014). Typicality-based VCA reproduces finite-temperature tDMRG spectra in 1D and agrees with earlier finite-$V$2 VCA results in 2D (Nishida et al., 2019). The cumulant Green’s functions method is benchmarked in 1D against thermodynamic Bethe ansatz and quantum transfer matrix results for the gap, the ground-state energy, and the double occupancy (Lira et al., 2022). eMBEX is benchmarked for the half-filled square-lattice Hubbard model against numerically exact diagrammatic Monte Carlo and shows very good agreement in the weak to intermediate-coupling regime (Kiese et al., 2024). Cluster mean-field plus second-order perturbation is tested against the exact Lieb–Wu solution in 1D and AFQMC, UCCSD, and related methods in 2D (Jiménez-Hoyos et al., 2015).

Experimental realizations now make the finite-cluster interpretation literal. Four PTCDA molecular anions on NaCl/Ag(111) are described by a 4-site extended Hubbard Hamiltonian whose occupation and transition energies match non-contact AFM, electrostatic force spectroscopy, and STM/STS data; in the asymmetric diamond cluster, different inter-site electrostatic interaction terms, on-site potentials, and asymmetric hopping terms are required, and with $V$3 the occupation asymmetry is driven by those terms independently of $V$4 (Tong et al., 6 Sep 2025). This development suggests that the cluster Hubbard model is no longer only a computational device: it is also an experimentally realizable finite correlated system, complementary to the embedded-cluster theories that approximate extended lattices.

Taken together, these results define the cluster Hubbard model as a unifying framework for strongly correlated electrons at multiple scales: a finite many-body Hamiltonian when the cluster is the system, and a short-range exact core of a larger approximation when the cluster is embedded in a lattice. Its enduring value lies in the controlled resolution of short-range correlation physics—local singlets, frustration, Mottness, orbital differentiation, and short-wavelength collective modes—together with explicit knowledge of where finite-cluster artifacts, bath approximations, or restricted ordering vectors become decisive.