betaFeedback

WorksMathematics

Dynamical low-rank methods for the Wigner equation I: separable difference potential

Abstract

Recent advances in dynamical low-rank approximation (DLRA) have demonstrated its effectiveness in high-dimensional simulations. However, existing DLRA algorithms still face significant challenges when handling systems that involve complex collision terms, including the pseudo-differential operator (Ψ{\rm Ψ}) in the Wigner equation, a representative operator characterized by nonlocality. It is deserving to carry out a series of works to develop the DLRA algorithms for solving the Wigner equation. As the first step in this series of works, we propose an efficient DLRA algorithm for the Wigner equation, using a separable decomposition of the difference potential. We combine this separable assumption with two often-used truncations of Ψ{\rm Ψ}, namely K\mathcal{K}-truncation and Y\mathcal{Y}-truncation, to obtain a kind of separated representation of Ψ{\rm Ψ}. Complexity analysis and several challenging experiments, including harmonic oscillators, Gaussian barrier scattering, electron-electron scattering, and a Helium-like system, all of which satisfy the separable assumption, confirm that the proposed DLRA algorithm has significant advantages, achieving a reduction in computational effort by one to two orders of magnitude in both runtime and memory requirements compared to the full-grid approach. It is worth noting that, even in the absence of a predetermined low-rank structure for the solution, DLRA can still serve as a numerical scheme that balances efficiency and accuracy.

Funding. This research was supported by the National Natural Science Foundation of China (Nos. 12325112, 12288101) and the High-performance Computing Platform of Peking University.

Introduction

Since its introduction in 1932 [44], the Wigner equation provides a convenient way to characterize fundamental phase-space formulation of quantum mechanics, analogous to the Boltzmann equation in classical kinetic theory. Although the Wigner function may exhibit negative values, this quasi-distribution function remains an important way to understand and forecast quantum phenomena in many fields, such as nanoelectronics [43, 20], high energy density physics [21], quantum optics [38], quantum transport [28, 27], etc.

The Wigner equation

The Wigner equation with a time-independent potential considered in this paper is

∂∂tf(x,k,t)=−ℏkm⋅∇xf(x,k,t)+ΘV[f](x,k,t):=Lf,(1)\frac{\partial}{\partial t}f(x,k,t)=-\frac{\hbar k}{m}\cdot\nabla_x f(x,k,t)+\Theta_V[f](x,k,t):=\mathcal{L}f, \tag*{(1)}

where x∈Ωx⊆Rdx\in\Omega_x\subseteq\mathbb{R}^d is the spatial variable, k∈Ωk⊆Rdk\in\Omega_k\subseteq\mathbb{R}^d is the Fourier conjugated wave vector, dd is the dimension, mm is the mass, ℏ\hbar is the reduced Planck constant and the nonlocal pseudodifferential operator (Ψ\PsiDO) is defined as

ΘV[f](x,k,t)=∫Rdf(x,k′,t)Vw(x,k−k′)dk′(2)\Theta_V[f](\boldsymbol{x},\boldsymbol{k},t)=\int_{\mathbb{R}^d}f(\boldsymbol{x},\boldsymbol{k}',t)V_w(\boldsymbol{x},\boldsymbol{k}-\boldsymbol{k}')\mathrm{d}\boldsymbol{k}' \tag*{(2)}
Vw(x,k)=1iℏ(2π)d∫Rde−ik⋅yDV(x,y)dy(3)V_w(\boldsymbol{x},\boldsymbol{k})=\frac{1}{i\hbar(2\pi)^d}\int_{\mathbb{R}^d}e^{-i\boldsymbol{k}\cdot\boldsymbol{y}}D_V(\boldsymbol{x},\boldsymbol{y})\mathrm{d}\boldsymbol{y} \tag*{(3)}
DV(x,y)=V(x+y2)−V(x−y2).(4)D_V(\boldsymbol{x},\boldsymbol{y})=V\left(\boldsymbol{x}+\frac{\boldsymbol{y}}{2}\right)-V\left(\boldsymbol{x}-\frac{\boldsymbol{y}}{2}\right). \tag*{(4)}

Here DV(x,y)D_V(\boldsymbol{x},\boldsymbol{y}) is central difference of potential function V(x)V(\boldsymbol{x}), the Wigner kernel Vw(x,k)V_w(\boldsymbol{x},\boldsymbol{k}) is Fourier transformation of DV(x,y)D_V(\boldsymbol{x},\boldsymbol{y}) and ΘV[f](x,k,t)\Theta_V[f](\boldsymbol{x},\boldsymbol{k},t) is the convolution of f(x,k,t)f(\boldsymbol{x},\boldsymbol{k},t) and Vw(x,k)V_w(\boldsymbol{x},\boldsymbol{k}). We can obtain another form of Ψ\PsiDO by changing the order of integration

ΘV[f](x,k,t)=1iℏ∫Rde−ik⋅yDV(x,y)ρ(x,y,t)dy,(5)\Theta_V[f](\boldsymbol{x},\boldsymbol{k},t)=\frac{1}{i\hbar}\int_{\mathbb{R}^d}e^{-i\boldsymbol{k}\cdot\boldsymbol{y}}D_V(\boldsymbol{x},\boldsymbol{y})\rho(\boldsymbol{x},\boldsymbol{y},t)\mathrm{d}\boldsymbol{y}, \tag*{(5)}
ρ(x,y,t)=1(2π)d∫Rdf(x,k′,t)eik′⋅ydk′.(6)\rho(\boldsymbol{x},\boldsymbol{y},t)=\frac{1}{(2\pi)^d}\int_{\mathbb{R}^d}f(\boldsymbol{x},\boldsymbol{k}',t)e^{i\boldsymbol{k}'\cdot\boldsymbol{y}}\mathrm{d}\boldsymbol{k}'. \tag*{(6)}

The Ψ\PsiDO is the most complex and important component in the Wigner equation, representing the fundamental distinction between the classical Vlasov equation and the Wigner equation. This equation evolves in a 2d2d-dimensional phase space, thus confronting the severe challenge of the curse of dimensionality (CoD) along with the nonlocality.

Despite the long history and broad application of the Wigner equation, its numerical solution continues to present significant challenges nowadays. Over the past three decades, significant progress has been made in developing Wigner equation solvers, which can be roughly classified into particle-based stochastic and grid-based deterministic approaches. The particle-based stochastic approaches such as Monte Carlo method [32, 39] and stochastic particle method [42] typically achieve only O(N−1/2)O(N^{-1/2}) convergence due to their inherent statistical nature, where NN is the number of samples. In contrast, the grid-based deterministic approaches exhibit superior convergence rates because of their concise mathematical principle, including the finite difference method [19], operator splitting method [37, 2], spectral element method (SEM) [40, 45, 7, 46], moment method [34], weighted essentially non-oscillatory (WENO) method [13], etc. While CoD confined deterministic methods to low-dimensional problems, rapid advances in high-performance computing, alongside techniques such as tensor approximations that mitigate CoD, have enabled deterministic methods to tackle 5D and 6D problems. The dynamical low-rank approximation (DLRA) method we consider next serves as one such approach for mitigating CoD.

DLRA: the projector-splitting integrator

Emerging from the Dirac-Frenkel time-dependent variational principle [12], the DLRA method [30] has evolved into a powerful class of numerical integrators. This method has been widely used in kinetic theory, including the Vlasov equation [18, 17], Boltzmann equation [16, 24] and radiation transfer equation [36, 15]. The main idea is to constrain the unknown function ff to a low-rank manifold, then deriving evolution equations of basis functions through tangent space projection. This approach reduces the original 2d2d-dimensional problem to several dd-dimensional subproblems, effectively alleviating CoD in traditional grid-based methods.

The DLRA method begins by assuming that the Wigner function f(x,k,t)f(\boldsymbol{x},\boldsymbol{k},t) admits a separable low-rank representation of the following form

f(x,k,t)≈∑i,j=1rXi(x,t)Kj(k,t)Sij(t),(7)f(x,k,t) \approx\sum_{i,j=1}^{r} X_i(x,t)K_j(k,t)S_{ij}(t), \tag*{(7)}

where rr is the rank, the coefficient matrix S=(Sij)S=(S_{ij}) is invertible, the basis function Xi(x,t)∈L2(Ωx)X_i(x,t)\in L^2(\Omega_x) and Kj(k,t)∈L2(Ωk)K_j(k,t)\in L^2(\Omega_k) are orthonormal

⟨Xi,Xj⟩x=δij,⟨Ki,Kj⟩k=δij,(8)\langle X_i,X_j\rangle_x=\delta_{ij}, \qquad\langle K_i,K_j\rangle_k=\delta_{ij}, \tag*{(8)}

with ⟨⋅,⋅⟩x\langle\cdot,\cdot\rangle_x and ⟨⋅,⋅⟩k\langle\cdot,\cdot\rangle_k being the standard L2L^2 inner products on L2(Ωx)L^2(\Omega_x) and L2(Ωk)L^2(\Omega_k), respectively, and δij\delta_{ij} denotes the Kronecker delta. Therefore, for fixed time tt, the Wigner function f(x,k,t)f(x,k,t) lies in the fixed rank-rr manifold

Mr={g∈L2(Ω)∣g(x,k)=∑i,j=1rXi(x)SijKj(k) with invertible S=(Sij),\mathcal{M}_r=\left\{g\in L^2(\Omega)\mid g(x,k)=\sum_{i,j=1}^{r}X_i(x)S_{ij}K_j(k)\text{ with invertible }S=(S_{ij}), \right.
Xi∈L2(Ωx),Kj∈L2(Ωk) with ⟨Xi,Xj⟩x=δij,⟨Ki,Kj⟩k=δij}.(9)\left. X_i\in L^2(\Omega_x),K_j\in L^2(\Omega_k)\text{ with }\langle X_i,X_j\rangle_x=\delta_{ij},\langle K_i,K_j\rangle_k=\delta_{ij}\right\}. \tag*{(9)}

Then project Lf\mathcal{L}f in (1.1) onto the tangent space of Mr\mathcal{M}_r, yielding the reduced system

∂∂tf(x,k,t)=Pf(Lf),(10)\frac{\partial}{\partial t}f(x,k,t)=\mathcal{P}_f(\mathcal{L}f), \tag*{(10)}

where the projection operator Pf\mathcal{P}_f acting on a generic function gg can be written as [30]

Pfg=PK‾g−PK‾PX‾g+PX‾g.(11)\mathcal{P}_f g=\mathcal{P}_{\overline{K}}g-\mathcal{P}_{\overline{K}}\mathcal{P}_{\overline{X}}g+\mathcal{P}_{\overline{X}}g. \tag*{(11)}

Here, K‾\overline{K} and X‾\overline{X} denote the vector spaces spanned by the basis functions KjK_j and XiX_i,

K‾=span⁡{Kj∣j=1,2,⋯ ,r},X‾=span⁡{Xi∣i=1,2,⋯ ,r},(12)\overline{K}=\operatorname{span}\{K_j\mid j=1,2,\cdots,r\},\qquad\overline{X}=\operatorname{span}\{X_i\mid i=1,2,\cdots,r\}, \tag*{(12)}

and PK‾,PX‾\mathcal{P}_{\overline{K}},\mathcal{P}_{\overline{X}} are the orthogonal projectors onto X‾\overline{X} and K‾\overline{K}, respectively,

PK‾g=∑j=1r⟨Kj,g⟩kKj,PX‾g=∑i=1rXi⟨Xi,g⟩x.(13)\mathcal{P}_{\overline{K}}g=\sum_{j=1}^{r}\langle K_j,g\rangle_k K_j,\qquad\mathcal{P}_{\overline{X}}g=\sum_{i=1}^{r}X_i\langle X_i,g\rangle_x. \tag*{(13)}

To avoid the possible ill-conditioning of the coefficient matrix SS [30, 29], the projector-splitting integrator [35] substitutes (1.11) into the reduced Wigner system (1.10) and decomposes (1.10) into three subproblems:

∂∂tf(x,k,t)=PK‾(Lf),(14)\frac{\partial}{\partial t}f(x,k,t)=\mathcal{P}_{\overline{K}}(\mathcal{L}f), \tag*{(14)}
∂∂tf(x,k,t)=−PK‾PX‾(Lf),(15)\frac{\partial}{\partial t}f(x,k,t)=-\mathcal{P}_{\overline{K}}\mathcal{P}_{\overline{X}}(\mathcal{L}f), \tag*{(15)}
∂∂tf(x,k,t)=PX‾(Lf).(16)\frac{\partial}{\partial t}f(x,k,t)=\mathcal{P}_{\overline{X}}(\mathcal{L}f). \tag*{(16)}

The decomposition of the projection operator Pf\mathcal{P}_f forms the basis of projector-splitting integrator. Exploiting orthogonality conditions, we can further simplify each subproblems and proceed in the following three substeps, respectively: Z-step. We can represent the low-rank assumption for ff in (1.7) as f(x,k,t)=∑j=1rZj(x,t)Kj(k,t)f(x,k,t)=\sum_{j=1}^{r}Z_j(x,t)K_j(k,t) with Zj(x,t)=∑i=1rXi(x,t)Sij(t)Z_j(x,t)=\sum_{i=1}^{r}X_i(x,t)S_{ij}(t), then rewrite Eq. (14) as

∑j=1r(∂∂tZjKj+Zj∂∂tKj)=∑j=1r⟨Kj,Lf⟩kKj.\sum_{j=1}^{r}\left(\frac{\partial}{\partial t}Z_jK_j+Z_j\frac{\partial}{\partial t}K_j\right)=\sum_{j=1}^{r}\langle K_j,\mathcal{L}f\rangle_kK_j.

Making an important observation that {Kj(k,t)}1≤j≤r\{K_j(k,t)\}_{1\leq j\leq r} are independent of time tt (see [35 Lemma 3.1]) in this substep, and using orthogonality conditions ⟨Ki,Kj⟩k=δij\langle K_i,K_j\rangle_k=\delta_{ij}, we derive the following differential equation for Zj(x,t)Z_j(x,t) as

∂∂tZj(x,t)=⟨Kj,Lf⟩k,j=1,…,r.(17)\frac{\partial}{\partial t}Z_j(x,t)=\langle K_j,\mathcal{L}f\rangle_k,\qquad j=1,\ldots,r. \tag*{(17)}

S-step. We can argue similarly to obtain that Eq. (15) is equivalent to

∂∂tSij(t)=⟨XiKj,Lf⟩x,k,i,j=1,…,r.(18)\frac{\partial}{\partial t}S_{ij}(t)=\langle X_iK_j,\mathcal{L}f\rangle_{x,k},\qquad i,j=1,\ldots,r. \tag*{(18)}

During this substep, both {Kj(k,t)}1≤j≤r\{K_j(k,t)\}_{1\leq j\leq r} and {Xi(x,t)}1≤i≤r\{X_i(x,t)\}_{1\leq i\leq r} remain unchanged.

L-step. Similarly, we can represent (1.7) as f(x,k,t)=∑i=1rXi(x,t)Li(k,t)f(x,k,t)=\sum_{i=1}^{r}X_i(x,t)L_i(k,t) with Li(k,t)=∑j=1rSij(t)Kj(k,t)L_i(k,t)=\sum_{j=1}^{r}S_{ij}(t)K_j(k,t). By similar arguments, Eq. (16) is equivalent to

∂∂tLi(k,t)=⟨Xi,Lf⟩x,i=1,…,r.(19)\frac{\partial}{\partial t}L_i(k,t)=\langle X_i,\mathcal{L}f\rangle_x,\qquad i=1,\ldots,r. \tag*{(19)}

During this substep, {Xi(x,t)}1≤i≤r\{X_i(x,t)\}_{1\leq i\leq r} remains unchanged.

This projector-splitting integrator has been shown to be robust to small singular values of SS and own first order rate [29]. For more details, we refer the reader to [18] which provides similar approach for the Vlasov-Poisson equation.

The main challenge

While (1.14)–(1.16) appear to offer a framework for applying DLRA, the efficiency relies on the separability of the right-hand side Lf\mathcal{L}f. The vast majority of problems solved using DLRA possess separability in their right-hand side terms, such as the Vlasov equation [18, 17] and the radiative transfer equation [36, 15]. Some studies utilize the steady-state properties to simplify the problem, for instance, the Maxwellian distribution is employed to reduce complexity when solving the Boltzmann-BGK equation [16]. However, this approach is not well suited for solving non-steady-state problems.

The so-called separability here refers to the fact that, under the low-rank assumption (1.7) satisfied by ff, Lf\mathcal{L}f can also be naturally expressed in a form that separates the variables xx and kk. Separability allows the projection to be computed directly via the low-rank components XiX_i, SijS_{ij}, KjK_j, without the full tensor ff. For example, the advection term ℏkm⋅∇xf\frac{\hbar k}{m}\cdot\nabla_xf possesses separability, since

ℏkm⋅∇xf(x,k,t)=ℏm∑i′,j′=1r∇xXi′(x,t)⋅(kKj′(k,t))Si′j′(t),\frac{\hbar k}{m}\cdot\nabla_xf(x,k,t)=\frac{\hbar}{m}\sum_{i',j'=1}^{r}\nabla_xX_{i'}(x,t)\cdot\left(kK_{j'}(k,t)\right)S_{i'j'}(t),

and then the projection onto the {Kj(k,t)}1≤j≤r\{K_j(k,t)\}_{1\leq j\leq r} becomes

⟨Kj,ℏkm⋅∇xf(x,k,t)⟩k=ℏm∑i′,j′=1r∇xXi′(x,t)⋅⟨Kj,kKj′(k,t)⟩kSi′j′(t),\left\langle K_j,\frac{\hbar k}{m}\cdot\nabla_xf(x,k,t)\right\rangle_k=\frac{\hbar}{m}\sum_{i',j'=1}^{r}\nabla_xX_{i'}(x,t)\cdot\langle K_j,kK_{j'}(k,t)\rangle_kS_{i'j'}(t),

where xx and kk dependent terms can be computed separately. In contrast, the intrinsic nonlocality of Ψ\PsiDO introduces strong coupling between the variables xx and kk, preventing it from possessing the separability. Consequently, evaluating its projection would typically necessitate reconstructing the full tensor, which undermines the efficiency aims of the low-rank approach. This strong coupling within Ψ\PsiDO represents the main difficulty to applying DLRA to the Wigner equation.

Organization

The rest of this paper is organized as follows. Section 2 introduces a separable assumption for the difference potential DV(x,y)D_V(\mathbf{x},\mathbf{y}) as shown in (22). By introducing this separable assumption, Ψ\PsiDO can be transformed into a separated form through KK-truncation and yy-truncation. The strong coupling between xx and kk in Ψ\PsiDO can be transferred to the separation rank RR. Regarding the applicability of this assumption, it effectively covers scenarios where the potential V(x)V(x) is real analytic, including typical potentials such as the harmonic oscillator, Gaussian barrier, and soft Coulomb discussed in subsequent numerical examples.

In Section 3, we derive the evolution equations for basis functions from the Wigner equation, detail their numerical implementation and analyze the overall complexity of the algorithm. In Section 4, we validate the efficiency of our algorithm through a series of typical numerical examples. We first choose several non-stationary and atypical low-rank numerical experiments such as harmonic oscillator and Gaussian barrier scattering. Our numerical experiments indicate that the DLRA method can still yield numerical solutions of sufficient accuracy, even when the solution lacks an obvious low-rank structure. We further validate the method through more complex numerical examples, including electron-electron scattering and Helium-like systems, successfully capturing fundamental quantum phenomena such as the Pauli exclusion principle and the Heisenberg uncertainty principle [22]. The numerical results confirm that the proposed algorithm can achieve a reduction in computational effort by one to two orders of magnitude in both runtime and memory requirements compared to the full-grid approach. These results clearly demonstrate the broad applicability and promising potential of the DLRA method in quantum dynamical simulations. Section 5 makes a summary and discusses future research directions.

Separable assumption

The intrinsic nonlocality of Ψ\PsiDO in (2) places the first obstacle for any numerical solver and some kinds of truncation have been introduced in the literature. If the potential V(x)V(x) admits a Taylor expansion in xx-space, then the difference potential DV(x,y)D_V(x,y) can be written as

DV(x,y)=∑l=0+∞2(2l+1)!∇x(2l+1)V(x)(y2)2l+1.(20)D_V(x,y)=\sum_{l=0}^{+\infty}\frac{2}{(2l+1)!}\nabla_x^{(2l+1)}V(x)\left(\frac{y}{2}\right)^{2l+1}. \tag*{(20)}

Substituting (20) into (3) and using the basic properties of Fourier transform yields directly the so-called Moyal expansion of Ψ\PsiDO

ΘVM[f](x,k,t)=∑l=0+∞(−1)l(2l+1)!(ℏ2)2l∇x2l+1V⋅∇p2l+1f,(21)\Theta_V^M[f](x,k,t)=\sum_{l=0}^{+\infty}\frac{(-1)^l}{(2l+1)!}\left(\frac{\hbar}{2}\right)^{2l}\nabla_x^{2l+1}V\cdot\nabla_p^{2l+1}f, \tag*{(21)}

where p=ℏkp=\hbar k is the momentum. In particular, the first term in the summation on the right-hand side is ∇xV⋅∇pf\nabla_x V\cdot\nabla_p f, indicating that the Wigner equation reduces to the classical Vlasov equation when ℏ\hbar tends to zero. In the early stage, one has to truncate the infinite series summation (21) to a finite number of terms [25, 26]. Obviously, the resulting scheme possesses certain separability and the primary reason is that in Eq. (2.1) for the difference potential DV(x,y)D_V(\mathbf{x},\mathbf{y}), the variables x\mathbf{x} and y\mathbf{y} are separated. Based on this observation, we introduce the separable assumption for the difference potential DV(x,y)D_V(\mathbf{x},\mathbf{y})

DV(x,y)≈∑s=1RDX,s(x)DY,s(y),(22)D_V(\mathbf{x},\mathbf{y}) \approx\sum_{s=1}^{R}D_{X,s}(\mathbf{x})D_{Y,s}(\mathbf{y}), \tag*{(22)}

where RR denotes the separation rank, and DX,s,DY,sD_{X,s},D_{Y,s} are functions that depend only on x,y\mathbf{x},\mathbf{y}, respectively.

Nowadays, the following two kinds of truncation, named after the truncation of the outermost infinite integral in Ψ\PsiDO, are often used [7, 8, 9, 45]:

  • Y\mathcal{Y}-truncation: starting from (1.5) and (1.6), computations are performed via the intermediate variable ρ\boldsymbol{\rho};

  • K\mathcal{K}-truncation: starting from (1.2) and (1.3), computations are performed via the intermediate variable VwV_w;

After applying the separable difference potential assumption (2.3), we have a separated approximation for Ψ\PsiDO

ΘV[f](x,k,t)=∑s=1R[∑i,j=1rX‾i,s(x,t)Sij(t)K‾j,s(k,t)],(23)\Theta_V[f](\mathbf{x},\mathbf{k},t)=\sum_{s=1}^{R}\left[\sum_{i,j=1}^{r}\overline{X}_{i,s}(\mathbf{x},t)S_{ij}(t)\overline{K}_{j,s}(\mathbf{k},t)\right], \tag*{(23)}
X‾i,s(x,t)=DX,s(x)Xi(x,t),(24)\overline{X}_{i,s}(\mathbf{x},t)=D_{X,s}(\mathbf{x})X_i(\mathbf{x},t), \tag*{(24)}

and the resulting K‾j,s(k,t)\overline{K}_{j,s}(\mathbf{k},t) are listed in Table 1.

At the continuous level, two approaches yield identical results and no truncation is required. At the discrete level, two approaches require truncation of the infinite integral and exhibit almost no difference in results. However, their applicable situations differ. The Y\mathcal{Y}-truncation requires the low-rank components KjK_j to have a Fourier transform. This is typically guaranteed by the rapid decay of KjK_j in the k\mathbf{k}-space, which is in turn induced by the rapid decay of the function ff in the k\mathbf{k}-space. In contrast, the K\mathcal{K}-truncation requires the function DY,s(y)D_{Y,s}(\mathbf{y}) to have a Fourier transform, which is not always guaranteed (for example, DY,1(y)=yD_{Y,1}(\mathbf{y})=\mathbf{y} in the harmonic oscillator).

Dynamical low-rank approximation for the Wigner equation

In this section, we substitute (2.4) into three substep equations (1.17)-(1.19) to derive more precise forms of each part, and design their numerical scheme.

Z-step. Plugging (2.4) into (17) to obtain

∂∂tZj(x,t)=∑s=1R∑j′=1rDX,s(x)cj,j′,sZj′(x,t)−ℏm∑j′=1raj,j′⋅∇xZj′(x,t).(25)\frac{\partial}{\partial t}Z_j(x,t)=\sum_{s=1}^{R}\sum_{j'=1}^{r}D_{X,s}(x)c_{j,j',s}Z_{j'}(x,t)-\frac{\hbar}{m}\sum_{j'=1}^{r}a_{j,j'}\cdot\nabla_x Z_{j'}(x,t). \tag*{(25)}

Here we define time-independent coefficients

Au∈Rr×r,(Au)i,j=(ai,j)u=⟨kuKi,Kj⟩k,(u=1,…,d),(26)A_u\in\mathbb{R}^{r\times r},\qquad(A_u)_{i,j}=(a_{i,j})_u=\langle k_uK_i,K_j\rangle_k,\qquad(u=1,\ldots,d), \tag*{(26)}
Cs∈Rr×r,(Cs)i,j=ci,j,s=⟨Ki,Kj,s‾⟩k,(s=1,…,R).(27)C_s\in\mathbb{R}^{r\times r},\qquad(C_s)_{i,j}=c_{i,j,s}=\langle K_i,\overline{K_{j,s}}\rangle_k,\qquad(s=1,\ldots,R). \tag*{(27)}

In this ZZ-step, the coefficients aj,j′a_{j,j'} and cj,j′,sc_{j,j',s} are computed first, and equation (25) is then solved until the next time step. Subsequently, a standard QR decomposition is applied to ZjZ_j to obtain new orthonormal XiX_i and updated SijS_{ij} in the end. This completes the first ZZ-step.

S-step. Plugging (2.4) into (18) to obtain

∂∂tSij(t)=ℏm∑i′,j′=1raj,j′⋅bi,i′Si′j′(t)−∑s=1R∑i′,j′=1rcj,j′,sdi,i′,sSi′j′(t).(28)\frac{\partial}{\partial t}S_{ij}(t)=\frac{\hbar}{m}\sum_{i',j'=1}^{r}a_{j,j'}\cdot b_{i,i'}S_{i'j'}(t)-\sum_{s=1}^{R}\sum_{i',j'=1}^{r}c_{j,j',s}d_{i,i',s}S_{i'j'}(t). \tag*{(28)}

Here we can use time-independent coefficients aj,j′a_{j,j'} and cj,j′,sc_{j,j',s} defined in (26) and (27) combined with

Bu∈Rr×r,(Bu)i,j=(bi,j)u=⟨Xi,∇xuXj⟩x,(u=1,…,d),(29)B_u\in\mathbb{R}^{r\times r},\qquad(B_u)_{i,j}=(b_{i,j})_u=\langle X_i,\nabla_{x_u}X_j\rangle_x,\qquad(u=1,\ldots,d), \tag*{(29)}
Ds∈Rr×r,(Ds)i,j=di,j,s=⟨Xi,Xj,s‾⟩x,(s=1,…,R).(30)D_s\in\mathbb{R}^{r\times r},\qquad(D_s)_{i,j}=d_{i,j,s}=\langle X_i,\overline{X_{j,s}}\rangle_x,\qquad(s=1,\ldots,R). \tag*{(30)}

In the SS-step, since the basis functions {Kj(k,t)}1≤j≤r\{K_j(k,t)\}_{1\leq j\leq r} remain unchanged during both the ZZ-step and SS-step, the precomputed values of aj,j′a_{j,j'} and cj,j′,sc_{j,j',s} from the previous ZZ-step can be used directly. The newly introduced coefficients bi,i′b_{i,i'} and di,i′,sd_{i,i',s} should be computed first, the SS-step equation (28) can then be solved to obtain the updated SS. This completes the second SS-step.

L-step. Plugging (2.4) into (19) to obtain

∂∂tLi(k,t)=∑s=1R∑i′,j′=1rdi,i′,sSi′j′(t)Kj′,s‾(k,t)−ℏm∑i′=1rk⋅bi,i′Li′(k,t).(31)\frac{\partial}{\partial t}L_i(k,t)=\sum_{s=1}^{R}\sum_{i',j'=1}^{r}d_{i,i',s}S_{i'j'}(t)\overline{K_{j',s}}(k,t)-\frac{\hbar}{m}\sum_{i'=1}^{r}k\cdot b_{i,i'}L_{i'}(k,t). \tag*{(31)}

with coefficients bi,i′b_{i,i'} and di,i′,sd_{i,i',s} have defined in (3.5) and (3.6). Since the basis functions {Xi(x,t)}1≤i≤r\{X_i(x,t)\}_{1\leq i\leq r} remain unchanged during SS-step, the precomputed values of bi,i′b_{i,i'} and di,i′,sd_{i,i',s} from the previous SS-step can be used directly to solve (31). Then apply standard QR decomposition to LiL_i to get new orthonormal KjK_j and updated SijS_{ij} in the end. This completes the last LL-step.

Remark 3.1. Coefficient matrices AuA_u and DsD_s are symmetric by construction, which ensures hyperbolicity of whole system.

A detailed account of first order accurate DLRA for the Wigner equation is given in Algorithm 3.1. The solutions corresponding to the ZZ-step, the SS-step and the LL-step are combined by a first order Lie splitting scheme in order to recover f(x,k,t)f(x,k,t). The sequence of these three steps is not unique [35], but we only consider the sequence of ZSLZSL in this work.

Algorithm 3.1 A first-order accurate dynamical low-rank algorithm for the Wigner equation

Input: DX,sD_{X,s}, DY,sD_{Y,s}, Xi0X_i^0, Sij0S_{ij}^0, Kj0K_j^0 such that f(x,k,0)=∑i,j=1rXi0(x)Sij0Kj0(k)f(x,k,0)=\sum_{i,j=1}^{r}X_i^0(x)S_{ij}^0K_j^0(k) with time step τ\tau.

Output: Xi1X_i^1, Sij3S_{ij}^3, Kj1K_j^1 such that f(x,k,τ)=∑i,j=1rXi1(x)Sij3Kj1(k)f(x,k,\tau)=\sum_{i,j=1}^{r}X_i^1(x)S_{ij}^3K_j^1(k).

  1. Compute Zi0Z_i^0 using Xi0X_i^0 and Sij0S_{ij}^0.

  2. Compute K‾j,s\overline{K}_{j,s} using DX,sD_{X,s}, DY,sD_{Y,s}, and Kj0K_j^0 according to Table 1.

  3. Compute aj,j′a_{j,j'} and cj,j′,sc_{j,j',s} using Kj0K_j^0 and K‾j,s\overline{K}_{j,s}.

  4. Solve (3.1) with initial value Zi0Z_i^0 until time τ\tau to obtain Zi1Z_i^1.

  5. Perform a QR decomposition of Zi1Z_i^1 to obtain Xi1X_i^1 and Sij1S_{ij}^1.

  6. Compute bi,i′b_{i,i'} and di,i′,sd_{i,i',s} using Xj0X_j^0.

  7. Solve (3.4) with initial value Sij1S_{ij}^1 until time τ\tau to obtain Sij2S_{ij}^2.

  8. Compute Li0L_i^0 using Ki0K_i^0 and Sij2S_{ij}^2.

  9. Solve (3.7) with initial value Li0L_i^0 until time τ\tau to obtain Li1L_i^1.

  10. Perform a QR decomposition of Li1L_i^1 to obtain Ki1K_i^1 and Sij3S_{ij}^3.

We have restricted our analysis to first-order Lie splitting. In principle, extending the method to second-order Strang splitting is straightforward; further implementation details can be found in [18, 4]. Although both first-order Lie splitting and second-order Strang splitting have been observed to achieve their respective convergence orders in practice, existing theoretical analyses can only guarantee first-order accuracy [29]. This limitation arises because the projection operator lacks Lipschitz continuity [30]. To date, this phenomenon remains not fully understood theoretically [5]. The numerical results presented in this work indicate that, even when employing a first-order Lie splitting method, second-order accuracy can still be attained provided that each splitting substep itself achieves second-order accuracy. This numerical observation warrants further theoretical investigation in future work.

Discretization

For the purpose of numerical solution, we first truncate the computational domain to a bounded rectangular region Ω=Ωx×Ωk\Omega=\Omega_x\times\Omega_k, then we consider the uniform grid mesh in xx-space and kk-space, xx-space using NxN_x grid points in every coordinate direction with the spacing Δxu\Delta x_u and kk-space using NkN_k grid points in every coordinate direction with the spacing Δku\Delta k_u. For the dual yy-space, we choose NyN_y grid points in every coordinate direction with the spacing Δyu\Delta y_u.

Once the grid discretization is determined, for a fixed time tt, we can obtain X(t)∈RNxd×rX(t)\in\mathbb{R}^{N_x^d\times r} to be the matrix composed of row vectors collected from XiX_i on the chosen spatial grid. It’s obvious that the resulting matrix entries depend on both the discretization performed and the ordering of the grid points. Similarly, K(t)∈RNkd×rK(t)\in\mathbb{R}^{N_k^d\times r} can be defined in the same way. Then, the low-rank approximation of the Wigner function can be reconstructed at any time from the low-rank factors via the computation

F(t)=X(t)S(t)KT(t)∈RNxd×Nkd.(32)F(t)=X(t)S(t)K^{T}(t)\in\mathbb{R}^{N_x^d\times N_k^d}. \tag*{(32)}

However, due to its prohibitive cost, this reconstruction is unnecessary for our algorithm. We only employ it in combination with Singular Value Decomposition (SVD) [14] to obtain the initial values of the low-rank factors and to reconstruct the solution at selected critical time points. Similarly, the separable assumption (2.3) for the difference potential DVD_V can be computed via SVD, and this computation is required only once during the initialization stage. In each time step, since a uniform grid is employed, K‾j,s\overline{K}_{j,s} can be efficiently computed via the standard Fast Fourier transform (FFT). It is also worth noting that the Poisson summation formula is often used in conjunction with truncation [7, 8, 9, 45], however, it is not required in the implementation of our DLRA algorithm.

Numerical scheme

In this subsection, we provide a detailed description of the procedure for solving (3.1), (3.4), and (3.7).

We first consider the SS-step (3.4). It is an ordinary differential equation and there is no source of stiffness in it. Hence, we perform its time integration by classical explicit fourth order Runge-Kutta scheme RK4.

Next, let us consider the ZZ-step (3.1) and use the vector notation Z=(Z1,…,Zr)Z=(Z_1,\ldots,Z_r), we can rewrite the evolution equation for the ZZ-step (3.1) in matrix formulation as follows:

∂∂tZ(x,t)=∑s=1RDX,s(x)Z(x,t)CsT−ℏm∑u=1d∇xuZ(x,t)AuT.(33)\frac{\partial}{\partial t}Z(x,t)=\sum_{s=1}^{R}D_{X,s}(x)Z(x,t)C_s^T-\frac{\hbar}{m}\sum_{u=1}^{d}\nabla_{x_u}Z(x,t)A_u^T. \tag*{(33)}

For simplicity, we only consider the two-dimensional case here. By further splitting the ZZ-step, we obtain the following two subflows:

∂∂tZ(x,t)=−ℏm∇x1Z(x,t)A1T,(34)\frac{\partial}{\partial t}Z(x,t)=-\frac{\hbar}{m}\nabla_{x_1}Z(x,t)A_1^T, \tag*{(34)}
∂∂tZ(x,t)=−ℏm∇x2Z(x,t)A2T+∑s=1RDX,s(x)Z(x,t)CsT.(35)\frac{\partial}{\partial t}Z(x,t)=-\frac{\hbar}{m}\nabla_{x_2}Z(x,t)A_2^T+\sum_{s=1}^{R}D_{X,s}(x)Z(x,t)C_s^T. \tag*{(35)}

As we mentioned before, the coefficients matrix A1A_1 and A2A_2 are symmetric, thus equations (3.10) and (3.11) are hyperbolic. It is evident that (3.10) can be solved in time by the method of characteristics, while (3.11) can be efficiently solved via an exponential integrator (see [23] for more details).

Finally, we consider the LL-step (3.7). Extending the vector notation similar to ZZ to XX, KK, YY, LL, SS, and K‾s\overline{K}_s, we can obtain

∂∂tL(k,t)=∑s=1RK‾sSTDsT−ℏm∑u=1dkuL(k,t)BuT.(36)\frac{\partial}{\partial t}L(k,t)=\sum_{s=1}^{R}\overline{K}_sS^TD_s^T-\frac{\hbar}{m}\sum_{u=1}^{d}k_uL(k,t)B_u^T. \tag*{(36)}

Unlike the ZZ-step, the hyperbolicity is maintained by the coefficient matrix DsD_s, hence it is necessary to examine the relationship between LL and KsSTK_sS^T and to exploit this connection in solving the LL-step equation.

For example, for Y\mathcal{Y}-truncation, using the relationship provided in Table 1, apply the Fourier transform in kk-direction to (3.12) and define

TruncationK‾j,s(k,t)\overline{K}_{j,s}(\mathbf{k},t)
Y\mathcal{Y}Yj(y,t)=1(2π)d∫Rdeik′⋅yKj(k′,t) dk′,K‾j,s(k,t)=1iℏ∫Rde−ik⋅yYj(y,t)DY,s(y) dy.\begin{aligned}Y_j(\mathbf{y},t)&=\frac{1}{(2\pi)^d}\int_{\mathbb{R}^d}e^{i\mathbf{k}'\cdot\mathbf{y}}K_j(\mathbf{k}',t)\,d\mathbf{k}',\\ \overline{K}_{j,s}(\mathbf{k},t)&=\frac{1}{i\hbar}\int_{\mathbb{R}^d}e^{-i\mathbf{k}\cdot\mathbf{y}}Y_j(\mathbf{y},t)D_{Y,s}(\mathbf{y})\,d\mathbf{y}.\end{aligned}
K\mathcal{K}DK,s(k)=1(2π)d∫Rde−ik⋅yDY,s(y) dy,K‾j,s(k,t)=1iℏ∫RdKj(k′,t)DK,s(k−k′) dk′.\begin{aligned}D_{K,s}(\mathbf{k})&=\frac{1}{(2\pi)^d}\int_{\mathbb{R}^d}e^{-i\mathbf{k}\cdot\mathbf{y}}D_{Y,s}(\mathbf{y})\,d\mathbf{y},\\ \overline{K}_{j,s}(\mathbf{k},t)&=\frac{1}{i\hbar}\int_{\mathbb{R}^d}K_j(\mathbf{k}',t)D_{K,s}(\mathbf{k}-\mathbf{k}')\,d\mathbf{k}'.\end{aligned}

Table 1. The resulting K‾j,s(k,t)\overline{K}_{j,s}(\mathbf{k},t) corresponding to truncation methods.

L(k,t)=K(k,t)ST=∫Rde−ik⋅yU(y,t) dy,(37)L(k,t)=K(k,t)S^T=\int_{\mathbb{R}^d}e^{-i k\cdot y}U(y,t)\,\mathrm{d}y, \tag*{(37)}
ℏm∑s=1dksL(k,t)BsT=∫Rde−ik⋅yV(y,t) dy,(38)\frac{\hbar}{m}\sum_{s=1}^{d}k_sL(k,t)B_s^T=\int_{\mathbb{R}^d}e^{-i k\cdot y}V(y,t)\,\mathrm{d}y, \tag*{(38)}

then the LL-step (3.12) turns to

∂∂tU(y,t)=1iℏ∑s=1RDY,s(y)U(y,t)DsT−V(y,t).(39)\frac{\partial}{\partial t}U(y,t)=\frac{1}{i\hbar}\sum_{s=1}^{R}D_{Y,s}(y)U(y,t)D_s^T-V(y,t). \tag*{(39)}

Therefore, we proceed to solve equation (3.15) involving the variable U(y,t)U(y,t), where V(y,t)V(y,t) can be regarded as a function of U(y,t)U(y,t). When RR is approximately the same size as the dimension dd, similar to the situation in the Z-step, we can likewise exploit the symmetry of the coefficient matrix DsD_s and combine a splitting method with an exponential integrator to solve the corresponding equations. Conversely, when RR is relatively large, the Runge–Kutta method is more suitable for (3.15). Similarly, for K\mathcal{K}-truncation, utilizing the convolution property of Fourier transform and the notation in (3.13) and (3.14), we can obtain the evolution equation for U(y,t)U(y,t) as same as (3.15).

Complexity analysis

Then we considered the time complexity of the proposed Algorithm 3.1. Here, we consider the case of Ny=O(Nk)N_y=\mathcal{O}(N_k).

We first consider the computational complexity required for calculating K‾j,s(k,t)\overline{K}_{j,s}(k,t). The additional storage requirements are all of order O(Nkd(R+r))\mathcal{O}(N_k^d(R+r)) and the required time complexity is O(Nkdlog⁡(Nk)Rr)\mathcal{O}(N_k^d\log(N_k)Rr). Then consider the other steps in Algorithm 3.1, the storage cost and time complexity for computing the coefficients and solving the equations is shown in Table 2. For completeness, the time complexity of both QR decomposition and matrix multiplication can be summarized as O((Nkd+Nxd)r2)\mathcal{O}((N_k^d+N_x^d)r^2). In summary, compared to the full-grid method, which exhibits a time and space complexity of O(NxdNkd)\mathcal{O}(N_x^dN_k^d), our DLRA method offers significant advantages.

StepsStorage costTime complexity
AuA_uO(r2)\mathcal{O}(r^2)O(Nkdr2)\mathcal{O}(N_k^d r^2)
BuB_uO(r2)\mathcal{O}(r^2)O(Nxdlog⁡(Nx)r+Nxdr2)\mathcal{O}(N_x^d\log(N_x)r+N_x^d r^2)
Cs,DsC_s,D_sO(Rr2)\mathcal{O}(Rr^2)O(NkdRr2)\mathcal{O}(N_k^d Rr^2)
Z-stepO(Nxdr)\mathcal{O}(N_x^d r)O(Nxdlog⁡(Nx)r+NxdRr2)\mathcal{O}(N_x^d\log(N_x)r+N_x^d Rr^2)
S-step—O(Rr4)\mathcal{O}(Rr^4)
L-stepO(Nkdr)\mathcal{O}(N_k^d r)O(Nkdlog⁡(Nk)r+NkdRr2)\mathcal{O}(N_k^d\log(N_k)r+N_k^d Rr^2)

Table 2. Complexity analysis of the main steps in the Algorithm 3.1.

Remark 3.2. Due to the convolution involved, K\mathcal{K}-truncation requires zero-padding of factor 2 to avoid aliasing errors; however, its complexity does not differ from that of the Y\mathcal{Y}-truncation.

Numerical experiments

In this section, we evaluate the accuracy and efficiency of the proposed DLRA method. As we mentioned in Section 2, we adopt the Y\mathcal{Y}-truncation in the subsequent numerical experiments.

We assess the accuracy of our algorithm using a series of typical benchmark tests for the Wigner equation, including the quantum harmonic oscillator and Gaussian barrier scattering. These test cases are characterized by non-stationary and non-low-rank features, which are significant for extending the applicability of the DLRA method. Once numerical accuracy has been verified, we can investigate typical quantum systems such as the electron-electron scattering and Helium-like system, and reveal fundamental quantum effects such as the Pauli exclusion principle and the Heisenberg uncertainty principle.

The numerical performance is evaluated using the L2L^2 error ε2(t)\varepsilon_2(t) and the L∞L^\infty error ε∞(t)\varepsilon_{\infty}(t), defined as follows:

ε2(t)=[∬X×K(fref(x,k,t)−fnum(x,k,t))2dxdk]12,(40)\varepsilon_{2}(t)=\left[\iint_{\mathcal{X}\times\mathcal{K}}\left(f^{\mathrm{ref}}(\boldsymbol{x},\boldsymbol{k},t)-f^{\mathrm{num}}(\boldsymbol{x},\boldsymbol{k},t)\right)^2\mathrm{d}\boldsymbol{x}\mathrm{d}\boldsymbol{k}\right]^{\frac{1}{2}}, \tag*{(40)}
ε∞(t)=max⁡(x,k)∈X×K∣fref(x,k,t)−fnum(x,k,t)∣,(41)\varepsilon_{\infty}(t)=\max_{(\boldsymbol{x},\boldsymbol{k})\in\mathcal{X}\times\mathcal{K}}\left|f^{\mathrm{ref}}(\boldsymbol{x},\boldsymbol{k},t)-f^{\mathrm{num}}(\boldsymbol{x},\boldsymbol{k},t)\right|, \tag*{(41)}

where freff^{\mathrm{ref}} and fnumf^{\mathrm{num}} denote the reference and DLRA numerical solution, respectively. For 4D problem, we adopt the reduced Wigner function onto the 2D phase-space plane (xj,kj)(x_j,k_j),

Ws(x,k,t)=∬R×Rf(x,k,t)dx3−sdk3−s,s=1,2,(42)W_s(x,k,t)=\iint_{\mathbb{R}\times\mathbb{R}}f(\boldsymbol{x},\boldsymbol{k},t)\mathrm{d}x_{3-s}\mathrm{d}k_{3-s},\quad s=1,2, \tag*{(42)}

or reduced one-body Wigner function [3]

W(x,k,t)=W1(x,k,t)+W2(x,k,t).(43)W(x,k,t)=W_1(x,k,t)+W_2(x,k,t). \tag*{(43)}

In practice, the low-rank assumption of f(x,k,t)f(\boldsymbol{x},\boldsymbol{k},t) in (1.7) yields

Ws(x,k,t)=∑i,j=1rSij(t)∫RXi(x,t)dx3−s∫RKj(k,t)dk3−s,(44)W_s(x,k,t)=\sum_{i,j=1}^{r}S_{ij}(t)\int_{\mathbb{R}}X_i(\boldsymbol{x},t)\mathrm{d}x_{3-s}\int_{\mathbb{R}}K_j(\boldsymbol{k},t)\mathrm{d}k_{3-s}, \tag*{(44)}

so that WsW_s can be obtained directly from the low-rank components XiX_i, SijS_{ij} and KjK_j without first recovering the full-grid numerical solution fnumf^{\mathrm{num}}.

Quantum harmonic oscillator

In this subsection, we consider 2D quantum harmonic oscillator V(x)=mωx2/2V(x)=m\omega x^2/2 with R=1R=1. The exact solution can be obtained as f(x,k,t)=f0(x(t),k(t))f(x,k,t)=f_0(x(t),k(t)), where (x(t),k(t))(x(t),k(t)) are the characteristic lines given by

x(t)=cos⁡(ωt)x(0)−ℏmωsin⁡(ωt)k(0),k(t)=mωℏsin⁡(ωt)x(0)+cos⁡(ωt)k(0).(45)\begin{aligned} x(t)&=\cos(\sqrt{\omega}t)x(0)-\frac{\hbar}{m\sqrt{\omega}}\sin(\sqrt{\omega}t)k(0),\\ k(t)&=\frac{m\sqrt{\omega}}{\hbar}\sin(\sqrt{\omega}t)x(0)+\cos(\sqrt{\omega}t)k(0). \tag*{(45)} \end{aligned}

For simplicity, we still choose parameters ℏ=m=ω=1\hbar=m=\omega=1, equip with an initial standard Gaussian wave packet (GWP)

f0(x,k)=1πe−(x−x0)22σ2−2σ2(k−k0)2,(46)f_0(x,k)=\frac{1}{\pi}\mathrm{e}^{-\frac{(x-x_0)^2}{2\sigma^2}-2\sigma^2(k-k_0)^2}, \tag*{(46)}

where x0=1x_0=1 is the center of GWP and k0=0k_0=0 is the initial wave number. Then reference solution can be written as

f(x,k,t)=f0(cos⁡(t)x−sin⁡(t)k,sin⁡(t)x+cos⁡(t)k),(47)f(x,k,t)=f_0(\cos(t)x-\sin(t)k,\sin(t)x+\cos(t)k), \tag*{(47)}

which means that reference solution has a period of 2π2\pi, thus we choose final time T=2πT=2\pi. The computational domain is Ω=Ωx×Ωk=[−12,12]×[−6.4,6.4]\Omega=\Omega_x\times\Omega_k=[-12,12]\times[-6.4,6.4] with Nx=Nk=128N_x=N_k=128. Since we mainly focus on the convergence with respect to time step Δt=T/Nt\Delta t=T/N_t and rank rr, several groups of simulations under Nt=5000,10000,20000,40000,80000N_t=5000,10000,20000,40000,80000 and r=5,7,9,10,15,20,25,30r=5,7,9,10,15,20,25,30 are performed.

First, we examine the rank of the reference solution. Once the spatial discretization is fixed, the reference solution can be treated as a matrix, and its effective rank can be estimated via SVD. The L2L_2 norm of the error between the reference solution matrix and its best rank-rr approximation for r=5,10,15,20,25,30r=5,10,15,20,25,30 is shown in Figure 1(a), which is equal to (r+1)(r+1)-th singular value of the error matrix. Therefore, we observe that the singular values of the solution decay in the range of machine accuracy without a distinct gap, except 4 specific times t=0,π2,π,3π2t=0,\frac{\pi}{2},\pi,\frac{3\pi}{2} in one period, which the exact rank equal to 1. Note that the solution from our DLRA algorithm can not be more accurate than the best rank rr approximation obtained via SVD. Hence, this numerical test either does not exhibit a genuinely low-rank structure, or its effective rank is relatively high, a behavior that differs from typical DLRA numerical experiments.

Figure 1. (a) ∥⋅∥2\lVert\cdot\rVert_2 for best rank-rr approximation. (b) Convergence with respect to Δt\Delta t. (c) ε2(t)\varepsilon_2(t) under Nt=10000N_t=10000 and different rr. (d) ε∞(t)\varepsilon_\infty(t) under Nt=10000N_t=10000 and different rr. (e) ε2(t)\varepsilon_2(t) under r=20r=20 and different NtN_t. (f) ε∞(t)\varepsilon_\infty(t) under r=20r=20 and different NtN_t. Fig. 1. 2D quantum harmonic oscillator.

Second, we examine the convergence with respect to the rank rr. We fix time discretization parameters Nt=10000N_t=10000 and variate the rank rr, the resulting numerical error ε2(t)\varepsilon_2(t) and ε∞(t)\varepsilon_\infty(t) are shown in the Figure 1(c) and Figure 1(d), respectively. It can be observed that ε2(t)\varepsilon_2(t) and ε∞(t)\varepsilon_\infty(t) behave roughly similarly. When rr increases from 5 to 15, the numerical error decreases significantly. However, as rr grows further from 15 to 30, the reduction slows down or even stops, this is because the error at this time is mainly controlled by time step Δt\Delta t.

Last, we investigate the convergence with respect to the time step Δt\Delta t. Error indicators ε2(t)\varepsilon_2(t) and ε∞(t)\varepsilon_\infty(t) for a fixed rank r=20r=20 and different time discretization NtN_t are shown in the Figure 1(e) and Figure 1(f). We observe that the performances of ε2(t)\varepsilon_2(t) and ε∞(t)\varepsilon_\infty(t) are basically the same, they show a clear convergence trend as we increase the time partition parameters NtN_t and the convergence rate is plotted in Figure 1(b). However, Figure 1(b) reveals two distinct convergence orders. The difference stems from the numerical schemes employed for the ZZ-step (3.9) and LL-step (3.12): when these steps are solved with a second-order Runge–Kutta (2nd-RK) or a second-order Lawson (2nd-Lawson) method, the numerical error shows second-order convergence; when a first-order Lawson (1st-Lawson) method is used, only first-order convergence is obtained. This indicates that the projector-splitting error actually exhibits a numerical accuracy no lower than second order, even though a first-order projector-splitting algorithm is employed in this setting.

Gaussian barrier scattering

To further validate the overall performance of the DLRA for Wigner equation, we simulate the Gaussian barrier scattering [40, 41, 45] of the GWP to investigate its convergence rate. In this numerical experiment, a Gaussian potential barrier is introduced into the spatial domain, given by

V(x)=HPexp⁡(−(x−xB)22w2),(48)V(x)=H_P\exp\left(-\frac{(x-x_B)^2}{2w^2}\right), \tag*{(48)}

where HP=0.3H_P=0.3 is the height of the potential barrier, xB=30x_B=30 is the center of the computational domain and w=1w=1 is the dispersion of the potential barrier. We adopt the computational domain Ω=Ωx⊗Ωk=[0,60]×[−4,4]\Omega=\Omega_x\otimes\Omega_k=[0,60]\times[-4,4] with spacing Δy=π4\Delta y=\frac{\pi}{4}. The initial Wigner function is chosen to be the GWP (4.7) with x0=15x_0=15, k0=0.7k_0=0.7 and σ=2.825\sigma=2.825. And the other parameters are: the reduced Planck constant ℏ=0.658211899\hbar=0.658211899, the effective mass m=0.0665mem=0.0665m_e, the stationary electron mass me=5.68562966m_e=5.68562966 and the final time T=20T=20.

A quantitative observation is conducted in Figure 2. In this numerical example, the grid parameters are set as Nx=200,Nk=400N_x=200,N_k=400, and R=80R=80 is estimated via SVD. For reference solution, the same spatial grid subdivision parameters $N_x,N_k,N_y and the fixed time step Δt=1/200\Delta t=1/200 are adopted as before. Figure 2(a) displays the L2L_2 error between the reference SEM solution and its best rank-rr approximation for r=20,40,60,80,100,120,140r=20,40,60,80,100,120,140. From this we can observe that the rank of

Figure 2. (a) ∥⋅∥2\| \cdot\|_2 for best rank-rr approximation. (b) Convergence with respect to Δt\Delta t. (c) ε2(t)\varepsilon_2(t) under Nt=10000N_t = 10000 and different rr. (d) ε2(t)\varepsilon_2(t) under r=100r = 100 and different NtN_t. 2D Gaussian barrier scattering.

this numerical test grows very rapidly over time, and relative to the grid size, the problem can be considered non-low-rank. For DLRA solution, we mainly focus on the convergence with respect to time step Δt\Delta t and rank rr, several groups of under Δt=1/200,1/300,1/400,1/500\Delta t = 1/200, 1/300, 1/400, 1/500 and r=30,40,50,60,70,80,100,120r = 30, 40, 50, 60, 70, 80, 100, 120 are performed. Figure 2(c) gives the relationship between error ε2(t)\varepsilon_2(t) and rank rr under the same time step Δt=1/500\Delta t = 1/500. When the rank r≤100r \le100, the error ε2(t)\varepsilon_2(t) decreases noticeably as rr increase. However, the reduction of ε2(t)\varepsilon_2(t) becomes much less pronounced or even ceases for r>100r > 100. This behavior occurs because the error is now dominated by the time step size Δt\Delta t. We also examine the influence of the time step Δt\Delta t on the error, then Figure 2(d) shows the dependence of the error ε2(t)\varepsilon_2(t) on Δt\Delta t for a fixed rank r=100r = 100. We observe that errors accumulate progressively over time, and as the time step is reduced, ε2(t)\varepsilon_2(t) decreases substantially while exhibiting a clear order of convergence. The behavior of the other error indicator ε∞(t)\varepsilon_\infty(t) is similar to ε2(t)\varepsilon_2(t). Moreover, the convergence rate with respect to Δt\Delta t presented in Figure 2(b), clearly exhibits second-order convergence as well.

Such Gaussian barrier scattering can be extended to the 4D situation easily. The Wigner equation f(x1,x2,k1,k2,t)f(x_1, x_2, k_1, k_2, t) with 2D potential V(x1,x2)=V1(x1)+V2(x2)V(x_1, x_2) = V_1(x_1) + V_2(x_2) has the form

f(x1,x2,k1,k2,t)=f1(x1,k1,t)f2(x2,k2,t),(49)f(x_1,x_2,k_1,k_2,t)=f_1(x_1,k_1,t)f_2(x_2,k_2,t), \tag*{(49)}

where fi(xi,ki,t)f_i(x_i,k_i,t) is 2D Wigner function corresponding to the 1D potential Vi(xi)V_i(x_i) for i=1,2i=1,2. For instance, we adopt computational domain Ω=Ωx×Ωk=[0,60]2×[−4,4]2\Omega=\Omega_x\times\Omega_k=[0,60]^2\times[-4,4]^2 with spacing Δy1=Δy2=π/4\Delta y_1=\Delta y_2=\pi/4. The potential function is chosen as

V(x1,x2)=HPexp⁡[−(x1−x1,B)22w2],(50)V(x_1,x_2)=H_P\exp\left[-\frac{(x_1-x_{1,B})^2}{2w^2}\right], \tag*{(50)}

where HP=0.3H_P=0.3, x1,B=30x_{1,B}=30 and w=1w=1. The initial Wigner function is set as degenerate function

f(x1,x2,k1,k2,0)=1π1Lx21Lk2exp⁡[−(x1−x1,0)22a12]exp⁡[−2a12(k1−k1,0)2],(51)f(x_1,x_2,k_1,k_2,0)=\frac{1}{\pi}\frac{1}{L_{x_2}}\frac{1}{L_{k_2}}\exp\left[-\frac{(x_1-x_{1,0})^2}{2a_1^2}\right]\exp\left[-2a_1^2(k_1-k_{1,0})^2\right], \tag*{(51)}

where x1,0=15x_{1,0}=15, k1,0=0.7k_{1,0}=0.7, a1=2.825a_1=2.825, Lx2=∣Ωx2∣=60L_{x_2}=|\Omega_{x_2}|=60 and Lk2=∣Ωk2∣=4L_{k_2}=|\Omega_{k_2}|=4. In this case, the Wigner function f(x1,x2,k1,k2,t)f(x_1,x_2,k_1,k_2,t) is degenerate in x2x_2-direction and k2k_2-direction, and the reduced Wigner function W1(x,k,t)W_1(x,k,t) is the same as that of the previous numerical test. The remaining parameters are the same as the previous 2D case. To study the convergence rate with respect to time step Δt\Delta t, we fix Nx1=Nx2=Nk1=Nk2=100N_{x_1}=N_{x_2}=N_{k_1}=N_{k_2}=100 and R=63R=63 is estimated by random SVD. The order of convergence is presented in Figure 3(a), we can clearly observe the second-order convergence order as well.

Figure 3. 4D4D Gaussian barrier scattering: The convergence order with respect to the time step Δt\Delta t. The left figure shows the convergence order for the case with degeneracy in the x2x_2-k2k_2 direction, while the right figure displays the convergence order for the free advection in the x2x_2-k2k_2 direction. In both cases, second-order numerical accuracy can be observed.

Similarly, if we choose the same potential function in (50) with initial Wigner function as

f(x1,x2,k1,k2,0)=1π2exp⁡[−(x1−x1,0)22a12−(x2−x2,0)22a22]×exp⁡[−2a12(k1−k1,0)2−2a22(k2−k2,0)2],(52)\begin{aligned} f(x_1,x_2,k_1,k_2,0)&=\frac{1}{\pi^2}\exp\left[-\frac{(x_1-x_{1,0})^2}{2a_1^2}-\frac{(x_2-x_{2,0})^2}{2a_2^2}\right]\\ &\quad\times\exp\left[-2a_1^2(k_1-k_{1,0})^2-2a_2^2(k_2-k_{2,0})^2\right], \tag*{(52)} \end{aligned}

where x1,0x_{1,0}, k1,0k_{1,0} and a1a_1 are in the same way as before, x2,0=45x_{2,0}=45, k2,0=−0.7k_{2,0}=-0.7 and a2=2.825a_2=2.825. In this case, the reduced Wigner function W1(x,k,t)W_1(x,k,t) is the same as that of the previous numerical test and W2(x,k,t)W_2(x,k,t) evolves as the free convection of an initial GWP over time. The remaining parameters are the same as the previous case except the rank rr is chosen as r=400r=400, which is much larger than the previous. The convergence order, shown in Figure 3(b), also clearly exhibits second-order accuracy.

Electron-electron scattering

Now we turn to discuss a more challenging problem. Consider that two electrons interacting through the repulsive Coulomb force, the potential function is given by the soft Coulomb potential

Vs(x1,x2)=1(x1−x2)2+ϵ,(53)V_s(x_1,x_2)=\frac{1}{\sqrt{(x_1-x_2)^2+\epsilon}}, \tag*{(53)}

where ϵ=1\epsilon=1 is the soft parameter. The initial value of the Wigner function to simulate two electronic states is chosen as a Fermion system

f0Fermion(x1,x2,k1,k2)=12π2exp⁡[−(x1−x10)22a2−(x2−x20)22a2−2a2(k1−k10)2−2a2(k2−k20)2]+12π2exp⁡[−(x1−x20)22a2−(x2−x10)22a2−2a2(k1−k20)2−2a2(k2−k10)2]−1π2exp⁡[−(x1−x10)2+(x1−x20)2+(x2−x10)2+(x2−x20)24a2]×exp⁡[(x10−x20)24a2−2a2(k1−k10+k202)2−2a2(k2−k10+k202)2]×cos⁡[(x10−x20)(k1−k2)−(k10−k20)(x1−x2)].(54)\begin{aligned} f_0^{\mathrm{Fermion}}(x_1,x_2,k_1,k_2) ={}&\frac{1}{2\pi^2}\exp\left[-\frac{(x_1-x_1^0)^2}{2a^2}-\frac{(x_2-x_2^0)^2}{2a^2}-2a^2(k_1-k_1^0)^2-2a^2(k_2-k_2^0)^2\right]\\ &+\frac{1}{2\pi^2}\exp\left[-\frac{(x_1-x_2^0)^2}{2a^2}-\frac{(x_2-x_1^0)^2}{2a^2}-2a^2(k_1-k_2^0)^2-2a^2(k_2-k_1^0)^2\right]\\ &-\frac{1}{\pi^2}\exp\left[-\frac{(x_1-x_1^0)^2+(x_1-x_2^0)^2+(x_2-x_1^0)^2+(x_2-x_2^0)^2}{4a^2}\right]\\ &\times\exp\left[\frac{(x_1^0-x_2^0)^2}{4a^2}-2a^2\left(k_1-\frac{k_1^0+k_2^0}{2}\right)^2-2a^2\left(k_2-\frac{k_1^0+k_2^0}{2}\right)^2\right]\\ &\times\cos\left[(x_1^0-x_2^0)(k_1-k_2)-(k_1^0-k_2^0)(x_1-x_2)\right]. \tag*{(54)} \end{aligned}

In the following computations, the Fermion system is taken as the initial condition, with the parameters set as x10=−2,x20=2,k10=0.5,k20=−0.5,a=2x_1^0=-2,x_2^0=2,k_1^0=0.5,k_2^0=-0.5,a=2, as illustrated in Figure 4(a).

Initial value for two fermions and convergence order with respect to the time step

Figure 4. The reduced Wigner function W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) for two fermions and the convergence order.

The first run is to demonstrate the numerical convergence of the proposed DLRA method by setting computational domain Ω=Ωx×Ωk=[−15,15]2×[−5π/6,5π/6]2\Omega=\Omega_x\times\Omega_k=[-15,15]^2\times[-5\pi/6,5\pi/6]^2 with spacings Δy1=Δy2=1.2\Delta y_1=\Delta y_2=1.2. The number of grid points in each spatial direction is set to Nx1=Nx2=150N_{x_1}=N_{x_2}=150, and in each momentum direction to Nk1=Nk2=100N_{k_1}=N_{k_2}=100 with final time T=4T=4. The rank r=200r=200 is chosen sufficiently large to reveal the convergence order with respect to the time step Δt\Delta t, while R=58R=58 is estimated by random SVD. The reference solution is computed using advective-spectral-mixed (ASM) scheme [45] with the same parameters listed above. Figure 4(b) presents the results, clearly showing that our DLRA algorithm achieves second-order convergence. However, choosing an excessively large rank rr leads to high computational cost, especially in the SS-step, which contradicts the original purpose of adopting a low-rank approach. In this nu-

merical test, a rank of r=150r = 150 is sufficient to produce an approximate solution that is visually indistinguishable from the reference.

Then, Figure 5 shows the evolution of the reduced Wigner function W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) under the same parameters listed above, up to final time T=6T = 6. The reduced Wigner function of the two electrons moving towards each other and decelerate before t=4t = 4 is suppressed in the ∣k∣≤1|k| \le1 region and has greater expansion in the xx-space due to Coulomb potential. Since then, two electrons tend to scattered and separated from each other due to repulsive interaction, which is line with our expectations. Moreover, it should be emphasized that at t=0t = 0, a hole structure known as the Fermi hole, appears at the center ((x10+x20)/2,(k10+k20)/2)\left((x_1^0+x_2^0)/2,(k_1^0+k_2^0)/2\right) between the two electrons as a direct manifestation of the Pauli exclusion principle. Furthermore, this Fermi hole persists throughout the evolution due to the strong correlation between the electrons.

Electron-electron scattering at different times, with panels (a) $t=1$, (b) $t=2$, (c) $t=3$, (d) $t=4$, (e) $t=5$, and (f) $t=6$

Figure 5. Electron-electron scattering: W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) at different time.

Finally, we employ this example to illustrate the advantages of the DLRA method in handling such high-dimensional problems. Table 3 compares the theoretical memory requirement and the running time per step for this problem under different ranks rr, using the same parameters as above, for both the reference ASM scheme and our DLRA method. It can be observed that the DLRA method outperforms the reference full-grid method in both theoretical memory requirement and computational time. For example, with r=150r = 150 being sufficient for this problem, the memory required by the DLRA algorithm is two orders of magnitude smaller, and the running time is one order of magnitude faster.

Full grids (ASM)r=50r = 50r=100r = 100r=150r = 150r=200r = 200
Theoretical memory requirement (GB)23.450.040.070.100.13
Running time per step (s)75.762.585.348.3511.34

Table 3. Comparison of computational cost for electron–electron scattering simulations: full-grid (ASM) vs. DLRA with different ranks rr.

A Helium-like system

As the final example, we consider a Helium-like system composed of two electrons. Besides the repulsive Coulomb force, they are both attracted by a Helium atom at x=0x = 0. To describe the nucleon-electron interaction, we adopt the attractive soft-Coulomb potential [33]

Vhe(x1,x2)=−Z∣x1∣2+ϵ−Z∣x2∣2+ϵ,(55)V_{he}(x_1,x_2) = -\frac{Z}{\sqrt{|x_1|^2+\epsilon}}-\frac{Z}{\sqrt{|x_2|^2+\epsilon}}, \tag*{(55)}

with the atomic number Z=2Z = 2 and the soft parameter ϵ=1\epsilon= 1. The initial value is taken to be the same Fermion system as in the previous case, using the identical parameter set, shown in Figure 4(a). And the other parameters are: the computational domain Ω=Ωx×Ωk=[−15,15]2×[−5π/3,5π/3]2\Omega= \Omega_x \times\Omega_k = [-15, 15]^2 \times[-5\pi/3, 5\pi/3]^2 with spacings Δy1=Δy2=0.6\Delta y_1 = \Delta y_2 = 0.6, the number of grid points in each spatial direction is set to Nx1=Nx2=150N_{x_1} = N_{x_2} = 150, and in each momentum direction to Nk1=Nk2=100N_{k_1} = N_{k_2} = 100, with R=106R = 10^6 is estimated by random SVD, while r=400r = 400 is chosen sufficiently large.

The first run is carried out to demonstrate the numerical convergence of the proposed DLRA method. Figure 6 presents the results, clearly showing that our DLRA method achieves second-order convergence at time T=1T = 1. Similarly, in this test case, a sufficiently large rank is required to reliably measure the convergence order; rank r=150r = 150 is sufficient to compute an approximate solution that captures the relevant physical properties.

Figure 6. The reduced Wigner function W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) for two fermions and the convergence order.

To demonstrate the dynamics of electrons more clearly, we take snapshots of the reduced Wigner function W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) from t=1t = 1 to t=15t = 15, as shown Figure 7. At early times (t<5t < 5), the reduced Wigner function remains localized in the central region because of the nucleon–electron interaction. Afterwards, both dispersion and correlation become clearly visible, producing a highly oscillatory structure in phase space. The region over which the reduced Wigner function evolves expands progressively with time, reaching the preset computational domain boundary by t=15t = 15. It is worth noting that a region capable of taking negative values, referred to as the Fermi hole, persists around the centers of both wave packets. This is a direct consequence of the electron–electron interaction, since negative distributions are associated with areas inaccessible to experimental observation due to the constraints imposed by the uncertainty principle. The branching behavior of the Fermi hole varies at different times: for example, it exhibits two branches at t=2t = 2, while only one branch is visible at t=10t = 10. Nevertheless, it persists throughout, which agrees with the results from the ASM method [45] but differs from those obtained by the signed Monte Carlo method [39]. This discrepancy may stem from the limited accuracy of the signed Monte Carlo method, which is insufficient to precisely capture quantum properties such as quantum interference. This aspect warrants further investigation in the future.

Figure 7. The Helium-like system: W1(x,k,t)+W2(x,k,t)W_1(x,k,t) + W_2(x,k,t) at different times.

Conclusion and discussion

This work presents the first application of DLRA to the Wigner equation. Our key innovation lies in introducing separable assumption (2.3) and combine it with two established approaches for Ψ\PsiDO to transform Ψ\PsiDO into a separated form (2.4). Comprehensive numerical tests on challenging non-stationary systems demonstrate both the accuracy and efficiency of our approach. Even in the absence of a predetermined low-rank structure for the solution, DLRA can still serve as a numerical scheme that balances efficiency and accuracy. These findings open new possibilities for simulating high-dimensional quantum systems.

While the current framework demonstrates promising results, several advanced strategies could further enhance the algorithm’s performance. First, the second-order robust BUG integrator [5] provides stable second-order accuracy and enables feasible adaptive rank control. Second, changing the way of low-rank such as dynamical tensor approximation [31] and hierarchical low-rank approximation [1] may help reduce the required rank. Third, emerging approximate interpolation techniques [6, 10, 11] could be adopted to accelerate the computation. Finally, beyond the extension in computational efficiency, we will consider more complex potentials in subsequent work, including singular potentials, discontinuous potentials, and even pure Coulomb potentials. These directions will be explored in our future research.

References

  1. [1]F. Allmann-Rahn, R. Grauer, and K. Kormann, A parallel low-rank solver for the six-dimensional Vlasov-Maxwell equations, Journal of Computational Physics, 469 (2022), p. 111562.
  2. [2]A. Arnold and C. Ringhofer, Operator splitting methods applied to spectral discretizations of quantum transport equations, SIAM Journal on Numerical Analysis, 32 (1995), pp. 1876–1894.DOI
  3. [3]E. Cancellieri, P. Bordone, and C. Jacoboni, Effect of symmetry in the many-particle Wigner function, Physical Review B, 76 (2007), p. 214301.DOI
  4. [4]F. Cassini and L. Einkemmer, Efficient 6D Vlasov simulation using the dynamical low-rank framework, Computer Physics Communications, 280 (2022), p. 108489.DOI
  5. [5]G. Ceruti, L. Einkemmer, J. Kusch, and C. Lubich, A robust second-order low-rank BUG integrator based on the midpoint rule, BIT Numerical Mathematics, 64 (2024), p. 30.arxiv.org/abs/2402.08607
  6. [6]S. Chaturantabut and D. C. Sorensen, Nonlinear model reduction via discrete empirical interpolation, SIAM Journal on Scientific Computing, 32 (2010), pp. 2737–2764.DOI
  7. [7]Z. Chen, H. Jiang, and S. Shao, A higher-order accurate operator splitting spectral method for the Wigner–Poisson system, Journal of Computational Electronics, 21 (2022), pp. 756–770.
  8. [8]Z. Chen, S. Shao, and W. Cai, A high order efficient numerical method for 4-D Wigner equation of quantum double-slit interferences, Journal of Computational Physics, 396 (2019), pp. 54–71.DOI
  9. [9]Z. Chen, Y. Xiong, and S. Shao, Numerical methods for the Wigner equation with unbounded potential, Journal of Scientific Computing, 79 (2019), pp. 345–368.DOI
  10. [10]A. Christlieb, S. Gong, J.-M. Qiu, and N. Zheng, A sampling-based adaptive rank approach to the Wigner-Poisson system, June 2025, https://arxiv.org/abs/2506.21314.
  11. [11]A. Dektor and L. Einkemmer, Interpolatory dynamical rank approximation for the 3+3d Boltzmann–BGK equation, Journal of Computational Physics, (2025), p. 114515.arxiv.org/abs/2411.15990
  12. [12]P. A. M. Dirac, Note on exchange phenomena in the Thomas atom, in Mathematical Proceedings of the Cambridge Philosophical Society, vol. 26, Cambridge University Press, 1930, pp. 376–385.DOI
  13. [13]A. Dorda and F. Schürrer, A WENO-solver combined with adaptive momentum discretization for the Wigner transport equation and its application to resonant tunneling diodes, Journal of Computational Physics, 284 (2015), pp. 95–116.pubmed.ncbi.nlm.nih.gov/25892748
  14. [14]C. Eckart and G. Young, The approximation of one matrix by another of lower rank, Psychometrika, 1 (1936), pp. 211–218.DOI
  15. [15]L. Einkemmer, J. Hu, and Y. Wang, An asymptotic-preserving dynamical low-rank method for the multi-scale multi-dimensional linear transport equation, Journal of Computational Physics, 439 (2021), p. 110353.DOI
  16. [16]L. Einkemmer, J. Hu, and L. Ying, An efficient dynamical low-rank algorithm for the Boltzmann-BGK equation close to the compressible viscous flow regime, SIAM Journal on Scientific Computing, 43 (2021), pp. B1057–B1080.arxiv.org/abs/2101.07104
  17. [17]L. Einkemmer and I. Joseph, A mass, momentum, and energy conservative dynamical low-rank scheme for the Vlasov equation, Journal of Computational Physics, 443 (2021), p. 110495.DOI
  18. [18]L. Einkemmer and C. Lubich, A low-rank projector-splitting integrator for the Vlasov–Poisson equation, SIAM Journal on Scientific Computing, 40 (2018), pp. B1330–B1360.arxiv.org/abs/1801.01103
  19. [19]W. R. Frensley, Effect of inelastic processes on the self-consistent potential in the resonant-tunneling diode, Solid-State Electronics, 32 (1989), pp. 1235–1239.DOI
  20. [20]A. Gehring and H. Kosina, Wigner function-based simulation of quantum transport in scaled DG-MOSFETs using a Monte Carlo method, Journal of Computational Electronics, 4 (2005), pp. 67–70.DOI
  21. [21]F. R. Graziani, J. D. Bauer, and M. S. Murillo, Kinetic theory molecular dynamics and hot dense matter: Theoretical foundations, Physical Review E, 90 (2014), p. 033104.pubmed.ncbi.nlm.nih.gov/25314544
  22. [22]W. Heisenberg, Über den anschaulichen Inhalt der quantentheoretischen Kinematik und Mechanik, Zeitschrift für Physik, 43 (1927), pp. 172–198.
  23. [23]M. Hochbruck and A. Ostermann, Exponential integrators, Acta Numerica, 19 (2010), pp. 209–286.DOI
  24. [24]J. Hu and Y. Wang, An adaptive dynamical low rank method for the nonlinear Boltzmann equation, Journal of Scientific Computing, 92 (2022), p. 75.arxiv.org/abs/2112.02695
  25. [25]M. Hug, C. Menke, and W. P. Schleich, Modified spectral method in phase space: Calculation of the Wigner function. I. Fundamentals, Physical Review A, 57 (1998), pp. 3188–3205.
  26. [26]M. Hug, C. Menke, and W. P. Schleich, Modified spectral method in phase space: Calculation of the Wigner function. II. Generalizations, Physical Review A, 57 (1998), pp. 3206–3224.DOI
  27. [27]C. Jacoboni and P. Bordone, The Wigner-function approach to non-equilibrium electron transport, Reports on Progress in Physics, 67 (2004), p. 1033.DOI
  28. [28]C. Jacoboni, R. Brunetti, P. Bordone, and A. Bertoni, Quantum transport and its simulation with the Wigner-function approach, International Journal of High Speed Electronics and Systems, 11 (2001), pp. 387–423.
  29. [29]E. Kieri, C. Lubich, and H. Walach, Discretized dynamical low-rank approximation in the presence of small singular values, SIAM Journal on Numerical Analysis, 54 (2016), pp. 1020–1038.DOI
  30. [30]O. Koch and C. Lubich, Dynamical low-rank approximation, SIAM Journal on Matrix Analysis and Applications, 29 (2007), pp. 434–454.DOI
  31. [31]O. Koch and C. Lubich, Dynamical tensor approximation, SIAM Journal on Matrix Analysis and Applications, 31 (2010), pp. 2360–2375.DOI
  32. [32]H. Kosina, M. Nedjalkov, and S. Selberherr, A Monte Carlo method seamlessly linking quantum and classical transport calculations, Journal of Computational Electronics, 2 (2003), pp. 147–151.DOI
  33. [33]M. Lein, T. K. Kreibich, E. K. U. Gross, and V. Engel, Strong-field ionization dynamics of a model H₂ molecule, Physical Review A, 65 (2002), p. 033403.DOI
  34. [34]R. Li, T. Lu, Y. Wang, and W. Yao, Numerical validation for high order hyperbolic moment system of Wigner equation, Communications in Computational Physics, 15 (2014), pp. 569–595.DOI
  35. [35]C. Lubich and I. V. Oseledets, A projector-splitting integrator for dynamical low-rank approximation, BIT Numerical Mathematics, 54 (2014), pp. 171–188.arxiv.org/abs/1301.1058
  36. [36]Z. Peng, R. G. McClarren, and M. F. Frank, A low-rank method for two-dimensional time-dependent radiation transport calculations, Journal of Computational Physics, 421 (2020), p. 109735.DOI
  37. [37]C. Ringhofer, A spectral method for the numerical simulation of quantum tunneling phenomena, SIAM Journal on Numerical Analysis, 27 (1990), pp. 32–50.DOI
  38. [38]W. P. Schleich, Quantum Optics in Phase Space, John Wiley & Sons, 2015.
  39. [39]J. M. Sellier and I. Dimov, On the simulation of indistinguishable fermions in the many-body Wigner formalism, Journal of Computational Physics, 280 (2015), pp. 287–294.DOI
  40. [40]S. Shao, T. Lu, and W. Cai, Adaptive conservative cell average spectral element methods for transient Wigner equation in quantum transport, Communications in Computational Physics, 9 (2011), pp. 711–739.DOI
  41. [41]S. Shao and J. M. Sellier, Comparison of deterministic and stochastic methods for time- dependent Wigner simulations, Journal of Computational Physics, 300 (2015), pp. 167–185.
  42. [42]S. SHAO AND Y. XIONG, A branching random walk method for many-body Wigner quantum dynamics, Numerical Mathematics: Theory, Methods and Applications, 12 (2019), pp. 21–71.DOI
  43. [43]M. L. VAN DE PUT, B. SORÉE, AND W. MAGNUS, Efficient solution of the Wigner–Liouville equation using a spectral decomposition of the force field, Journal of Computational Physics, 350 (2017), pp. 314–325.DOI
  44. [44]E. WIGNER, On the quantum correction for thermodynamic equilibrium, Physical Review, 40 (1932), pp. 749–759.
  45. [45]Y. XIONG, Z. CHEN, AND S. SHAO, An advective-spectral-mixed method for time-dependent many-body Wigner simulations, SIAM Journal on Scientific Computing, 38 (2016), pp. B491–B520.arxiv.org/abs/1602.08853
  46. [46]Y. XIONG, Y. ZHANG, AND S. SHAO, A characteristic-spectral-mixed scheme for six-dimensional Wigner–Coulomb dynamics, SIAM Journal on Scientific Computing, 45 (2023), pp. B906–B931.arxiv.org/abs/2205.02380

Paper details

vlnc:W_vb3wdjjc
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.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents