Periodic Gaussian Basis Set for Solids

• Periodic Gaussian basis sets are localized Gaussian functions modified via Bloch sums to enforce translational symmetry in crystalline systems.
• They utilize a dual real- and reciprocal-space integration strategy to efficiently evaluate overlap, kinetic, and Coulomb integrals for methods like Hartree–Fock and DFT.
• Design strategies focus on optimizing conditioning, transferability, and virtual-space accuracy for correlated and quasiparticle calculations in diverse solid materials.

A periodic Gaussian basis set is a basis construction in which localized Gaussian functions are adapted to translational symmetry, typically by forming Bloch sums of atom-centered Gaussian orbitals or by lattice-summing a Gaussian over all unit-cell translations. In this form, the basis retains the compactness, chemical locality, and analytic integral structure of Gaussian orbitals while satisfying periodic boundary conditions and supporting explicit Brillouin-zone sampling. Periodic Gaussian bases underlie Hartree–Fock, density-functional, post-Hartree–Fock, and many-body Green’s-function calculations for solids, and they also appear in mixed Gaussian–plane-wave, pseudized, and explicitly correlated formulations (Zhu et al., 2020, Sharma et al., 2020, Sun et al., 2017).

1. Formal construction and periodicity

In crystalline Gaussian electronic-structure methods, the one-particle basis functions are periodic Bloch sums of ordinary Gaussian atomic orbitals,

$$\phi_{\mu \mathbf{k}}(\mathbf{r})=\sum_{\mathbf{T}} e^{i\mathbf{k}\cdot \mathbf{T}}\,\tilde{\phi}_\mu(\mathbf{r}-\mathbf{T}),$$

where $\mathbf{T}$ runs over lattice translation vectors. This construction enforces lattice periodicity through Bloch’s theorem while preserving the real-space localization of the underlying atomic orbital. The resulting crystalline molecular orbitals are expanded as

$$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$

and the periodic Roothaan–Hall equation is written as

$$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$

This is the standard localized-orbital analogue of a plane-wave basis with $k$-point sampling (Zhu et al., 2020, Lee et al., 2021).

An equivalent viewpoint starts from a molecular Gaussian-type orbital and periodizes it directly by summing over all Bravais-lattice vectors. For a solid-harmonic Gaussian

$$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$

the periodic version is

$$D_a^{\mathrm{per}}(\mathbf r)=\sum_{\mathbf P} D_a(\mathbf r+\mathbf P).$$

Under this definition, a periodic Gaussian basis function is the infinite crystal repetition of a single Gaussian centered in a reference cell. Periodic two-center matrix elements then become lattice sums of molecular-like integrals over translated images, and crystal momentum conservation enters through relations such as

$$\mathbf{k}_{pq}=\mathbf{k}_p-\mathbf{k}_q+N\mathbf{b},$$

with $\mathbf{b}$ a reciprocal lattice vector (Sharma et al., 2020, Lei et al., 2022).

This dual description—Bloch-summed atomic orbitals or lattice-summed Gaussians—clarifies why periodic Gaussian bases are technically distinct from both isolated-molecule Gaussian bases and pure plane-wave discretizations. The basis is local in real space, but periodicity is imposed analytically rather than by a supercell grid.

2. Integral evaluation, Coulomb treatment, and density fitting

The central computational difficulty of periodic Gaussian bases is not the basis construction itself but the evaluation of lattice-summed overlap, kinetic, and Coulomb integrals. A general solution is to split the periodic lattice sum between real and reciprocal space using Poisson’s summation formula. In the two-center case, the real-space contribution converges rapidly for tight Gaussians, while the reciprocal-space contribution converges rapidly for diffuse Gaussians. Real-space derivatives are assembled with the McMurchie–Davidson recurrence relation, and reciprocal-space expressions reuse Gaussian Fourier factors efficiently. In the implementation reported for overlap, kinetic, and Coulomb kernels, the periodic algorithm is only about $5$–$\mathbf{T}$0 times slower than the corresponding molecular integral evaluation, which indicates that both real-space and reciprocal-space summations can be kept short in practice (Sharma et al., 2020).

For three-center Coulomb quantities and four-center electron-repulsion integrals, density fitting is the dominant acceleration strategy. In periodic Gaussian density fitting with the Coulomb metric, pair densities

$\mathbf{T}$1

are expanded in an auxiliary Gaussian basis. Range-separated Gaussian density fitting splits the Coulomb kernel into short-range and long-range parts,

$\mathbf{T}$2

with the short-range piece evaluated in real space and the long-range piece in reciprocal space. This range-separated GDF yields about $\mathbf{T}$3-fold speedups over the previously developed periodic GDF, shows sublinear to linear scaling with the number of $\mathbf{T}$4-points for small to medium meshes, and introduces errors of about $\mathbf{T}$5 in converged Hartree–Fock energies with default auxiliary basis sets (Ye et al., 2021).

A more elaborate variant is mixed Gaussian and plane-wave density fitting, in which an AO pair density is expanded in periodic Gaussian auxiliary functions together with a plane-wave component. The Gaussian sector captures compact, localized density near nuclei, while the plane-wave sector represents smooth long-range density and treats the periodic Coulomb singularity naturally. In hydrogen-crystal benchmarks, mixed density fitting with analytical Fourier transforms was reported to be $\mathbf{T}$6–$\mathbf{T}$7 orders of magnitude more accurate than pure plane-wave fitting with the same number of plane waves; in silicon, with $\mathbf{T}$8 plane waves and a linear-dependence threshold of $\mathbf{T}$9, the Hartree–Fock total-energy error was about $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$0 m$$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$1 (Sun et al., 2017).

Exchange evaluation is another bottleneck. A robust pseudospectral method for periodic Gaussian exchange avoids FFTs during each exchange build by representing products of Gaussian basis functions in a compact auxiliary basis constructed with ISDF and using occ-RI exchange. The formal scaling remains cubic, but the prefactor is reduced substantially, and the final-energy error decreases exponentially rapidly with the number of auxiliary functions (Sharma et al., 2022).

3. Basis-set design for solids: conditioning, transferability, and material specificity

Periodic Gaussian basis-set design is governed by a tension between flexibility and numerical stability. In solids, diffuse primitives overlap strongly with their periodic images, so the crystalline overlap matrix can become nearly singular. This linear-dependence problem is more severe than in molecular calculations, particularly in close-packed systems and at higher zeta levels. As a result, periodic Gaussian basis development has emphasized condition-number control, compact valence spaces, and explicit benchmarking against plane-wave references (Ye et al., 2021, Lee et al., 2021).

One strategy is material-specific optimization. A two-stage workflow first optimizes an atomic basis on the isolated atom using HF and then CISD, with pruning of nearly empty contractions and very diffuse primitives, and then re-optimizes the valence part in the target solid. The optimization objective combines the DFT total energy, a band-structure matching term against converged plane-wave data, and a penalty on overlap-matrix ill-conditioning,

$$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$2

This approach was tested on diamond, graphite, silicon, MoS$$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$3, and NiO. It improved band eigenvalues relative to existing periodic Gaussian bases and explicitly controlled overlap conditioning, with acceptable overlap condition numbers typically below $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$4 (Zhou et al., 2021).

A second strategy emphasizes transferability rather than material specificity. The unc-def2-GTH family is built by uncontracting def2 basis functions, removing primitives with exponent $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$5, and taking the union with uncontracted SZV-MOLOPT-SR-GTH primitives. The resulting uncontracted bases were benchmarked on $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$6 simple semiconductors. For unc-def2-QZVP-GTH, the reported basis-set incompleteness error is smaller than $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$7 m$$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$8 per atom in total energies and $$\psi_{n\mathbf{k}}(\mathbf{r})=\sum_{\mu} C_{\mu n}(\mathbf{k})\, \phi_{\mu\mathbf{k}}(\mathbf{r}),$$9 meV in band gaps for all systems considered. Linear dependencies are handled by canonical orthogonalization with $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$0 (Lee et al., 2021).

A third strategy revisits correlation-consistent basis design for solids. The GTH-cc-pV$$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$1Z basis sets adapt Dunning’s hierarchy to periodic calculations by limiting the number of primitive functions and making the valence space less diffuse than in molecular analogues, especially for atoms on the left side of the periodic table. The purpose is to avoid problematic small exponents while still achieving smooth convergence to the complete-basis-set limit. The paper identifies overlap-matrix condition numbers above about $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$2 as problematic in practice and validates the resulting bases in Hartree–Fock and MP2 calculations on $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$3 semiconductors and insulators (Ye et al., 2021).

These three approaches exemplify the current design space: one can optimize for a target material, optimize for broad transferability, or optimize for systematic correlation-consistent convergence. The literature suggests that no single criterion is sufficient by itself.

4. Role in correlated and quasiparticle methods

Periodic Gaussian bases are especially important in methods where virtual-space compactness matters. In all-electron periodic $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$4, a crystalline Gaussian basis allows direct inclusion of all virtual states in the basis without the explicit empty-band truncation common in plane-wave formulations. For cubic ZnO, the reported basis sizes were $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$5 functions per unit cell in cc-pVTZ and $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$6 in cc-pVQZ, and “no virtual band truncation is needed to compute the polarizability.” Across semiconductors and rare-gas solids, extrapolated $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$7@PBE band gaps showed a mean absolute relative error of $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$8 versus experiment. For core excitation binding energies, $$\mathbf F^{\mathbf k} \mathbf C^{\mathbf k} = \mathbf S^{\mathbf k} \mathbf C^{\mathbf k} \mathbf \epsilon^{\mathbf k}.$$9 yielded a mean absolute error of only $k$0 eV for the tested solids (Zhu et al., 2020).

The same basis philosophy extends to quasiparticle self-consistent $k$1. In periodic QSGW with crystalline Gaussian bases, Gaussian density fitting, full-frequency analytic continuation, and Brillouin-zone sampling are combined with a static effective potential $k$2. The method was benchmarked for semiconductors, rare-gas solids, and transition-metal oxides including MnO, FeO, CoO, and NiO. The reported mean absolute relative error in band gaps for the semiconductor and rare-gas benchmark set was $k$3, compared with $k$4 for $k$5-PBE and $k$6 for PBE. The same study found that QSGW systematically overestimates band gaps but removes dependence on the starting density functional and gives qualitatively correct insulating behavior for all four tested transition-metal oxides (Lei et al., 2022).

At the wavefunction level, periodic Gaussian bases have also been shown to support reliable MP2 convergence. The GTH-cc-pV$k$7Z family yields smooth and systematic convergence toward the complete-basis-set limit in periodic MP2, and the Gaussian basis can be used not only as a stand-alone periodic basis but also as a projected virtual space that accelerates plane-wave MP2 convergence (Ye et al., 2021). This suggests that periodic Gaussian bases are not merely a localized alternative to plane waves; they are a virtual-space compression tool for correlated electronic structure.

5. Pseudized, mixed, and explicitly correlated variants

A distinct line of development embeds Gaussian basis functions inside a plane-wave/PAW framework. In this approach, atom-centered contracted Gaussian functions are pseudized within the PAW cutoff radius so that their Fourier expansion converges at a moderate kinetic-energy cutoff. The pseudized Gaussian radial function is replaced inside $k$8 by a smooth combination of spherical Bessel functions chosen to match value, logarithmic derivative, and norm conservation at the cutoff. The resulting pseudized Gaussian-type orbitals are represented by their plane-wave coefficients in a periodic code, and a hybrid construction uses a converged plane-wave occupied space together with an orthogonalized Gaussian virtual space for correlated methods. This procedure was demonstrated for the water dimer, solid neon, and water adsorption on LiH at the MP2 level (Booth et al., 2016).

Periodic Gaussian constructions are not limited to one-particle orbital bases. In a 2026 explicitly correlated Gaussian formulation, shifted correlated Gaussian primitives

$k$9

are periodized by summing over all composite lattice translations,

$$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$0

and a generalized unfolding theorem reduces the resulting double lattice sums in matrix elements to single sums. Closed-form overlap, kinetic, Coulomb, and contact-density expressions are then obtained within a unified framework. For an infinite one-dimensional hydrogen chain, the reported $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$1-point energy with $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$2 optimized basis states at $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$3 was $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$4 Hartree, in close agreement with the benchmark variational value $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$5 Hartree (Varga, 12 May 2026).

These variants show that “periodic Gaussian basis set” is broader than the standard crystalline atomic-orbital picture. It includes pseudized Gaussian orbitals represented in reciprocal space, mixed occupied/virtual constructions, and explicitly correlated Gaussian functions periodized at the many-electron level.

6. Conceptual boundaries, misconceptions, and recurring limitations

The term “periodic Gaussian basis set” is often used loosely, but the literature distinguishes several non-equivalent constructions. “Sigma basis sets” are a family of contracted GTO basis sets for molecular calculations; they were introduced and tested only for atoms and homonuclear diatomics, and the paper explicitly states that it does not present periodic-solid calculations, crystal calculations, or explicit Bloch-periodic Gaussian basis development (Ema et al., 2022). Similarly, the “universal lattice basis” is a periodic Shannon/Fourier lattice basis that represents Gaussian-like functions with quadratic convergence, but it is not a Gaussian basis in the standard quantum-chemistry sense (Jerke et al., 2013). “Gausslets” are orthogonal, smooth, grid-based functions built from sums of Gaussians; they are translationally repeated on a lattice, yet the paper does not focus on Bloch-periodic crystalline Gaussian bases (White, 2017).

There is, however, at least one periodic Gaussian construction outside the crystalline atomic-orbital tradition that is still properly Gaussian: the periodic von Neumann basis. Its basis functions are phase-space Gaussians projected onto a periodic Fourier grid, producing the periodic von Neumann basis

$$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$6

which is exactly equivalent to the Fourier Grid Hamiltonian space on the same finite periodic domain. The same work shows that a biorthogonal variant can be pruned without loss of accuracy and can approach the efficiency of $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$7 basis function per $$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$8 eigenstate in the classical limit (Shimshovitz et al., 2010). This is a periodic Gaussian basis, but not a conventional solid-state LCAO basis.

The principal limitations of conventional periodic Gaussian bases recur across the literature. Diffuse primitives can induce near-linear dependence; larger nominal zeta quality is not automatically better; conduction bands and low-lying virtual states can be described poorly even when occupied-state energetics are acceptable; and compressed geometries exacerbate conditioning problems. Material-specific optimization may repair band crossings or missing orbital character, as shown for MoS$$D_a(\mathbf r)= S_{l_a m_a}(\mathbf r-\mathbf A)\, e^{-a|\mathbf r-\mathbf A|^2},$$9 and NiO, but this reduces universality (Zhou et al., 2021). Highly transferable uncontracted bases can approach the basis-set limit, but they are large and heavily linearly dependent (Lee et al., 2021). Correlation-consistent periodic families improve stability by becoming less diffuse than molecular analogues, which suggests that the molecular intuition “more diffuse means systematically better” is often incorrect in dense periodic environments (Ye et al., 2021).

Taken together, these results establish the periodic Gaussian basis set as a family of localized, symmetry-adapted basis constructions rather than a single canonical object. Its defining feature is not merely the use of Gaussians, but the controlled reconciliation of locality, translational symmetry, Coulomb singularities, virtual-space completeness, and overlap-matrix conditioning in extended systems.