Direct minimization versus iterative embedding in the ghost-Gutzwiller method: a comparative study of magnetism in Mott insulators

Source: arXiv:2607.27156 · Published 2026-07-29 · By Antonio Maria Tagliente, Ivan Pasqua, Michele Fabrizio

TL;DR

This paper studies numerical solution strategies for the ghost-Gutzwiller (ghost-GA) approximation applied to Mott insulating phases of the single-band Hubbard model, comparing iterative embedding schemes versus direct energy minimization approaches. The ghost-GA method can be framed equivalently as a variational problem over a wavefunction or as a quantum impurity embedding model solved iteratively. While the iterative embedding approach is computationally efficient and works well in symmetric metallic or symmetry-broken antiferromagnetic phases, it suffers from instabilities and unphysical behaviors in the paramagnetic Mott insulating phase, especially under a Zeeman magnetic field. These failings stem from degeneracies and the constraint relaxation inherent in the embedding scheme, leading to discontinuous energy jumps and spurious fully polarized insulating solutions. In contrast, directly minimizing the ghost-GA energy functional over impurity wavefunction parameters avoids these pathologies, stabilizing smooth, physically correct paramagnetic Mott states with finite spin susceptibility. The paper carefully analyzes these contrasting behaviors, delineates the limitations of the iterative approach, and shows that its reliability depends on whether symmetry breaking occurs. When antiferromagnetism is included, iterative embedding matches DMFT results well, confirming its utility. Overall, the study provides crucial insight into when and how different numerical solution schemes for quantum embedding methods can be trusted for correlated electron problems near and beyond the Mott transition.

Key findings

• The iterative embedding scheme converges smoothly in the paramagnetic metallic phase (U < Uc) but suffers degeneracies driving instability past the Mott transition (U > Uc).
• At U > Uc under zero magnetic field, embedding solutions require ad-hoc fixes to enforce paramagnetic symmetry; otherwise, magnetization flips appear iteration-to-iteration.
• Applying a weak Zeeman field h=0.01 causes the iterative embedding scheme’s energy to jump discontinuously and leads to spurious fully polarized (m=1) insulating solutions.
• Direct minimization of the ghost-GA energy functional yields smooth energy curves, stable partially polarized Mott insulating states (m < 1), and finite zero-field spin susceptibilities across the transition.
• In the antiferromagnetic phase allowing symmetry breaking, the iterative embedding scheme produces energy and magnetization results closely matching DMFT and outperforming direct minimization.
• The direct minimization method facilitates access to metastable states with fixed order parameter and arbitrary staggered magnetization not accessible with the iterative approach.
• Mutual information measures show the iterative embedding reduces to Hartree-Fock-like descriptions at strong coupling, missing quantum fluctuation effects better captured variationally.
• The critical interaction strength Uc for the Mott transition and quasiparticle weights computed via ghost-GA agree well between iterative and direct methods in symmetric phases without perturbations.

Methodology — deep read

The authors study the single-band Hubbard model on a Bethe lattice with infinite coordination (D=2), using the ghost-Gutzwiller approximation (ghost-GA), a variational wavefunction approach extended with auxiliary 'ghost' orbitals representing baths.

The ghost-GA wavefunction is constructed as a projector P applied to an uncorrelated Slater determinant |ψ*> defined on an auxiliary Hilbert space expanded by Nghosts ghost orbitals beyond the physical orbitals. The projector P is parametrized by local linear operators mapping auxiliary fermion states to physical states. The number of ghosts is fixed at 2 for impurity models with one physical site and three bath orbitals.

The core numerical problem is minimizing the variational energy functional E(ψ*, ϕ) that depends on the Slater determinant |ψ*> and an impurity wavefunction |ϕ⟩ describing the local impurity plus baths. This functional includes kinetic energy evaluated analytically on infinite-coordination lattices and local Hubbard interaction terms.

Two solution schemes are employed:

1. Iterative embedding scheme: Using Lagrange multipliers and relaxing constraints, the problem is recast into coupled eigenvalue problems resembling a DMFT-like self-consistency loop where the impurity model's ground state is solved iteratively to update bath parameters and density matrices.
2. Direct minimization: The impurity wavefunction is parametrized explicitly with five variational parameters encoding impurity-bath hybridization and magnetization. The full energy functional is minimized directly over these parameters at fixed magnetization sectors without enforcing the impurity wavefunction to be the ground state of an auxiliary linear problem.

The critical interaction Uc for the Mott transition is determined by monitoring the quasiparticle residue Z. Solutions are computed both in the paramagnetic case (symmetry-preserving) and with allowed antiferromagnetic order (symmetry broken). To test robustness, a small Zeeman field h=0.01 is applied to probe spin susceptibility and degeneracy lifting effects.

For the paramagnetic Mott phase, the iterative scheme exhibits convergence difficulties and unstable magnetization flips due to ground state degeneracies involving decoupled bath orbitals. Direct minimization stabilizes physically correct paramagnetic insulating states with finite susceptibility.

In the antiferromagnetic phase, iterative embedding converges reliably and its predictions align closely with DMFT results obtained from exact diagonalization solvers, validating this approach when symmetry breaking is present.

Comparison plots show total energies, local interaction energies, kinetic energies, impurity magnetizations, and mutual information across U values ranging from metallic to strong Mott insulating regimes. The benchmarking includes standard Gutzwiller approximation and Hartree-Fock results for context.

The paper also discusses representability constraints on one-body density matrices, gauge symmetries causing degeneracies, and how these impact the consistency and stability of solutions in each method. Code and exact numerical parameters are not noted as publicly available, but description of wavefunction parametrizations and optimization procedures is detailed enough to allow reproduction by experts.

Technical innovations

• Identification and explicit resolution of degeneracies in the impurity wavefunction manifold causing iterative embedding instabilities in paramagnetic Mott insulators.
• Construction of a compact variational parametrization of the impurity wavefunction including decoupled baths allowing direct minimization to avoid embedding artifacts.
• Demonstration that iterative embedding schemes can fail under perturbations (e.g., Zeeman field) that lift impurity ground state degeneracies, while direct minimization remains robust.
• Extension of ghost-Gutzwiller approximation to handle symmetry-broken antiferromagnetic phases via two distinct impurity wavefunctions for sublattices, reproducing DMFT results accurately.
• Clarification of conditions under which the embedding scheme is reliable and when direct minimization is essential for physically meaningful quantum embedding solutions.

Baselines vs proposed

• Standard Gutzwiller approximation: finite magnetic susceptibility diverges at Mott transition; ghost-GA direct minimization yields finite susceptibility smoothly evolving across Uc.
• Iterative embedding vs direct minimization (paramagnetic Mott insulator with Zeeman field): embedding suffers discontinuous energy jump and fully polarized m=1 state; direct minimization energy is lower and magnetization m < 1.
• Ghost-GA iterative embedding vs DMFT (antiferromagnetic phase): total energies closely match across U values; iterative embedding outperforms direct minimization in energy and magnetization accuracy.
• Hartree-Fock and conventional GA compared to ghost-GA and DMFT (antiferromagnetic phase): ghost-GA and DMFT lower energy at small U; differences diminish at large U.

Limitations

• Direct minimization approach is computationally less efficient and more complex than iterative embedding, limiting practical scalability.
• Methodological assessment restricted to infinite-coordination Bethe lattice; finite-dimensional lattice effects, frustration, and disorder are not addressed.
• No explicit adversarial or robustness tests beyond small magnetic field perturbations; generalization to other perturbations like crystal-field splitting or Hund’s coupling is not demonstrated.
• Code and datasets are not publicly released, potentially limiting reproducibility by external researchers without reimplementation.
• The impurity wavefunction parametrization in direct minimization is limited to fixed low-dimensional subspace, possibly neglecting more complex bath correlations or dynamics.
• Iterative embedding failures stem from relaxed representability constraints which may not capture full complexity of feasible density matrices.

Open questions / follow-ons

• Can the direct minimization scheme be accelerated or approximated to make it more computationally competitive with iterative embedding methods?
• How does the ghost-Gutzwiller approach generalize to multi-band systems with more complex local interactions such as Hund's coupling or crystal-field splitting?
• What is the behavior of these methods on frustrated or low-dimensional lattices where spin-liquid states and strong quantum fluctuations dominate?
• Can hybrid methods be designed to combine iterative embedding efficiency with the robustness of direct minimization to exploit their complementary strengths?

Why it matters for bot defense

Although the paper focuses on quantum embedding methods for strongly correlated electronic systems rather than bot-defense or CAPTCHA techniques, it provides important general insights into iterative versus direct optimization schemes for constrained variational functionals. Bot-defense practitioners developing CAPTCHA systems relying on iterative model training or embedding layers may learn from the demonstrated instability and degeneracy issues that arise in iterative schemes under certain degenerate or symmetry-related conditions. The finding that direct minimization of an energy-like functional can avoid pathological jumps and enforce physical symmetries might inspire more robust training or optimization strategies for CAPTCHA-related models, especially when the problem domain has competing or degenerate solutions. Additionally, the careful treatment of enforcement of representability and symmetry constraints in high-dimensional parameter spaces may analogously inform methods to enhance adversarial robustness or state consistency in bot-detection classifiers or puzzle generators. In short, the detailed investigation of optimization pathologies and the conditions ensuring stable and physical solutions has likely conceptual utility for designing trustworthy, stable iterative or variational methods in bot-defense machine learning pipelines.

Cite

@article{arxiv2607_27156,
  title={ Direct minimization versus iterative embedding in the ghost-Gutzwiller method: a comparative study of magnetism in Mott insulators },
  author={ Antonio Maria Tagliente and Ivan Pasqua and Michele Fabrizio },
  journal={arXiv preprint arXiv:2607.27156},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.27156}
}

Read the full paper

• arXiv:2607.27156 (abstract)
• arXiv:2607.27156 (PDF)