Position dipoles of the PAO basis in the velocity operator#

The problem#

PAOFLOW’s velocity operator in the orthonormal PAO Bloch basis is the k-derivative of the Hamiltonian plus the inter-site term, v = dH/dk + i[H, τ] (the Dnm term of hamiltonian/do_gradient): the position (elphon/elph_bloch.band_velocities, the Dnm term of hamiltonian/do_gradient): the position operator is taken diagonal on the orbital centres τ_i. For orthonormal localised orbitals ω the exact tight-binding velocity has a third term,

v(k) = dH/dk + i[H, τ] + i[H, d(k)],
d_ij(R) = <ω_i(r − τ_i) | r − R − τ_j | ω_j(r − R − τ_j)>,

whose band-basis form i(ε_n − ε_m) d_nm is the analogue of EPW’s Berry-connection term in vme = 'wannier'. Its largest part is the on-site Δl = ±1 dipole (⟨s|r|p⟩, ⟨p|r|d⟩) of the same atom. It vanishes on the diagonal and inside degenerate blocks, so group velocities (SERTA/MRTA, adaptive smearing, Drude) are untouched.

The implementation (hamiltonian/position_dipoles.py)#

PAOFLOW’s basis functions are the atomic orbitals φ symmetrically (Löwdin) orthogonalised, ω = φ S^{-1/2}. With the two-centre tables of the atomic orbitals — the overlap S(R) and the intrinsic dipole X(R) with the origin on the bra site (_two_center, the PAO radial function as bra with g = rR at l ± 1, vectorised per site pair in two_center_blocks.py) — the Bloch-space dipole of the orthonormal basis is

d(k) = S^{-1/2} X S^{-1/2} + S^{-1/2} τ S^{1/2} − τ + i S^{-1/2} ∂_k S^{1/2}

(all at k; S(k) = Σ_R e^{ik·R} S(R), likewise X; ∂_k S^{1/2} from the eigendecomposition of S). The three extra terms come from the cell-position operator, which does not commute with S^{±1/2}, and from the bra-site origin of X; d(k) is Hermitian.

function

role

enumerate_pao_pairs

(bra site, ket site, ΔR) within the orbital cutoffs

build_position_tables

S and X blocks per pair, one vectorised evaluation per pair (two_center_blocks.two_center_pair_blocks: all channels, L and m at once; Si 44 orbitals, 550 pairs: 3 s instead of 41 s)

assemble_position_kspace

S(k), ∂_k S(k), X(k) by Bloch sums e^{ik·ΔR}

lowdin_dipole_kspace

the formula above; level='intra-atomic' keeps the on-site blocks with S ≡ 1

dipole_velocity_term

i[H(k), d(k)] in the orbital basis

position_dipole_operator

data-controller-facing callable K_cryst -> d(k)

Hooks: gradient_and_momenta(position_dipoles=True, position_dipole_level='full') (or attr['position_dipoles'] = True; off by default, see the dHksp route below), raman_spectrum(position_dipoles=True) (default on) and the optical run emitted by paoflow-gen (gradient_and_momenta(position_dipoles=True)). | position_dipole_operator | data-controller-facing callable K_cryst -> d(k), the position_dipoles input of band_velocities |

Hooks: band_velocities(..., position_dipoles=), phonon_assisted_absorption_dense_q(..., position_dipoles=data_controller, position_dipole_level='full'), the Si example flag POSITION_DIPOLES (--position-dipoles, --position-dipole-level). Off by default; carrier_scattering is unaffected (diagonal velocities). Scalar-relativistic bases only (NotImplementedError for dftSO).

Conventions (measured, not assumed)#

PAOFLOW’s atomic Bloch orbitals (calc_atwfc_k) are built in reciprocal space with the phase i^l and the calc_ylmg harmonics, which carry an extra −1 on odd |m|. Transformed back to real space they are (−1)^l (−1)^{|m| mod 2} times R(r) Y_lm(r̂) with the tesseral harmonics of _two_center. tesseral_signs applies both factors on both orbital indices. Verified on Si (3S 3P 3D 4S 4P, 26 orbitals): the two-centre S(k) reproduces the plane-wave overlap of the PAOFLOW orbitals to 1e-3 (raw values up to 8.9 at Γ for the diffuse 4S/4P/3D Bloch sums), and the Löwdin dipoles agree with the exact Berry connection i⟨u_i|∂_k u_j⟩ of the Löwdin plane-wave orbitals to 2 % (max |d| 3.9 bohr). The odd-m rule alone (the convention of the former non-local term) flips every s–p / p–d element.

The on-site blocks alone ('intra-atomic') differ from the exact d by up to 2.7 bohr for this basis: with diffuse shells the Löwdin terms are the bulk of d, not a correction.

Validation (EPW tutorial-06 Si data, EPW_opt_audit/compare/pw_velocity_check.py, dipole_pw_diagnostic.py)#

At the coarse nscf k-points the exact plane-wave velocity of the Kohn–Sham states, v = 2⟨n|k+G|m⟩ + ⟨n|i[V_NL, r]|m⟩ (Ry bohr; KB commutator from the k-derivative of the projector form factors, checked by Hellmann–Feynman against the PAO group velocities), is compared with the PAO velocity in gauge-invariant block sums of |v|²:

pairs

plain PAO / exact

with d / exact

valence → CB1–CB4

1.4–2.3

0.96–1.13

valence top → CB1

0.9–1.6

0.9–1.1 (1.45 at Γ)

CB1 → CB2 at X

4

0.5

Direct ε₂ of Si at 12³ k against EPW (vme='wannier', which contains the Berry term): plain 1.06, with d 0.98 (ω-weighted integral 2.5–5 eV). The CB1 → CB2 transition at the conduction minimum is symmetry-forbidden on the valley axis in both the PAO and the plane-wave velocity; off-axis it has the plane-wave magnitude already without d.

The dHksp route (gradient_and_momenta(position_dipoles=True))#

PAOFLOW.position_dipole_correction() builds d(k) on the FFT k-grid (position_dipoles_on_grid: scalar, ad-hoc spin-orbit and fully relativistic (j, m_j) bases, the latter through scalar_to_jm_matrix with the per-channel (−1)^l) and adds −i[H(k), d(k)] to dHksp (inject_dipole_velocity_into_dHksp; H(k) = V E V† from pao_eigh, since do_gradient overwrites Hksp). Every property built on pksp then carries the full commutator: dielectric_tensor, EELS, optical conductivity, anomalous and spin Hall, the ac Hall conductivities, Rashba–Edelstein. Checks: on Si the pipeline reproduces the electron–phonon implementation (∫ε₂ ×0.957 / ×0.993 over 2.5–5 / 2.5–10 eV against ×0.953 / ×0.995) and ε₁(0) 13.54 against the exact 13.62; Fe AHC and Pt SHC (12³ scf, 32³ interpolation, Workspace/PAOFLOW/validation/position_dipoles_hall/): σ_xy(E_F) 494 → 490 S/cm and σ^z_xy(E_F) 2045 → 2117 (ħ/e) S/cm, i.e. −1 % and +3.5 %, with the curves unchanged in shape — the few-per-cent effect expected from the Kubo structure (the dipole cross term carries one power of 1/Δε, the main term two). The former non-local term (repaired, before its removal) on the same runs moved them by +10 % / −7 % at E_F and doubles the Fe peak at −1.35 eV (1700 → 2920 S/cm), the signature of the double counting amplified by 1/Δε² at the d-band near-degeneracies.

QE epsilon.x on the same system (full 12³ grid, nosym/noinv) reproduces the plane-wave local-only kernel to 2 % (∫ε₂ 2.5–10 eV: 91.1 vs 92.9; exact 78.4): it is the local-momentum reference and lies 15–19 % above the exact spectrum, which is why agreement with epsilon.x is not a validation of a velocity operator.

Open points#

  • Any pseudopotential type: the catalog of load_pao_orbitals reads only the atomic orbitals (the basis tables of the projection, else the UPF PP_PSWFC block), so ultrasoft and PAW saves run the dipole route like norm-conserving ones (the norm-conserving guard of the removed term went with it on 2026-10-22; checked on ultrasoft Si and PAW GaAs).

  • PAW augmentation: the partial-wave correction of the position matrix (PP_FULL_WFC) and the q_ij term of the overlap (ProjectorAugmentation, build_projector_augmentation, augmentation_kspace) are added to S(k), ∂_k S and X(k) by position_dipole_operator(augmentation='auto'), scalar-relativistic only. The dHksp route (position_dipoles_on_grid) does not apply it yet. With an all-electron-only basis S + ΔS can be nearly singular (a RuntimeWarning flags ill-conditioned PAO overlaps); the default basis route of projections(pseudo_orbitals='auto') takes the valence shells of ultrasoft / PAW species from the UPF pseudo-orbitals (internal-basis.md).

  • The former non-local correction Δp was removed in October 2026 (nonlocal-velocity.md): v = i[H, τ + d] is already the full commutator of a Hamiltonian that contains V_NL, and adding Δp double counted (exact plane-wave ε₂ of Si: dipoles 0.99–1.01, dipoles + Δp 0.87).

  • The Hubbard term of eACBN0 electrons is inside H(R), so i[H, τ + d] contains its velocity commutator; nothing else is needed.

  • VASP input (November 2026): build_projector_augmentation reads the partial waves, projectors and charges of the PAW datasets of the POTCAR through inputs/pseudopotentials.py (the real-space projectors are the inverse transforms of the reciprocal tables, re-biorthogonalised to the pseudo partial waves, see internal-basis.md), so the PAW augmentation of the dipoles applies to VASP runs as to QE PAW runs; not validated against a reference yet.

References#

Here are the references, mapped to each statement of the paragraph.

Velocity in a localised basis: ∂H/∂k plus the intersite term, position operator diagonal on the orbital centres

  • M. Graf and P. Vogl, “Electromagnetic fields and dielectric response in empirical tight-binding theory”, Phys. Rev. B 51, 4940 (1995). The standard derivation of the tight-binding velocity from the k-derivative of H(k) with the Peierls coupling, with the position operator diagonal on the atomic sites; discusses what that approximation drops.

  • L. C. Lew Yan Voon and L. R. Ram-Mohan, “Tight-binding representation of the optical matrix elements: theory and applications”, Phys. Rev. B 47, 15500 (1993). Momentum matrix elements in an orthogonal tight-binding basis, including the intra-atomic contributions.

  • B. A. Foreman, “Consequences of local gauge symmetry in empirical tight-binding theory”, Phys. Rev. B 66, 165212 (2002). The gauge-consistent form of the coupling and why the site-diagonal position operator is an approximation.

The third term, the intra-atomic (on-site Δl = ±1) dipoles ⟨s|r|p⟩, ⟨p|r|d⟩

  • T. G. Pedersen, K. Pedersen and T. B. Kriestensen, “Optical matrix elements in tight-binding calculations”, Phys. Rev. B 63, 201101(R) (2001). Shows that the intra-atomic dipole matrix elements are the largest correction to the site-diagonal position operator and quantifies them for semiconductors.

  • T. B. Boykin and P. Vogl, “Dielectric response of molecules in empirical tight-binding theory”, Phys. Rev. B 65, 035202 (2001). Intra-atomic dipoles in the dielectric response.

  • T. Sandu, “Optical matrix elements in tight-binding models with overlap”, Phys. Rev. B 72, 125105 (2005). The same construction for a non-orthogonal basis, which is the Löwdin-orthogonalised case of the PAO basis (S^{-1/2} terms).

  • C. D. Pemmaraju et al., “Velocity-gauge real-time TDDFT within a numerical atomic-orbital basis set”, Comput. Phys. Commun. 226, 30 (2018). The complete velocity operator in an atomic-orbital DFT basis, with the position-dependent (intra-atomic) terms.

  • P.-O. Löwdin, J. Chem. Phys. 18, 365 (1950), for the symmetric orthogonalisation ω = φ S^{-1/2} itself.

Band-basis form i(ε_n − ε_m) d_nm and its identity with the Berry-connection term of the Wannier velocity (EPW vme = ‘wannier’)

  • E. I. Blount, “Formalisms of band theory”, Solid State Physics 13, 305 (1962). The position operator in the Bloch representation, r = i∂_k + A, and the interband relation between position and velocity matrix elements.

  • X. Wang, J. R. Yates, I. Souza and D. Vanderbilt, “Ab initio calculation of the anomalous Hall conductivity by Wannier interpolation”, Phys. Rev. B 74, 195118 (2006), and J. R. Yates, X. Wang, D. Vanderbilt and I. Souza, “Spectral and Fermi surface properties from Wannier interpolation”, Phys. Rev. B 75, 195121 (2007). The Wannier-gauge velocity ∂H^W/∂k − i(ε_n − ε_m)A_nm with A the Wannier Berry connection; the diagonal (group velocity) is untouched by the A term.

  • N. Marzari and D. Vanderbilt, Phys. Rev. B 56, 12847 (1997), for the position matrix elements ⟨0n|r|Rm⟩ between localised functions that define A.

  • J. Ibañez-Azpiroz, S. S. Tsirkin and I. Souza, “Ab initio calculation of the shift photocurrent by Wannier interpolation”, Phys. Rev. B 97, 245143 (2018), for the explicit use of r_nm = v_nm/(iω_nm) with the Wannier position matrix elements.

  • EPW implementation: S. Poncé, E. R. Margine, C. Verdi and F. Giustino, Comput. Phys. Commun. 209, 116 (2016); H. Lee et al., “Electron-phonon physics from first principles using the EPW code”, npj Comput. Mater. 9, 156 (2023), which documents the vme = ‘wannier’ and ‘dipole’ options (the routine is vmewan2bloch in wannier2bloch.f90, following Yates et al.).

Diagonal elements are the group velocities and the dipole term vanishes inside degenerate blocks

  • Hellmann–Feynman for bands: Blount (1962) above; also Yates et al. (2007), where the band velocity is the diagonal of the Bloch-gauge ∂H/∂k and the Berry term drops out of it. The vanishing within degenerate blocks follows from the prefactor (ε_n − ε_m) and is stated in the same references.

Length-velocity relation, non-local potentials and the scissor (the context in which the term was identified)

  • A. F. Starace, Phys. Rev. A 3, 1242 (1971): length and velocity forms are equivalent only if the commutator [V, r] of non-local potentials is included.

  • A. J. Read and R. J. Needs, Phys. Rev. B 44, 13071 (1991): optical matrix elements with non-local pseudopotentials (the i[V_NL, r] term in the plane-wave velocity, which is contained in the PAO Hamiltonian and must not be added again).

  • R. Del Sole and R. Girlanda, Phys. Rev. B 48, 11789 (1993), and Z. H. Levine and D. C. Allan, Phys. Rev. Lett. 63, 1719 (1989): velocity renormalisation under a scissor (the Levine–Allan factor used in the study).

For the manuscript, the minimal citation set for that paragraph is Graf–Vogl 1995 (site-diagonal position operator), Pedersen 2001 (intra-atomic dipoles), Sandu 2005 (non-orthogonal basis), Wang 2006 and Yates 2007 (Berry-connection form of the Wannier velocity), and Lee 2023 (EPW’s vme = ‘wannier’).