betaFeedback

WorksPhysics and Astronomy

Good-ΓΓ Numbers: A Shell-Optimized Route to the Thermodynamic Limit in Quantum Monte Carlo Calculations of the High-Density Two-Dimensional Electron Gas

Abstract

Single-particle finite-size errors remain a central obstacle in quantum Monte Carlo (QMC) simulations of periodic systems, conventionally reduced via twist averaging (TA). Here we introduce Good-ΓΓ (GG) particle numbers as system sizes NN at which the ΓΓ-point Hartree–Fock (HF) energy of the high-density two-dimensional uniform electron gas (2D UEG) lies anomalously close to its thermodynamic-limit (TL) value. All results in this work concern the 2D UEG in its high-density Fermi-liquid regime, rs=1r_s=1–44. We systematically investigate the kinetic- and exchange-energy errors, ΔKE(N)ΔKE(N) and ΔEX(N)ΔEX(N), that underlie this one-body finite-size-error cancellation across four densities and both spin polarizations. We then compare diffusion Monte Carlo (DMC) energies extrapolated to the TL from GG-NN sequences against twist-averaged extrapolations at the same densities and polarizations, finding agreement to within a fraction of a millihartree in every case, independent of whether a 1/N1/N or an N−5/4N^{-5/4} extrapolation form is used.

Introduction

Quantum Monte Carlo (QMC) simulations, including variational and diffusion quantum Monte Carlo (VMC and DMC), of periodic systems are usually performed in finite simulation cells subject to periodic boundary conditions. As a result, they suffer from finite-size (FS) errors, which arise from using a finite number of particles and a finite simulation-cell volume to represent the thermodynamic limit [1, 2, 3, 4, 5, 6]. Controlling and correcting these errors is essential for applying QMC methods to model Hamiltonians such as the electron liquid, and to real periodic systems, particularly when high accuracy is required.

FS errors are conventionally separated into two categories. The first consists of independent-particle, or one-body, FS errors, associated with the discretization of momentum states in a finite simulation cell. Analogous errors appear in density-functional theory (DFT) calculations of periodic systems and are reduced there by dense kk-point sampling of the Brillouin zone (BZ); an alternative, dating back to Baldereschi’s mean-value point [7] and the special kk-point methods of Rajagopal et al. [8, 9], is to replace the full BZ sum with one or a few carefully chosen points. In QMC, the analogous procedure to dense sampling is twist averaging (TA), in which the energy is averaged over boundary-condition phases (twists) to reduce shell-filling effects [10]. The special-point idea has itself been developed further within QMC: Kwee, Zhang, and Krakauer [11] used a mean-field proxy to construct an external, post-processing FS correction, while Dagrada et al. [12] introduced the “exact special twist” (EST) method, which continuously adjusts a single twist angle so that the mean-field energy at that twist reproduces the thermodynamic limit (TL) value to arbitrary numerical precision, for a fixed NN. The present work is closely related to this line of research but explores a different complementary lever. Rather than tuning the twist at fixed NN, we fix the twist at Γ\Gamma and search over NN for particle numbers at which the Γ\Gamma-point HF energy happens to lie close to its TL value. We refer to such NN as Good-Γ\Gamma (GG) particle numbers. As we show below (Sec. III D), selecting GG numbers usefully for thermodynamic-limit extrapolation requires more than ranking candidate NN by ∣ΔEHF(N)∣|\Delta E^{\mathrm{HF}}(N)| alone, as the selected NN must also be reasonably spread in 1/N1/N, since the deepest cancellations can otherwise cluster at nearly adjacent NN, leaving a linear fit with almost no leverage. We take this spacing requirement to be part of a practically useful definition of Good-Γ\Gamma numbers, not an optional refinement.

The second category of FS error consists of many-body contributions, arising from the long-range nature of the Coulomb interaction, the finite extent of the exchange-correlation hole, and the artificial periodicity of the simulation cell. Unlike one-body FS errors, many-body FS errors are not removed by twist averaging and require separate correction schemes or extrapolation to the TL [8, 9, 13, 14, 15, 16].

In this work we restrict our attention to the two-dimensional uniform electron gas (2D UEG) [17] in its high-density Fermi-liquid regime, rs=1r_s = 1–44, which we regard as a controlled, model-system setting for investigating why a HF-level shell-cancellation criterion behaves the way it does, and what it can and cannot be expected to deliver at the correlated level – not as a proposal for a general-purpose replacement for twist averaging. We demonstrate the GG approach at rs=1,2,3,4r_s = 1, 2, 3, 4, for both unpolarized (ζ=0\zeta= 0) and fully spin-polarized (ζ=1\zeta= 1) polarization, at both the HF and the correlated diffusion quantum Monte Carlo (DMC) level, and, by comparing HF, Γ-point DMC, and TA-DMC calculations as functions of system size, determine precisely what the GG criterion does and does not achieve.

Model and computational method

We consider the 2D UEG, at both unpolarized (ζ=0\zeta= 0) and fully spin-polarized (ζ=1\zeta= 1) polarization, characterized by the density parameter rsr_s, using hexagonal simulation cells with periodic boundary conditions. We used standard Slater-Jastrow-Backflow trial wave functions, described in detail in our previous work [17, 18, 19]. Unlike that previous work [17], we did not include the pp and π\pi terms in the Jastrow and backflow functions, in order to reduce computational cost. The wave functions were optimized by variance minimization [20] followed by energy minimization [21]. We used a target population of 1920 walkers in our DMC calculations, with a time step of 0.01 (1/Ha); the same walker population and time step were used at both polarizations, ζ=0\zeta= 0 and ζ=1\zeta= 1, at each density. Twist-averaged energies were obtained using 100 random twists, shifted from Γ and drawn from a uniform distribution over the BZ, for N≤200N \le200, and 50 twists for N>200N > 200. We did not repeat an explicit twist-count convergence test for the present calculations; this choice is based on our previous systematic studies of finite-size and twist-averaging behaviour in the uniform electron gas [6, 17, 22, 23, 24, 25], where this number of twists was established to be sufficient for the densities and system sizes considered here. All QMC calculations were performed using the casino package [26].

Hartree-Fock

We define the finite-cell HF error as

ΔEHF(N)=ENHF−E∞HF,(1)\Delta E^{\mathrm{HF}}(N) = E_N^{\mathrm{HF}} - E_\infty^{\mathrm{HF}}, \tag*{(1)}

and identify Good-Γ (GG) particle numbers as those NN, within a chosen range, for which ∣ΔEHF(N)∣\lvert\Delta E^{\mathrm{HF}}(N)\rvert is minimized. All results in this section use N>100N > 100 throughout. Figure 1 and Figure 2 show ΔEHF(N)\Delta E^{\mathrm{HF}}(N) for Γ-point and TA HF calculations at rs=1,2,3,4r_s = 1, 2, 3, 4, over N=102N = 102 to 5000, for ζ=0\zeta= 0 and ζ=1\zeta= 1 respectively obtained using 10410^4 twists. In both cases, the Γ-point results show the large shell-filling oscillations that are well known for Γ-point calculations of the UEG [1, 10], while the TA results vary smoothly with NN and do not minimize ∣ΔEHF(N)∣\lvert\Delta E^{\mathrm{HF}}(N)\rvert

Finite-cell HF energy error for the unpolarized 2D UEG at four densities, showing Γ-point and twist-averaged results

Figure 1. Finite-cell HF energy error, ΔEHF=ENHF−E∞HF\Delta E^{\mathrm{HF}} = E_N^{\mathrm{HF}} - E_\infty^{\mathrm{HF}} (Ha/electron), as a function of 1/N1/N for the unpolarized (ζ=0\zeta= 0) 2D UEG at rs=1,2,3,4r_s = 1, 2, 3, 4, using hexagonal simulation cells, N=102N = 102–5000. Γ-point (left column) and TA (right column) results are shown separately, one row per density. The Γ-point results oscillate in sign; the TA results remain single-signed throughout.

Plots of $\Delta HF=\mathrm{HF}_N-\mathrm{HF}_\infty$ versus $1/N$ for $\zeta=1$ 2D UEG

Figure 2. As Fig. 1, for the fully spin-polarized (ζ=1\zeta=1) 2D UEG.

at any particular NN. The oscillation itself is not new; what we examine here is its decomposition into kinetic- and exchange-energy components, and the consequences of that decomposition for the particular NN at which the total error nearly vanishes.

Cancellation between ΔKE and ΔEX at Good-Γ numbers

At Good-Γ numbers, ΔKE(N)\Delta KE(N) and ΔEX(N)\Delta EX(N) are individually large and opposite in sign, canceling in the total ΔEHF(N)\Delta E^{\mathrm{HF}}(N). We tested this directly at every GG number identified across N=102N=102–5000, at all four densities and both polarizations and found that ΔKE(N)\Delta KE(N) and ΔEX(N)\Delta EX(N) are opposite in sign and individually 5× to more than 10410^4× larger in magnitude than their sum, with the smaller ratios typical and the largest reflecting a small number of near-exact cancellations.

ζ\zetarsr_smedian ΔEX/ΔKE\Delta EX/\Delta KE
01−0.727
02−1.454
03−2.191
04−2.909
11−0.654
12−1.309
13−1.965
14−2.617

Table 1. Median ΔEX(N)/ΔKE(N)\Delta EX(N)/\Delta KE(N) over N=102N=102–5000, restricted at each density to the half of that range with the larger ∣ΔKE(N)∣|\Delta KE(N)| (a rank-based, not absolute, cut, chosen so the same fraction and count of points enters the median at every density – an absolute cut biases the sample toward progressively fewer, smaller-NN points as rsr_s increases, since ΔKE\Delta KE shrinks with rsr_s). The ratio is negative at every density and both polarizations tested, and its magnitude increases with rsr_s in close to direct proportion (1:2:3:4) at both polarizations.

Table I shows that ΔEX(N)/ΔKE(N)\Delta EX(N)/\Delta KE(N) is negative at every density and both polarizations, confirming that the two errors are consistently opposite in sign, not just occasionally so. Its magnitude grows with rsr_s in close to exact proportion (1:2:3:4 at both polarizations), and is somewhat larger at ζ=0\zeta=0 than at ζ=1\zeta=1 for a given rsr_s. We report the median rather than the mean because ΔKE(N)\Delta KE(N) oscillates through zero, so a small number of points close to a zero crossing produce near-singular ratios that can dominate a mean (at rs=1r_s=1, ζ=0\zeta=0, for example, the mean of the same subset has the wrong sign) but leave the median unaffected.

One-body structure beyond the cancellation ratio

Three further HF-level properties, established from the N=102N=102–5000 dataset, sharpen this picture. First, it is the kinetic-energy error, not the exchange-energy error, that qualitatively distinguishes Γ\Gamma-point from TA behavior: over this range, TA’s ΔKE(N)\Delta KE(N) and ΔEX(N)\Delta EX(N) never change sign, and neither does the Γ\Gamma-point ΔEX(N)\Delta EX(N); only Γ\Gamma-point ΔKE(N)\Delta KE(N) changes sign, repeatedly, 58 times for ζ=0\zeta=0 and 81 for ζ=1\zeta=1, identically at all four densities in each case. Second, at N≥4000N\ge4000 the Γ\Gamma-point kinetic-energy error is 27–28× larger than TA’s at both polarizations, while its exchange-energy error is only 0.91–0.96× TA’s; TA’s advantage over a generic Γ\Gamma-point calculation is thus overwhelmingly a kinetic-energy effect, and for exchange alone a single well-chosen Γ\Gamma-point calculation remains essentially as converged as TA even at the largest NN studied. These figures hold to the precision quoted regardless of density or polarization, indicating they are intrinsic to the Γ\Gamma-point occupation rather than incidental to any one case.

Figures 3 and 4 show ΔKE(N)\Delta KE(N) and ΔEX(N)\Delta EX(N) directly, for ζ=0\zeta=0 and ζ=1\zeta=1 respectively. The contrast is visually immediate in both cases, with ΔEX(N)\Delta EX(N) tracing a single, smoothly decaying, single-signed curve at every density while ΔKE(N)\Delta KE(N) oscillates through zero with a slowly decaying envelope. Figure 5 shows the same two quantities for the TA energies, both of which decay smoothly and remain single-signed throughout, with none of the Γ\Gamma-point oscillation visible above, which is a direct visual confirmation of the sign-change counts just described.

Third, the set of GG-NN is density-dependent, meaning no single NN ranks among the best Good-Γ\Gamma candidates at all four densities simultaneously, at either polarization, and the overlap between the GG sets at two densities is larger the closer those densities are (the overlap counts are tabulated in the Supplemental Material [27]). A practical consequence is that GG numbers must be identified separately at each rsr_s and ζ\zeta from the Γ\Gamma-point HF scan, which is inexpensive; there is no universal GG set.

Per-twist structure: testing the sign asymmetry directly

Two results (Table II) are robust across the eight NN considered, including the four chosen with no reference to ∣ΔEHF(N)∣\lvert\Delta E^{\mathrm{HF}}(N)\rvert, and are therefore properties of the TA distribution itself rather than artifacts of the Good-Γ\Gamma selection: (i) ΔEX\Delta EX is negative at every individual twist, not merely on average – a stronger statement than the twist-averaged result of Sec. III B; and (ii) ΔKE\Delta KE is positive for a majority of twists at every NN tested, confirming that the positive sign of the twist-averaged ΔKE\Delta KE (dashed line, Figure 6) reflects a genuine majority rather than a small number of large outliers. Figure 6 further shows that ΔKE\Delta KE and ΔEX\Delta EX depend on the twist vector itself very differently at the two kinds of NN studied.

Plots of $\Delta KE$ and $\Delta EX$ versus twist-vector magnitude

Figure 6. ΔKE\Delta KE (left) and ΔEX\Delta EX (right) vs. twist-vector magnitude ∣ks∣|k_s|, rs=1r_s = 1, ζ=0\zeta= 0, 5000 twists per panel, one row per NN (top four: Good-Γ\Gamma NN with QMC data; bottom four: NN chosen without reference to ΔEHF\Delta E^{\mathrm{HF}}, within the range most relevant to QMC calculations). The dashed line in each ΔKE\Delta KE panel marks the twist-averaged value.

$\Delta KE(N)$ and $\Delta EX(N)$ for the fully spin-polarized 2D UEG

Figure 4. As Figure 3, for the fully spin-polarized (ζ=1\zeta= 1) 2D UEG.

Γ-point $\Delta KE(N)$ and $\Delta EX(N)$ versus $1/N$ for the unpolarized 2D UEG

Figure 3. Γ-point ΔKE(N)\Delta KE(N) (left) and ΔEX(N)\Delta EX(N) (right) vs. 1/N1/N, unpolarized (ζ=0\zeta= 0) 2D UEG, rs=1,2,3,4r_s = 1, 2, 3, 4 (one row per density), N=102N = 102–50005000. ΔEX(N)\Delta EX(N) decays smoothly and is single-signed at every density; ΔKE(N)\Delta KE(N) oscillates through zero repeatedly, with an envelope that decays much more slowly.

Thermodynamic-limit extrapolation within the QMC-relevant range

We restrict this comparison to 100<N<500100 < N < 500, the range within which QMC calculations in this work are actually performed. For each density, we linearly extrapolate (1/N1/N) the TA-HF energy using all available NN in this range, and separately extrapolate the Γ\Gamma-point HF energy using four Good-Γ\Gamma NN selected to have ∣ΔEΓHF∣\lvert\Delta E^{\mathrm{HF}}_{\Gamma}\rvert as small as possible subject to a minimum spacing of 50 in NN between any two selected points, chosen so the linear fit is not dominated by nearly-degenerate points. The four selected NN are {214,290,372,468}\{214, 290, 372, 468\}, {208,266,368,454}\{208, 266, 368, 454\}, {138,204,274,350}\{138, 204, 274, 350\}, and {202,274,354,450}\{202, 274, 354, 450\} for rs=1,2,3,4r_s=1,2,3,4 respectively; the minimum gap between consecutive selected NN is 58–96, comfortably spread across the 400-wide window at every density. Both extrapolations are compared against the exact analytic thermodynamic-limit value.

Twist-averaged $\Delta KE_{\mathrm{TAV}}(N)$ and $\Delta EX_{\mathrm{TAV}}(N)$ for the unpolarized ($\zeta=0$) 2D UEG

Figure 5. As Fig. 3, but for the twist-averaged ΔKETAV(N)\Delta KE_{\mathrm{TAV}}(N) and ΔEXTAV(N)\Delta EX_{\mathrm{TAV}}(N) rather than the Γ\Gamma-point values, unpolarized (ζ=0\zeta=0) 2D UEG.

Both extrapolations agree with the analytic value to well within 0.2 mHa at every density (Table III, Figure 7), and neither is uniformly better than the other: GG-NN is more accurate at rs=2r_s=2, comparably accurate at rs=3r_s=3,

HF energy per electron versus $1/N$ for $r_s=1,2,3,4$ and $\zeta=0$

Figure 7. HF energy per electron vs. 1/N1/N, rs=1,2,3,4r_s = 1, 2, 3, 4, ζ=0\zeta= 0, 100<N<500100 < N < 500: twist-averaged energy (all available NN) and Γ\Gamma-point energy at four well-spaced Good-Γ\Gamma NN, each linearly extrapolated to 1/N=01/N = 0 and compared against the exact analytic value (dotted line).

NNtypeΔEΓHF\Delta E^{\mathrm{HF}}_{\Gamma}ΔKE\Delta KE % pos. (5k)ΔKE\Delta KE % pos. (10k)
104GG−4.00×10−4-4.00 \times10^{-4}81.0%80.5%
128GG−2.87×10−4-2.87 \times10^{-4}82.6%82.3%
158GG−1.23×10−5-1.23 \times10^{-5}90.4%91.0%
190GG+2.61×10−4+2.61 \times10^{-4}51.8%51.8%
250arbitrary−1.79×10−3-1.79 \times10^{-3}72.0%72.3%
300arbitrary−1.84×10−3-1.84 \times10^{-3}75.0%75.6%
400arbitrary−2.11×10−3-2.11 \times10^{-3}53.9%54.8%
500arbitrary−1.50×10−3-1.50 \times10^{-3}63.1%63.7%

Table 2. Per-twist statistics of ΔKE\Delta KE and ΔEX\Delta EX at rs=1r_s=1, ζ=0\zeta=0, for four Good-Γ\Gamma and four arbitrary NN, at 5000 and 10000 twists per NN. ΔEX\Delta EX is negative at every individual twist at every NN; ΔKE\Delta KE is positive for a majority of twists at every NN. The fraction of positive ΔKE\Delta KE changes by less than one percentage point between 5000 and 10000 twists, consistent with statistical noise, so it is converged at 5000 twists.

rsr_sTAV (all NN)GG-NN (4 points)
1−0.113-0.113+0.146+0.146
2−0.042-0.042+0.005+0.005
3−0.025-0.025+0.025+0.025
4−0.017-0.017−0.047-0.047
rsr_sTAV (all NN)GG-NN (4 points)
1−0.113-0.113+0.146+0.146
2−0.042-0.042+0.005+0.005
3−0.025-0.025+0.025+0.025
4−0.017-0.017−0.047-0.047

Table 3. Difference (mHa) between the linearly extrapolated (1/N1/N) HF energy and the exact analytic value, 100<N<500100<N<500, ζ=0\zeta=0.

and less accurate than TAV at rs=1r_s=1 and rs=4r_s=4. This shows that, once the Good-Γ\Gamma NN are chosen to be well separated in 1/N1/N rather than selected by rank alone (Sec. III B), a handful of Γ\Gamma-point calculations can extrapolate to the thermodynamic limit about as reliably as a full TA sequence over the same range, at the HF level.

Quantum Monte Carlo

We test whether the HF-level GG criterion, shown in Sec. III, provides a practical route to the DMC thermodynamic limit (TL). Throughout this section we compare two DMC sequences evaluated at the same Good-Γ\Gamma particle numbers, including the Γ\Gamma-point DMC energies themselves (“GG”), and the twist-averaged DMC energies at those same NN (“TAV-GG”, twist-averaged at Good-Γ\Gamma NN). The latter serves as the reference against which the Γ\Gamma-point sequence is judged, since it removes the one-body shell error by averaging rather than by cancellation. The result is that a GG-based DMC sequence extrapolates to the TL in close agreement with the TAV-GG extrapolation, at every density and polarization tested (Sec. IV B), while requiring substantially fewer independent DMC calculations to obtain (§ 1.4). This does not mean single-NN GG accuracy is guaranteed at the DMC level, as the correlated-energy finite-size error is expected to be dominated by many-body effects that a purely one-body criterion does not address, and we do not find that it is guaranteed. What matters for a converged benchmark, however, is the extrapolated result, and that is where GG performs reliably. A direct test isolating the one-body contribution to the correlated-energy error, contrasting GG against generic (“Bad-Γ\Gamma”) particle numbers known to be poor at the one-body level, is given in the Supplemental Material [27].

Resolving the opposite-slope puzzle

The GG and TA-DMC sequences approach the TL from opposite sides as functions of 1/N1/N. Our results show that TA-DMC energies satisfy ETAV(N)−E(∞)<0E_{\mathrm{TAV}}(N)-E(\infty)<0 (the finite cell overbinds), while GG DMC energies satisfy EGG(N)−E(∞)>0E_{\mathrm{GG}}(N)-E(\infty)>0. This sign reflects an over correction, as the many-body finite-size correction beyond the one-body error is generically positive [13, 15], while the one-body cancellation at GG numbers leaves no reason for the corresponding interacting-system one-body bias to vanish at the same NN. When this residual one-body bias is larger in magnitude than, and opposite in sign to, the true positive many-body correction, it flips the net sign of the observed error, producing the anomalous EGG(N)−E(∞)>0E_{\mathrm{GG}}(N)-E(\infty)>0 trend.

The positive sign of the many-body correction, established in the literature [13, 15], can also be tested directly from our own data, without relying on an external correction scheme. We define the correlation energy as Ecorr(N)=EDMC(N)−EHF(N)E_{\mathrm{corr}}(N)=E_{\mathrm{DMC}}(N)-E_{\mathrm{HF}}(N), computed separately along the GG and TAV sequences using each sequence’s own Γ\Gamma-point or TA-HF energy at the same NN, and examine its dependence on 1/N1/N (Figure 8, Table IV). At every density, both polarizations, and both sequences – 16 independent tests in total – a linear fit to Ecorr(N)E_{\mathrm{corr}}(N) vs. 1/N1/N has a positive slope, i.e. the trend is toward a more negative (more strongly bound) correlation energy as NN increases, with no exceptions. Two of the sixteen raw sequences (each with only 3–4 available NN) show a small point-to-point deviation from strict monotonicity, plausibly ordinary scatter given the small sample size; nonetheless, every individual finite-NN data point in all sixteen sequences is less negative than its own linearly extrapolated N→∞N\to\infty value, directly confirming a positive correlation-energy finite-size error using only our own raw, uncorrected DMC and HF data.

Eight-panel plot of correlation energy versus inverse particle number for GG and TAV sequences

Figure 8. Correlation energy Ecorr=EDMC−EHFE_{\mathrm{corr}} = E_{\mathrm{DMC}} - E_{\mathrm{HF}} vs. 1/N1/N, computed separately along the GG and TAV sequences, all four densities and both polarizations. In every one of the 16 panel/sequence combinations, a linear fit to Ecorr(N)E_{\mathrm{corr}}(N) has a positive slope.

Linearity and thermodynamic-limit extrapolation

We fit E(N)=E(∞)+b/NE(N)=E(\infty)+b/N to the GG and TAV-GG DMC sequences at each density, at both polarizations. The fits are shown in Figure 9 and Figure 10, and in Table V reports the extrapolated N→∞N\to\infty energies for both sequences; the quoted uncertainties are the standard errors of the fitted intercepts.

Four-panel linear extrapolation plots of GG and TAV-GG DMC energies

Figure 9. Linear extrapolation of GG and TAV-GG DMC energies to 1/N=01/N = 0, 2D UEG at rs=1,2,3,4r_s = 1, 2, 3, 4, ζ=0\zeta= 0.

Linear extrapolation plots for $r_s=1,2,3,4$ and $\zeta=1$

Figure 10. Linear extrapolation of GG and TAV-GG DMC energies to 1/N=01/N=0, 2D UEG at rs=1,2,3,4r_s=1,2,3,4, ζ=1\zeta=1.

ζ\zetarsr_sslope, GG (mHa)slope, TAV (mHa)
01306.7257.9
02116.8146.1
0357.8114.8
0417.665.9
11156.3180.3
1277.399.1
1348.758.3
1437.754.4

Table 4. Slope bb of the linear fit Ecorr(N)=Ecorr(∞)+b/NE_{\mathrm{corr}}(N)=E_{\mathrm{corr}}(\infty)+b/N to the GG and TAV correlation-energy sequences of Fig. 8. The slope is positive at every density, polarization, and sequence tested (16/16), confirming a positive correlation-energy finite-size error directly from our own data.

intercepts.

ζ\zetarsr_sE∞GGE_{\infty}^{\mathrm{GG}} (Ha)E∞TAV−GGE_{\infty}^{\mathrm{TAV-GG}} (Ha)Δ\Delta (mHa)
01−0.209980(888)−0.210289(438)0.31
02−0.258570(109)−0.258211(112)0.36
03−0.212207(78)−0.211999(47)0.21
04−0.175730(136)−0.175685(101)0.05
110.126332(31)0.126412(36)0.08
12−0.194610(40)−0.194575(69)0.04
13−0.189244(20)−0.189146(24)0.10
14−0.165168(21)−0.165115(14)0.05

Table 5. Thermodynamic-limit DMC energies per electron extrapolated from GG and TAV-GG sequences, all four densities, both polarizations. The two extrapolations agree to 0.04–0.4 mHa, comparable to or smaller than typical statistical uncertainties in DMC benchmarks of this system, and are, if anything, slightly tighter at ζ=1\zeta=1 than at ζ=0\zeta=0.

Choice of extrapolation form. The leading many-body finite-size correction to the energy of the 2D UEG is expected to scale as N−5/4N^{-5/4} rather than 1/N1/N, because the static structure factor behaves as S(k)∝k3/2S(k)\propto k^{3/2} at small kk in two dimensions, as a consequence of the k\sqrt{k} plasmon dispersion [3, 4, 13]. The 1/N1/N form used above is the conventional empirical choice, and over the range 100<N<500100<N<500 spanned by 3–5 points per sequence the two forms are not easily distinguished. To establish that our conclusions do not depend on this choice, we repeated every extrapolation with E(N)=E(∞)+bN−5/4E(N)=E(\infty)+bN^{-5/4} (Table VI). The GG and TAV-GG intercepts again agree to within a fraction of a millihartree in every case, and the agreement is in fact tighter with N−5/4N^{-5/4} at six of the eight density/polarization combinations; the one case in which it is looser (rs=1r_s=1, ζ=0\zeta=0) is the sequence with the largest fit residual. We retain 1/N1/N as the primary form in the remainder of the paper for two reasons: it is the form against which the slope analysis of Sec. IV A is most transparently interpreted, and the GG sequences are Γ\Gammapoint rather than twist-averaged energies, so they carry a residual one-body term (Sec. IV A) for which no particular power law is expected. The central result, agreement between the GG and TAV-GG thermodynamic limits, is the same under either form.

ζ\zetarsr_sE∞GGE_{\infty}^{\mathrm{GG}} (Ha)E∞TAV−GGE_{\infty}^{\mathrm{TAV-GG}} (Ha)Δ\Delta (mHa)
01−0.209754(724)-0.209754(724)−0.210385(350)-0.210385(350)0.63
02−0.258413(86)-0.258413(86)−0.258264(88)-0.258264(88)0.15
03−0.212125(57)-0.212125(57)−0.212031(37)-0.212031(37)0.09
04−0.175740(110)-0.175740(110)−0.175723(79)-0.175723(79)0.02
110.126478(8)0.126478(8)0.126349(38)0.126349(38)0.13
12−0.194560(27)-0.194560(27)−0.194610(61)-0.194610(61)0.05
13−0.189203(14)-0.189203(14)−0.189181(19)-0.189181(19)0.02
14−0.165139(15)-0.165139(15)−0.165132(13)-0.165132(13)0.01

Table VI. As Table V, but using the extrapolation form E(N)=E(∞)+bN−5/4E(N)=E(\infty)+bN^{-5/4} appropriate to the leading many-body finite-size correction in two dimensions [4]. The GG and TAV-GG extrapolations agree to 0.01–0.6 mHa; the conclusions drawn from Table V are unchanged.

The full VMC, DMC, and HF data at the GG particle numbers underlying the extrapolations above, for every density and both polarizations, are provided in the Supplemental Material [27].

Testing a one-body correction to the DMC extrapolation

Since Sec. III A establishes that ΔEΓHF(N)\Delta E_{\Gamma}^{\mathrm{HF}}(N) at Good-Γ\Gamma numbers is a one-body cancellation, we tested directly whether adding this quantity to the DMC energy improves the 1/N1/N extrapolation, using the same GG-NN sequences as Table V. We considered EDMC(N)+ΔEΓHF(N)E_{\mathrm{DMC}}(N)+\Delta E_{\Gamma}^{\mathrm{HF}}(N) (signed) and EDMC(N)+∣ΔEΓHF(N)∣E_{\mathrm{DMC}}(N)+|\Delta E_{\Gamma}^{\mathrm{HF}}(N)| (absolute value), each fit linearly in 1/N1/N as in Sec. IV B (Table VII).

rsr_squantityRMS resid. (mHa)intercept (Ha)
1EDMCE_{\mathrm{DMC}}0.277−0.209980(889)-0.209980(889)
1EDMC+ΔEΓHFE_{\mathrm{DMC}}+\Delta E_{\Gamma}^{\mathrm{HF}}0.288−0.208997(925)-0.208997(925)
1EDMC+∣ΔEΓHF∣E_{\mathrm{DMC}}+|\Delta E_{\Gamma}^{\mathrm{HF}}|0.197−0.210128(632)-0.210128(632)
2EDMCE_{\mathrm{DMC}}0.059−0.258570(109)-0.258570(109)
2EDMC+ΔEΓHFE_{\mathrm{DMC}}+\Delta E_{\Gamma}^{\mathrm{HF}}0.103−0.258640(190)-0.258640(190)
2EDMC+∣ΔEΓHF∣E_{\mathrm{DMC}}+|\Delta E_{\Gamma}^{\mathrm{HF}}|0.059−0.258549(110)-0.258549(110)
3EDMCE_{\mathrm{DMC}}0.040−0.212206(78)-0.212206(78)
3EDMC+ΔEΓHFE_{\mathrm{DMC}}+\Delta E_{\Gamma}^{\mathrm{HF}}0.064−0.212244(126)-0.212244(126)
3EDMC+∣ΔEΓHF∣E_{\mathrm{DMC}}+|\Delta E_{\Gamma}^{\mathrm{HF}}|0.047−0.212241(92)-0.212241(92)

Table VII. RMS residual of a linear-in-1/N1/N fit, and the resulting extrapolated intercept, for the raw DMC energy and for two variants corrected by the Γ\Gamma-point HF finite-size error at the same GG-NN sequences as Table V, ζ=0\zeta=0. Neither correction gives a reliable, density-independent improvement (see text).

The signed correction worsens linearity at every density (the RMS residual increases by 4%, 74%, and 60% at rs=1r_s=1, 2, 3). The absolute-value correction improves linearity at rs=1r_s=1 (1.4× smaller residual, with the intercept moving closer to the TAV-GG reference of Table V, 0.31 → 0.16 mHa), but leaves it unchanged at rs=2r_s=2 and slightly worsens it at rs=3r_s=3. In no case does either correction shift the extrapolated intercept by more than its own fitted uncertainty.

This negative result is itself consistent with the central finding of Secs. III A–III B, as ΔEΓHF(N)\Delta E_{\Gamma}^{\mathrm{HF}}(N) at Good-Γ\Gamma numbers is small by cancellation and oscillates in sign with NN, even within a single density’s own GG-NN set. It is therefore not a smooth, physically meaningful correction, and there is no reason to expect it to track, or usefully correct for, the DMC energy’s own finite-size error. Our results confirm that it does not do so reliably.

Discussion and scope

Scope. All results in this work concern the high-density 2D UEG, rs=1r_s=1–4, in its Fermi-liquid regime. We do not claim, and do not test, applicability to real materials, where the number of electrons in a simulation cell is fixed by the lattice and unit-cell choice and the Fermi surface need not be isotropic. Nor do we claim, or test, applicability to the three-dimensional UEG, to the low-density 2D UEG, or to any other model system. The empirical characterization of Sec. III A was obtained specifically for the high-density 2D case, and the quantitative behavior established in Sec. IV should not be assumed to carry over. Extending the GG concept to three dimensions, to lower densities, to real materials, and to anisotropic or multi-sheet Fermi surfaces are each left to separate, future studies. Every statement below is to be read within this scope.

What the GG approach offers within that scope. For the 2D UEG, the GG approach has three practical properties relative to twist averaging. First, the Γ\Gamma point is a valid, high-symmetry reciprocal-lattice point for any simulation-cell shape, so identifying GG numbers requires no lattice-specific construction analogous to Baldereschi’s mean-value point [7] or the special kk-point methods of Rajagopal et al. [8, 9]; the same procedure, scan the HF energy over NN at Γ\Gamma, applies unchanged regardless of cell geometry. Second, the Bloch orbitals at Γ\Gamma can always be chosen real, whereas orbitals at a generic twist are complex; working with real rather than complex orbitals reduces both the memory footprint and the arithmetic cost of every VMC and DMC step. Third, the GG-based sequences used in Sec. IV B require only 3–5 independent, real-valued Γ\Gamma-point DMC calculations per density, whereas a TA sequence spanning the same range of NN requires 50–100 independent, generally complex, twist calculations at each of those NN (Sec. II). A GG-based extrapolation therefore requires far fewer inde- @references 񟿿 pendent calculations and far less workflow overhead than a TA one covering the same NN range. Whether this translates into a proportionate saving in total CPU time depends on how sampling effort is distributed among twists, as discussed next, and we make no claim that it does; we also emphasize again that GG is not more accurate than TA at any single fixed NN.

Computational cost of twist averaging. For a fixed target statistical uncertainty, Monte Carlo sampling effort can in principle be distributed among twists, so the cost of twist averaging does not scale linearly with the number of twists, and the total number of Monte Carlo steps needed for a TA energy of given precision need not greatly exceed that for a single Γ\Gamma-point energy of the same precision. Twist averaging nonetheless requires separate equilibration, walker populations, orbital sets, and data handling for each twist, and may require twist-dependent wave-function preparation or optimization; it therefore carries non-negligible setup and workflow overhead relative to a single, carefully selected Γ\Gamma-point calculation. This overhead is system- and implementation-dependent, and is relatively modest for the UEG studied here. The honest summary is that GG reduces the number and complexity of calculations, not necessarily the number of Monte Carlo samples.

Relation to prior special-point methods. The GG criterion explored here is conceptually related to, but distinct from, the exact special twist (EST) method of Dagrada et al. [12] and the mean-field-based correction of Kwee, Zhang, and Krakauer [11]. EST continuously tunes a single twist angle at fixed NN to drive the mean-field FS error to an arbitrarily small, controlled value; it is, by construction, always available and always accurate to the desired tolerance. GG, by contrast, fixes the twist at Γ\Gamma and searches over the discrete set of accessible NN, achieving a small one-body error only at specific, sparsely distributed particle numbers, and, as shown in Sec. III, by cancellation rather than by directly minimizing the error. We do not claim GG is more accurate or more broadly applicable than EST or than twist averaging. The shell oscillation of the Γ\Gamma-point energy (Figure 1) has been known since the introduction of twist averaging [10]; what we add is its decomposition into kinetic and exchange components at the particle numbers where it nearly vanishes, the demonstration that this vanishing is a cancellation rather than a convergence, and a direct test, against twist-averaged DMC at eight density/polarization combinations, of what such particle numbers can and cannot deliver at the correlated level (Sec. IV).

Conclusion

We have characterized Good-Γ\Gamma (GG) particle numbers – system sizes at which the Γ\Gamma-point HF energy of the high-density 2D UEG lies anomalously close to its thermodynamic-limit value – and shown, at rs=1,2,3,4r_s=1,2,3,4 and at both unpolarized (ζ=0\zeta=0) and fully spin-polarized (ζ=1\zeta=1) polarization, that this near-cancellation involves individually large kinetic- and exchange-energy errors that cancel in the total. The ratio between them is negative and grows with rsr_s in close to direct proportion at every density and polarization studied, though its overall scale is somewhat larger at ζ=0\zeta=0 than at ζ=1\zeta=1 at the same rsr_s. At all four densities and both polarizations, we additionally tested the correlated DMC level and found that a GG-based sequence, once selected with NN reasonably spread in 1/N1/N, extrapolates to the thermodynamic limit in good agreement with twist-averaged DMC extrapolation over the same range, whether a 1/N1/N or an N−5/4N^{-5/4} extrapolation form is used. We clarify the reason behind the opposite-sign finite-size trends of GG and TA sequences. Within the scope of this work – the high-density 2D UEG – a well-spaced GG sequence offers a practical route to the thermodynamic limit that requires no lattice-specific construction, always admits a real wave function, and needs far fewer independent DMC calculations than a TA sequence covering the same NN range.

Acknowledgment

S.A, A.P, and M.S.B acknowledge the support of the Leverhulme Trust under the grant agreement RPG-2023-253. S. Azadi and T.D. Kühne acknowledge the computing time provided to them on the high-performance computers Noctua2 at the NHR Center in Paderborn (PC2). The data that support the findings of this article are openly available [28], embargo periods may apply.

References

References

  1. [1]W. M. C. Foulkes, L. Mitas, R. J. Needs, and G. Rajagopal, Quantum monte carlo simulations of solids, Rev. Mod. Phys. 73, 33 (2001).DOI
  2. [2]M. Holzmann, B. Bernu, and D. M. Ceperley, Finite-size analysis of the fermi liquid properties of the homogeneous electron gas, Journal of Physics: Conference Series 321, 012020 (2011).DOI
  3. [3]M. Holzmann, R. C. Clay, III, M. A. Morales, N. M. Tubman, D. M. Ceperley, and C. Pierleoni, Theory of finite size effects for electronic quantum Monte Carlo calculations of liquids and solids, Phys. Rev. B 94, 035126 (2016).DOI
  4. [4]N. D. Drummond, R. J. Needs, A. Sorouri, and W. M. C. Foulkes, Finite-size errors in continuum quantum monte carlo calculations, Phys. Rev. B 78, 125106 (2008).DOI
  5. [5]S. Azadi and W. M. C. Foulkes, Systematic study of finite-size effects in quantum monte carlo calculations of real metallic systems, J. Chem. Phys. 143, 102807 (2015).DOI
  6. [6]S. Azadi and W. M. C. Foulkes, Efficient method for grand-canonical twist averaging in quantum monte carlo calculations, Phys. Rev. B 100, 245142 (2019).DOI
  7. [7]A. Baldereschi, Mean-value point in the brillouin zone, Phys. Rev. B 7, 5212 (1973).DOI
  8. [8]G. Rajagopal, R. Needs, S. Kenny, W. M. C. Foulkes, and A. James, Quantum monte carlo calculations for solids using special k points methods, Phys. Rev. Lett. 73, 1959 (1994).pubmed.ncbi.nlm.nih.gov/10056932
  9. [9]G. Rajagopal, R. Needs, A. James, S. Kenny, and W. M. C. Foulkes, Variational and diffusion quantum monte carlo calculations at nonzero wave vectors: Theory and application to diamond-structure germanium, Phys. Rev. B 51, 10591 (1995).pubmed.ncbi.nlm.nih.gov/9977754
  10. [10]C. Lin, F. H. Zong, and D. M. Ceperley, Twist-averaged boundary conditions in continuum quantum monte carlo algorithms, Phys. Rev. E 64, 016702 (2001).pubmed.ncbi.nlm.nih.gov/11461437
  11. [11]H. Kwee, S. Zhang, and H. Krakauer, Finite-size correction in many-body electronic structure calculations, Phys. Rev. Lett. 100, 126404 (2008).pubmed.ncbi.nlm.nih.gov/18517892
  12. [12]M. Dagrada, S. Karakuzu, V. L. Vildosola, M. Casula, and S. Sorella, Exact special twist method for quantum Monte Carlo simulations, Phys. Rev. B 94, 245108 (2016).DOI
  13. [13]S. Chiesa, D. M. Ceperley, R. M. Martin, and M. Holzmann, Finite-size error in many-body simulations with long-range interactions, Phys. Rev. Lett. 97, 076404 (2006).pubmed.ncbi.nlm.nih.gov/17026255
  14. [14]L. M. Fraser, W. M. C. Foulkes, G. Rajagopal, R. Needs, S. Kenny, and A. Williamson, Finite-size effects and coulomb interactions in quantum monte carlo calculations for homogeneous systems with periodic boundary conditions, Phys. Rev. B 53, 1814 (1996).pubmed.ncbi.nlm.nih.gov/9983641
  15. [15]P. R. C. Kent, R. Q. Hood, A. J. Williamson, R. J. Needs, W. M. C. Foulkes, and G. Rajagopal, Finite-size errors in quantum many-body simulations of extended systems, Phys. Rev. B 59, 1917 (1999).DOI
  16. [16]A. J. Williamson, G. Rajagopal, R. J. Needs, L. M. Fraser, W. M. C. Foulkes, Y. Wang, and M.-Y. Chou, Elimination of coulomb finite-size effects in quantum many-body simulations, Phys. Rev. B 55, R4851 (1997).DOI
  17. [17]S. Azadi, N. Drummond, and S. Vinko, Quantum monte carlo study of the phase diagram of the two-dimensional uniform electron liquid, Phys. Rev. B 110, 245145 (2024).
  18. [18]N. D. Drummond, M. D. Towler, and R. J. Needs, Jastrow correlation factor for atoms, molecules, and solids, Phys. Rev. B 70 (2004).DOI
  19. [19]P. López Ríos, P. Seth, N. D. Drummond, and R. J. Needs, Framework for constructing generic Jastrow correlation factors, Phys. Rev. E 86, 036703 (2012).DOI
  20. [20]N. D. Drummond and R. J. Needs, Variance-minimization scheme for optimizing Jastrow factors, Phys. Rev. B 72, 085124 (2005).DOI
  21. [21]J. Toulouse and C. J. Umrigar, Optimization of quantum monte carlo wave functions by energy minimization, J. Chem. Phys. 126, 084102 (2007).
  22. [22]S. Azadi, N. D. Drummond, and W. M. C. Foulkes, Quasiparticle effective mass of the three-dimensional fermi liquid by quantum monte carlo, Phys. Rev. Lett. 127, 086401 (2021).DOI
  23. [23]S. Azadi and N. D. Drummond, Low-density phase diagram of the three-dimensional electron gas, Phys. Rev. B 105, 245135 (2022).DOI
  24. [24]S. Azadi, N. D. Drummond, and S. M. Vinko, Correlation energy of the paramagnetic electron gas at the thermodynamic limit, Phys. Rev. B 107, L121105 (2023).DOI
  25. [25]S. Azadi, N. Drummond, and S. M. Vinko, Correlation energy of the spin-polarized electron liquid studied using quantum monte carlo simulations, Phys. Rev. B 108, 115134 (2023).DOI
  26. [26]R. J. Needs, M. D. Towler, N. D. Drummond, P. López Ríos, and J. R. Trail, Variational and diffusion quantum Monte Carlo calculations with the CASINO code, J. Chem. Phys. 152, 154106 (2020).
  27. [27]See supplemental material, which provides the full vmc, dmc, and hf numerical data at the good-γ particle numbers underlying the thermodynamic-limit extrapolation, together with a direct comparison against generic (“bad-γ”) particle numbers presented to isolate the effect of the one-body finite-size error on the correlated (dmc) energy (2026).
  28. [28]S. Azadi, Github.

Paper details

vlnc:W_q6qahmba
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

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

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents