Stochastic GW (sGW) Overview

• Stochastic GW (sGW) is a reformulation of the GW many-body perturbation method that replaces explicit orbital sums with stochastic sampling for evaluating Green’s functions and screening.
• It employs stochastic time-dependent Hartree propagation and matrix compression to achieve near-linear scaling in quasiparticle energy calculations.
• The method accurately predicts ionization energies for molecules and nanostructures and extends to fully spin-polarized and non-collinear systems.

Stochastic GW (sGW) is a reformulation of the GW many-body perturbation correction to density functional theory in which occupied and virtual Kohn–Sham orbitals are replaced by stochastic orbitals for the evaluation of the Green function, the polarization potential, and the GW self-energy. In its original real-space, time-domain formulation, sGW combines stochastic time-dependent Hartree propagation, stochastic matrix compression, and spatial/temporal stochastic decoupling to enable linear-scaling or near-linear-scaling quasiparticle calculations while circumventing the need for energy cutoff approximations (Neuhauser et al., 2014). Subsequent work established its accuracy for molecular ionization energies and extended the framework to fully spin-polarized collinear and non-collinear spinor systems (Vlcek et al., 2016). In unrelated cosmological usage, closely similar abbreviations can denote stochastic gravitational waves or the stochastic gravitational-wave background (Xie et al., 2024).

1. Definition, scope, and development

The GW approximation evaluates the electronic self-energy as the product of the one-particle Green’s function and the dynamically screened Coulomb interaction,

$\Sigma(1,2) = i\,G(1,2)\,W(1^+,2),$

and is widely used to compute quasiparticle excitation energies beyond Kohn–Sham DFT. Conventional deterministic $G_0W_0$ implementations require explicit occupied and large virtual manifolds, large dielectric or polarizability objects, and repeated frequency-dependent matrix operations, which in practice lead to steep computational scaling. sGW replaces those explicit orbital sums and large matrix constructions by stochastic sampling on a real-space grid, preserving the GW formalism while changing the numerical representation of $G_0$, $\chi$, $W$, and $\Sigma$ (Neuhauser et al., 2014).

The 2014 formulation was presented as a method to calculate quasiparticle energies “within the GW many-body perturbation correction to the density functional theory (DFT),” with the formalism illustrated for silicon nanocrystals of varying sizes with over 3000 electrons (Neuhauser et al., 2014). A 2016 study focused on molecules and showed that the method is very accurate for vertical ionization energies across a benchmark set of ten systems, thereby validating the statistical formulation against deterministic GW codes (Vlcek et al., 2016). A 2025 extension generalized sGW to fully spin-polarized systems, including both collinear and non-collinear spin configurations, so that large-scale magnetic and spin-orbit-coupled materials can be treated within a unified stochastic framework (Jiang et al., 18 Sep 2025).

2. Placement within the GW quasiparticle formalism

In the stochastic formulation, the target quantity remains the quasiparticle correction to the Kohn–Sham spectrum. One starts from a Kohn–Sham Hamiltonian $\hat h_{KS}$ and the usual GW partition of the self-energy into exchange and polarization parts. In the 2014 formalism, the polarization self-energy is written as

$$\Sigma^{P}(\mathbf{r}_1,\mathbf{r}_2,t;\varepsilon) = i\hbar\, G_{0}(\mathbf{r}_1,\mathbf{r}_2,t)\, W^{P}(\mathbf{r}_1,\mathbf{r}_2,t;\varepsilon),$$

with

$$W^{P}(\mathbf{r}_1,\mathbf{r}_2,t;\varepsilon) \equiv \langle \mathbf{r}_1 | u_C \otimes \chi(t;\varepsilon) \otimes u_C | \mathbf{r}_2 \rangle,$$

and the quasiparticle energy is represented as

$$\varepsilon_{QP}(\varepsilon) = \varepsilon + \tilde{\Sigma}^{P}(\omega_{QP};\varepsilon) + \Sigma^{X}(\varepsilon) - \Sigma^{XC}(\varepsilon).$$

Here $G_0W_0$0 is the Fourier transform of the time-domain polarization self-energy, $G_0W_0$1 is the exact-exchange self-energy, and $G_0W_0$2 is the Kohn–Sham exchange–correlation contribution (Neuhauser et al., 2014).

The molecular formulation is explicitly framed as a reformulation of standard $G_0W_0$3. In that presentation, the quasiparticle energies $G_0W_0$4 are obtained from

$G_0W_0$5

where

$G_0W_0$6

Thus, sGW does not alter the physical content of GW; it alters the numerical evaluation of the same quasiparticle self-energy objects (Vlcek et al., 2016).

3. Stochastic representation of $G_0W_0$7, screening, and self-energy

The central device of sGW is a stochastic resolution of the identity on a real-space grid. A stochastic orbital is defined by assigning to each grid point a random value $G_0W_0$8 or $G_0W_0$9 with equal probability, so that

$G_0$0

This identity property converts traces and two-point kernels into averages over products of one-point stochastic fields. Energy filtering is introduced by

$G_0$1

with $G_0$2 a Gaussian filter implemented by Chebyshev expansion (Neuhauser et al., 2014).

The non-interacting Green’s function is then written stochastically as

$G_0$3

where $G_0$4 is a projected, time-propagated stochastic orbital. This converts the full two-point Green’s function into an average over separable products and is one aspect of what the original paper calls stochastic matrix compression (Neuhauser et al., 2014).

The polarization self-energy is likewise recast as a stochastic average. Starting from a trace form for $G_0$5, the formalism introduces additional stochastic orbitals and a spatial/temporal stochastic decoupling,

$G_0$6

This factorization separates the time dependence carried by $G_0$7 from the action of the screened interaction and is one of the defining features of the original formulation (Neuhauser et al., 2014).

Screening is obtained through stochastic time-dependent Hartree propagation, and in later formulations through stochastic time-dependent Hartree or TDDFT propagation. Instead of propagating all occupied orbitals, one introduces occupied projected stochastic orbitals and computes the density as

$G_0$8

The density response to an impulsive perturbation yields the retarded response function and therefore the screened interaction $G_0$9, without ever forming a dielectric matrix explicitly (Neuhauser et al., 2014). In the molecular implementation, the same logic appears in time-domain form: the screened interaction is reconstructed from the density response $\chi$0 generated by stochastic occupied orbitals propagated under TDH/RPA, and the self-energy is assembled from stochastic products of Green’s-function and screening factors (Vlcek et al., 2016).

4. Algorithmic workflow and scaling

An sGW calculation proceeds from a Kohn–Sham Hamiltonian on a real-space grid, often obtained from a stochastic DFT or grid-based KS calculation, and selects an energy window around the target quasiparticle state. Filtered stochastic orbitals are constructed, propagated in time to obtain $\chi$1, and coupled to a stochastic TDH or TDDFT response calculation that supplies the screened interaction. The polarization self-energy is accumulated in the time domain, Fourier transformed, and combined with exchange and exchange–correlation terms to solve the quasiparticle equation (Neuhauser et al., 2014).

What distinguishes sGW computationally is the removal of explicit occupied and virtual orbital summations and the avoidance of large matrix storage. The 2014 work states that the method “enables linear scaling GW calculations breaking the theoretical scaling limit for GW as well as circumventing the need for energy cutoff approximations” (Neuhauser et al., 2014). In the nanocrystal calculations summarized there, CPU time for exchange, exchange–correlation, and polarization self-energy scales nearly linearly with the number of electrons, while storage scales linearly with the number of grid points because $\chi$2, $\chi$3, $\chi$4, and $\chi$5 are never stored as full two-point matrices (Neuhauser et al., 2014).

The molecular implementation makes the same point in the language of $\chi$6: explicit sums over unoccupied states, explicit dielectric matrices, and high-order deterministic scaling are replaced by Monte Carlo sampling with random orbitals, implemented in real time on a real-space grid (Vlcek et al., 2016). Statistical errors decrease as $\chi$7, where $\chi$8 is the number of stochastic iterations, so convergence is controlled by sampling rather than by basis-set completeness in the virtual manifold (Neuhauser et al., 2014). The method is therefore especially attractive for extended and open-boundary systems, where deterministic GW becomes prohibitive because of virtual-state counts and dielectric-matrix size (Vlcek et al., 2016).

5. Benchmarks, accuracy, and application range

The first large-scale demonstrations targeted hydrogen-passivated silicon nanocrystals ranging from $\chi$9 to $W$0, reaching 3120 electrons. The reported quasiparticle gaps decrease with nanocrystal size, consistent with quantum confinement, and the sGW results are in very good agreement with $W$1SCF calculations and literature GW values. For almost the entire range of nanocrystal sizes considered, the original report states that sGW was cheaper than sDFT itself (Neuhauser et al., 2014).

Accuracy was examined systematically for molecules in “Stochastic GW calculations for molecules” (Vlcek et al., 2016). Using a set of 10 molecules, the study reports that sGW provides reliable vertical ionization energies in close agreement with benchmark deterministic GW results, with mean deviation $W$2 eV and mean absolute deviation $W$3 eV (Vlcek et al., 2016). The work frames this as an explicit validation of the earlier near-linear-scaling formalism, showing that the stochastic approximation does not merely improve scaling but also preserves predictive accuracy at the level relevant to GW benchmarking.

The method was also applied to PCBM in the 2014 study. For the hole state, sGW yields $W$4 eV, compared to the experimental ionization potential $W$5 eV. For the electron state, the calculation gives $W$6 eV versus experimental $W$7 eV with RPA screening, improving to $W$8 eV when TDDFT screening is used (Neuhauser et al., 2014). These numbers illustrate a recurring theme in sGW applications: the stochastic machinery addresses the scaling bottleneck, while the physical accuracy remains tied to the underlying GW level and the quality of the screening approximation.

6. Spin-dependent and non-collinear extensions

The 2025 extension generalizes sGW to fully spin-polarized systems, including both collinear and non-collinear spin configurations (Jiang et al., 18 Sep 2025). In collinear systems, the Hamiltonian remains block-diagonal in spin, whereas non-collinear systems require complex two-component spinors and include spin mixing from spin–orbit coupling. The extension keeps the screened interaction spin-diagonal in the usual RPA sense but allows $W$9 and $\Sigma$0 to acquire spin off-diagonal structure through the spinor Kohn–Sham basis (Jiang et al., 18 Sep 2025).

A technical issue in the non-collinear case is that the source charge associated with a stochastic orbital and a complex spinor state is itself complex. The 2025 work resolves this by introducing a complex-valued stochastic basis that preserves the real-valued external stochastic charge applied at time zero. This permits an unbiased evaluation of the random-phase approximation screened interaction for spinors without duplicating the TDH evolution for separate real and imaginary sources (Jiang et al., 18 Sep 2025).

The reported performance is explicitly quantified. Error analysis and benchmarks on real materials show that collinear sGW retains the same time complexity as the spin-unpolarized method, while non-collinear sGW incurs a computational cost two to three times higher than the spin-unpolarized version, while preserving linear scaling with low multiplicity (Jiang et al., 18 Sep 2025). The paper validates the extension on systems including monolayer $\Sigma$1, $\Sigma$2, $\Sigma$3, and bulk AlSb, with quasiparticle gaps and relativistic corrections consistent with deterministic spinor GW benchmarks (Jiang et al., 18 Sep 2025).

In this expanded form, sGW is best understood as a statistically controlled, real-space implementation of $\Sigma$4 in which the dominant cost is shifted from virtual-state enumeration and dielectric-matrix algebra to stochastic sampling and time propagation. Its main advantages—near-linear or linear scaling, low memory footprint, and the elimination of explicit energy cutoffs for unoccupied states—have been demonstrated for molecules, nanostructures, and spinor materials alike, while its limitations remain those of the underlying GW level and the chosen screening model (Neuhauser et al., 2014).