The flaw that interpolation is meant to remove
Cluster extensions of dynamical mean-field theory capture short-ranged non-local correlations that single-site DMFT misses, but they pay for it with self-energies that are unphysical in a specific, structural way. Cellular DMFT (CDMFT) couples a real-space cluster of sites to self-consistent baths, which breaks translation symmetry, so its real-space self-energy is not translationally invariant. The dynamical cluster approximation (DCA) instead coarse-grains the Brillouin zone into patches, which introduces artificial discontinuities in the momentum dependence of the self-energy.
Neither defect is a statement about the physics; both are artifacts of how the cluster is embedded. The standard remedy is post-processing: periodization for CDMFT, interpolation for DCA, which the paper groups under the single umbrella term interpolation. The goal is to reconstruct a smooth, translationally invariant lattice quantity from cluster data that are defined only on a few sites or patches.
Why interpolating G, Sigma or M can break causality
The established schemes interpolate the Green's function G(omega), the self-energy Sigma(omega), or the cumulant M(omega). All three are frequency-dependent functions, and the paper states that they all have problematic properties, sometimes including violation of causality. The reason is that interpolation is a smoothing operation applied across cluster or patch boundaries, and nothing in that operation enforces the analytic structure a physical response function must have.
The practical failure is visible in the hole-doped square-lattice Hubbard model, where traditional approaches struggle to reliably capture intricate structures such as hole pockets. So the problem is not merely formal: the schemes that are supposed to repair the cluster result can distort exactly the momentum-space features readers want to resolve.
Interpolating frequency-independent Liouvillian elements instead
The paper's proposal is to interpolate a different object. The Liouvillian L is the generator of time translations in the space of all linear operators acting in Fock space, and its matrix elements can be obtained from a continued-fraction expansion (CFE) of the retarded self-energy. The authors interpolate the frequency-independent matrix elements of the single-particle irreducible part of L, then reinsert the interpolated elements into the self-energy CFE.
Two properties make these elements better candidates than G, Sigma or M. First, they do not depend on frequency, so the interpolation never touches the frequency axis where analytic structure lives. Second, the paper demonstrates that the ingredients of the CFE possess a more local Fourier expansion than the functions used in traditional schemes, meaning they typically have a shorter range in real space. Because the causal structure is carried by the CFE rather than by the interpolated quantity, the resulting self-energy is causal by construction, with a minor caveat the paper addresses on p. 11. The authors also note the method is technically not more demanding than the alternatives; its only additional ingredient is the CFE.
Faster convergence with cluster size in the 1D test
The first test uses the one-dimensional Hubbard model with CDMFT as a controlled setting where cluster size can be varied. The paper reports that L-interpolation yields faster convergence with cluster size than Sigma-, G- or M-interpolation. This is a comparison of convergence rates across schemes applied to the same cluster data, not a claim about the 1D model's physics, and it establishes that the method approaches the large-cluster limit more efficiently than the standard alternatives.
Fermi and Luttinger arcs in the 2D pseudogap regime
The physical application is a four-patch DCA treatment of the two-dimensional Hubbard model in the pseudogap regime, solved with the numerical renormalization group, alongside a four-site CDMFT calculation using exact diagonalization. The teaser comparison shows the raw DCA result with its Brillouin-zone discontinuities next to interpolations based on Sigma, M and L. The paper states that L-interpolation is the scheme that most convincingly mimics the original discontinuous cluster results.
The main physical result follows from that reconstruction: depending on doping, L-interpolation can yield Fermi arcs and Luttinger arcs that together seem to form a closed surface, with a seemingly smooth crossover between the two. The word seem is doing real work here. The closed surface and the smooth crossover are features of the interpolated quantity, and the paper presents them as the outcome of the scheme rather than as an independently measured property of the model.
What the method does not settle
The paper is explicit that its goal is mainly methodological: to describe the concepts and computational steps of L-interpolation in detail, with the 2D application included to illustrate its power. It deliberately refrains from adjudicating between CDMFT and DCA, noting that problem-specific reasons may favor one approach. The authors suggest that where both are feasible at comparable cost, such as four-site CDMFT versus four-patch DCA, a systematic comparison of the resulting Fermi surface properties would be worthwhile.
Scope also limits the demonstrated cases. The illustrations cover the 1D Hubbard model with CDMFT and the 2D Hubbard model with four-site CDMFT and four-patch DCA; the symmetry-broken self-energies needed for d-wave superconductivity and antiferromagnetism would require generalizing the formulation, which the paper lists as future work. Interpolation here is a post-processing step applied after DMFT self-consistency, not something performed inside the cluster-DMFT loop.
Where Liouvillian interpolation earns its place
The decision the paper supports is narrow and concrete. When the target is good momentum resolution of spectral features from cluster-DMFT data, and the choice is between interpolating G, Sigma, M or the Liouvillian, the evidence favors the Liouvillian: it conserves causality by construction, converges faster with cluster size in the 1D CDMFT test, and in the four-patch DCA pseudogap calculation produces the Fermi and Luttinger arcs that the other schemes do not convincingly reproduce. The cost of that choice is one extra technical ingredient, the continued-fraction expansion.
The priority is therefore not a verdict on CDMFT versus DCA, which the authors leave open, but on which quantity to interpolate once cluster data exist. The most consequential claim for readers is the combination of guaranteed causality and the reconstructed arc structure in the pseudogap regime, since that is where the older schemes visibly struggle.
Sources used in this answer
Liouvillian interpolation of the self-energy of cluster dynamical mean-field theories
Interpolating frequency-independent matrix elements of the single-particle irreducible Liouvillian, obtained from a continued-fraction expansion of the cluster self-energy, yields a causal-by-construction interpolated self-energy that converges faster with cluster size than Sigma-, G- or M-interpolation in 1D CDMFT and produces Fermi and Luttinger arcs in four-patch DCA for the 2D Hubbard model.