Compressed Wannier modes found from an $L_1$ regularized energy functional

Farzin Barekat, Ke Yin, Russel E. Caflisch, Stanley Osher, Rongjie Lai, Vidvuds OzoliņšarXiv 2014 · DOI · arXiv

Abstract

We propose a method for calculating Wannier functions of periodic solids directly from a modified variational principle for the energy, subject to the requirement that the Wannier functions are orthogonal to all their translations ("shift-orthogonality"). Localization is achieved by adding an $L_1$ regularization term to the energy functional. This approach results in "compressed" Wannier modes with compact support, where one parameter $μ$ controls the trade-off between the accuracy of the total energy and the size of the support of the Wannier modes. Efficient algorithms for shift-orthogonalization and solution of the variational minimization problem are demonstrated.

Introduction

In the study of periodic solids, electronic states are traditionally described using Bloch waves. While these delocalized functions are eigenfunctions of the Hamiltonian, they lack the intuitive spatial localization necessary to understand local phenomena such as chemical bonding, electric polarization, or the development of linear-scaling electronic structure methods. Wannier functions (WFs) provide a localized alternative, but their construction typically involves a post-processing step: first calculating delocalized Bloch waves and then applying a transformation (like the Maximally Localized Wannier Function method) to achieve localization. This process is often computationally expensive, sensitive to initial guesses, and results in functions with infinite "tails."

This paper introduces a direct variational approach to finding localized states by incorporating $L_1$ regularization—a technique borrowed from compressed sensing—directly into the energy functional.

Methodology: $L_1$ Regularization and Compressed Modes

The core of the proposed method is a modified variational principle. Instead of minimizing the standard energy functional, the authors minimize a functional $J(\psi)$ that includes an $L_1$ penalty: $$J(\psi) = \frac{1}{\mu} \|\psi\|_1 + \langle \psi | \hat{H} | \psi \rangle$$ The $L_1$ norm (where $\|\psi\|_1 = \int |\psi(x)| dx$) acts as a sparsity-promoting term. In this context, "sparsity" translates to spatial localization, specifically forcing the wave function to have compact support—meaning it becomes exactly zero outside a finite region.

To ensure these functions represent a valid basis for the crystal, two additional constraints are applied:

  1. Shift-Orthogonality: Each mode must be orthogonal to all its translations by lattice vectors.
  2. Split Bregman Iteration: This numerical scheme efficiently handles the $L_1$ term by decoupling the Hamiltonian physics from the sparsity-inducing soft-thresholding operation.
The periodic potential (top) and the first eight compressed Wannier modes (Level 1-8), demonstrating compact support and localization within specific potential wells.

As shown in Figure 2, the algorithm successfully identifies modes localized within the potential wells of a 1D model. Lower-energy "semi-core" states are highly localized in the deepest wells, while higher-energy states spread across multiple wells but remain strictly confined to a finite spatial domain.

Results and Discussion

The effectiveness of these compressed modes is evaluated by their ability to reconstruct the underlying physics of the system. The parameter $\mu$ serves as a tuning knob: a larger $\mu$ prioritizes energy accuracy, while a smaller $\mu$ enforces tighter spatial compression.

Band Structure Reconstruction

A critical test for any Wannier-like basis is the reconstruction of the energy dispersion (band structure). Using only eight compressed modes derived from the 1D model, the authors calculated the energy eigenvalues $E(k)$ across the Brillouin zone.

Comparison of the energy dispersion $E(k)$ calculated using eight compressed Wannier modes (red lines) versus exact diagonalization (blue stars).

As illustrated in Figure 1, the compressed modes provide an excellent approximation of the exact band structure. The lower bands show near-perfect agreement, with only minor deviations appearing in the highest band, which is expected given the finite number of modes used.

Advantages over Traditional Methods

  • Directness: The method bypasses the need for Bloch wave calculation, potentially avoiding local minima issues found in traditional Wannierization.
  • Generality: The approach is applicable to both insulating and metallic systems, whereas traditional methods often require specialized handling for metals.
  • Computational Efficiency: The use of Fourier-based projections for shift-orthogonality ensures the method scales favorably ($M \log M$), making it competitive with existing plane-wave density functional theory (DFT) codes.

Conclusion and Future Work

The use of $L_1$ regularization provides a robust framework for generating Wannier functions with strictly compact support. While demonstrated here in one dimension, the method holds significant promise for 3D materials science applications. Future work will focus on extending the algorithm to realistic 3D atomic potentials and exploring its utility in computationally intensive "beyond-DFT" methods, such as Quantum Monte Carlo or screened exchange, where the finite range of compressed modes could offer substantial speedups.

Summary updated

Abstract

We propose a method for calculating Wannier functions of periodic solids directly from a modified variational principle for the energy, subject to the requirement that the Wannier functions are orthogonal to all their translations ("shift-orthogonality"). Localization is achieved by adding an $L_1$ regularization term to the energy functional. This approach results in "compressed" Wannier modes with compact support, where one parameter $μ$ controls the trade-off between the accuracy of the total energy and the size of the support of the Wannier modes. Efficient algorithms for shift-orthogonalization and solution of the variational minimization problem are demonstrated.
BibTeX
@misc{barekat2014compressed,
  author = {Farzin Barekat and Ke Yin and Russel E. Caflisch and Stanley Osher and Rongjie Lai and Vidvuds Ozoliņš},
  title = {{Compressed Wannier modes found from an $L\_1$ regularized energy functional}},
  year = {2014},
  eprint = {1403.6883},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1403.6883},
}
Figure