Maximally Localized Wannier Functions

• MLWFs are optimally localized real-space orbitals obtained by unitary mixing of Bloch functions in periodic systems.
• They minimize the Marzari–Vanderbilt spread functional, enabling precise band structure interpolation and tight-binding model extraction.
• MLWFs facilitate a range of applications from electronic structure analysis and quantum transport to automated high-throughput materials simulations.

Maximally Localized Wannier Functions (MLWFs) are real-space orbitals obtained from the unitary mixing and Fourier transform of Bloch eigenfunctions of a periodic Hamiltonian. MLWFs furnish an optimally compact, physically transparent basis for the representation and interpolation of the electronic structure in solids. The global localization criterion, the Marzari–Vanderbilt spread functional, formalizes the selection of a unique set of Wannier functions distinguished by minimal spatial variance, facilitating post-processing, tight-binding construction, property interpolation, and orbital analysis across applications in condensed-matter, photonic, and quantum transport contexts.

1. Mathematical Formulation and Localization Criterion

In a periodic crystal, the Hamiltonian eigenstates are Bloch functions $$\psi_{n\mathbf{k}}(\mathbf{r}) = e^{i\mathbf{k}\cdot\mathbf{r}}u_{n\mathbf{k}}(\mathbf{r})$$, where $u_{n\mathbf{k}}(\mathbf{r})$ is cell-periodic, $n$ is the band index, and $\mathbf{k}$ is a Brillouin zone wavevector. Wannier functions are defined via inverse Fourier transformation: $$w_{n\mathbf{R}}(\mathbf{r}) = \frac{V}{(2\pi)^3}\int_{\mathrm{BZ}}d\mathbf{k}\;e^{-i\mathbf{k}\cdot\mathbf{R}}\sum_{m=1}^{J}U_{mn}(\mathbf{k})\psi_{m\mathbf{k}}(\mathbf{r}),$$ where $U_{mn}(\mathbf{k})$ is a $J\times J$ unitary ("gauge") matrix at each $\mathbf{k}$, $\mathbf{R}$ a lattice vector, and $J$ the size of the target manifold.

Localization is quantified by the total quadratic spread functional (Marzari et al., 2011, Shelley et al., 2011): $$\Omega = \sum_{n=1}^J \left[\langle w_{n\mathbf{0}}|r^2|w_{n\mathbf{0}}\rangle - |\langle w_{n\mathbf{0}}|{\bf r}|w_{n\mathbf{0}}\rangle|^2 \right].$$ Maximally localized Wannier functions (MLWFs) are the set minimizing $\Omega$ over all allowed $U_{mn}(\mathbf{k})$. Marzari and Vanderbilt decomposed the spread into gauge-invariant ($\Omega_I$), diagonal ($\Omega_D$), and off-diagonal ($\Omega_{OD}$) parts, which can be explicitly formulated in terms of nearest-neighbor $k$-space overlaps $$M_{mn}^{(\mathbf{k},\mathbf{b})} = \langle u_{m\mathbf{k}}|u_{n,\mathbf{k}+\mathbf{b}}\rangle$$ for a discrete mesh.

2. Construction, Algorithms, and Disentanglement

For an isolated group of $J$ bands, one directly optimizes the gauge matrices $U^{(\mathbf{k})}$ to minimize $\Omega$ using iterative (conjugate-gradient or steepest-descent) schemes (Marzari et al., 2011, Shelley et al., 2011). For entangled bands, a two-step disentanglement is necessary (Shelley et al., 2011, Damle et al., 2018, Qiao et al., 2023):

• Subspace selection (disentanglement): An "outer energy window" is defined ($J\geq N$ bands at each $k$). Minimization of the gauge-invariant part $\Omega_I$ yields a smooth $N$-dimensional subspace, optionally using projectability-based (Jiang et al., 9 Jul 2025, Qiao et al., 2023) or SCDM-based (Vitale et al., 2019, Damle et al., 2018) schemes.
• Spread minimization: Final gauge rotations $U^{(k)}$ within the selected subspace reduce $\Omega_{OD}+\Omega_D$.

Advanced, parameter-free algorithms bypass explicit projection guesses (Mustafa et al., 2015, Cancès et al., 2016), employing optimized projection functions (OPF). These approaches systematically select a semi-unitary matrix $W$ generating trial orbitals as linear combinations of an overspanning atomic-like set, then minimize a suitable localization Lagrangian. Self-projection cycles further enhance localization for entangled bands (Tillack et al., 5 Feb 2025).

In one dimension, an exact procedure for MLWFs avoids numerical minimization, using analytic phase fixing and projected position operators, guaranteeing uniqueness and optimality (Lensky et al., 2014).

3. Properties, Realness, and Symmetry

For real, time-reversal invariant Hamiltonians with gapped, smooth band manifolds, MLWFs can always be chosen real (up to global phase), as proven in (Ri et al., 2014). The proof utilizes decomposition of complex Wannier functions into algebraic averages of real ones, showing the spread is minimized within the real gauge. Restricting to half the Brillouin zone further enhances efficiency for such cases.

In general, MLWFs do not preserve local chemical or symmetry properties—e.g., $\sigma/\pi$ separation in planar carbons—due to the full minimization of the Foster–Boys spread functional, resulting in "banana"-shaped orbitals (equal $\sigma$ and $\pi$ mixing). Alternative localization criteria, e.g., the Pipek–Mezey objective, strictly preserve $\sigma/\pi$ distinction but MLWFs (Foster–Boys) are nevertheless as localized as Pipek–Mezey Wannier functions according to either criterion (Jónsson et al., 2016).

Extensions allow selective localization (Wang et al., 2014), fixing centers, and enforcing point-group symmetry of subsets (OWFs) through partial spread functionals and Lagrange constraints.

4. Applications: Electronic Structure, Quantum Transport, and Beyond

MLWFs underpin (i) Wannier interpolation of band structures, density of states, and Fermi surfaces, (ii) calculation of electric polarization (modern theory), orbital magnetization, and Berry curvature (Marzari et al., 2011), and (iii) ab initio parameterization of tight-binding and Hubbard-like models in strongly correlated systems (Franchini et al., 2011).

In quantum transport, MLWF Hamiltonians provide the basis for Landauer–Büttiker calculations of coherent transport. The block-structured Hamiltonian enables massive system assembly using a "nearsightedness" principle, allowing the simulation of devices with $>10^4$ atoms (Shelley et al., 2011).

MLWFs are also used for:

• Phonon tight-binding models (localized vibrational modes).
• Photonic crystals (localized photon orbitals).
• Bose-Hubbard models in cold-atom lattices.
• High harmonic generation with smooth Berry connection for stable time-dependent simulations (Silva et al., 2019).

5. Automated, High-Throughput, and Robust Wannierization

Recent progress has focused on fully automated workflows for MLWF generation compatible with high-throughput frameworks (AiiDA, Wannier90), robust even for nontrivial band entanglement, topological materials, and systems with spin–orbit coupling or magnetization (Vitale et al., 2019, Jiang et al., 9 Jul 2025, Qiao et al., 2023, Mustafa et al., 2016). Core methods include:

• SCDM: Selected Columns of the Density Matrix provide physically motivated, parameter-free initial projections (Vitale et al., 2019).
• Projectability-disentanglement: Automated identification of strongly and weakly projectable states, followed by robust disentanglement/mobile window adjustment (Qiao et al., 2023, Jiang et al., 9 Jul 2025).
• OPF and self-projection: Elaboration of the OPF approach with exact gradient descent and trial basis augmentation (Tillack et al., 5 Feb 2025, Mustafa et al., 2015).

Benchmarks across hundreds to tens of thousands of materials validate that these schemes routinely achieve sub-20 meV interpolative accuracy up to 2 eV above the Fermi level (Jiang et al., 9 Jul 2025, Qiao et al., 2023, Vitale et al., 2019).

A key caveat is that for systems with nonvanishing Chern numbers, exponentially localized MLWFs are topologically forbidden (Mustafa et al., 2016, Cancès et al., 2016). For $\mathbb{Z}_2$ topological insulators, specialized OPF methods can nevertheless yield delocalized complex MLWFs spanning the entire nontrivial composite manifold, at the cost of increased spatial range and necessary complex phases.

6. Practical Implementation and Interoperability

MLWFs are constructed in practice using codes such as Wannier90, which interface with mainstream first-principles software (Quantum ESPRESSO, VASP, ABINIT, GPAW, Wien2k, SIESTA, FLEUR) (Shelley et al., 2011). The algorithmic pipeline typically involves:

1. Bloch eigenstate calculation on dense $k$-meshes.
2. Overlap/projection matrix computation (atomic orbitals, SCDM, PAOs, hydrogenic augmentations).
3. Optional disentanglement (outer and inner energy/projectability windows).
4. Initialization of the gauge (projections, SCDM, continuous transport, or OPF).
5. Iterative minimization of the spread functional, often with analytic gradient formulas.
6. Construction and storage of MLWFs for post-processing.

ASE/GPAW infrastructure (and ABINIT, NWChem, VASP) supports MLWF and Pipek–Mezey Wannier function routines in multiple wavefunction representations (real-space finite difference, plane-wave, LCAO, PAW) and fully enables post-processing, structural analysis, and seamless interoperability (Jónsson et al., 2016).

Advanced workflow integration supports automated defect/surface Hamiltonian "stitching" (Lihm et al., 2019), sub-manifold remixing for targeted projections (Qiao et al., 2023), and the use of extended projectors (PAO+hydrogenic) for challenging models (Jiang et al., 9 Jul 2025).

7. Uniqueness, Limitations, and Future Directions

In one dimension, the minimization of $\Omega$ is analytically reducible to phase fixing and projected position operator diagonalization, yielding a unique MLWF basis—the "1D determinacy principle" (Lensky et al., 2014).

In higher dimensions, nonconvexity can trap norm-based projection algorithms in false minima if initial projections are poor; parameter-free continuous gauge construction ensures correct topological sector selection (Cancès et al., 2016). For systems with topological obstructions (finite Chern numbers), exponentially localized MLWFs do not exist; for $\mathbb{Z}_2$ topological insulators, OPF algorithms produce extended, complex MLWFs capturing the nontrivial topology (Mustafa et al., 2016).

Future work includes robust automation for MLWF construction in systems with coexisting spin–orbit coupling and noncollinear magnetism, high-throughput parametrization for machine-learned materials models, and efficient algorithms for multiband, multi-manifold separation with symmetry and center constraints (Jiang et al., 9 Jul 2025, Wang et al., 2014, Qiao et al., 2023).