Subbath CDMFT: Efficient Bath Decomposition

• SB-CDMFT is a variant of cluster DMFT that splits the bath into independent subbaths, enabling separate solution of smaller Anderson impurity problems.
• The method reduces the exponential Hilbert-space scaling of ED by lowering the effective bath size while maintaining the overall CDMFT framework.
• Benchmark studies on 1D and 2D Hubbard models demonstrate improved low-energy spectral resolution and accurate reproduction of key physical observables.

Searching arXiv for the cited SB-CDMFT paper and closely related CDMFT bath-parameterization background. Subbath CDMFT (SB-CDMFT) is a variant of cluster dynamical mean-field theory with an exact-diagonalization impurity solver in which the discrete bath is not treated as a single Anderson bath, but is split into several independent subbaths. Each subbath defines its own smaller Anderson impurity problem containing the full interacting cluster plus only a fraction of the bath orbitals. The central approximation is to solve these smaller impurity models separately and combine their outputs—primarily the cluster self-energies and hybridization functions—by averaging. In this way, SB-CDMFT retains the outer CDMFT self-consistency structure while replacing one large impurity problem by multiple smaller separate ones, with the stated purpose of overcoming the exponential Hilbert-space bottleneck of conventional CDMFT-ED (Lagrave et al., 9 Sep 2025).

1. Conceptual position within cluster DMFT

SB-CDMFT is formulated as a modification of the impurity side of cellular dynamical mean-field theory, not as a change in the surrounding cluster-DMFT philosophy. In standard CDMFT, the infinite lattice is tiled into clusters of $N_{\rm c}$ sites, and the lattice self-energy is approximated by the cluster self-energy. The cluster Green’s function is written as

$$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$

while the lattice Green’s function is approximated as

$$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$

The impurity problem is then an Anderson model in which the interacting cluster is coupled to a finite set of uncorrelated bath orbitals, and the bath parameters are determined self-consistently by fitting a finite hybridization function to a target Weiss field on a Matsubara grid (0806.2690).

The bottleneck that motivates SB-CDMFT is the Hilbert-space growth of the ED impurity problem. In conventional CDMFT with ED, a cluster of $N_{\rm c}$ interacting sites is embedded in a bath of $N_{\rm b}$ noninteracting orbitals, and the Hilbert-space dimension scales as

$4^{N_{\rm c}+N_{\rm b}},$

since each spinful orbital has four local states. This makes enlarging the bath rapidly prohibitive even though a larger bath improves the representation of the dynamical environment. SB-CDMFT is designed to allow a much larger effective bath representation at fixed ED cost by solving $N_{\rm sb}$ impurity problems of size $N_{\rm c}+N_{\rm b}/N_{\rm sb}$ instead of one problem of size $N_{\rm c}+N_{\rm b}$ (Lagrave et al., 9 Sep 2025).

This construction has a clear relation to earlier symmetry-structured bath ideas in CDMFT. Prior work on bath parametrization emphasized that the bath Green matrix can be decomposed into irreducible symmetry blocks and that bath sites can be arranged into sets corresponding to those irreducible representations, with each block fitted independently (0804.3320). SB-CDMFT generalizes the idea from a bath organization principle to a full decomposition-and-recombination impurity algorithm.

2. Formal construction

The reference lattice model in the SB-CDMFT paper is the one-band Hubbard model,

$$H=-t\sum_{\langle ij\rangle,\sigma}(c^\dagger_{i\sigma}c_{j\sigma}+\text{h.c.})+U\sum_i n_{i\uparrow}n_{i\downarrow}-\mu\sum_{i,\sigma}n_{i\sigma},$$

with the usual meanings of $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$0, $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$1, and $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$2. In conventional CDMFT, the associated impurity Hamiltonian is

$$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$3

and the hybridization matrix is

$$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$4

The finite bath is fitted to the projected lattice Green function through a distance function written in inverse-Green-function or hybridization form, with Matsubara frequencies defined using a fictitious inverse temperature $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$5 and a weight $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$6 taken constant below a cutoff $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$7 and zero above (Lagrave et al., 9 Sep 2025).

SB-CDMFT partitions the full bath of $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$8 orbitals into $$\mathbf{G}_c^{-1}(z)=z-\mathbf{t}_c-\mathbf{\Sigma}_c(z),$$9 subbaths. For an even split, each subbath contains $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$0 bath orbitals. Each subbath $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$1 defines its own Anderson impurity Hamiltonian,

$$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$2

The cluster is repeated identically in every subbath problem, but only one bath sector is attached at a time. Each of these impurity models has its own hybridization function

$$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$3

and its own cluster Green function

$$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$4

The phrase that each subbath has a distinct hybridization function means that the bath parameters $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$5 and $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$6 are not shared across subbaths. Each subbath represents a different component of the total dynamical environment seen by the cluster. This distinction can be arbitrary or symmetry-based. In symmetry-adapted constructions, the hybridization matrix is block-diagonalized according to irreducible representations of the cluster point group, and each block is treated as a subbath. The paper gives as examples a one-dimensional four-site cluster with mirror symmetry $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$7, where the bath decomposes into symmetric $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$8 and antisymmetric $$\mathbf{G}^{-1}(\tilde{\mathbf{k}},z)=z-\mathbf{t}(\tilde{\mathbf{k}})-\mathbf{\Sigma}_c(z).$$9 sectors, and the $N_{\rm c}$0 plaquette, where four subbaths can correspond to the four irreducible representations of $N_{\rm c}$1: $N_{\rm c}$2, $N_{\rm c}$3, $N_{\rm c}$4, and $N_{\rm c}$5. Symmetry is not required: the method is also reported to work for unconstrained, symmetry-broken bath parametrizations (Lagrave et al., 9 Sep 2025).

3. Recombination and self-consistency

The defining approximation of SB-CDMFT is the recombination of separate impurity solutions by averaging. The default choice is a uniform average of self-energies,

$N_{\rm c}$6

together with a uniform average of hybridization functions,

$N_{\rm c}$7

The factor $N_{\rm c}$8 is motivated by the fact that the cluster appears in every impurity model and should not be counted multiple times. The averaged self-energy is used in the lattice Dyson equation and in the CDMFT self-consistency loop, while the averaged hybridization replaces the conventional $N_{\rm c}$9 in the fit to the target Weiss field (Lagrave et al., 9 Sep 2025).

The authors briefly consider a more general weighted recombination,

$N_{\rm b}$0

with $N_{\rm b}$1 as additional variational parameters, but report severe instabilities and pathological solutions. In the rare converged cases, the optimized $N_{\rm b}$2 reverted to the uniform distribution, so the practical method uses uniform averaging.

Algorithmically, SB-CDMFT remains an iterative self-consistent CDMFT loop. The workflow described in the paper is:

1. Choose the cluster geometry, total bath size $N_{\rm b}$3, number of subbaths $N_{\rm b}$4, and optionally a symmetry-based parametrization of the bath.
2. Initialize all subbath bath parameters $N_{\rm b}$5.
3. For each subbath $N_{\rm b}$6, build the reduced Anderson impurity Hamiltonian $N_{\rm b}$7.
4. Solve each $N_{\rm b}$8 independently by exact diagonalization to obtain $N_{\rm b}$9 and $4^{N_{\rm c}+N_{\rm b}},$0.
5. Form the averaged self-energy $4^{N_{\rm c}+N_{\rm b}},$1.
6. Insert $4^{N_{\rm c}+N_{\rm b}},$2 into the lattice Green function and compute the projected Green function $4^{N_{\rm c}+N_{\rm b}},$3.
7. Construct the target hybridization $4^{N_{\rm c}+N_{\rm b}},$4 from the cluster Dyson relation.
8. Form the combined impurity hybridization $4^{N_{\rm c}+N_{\rm b}},$5.
9. Minimize the distance

$4^{N_{\rm c}+N_{\rm b}},$6

with respect to all subbath parameters simultaneously.

1. Update the bath parameters and repeat until convergence (Lagrave et al., 9 Sep 2025).

Two implementation points are emphasized. First, the ED solutions of the different subbaths are independent and therefore naturally parallelizable. Second, the optimization is not independent subbath by subbath: all variational parameters across all subbaths are fitted together because only their total averaged hybridization is required to match the CDMFT Weiss field. Earlier CDMFT work provides the general background for this finite-bath fitting strategy: the target Weiss field is approximated, not matched exactly, because a finite bath cannot satisfy self-consistency at all frequencies (0806.2690).

4. Scaling, bath structure, and implementation regimes

The stated computational gain is the main practical argument for SB-CDMFT. The ED cost per impurity problem is reduced from

$4^{N_{\rm c}+N_{\rm b}},$7

to

$4^{N_{\rm c}+N_{\rm b}},$8

and because there are $4^{N_{\rm c}+N_{\rm b}},$9 such problems, the total cost scales as

$N_{\rm sb}$0

The paper defines the Hilbert-space reduction factor as

$N_{\rm sb}$1

For $N_{\rm sb}$2 and $N_{\rm sb}$3, this gives $N_{\rm sb}$4. This scaling is used to motivate calculations corresponding to full-bath sizes whose total Hilbert spaces would be of order $N_{\rm sb}$5 and $N_{\rm sb}$6, described as inaccessible to standard ED-CDMFT (Lagrave et al., 9 Sep 2025).

The method has a trivial exact limit: when $N_{\rm sb}$7, SB-CDMFT reduces to ordinary CDMFT-ED. Beyond that limit, the paper states that it should work best when the bath can be naturally decomposed into relatively weakly entangled sectors, especially symmetry sectors, so that separate treatment plus averaging is a good approximation. Deviations are expected when simultaneous multi-sector bath correlations matter strongly, or when the optimization landscape becomes difficult because the number of subbaths and variational parameters becomes too large. Larger subbaths are observed to improve the low-energy fit and spectral resolution, consistent with the idea that the approximation weakens as each subbath becomes a better local representation of its piece of the environment (Lagrave et al., 9 Sep 2025).

The specific implementations studied are:

Model and cluster	Standard bath	SB-CDMFT variants
1D Hubbard, four-site cluster	one 8-site bath	two 4-site subbaths; four 2-site subbaths; four 4-site subbaths
2D Hubbard, $N_{\rm sb}$8 cluster	one 8-site bath	four 2-site subbaths; four 4-site subbaths; four 8-site subbaths
Symmetry-broken two-site test	one 8-site bath	two 4-site subbaths

The bath-fitting uses Matsubara frequencies with fictitious $N_{\rm sb}$9, with $N_{\rm c}+N_{\rm b}/N_{\rm sb}$0 constant below $N_{\rm c}+N_{\rm b}/N_{\rm sb}$1 and zero above. The paper reports $N_{\rm c}+N_{\rm b}/N_{\rm sb}$2 in some 1D benchmarks and $N_{\rm c}+N_{\rm b}/N_{\rm sb}$3 in large-system examples. ED is the impurity solver at each subbath stage, and a small physical temperature $N_{\rm c}+N_{\rm b}/N_{\rm sb}$4 is introduced in doped cases to smooth sector changes (Lagrave et al., 9 Sep 2025).

The broader literature provides two closely related backgrounds. First, symmetry-adapted bath parametrization in CDMFT and DCA already established that bath sites can be arranged into sets corresponding to irreducible representations and fitted block by block (0804.3320). Second, later work on bath parameterization in multi-band cluster DMFT studies a partition of the bath into a collection of smaller bath sets, where each set contains as many bath orbitals as there are impurity orbitals; that work is not SB-CDMFT by name, but it supplies a closely related implementation template for structured bath subsets (Florez-Ablan et al., 6 Jun 2025).

5. Benchmarks and physical content

The validation presented for SB-CDMFT focuses on the one-dimensional Hubbard model, the two-dimensional Hubbard model on a $N_{\rm c}+N_{\rm b}/N_{\rm sb}$5 plaquette, and a symmetry-broken two-site test case. In the one-dimensional four-site cluster benchmark, the comparison is between standard CDMFT with eight bath sites and SB-CDMFT with either two subbaths of four sites or four subbaths of two sites. The observables are local spectral functions and a Kolmogorov–Smirnov metric computed from cumulative spectral functions,

$N_{\rm c}+N_{\rm b}/N_{\rm sb}$6

where the KS number is the maximum absolute difference between cumulative distributions. The spectral functions are reported to show very good agreement, especially at low frequencies, and the Mott gap is reproduced (Lagrave et al., 9 Sep 2025).

A notable result in the one-dimensional half-filled case is that particle-hole symmetry is preserved by SB-CDMFT even though it is not explicitly enforced. The paper highlights that, in this case, the particle-hole transformation mixes subbaths, so symmetry preservation is presented as an emergent validation of the averaging scheme. The same benchmark also shows that increasing subbath size improves low-energy spectral resolution because the Green function acquires more poles (Lagrave et al., 9 Sep 2025).

The doped one-dimensional study compares density versus chemical potential with the exact Lieb–Wu solution. The observables include the local electron density $N_{\rm c}+N_{\rm b}/N_{\rm sb}$7, the insulator-to-metal crossover near the Mott transition, and spectral functions near the critical $N_{\rm c}+N_{\rm b}/N_{\rm sb}$8. The two-subbath SB-CDMFT is reported to reproduce the exact density trend well and in some respects to improve on the single-bath discretization. At the same time, the paper identifies a practical sector-change problem caused by bath discreteness: the impurity electron number $N_{\rm c}+N_{\rm b}/N_{\rm sb}$9 changes in jumps as $N_{\rm c}+N_{\rm b}$0 varies, and in SB-CDMFT different subbaths may converge to different mixed states with different average $N_{\rm c}+N_{\rm b}$1. For this reason a small physical temperature $N_{\rm c}+N_{\rm b}$2 is used to smooth the transitions and improve convergence (Lagrave et al., 9 Sep 2025).

A generality-oriented benchmark breaks reflection and particle-hole symmetry in a two-site cluster by assigning $N_{\rm c}+N_{\rm b}$3 on one site and $N_{\rm c}+N_{\rm b}$4 on the other. Agreement with standard CDMFT remains good, which is used to show that the method does not rely on symmetry constraints.

For the two-dimensional Hubbard model on a four-site square cluster, standard CDMFT with eight bath sites is compared with SB-CDMFT using four subbaths of two sites each. The main observables are spectral functions, Mott-gap structure, particle-hole symmetry, and KS similarity. The method reproduces the main spectral features and the expected particle-hole symmetry. The main quantitative deviation reported is a slight overestimation of the Mott gap in SB-CDMFT (Lagrave et al., 9 Sep 2025).

The large-system calculations are presented as the clearest demonstration of practical reach. The paper studies a four-site one-dimensional cluster coupled to four subbaths of four sites each, a $N_{\rm c}+N_{\rm b}$5 plaquette coupled to four subbaths of four bath sites each, and then a $N_{\rm c}+N_{\rm b}$6 plaquette coupled to four subbaths of eight bath sites each. In these cases the observables are local spectral functions and the trace of the converged hybridization functions along the imaginary axis. The method continues to show clear Mott gaps and, in the two-dimensional examples, precise particle-hole symmetry (Lagrave et al., 9 Sep 2025).

6. Limitations, interpretation, and relation to adjacent developments

SB-CDMFT is explicitly approximate because the full bath is never present simultaneously in the impurity Hilbert space. The recombination by simple averaging is not derived from a controlled expansion; it is presented as an empirically successful ansatz. The approximation consists precisely in the decomposition-and-recombination procedure: correlations induced by the simultaneous presence of all bath orbitals in a single impurity Hilbert space are not treated exactly. Instead, each subbath sees the cluster separately, and inter-subbath effects are only mediated through the global self-consistency loop and the averaged quantities (Lagrave et al., 9 Sep 2025).

The paper also identifies several practical limitations. Optimization becomes harder as the number of subbaths grows because the variational space across all bath parameters grows rapidly even though each ED solve is cheap. “Redundant” subbaths may appear when $N_{\rm c}+N_{\rm b}$7 exceeds the minimum needed by symmetry, in which case some subbaths may converge to negligible hybridization; the authors suggest monitoring this and changing initial bath parameters when necessary. Doped systems are particularly delicate because of bath discreteness and sector changes. When subbaths are chosen according to symmetry, convergence requires that each subbath belong to a single irreducible representation, which is why $N_{\rm c}+N_{\rm b}$8 is typically chosen as a multiple of the number of irreducible representations $N_{\rm c}+N_{\rm b}$9. Variable subbath weights $$H=-t\sum_{\langle ij\rangle,\sigma}(c^\dagger_{i\sigma}c_{j\sigma}+\text{h.c.})+U\sum_i n_{i\uparrow}n_{i\downarrow}-\mu\sum_{i,\sigma}n_{i\sigma},$$0 were found unstable (Lagrave et al., 9 Sep 2025).

Within the wider CDMFT literature, SB-CDMFT occupies a specific methodological niche. It is distinct from changing impurity solvers, and it is distinct from simply imposing stronger symmetry reduction. Earlier work had already shown that finite-bath CDMFT depends on the form of the bath fit, the weighting of low frequencies, the Matsubara cutoff, and the fictitious inverse temperature, so bath organization and parametrization are substantive methodological choices rather than implementation details (0806.2690). More recent work on multi-band cluster DMFT likewise finds that, for small bath sizes, the choice of parameterization can significantly influence the solution and that structured bath partitions can change which metastable or symmetry-broken solution is found (Florez-Ablan et al., 6 Jun 2025). This suggests that SB-CDMFT should be assessed not only by cost reduction but also by systematic convergence checks with respect to subbath number, subbath size, and bath structure.

A further contextual issue is that CDMFT itself breaks translation invariance by construction, because the lattice self-energy is approximated by a cluster self-energy that is nonzero only within a finite cluster and zero between clusters. That artifact can generate cluster-shaped density waves and related low-energy spectral signatures in standard CDMFT (Verret et al., 2019). The SB-CDMFT paper does not claim to remove this intrinsic feature of CDMFT. A plausible implication is that SB-CDMFT primarily addresses the ED bath-size bottleneck while leaving the underlying real-space cluster self-energy ansatz unchanged.

What SB-CDMFT contributes, in the formulation of its authors, is a new bath-factorization strategy for ED-based cluster DMFT: rather than reducing the bath to keep ED feasible, it keeps a richer bath description by splitting it into independently solvable pieces and recombining them self-consistently. It is best viewed as an approximate but practical route to extending ED-CDMFT to larger effective bath sizes while retaining ED’s direct real-frequency access and zero-temperature character. The paper further suggests that similar ideas may be useful for DCA, where one could imagine associating subbaths to cluster momenta $$H=-t\sum_{\langle ij\rangle,\sigma}(c^\dagger_{i\sigma}c_{j\sigma}+\text{h.c.})+U\sum_i n_{i\uparrow}n_{i\downarrow}-\mu\sum_{i,\sigma}n_{i\sigma},$$1, potentially increasing accessible ED cluster sizes without incurring Monte Carlo sign problems (Lagrave et al., 9 Sep 2025).