betaFeedback

WorksPhysics and Astronomy

Semilocal exchange functionals from the exact-exchange condition for the hydrogen atom: Hydrogenic exactness and recovery of Rydberg-like bound states

Abstract

In semilocal density functionals the exchange potential decays too rapidly outside atoms and molecules and lacks the correct −1/r-1/r tail; as a consequence, functionals from PBE to modern meta-GGAs such as SCAN support no bound Rydberg-like state of an atom. We revisit the gradient-corrected exchange functional that Gill and Pople (GP93) constructed to reproduce the exact exchange potential of the hydrogen atom, solve their equation as an inverse problem, and use the resulting enhancement factor – the GP93 factor – as a tail-generating ingredient of a switched semilocal functional. The GP93 factor is exact on the hydrogen 1s1s density and, by uniform coordinate scaling, on hydrogenic 1s1s ions, where it yields the eigenvalue −Z2/2-Z^2/2 exactly (hydrogenic exactness). Its large-gradient growth has the hydrogen-exact form O[s(ln⁡s)2/3]O[s(\ln s)^{2/3}], distinct from the sln⁡ss\ln s form of Armiento and Kümmel. The factor is combined with a kinetic-energy-density indicator so that its divergent branch acts only in one-electron-like regions; the GP93 ingredient is nonempirical, while a few switching parameters are calibrated to balance tail recovery against self-consistent-field stability. The resulting functional produces Rydberg-like series of bound virtual Kohn–Sham states in systems whose outermost shell is a one-electron-like ss shell – all-electron H and He, and Li, Na, and K with large-core pseudopotentials – and moves the highest-occupied eigenvalue toward the experimental ionization energy. For pp-shell atoms (Ne, Ar) the switch closes, so the domain of applicability is fixed by design. Fixed-density tests indicate that a reduced-Laplacian switch distinguishes atomic tails from covalent bond centers, and that a density-only kinetic-energy functional renders the entire switch orbital-independent, so that no generalized Kohn–Sham solver is required.

Introduction

The semilocal exchange-correlation functionals that power practical Kohn–Sham [1, 2] density-functional theory—GGAs such as PBE [7] and meta-GGAs such as TPSS [12] and SCAN [13]—have been remarkably successful for thermochemistry, structure, and cohesive properties. Their exchange potential vx(r)v_x(\mathbf{r}), however, decays exponentially in the outer region of a finite system, where the density itself decays exponentially, and therefore lacks the exact Coulombic tail

vx(r)⟶r→∞−1r.(1)v_x(\mathbf{r}) \underset{r\to\infty}{\longrightarrow} -\frac{1}{r}. \tag*{(1)}

A well-known family of failures follows from this single structural defect: unbound Rydberg states, unstable negative ions, systematically underestimated ionization energies [4, 29], and the excessive charge delocalization caused by the self-interaction error (SIE). We begin with the numerical fact that motivates this work: in the Kohn–Sham virtual spectrum of the He atom, PBE, SCAN, TPSS, and M06-L support no bound Rydberg orbital at all (Sec. VI, Figure 3). Even the latest meta-GGAs, despite the additional variable τ\tau, do not restore Eq. (1).

Figure 3. Helium atom. (A) −εHOMO-\varepsilon_{\mathrm{HOMO}} and number of bound Rydberg orbitals: PBE, SCAN, TPSS, and M06-L all bind zero, while the integrated FmixF_{\mathrm{mix}} binds eight with −εHOMO=25.1-\varepsilon_{\mathrm{HOMO}}=25.1 eV (experiment 24.59). (B) The two bound levels (neff=1.84n_{\mathrm{eff}}=1.84 and 3.63). FmixF_{\mathrm{mix}} binds the lowest two Rydberg-like levels; shallower high-nn levels remain unbound owing to the residual of the potential tail (a systematic error of the hydrogen-referenced calibration). Even so, the existence of bound levels is a qualitative difference from the semilocal and hybrid functionals, all of which bind none. (C) −rvx(r)-r v_x(r): FmixF_{\mathrm{mix}} retains a finite tail, PBE collapses to zero. That FmixF_{\mathrm{mix}} does not reach the exact value 1 asymptotically is the same systematic calibration error responsible for the slight overestimate of the IP (disclosed as part of the scope).

Existing attempts to repair the asymptotic fall into two broad classes. Model potentials (LB94 [8], mBJ [10]) construct the shape of vxv_x directly but possess no parent energy functional and hence lack variational consistency. Range-separated hybrids [22, 23] restore the long-range tail through nonlocal exact exchange, at the corresponding computational cost. Within the semilocal energy form, Armiento and Kümel (AK13) [11] pursued improved asymptotics with an enhancement factor of sln⁡ss\ln s growth; their construction, however, is independent of any exactness condition for hydrogen.

This paper takes a third route. In 1993 Gill and Pople derived the ordinary differential equation that a GGA must satisfy in order to reproduce the exact exchange potential on the hydrogen atom (GP93) [3]. We solve this equation completely as an inverse problem and use the resulting function w(z)w(z)—the enhancement factor it defines, FGP93F_{\mathrm{GP93}} (hereafter the GP93 factor)—as the single nonempirical input of the design. The consequences come in three layers.

(i) Hydrogenic exactness (Sec. II). A GGA built on the GP93 factor satisfies vx=−vHv_x=-v_{\mathrm{H}} on the hydrogen 1s density by construction (hydrogen-exactness theorem), and scale covariance then yields the exact eigenvalue ε1s=−Z2/2\varepsilon_{1s}=-Z^2/2 for every hydrogenic 1s ion (eigenvalue theorem). The GP93 factor itself contains no empirical parameters; the switching parameters introduced later are calibrated (Sec. III).

(ii) Asymptotics and Rydberg-like bound states (Sec. III). In the large-gradient regime the GP93 factor is unbounded, with the hydrogen-exact growth F=O[s(ln⁡s)2/3]F=O[s(\ln s)^{2/3}]—distinct from the AK13 sln⁡ss\ln s form—and on the hydrogen 1s density it endows the exchange potential with the exact −1/r-1/r tail by construction. An integrated functional in which this tail-carrying FmixF_{\mathrm{mix}} is switched off in many-electron regions, using the meta-GGA indicator α=(τ−τW)/τunif\alpha=(\tau-\tau^{W})/\tau^{\mathrm{unif}} of one-electron character, produces Rydberg-like series of bound virtual states.

(iii) Predicted domain of applicability (Sec. IV). The switch activates the GP93 term only where α≈0\alpha\approx0, i.e., where the density is described by a single orbital. The effect therefore reaches only systems whose highest occupied molecular orbital (HOMO) is a one-electron-like ss shell. Numerically, Rydberg-like bound states are recovered (4–8 levels) in all five such systems—all-electron H and He, and Li, Na, and K reduced to a single valence electron by large-core pseudopotentials—and not in the pp-shell atoms Ne and Ar, exactly as the design predicts. That pseudopotentials render the valence electron one-electron-like plays an essential role here, which suggests a natural affinity with real-space pseudopotential implementations.

The claims of this paper are limited to one-electron properties (Rydberg states, asymptotics, hydrogenic levels, SIE). We make no claim about thermochemistry (atomization energies): hydrogenic exactness pins the absolute energy of the H atom to its exact value, which differs from the error-compensated value that PBE was optimized to produce, so the two cannot be reconciled within a semilocal form (§ 1.7). This explicitly delimited scope follows the precedent of the LB94 [8] and mBJ [10] lineage (potential-oriented functionals that do not target thermochemistry).

Solution of the GP93 equation and hydrogenic exactness

For a spin-unpolarized density we write the GGA exchange energy as

Ex[n]=Ax∫d3r n4/3F(s),Ax=−34(3π)1/3,E_{\mathrm{x}}[n] = A_{\mathrm{x}} \int\mathrm{d}^{3}r\, n^{4/3} F(s), \qquad A_{\mathrm{x}} = -\frac{3}{4}\left(\frac{3}{\pi}\right)^{1/3},
s=∣∇n∣2(3π2)1/3n4/3(2)s = \frac{|\nabla n|}{2(3\pi^{2})^{1/3}n^{4/3}} \tag*{(2)}

with spin polarization restored by the spin-scaling relation. On the hydrogen 1s1s density the exact-exchange condition reads

Ex[n1el]=−EH[n1el],vx=δExδn=−vH.(3)E_{\mathrm{x}}[n_{\mathrm{1el}}] = -E_{\mathrm{H}}[n_{\mathrm{1el}}], \qquad v_{\mathrm{x}} = \frac{\delta E_{\mathrm{x}}}{\delta n} = -v_{\mathrm{H}}. \tag*{(3)}

For hydrogen, nH=π−1e−2rn_{\mathrm{H}} = \pi^{-1}e^{-2r}, vH(r)=r−1[1−e−2r(1+r)]v_{\mathrm{H}}(r) = r^{-1}[1-e^{-2r}(1+r)], and vxexact→−1/rv_{\mathrm{x}}^{\mathrm{exact}} \to-1/r.

In the GP93 variables x=∣∇n∣/n4/3x=|\nabla n|/n^{4/3} and z=3ln⁡[x/(2π1/3)]z=3\ln[x/(2\pi^{1/3})], with the semilocal factor g(x)=−xw(z)g(x)=-xw(z), the enhancement factor is

AxF(s)=g(x)=−xw(z),A_{\mathrm{x}}F(s)=g(x)=-xw(z),
F(s)=−xw(z)Ax=8π3s w(z)(4)F(s)=-\frac{xw(z)}{A_{\mathrm{x}}}=\frac{8\pi}{3}s\,w(z) \tag*{(4)}

with z=3ln⁡s+ln⁡6πz=3\ln s+\ln6\pi. On the hydrogen density x=2π1/3e2r/3x=2\pi^{1/3}e^{2r/3}, so that

z=2r  (≥0)(5)z=2r\;(\ge0) \tag*{(5)}

i.e., the physical domain is restricted to z≥0z\ge0. Imposing Eq. (3), ww obeys the inhomogeneous Kummer-type equation

zw′′+(2−z)w′+23w=12[2−e−z(z+2)](6)zw''+(2-z)w'+\frac{2}{3}w=\frac{1}{2}\left[2-e^{-z}(z+2)\right] \tag*{(6)}

whose homogeneous part has parameters a=−2/3a=-2/3, b=2b=2 with fundamental solutions M(a,b,z)M(a,b,z) (Kummer) and U(a,b,z)U(a,b,z) (Tricomi).

Variation of constants gives

w(z)=12−Γ(1/3)4[U(z)IM(z)+M(z)IU(z)](7)w(z)=\frac{1}{2}-\frac{\Gamma(1/3)}{4}\left[U(z)I_M(z)+M(z)I_U(z)\right] \tag*{(7)}
IM=∫0zWM dt,IU=∫z∞WU dt,I_M=\int_{0}^{z}WM\,\mathrm{d}t,\qquad I_U=\int_{z}^{\infty}WU\,\mathrm{d}t,

with W=t(t+2)e−2tW=t(t+2)e^{-2t}. As z→∞z\to\infty, IU=e−2zz8/3Φ(1/z)I_U=e^{-2z}z^{8/3}\Phi(1/z) with Φ(t)=12t−19t2+181t3+102187t4+O(t4)\Phi(t)=\frac{1}{2}t-\frac{1}{9}t^{2}+\frac{1}{81}t^{3}+\frac{10}{2187}t^{4}+O(t^{4}) (endpoint Laplace expansion); Appendix A is exponentially small and IM→CMI_M\to C_M converges, so

w(z)=12−Γ(1/3)4CMU(z)+O(e−z),U(z)∼z2/3(8)w(z)=\frac{1}{2}-\frac{\Gamma(1/3)}{4}C_MU(z)+O(e^{-z}),\qquad U(z)\sim z^{2/3} \tag*{(8)}

This z2/3z^{2/3} growth generates, through Eq. (4), F∼sln⁡sF\sim s\ln s and hence vx∼−1/rv_{\mathrm{x}}\sim-1/r (Sec. III).

Because UU has a branch cut on the negative real axis, its continuation to z<0z<0 is not unique. Equation (6), however, has a regular singular point at z=0z=0 (indices ρ=0,−1\rho=0,-1), and demanding analyticity at z=0z=0 singles out the regular branch. For the regular series w=∑ncnznw=\sum_n c_nz^n the recursion is

(m+1)(m+2)cm+1+(−m+23)cm=rm(9)(m+1)(m+2)c_{m+1}+\left(-m+\frac{2}{3}\right)c_m=r_m \tag*{(9)}

where 12[2−e−z(z+2)]=∑mrmzm\frac{1}{2}[2-e^{-z}(z+2)]=\sum_m r_mz^m. The single free constant c0=w(0)c_0=w(0) is fixed by matching to the z>0z>0 representation Eq. (7), giving

c0=w(0)=0.68928799(10)c_0=w(0)=0.68928799 \tag*{(10)}

(matching points zm∈{2,3,4,5}z_m\in\{2,3,4,5\}; stable to eight digits). The GP93 enhancement factor FF is thereby uniquely determined on the whole domain (Figure 2).

GP93 auxiliary function and enhancement factors

Figure 2. (a) The GP93 auxiliary function w(z)w(z): the Frobenius series (z<0z<0, dashed), the exact z>0z>0 representation of Eq. (7) (solid), and the asymptotic form of Eq. (8) (dotted) join uniquely at z=0z=0 (c0=0.68929c_0=0.68929). (b) Enhancement factors F(s)F(s): PBE is bounded and therefore lacks the hydrogenic tail. AK13 has an sln⁡ss\ln s-type growth, whereas the GP93 factor carried by FmixF_{\mathrm{mix}} (s0=0.16s_0=0.16) has the hydrogen-exact large-gradient form O[s(ln⁡s)2/3]O[s(\ln s)^{2/3}] of Eq. (12) (Proposition 3). Both are unbounded, but their asymptotic structures and design principles differ; FmixF_{\mathrm{mix}} is constructed from hydrogen exactness (Theorem 1) as its nonempirical input.

Theorem 1 (Hydrogen-exactness theorem). The GGA of Eq. (2) built from FF of Eqs. (4) and (7) satisfies vx(r)=−vH(r)v_{\mathrm{x}}(r)=-v_{\mathrm{H}}(r) on the hydrogen density by construction. If the range where the GP93 term acts is restricted to a finite radius by the outer switch of Sec. III, exactness is preserved within that radius.

Proof. Equation (6) is the identity derived from the requirement vx[nH]=−vHv_{\mathrm{x}}[n_{\mathrm{H}}]=-v_{\mathrm{H}}; it holds as long as FF is related to its solution ww by Eq. (4). □

Theorem 2 (Scale covariance: hydrogenic eigenvalue theorem). If FF depends only on scale-invariant descriptors (ss and, below, α\alpha and qq), then under uniform scaling nk(r)=k3n(kr)n_k(r)=k^3n(kr) one has Ex[nk]=kEx[n]E_{\mathrm{x}}[n_k]=kE_{\mathrm{x}}[n]. Consequently, for the hydrogenic densities nZ=(Z3/π)e−2Zrn_Z=(Z^3/\pi)e^{-2Zr} the exchange-only Kohn–Sham equation yields ε1s=−Z2/2\varepsilon_{1s}=-Z^2/2 exactly.

Proof. From s[nκ](r)=s[n](κr)s[n_\kappa](r)=s[n](\kappa r) it follows that ∫nκ4/3F=κ∫n4/3F\int n_\kappa^{4/3}F=\kappa\int n^{4/3}F; the statement follows from consistency with the external potential −Z/r-Z/r under r→Zrr\to Zr. Numerically the eigenvalues agree at the mHa level for Z=1Z=1–55 (Table I). □

Asymptotic −1/r-1/r and the mixed functional

The GGA exchange potential involves FF, F′F', and F′′F''. In the outer region of an exponentially decaying density n∼e−2βrn\sim e^{-2\beta r} one has s∝e2βr/3→∞s\propto e^{2\beta r/3}\to\infty, so the asymptotics are governed by the large-ss behavior of FF.

Proposition 3 (Large-gradient behavior). If FF is bounded, vxv_x vanishes exponentially in the tail of any exponentially decaying density; an unbounded large-ss enhancement is therefore necessary for any nonvanishing long-range exchange potential. AK13 [11] employs the logarithmic growth

FAK13(s)∼B1sln⁡s(s→∞).(11)F_{\mathrm{AK13}}(s)\sim B_1s\ln s\qquad(s\to\infty). \tag*{(11)}

The GP93 factor is likewise unbounded, but with a different, hydrogen-exact structure: inserting Eq. (8), w(z)=12−C~z2/3+O(z−1/3)+O(e−z)w(z)=\frac{1}{2}-\widetilde{C}z^{2/3}+O(z^{-1/3})+O(e^{-z}) with C~=Γ(1/3)CM/4\widetilde{C}=\Gamma(1/3)C_M/4, into Eq. (4) gives

FGP93(s)=4π3s−8π3C~s(3ln⁡s)2/3+⋯ ,(12)F_{\mathrm{GP93}}(s)=\frac{4\pi}{3}s-\frac{8\pi}{3}\widetilde{C}s(3\ln s)^{2/3}+\cdots, \tag*{(12)}

i.e., FGP93=O[s(ln⁡s)2/3]F_{\mathrm{GP93}}=O[s(\ln s)^{2/3}], in agreement with the O[x(ln⁡x)2/3]O[x(\ln x)^{2/3}] growth noted by Gill and Pople [3]. On the hydrogen 1s1s density this growth is not merely an asymptotic ansatz: the GP93 equation itself ensures vx=−vHv_x=-v_{\mathrm H} and hence the Coulombic −1/r-1/r tail by construction (Theorem 1).

Remark 4 (Asymptotics–stability trade-off). The unbounded growth of the enhancement factor competes with many-electron SCF stability and energy boundedness in the very same large-ss region. Bounding FF (reverting to a PBE-like bounded form) restores stability but destroys Eq. (1). The two cannot coexist within a semilocal form; this dictates the switch design of Sec. IV and the scope discussed in § 1.7.

The small-ss (GE2/PBE), z≈0z\approx0 (Frobenius series), and large-zz (divergent branch) regimes are joined by an inner switch η(z)\eta(z) and an outer switch

χ(s)=1−exp⁡[−(s/s0)p],p≥4,(13)\chi(s)=1-\exp[-(s/s_0)^p],\qquad p\ge4, \tag*{(13)}

yielding Fmix=[1−χ]FPBE+χFGP93F_{\mathrm{mix}}=[1-\chi]F_{\mathrm{PBE}}+\chi F_{\mathrm{GP93}}. The property χ=O(sp)\chi=O(s^p) (p≥4p\ge4) protects the second-order gradient expansion, while on the hydrogen density s∝e2r/3s\propto e^{2r/3} implies that 1−χ1-\chi vanishes doubly exponentially in rr, so the tail is dominated by the divergent branch. The parameters (p,s0)(p,s_0) of the outer switch are not an empirical fit: they are calibrated against the very purpose of the functional—the recovery of bound Rydberg states and the agreement of −εHOMO-\varepsilon_{\mathrm{HOMO}} with experimental ionization energies. We minimize the mean absolute error MAE =15∑∣−εHOMO−Iexp∣=\frac{1}{5}\sum|-\varepsilon_{\mathrm{HOMO}}-I_{\mathrm{exp}}| over the five systems with one-electron-like ss-shell HOMOs (H, He, Li, Na, K), within the region where all systems converge self-consistently (Figure 1). For p≳8p\gtrsim8 the results saturate and the pp dependence is negligible. Decreasing s0s_0 improves the MAE, but for s0≲0.15s_0\lesssim0.15 the SCF of the alkali valence electron diverges. The optimum over the stable domain is

Mean absolute error of the Rydberg series versus the outer-switch scale $s_0$

Figure 1. Mean absolute error of the Rydberg series versus the outer-switch scale s0s_0 (five ss-shell systems: H, He, Li, Na, K). The minimum occurs at s0=0.16s_0=0.16; for s0≲0.15s_0\lesssim0.15 (shaded) the SCF of the alkali valence electron diverges. The exponent is fixed to p=8p=8, beyond which the results saturate.

(p∗,s0∗)=(8,0.16),MAE=0.80 eV,(14)(p^\ast,s_0^\ast)=(8,0.16),\qquad\mathrm{MAE}=0.80\ \mathrm{eV}, \tag*{(14)}

which we use throughout. At this s0s_0 the exchange-only SCF density of hydrogen reproduces the exact density closely (∫r<2∣n−nH∣ dr≃10−3\int_{r<2}|n-n_{\mathrm H}|\,\mathrm{d}r\simeq10^{-3}) with ε1s=−0.4998\varepsilon_{1s}=-0.4998; the Rydberg-oriented calibration is thus compatible with hydrogenic exactness.

Many-electron switching and the integrated functional

From τ=12∑i∣∇ψi∣2\tau=\frac{1}{2}\sum_i|\nabla\psi_i|^2, τW=∣∇n∣2/8n\tau^W=|\nabla n|^2/8n, and τunif=310(3π2)2/3n5/3\tau^{\mathrm{unif}}=\frac{3}{10}(3\pi^2)^{2/3}n^{5/3} we define

α=τ−τWτunif≥0,q=∇2n4(3π2)2/3n5/3.(15)\alpha=\frac{\tau-\tau^W}{\tau^{\mathrm{unif}}}\ge0,\qquad q=\frac{\nabla^2n}{4(3\pi^2)^{2/3}n^{5/3}}. \tag*{(15)}

In any single-orbital region τ=τW\tau=\tau^W, hence α=0\alpha=0: α\alpha is an exact detector of one-electron character.

Remark 5 (Separation limit of finite descriptor sets). For any finite set of scale-invariant descriptors {s,q,…,pK}\{s,q,\ldots,p_K\} one can construct mimicking exponential densities that agree with hydrogen up to order KK as z→∞z \to\infty. Separation is complete only in the limit K→∞K \to\infty; in particular, with (s,q)(s,q) alone parts of heavy-atom cores are degenerate with hydrogen in the high-density, small-ss regime, and the n4/3n^{4/3} weight then defeats the separation of exchange energies.

The variables α\alpha and qq carry complementary local information: α\alpha distinguishes the number of orbitals (single vs. many) but is blind to the shape of a single orbital (Be-type: non-hydrogenic valence with α≡0\alpha\equiv0), whereas qq captures the curvature of the density but is insensitive to orbital count. The pair (s,α,q)(s,\alpha,q) therefore provides a more discriminating local description than (s,α)(s,\alpha) alone; in the SCAN [13] variable space (s,α)(s,\alpha) the Be-type case cannot be separated.

The integrated functional is

F~(s,α)=FPBE(s)+G(α)[Fmix(s)−FPBE(s)],(16)\widetilde{F}(s,\alpha)=F_{\mathrm{PBE}}(s)+G(\alpha)\left[F_{\mathrm{mix}}(s)-F_{\mathrm{PBE}}(s)\right], \tag*{(16)}

with G=exp⁡[−(α/wα)2]G=\exp[-(\alpha/w_\alpha)^2] and wα=0.10w_\alpha=0.10. Since G(0)=1G(0)=1 and G′(0)=0G'(0)=0, the switch does not perturb vxv_{\mathrm{x}} in one-electron regions. Two constructive properties follow: (i) one-electron limit—for α≡0\alpha\equiv0 one has F~=Fmix\widetilde{F}=F_{\mathrm{mix}}, so the GP93 term and its asymptotics are fully active (inheriting Theorems 1 and 3); (ii) suppression in many-electron regions—where α>0\alpha>0, F~→FPBE\widetilde{F}\to F_{\mathrm{PBE}}, preventing the divergence of FmixF_{\mathrm{mix}} from destroying many-electron energies (indeed, the SCF of the HF molecule, which diverges with bare FmixF_{\mathrm{mix}}, converges with the integrated form).

A prediction of the domain of applicability follows immediately: the asymptotic improvement reaches the HOMO only if the HOMO itself is one-electron-like (α≈0\alpha\approx0), i.e., only in systems whose outermost shell is occupied by a single nsns electron. For pp-shell HOMOs (α>0\alpha>0) the switch suppresses the GP93 term and no effect appears. This prediction is verified numerically in Sec. VI.

q-resolved switching toward molecules and solids

The α\alpha switch of Eq. (16) distinguishes the type of atomic HOMO (ss vs. pp shell) but can be insufficient in molecules and solids. At the center of a covalent bond, the overlap of two atomic orbitals produces a low-density, large-ss region with α≈0\alpha\approx0, where the divergent branch of FmixF_{\mathrm{mix}} causes numerical instability. This region cannot be distinguished from the atomic tail (the source of the Rydberg states, where the GP93 term should act) by α\alpha alone.

The discriminating key is the reduced Laplacian of Eq. (15). That qq carries information independent of α\alpha can be verified explicitly on an exponentially decaying atomic tail. For n(r)∼e−2κrn(r)\sim e^{-2\kappa r} the reduced gradient s=∣∇n∣/[2(3π2)1/3n4/3]s=|\nabla n|/[2(3\pi^2)^{1/3}n^{4/3}] grows with rr, and ∇2n=n′′+2rn′=(4κ2−4κr)n>0\nabla^2n=n''+\frac{2}{r}n'=(4\kappa^2-\frac{4\kappa}{r})n>0, so

q(r)⟶r→∞κ2(3π2)2/3n2/3⟶+∞,(17)q(r)\underset{r\to\infty}{\longrightarrow}\frac{\kappa^2}{(3\pi^2)^{2/3}n^{2/3}}\longrightarrow+\infty, \tag*{(17)}

i.e., qq diverges to large positive values (the density is convex). At a covalent bond center, by contrast, the overlapping densities form a saddle with ∇2n≤0\nabla^2n\leq0, hence q≲0q\lesssim0. This sign difference arises even when α=0\alpha=0 in both regions, because α\alpha measures the one-electron character of the kinetic energy while qq measures the curvature of the density—independent pieces of information. Numerically, atomic tails show q∼10q\sim10–10210^2 while bond centers show q∼−0.5q\sim-0.5–−1-1 (See Sec. VI, Table V). Exploiting this independence we define the triple-switched form

MoleculeSwitchqqα\alphaClosing mechanism
H2+_2^+0.008−0.45-0.450.00qq (covalent)
H2_20.012−0.31-0.310.00qq (covalent)
BH0.005−0.36-0.360.09qq (covalent)
CO0.008−0.31-0.310.07qq (covalent)
N2_20.000−0.08-0.080.55q+αq+\alpha
LiH0.000+2.10+2.106.06α\alpha (ionic)
HF0.000+0.19+0.190.62α\alpha (polar)
LiF0.000+5.50+5.5013.1α\alpha (ionic)
HeH+^+0.064+0.20+0.200.00qq
He2+_2^+0.114+0.63+0.630.08borderline

Table V. Triple-switch values at molecular bond centers (non-self-consistent, near equilibrium). Values close to 0 mean the GP93 term is suppressed. Covalent bonds are switched off via qq; polar and ionic bonds via α\alpha.

F~(s,α,q)=FPBE+D(q)G(α)[Fmix−FPBE],(18)\widetilde{F}(s,\alpha,q)=F_{\mathrm{PBE}}+D(q)G(\alpha)\left[F_{\mathrm{mix}}-F_{\mathrm{PBE}}\right], \tag*{(18)}

with D(q)=12[1+tanh⁡((q−qc)/wq)]D(q)=\frac{1}{2}[1+\tanh((q-q_c)/w_q)], qc=1q_c=1, and wq=0.6w_q=0.6. The switch DD has the two limits

D(q)⟶q≫qc1,D(q)⟶q≪qc0.(19)D(q)\underset{q\gg q_c}{\longrightarrow}1,\qquad D(q)\underset{q\ll q_c}{\longrightarrow}0. \tag*{(19)}

By Eq. (17), atomic tails have q→+∞q\to+\infty and thus D→1D\to1: the α\alpha-switched form of Eq. (16) is recovered and the GP93 term remains active. At covalent bond centers q≲0<qcq\lesssim0<q_c gives D→0D\to0, hence F~→FPBE\widetilde{F}\to F_{\mathrm{PBE}} and the divergent branch of FmixF_{\mathrm{mix}} is removed. At polar and ionic bond centers qq may be positive, but α→0\alpha\to0 already suppresses the GP93 term through G(α)→0G(\alpha)\to0. The product D(q)G(α)D(q)G(\alpha) closes whenever either factor vanishes,

D(q)G(α)→0ifq≲qc or α>0,(20)D(q)G(\alpha)\to0\quad\text{if}\quad q\lesssim q_c\ \text{or}\ \alpha>0, \tag*{(20)}

suppressing the divergence in the representative bonding environments tested here (fixed-density tests for ten molecules in Sec. VI).

Importantly, the qq switch does not damage the atomic −1/r-1/r tail at all. This follows from the separation of roles between qq and ss: an atomic tail has s→∞s\to\infty and q→+∞q\to+\infty, and by Eq. (19) the switch is saturated there, ∂D/∂qc→0\partial D/\partial q_c\to0, so that varying qcq_c leaves the atomic −1/r-1/r tail unchanged; only near bond centers (q∼qcq\sim q_c) does ∂D/∂qc≠0\partial D/\partial q_c\ne0, where suppression is tuned. That is,

∂(−rvx)∂qc∣tail≈0,∂[DG]∂qc∣bond≠0,(21)\left.\frac{\partial(-r v_{\mathrm{x}})}{\partial q_c}\right|_{\mathrm{tail}}\approx0,\qquad\left.\frac{\partial[D G]}{\partial q_c}\right|_{\mathrm{bond}}\ne0, \tag*{(21)}

i.e., in the cases tested the atomic tail is insensitive to qcq_c while the bond-center suppression is tuned by it (Sec. VI). This is in sharp contrast to an ss-dependent switch, which would also suppress the large-ss atomic tail and destroy the Rydberg physics; the sign structure of Eq. (17) makes qq essential.

Including qq (the Laplacian) in the energy introduces the functional derivative vlapl=∂ex/∂(∇2n)v_{\mathrm{lapl}}=\partial e_{\mathrm{x}}/\partial(\nabla^2n), which—as for SCAN-type meta-GGAs—is numerically delicate in a self-consistent potential and unstable in Gaussian-basis implementations. In this paper we therefore demonstrate the discriminating power of the triple switch non-self-consistently (on fixed densities) and leave the self-consistent implementation to real-space methods, where vlaplv_{\mathrm{lapl}} can be handled stably on a grid (§ 1.7).

In all-electron calculations the cores of heavy atoms pass through the enhancement region of the GP93 factor and over-enhance the exchange (the degeneracy noted in Remark 5); for Be-type systems the SCF diverges. Large-core pseudopotentials remove the core, thereby (i) eliminating the source of over-enhancement and (ii) rendering the alkali valence literally one-electron (ns1ns^1), which makes α≈0\alpha\approx0 exact. Large-core pseudopotentials thus play a constructive role in the present framework by isolating a one-electron-like valence density.

The exchange is combined with an independent correlation functional (as for AK13).

Proposition 6 (LYP compatibility). LYP correlation [6] vanishes for any one-electron density (Ec=0E_{\mathrm{c}} = 0) and therefore preserves the hydrogenic exactness of the present exchange (Theorems 1 and 2). Numerically, the H-atom eigenvalue with the integrated exchange plus LYP is ϵ1s=−0.4994\epsilon_{1s} = -0.4994, essentially identical to the exchange-only value −0.4999-0.4999, whereas PBE correlation shifts it to −0.5085-0.5085 (Table VI).

+LYP+PBEcReference
ReR_{\mathrm{e}} [Å]0.700.70FCI/expt. 0.74
Well depth [eV]3.63.6FCI 4.7
H atom ε1s\varepsilon_{1s}−0.4994-0.4994−0.5085-0.5085x-only −0.4999-0.4999

Table VI. Correlation functionals compared in the H2_2 binding region (Proposition 6). LYP vanishes for one-electron systems and keeps the H-atom eigenvalue at the exchange-only value (−0.4999-0.4999), whereas PBE correlation injects finite correlation. Bond length and well depth are equivalent; the well being shallower than FCI is the generic limitation of an exchange-focused semilocal functional (thermochemistry is not claimed).

Computational methods

All calculations use the custom-functional interface (define_xc_) of PySCF [17] (version 2.13), with correlation taken from libxc [18] (version 7.0; PBE or LYP). The exchange is implemented with analytic channel derivatives vρ,vσ,vτv_\rho, v_\sigma, v_\tau, with FF and F′F' represented by cubic splines in ln⁡s\ln s over s∈[10−5,104]s \in[10^{-5}, 10^4] and clipped outside this range; densities are floored at 10−25010^{-250} in the evaluation of reduced variables. Default PySCF quadrature grids of level 5–7 are used, with SCF convergence thresholds of 10−810^{-8}–10−910^{-9} Ha; where convergence is difficult, damping (0.2–0.3) and level shifts (0.2 Ha) are applied in a first stage and removed for the final iterations. Spin-restricted (RKS) and spin-unpolarized densities are treated through the spin-scaling relation; open-shell systems use unrestricted (UKS) calculations. Rydberg-like bound states are identified in exchange-only calculations, where the presence or absence of the potential tail is reflected directly, by counting negative Kohn–Sham virtual eigenvalues (ϵvirt<0\epsilon_{\mathrm{virt}} < 0). Diffuse basis sets [24, 25] are essential to represent the tail (aug-cc-pV5Z for H and He; aug-cc-pVQZ/TZ for Ne and Ar). Because a finite Gaussian basis provides only a discretized representation of the near-continuum spectrum, the precise count of such states can depend on the diffuseness of the basis; the qualitative contrast reported below (zero bound states for the reference functionals versus a series of bound states for the present one, stable over the calibration range s0=0.13s_0 = 0.13–0.28) is the basis-robust content of our claim. Alkali atoms are reduced to a single valence electron by large-core pseudopotentials (CRENBL for Li [26]; SBKJC for Na and K [27, 28]). Experimental ionization energies are taken from the NIST Atomic Spectra Database [30]. The α\alpha switch converges with standard DIIS because vτv_\tau enters as a Hermitian kinetic-energy-like operator; where convergence is difficult, staged damping and level shifting are used.

The radial exchange potential −r vx(r)-r\,v_x(r) of Figure 6 is evaluated on densities solved self-consistently with correlation for each functional, taking the exchange part only,

Radial exchange potential for H, He, and Na

Figure 6. Radial exchange potential −rvx(r)-r v_{\mathrm{x}}(r) evaluated on self-consistent densities obtained with correlation, taking the exchange part only (finite-difference construction). In each panel the red curve is the exact exchange reference computed from the same density: for H (exact density) vx=−vHv_{\mathrm{x}}=-v_{\mathrm{H}} (Theorem 1); for He, a closed-shell two-electron system, the exact relation vx=−vH/2v_{\mathrm{x}}=-v_{\mathrm{H}}/2 holds; for Na, a one-electron valence system, vx=−vHv_{\mathrm{x}}=-v_{\mathrm{H}} holds, with rvH[n]r v_{\mathrm{H}}[n] (halved for He) constructed by radial integration. For H, FmixF_{\mathrm{mix}} coincides with the exact curve. For He, FmixF_{\mathrm{mix}} maintains −rvx≃1-r v_{\mathrm{x}}\simeq1 while PBE and BLYP [5, 6] collapse to 0.3–0.5. For Na, FmixF_{\mathrm{mix}} approaches the exact reference outside the valence-density maximum, reaching 0.98 at r=5.5r=5.5 (exact 0.97). For He and Na the curves are restricted to the range where the Gaussian basis is reliable. Only FmixF_{\mathrm{mix}} recovers the −1/r-1/r tail, which is the origin of the Rydberg-like bound states.

vx(r)=vρ−[f′(r)+2rf(r)],f(r)=2vσ ∂rn,(22)v_x(\mathbf{r}) = v_\rho- \left[f'(r) + \frac{2}{r}f(r)\right], \qquad f(r) = 2v_\sigma\,\partial_r n, \tag*{(22)}

with the divergence term constructed by finite differences (the same procedure as in a real-space code). On the hydrogen density Theorem 1 gives vx=−vHv_x = -v_{\mathrm{H}} exactly, so −r vx=1−e−2r(1+r)→1-r\,v_x = 1 - e^{-2r}(1+r) \to1 is fixed analytically. When the self-consistent Fmin⁡F_{\min} density is expanded in Gaussians, however, the outer region deviates from the exact density and −r vx-r\,v_x departs from the exact curve. This is a mathematical property of the basis: hydrogen is a one-electron system, so Hartree–Fock equals exact exchange and the exact density n=e−2r/πn = e^{-2r}/\pi is known; Gaussian HF densities reproduce it to within 0.3% relative error in the intermediate region (r≲3.5r \lesssim3.5) at the aug-cc-pV5Z level, but in the far region (r≳4r \gtrsim4, n≲10−5n \lesssim10^{-5}) a finite Gaussian set cannot represent the exponential tail e−2re^{-2r} (even though the total energy is converged to −0.499995-0.499995 Ha at V5Z). Figure 6 is therefore restricted to the inner regions where the basis is reliable. A quantitative evaluation of the potential tail for all systems, and the self-consistent implementation of the qq-switched form (which requires a stable treatment of vlaplv_{\mathrm{lapl}}), are better suited to real-space pseudopotential methods; in this paper the primary evidence is the bound Rydberg series, which is the physical consequence of the asymptotic tail.

Results

Hydrogenic exactness

Table I lists the exchange-only eigenvalues of the hydrogenic series. As required by Theorem 2, ϵ1s=−Z2/2\epsilon_{1s} = -Z^2/2 is reproduced at the mHa level; no empirical parameter enters the GP93 factor itself.

Rydberg-like bound states (main result)

For the He atom (Figure 3), PBE, SCAN, TPSS, and M06-L all fail to bind any Rydberg state (−ϵHOMO≈16-\epsilon_{\mathrm{HOMO}} \approx16 eV, more than 8 eV below the experimental 24.59).

System−ϵ1s-\epsilon_{1s} (calc.)Z2/2Z^2/2
H0.499720.5
He+^+1.999892
Li2+^{2+}4.501444.5

Table 1. Hydrogenic series (exchange only): numerical demonstration of Theorem 2. Energies in hartree.

while the integrated FmixF_{\mathrm{mix}} alone yields eight bound orbitals and −εHOMO=25.1-\varepsilon_{\mathrm{HOMO}}=25.1 eV. The independent bound levels are two, at −0.147-0.147 and −0.038-0.038 Ha; their effective quantum numbers neff=1/2∣ε∣=1.84n_{\mathrm{eff}}=1/\sqrt{2|\varepsilon|}=1.84 and 3.633.63 place them close to a hydrogenic progression. The count of eight orbitals includes the angular-momentum partners that accompany these two levels in the diffuse basis; shallower high-nn levels are not bound because of the residual of the potential tail. Even so, the existence of bound levels is a qualitative difference from the semilocal and hybrid functionals, which bind none (Table II; Figure 3B). The radial −rvx-rv_x [Figure 3(C)] visualizes the origin directly at the level of the potential: FmixF_{\mathrm{mix}} retains a finite tail while PBE collapses to 4×10−34\times10^{-3} by r=9r=9 bohr.

The He/Ne contrast in Table II shows that the effect is not specific to He but follows from the design principle (one-electron-like ss-shell HOMO). Li illustrates a different regime: with a single valence electron and small ZeffZ_{\mathrm{eff}}, other functionals also bind Rydberg-like states (PBE 3, SCAN 7), but FmixF_{\mathrm{mix}} is closest to the experimental ionization energy (4.35 versus 5.39 eV) with among the largest bound counts—the advantage shifts from “the only functional that binds” (He) to “the best description” (Li).

The atomic systematics (Figure 4, Table III) constitute the central result of this paper. In all five systems with one-electron-like ss-shell HOMOs—H (1s), He (1s), and Li (2s), Na (3s), K (4s) reduced to one valence electron by large-core pseudopotentials—Rydberg-like series of bound states (4–8 levels) are recovered and −εHOMO-\varepsilon_{\mathrm{HOMO}} is consistently closer to the experimental IP than PBE. In the pp-shell atoms Ne (2p) and Ar (3p) the HOMO region has α>0\alpha>0, the switch suppresses the GP93 term, and the bound count remains zero. This “ss-shell-only” selectivity is consistent with the intended design domain: the α\alpha switch selects one-electron-like HOMOs (Sec. IV), so that the domain of applicability follows from the design itself.

Atomic systematics bar chart

Figure 4. Atomic systematics, all-functional comparison (number above each bar = bound Rydberg orbitals). −εHOMO-\varepsilon_{\mathrm{HOMO}} of PBE [7], SCAN [13], B3LYP [14], and FmixF_{\mathrm{mix}} (dark blue), with experimental ionization potentials (red diamonds). In all five systems (H, He, Li, Na, K; one-electron-like ss-shell HOMOs) FmixF_{\mathrm{mix}} is closest to experiment. The mean absolute HOMO–IP errors are PBE 4.19, SCAN 3.95, B3LYP 3.30, and FmixF_{\mathrm{mix}} 0.79 eV, four to five times better than the other functionals. For H and He, FmixF_{\mathrm{mix}} also stands out in the number of bound levels (4–8 versus 0–2). Li uses CRENBL, Na and K use SBKJC (one valence electron).

Many-electron stability and H2

The α\alpha switch of Eq. (16) stabilizes molecular SCF calculations—which diverge with bare FmixF_{\mathrm{mix}} in the low-density, large-ss bonding region—wherever the density localizes near the nuclei. In the binding region of H2_2 (Figure 5), combined with LYP correlation, the equilibrium bond length is Re≃0.70R_e\simeq0.70 Å (FCI/experiment 0.74), while the well depth of 3.6 eV is shallower than the FCI value of 4.7 eV—the generic limitation of an exchange-focused semilocal functional (thermochemistry is outside the claims of this paper). On the dissociation side (R≳1.5R\gtrsim1.5 Å) the broken-symmetry UKS diverges in the low-density internuclear region and cannot be converged: this is precisely the region where the divergent branch of FmixF_{\mathrm{mix}} is not fully suppressed by the α\alpha switch alone, motivating the qq switch of Sec. IV A and the real-space implementation.

Binding-region potential curve of H₂ with the integrated functional

Figure 5. Binding-region potential curve of H₂ with the integrated functional (LYP correlation; broken-symmetry UKS). The bond length Re≃0.70R_e \simeq0.70 Å is close to FCI and experiment (0.74 Å); the well depth of 3.6 eV is shallower than the FCI value of 4.7 eV. Shown is the binding region in which stable self-consistent solutions exist.

Systematic tests of the triple switch (q resolution)

That the qq switch of Eq. (18) suppresses the divergence where the α\alpha switch is insufficient in molecules is verified systematically in non-self-consistent evaluations (on fixed PBE or HF densities). A self-consistent qq switch is unstable in Gaussian bases owing to the νlapl\nu_{\mathrm{lapl}} problem (§ 1.7), but the discriminating power itself is unambiguous on fixed densities. The quantity monitored is whether the triple switch D(q)G(α)D(q)G(\alpha) opens (→1\to1) or closes (→0\to0) in each region.

For atoms, Table IV lists α\alpha near the HOMO of the second-period elements together with whether the exchange potential retains its −1/r-1/r tail (GP93 term active). The ss-shell HOMOs of H and He have α=0\alpha=0; with qq large and positive in the tail (1010–10210^2), the GP93 term acts and −rvx≃0.9-r v_x \simeq0.9 is maintained. The pp-shell HOMOs from B to Ne have α=0.7\alpha=0.7–4040, so G(α)≈0G(\alpha)\approx0 suppresses the GP93 term completely and vxv_x reverts to PBE. This confirms the prediction of Sec. IV (the GP93 term acts only for ss shells) across the entire second period.

For molecules, Table V lists the triple-switch value at the bond center near equilibrium for ten molecules representing diverse bonding types. In covalent bonds (H2_2, BH, CO, N2_2) the bond-center qq is small or negative (−0.5-0.5–00) and D(q)D(q) closes; in polar and ionic bonds (LiH, HF, LiF) α>0\alpha>0 and G(α)G(\alpha) closes. In total, the bond-center switch is below 0.1 for nine of the ten molecules, suppressing the divergent branch of FmixF_{\mathrm{mix}}. The only borderline case is He2+_2^+ (both qq and α\alpha intermediate; switch 0.11). The two switches thus act complementarily and suppress the divergence in the representative bonding environments tested here.

That these two mechanisms leave the atomic −1/r-1/r tail intact is verified by scanning qcq_c. Moving qcq_c from 0 to 1 leaves the He tail fixed at −rvx=0.90-r v_x=0.90, while the H2+_2^+ bond-center switch decreases monotonically from 0.33 to 0.017. Because the atomic tail (q∼30q\sim30) and the bond center (q∼−0.2q\sim-0.2) are completely separated in qq, one can be tuned without sacrificing the other. This is fundamentally different from an ss-dependent switch, which would also suppress the large-ss atomic tail and destroy the Rydberg states, and it establishes that discrimination by the reduced Laplacian qq is essential for the extension to molecules and solids.

FunctionalTypeHe (1s1s) −εH-\varepsilon_{\mathrm{H}}He (1s1s) boundNe (2p2p) −εH-\varepsilon_{\mathrm{H}}Ne (2p2p) boundLiECP^{\mathrm{ECP}} (2s2s) −εH-\varepsilon_{\mathrm{H}}LiECP^{\mathrm{ECP}} (2s2s) bound
PBEGGA15.76013.3503.213
BLYPGGA15.92013.3703.004
SCANmeta-GGA16.54014.0003.247
B3LYPhybrid (20%)18.00015.6503.614
PBE0hybrid (25%)18.22016.0003.803
SCAN0hybrid (25%)18.77016.480——
FmixF_{\mathrm{mix}}integrated meta-GGA25.12811.8504.357
Experiment—24.59—21.56—5.39—

Table 2. Rydberg-like bound states, all-functional comparison. Each column pair gives −εH≡−εHOMO-\varepsilon_{\mathrm{H}} \equiv-\varepsilon_{\mathrm{HOMO}} [eV] and the number of bound virtual orbitals (including angular-momentum partners). Experimental ionization energies [30] in the last row. See the text for discussion.

Effect of the correlation functional

Table VI shows the correlation dependence. Because LYP vanishes for one-electron systems, the H-atom eigenvalue ε1s=−0.4994\varepsilon_{1s} = -0.4994 essentially coincides with the exchange-only value (−0.4999-0.4999), preserving hydrogenic exactness (Proposition 6); PBE correlation injects a finite correlation into the one-electron system and shifts the eigenvalue to −0.5085-0.5085. Bond length and well depth are essentially identical for the two correlations, so LYP, which preserves hydrogenic exactness, is the recommended companion of the present exchange.

Orbital-independent triple switch

The indicator α=(τ−τW)/τunif\alpha= (\tau-\tau^{W})/\tau^{\mathrm{unif}} in the triple switch of Eq. (18) depends on the Kohn–Sham orbitals through the kinetic-energy density τ\tau and therefore requires a generalized Kohn–Sham (gKS) solver. In this section we show that τ\tau can be replaced by an orbital-free kinetic-energy-density functional (KEDF), turning α\alpha into a semilocal density functional α(s,q)\alpha(s,q). The entire triple switch then becomes a pure density functional: its exchange potential is an ordinary local (multiplicative) potential common to all orbitals, so that no gKS solver is required in any Kohn–Sham code. Its only remaining nonstandard ingredient is the Laplacian ∇2n\nabla^{2}n, whose potential term vlaplv_{\mathrm{lapl}} is unstable in Gaussian bases but is evaluated naturally on real-space grids (§ 1.7); real-space pseudopotential codes are therefore the natural target.

Writing the kinetic-energy density through an enhancement factor, τ=τunifFs\tau=\tau_{\mathrm{unif}}F_s with τunif=310(3π2)2/3n5/3\tau_{\mathrm{unif}}=\frac{3}{10}(3\pi^2)^{2/3}n^{5/3}, τW=τunifFsW\tau^W=\tau_{\mathrm{unif}}F_s^W, and FsW=53pF_s^W=\frac{5}{3}p (p=s2p=s^2), we have

α=τ−τWτunif=Fs−FsW.(23)\alpha=\frac{\tau-\tau^W}{\tau_{\mathrm{unif}}}=F_s-F_s^W . \tag*{(23)}

Making α\alpha orbital-free thus reduces to approximating the true (orbital) enhancement factor FsF_s by a semilocal density functional Fs(s,q)F_s(s,q). The exact iso-orbital condition α=0\alpha=0 corresponds to Fs=FsWF_s=F_s^W, i.e., to saturation of the lower bound τ=τW\tau=\tau^W.

AtomHOMOExpt. IPPBESCANB3LYPFmixF_{\mathrm{mix}}
H1s13.607.59/08.02/28.77/113.59/4
He1s24.5915.76/016.54/018.00/025.12/8
LiECP^{\mathrm{ECP}}2s5.393.21/33.24/73.61/44.35/7
NaECP^{\mathrm{ECP}}3s5.143.02/33.03/73.38/43.99/6
KECP^{\mathrm{ECP}}4s4.342.52/32.50/102.80/43.12/6

Table 3. Atomic systematics of bound Rydberg states, all-functional comparison. Each cell gives −εHOMO-\varepsilon_{\mathrm{HOMO}} [eV] / number of bound Rydberg orbitals. In all five systems FmixF_{\mathrm{mix}} brings −εHOMO-\varepsilon_{\mathrm{HOMO}} closest to the experimental IP (compared not only with PBE but also with SCAN and B3LYP). For H and He, FmixF_{\mathrm{mix}} also stands out in bound counts (4–8 versus 0–2). For the valence-only alkalis (Li–K) other functionals also bind states, but the −εHOMO-\varepsilon_{\mathrm{HOMO}} of FmixF_{\mathrm{mix}} is consistently the best. Li: CRENBL; Na, K: SBKJC (one valence electron each). s0=0.16s_0=0.16 (Rydberg-accuracy calibration); the MAE of HOMO–IP over the five ss-shell systems is 0.79 eV. For the pp-shell references (Ne, Ar) see the text and Table II.

Naive choices fail. The second-order gradient expansion, and Pauli–Gaussian-type KEDFs [16] (FsP=e−μp+βq2F_s^P=e^{-\mu p}+\beta q^2 with FsP≡Fs−FsWF_s^P\equiv F_s-F_s^W), are designed for the accuracy of the integrated kinetic energy Ts=∫τT_s=\int\tau and do not guarantee the pointwise bound Fs≥FsWF_s\ge F_s^W. Indeed, the βq2\beta q^2 term diverges in the atomic tail, where q→+∞q\to+\infty [Eq. (17)], so α=e−μp+βq2→∞≠0\alpha=e^{-\mu p}+\beta q^2\to\infty\ne0, violating one-electron exactness.

The Laplacian-level KEDF of Perdew and Constantin (PC07) [15] is essential here precisely because it imposes the pointwise bound τW≤τ\tau^W\le\tau (Fs≥FsWF_s\ge F_s^W) constructively. The starting point is the fourth-order gradient expansion

FsGE4=1+527p+209q+Δ,Δ=881q2−19pq+8243p2,(24)F_s^{\mathrm{GE4}}=1+\frac{5}{27}p+\frac{20}{9}q+\Delta,\qquad\Delta=\frac{8}{81}q^2-\frac{1}{9}pq+\frac{8}{243}p^2, \tag*{(24)}

which, however, diverges as Δ∼881q2\Delta\sim\frac{8}{81}q^2 for q→∞q\to\infty. Its bounded modification,

FsGE4M=FsGE41+[Δ/(1+53p)]2⟶∣q∣→∞1+53p=1+FsW,(25)F_s^{\mathrm{GE4M}}=\frac{F_s^{\mathrm{GE4}}}{\sqrt{1+\left[\Delta/(1+\frac{5}{3}p)\right]^2}}\underset{|q|\to\infty}{\longrightarrow}1+\frac{5}{3}p=1+F_s^W, \tag*{(25)}

saturates onto FsWF_s^W as ∣q∣→∞|q|\to\infty (the denominator grows like ∣Δ∣|\Delta|). Finally, FsPCF_s^{\mathrm{PC}} is smoothly interpolated to the lower bound FsWF_s^W,

FsPC=FsW+(FsGE4M−FsW)fab(FsGE4M−FsW),(26)F_s^{\mathrm{PC}}=F_s^W+\left(F_s^{\mathrm{GE4M}}-F_s^W\right)f_{ab}\left(F_s^{\mathrm{GE4M}}-F_s^W\right), \tag*{(26)}
fab(z)={[1+ea/(a−z)ea/z+ea/(a−z)]b,0<z<a,1,z≥a,(27)f_{ab}(z)=\begin{cases} \left[\dfrac{1+e^{a/(a-z)}}{e^{a/z}+e^{a/(a-z)}}\right]^b, & 0<z<a,\\ 1, & z\ge a, \end{cases} \tag*{(27)}

with a=0.5389a=0.5389 and b=3b=3. Because fab(0)=0f_{ab}(0)=0, one has FsPC=FsWF_s^{\mathrm{PC}}=F_s^W wherever FsGE4M=FsWF_s^{\mathrm{GE4M}}=F_s^W (one-electron regions), so that by Eq. (23) αPC=(FsGE4M−FsW)fab(⋅)≥0\alpha^{\mathrm{PC}}=(F_s^{\mathrm{GE4M}}-F_s^W)f_{ab}(\cdot)\ge0: the lower bound is guaranteed. This construction reproduces the orbital indicator at covalent bond centers (α≈0\alpha\approx0), in pp-shell atoms (α>0\alpha>0), and in the inner one-electron region (α=0\alpha=0) (Table VII).

RegionαKS\alpha^{\mathrm{KS}}αPC(s,q)\alpha^{\mathrm{PC}}(s,q)
H tail (r≥2r \ge2)0.000.00
He tail (r≥3r \ge3)0.000.00
Ne (2p shell)0.44–0.730.68–0.76
H2_2 bond center0.000.00
N2_2 bond center0.550.545
CO bond center0.070.003
LiH bond center6.065.58

Table VII. Orbital-independent αPC(s,q)\alpha^{\mathrm{PC}}(s,q) versus the orbital indicator αKS\alpha^{\mathrm{KS}} (non-self-consistent, on PBE/HF densities). Covalent bond centers, the pp shell, and one-electron tails are reproduced.

AtomHOMOα\alpha (HOMO region)−1/r-1/r tail
H1s0.00yes (−rvx≈0.94-r v_x\approx0.94)
He1s0.00yes (−rvx≈0.88-r v_x\approx0.88)
B2p1.08no (reverts to PBE)
C2p1.04no
N2p1.41no
O2p0.59no
F2p1.58no
Ne2p5.56no

Table 4. Triple-switch behavior near atomic HOMOs (non-self-consistent). α\alpha is the value in the HOMO region; the last column indicates whether vxv_x keeps the −1/r-1/r tail: yes for ss shells (α=0\alpha=0), no for pp shells (α>0\alpha>0).

In the outermost atomic tail, however, the bounded form Eq. (25) does not return to the von Weizsäcker limit. For an exponentially decaying density both pp and qq grow without bound, with q/p=1−1/(κr)→1q/p=1-1/(\kappa r)\to1; in this regime Δ≃5243p2>0\Delta\simeq\frac{5}{243}p^{2}>0 dominates, so that FsGE4M→1+FsWF_{\mathrm{s}}^{\mathrm{GE4M}} \to1+F_{\mathrm{s}}^{\mathrm{W}} rather than FsWF_{\mathrm{s}}^{\mathrm{W}}, and consequently αPC→1\alpha^{\mathrm{PC}}\to1 instead of the exact value 0. At the finite radii where Gaussian densities are meaningful (κr≃3\kappa r\simeq3–6, q/p≃0.7q/p\simeq0.7–0.85) the same saturation already produces a positive residual. As a pragmatic regularization of the density-only approximation—not a parameter-free consequence of the GP93 construction—we impose a large-pp damping factor,

αPC(s,q)=(FsGE4M−FsW)fab(⋅)e−(p/pc)2,pc=5,(28)\alpha^{\mathrm{PC}}(s,q)=\left(F_{\mathrm{s}}^{\mathrm{GE4M}}-F_{\mathrm{s}}^{\mathrm{W}}\right)f_{\mathrm{ab}}(\cdot)e^{-(p/p_{\mathrm{c}})^2},\qquad p_{\mathrm{c}}=5, \tag*{(28)}

forcing αPC→0\alpha^{\mathrm{PC}}\to0 as p→∞p\to\infty (outermost tail), which restores α=0\alpha=0 (GP93 term active) in the asymptotic region that determines the Rydberg levels.

The integrated functional built on the orbital-independent αPC\alpha^{\mathrm{PC}} yields, in the outer tail of He (r≥3r\geq3, the region that determines the Rydberg levels), exactly the same exchange-potential tail as the orbital version (−rvx≃0.90-r v_{\mathrm{x}}\simeq0.90): the Rydberg asymptotics are recovered without orbitals. In the intermediate region (r≃1.5r\simeq1.5–2, p≃2p\simeq2–8), by contrast, the second-derivative structure of the density (p,q,q/p)(p,q,q/p) of one-electron-like He happens to be close to that of pp-shell Ne; local quantities alone cannot fully decide the one-electron character there and a residual of about 0.3 remains in αPC\alpha^{\mathrm{PC}}. This residual mildly perturbs vxv_{\mathrm{x}} near r≃2r\simeq2 but recovers for r≳3r\gtrsim3, so the impact on HOMO and Rydberg levels—determined in the outer region—is limited. This is an intrinsic limitation of local discrimination by the density; refinement with higher-order descriptors is left for future work.

The methodological consequence is significant: the one-electron detection at the core of meta-GGA function-

ality can be approximated as a density functional by an orbital-free KEDF possessing the τW≤τ\tau_W \le\tau bound structure, so that the entire triple switch requires no gKS solver—only a standard local exchange potential; combined with the grid-friendly treatment of ∇2n\nabla^2 n, this makes self-consistent implementation in real-space Kohn–Sham codes feasible. This opens the path from the non-self-consistent demonstrations of Sec. VI D to full self-consistency.

Discussion

LB94 and mBJ construct potentials directly and lack variational consistency. AK13 realized Eq. (11) within the GGA energy form, but its construction is independent of hydrogenic exactness. The present framework differs in that a single nonempirical input—the GP93 hydrogen-exactness condition—produces both hydrogenic exactness (Theorems 1 and 2) and the unbounded large-gradient growth (Proposition 3) from the same function w(z)w(z). The contrast with SCAN is direct: SCAN refines the intermediate-range description with α\alpha but its potential tail remains exponential [Fig. 3(A)]; the present framework uses the same variable α\alpha as the criterion for suppressing the GP93 term, thereby recovering bound Rydberg states at comparable computational cost.

Two independent checks show that the bound Rydberg states are not an artifact of the calibration. First, predictivity on one-electron cations not used in the calibration: for He+^+ (Z=2Z=2) and Li2+^{2+} (Z=3Z=3) the integrated FmixF_{\mathrm{mix}} gives −ϵ1s=54.44-\epsilon_{1s}=54.44 and 122.50122.50 eV, in essentially exact agreement with Z2/2Z^2/2 (54.42 and 122.45 eV), whereas PBE gives 42.0 and 103.7 eV. This is simultaneously a numerical demonstration of Theorem 2 and evidence that exact one-electron IPs are reproduced independently of the calibration parameters. Second, robustness of the He Rydberg states against s0s_0: over the range s0=0.13s_0=0.13–0.280.28, both −ϵHOMO=25.12-\epsilon_{\mathrm{HOMO}}=25.12 eV and the count of eight bound orbitals are unchanged; s0s_0 only fine-tunes the alkali −ϵHOMO-\epsilon_{\mathrm{HOMO}}, and the qualitative binding in one-electron-like HOMO systems does not depend on it. The bound

Rydberg series is therefore a consequence of the structural property vx→−1/rv_x\to-1/r, not an artifact of fitting.

Table II shows that even global hybrid functionals containing partial exact exchange (B3LYP [14], PBE0 [20], SCAN0 [21]; 20–25% mixing) bind no He Rydberg-like state in the present calculations. Global hybrids improve the HOMO eigenvalue (here from about 16 to 18 eV) by introducing a fraction of nonlocal exact exchange, but their long-range potential contains only the corresponding fraction of the full Coulombic tail and therefore does not reach the full −1/r-1/r Kohn–Sham asymptote; range-separated hybrids [22, 23] are usually required when a quantitatively correct long-range potential is the primary target. The present framework, by contrast, recovers the −1/r-1/r behavior within a purely semilocal form, without exact-exchange admixture and its attendant nonlocal cost.

The limitations of the framework are structural, not incidental. (i) Thermochemistry: hydrogenic exactness pins the absolute energy of the H atom, which differs from the error-compensated PBE value, so atomization energies involving H are necessarily shifted. (ii) pp-shell HOMOs: since the one-electron detection is based on α\alpha, the asymptotic improvement does not reach many-electron HOMOs (Ne, Ar). (iii) Quantitative IPs of heavier systems: owing to the systematic error of the hydrogen-referenced calibration, −ϵHOMO-\epsilon_{\mathrm{HOMO}} is highly accurate for He but not quantitative in general. (iv) Response properties: because of the density sensitivity of the switch boundaries, higher derivatives such as polarizabilities are outside the scope. With these limitations made explicit, the claim—that Rydberg-like bound states, asymptotics, and hydrogenic levels of ss-shell one-electron-like HOMO systems are recovered—with a nonempirical tail-generating ingredient and a small number of calibrated switching parameters, at semilocal cost—where all existing semilocal functionals including SCAN fail qualitatively—is consistent with every number presented.

The Li–K systematics (Table III) showed that large-core pseudopotentials activate the GP93 term by rendering the valence electron one-electron-like. This suggests that the functional operates naturally in real-space pseudopotential implementations. The fully variational implementation of the qq-switched form of Eq. (18) is also expected to realize its full value on real-space grids, where vlaplv_{\mathrm{lapl}} can be assembled naturally.

In this paper the discriminating power of the triple switch was demonstrated non-self-consistently (Sec. VII). When a self-consistent implementation is attempted in a Gaussian basis, dropping the vlaplv_{\mathrm{lapl}} term associated with qq causes the SCF to diverge through the inconsistency between energy and potential, while including it correctly requires a nonlocal VxcV_{\mathrm{xc}} assembly involving second derivatives of the basis functions. This is the same class of difficulty that Laplacian-level meta-GGAs face, and it is impractical in Gaussian bases. On a real-space grid, by contrast, ∇2n\nabla^2 n and its functional derivative are evaluated naturally by finite differences and vlaplv_{\mathrm{lapl}} can be handled stably. The self-consistent implementation of the triple switch therefore belongs to real-space methods, where the discriminating power established here—no trade-off; atomic tails preserved while molecular divergences are suppressed—should be realized in full.

Finally, in extended systems the −1/r-1/r correction is switched off in the smooth bulk interior through the α\alpha and qq switches, so the functional reduces there to PBE (neither benefit nor breakdown). The correction acts in low-density, one-electron-like regions—the vacuum side of surfaces and the exterior of low-dimensional systems—exactly where an image-potential-like −1/r-1/r asymptote is physically required and systematically missed by semilocal functionals. In preliminary periodic-boundary calculations a simplified switched variant converged stably for solid He over a range of lattice constants, confirming numerical robustness under periodic boundary conditions. Once the real-space implementation of the triple switch is realized, applications are anticipated to problems where the vacuum-side asymptote affects adsorption energetics, such as hydrogen adsorption on metal and semiconductor surfaces—an application domain complementary to bulk thermochemistry, which this functional does not target.

Conclusions

We solved the GP93 hydrogen-exactness condition completely as an inverse problem and, from the GP93 factor defined by its solution, constructed a semilocal exchange ingredient that is exact on the hydrogen 1s1s density and on its coordinate-scaled hydrogenic family, with no empirical parameter entering the ingredient itself. Embedded in a switched meta-GGA form with a small number of calibrated switching parameters, this ingredient produces Rydberg-like series of bound virtual states in all of H, He, Li, Na, and K—systems whose highest occupied orbital is a one-electron-like ss shell—where existing semilocal functionals from PBE to SCAN bind none in the same calculations. The effective domain (ss shells yes, pp shells no) is predicted by the design and confirmed numerically; pseudopotentials that render the valence electron one-electron-like are the key that activates the GP93 term. As an extension toward molecules and solids we introduced a triple switch that adds the reduced Laplacian qq as a third indicator, and showed on fixed densities—for the second-period atoms and ten molecules—that it distinguishes atomic tails from covalent bond centers and suppresses the molecular divergence. Finally, replacing the kinetic-energy density τ\tau by an orbital-free functional with the τW≤τ\tau^{\mathrm{W}} \le\tau bound (Perdew–Constantin type) renders the entire switch a density-only functional and reproduces the Rydberg asymptotics without orbitals, removing the need for a generalized Kohn–Sham solver, while the remaining Laplacian ingredient is naturally handled on real-space grids. A fully self-consistent real-space pseudopotential implementation, systematic basis- and grid-convergence tests of the bound-state counts, and applications to problems where the vacuum-side asymptote matters—such as hydrogen adsorption on surfaces—remain necessary next steps.

Acknowledgments

This work received no funding or external support. During the research, AI assistants—Claude (Claude Opus 4.8, Anthropic) and GPT-5.5 (OpenAI)—were used interactively to help explore analytic manipulations. The author independently verified all derivations and results and bears sole responsibility for the content.

Endpoint Laplace expansion of the tail integral

Substituting U(t)∼t2/3(1−109t−1081t2+⋯ )U(t) \sim t^{2/3}\left(1-\frac{10}{9t}-\frac{10}{81t^2}+\cdots\right) into IU=∫z∞W U dtI_U = \int_z^\infty W\,U\,\mathrm{d}t and applying termwise

∫z∞e−2ttν dt=12e−2zzν[1+ν2z+ν(ν−1)(2z)2+⋯ ](29)\int_z^\infty e^{-2t}t^\nu\,\mathrm{d}t = \frac{1}{2}e^{-2z}z^\nu\left[1+\frac{\nu}{2z}+\frac{\nu(\nu-1)}{(2z)^2}+\cdots\right] \tag*{(29)}

yields Φ(t)=12+109t−2081t2−4302187t3+O(t4)\Phi(t)=\frac{1}{2}+\frac{10}{9}t-\frac{20}{81}t^2-\frac{430}{2187}t^3+O(t^4). The relative error of the three-/four-term truncation is below 2.3×10−2/2.1×10−32.3\times10^{-2}/2.1\times10^{-3} for z∈[2,20]z\in[2,20].

Frobenius coefficients and determination of c0c_0

Decomposing the recursion Eq. (9) as cn=Anc0+Bnc_n=A_nc_0+B_n (A0=1A_0=1, B0=0B_0=0) gives Am+1=(m−2/3)Am(m+1)(m+2)A_{m+1}=\frac{(m-2/3)A_m}{(m+1)(m+2)} and Bm+1=rm+(m−2/3)Bm(m+1)(m+2)B_{m+1}=\frac{r_m+(m-2/3)B_m}{(m+1)(m+2)}. Then c0=[wexact(zm)−ψN(zm)]/ϕN(zm)c_0=[w_{\mathrm{exact}}(z_m)-\psi_N(z_m)]/\phi_N(z_m), with ϕN=∑Anzn\phi_N=\sum A_nz^n, ψN=∑Bnzn\psi_N=\sum B_nz^n, and N=60N=60, is stable to eight digits for zm∈{2,3,4,5}z_m\in\{2,3,4,5\}.

Verification of the implementation

The numerical correctness of the FmixF_{\mathrm{mix}} implementation was verified in several independent layers. The analytic functional derivatives vρv_\rho and vσv_\sigma agree with numerical differentiation to better than 10−910^{-9}. Scale covariance (Theorem 2) is confirmed on the hydrogenic series, with ϵ1s=−Z2/2\epsilon_{1s}=-Z^2/2 reproduced to 0.130.13 mHa for H, 0.100.10 mHa for He, and 1.61.6 mHa for Li. The exchange energy from direct grid integration coincides with the value reported by PySCF to machine precision (2×10−162\times10^{-16}). The SCF density is a true energy minimum (energy increases under occupied–virtual rotations), and three different initial densities converge to the same solution. These checks strongly suggest that the reported trends are not implementation artifacts.

References

References

  1. [1]P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 (1964).
  2. [2]W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 (1965).
  3. [3]P. M. W. Gill and J. A. Pople, Phys. Rev. A 47, 2383 (1993).
  4. [4]J. P. Perdew, R. G. Parr, M. Levy, and J. L. Balduz, Phys. Rev. Lett. 49, 1691 (1982).
  5. [5]A. D. Becke, Phys. Rev. A 38, 3098 (1988).
  6. [6]C. Lee, W. Yang, and R. G. Parr, Phys. Rev. B 37, 785 (1988).
  7. [7]J. P. Perdew, K. Burke, and M. Ernzerhof, Phys. Rev. Lett. 77, 3865 (1996).
  8. [8]R. van Leeuwen and E. J. Baerends, Phys. Rev. A 49, 2421 (1994).
  9. [10]F. Tran and P. Blaha, Phys. Rev. Lett. 102, 226401 (2009).
  10. [11]R. Armiento and S. Kümmel, Phys. Rev. Lett. 111, 036402 (2013).
  11. [12]J. Tao, J. P. Perdew, V. N. Staroverov, and G. E. Scuseria, Phys. Rev. Lett. 91, 146401 (2003).
  12. [13]J. Sun, A. Ruzsinszky, and J. P. Perdew, Phys. Rev. Lett. 115, 036402 (2015).
  13. [14]A. D. Becke, J. Chem. Phys. 98, 5648 (1993).
  14. [15]J. P. Perdew and L. A. Constantin, Phys. Rev. B 75, 155109 (2007).
  15. [16]F. Della Sala, E. Fabiano, and L. A. Constantin, Int. J. Quantum Chem. 116, 1641 (2016).
  16. [17]Q. Sun et al., J. Chem. Phys. 153, 024109 (2020).
  17. [18]S. Lehtola, C. Steigemann, M. J. T. Oliveira, and M. A. L. Marques, SoftwareX 7, 1 (2018).
  18. [20]C. Adamo and V. Barone, J. Chem. Phys. 110, 6158 (1999).
  19. [21]K. Hui and J.-D. Chai, J. Chem. Phys. 144, 0441114 (2016).
  20. [22]H. Hikura, T. Tsuneda, T. Yanai, and K. Hirao, J. Chem. Phys. 101, 3540 (2001).
  21. [23]T. Yanai, D. P. Tew, and N. C. Handy, Chem. Phys. Lett. 393, 51 (2004).
  22. [24]T. H. Dunning, Jr., J. Chem. Phys. 90, 1007 (1989).
  23. [25]R. A. Kendall, T. H. Dunning, Jr., and R. J. Harrison, J. Chem. Phys. 96, 6796 (1992).
  24. [26]L. F. Facios and P. A. Christiansen, J. Chem. Phys. 82, 2664 (1985).
  25. [27]W. J. Stevens, H. Basch, and M. Krauss, J. Chem. Phys. 81, 6026 (1984).
  26. [28]W. J. Stevens, M. Krauss, H. Basch, and P. G. Jasien, Can. J. Chem. 70, 612 (1992).
  27. [29]J. F. Janak, Phys. Rev. B 18, 7165 (1978).
  28. [30]A. Kramida, Yu. Ralchenko, J. Reader, and NIST ASD Team, NIST Atomic Spectra Database, National Institute of Standards and Technology, Gaithersburg, MD, https://physics.nist.gov/asd.

Paper details

vlnc:W_35nee6f5
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Claude Opus 4.8 (bare), GPT-5.5 (bare)

View statement
Looking for related papers…
Contents