SciPost logo

SciPost Submission Page

Thermodynamics of the metal-insulator transition in the extended Hubbard model

by M. Schüler, E. G. C. P. van Loon, M. I. Katsnelson, T. O. Wehling

This Submission thread is now published as

Submission summary

Authors (as registered SciPost users): Malte Schüler · Erik van Loon
Submission information
Preprint Link:  (pdf)
Date accepted: 2019-06-06
Date submitted: 2019-06-03 02:00
Submitted by: Schüler, Malte
Submitted to: SciPost Physics
Ontological classification
Academic field: Physics
  • Condensed Matter Physics - Theory
  • Condensed Matter Physics - Computational
Approach: Theoretical


In contrast to the Hubbard model, the extended Hubbard model, which additionally accounts for non-local interactions, lacks systemic studies of thermodynamic properties especially across the metal-insulator transition. Using a variational principle, we perform such a systematic study and describe how non-local interactions screen local correlations differently in the Fermi-liquid and in the insulator. The thermodynamics reveal that non-local interactions are at least in parts responsible for first-order metal-insulator transitions in real materials.

Published as SciPost Phys. 6, 067 (2019)

Author comments upon resubmission

We thank the referees for their careful reading of the manuscript and for their useful questions.

Based on the comments we have carefully revised our manuscript to make it clearer and more convincing. Both referees remarked that the magnitude of our findings on first-order transitions are warningly small. We have addressed this issue by adding a discussion on the significance of our findings in relation to the error estimates which have both numerical and methodological sources.

Here we provide point-by-point answers to all referees comments.

Referee 1:
\emph{1) It would be better to explicitly mention that the change in the double occupancy plotted in Fig. 2 is very small. Even in the case of $V/t=0.3$, the jump is only 0.0003. The energy scale of the structure plotted in Fig.4 is also 0.0001.}
Following also the critique of the second referee, we now highlight the magnitude of the changes in the text and have added a discussion on the significance of the magnitude of the discontinuity with respect to the error estimates shown in the appendix.

\emph{I understand that they performed the calculation very carefully, and the variational method which they employed worked quite successfully in their previous works. However, their approach is not the unique way to map the extended Hubbard model onto the Hubbard model. We may think about the possibility where $\tilde U$ in the effective model has $\omega$ dependence. I wonder how the present result is modified when we improve the mapping from the extended Hubbard model to the on-site Hubbard model.}
The extension of the effective model by a frequency dependent $\tilde U (\omega)$ is an interesting approach to improve the mapping. Unfortunately this is out of reach in our current setup and probably worth a separate publication. Our method crucially depends on the possibility to be able to solve the effective model. This is only true for the Hubbard model with static interaction, by using the DQMC method.
Typically, mappings of an extended Hubbard model to a model with frequency dependent local interaction come with severe approximations which trace back to neglecting nonlocal correlations, like in extended dynamical mean-field theory (EDMFT). Diagrammatic extensions to EDMFT, which reintroduce these nonlocal correlations (like GW+EDMFT or the dual boson approach) are typically themselves approximative and computationally demanding.

Referee 2:
\emph{The major claim of this paper is based on the very small jumps in the double occupancy and entropy $\sim 10^{-4}$. It is really not clear whether it is a numerical artifact, the noise, or the artifact of the particular variational method employed in this paper.}
As also pointed out by the other referee the discontinuities in, e.g., the double occupancy are rather small. They are, however, significantly larger than the error estimates presented in figure 8 in the appendix. We have added corresponding discussions to the manuscript.

\emph{The authors state that they do not observe such first order transition in variational method is used only for the U term (with no hopping renormalization). Hence, it is not clear how valid this varational approach is, some benchmarking is really necessary.}
Unfortunately, we cannot follow the referee's line of argumentation here: How does not finding a feature in an inferior version of a method weaken the validity of a finding in the superior version? Our statement that lower variational degrees of freedom do not show a first-order transition is supposed to highlight that the physical mechanism behind this transition is the competition of kinetic and potential energy minimization, described in our variational approach by an abrupt change from a regime where mostly the effective interaction changes to one where mostly the effective hopping changes with $V$.
We have added a short discussion to the corresponding paragraph, clarifying the physical reason of why two parameter are crucial.
Related to this comment, we have showed in an earlier work (Phys. Rev. B 97, 165135 (2018)), that the variational principle with only one variational parameter does give a first-order transition, but this happens at much larger values of the non-local interaction $V$, where charge-density waves become relevant.
Concerning benchmarking of the variational method, we refer to our publication (Phys. Rev. B 94, 165141 (2016)) where we have performed promising benchmarks against dual boson calculations for the case of a single variational parameter on the square lattice. Additionally, our recent preprint ( shows benchmarks of the method for the case of two variational parameters for exactly solvable systems.

\emph{1) The authors should say explicitly how the entropy calculations was done. }
We have included details of how we calculate the entropy to Appendix A on the free energy calculation.

\emph{2) In Fig. 4 and Fig. 5, the data are presented in $\tilde t$ and $\tilde U$ parameter space. It is really not clear how these numbers are related to real $U$ and $t$ values.}
In Fig. 4, the parameters of the original model (the real parameters $U$, $t$) are highlighted as blue lines in the space of variational parameters. This directly shows the influence nonlocal interactions have on the variational parameters. In Fig. 5, the parameters of the original model are those corresponding to the square markers, since $V=0$ means $\tilde U = U$ and $\tilde t = t$. For clarity, we have added the parameters of the original model in the caption of Fig. 5.

\emph{3) The authors claim that they find the band broadening effect of non-local interactions $V$. Their claim is based on some speculations of free energy behavior in $\tilde t$ parameter space. If the authors mean they see a similar effect as in Ref.~\url{}, they should perhaps find a better way to support their claim.}
We actually do not mention band-widening in our manuscript. We did so on purpose although it may seem tempting to interpret a renormalization of the effective parameter $\tilde t$ as a band renormalization. First, however, $\tilde t$ only determines the band-width in the non-interacting (and weakly interacting) case. Second, the variational principle we use leads to variational parameters which optimize the free energy, i.e., ground-state properties, and not spectral properties.

\emph{4) Why in eq. 7, the original lattice dispersion function ϵk and not for the effective Hubbard model $\tilde \varepsilon_k$ is used.}
Eq. 7 of course contains the dispersion of the effective model. We have fixed our ambiguous notation in the revised manuscript.

\emph{5) Why the energy units (hopping t) are changes through the paper? In Fig. 6 the authors used $t=1.26$ and in Fig. 9, $t=1$ is used. This way your are studying the original extended Hubbard model for different bandwidth values.}
In our work, we decided to fix the energy units by the temperature ($T=0.1$) rather than a fixed hopping. Using the variational principle with the hopping $\tilde t$ as an effective parameter, it is quite natural to also consider a variable original hopping $t$.

List of changes

Added discussion in Sec. 3.1 on the significance of the discontinuities presented in Figs. 2 and 3 with respect to error estimates presented in Fig. 8.
Added the parameters of the presented original models to the caption of Fig. 5.
Added discussion in Sec. 3.3 on why two variational parameters are crucial for describing the first-order transition.
Clarified notation of Eq. 7 in Appendix A.
Extended appendix A on the calculation of the free energy with details of how we calculate the entropy (with a new Equation 8).

Reports on this Submission

Anonymous Report 2 on 2019-6-4 (Invited Report)


This version of the manuscript is improved. I appreciate the authors taking the time and taking into account the comments/suggestions.


I appreciate that authors answered and addressed all the comments/suggestions.
This version of the manuscript is much more improved and is suitable for publication.

  • validity: good
  • significance: good
  • originality: top
  • clarity: good
  • formatting: perfect
  • grammar: perfect

Anonymous Report 1 on 2019-6-3 (Invited Report)


I think that the authors responded to all the criticisms raised in the previous report. Now the manuscript is suitable for publication.

  • validity: -
  • significance: -
  • originality: -
  • clarity: -
  • formatting: -
  • grammar: -

Login to report or comment