$\texttt{iNORG}$: An open-source quantum impurity solver package based on the natural orbitals renormalization group

Source: arXiv:2607.13993 · Published 2026-07-15 · By Jia-Ming Wang, Yi-Heng Tian, Yin Chen, Ru Zheng, Rong-Qiang He, Zhong-Yi Lu

TL;DR

This paper presents iNORG, an open-source quantum impurity solver based on the Natural Orbitals Renormalization Group (NORG) method, designed to efficiently solve Anderson impurity models within dynamical mean-field theory (DMFT) for strongly correlated electron systems. Traditional impurity solvers like exact diagonalization (ED), quantum Monte Carlo (QMC), and numerical renormalization group (NRG) struggle to accurately and efficiently handle large baths with many orbitals, especially at zero temperature and for complex interactions. iNORG leverages an optimized natural orbital basis extracted from the single-particle density matrix to compress the Hilbert space by freezing nearly full or empty bath orbitals while explicitly treating active orbitals with significant quantum fluctuations. This selective Hilbert space truncation yields polynomial scaling (roughly cubic with bath size), allowing solution of larger impurity problems with ED-level accuracy but significantly reduced cost.

The paper carefully details the iterative NORG algorithm, including bath discretization via hybridization fitting using Levenberg–Marquardt nonlinear least squares, construction of a Hilbert subspace guided by natural orbital occupancy constraints (NOOC), ground state solution by Lanczos diagonalization in the compressed space, and calculation of physical observables like single-particle Green's functions. The authors demonstrate the software’s computational workflow, modular C++ implementation with multi-threading and BLAS/LAPACK support, and integration paths into DMFT contexts. Benchmark examples illustrate that the fitting error for representing the bath real-frequency hybridization function drops exponentially with the number of bath orbitals used, validating the approach. Overall, iNORG provides a robust, scalable solver enabling high-resolution zero-temperature DMFT studies of multi-orbital quantum impurity problems previously intractable for standard methods.

Key findings

• Natural orbital based Hilbert space truncation reduces computational scaling from exponential in full ED to approximately O(N_bath^3), enabling large bath sizes of tens to hundreds of orbitals.
• Hybridization function fitting error decreases exponentially with bath size Nbath, achieving errors below 10^-4 for Nbath ≥ 7 (Fig. 3).
• NOOC (natural orbital occupancy constraint) dynamically freezes orbitals with occupations close to 0 or 1, retaining accuracy while compressing Hilbert space.
• Lanczos algorithm with Krylov subspace of m ~200 iterations efficiently converges to ground state energy and wavefunction within the active subspace.
• Bath orbital transformation only applied to non-interacting bath orbitals preserves sparse Hamiltonian structure and improves computational efficiency compared to full orbital basis rotation.
• Iterative self-consistent NORG cycle converges both the natural orbital occupations and ground state energy until user-defined criteria (energy difference, occupancy norm difference) are satisfied.
• The iNORG implementation supports multi-orbital Bethe lattice models and is compatible with DFT+DMFT software frameworks via configuration files.
• Current iNORG version is limited to zero temperature, single-site impurity models with density-density Coulomb interactions.

Threat model

Not applicable; the paper addresses computational complexity and algorithmic efficiency challenges in solving quantum impurity models within DMFT, rather than security adversaries.

Methodology — deep read

1. Threat model & assumptions: The paper addresses the computational problem of efficiently and accurately solving Anderson impurity models appearing in DMFT, focusing on single-site, zero-temperature impurity problems with density-density Coulomb interactions. The adversary in this context is the exponential Hilbert space growth that hampers exact diagonalization and many standard solvers. The solver assumes a discretized bath representation fit from continuous hybridization functions. It does not address adversarial noise or direct attack scenarios.

2. Data: The impurity model is defined by local Hamiltonian parameters and a discretized bath obtained by fitting the continuous local hybridization function Γloc(z), typically from DMFT calculations or model systems like a Bethe lattice. The fitting uses the Levenberg–Marquardt algorithm on Matsubara frequency points with constraints enforcing sum rules and preventing parameter divergence. No large external datasets are used; instead, the bath parameters are organically generated.

3. Architecture / algorithm: The core algorithm iteratively optimizes the single-particle orbital basis to find natural orbitals (NOs) via diagonalization of the single-particle density matrix (SPDM) from the current approximation to the ground state. The natural orbitals separate into active orbitals (occupation away from 0 or 1) and inactive orbitals (near 0 or 1). The Hilbert space is truncated by freezing inactive orbitals, significantly reducing dimension. The effective Hamiltonian within the selected subspace is constructed and solved by the Lanczos algorithm, which builds a Krylov subspace to approximate the ground state with ~200 iterations. After each solve, the SPDM is recomputed from the ground state, the NO basis updated, and the active space refined according to the natural orbital occupancy constraint (NOOC). This self-consistent loop proceeds until ground state energy and NO occupations converge. Bath orbital transformation applies only to bath (non-interacting) orbitals, keeping impurity orbitals unchanged to preserve sparse Hamiltonian structure and computational efficiency.

4. Training regime: Not applicable as this is a numerical physics solver. The algorithms run single simulations iteratively within a self-consistency loop. The Lanczos iteration count, tolerances, and fitting hyperparameters are user-configurable. The authors implemented multi-threaded BLAS/LAPACK and parallelism via MPI to accelerate linear algebra and Hamiltonian operations.

5. Evaluation protocol: Fit quality is evaluated by residual squared error reduction in hybridization function fitting. Solver accuracy is demonstrated by convergence of ground state energy and natural orbital occupations. The scalability and efficiency claims are supported by showing polynomial cubic scaling in bath size Nbath, and by examples resolving bath fitting errors below 10^-4 with Nbath ≥ 7. Benchmarks compare with traditional ED methods qualitatively but detailed quantitative benchmarking against other methods is not reported. No cross-validation or adversarial distribution shifts were evaluated.

6. Reproducibility: The full source code is publicly available in C++ under AGPL-3.0 at https://github.com/QTMEC-RUC/iNORG with installation instructions. The software integrates with DMFT workflows producing input parameters, enabling reproducible end-to-end calculations. The bath discretization fitting and NOOC algorithms are fully described for independent reimplementation.

Example: Given a DMFT output local hybridization function Γloc(z), iNORG fits this function via Levenberg–Marquardt nonlinear least squares to a discretized bath of Nbath orbitals. The hybridization amplitudes Vk and bath energies ϵk are optimized to minimize the fit residual. Using this discrete impurity Hamiltonian, the initial single particle basis is chosen as the bath orbitals. The solver constructs an initial Hilbert subspace by selecting orbitals with intermediate occupations. The Lanczos algorithm finds the approximate ground state in this subspace, from which the SPDM is computed and diagonalized to get a new natural orbital basis and occupancy numbers. Orbitals with nearly full or empty occupations are frozen, reducing the active subspace size. The Hamiltonian is rebuilt in the new basis subspace and the process repeats until the ground state energy and occupation norm difference between iterations falls below tolerance. Finally, dynamic observables like Green’s functions are computed using continued fractions from the Krylov basis. This iterative procedure typically converges within a few tens of cycles.

Technical innovations

• Use of natural orbital basis extracted from single-particle density matrix to compress the impurity plus bath Hilbert space with high accuracy.
• Natural orbital occupancy constraint (NOOC) scheme dynamically freezes orbitals near full or empty occupation to reduce Hilbert space exponentially while controlling truncation error.
• Selective bath-only orbital transformations preserve sparse Hamiltonian structure and enable polynomial scaling ∼O(N_bath^3) versus full orbital rotation with O(N^4) complexity.
• Automated hybridization function fitting via Levenberg–Marquardt nonlinear least squares with adaptive damping and physical constraints to discretize baths accurately.
• Implementation of a NORG-Table caching structure storing many-body operator relations to efficiently reconstruct Hamiltonian matrices during self-consistency iterations.

Baselines vs proposed

• Fitting error for hybridization function: Nbath=3, error ~1e-3; Nbath=7, error <1e-4, showing exponential error decay with bath size (Fig. 3).
• Computational scaling: ED scales exponentially with bath size; iNORG scales roughly as O(N_bath^3) due to NOOC and bath-only orbital transformations.

Limitations

• Currently restricted to zero-temperature impurity problems; finite-temperature calculations are planned but not yet supported.
• Supports only single-site impurity models; cluster DMFT and multi-impurity extensions are future work.
• Only density-density Coulomb interaction terms are implemented, excluding fully rotationally invariant interactions.
• Benchmarking primarily qualitative; lacks detailed quantitative comparison against other state-of-the-art impurity solvers across diverse problem sets.
• No adversarial robustness or distribution shift testing on bath discretization or solver convergence.
• The accuracy of bath discretization depends on initial guess and nonlinear optimization convergence; risk of local minima in fitting.

Open questions / follow-ons

• How to efficiently extend iNORG to finite-temperature settings while maintaining computational efficiency?
• What is the performance and accuracy impact when incorporating full Coulomb vertex interactions beyond density-density terms?
• Can the method be extended and optimized for cluster impurity models or multi-site DMFT problems?
• How does iNORG compare quantitatively to continuous-time QMC and NRG methods in benchmark spectral functions and physical observables?

Why it matters for bot defense

While not directly related to bot-defense or CAPTCHA technical challenges, the iNORG solver's core contribution lies in efficient dimensionality reduction and iterative optimization strategies to handle extremely large state spaces. Bot-defense engineers interested in scalable solutions for high-dimensional input spaces could draw inspiration from the natural orbital occupancy constraints that freeze inactive degrees of freedom, akin to feature selection in ML. The automated hybridization function fitting and adaptive subspace refinement mechanisms resemble iterative model compression and efficient representation learning algorithms used in adversarial example detection or human interaction evaluations. However, direct application in CAPTCHA or bot-detection systems would require adapting these quantum many-body methods for non-quantum signal or behavioral data. The paper’s principle of iterative basis optimization and sparse representation might inform future anti-bot system designs that need to efficiently model high-dimensional interaction spaces with minimal computational overhead.

Cite

@article{arxiv2607_13993,
  title={ $\texttt{iNORG}$: An open-source quantum impurity solver package based on the natural orbitals renormalization group },
  author={ Jia-Ming Wang and Yi-Heng Tian and Yin Chen and Ru Zheng and Rong-Qiang He and Zhong-Yi Lu },
  journal={arXiv preprint arXiv:2607.13993},
  year={ 2026 },
  url={https://arxiv.org/abs/2607.13993}
}

Read the full paper

• arXiv:2607.13993 (abstract)
• arXiv:2607.13993 (PDF)