Natural Atomic Orbitals (NAO)

• Natural atomic orbitals (NAOs) are defined as eigenfunctions of the one-particle density matrix, providing a compact and physically clear basis for electron distributions.
• They are constructed via atomic block diagonalization and orthogonalization, which facilitates efficient basis set compression and stable embedding in quantum methods.
• NAOs enable robust population analysis and serve as projectors in methods like DFT+DMFT, significantly reducing computational cost while maintaining high accuracy.

Natural atomic orbitals (NAOs) are orthonormal, atom-centered functions derived by diagonalizing atomic blocks of the one-particle density matrix, providing a compact, physically transparent, and mathematically rigorous basis for quantum chemistry and electronic structure theory. By construction, NAOs order themselves by occupation—offering a systematic approach to basis set compression, accurate population analysis, and stable projective embeddings in methods spanning correlated wavefunction theory to density functional theory (DFT) and beyond. Owing to their locality and controlled occupancy, NAOs are extensively utilized in basis set construction, DFT+DMFT embeddings, and as the foundation for lossy compression of large atomic orbital sets, enabling significant computational efficiency gains with well-controlled accuracy.

1. Mathematical Definition and Physical Properties of NAOs

NAOs are defined as the eigenfunctions of the one-particle density matrix, constructed either globally or within atomic subspaces. In a general orthonormal AO basis $\{\phi_p\}$, the one-particle density matrix for a correlated wavefunction $\Psi$ is

$$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$

where $a^\dagger_p$, $a_q$ are fermionic creation and annihilation operators. NAOs $\{\chi^i\}$ are the eigenfunctions of this matrix,

$\sum_q \gamma^1_{pq} \chi^i_q = n_i \chi^i_p,$

with occupation numbers $n_i \in [0,2]$ for spin-restricted systems, and expansion coefficients $\chi^i_p = \langle \phi_p | \chi^i \rangle$ (Petruzielo et al., 2010). In atom-centered schemes, the diagonalization is restricted to atomic blocks (or their one-center orthonormalized forms), yielding strictly local NAOs per atom (Lara et al., 27 Nov 2025).

NAOs are orthonormal and provide an optimally compact description of the single-particle density, as measured by the sum of leading occupation numbers. Orbitals with $n_i \approx 2$ (fully occupied) and $\Psi$0 (empty) can be systematically separated from partially occupied, strongly correlated orbitals—a property leveraged in population analyses and embedding strategies (Sim et al., 2019).

2. Construction Algorithms and One-Center Orthogonalization

The typical algorithm for NAO construction proceeds as follows (Lara et al., 27 Nov 2025):

• Remove near-linear dependencies from the full AO basis by iteratively deleting AOs that, if removed, raise the smallest eigenvalue of the overlap matrix $\Psi$1 above a numerical cutoff ($\Psi$2–$\Psi$3).
• Partition the basis by atom. For each atomic subset, form the atomic overlap block $\Psi$4 and perform symmetric Löwdin orthogonalization:

  $\Psi$5

• Construct one-center orthogonalized AOs $\Psi$6. Assemble the global orthogonalizer $\Psi$7.
• Transform the full (contravariant) density matrix $\Psi$8 into this representation: $\Psi$9.
• For each atom $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$0, extract the atomic block $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$1 and diagonalize:

  $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$2

  yielding NAO coefficients $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$3 and occupation numbers $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$4.

• Reconstruct each (global) NAO as $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$5.

This approach produces a set of atom-centered, orthonormal NAOs that diagonalize the occupied and virtual atomic density, with direct physical correspondence to atomic populations (Lara et al., 27 Nov 2025, Sim et al., 2019). In the more general nonorthogonal basis context, the NAO construction solves the generalized eigenproblem $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$6, where $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$7 is the orbital overlap matrix and $$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$8 is the one-particle density matrix in the chosen basis (Sim et al., 2019).

3. Basis Set Construction and Lossy Compression

NAOs are used both as the contracted components of atomic basis sets and as a foundation for optimal lossy compression of large AO spaces:

• In multi-configurational self-consistent field (MCSCF) and related wavefunction-based methods, NAOs are found by diagonalizing the one-particle density from a correlated atomic calculation, usually incorporating active spaces tailored to atomic valence and near-valence shells (e.g., CAS($$\gamma^1_{pq} = \langle \Psi | a_p^\dagger a_q | \Psi \rangle,$$9) procedures) (Petruzielo et al., 2010). Dominant NAOs for each angular momentum are used as contracted basis functions—supplemented by optimized primitives such as Gauss–Slater functions for smooth pseudopotentials.
• For large-basis DFT calculations, NAOs serve as the vehicle for atom-centered lossy compression: after constructing NAOs and ordering by occupation numbers, a threshold $a^\dagger_p$0 is chosen, and all NAOs with $a^\dagger_p$1 are discarded. This yields a reduced basis of $a^\dagger_p$2 functions, with compression factors $a^\dagger_p$3 typically 2–5 for quadruple-zeta (QZ) atomic bases (Lara et al., 27 Nov 2025). With $a^\dagger_p$4 ($a^\dagger_p$5), relative energies are preserved to $a^\dagger_p$6, and with $a^\dagger_p$7 accuracy tightens to $a^\dagger_p$8. After truncation, standard SCF procedures are executed in the compressed NAO basis, dramatically reducing cubic-scaling steps such as diagonalization and density-matrix updates.
• In numerical atomic orbital (NAO) schemes—especially in DFT—a hierarchy of NAOs can be built by contraction of nearly complete localized bases (e.g., truncated spherical waves), with the contraction optimized via variational minimization (e.g., of the kinetic-energy trace in the deficiency space) (Huang et al., 14 Mar 2026). Such schemes yield systematically improvable, strictly localized NAOs for both molecules and solids, enabling highly accurate electronic structure predictions with controlled transferability and avoidance of artifacts from periodic boundary conditions.

4. NAOs as Projectors and Embeddings in Correlated Subspaces

NAOs provide a mathematically natural, numerically stable choice for defining correlated subspaces in embedding and many-body methods. In DFT+DMFT frameworks, NAOs are used as projectors $a^\dagger_p$9 onto local correlated subspaces for dynamic mean-field calculations (Sim et al., 2019). The key steps:

• From a (possibly nonorthogonal) set of pseudo-atomic orbitals (PAOs), compute the overlap matrix $a_q$0 and one-particle density matrix $a_q$1.
• Solve $a_q$2, constructing the NAOs and their occupancies.
• Retain only NAOs with intermediate occupancies (e.g., $a_q$3), which correspond to strongly fluctuating, physically relevant correlated orbitals.
• The subspace projector is $a_q$4 for each atom $a_q$5 and orbital $a_q$6 of interest.
• During DFT+DMFT self-consistency, NAOs may be updated at each iteration, preserving the locality, occupation count, and orthonormality of the correlated subspace. The construction is phase-invariant and insensitive to PAO set completeness or extension.

NAOs also facilitate computational efficiency by enabling automatic "orbital splitting": orbitals with occupancy near 0 or 2 can be treated perturbatively, while those with $a_q$7 are solved with numerically demanding impurity solvers (e.g., CTQMC), thus reducing solver dimensionality and overall cost (Sim et al., 2019).

5. Numerical Benchmarks and Performance Assessment

The accuracy and efficiency of NAO-based compression and basis construction have been systematically assessed in multiple contexts:

• In QZ-level bases, NAO-based compression at $a_q$8 yields compression factors between 2.5 and 4.5, with energy errors (relative) typically less than $a_q$9. Tightening to $\{\chi^i\}$0 reduces absolute errors to $\{\chi^i\}$1–$\{\chi^i\}$2 a.u. and relative errors to $\{\chi^i\}$3 (Lara et al., 27 Nov 2025).
• For double- and triple-zeta NAO+GS bases, CCSD total energies improve by several mHartree per electron relative to Burkatzki–Filippi–Dolg sets; atomization energies recover an additional $\{\chi^i\}$4–$\{\chi^i\}$5 at DZ, with similar improvements observed across HF, B3LYP, and Quantum Monte Carlo (Petruzielo et al., 2010).
• In NAO sets built from contracted truncated spherical waves, the pVTZ hierarchy achieves total energy errors below $\{\chi^i\}$6 per atom, atomization energy mean absolute errors (MAEs) below $\{\chi^i\}$7, and bond-length MAEs below $\{\chi^i\}$8 for molecules. For bulk crystals, energy errors are within chemical accuracy, and band gap MAEs are $\{\chi^i\}$9 (Huang et al., 14 Mar 2026).

NAOs have also demonstrated high stability with respect to variations in the primitive set, robust localization properties (spread $\sum_q \gamma^1_{pq} \chi^i_q = n_i \chi^i_p,$0 values near those of maximally localized Wannier functions), and accurate charge populations in projected subspaces (Sim et al., 2019).

6. Comparison with Alternative Localized and Projected Constructs

NAOs are distinguished from alternative orbital-localization and projector-construction strategies such as maximally localized Wannier functions (MLWFs), Mulliken or Löwdin orthogonalizations, and PAO-Mulliken populations:

• NAO construction requires no energy window or trial orbital selection, is phase-invariant, and provides a unique decomposition of the density matrix.
• Population analysis in the NAO basis (direct sum of occupations) avoids ambiguities and overestimation present in PAO-Mulliken or Löwdin schemes.
• NAO projectors saturate integrated density-of-states metrics (IDOS) essentially as fast as MLWFs and without the convergence failures sometimes observed for MLWFs in entangled bands (Sim et al., 2019).
• In practice, NAOs generally yield robust localization, physicochemical assignments, and transferability between active-space choices and extended systems.

A plausible implication is that NAOs offer a physically meaningful middle ground between strictly local (but possibly over-complete) atomic bases and global (but delocalized) canonical orbitals, supporting both efficient computation and rigorous chemical interpretation across correlated and DFT-based quantum simulations.

7. Systematic Improvability and Guidelines for Practical Use

The systematic improvability of NAO-based schemes is a distinguishing feature:

• In numerical and analytic basis construction, completeness can be systematically achieved by raising the radial cutoff, increasing the angular momentum range, or including additional $\sum_q \gamma^1_{pq} \chi^i_q = n_i \chi^i_p,$1 functions per angular momentum channel (Huang et al., 14 Mar 2026).
• Hierarchies of NAO-based bases (minimal, double-, triple-, quadruple-zeta, and virtual-extended) are constructed to reach desired accuracy in total energies, structure, and electronic spectra, directly analogous to Dunning's correlation-consistent strategy (Petruzielo et al., 2010, Huang et al., 14 Mar 2026).
• For transferability beyond valence (e.g., conduction-band or unoccupied-state properties), inclusion of virtual states in contraction fitting or spillage minimization systematically improves high-lying state reproduction (Huang et al., 14 Mar 2026).

In summary, NAOs form the foundation for a wide array of computational strategies in electronic structure, spanning efficient basis reduction, robust embedding, and chemically interpretable electron populations, while offering systematically controlled accuracy and localization properties (Lara et al., 27 Nov 2025, Petruzielo et al., 2010, Sim et al., 2019, Huang et al., 14 Mar 2026).