betaFeedback

WorksMathematics

Full- and low-rank exponential Euler integrators for the Lindblad equation

Abstract

The Lindblad equation is a widely used quantum master equation to model the dynamical evolution of open quantum systems whose states are described by density matrices. These solution matrices are characterized by semi-positiveness and trace preserving properties, which must be guaranteed in any physically meaningful numerical simulation. In this paper, novel full- and low-rank exponential Euler integrators are developed for approximating the Lindblad equation that preserve positivity and trace unconditionally. Theoretical results are presented that provide sharp error estimates for the two classes of exponential integration methods. Results of numerical experiments are discussed that illustrate the effectiveness of the proposed schemes, beyond present state-of-the-art capabilities.

Introduction

The Lindblad equation is a widely used Markovian quantum master equation to model the dynamical evolution of open quantum systems [1, 2]. The quantum master equation plays a crucial role in many application areas including condensed matter physics, quantum physics and quantum information theory. In this paper, we consider quantum systems consisting of KK dephasing dd-level qudits undergoing Markovian open quantum dynamics. The time dynamics of these systems is given by the Gorini-Kossakowski-Sudarshan-Lindblad equation (or Lindblad equation for short) [3, 4]:

ρ˙(t)=−i[H,ρ(t)]+∑k=1Kγk(Lkρ(t)Lk†−12{Lk†Lk,ρ(t)}),(1)\dot{\rho}(t)=-i[H,\rho(t)]+\sum_{k=1}^{K}\gamma_k\left(L_k\rho(t)L_k^\dagger-\frac{1}{2}\left\{L_k^\dagger L_k,\rho(t)\right\}\right), \tag*{(1)}

Email addresses: [email protected] (Hao Chen), [email protected] (Alfio Borzì), [email protected] (Denis Janković), [email protected] (Jean-Gabriel Hartmann), [email protected] (Paul-Antoine Hervieux)

where ρ(t)∈Cm×m\rho(t) \in\mathbb{C}^{m \times m} is the density matrix describing the state of the system, H=H†H = H^{\dagger} is the Hamiltonian operator describing the unitary evolution of the qudit, LkL_k are the Lindblad or jump operators characterizing the dissipation channels, and γk≥0\gamma_k \ge0 are the decay parameters for each of the KK channels.

The Lindblad equation can be also written in a vectorized form, using Dirac’s bracket notation, as follows:

∣ρ˙(t)⟩=L∣ρ(t)⟩,(2)\lvert\dot{\rho}(t)\rangle= \mathcal{L}\lvert\rho(t)\rangle, \tag*{(2)}

with

L=−iI⊗H+iHT⊗I+∑k=1Kγk(Lk∗⊗Lk−12I⊗(Lk†Lk)−12(LkTLk∗)⊗I),\mathcal{L}=-iI\otimes H+iH^{\mathsf{T}}\otimes I+\sum_{k=1}^{K}\gamma_k\left(L_k^{*}\otimes L_k-\frac{1}{2}I\otimes(L_k^{\dagger}L_k)-\frac{1}{2}(L_k^{\mathsf{T}}L_k^{*})\otimes I\right),

where ∣ρ⟩\lvert\rho\rangle is the vectorized density matrix and L∈Cm2×m2\mathcal{L} \in\mathbb{C}^{m^2 \times m^2} is the matrix representation of the Lindblad generator. The superscripts †\dagger, T\mathsf{T} and ∗* denote the adjoint, transpose and complex conjugate operators, respectively.

The Lindblad equation possesses two important properties [3, 4]: it is positivity and trace preserving, which means that the solution ρ(t)\rho(t) is Hermitian positive semidefinite and has unit trace if the initial data is Hermitian positive semidefinite with unit trace. These properties of the density matrix are of fundamental physical significance, and whether they can be preserved at the discrete level is a significant issue in numerical simulations, and especially for long-time simulation [5]. The major goal of this work is to develop positivity and trace preserving numerical schemes for approximating the Lindblad equation (1) with guaranteed accuracy.

The literature on numerical methods for solving the Lindblad equation is relatively scarce. Explicit Runge-Kutta methods have been discussed in [5], where the authors show that these methods do not preserve the positivity of the density matrix. The Crank-Nicolson method has been studied in [6] for a two-level Lindblad equation and it is shown in [7] that the Crank-Nicolson scheme can not preserve the positivity in general. In fact, standard Runge-Kutta methods and multistep methods cannot be better than first-order accurate if they must preserve positivity; see, e.g., [8, 9]. On the other hand, there are methods based on matrix exponential [5, 10, 11] for solving the vector representation (2) of the Lindblad equations. This kind of methods are positivity and trace preserving since the differential equation (2) is solved analytically. However, the coefficient matrix L\mathcal{L} is of size m2×m2m^2 \times m^2 and the computation of the product of the associated matrix exponential with vectors is expensive if m2m^2 is large. Further, there are methods based on operator splitting techniques [12, 13] and low-rank operator splitting schemes [14, 15], where by means of a normalization procedure the low-rank splitting schemes preserve positivity and trace; however, rigorous error analysis is still unavailable. Moreover, Kraus representation approximation methods have been studied in [16], where it is shown that these methods are positivity and trace preserving, where the latter is achieved by a normalization procedure. In addition, a decomposition of nonunitary operators approach has been developed in [17] for the vectorized Lindblad equation.

Besides the above mentioned approaches, several numerical methods have been studied for large-scale problems [18, 19], including methods based on matrix product states [20, 21, 22, 23] and strategies based on neural networks [24, 25, 26]. Tree tensor networks [27, 28] have also been developed to simulate quantum many-body states.

Noting that, in general, positivity and trace preserving schemes are relatively less considered and rigorous error analysis on numerical integrators for the Lindblad equation is largely unavailable, the aim of this paper is to develop and analyze positivity and trace preserving numerical schemes for solving the Lindblad equation (1). Inspired by the exponential integrators for solving differential matrix Riccati equations [29, 30], we present novel full- and low-rank exponential Euler integrators for the Lindblad equation. We prove that our exponential Euler schemes preserve positivity and trace unconditionally, that is, the new methods are unconditionally stable. However, while the full-rank scheme does not require any normalization procedure to preserve unit trace, in the low-rank case normalization is still required. On the other hand, trace preservation in the full-rank case allows us to prove sharp accuracy estimates that are also instrumental for obtaining the accuracy estimates for the low-rank method presented in this paper. Clearly, one disadvantage of the low-rank approach is to introduce an additional approximation error due to the low-rank representation of the density matrix. On the other hand, we have the benefit of drastically reducing computational costs as we show in the section of numerical experiments.

This paper is organized as follows. In Section 2, we introduce our full-rank exponential Euler integrator for discretizing the Lindblad equation and discuss related implementation issues. In Section 3, we show that the proposed full-rank exponential scheme preserves positivity and unit trace of the solution unconditionally. Section 4 is dedicated to the error analysis of the full-rank exponential Euler integrator. In Section 5, we propose the low-rank exponential Euler scheme and provide its error estimate. In Section 6, many numerical experiments are performed for the Lindblad equation with various Hamiltonians, including the tests of convergence rate, positivity and unit trace preservation, and effectiveness of the low-rank algorithm. Also in this section, we compare our exponential Euler schemes with the Lindblad solvers of the well-known QuTiP framework for simulation of open quantum systems [31], showing that, while the latter may provide higher accuracy, they seem not able to provide the required positivity property, which is guaranteed by the proposed full- and low-rank exponential Euler schemes. Moreover, it is shown that the new low-rank exponential scheme performs much faster than all other schemes considered in this paper, as the dimensionality of the problem increases. A section of conclusion completes this work.

An exponential Euler integrator

In this section, we develop an exponential integrator to solve the Lindblad equation (1) on the time interval [0,T][0,T], for some T>0T > 0, equipped with the initial condition ρ(0)=ρ0\rho(0) = \rho_0, where ρ0\rho_0 is a Hermitian and positive semidefinite matrix with unit trace. First, let us define

A:=−iH−12∑k=1KγkLk†Lk.(3)A := -iH - \frac{1}{2}\sum_{k=1}^{K}\gamma_k L_k^\dagger L_k. \tag*{(3)}

Then, we can rewrite (1.1) as follows:

ρ˙=Aρ+ρA†+∑k=1KγkLkρLk†.(4)\dot{\rho}=A\rho+\rho A^\dagger+\sum_{k=1}^{K}\gamma_k L_k\rho L_k^\dagger. \tag*{(4)}

Given a positive integer NN, we divide the time interval by NN subintervals of size τ=TN\tau=\frac{T}{N}, with endpoints tn=nτt_n=n\tau, n=0,1,…,Nn=0,1,\ldots,N. Then, integrating the equation (2.2) from tnt_n to tn+1t_{n+1} and using the variation-of-constants formula yields

ρ(tn+1)=eτAρ(tn)eτA†+∑k=1Kγk∫0τe(τ−s)ALkρ(tn+s)Lk†e(τ−s)A† ds.(5)\rho(t_{n+1})=e^{\tau A}\rho(t_n)e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_{0}^{\tau}e^{(\tau-s)A}L_k\rho(t_n+s)L_k^\dagger e^{(\tau-s)A^\dagger}\,\mathrm{d}s. \tag*{(5)}

Applying this formula again for ρ(tn+s)\rho(t_n+s) within the integral, we obtain

ρ(tn+1)=eτAρ(tn)eτA†+∑k=1Kγk∫0τe(τ−s)ALkesAρ(tn)esA†Lk†e(τ−s)A† ds+Rn,1,(6)\rho(t_{n+1})=e^{\tau A}\rho(t_n)e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_{0}^{\tau}e^{(\tau-s)A}L_ke^{sA}\rho(t_n)e^{sA^\dagger}L_k^\dagger e^{(\tau-s)A^\dagger}\,\mathrm{d}s+R_{n,1}, \tag*{(6)}

where

Rn,1=∑k=1K∑j=1Kγkγj∫0τ(e(τ−s)ALk(∫0se(s−v)ALjρ(tn+v)Lj†e(s−v)A† dv)Lk†e(τ−s)A†) ds.(7)R_{n,1}=\sum_{k=1}^{K}\sum_{j=1}^{K}\gamma_k\gamma_j\int_{0}^{\tau}\left(e^{(\tau-s)A}L_k\left(\int_{0}^{s}e^{(s-v)A}L_j\rho(t_n+v)L_j^\dagger e^{(s-v)A^\dagger}\,\mathrm{d}v\right)L_k^\dagger e^{(\tau-s)A^\dagger}\right)\,\mathrm{d}s. \tag*{(7)}

By truncating the term Rn,1R_{n,1} and approximating the matrix exponentials by e(τ−s)A≈Ie^{(\tau-s)A}\approx I and e(τ−s)A†≈Ie^{(\tau-s)A^\dagger}\approx I in (2.4), we obtain our full-rank exponential Euler (FREE) integrator, for the approximation {ρn}n=0N\{\rho_n\}_{n=0}^{N} to {ρ(tn)}n=0N\{\rho(t_n)\}_{n=0}^{N}, as follows:

ρn+1=eτAρneτA†+∑k=1Kγk∫0τLkesAρnesA†Lk† ds,(8)\rho_{n+1}=e^{\tau A}\rho_n e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_{0}^{\tau}L_ke^{sA}\rho_n e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s, \tag*{(8)}

for 0≤n≤N−10\leq n\leq N-1.

Remark 2.1. If we make the approximation ρ(tn+s)≈ρn\rho(t_n+s)\approx\rho_n within the integral in (2.3), then we obtain the “standard” exponential Euler scheme for solving (1.1) given by

ρn+1=eτAρneτA†+∑k=1Kγk∫0τe(τ−s)ALkρnLk†e(τ−s)A† ds=eτAρneτA†+∑k=1Kγk∫0τesALkρnLk†esA† ds.(9)\begin{aligned} \rho_{n+1}=e^{\tau A}\rho_n e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_{0}^{\tau}e^{(\tau-s)A}L_k\rho_nL_k^\dagger e^{(\tau-s)A^\dagger}\,\mathrm{d}s \\ &=e^{\tau A}\rho_n e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_{0}^{\tau}e^{sA}L_k\rho_nL_k^\dagger e^{sA^\dagger}\,\mathrm{d}s. \tag*{(9)} \end{aligned}

Note that the “standard” exponential Euler scheme (2.7) can preserve the positive semidefinite property and is first-order convergent. However, this scheme can not preserve unit trace of the Lindblad equation. Similar “standard” exponential integrators have been proposed and analyzed for solving differential matrix Riccati equations in [29, 30]. Such matrix-valued exponential integrators are natural extensions of the vector-valued exponential integrators, for which we refer to the review [32] for details.

Remark 2.2. We remark that the exponential Euler integrator (2.6) can be extended straightforwardly to solve Lindblad equations with time-dependent Hamiltonians H(t)H(t), in which case the exponential Euler scheme reads

ρn+1=eτAnρneτAn†+∑k=1Kγk∫0τLkesAnρnesAn†Lk† ds.(10)\rho_{n+1}=e^{\tau A_n}\rho_n e^{\tau A_n^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA_n}\rho_n e^{sA_n^\dagger}L_k^\dagger\,\mathrm{d}s. \tag*{(10)}

where

An=−iH(tn)−12∑k=1KγkLk†Lk.A_n=-\mathrm{i}H(t_n)-\frac{1}{2}\sum_{k=1}^{K}\gamma_k L_k^\dagger L_k.

Since the analysis of this scheme is similar to the one with time-independent Hamiltonian, for ease of presentation, we will only present numerical analysis for the FREE scheme (2.6).

Now, we briefly illustrate the practical implementation of the proposed exponential Euler scheme (2.6). Define

Wn=∫0τesAρnesA† ds.W_n=\int_0^\tau e^{sA}\rho_n e^{sA^\dagger}\,\mathrm{d}s.

By using the properties of Kronecker products, we can show that WnW_n is the solution of the following algebraic Lyapunov equation (see [29, 30] for details)

AWn+WnA†=eτAρneτA†−ρn.AW_n+W_nA^\dagger=e^{\tau A}\rho_n e^{\tau A^\dagger}-\rho_n.

It follows that the exponential Euler integrator (2.6) is equivalent to

AWn+WnA†=eτAρneτA†−ρn,(11)AW_n+W_nA^\dagger=e^{\tau A}\rho_n e^{\tau A^\dagger}-\rho_n, \tag*{(11)}
ρn+1=eτAρneτA†+∑k=1KγkLkWnLk†.(12)\rho_{n+1}=e^{\tau A}\rho_n e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k L_kW_nL_k^\dagger. \tag*{(12)}

Note that at every time step of the exponential Euler integrator (2.9), we need to solve an algebraic Lyapunov equation and compute 2m2m matrix-vector products associated with matrix exponential eτAe^{\tau A}.

Positivity and trace preserving properties

In this section, we study the positivity and trace preserving properties of the exponential Euler integrator (2.6). Throughout the paper, we use the following notation. For any matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m\times m}, we denote Tr⁡(ϱ)\operatorname{Tr}(\varrho) the trace of the matrix ϱ\varrho. For the trace, we have the cyclic property that Tr⁡(BCD)=Tr⁡(CDB)=Tr⁡(DBC)\operatorname{Tr}(BCD)=\operatorname{Tr}(CDB)=\operatorname{Tr}(DBC). We write ϱ≥0\varrho\ge0 if the matrix ϱ\varrho is Hermitian and positive semidefinite.

For any matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m\times m}, the trace norm, or the Schatten-1 norm, is defined as ∥ϱ∥1=Tr⁡(ϱ†ϱ)=∑j=1mσj(ϱ)\|\varrho\|_1=\operatorname{Tr}\left(\sqrt{\varrho^\dagger\varrho}\right)=\sum_{j=1}^{m}\sigma_j(\varrho), where σ1(ϱ)≥σ2(ϱ)≥⋯≥σm(ϱ)\sigma_1(\varrho)\ge\sigma_2(\varrho)\ge\cdots\ge\sigma_m(\varrho) are the singular values of ϱ\varrho. If ϱ≥0\varrho\ge0, we have that ∥ϱ∥1=Tr⁡(ϱ)\|\varrho\|_1=\operatorname{Tr}(\varrho). We denote ∥ϱ∥F\|\varrho\|_F the Frobenius norm of ϱ\varrho.

Now, we study the positivity preserving property of the proposed FREE integrator.

Theorem 3.1. Assume that ρ0\rho_0 is Hermitian and positive semidefinite. Then, for any time step size τ>0\tau> 0, the solution of the full-rank exponential Euler scheme (2.6) given by {ρn}n=0N\{\rho_n\}_{n=0}^{N} is Hermitian and positive semidefinite.

Proof. We prove the theorem by induction. Assume that the result holds for n=jn = j, i.e. ρj\rho_j is Hermitian and positive semidefinite. Next, we prove that the result holds for n=j+1n = j+1. Since ρj\rho_j is Hermitian and positive semidefinite, we see that both terms on the right-hand side of (2.6) are Hermitian and positive semidefinite. It follows that ρj+1\rho_{j+1} is also Hermitian and positive semidefinite, which completes the proof. □

The following result will play a key role in the analysis of the trace preserving property of the exponential Euler scheme.

Lemma 3.2. For any matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m \times m} and any t≥0t \ge0, it holds that

Tr⁡(etAϱetA†+∑k=1Kγk∫0tLkesAϱesA†Lk† ds)=Tr⁡(ϱ).(13)\operatorname{Tr}\left(e^{tA}\varrho e^{tA^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^t L_k e^{sA}\varrho e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s\right)=\operatorname{Tr}(\varrho). \tag*{(13)}

Proof. Note that the solution of the differential equation

σ˙=Aσ+σA†,σ(0)=ϱ,(14)\dot{\sigma}=A\sigma+\sigma A^\dagger,\qquad\sigma(0)=\varrho, \tag*{(14)}

is given by

σ(t)=etAϱetA†.(15)\sigma(t)=e^{tA}\varrho e^{tA^\dagger}. \tag*{(15)}

Using the definition of AA in (2.1), we obtain that the differential equation (3.2) is equivalent to

σ˙=−iHσ+iσH−12∑k=1Kγk(Lk†Lkσ+σLk†Lk),σ(0)=ϱ.\dot{\sigma}=-iH\sigma+i\sigma H-\frac{1}{2}\sum_{k=1}^{K}\gamma_k\left(L_k^\dagger L_k\sigma+\sigma L_k^\dagger L_k\right),\qquad\sigma(0)=\varrho.

By the variation-of-constants formula, we have

σ(t)=e−itHϱeitH−12∑k=1Kγk∫0te−i(t−s)H(Lk†Lkσ(s)+σ(s)Lk†Lk)ei(t−s)H ds.\sigma(t)=e^{-itH}\varrho e^{itH}-\frac{1}{2}\sum_{k=1}^{K}\gamma_k\int_0^t e^{-i(t-s)H}\left(L_k^\dagger L_k\sigma(s)+\sigma(s)L_k^\dagger L_k\right)e^{i(t-s)H}\,\mathrm{d}s.

Using the cyclic property of the trace and (15), we obtain

Tr⁡(etAϱetA†)=Tr⁡(σ(t))=Tr⁡(e−itHϱeitH)−12∑k=1Kγk∫0tTr⁡(e−i(t−s)H(Lk†Lkσ(s)+σ(s)Lk†Lk)ei(t−s)H) ds=Tr⁡(ϱ)−12∑k=1Kγk∫0tTr⁡(Lk†Lkσ(s)+σ(s)Lk†Lk) ds=Tr⁡(ϱ)−∑k=1Kγk∫0tTr⁡(Lk†Lkσ(s)) ds=Tr⁡(ϱ)−∑k=1Kγk∫0tTr⁡(Lkσ(s)Lk†) ds=Tr⁡(ϱ)−∑k=1Kγk∫0tTr⁡(LkesAϱesA†Lk†) ds.(16)\begin{aligned} \operatorname{Tr}\left(e^{tA}\varrho e^{tA^\dagger}\right) &= \operatorname{Tr}(\sigma(t))=\operatorname{Tr}\left(e^{-itH}\varrho e^{itH}\right) \\ &\quad-\frac{1}{2}\sum_{k=1}^{K}\gamma_k\int_0^t \operatorname{Tr}\left(e^{-i(t-s)H}\left(L_k^\dagger L_k\sigma(s)+\sigma(s)L_k^\dagger L_k\right)e^{i(t-s)H}\right)\,\mathrm{d}s \\ &= \operatorname{Tr}(\varrho)-\frac{1}{2}\sum_{k=1}^{K}\gamma_k\int_0^t \operatorname{Tr}\left(L_k^\dagger L_k\sigma(s)+\sigma(s)L_k^\dagger L_k\right)\,\mathrm{d}s \\ &= \operatorname{Tr}(\varrho)-\sum_{k=1}^{K}\gamma_k\int_0^t \operatorname{Tr}\left(L_k^\dagger L_k\sigma(s)\right)\,\mathrm{d}s \\ &= \operatorname{Tr}(\varrho)-\sum_{k=1}^{K}\gamma_k\int_0^t \operatorname{Tr}\left(L_k\sigma(s)L_k^\dagger\right)\,\mathrm{d}s \\ &= \operatorname{Tr}(\varrho)-\sum_{k=1}^{K}\gamma_k\int_0^t \operatorname{Tr}\left(L_k e^{sA}\varrho e^{sA^\dagger}L_k^\dagger\right)\,\mathrm{d}s. \tag*{(16)} \end{aligned}

Then, the desired result (13) follows. □

Now, we are in the position to give the following theorem.

Theorem 3.3. Assume that the initial data satisfies Tr⁡(ρ0)=1\operatorname{Tr}(\rho_0)=1. Then, for any time step size τ>0\tau>0, the full-rank exponential Euler integrator (8) preserves unit trace in the sense that

Tr⁡(ρn)=1,n≥0.\operatorname{Tr}(\rho_n)=1,\qquad n\geq0.

Proof. We prove the theorem by induction. We have Tr⁡(ρ0)=1\operatorname{Tr}(\rho_0)=1. Assume that the result holds for n=jn=j, i.e., Tr⁡(ρj)=1\operatorname{Tr}(\rho_j)=1. Next we prove the result holds for n=j+1n=j+1. Using (8) and applying Lemma 3.2, we have Tr⁡(ρj+1)=Tr⁡(ρj)\operatorname{Tr}(\rho_{j+1})=\operatorname{Tr}(\rho_j), which completes the proof. □

Remark 3.4. Note from Theorems 3.1 and 3.3 that ∥ρn∥1=Tr⁡(ρn)=1\|\rho_n\|_1=\operatorname{Tr}(\rho_n)=1 for any n≥0n\geq0 and any τ>0\tau>0, which indicates that the proposed exponential Euler integrator (8) is unconditionally stable.

Accuracy of approximation

In this section, we investigate the convergence of the exponential Euler scheme (8). First, we present some preparatory results that will be useful in the error analysis.

Lemma 4.1. For any matrix B∈Cm×mB\in\mathbb{C}^{m\times m} and any Hermitian matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m\times m}, it holds that

∥BϱB†∥1≤∥B∣ϱ∣B†∥1,\|B\varrho B^\dagger\|_1\leq\|B|\varrho|B^\dagger\|_1,

where ∣ϱ∣=ϱ†ϱ|\varrho|=\sqrt{\varrho^\dagger\varrho}.

Proof. Note that for any Hermitian matrix ϱ\varrho, we have the spectral decomposition ϱ=QDQ†\varrho= QDQ^\dagger, where QQ is unitary and DD is diagonal. Note that ∣ϱ∣=Q∣D∣Q†|\varrho| = Q|D|Q^\dagger. So we only need to prove that the following inequality

∥PDP†∥1≤∥P∣D∣P†∥1\lVert PDP^\dagger\rVert_1 \leq\lVert P|D|P^\dagger\rVert_1

holds for any P=[p1,…,pm]∈Cm×mP = [p_1,\ldots,p_m] \in\mathbb{C}^{m \times m} and any diagonal matrix D=diag⁡(d1,…,dm)∈Rm×mD = \operatorname{diag}(d_1,\ldots,d_m) \in\mathbb{R}^{m \times m}.

Since PDP†PDP^\dagger is Hermitian, we have the spectral decomposition PDP†=UΛU†PDP^\dagger= U\Lambda U^\dagger, where U=[u1,…,um]U = [u_1,\ldots,u_m] is unitary and Λ=diag⁡(λ1,…,λm)\Lambda= \operatorname{diag}(\lambda_1,\ldots,\lambda_m) is the diagonal matrix containing eigenvalues of PDP†PDP^\dagger. Using λj=uj†PDP†uj\lambda_j = u_j^\dagger PDP^\dagger u_j, PDP†=∑k=1mdkpkpk†PDP^\dagger= \sum_{k=1}^{m} d_k p_k p_k^\dagger, and P∣D∣P†=∑k=1m∣dk∣pkpk†P|D|P^\dagger= \sum_{k=1}^{m} |d_k|p_kp_k^\dagger, we get

∥PDP†∥1=∑j=1mσj(PDP†)=∑j=1m∣λj∣=∑j=1m∣uj†PDP†uj∣=∑j=1m∣uj†(∑k=1mdkpkpk†)uj∣=∑j=1m∣∑k=1mdk(pk†uj)†(pk†uj)∣≤∑j=1m∑k=1m∣dk∣(pk†uj)†(pk†uj)=∑j=1muj†(∑k=1m∣dk∣pkpk†)uj=∑j=1muj†P∣D∣P†uj=Tr⁡(U†P∣D∣P†U)=Tr⁡(P∣D∣P†)=∥P∣D∣P†∥1,\begin{aligned} \lVert PDP^\dagger\rVert_1 &= \sum_{j=1}^{m}\sigma_j(PDP^\dagger) = \sum_{j=1}^{m}|\lambda_j| = \sum_{j=1}^{m}\left|u_j^\dagger PDP^\dagger u_j\right| \\ &= \sum_{j=1}^{m}\left|u_j^\dagger\left(\sum_{k=1}^{m}d_kp_kp_k^\dagger\right)u_j\right| \\ &= \sum_{j=1}^{m}\left|\sum_{k=1}^{m}d_k(p_k^\dagger u_j)^\dagger(p_k^\dagger u_j)\right| \\ &\leq\sum_{j=1}^{m}\sum_{k=1}^{m}|d_k|(p_k^\dagger u_j)^\dagger(p_k^\dagger u_j) \\ &= \sum_{j=1}^{m}u_j^\dagger\left(\sum_{k=1}^{m}|d_k|p_kp_k^\dagger\right)u_j \\ &= \sum_{j=1}^{m}u_j^\dagger P|D|P^\dagger u_j = \operatorname{Tr}(U^\dagger P|D|P^\dagger U) \\ &= \operatorname{Tr}(P|D|P^\dagger) = \lVert P|D|P^\dagger\rVert_1, \end{aligned}

and the required result follows. □\square

Lemma 4.2. For any Hermitian matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m \times m}, it holds that

∥etAϱetA†∥1≤∥ϱ∥1,t≥0.\left\lVert e^{tA}\varrho e^{tA^\dagger}\right\rVert_1 \leq\lVert\varrho\rVert_1,\quad t \geq0.

Proof. Let ∣ϱ∣=ϱ†ϱ|\varrho| = \sqrt{\varrho^\dagger\varrho} and we note that ∣ϱ∣≥0|\varrho| \geq0 and ∥ϱ∥1=Tr⁡(∣ϱ∣)\lVert\varrho\rVert_1 = \operatorname{Tr}(|\varrho|). For any t≥0t \geq0,

applying Lemma 4.1 and Lemma 3.2 yields

∥etAϱetA†∥1≤∥etA∣ϱ∣etA†∥1=Tr⁡(etA∣ϱ∣etA†)=Tr⁡(∣ϱ∣)−∑k=1Kγk∫0tTr⁡(LkesA∣ϱ∣esA†Lk†) ds=∥ϱ∥1−∑k=1Kγk∫0t∥LkesA∣ϱ∣esA†Lk†∥1 ds≤∥ϱ∥1.\begin{aligned} \left\|e^{tA}\varrho e^{tA^\dagger}\right\|_1 &\le\left\|e^{tA}|\varrho|e^{tA^\dagger}\right\|_1 \\ &= \operatorname{Tr}\left(e^{tA}|\varrho|e^{tA^\dagger}\right) \\ &= \operatorname{Tr}(|\varrho|)-\sum_{k=1}^{K}\gamma_k\int_0^t\operatorname{Tr}\left(L_ke^{sA}|\varrho|e^{sA^\dagger}L_k^\dagger\right)\,ds \\ &= \|\varrho\|_1-\sum_{k=1}^{K}\gamma_k\int_0^t\left\|L_ke^{sA}|\varrho|e^{sA^\dagger}L_k^\dagger\right\|_1\,ds \\ &\le\|\varrho\|_1. \end{aligned}

The conclusion follows. □

Lemma 4.3. For any Hermitian matrix ϱ∈Cm×m\varrho\in\mathbb{C}^{m\times m} and any t≥0t \ge0, it holds that

∥etAϱetA†+∑k=1Kγk∫0tLkesAϱesA†Lk† ds∥1≤∥ϱ∥1.(17)\left\|e^{tA}\varrho e^{tA^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^t L_ke^{sA}\varrho e^{sA^\dagger}L_k^\dagger\,ds\right\|_1\le\|\varrho\|_1. \tag*{(17)}

Proof. Using Lemma 4.1 and noting that ∣ϱ∣≥0|\varrho| \ge0, we have

∥etAϱetA†+∑k=1Kγk∫0tLkesAϱesA†Lk† ds∥1≤∥etAϱetA†∥1+∑k=1Kγk∫0t∥LkesAϱesA†Lk†∥1 ds≤∥etA∣ϱ∣etA†∥1+∑k=1Kγk∫0t∥LkesA∣ϱ∣esA†Lk†∥1 ds=Tr⁡(etA∣ϱ∣etA†)+∑k=1Kγk∫0tTr⁡(LkesA∣ϱ∣esA†Lk†) ds=Tr⁡(etA∣ϱ∣etA†+∑k=1Kγk∫0tLkesA∣ϱ∣esA†Lk† ds)=Tr⁡(∣ϱ∣),\begin{aligned} \left\|e^{tA}\varrho e^{tA^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^t L_ke^{sA}\varrho e^{sA^\dagger}L_k^\dagger\,ds\right\|_1 \le\left\|e^{tA}\varrho e^{tA^\dagger}\right\|_1+\sum_{k=1}^{K}\gamma_k\int_0^t\left\|L_ke^{sA}\varrho e^{sA^\dagger}L_k^\dagger\right\|_1\,ds \\ &\le\left\|e^{tA}|\varrho|e^{tA^\dagger}\right\|_1+\sum_{k=1}^{K}\gamma_k\int_0^t\left\|L_ke^{sA}|\varrho|e^{sA^\dagger}L_k^\dagger\right\|_1\,ds \\ &= \operatorname{Tr}\left(e^{tA}|\varrho|e^{tA^\dagger}\right)+\sum_{k=1}^{K}\gamma_k\int_0^t\operatorname{Tr}\left(L_ke^{sA}|\varrho|e^{sA^\dagger}L_k^\dagger\right)\,ds \\ &= \operatorname{Tr}\left(e^{tA}|\varrho|e^{tA^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^tL_ke^{sA}|\varrho|e^{sA^\dagger}L_k^\dagger\,ds\right) \\ &= \operatorname{Tr}(|\varrho|), \end{aligned}

where the last equality follows from Lemma 3.2. The desired result follows since Tr⁡(∣ϱ∣)=∥ϱ∥1\operatorname{Tr}(|\varrho|)=\|\varrho\|_1. □

For the discussion that follows, we define

C1:=∑k=1Kγk∥Lk∥12,C2:=∥A∥1.C_1:=\sum_{k=1}^{K}\gamma_k\|L_k\|_1^2,\qquad C_2:=\|A\|_1.

Now, we are ready to prove our error estimate of the full-rank exponential Euler scheme (8). We have

Theorem 4.4. Let ρ(t)\rho(t) be the solution of the Lindblad equation (1) with ρ(0)=ρ0\rho(0)=\rho_0, and {ρn}n=0N\{\rho_n\}_{n=0}^{N} be the corresponding numerical solution generated by the full-rank exponential Euler scheme (8). Then, it holds that

∥ρ(tn)−ρn∥1≤c1tnτ,0≤n≤N,\lVert\rho(t_n)-\rho_n\rVert_1 \le c_1 t_n \tau,\qquad0\le n\le N,

where the constant c1>0c_1>0 depends on C1C_1, C2C_2, TT, but is independent of τ\tau and nn.

Proof. Expanding e(τ−s)Ae^{(\tau-s)A} in a Taylor series with remainder in integral form yields

e(τ−s)A=I+∫sτAe(τ−v)A dv.e^{(\tau-s)A}=I+\int_s^\tau Ae^{(\tau-v)A}\,\mathrm{d}v.

Inserting the above equation into (6), we have

ρ(tn+1)=eτAρ(tn)eτA†+∑k=1Kγk∫0τLkesAρ(tn)esA†Lk† ds+∑j=14Rn,j,(18)\rho(t_{n+1})=e^{\tau A}\rho(t_n)e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA}\rho(t_n)e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s+\sum_{j=1}^{4}R_{n,j}, \tag*{(18)}

where Rn,1R_{n,1} is defined in (7), and

Rn,2=∑k=1Kγk∫0τ(∫sτAe(τ−v)A dv)LkesAρ(tn)esA†Lk† ds,R_{n,2}=\sum_{k=1}^{K}\gamma_k\int_0^\tau\left(\int_s^\tau Ae^{(\tau-v)A}\,\mathrm{d}v\right)L_k e^{sA}\rho(t_n)e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s,
Rn,3=∑k=1Kγk∫0τLkesAρ(tn)esA†Lk†(∫sτA†e(τ−v)A† dv)ds,R_{n,3}=\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA}\rho(t_n)e^{sA^\dagger}L_k^\dagger\left(\int_s^\tau A^\dagger e^{(\tau-v)A^\dagger}\,\mathrm{d}v\right)\mathrm{d}s,
Rn,4=∑k=1Kγk∫0τ(∫sτAe(τ−v)A dv)LkesAρ(tn)esA†Lk†(∫sτA†e(τ−v)A† dv)ds.R_{n,4}=\sum_{k=1}^{K}\gamma_k\int_0^\tau\left(\int_s^\tau Ae^{(\tau-v)A}\,\mathrm{d}v\right)L_k e^{sA}\rho(t_n)e^{sA^\dagger}L_k^\dagger\left(\int_s^\tau A^\dagger e^{(\tau-v)A^\dagger}\,\mathrm{d}v\right)\mathrm{d}s.

Denote the global error En:=ρ(tn)−ρnE_n:=\rho(t_n)-\rho_n. Subtracting (8) from (18), we obtain the error recursion

En+1=eτAEneτA†+∑k=1Kγk∫0τLkesAEnesA†Lk† ds+Rn,(19)E_{n+1}=e^{\tau A}E_ne^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA}E_n e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s+R_n, \tag*{(19)}

where Rn=∑j=14Rn,jR_n=\sum_{j=1}^{4}R_{n,j}.

Using Lemma 4.2 and noting that ∥ρ(t)∥1=1\|\rho(t)\|_1=1 for any t≥0t\ge0, we have

∥Rn,1∥1≤∑k,j=1Kγkγj∫0τ∥e(τ−s)ALk(∫0se(s−v)ALjρ(tn+v)Lj†e(s−v)A† dv)Lk†e(τ−s)A†∥1 ds≤∑k,j=1Kγkγj∥Lk∥12∫0τ∫0s∥e(s−v)ALjρ(tn+v)Lj†e(s−v)A†∥1 dv ds≤∑k,j=1Kγkγj∥Lk∥12∥Lj∥12∫0τ∫0s∥ρ(tn+v)∥1 dv ds=12τ2∑k,j=1Kγkγj∥Lk∥12∥Lj∥12=12C12τ2.(20)\begin{aligned} \|R_{n,1}\|_1 &\le\sum_{k,j=1}^{K}\gamma_k\gamma_j\int_0^\tau\left\|e^{(\tau-s)A}L_k\left(\int_0^s e^{(s-v)A}L_j\rho(t_n+v)L_j^\dagger e^{(s-v)A^\dagger}\,dv\right)L_k^\dagger e^{(\tau-s)A^\dagger}\right\|_1\,ds\\ &\le\sum_{k,j=1}^{K}\gamma_k\gamma_j\|L_k\|_1^2\int_0^\tau\int_0^s\left\|e^{(s-v)A}L_j\rho(t_n+v)L_j^\dagger e^{(s-v)A^\dagger}\right\|_1\,dv\,ds\\ &\le\sum_{k,j=1}^{K}\gamma_k\gamma_j\|L_k\|_1^2\|L_j\|_1^2\int_0^\tau\int_0^s\|\rho(t_n+v)\|_1\,dv\,ds\\ &=\frac{1}{2}\tau^2\sum_{k,j=1}^{K}\gamma_k\gamma_j\|L_k\|_1^2\|L_j\|_1^2=\frac{1}{2}C_1^2\tau^2. \tag*{(20)} \end{aligned}

Similarly, we can obtain

∥Rn,2∥1≤12C1C2τ2,∥Rn,3∥1≤12C1C2τ2,(21)\|R_{n,2}\|_1\le\frac{1}{2}C_1C_2\tau^2,\qquad\|R_{n,3}\|_1\le\frac{1}{2}C_1C_2\tau^2, \tag*{(21)}

and

∥Rn,4∥1≤13C1C22τ3.(22)\|R_{n,4}\|_1\le\frac{1}{3}C_1C_2^2\tau^3. \tag*{(22)}

Combining (20)–(22), we have

∥Rn∥1≤c1τ2,(23)\|R_n\|_1\le c_1\tau^2, \tag*{(23)}

where c1=12C12+C1C2+13C1C22Tc_1=\frac{1}{2}C_1^2+C_1C_2+\frac{1}{3}C_1C_2^2T.

It follows from (19), (23), E0=0E_0=0 and Lemma 4.3 that

∥En+1∥1≤∥eτAEneτA†+∑k=1Kγk∫0τLkesAEnesA†Lk† ds∥1+∥Rn∥1≤∥En∥1+c1τ2≤∥En−1∥1+2c1τ2≤∥E0∥1+c1(n+1)τ2=c1tn+1τ,\begin{aligned} \|E_{n+1}\|_1 &\le\left\|e^{\tau A}E_ne^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_ke^{sA}E_ne^{sA^\dagger}L_k^\dagger\,ds\right\|_1+\|R_n\|_1\\ &\le\|E_n\|_1+c_1\tau^2\le\|E_{n-1}\|_1+2c_1\tau^2\\ &\le\|E_0\|_1+c_1(n+1)\tau^2=c_1t_{n+1}\tau, \end{aligned}

which completes the proof. □

A low-rank exponential Euler integrator

When the Lindblad equation is large-sized, i.e. m≫1m\gg1, the computational cost of the full-rank exponential Euler integrator (8) would be expensive since algebraic Lyapunov equations with dimension mm need to be solved and 2m2m matrix-vector products associated with eτAe^{\tau A} are required at every time step. Now, in order to reduce computational costs while controlling accuracy, we develop a low-rank variant of the FREE integrator (8). The aim is to find matrices Zn∈Cm×rnZ_n \in\mathbb{C}^{m \times r_n} with rn≪mr_n \ll m such that the solution of the Lindblad equation can be well approximated as

ρ(tn)≈ZnZn†=:ϱn,\rho(t_n) \approx Z_n Z_n^\dagger=: \varrho_n,

where we denote with ϱn\varrho_n the numerical low-rank solution to the Lindblad equation in order to distinguish it from ρn\rho_n, the full-rank numerical solution of the same equation.

Let ρ0≈ϱ0=Z0Z0†\rho_0 \approx\varrho_0 = Z_0 Z_0^\dagger with Z0∈Cm×r0Z_0 \in\mathbb{C}^{m \times r_0} and Tr⁡(ϱ0)=1\operatorname{Tr}(\varrho_0) = 1. To derive our low-rank exponential Euler scheme, we first define

ϱ~n+1=eτAϱneτA†+∑k=1KγkLkW~nLk†,n=0,…,N−1,(24)\widetilde{\varrho}_{n+1} = e^{\tau A}\varrho_n e^{\tau A^\dagger} + \sum_{k=1}^{K} \gamma_k L_k \widetilde{W}_n L_k^\dagger,\qquad n=0,\ldots,N-1, \tag*{(24)}

where

W~n=∫0τesAϱnesA† ds.(25)\widetilde{W}_n = \int_0^\tau e^{sA}\varrho_n e^{sA^\dagger}\,\mathrm{d}s. \tag*{(25)}

Instead of computing W~n\widetilde{W}_n by solving an algebraic Lyapunov equation as in (11), we approximate the integral in (25) by the following quadrature formula

∫0τesAϱnesA† ds≈τeτAϱneτA†=τeτAZnZn†eτA†=:τVnVn†,(26)\int_0^\tau e^{sA}\varrho_n e^{sA^\dagger}\,\mathrm{d}s \approx\tau e^{\tau A}\varrho_n e^{\tau A^\dagger} = \tau e^{\tau A}Z_nZ_n^\dagger e^{\tau A^\dagger} =: \tau V_nV_n^\dagger, \tag*{(26)}

where Vn=eτAZnV_n=e^{\tau A}Z_n. It then follows from (24) that

ϱ~n+1≈eτAZnZn†eτA†+∑k=1KγkτLkVnVn†Lk†=:Z~n+1Z~n+1†,\widetilde{\varrho}_{n+1} \approx e^{\tau A}Z_nZ_n^\dagger e^{\tau A^\dagger} + \sum_{k=1}^{K}\gamma_k\tau L_kV_nV_n^\dagger L_k^\dagger=: \widetilde{Z}_{n+1}\widetilde{Z}_{n+1}^\dagger,

where

Z~n+1=[Vn,γ1τL1Vn,…,γKτLKVn].\widetilde{Z}_{n+1} = \left[V_n,\sqrt{\gamma_1\tau}L_1V_n,\ldots,\sqrt{\gamma_K\tau}L_KV_n\right].

By this notation we mean that the matrix VnV_n and the KK matrices γkτLkVn\sqrt{\gamma_k\tau}L_kV_n are placed side by side.

Note that for many problems, the matrix exponential eτAe^{\tau A} or the product of matrix exponential times vectors eτAZne^{\tau A}Z_n may be costly to compute and approximations may be required. In the low-rank algorithm, we denote with eτA~\widetilde{e^{\tau A}} (resp. eτAZn~\widetilde{e^{\tau A}Z_n}) an approximation of eτAe^{\tau A} (resp. eτAZne^{\tau A}Z_n). We will always assume that ∥eτA~σeτA†~−eτAσeτA†∥1≤C3 tol1∥σ∥1\left\|\widetilde{e^{\tau A}}\sigma\widetilde{e^{\tau A^\dagger}}-e^{\tau A}\sigma e^{\tau A^\dagger}\right\|_1 \le C_3\,tol_1\|\sigma\|_1 for any σ∈Cm×m\sigma\in\mathbb{C}^{m\times m}, where tol1>0tol_1>0 is the error tolerance of the underlying matrix exponential algorithm and C3>0C_3>0. In addition, note that the matrix Z~n+1\widetilde{Z}_{n+1} has more columns than ZnZ_n, and likely also more than its rank. A better low-rank approximation can be obtained by applying a column compression technique [33], which can be done by truncating the singular value decomposition (SVD) of the given matrix. We denote with Ttol2(⋅)\mathcal{T}_{tol_2}(\cdot) the truncated SVD of a matrix with error tolerance tol2>0tol_2>0, that is, Ttol2(Z)\mathcal{T}_{tol_2}(Z) represents the best rank rr approximation of the matrix Z∈Cm×qZ \in\mathbb{C}^{m\times q} in Frobenius norm, where rr is the minimal integer such that ∑j=r+1qσj2(Z)≤tol2\sum_{j=r+1}^{q}\sigma_j^2(Z)\le tol_2. We then get

∥ZZ†−Ttol2(Z)Ttol2(Z)†∥1=∑j=r+1qσj2(Z)≤tol2.\left\|ZZ^\dagger-\mathcal{T}_{tol_2}(Z)\mathcal{T}_{tol_2}(Z)^\dagger\right\|_1=\sum_{j=r+1}^{q}\sigma_j^2(Z)\le tol_2.

To summarize, our low-rank exponential Euler (LREE) scheme reads as follows:

Vn=eτAZn,(27)V_n=e^{\tau A}Z_n, \tag*{(27)}
Z~n+1=[Vn,γ1τL1Vn,…,γKτLKVn],(28)\widetilde{Z}_{n+1}=\left[V_n,\sqrt{\gamma_1\tau}L_1V_n,\ldots,\sqrt{\gamma_K\tau}L_KV_n\right], \tag*{(28)}
Z^n+1=Ttol2(Z~n+1),(29)\widehat{Z}_{n+1}=\mathcal{T}_{\mathrm{tol}_2}(\widetilde{Z}_{n+1}), \tag*{(29)}
Zn+1=Z^n+1∥Z^n+1∥F.(30)Z_{n+1}=\frac{\widehat{Z}_{n+1}}{\left\|\widehat{Z}_{n+1}\right\|_F}. \tag*{(30)}

Note that the LREE integrator requires to approximate rnr_n matrix-exponential-vector products and compute a SVD of a tall matrix with (K+1)rn(K+1)r_n columns at every time step. The computational cost is much cheaper than that of the FREE scheme, which requires to compute 2m2m (2m≫rn2m\gg r_n) matrix-exponential-vector products and solve an algebraic Lyapunov equation of dimension mm.

Remark 5.1. Note from ϱn+1=Zn+1Zn+1†\varrho_{n+1}=Z_{n+1}Z_{n+1}^{\dagger} and (30) that the LREE scheme (5.4) is positivity and trace preserving:

∥ϱn+1∥1=∥Zn+1Zn+1†∥1=Tr⁡(Zn+1Zn+1†)=Tr⁡(Z^n+1Z^n+1†)∥Z^n+1∥F2=1,n=0,…,N−1.(31)\left\|\varrho_{n+1}\right\|_1=\left\|Z_{n+1}Z_{n+1}^{\dagger}\right\|_1=\operatorname{Tr}(Z_{n+1}Z_{n+1}^{\dagger})=\frac{\operatorname{Tr}(\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{\dagger})}{\left\|\widehat{Z}_{n+1}\right\|_F^2}=1,\qquad n=0,\ldots,N-1. \tag*{(31)}

However, in this case normalization is performed.

We remark that the LREE (5.4) is equivalent to

ϱ~n+1=eτAϱneτA†+∑k=1Kγk∫0τLkesAϱnesA†Lk† ds,(32)\widetilde{\varrho}_{n+1}=e^{\tau A}\varrho_n e^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA}\varrho_n e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s, \tag*{(32)}
ϱ^n+1=ϱ~n+1−ϑn+1,(33)\widehat{\varrho}_{n+1}=\widetilde{\varrho}_{n+1}-\vartheta_{n+1}, \tag*{(33)}
ϱn+1=ϱ^n+1Tr⁡(ϱ^n+1),(34)\varrho_{n+1}=\frac{\widehat{\varrho}_{n+1}}{\operatorname{Tr}(\widehat{\varrho}_{n+1})}, \tag*{(34)}

where ϱ^n+1=Z^n+1Z^n+1†\widehat{\varrho}_{n+1}=\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{\dagger} and the Hermitian matrix ϑn+1\vartheta_{n+1} can be seen as the perturbation caused by the numerical quadrature (26), the approximation to the matrix exponential times vector (27) and the column compression procedure (29). A bound on ϑn+1\vartheta_{n+1} is given in the following theorem.

Theorem 5.2. Assume that the matrix exponential algorithm used in the LREE scheme (5.4) satisfies ∥eτAσeτA†−eτAσeτA†∥1≤C3tol⁡1∥σ∥1\left\|e^{\tau A}\sigma e^{\tau A^\dagger}-e^{\tau A}\sigma e^{\tau A^\dagger}\right\|_1\leq C_3\operatorname{tol}_1\left\|\sigma\right\|_1 for any σ∈Cm×m\sigma\in\mathbb{C}^{m\times m} with tolerance tol⁡1>0\operatorname{tol}_1>0. Let tol⁡2>0\operatorname{tol}_2>0 be the error tolerance of the column compression algorithm used in the LREE scheme (5.4). Then, it holds that

∥ϑn+1∥1≤C1C2τ2+c2tol⁡1+tol⁡2,n=0,1,…,N−1,(35)\left\|\vartheta_{n+1}\right\|_1\leq C_1C_2\tau^2+c_2\operatorname{tol}_1+\operatorname{tol}_2,\qquad n=0,1,\ldots,N-1, \tag*{(35)}

where the constant c2>0c_2>0 depends on C1C_1, C3C_3, TT, but is independent of τ\tau, nn, tol⁡1\operatorname{tol}_1 and $\operatorname{tol}_2.

Proof. Let us first define

ϕn+1=eτAϱneτA†+τ∑k=1KγkLkeτAϱneτA†Lk†,\phi_{n+1}=e^{\tau A}\varrho_n e^{\tau A^\dagger}+\tau\sum_{k=1}^{K}\gamma_k L_k e^{\tau A}\varrho_n e^{\tau A^\dagger}L_k^\dagger,

and

φn+1=eτAϱneτA†+τ∑k=1KγkLkeτAϱneτA†Lk†.\varphi_{n+1}=e^{\tau A}\varrho_n e^{\tau A^\dagger}+\tau\sum_{k=1}^{K}\gamma_k L_k e^{\tau A}\varrho_n e^{\tau A^\dagger}L_k^\dagger.

Using (5.6b), we see that

∥ϑn+1∥1=∥ϱ~n+1−ϕn+1+ϕn+1−φn+1+φn+1−ϱ^n+1∥1≤∥ϱ~n+1−ϕn+1∥1+∥ϕn+1−φn+1∥1+∥φn+1−ϱ^n+1∥1.(36)\begin{aligned} \|\vartheta_{n+1}\|_1 &=\|\widetilde{\varrho}_{n+1}-\phi_{n+1}+\phi_{n+1}-\varphi_{n+1}+\varphi_{n+1}-\widehat{\varrho}_{n+1}\|_1\\ &\leq\|\widetilde{\varrho}_{n+1}-\phi_{n+1}\|_1+\|\phi_{n+1}-\varphi_{n+1}\|_1+\|\varphi_{n+1}-\widehat{\varrho}_{n+1}\|_1. \tag*{(36)} \end{aligned}

Note that the above three components in the perturbation ϑn+1\vartheta_{n+1} correspond to errors induced by the numerical quadrature, the approximation to the matrix exponential and the column compression procedure, respectively.

By comparing (5.4) with (5.6), we obtain φn+1=Z~n+1Z~n+1†\varphi_{n+1}=\widetilde{Z}_{n+1}\widetilde{Z}_{n+1}^{\dagger}, ϱ^n+1=Z^n+1Z^n+1†\widehat{\varrho}_{n+1}=\widehat{Z}_{n+1}\widehat{Z}_{n+1}^{\dagger} and Z^n+1=Ttol2(Z~n+1)\widehat{Z}_{n+1}=\mathcal{T}_{\mathrm{tol}_2}(\widetilde{Z}_{n+1}). Then, it follows that

∥φn+1−ϱ^n+1∥1=∥Z~n+1Z~n+1†−Ttol2(Z~n+1)Ttol2(Z~n+1)†∥1≤tol2.(37)\|\varphi_{n+1}-\widehat{\varrho}_{n+1}\|_1=\left\|\widetilde{Z}_{n+1}\widetilde{Z}_{n+1}^{\dagger}-\mathcal{T}_{\mathrm{tol}_2}(\widetilde{Z}_{n+1})\mathcal{T}_{\mathrm{tol}_2}(\widetilde{Z}_{n+1})^\dagger\right\|_1\leq\mathrm{tol}_2. \tag*{(37)}

To estimate the second component in (5.8), we define υn=eτAϱneτA†−eτAϱneτA†\upsilon_n=e^{\tau A}\varrho_n e^{\tau A^\dagger}-e^{\tau A}\varrho_n e^{\tau A^\dagger}. Using (5.5) and the assumption on the matrix exponential algorithm, we get

∥υn∥1≤C3tol1∥ϱn∥1=C3tol1.\|\upsilon_n\|_1\leq C_3\mathrm{tol}_1\|\varrho_n\|_1=C_3\mathrm{tol}_1.

Then, we obtain

∥ϕn+1−φn+1∥1=∥υn+τ∑k=1KγkLkυnLk†∥1≤(1+τ∑k=1Kγk∥Lk∥12)∥υn∥1≤(1+τC1)C3tol1≤c2tol1,(38)\begin{aligned} \|\phi_{n+1}-\varphi_{n+1}\|_1 &=\left\|\upsilon_n+\tau\sum_{k=1}^{K}\gamma_kL_k\upsilon_nL_k^\dagger\right\|_1 \leq\left(1+\tau\sum_{k=1}^{K}\gamma_k\|L_k\|_1^2\right)\|\upsilon_n\|_1\\ &\leq(1+\tau C_1)C_3\mathrm{tol}_1\leq c_2\mathrm{tol}_1, \tag*{(38)} \end{aligned}

where c2=(1+TC1)C3c_2=(1+TC_1)C_3.

Now, we estimate the first component in (5.8). Expanding esAϱnesA†e^{sA}\varrho_n e^{sA^\dagger} in Taylor series with remainder in integral form yields

esAϱnesA†=eτAϱneτA†−∫sτeμA(Aϱn+ϱnA†)eμA† dμ.e^{sA}\varrho_n e^{sA^\dagger}=e^{\tau A}\varrho_n e^{\tau A^\dagger}-\int_s^\tau e^{\mu A}(A\varrho_n+\varrho_n A^\dagger)e^{\mu A^\dagger}\,d\mu.

Defining

ωn=∫0τesAϱnesA† ds−τeτAϱneτA†,\omega_n=\int_0^\tau e^{sA}\varrho_n e^{sA^\dagger}\,ds-\tau e^{\tau A}\varrho_n e^{\tau A^\dagger},

and using (31) and Lemma (17), we obtain

∥ωn∥1=∥∫0τ∫sτeμA(Aϱn+ϱnA†)eμA† dμ ds∥1≤∫0τ∫sτ2∥A∥1 dμ ds=C2τ2.\|\omega_n\|_1=\left\|\int_0^\tau\int_s^\tau e^{\mu A}(A\varrho_n+\varrho_n A^\dagger)e^{\mu A^\dagger}\,d\mu\,ds\right\|_1\leq\int_0^\tau\int_s^\tau2\|A\|_1\,d\mu\,ds=C_2\tau^2.

By the definitions of ϱ~n+1\widetilde{\varrho}_{n+1} and ϕn+1\phi_{n+1}, we get

∥ϱ~n+1−ϕn+1∥1=∥∑k=1KγkLkωnLk†∥1≤∑k=1Kγk∥Lk∥12∥ωn∥1≤C1C2τ2.(39)\left\|\widetilde{\varrho}_{n+1}-\phi_{n+1}\right\|_1=\left\|\sum_{k=1}^{K}\gamma_kL_k\omega_nL_k^\dagger\right\|_1\leq\sum_{k=1}^{K}\gamma_k\|L_k\|_1^2\|\omega_n\|_1\leq C_1C_2\tau^2. \tag*{(39)}

Finally, we conclude using (36)–(39). □

Now, we derive the error estimate of the LREE integrator (5.4) (or (5.6)). We assume that the initial low-rank approximation satisfies

∥ρ0−ϱ0∥1≤δ,\|\rho_0-\varrho_0\|_1\leq\delta,

for some δ>0\delta>0.

Theorem 5.3. Let ρ(t)\rho(t) be the solution of the Lindblad equation (1) with ρ(0)=ρ0\rho(0)=\rho_0, and {ϱn}n=0N\{\varrho_n\}_{n=0}^{N} be the corresponding numerical solution generated by the LREE scheme (5.4). Assume that the matrix exponential algorithm used in the LREE scheme (5.4) satisfies ∥eτAσeτA†−e~τAσe~τA†∥1≤C3tol⁡1∥σ∥1\left\|e^{\tau A}\sigma e^{\tau A^\dagger}-\widetilde{e}^{\tau A}\sigma\widetilde{e}^{\tau A^\dagger}\right\|_1\leq C_3\operatorname{tol}_1\|\sigma\|_1 for any σ∈Cm×m\sigma\in\mathbb{C}^{m\times m}. Let the error tolerances of the matrix exponential algorithm and the column compression satisfy tol⁡1=τϵ1\operatorname{tol}_1=\tau\epsilon_1 and tol⁡2=τϵ2\operatorname{tol}_2=\tau\epsilon_2 for some ϵ1,ϵ2>0\epsilon_1,\epsilon_2>0, respectively. Then, it holds that

∥ρ(tn)−ϱn∥1≤c~1tnτ+δ+2c2tnϵ1+2tnϵ2,0≤n≤N,\|\rho(t_n)-\varrho_n\|_1\leq\widetilde{c}_1t_n\tau+\delta+2c_2t_n\epsilon_1+2t_n\epsilon_2,\qquad0\leq n\leq N,

where c~1=c1+2C1C2\widetilde{c}_1=c_1+2C_1C_2 and the positive constants c1c_1 and c2c_2 are as defined in Theorem 4.4 and Theorem 5.2, respectively.

Proof. First, we split the global error ρ(tn+1)−ϱn+1\rho(t_{n+1})-\varrho_{n+1} as follows:

ρ(tn+1)−ϱn+1=(ρ(tn+1)−ρn+1)+(ρn+1−ρ~n+1)+(ρ~n+1−ϱn+1),(40)\rho(t_{n+1})-\varrho_{n+1}=\bigl(\rho(t_{n+1})-\rho_{n+1}\bigr)+\bigl(\rho_{n+1}-\widetilde{\rho}_{n+1}\bigr)+\bigl(\widetilde{\rho}_{n+1}-\varrho_{n+1}\bigr), \tag*{(40)}

where the auxiliary quantities ρ~n+1\widetilde{\rho}_{n+1} are derived from the FREE integrator with low-rank initial value ϱ0\varrho_0. That means

ρ~n+1=eτAρ~neτA†+∑k=1Kγk∫0τLkesAρ~nesA†Lk† ds,0≤n≤N−1,(41)\widetilde{\rho}_{n+1}=e^{\tau A}\widetilde{\rho}_ne^{\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_ke^{sA}\widetilde{\rho}_ne^{sA^\dagger}L_k^\dagger\,ds,\qquad0\leq n\leq N-1, \tag*{(41)}

where ρ~0=ϱ0\widetilde{\rho}_0=\varrho_0. Note that the global error is decomposed into three components, where the first component ρ(tn+1)−ρn+1\rho(t_{n+1})-\rho_{n+1} denotes the global error of the FREE integrator (8), and we have presented the error estimate for this part in Theorem 4.4. The second component ρn+1−ρ~n+1\rho_{n+1}-\widetilde{\rho}_{n+1} is the difference between the full-rank solutions with initial values ρ0\rho_0 and low-rank ρ0\rho_0. The third component ρ~n+1−ϱn+1\widetilde{\rho}_{n+1}-\varrho_{n+1} is the difference of the solutions obtained with the FREE integrator and the LREE integrator with the same initial value ϱ0\varrho_0.

Subtracting (5.13) from (2.6) and applying Lemma 4.3, we obtain

∥ρn+1−ρ~n+1∥1=∥e−τA(ρn−ρ~n)e−τA†+∑k=1Kγk∫0τLkesA(ρn−ρ~n)esA†Lk† ds∥1≤∥ρn−ρ~n∥1≤∥ρ0−ρ~0∥1=∥ρ0−ϱ0∥1=δ.(42)\begin{aligned} \left\|\rho_{n+1}-\widetilde{\rho}_{n+1}\right\|_1 &=\left\|e^{-\tau A}(\rho_n-\widetilde{\rho}_n)e^{-\tau A^\dagger}+\sum_{k=1}^{K}\gamma_k\int_0^\tau L_k e^{sA}(\rho_n-\widetilde{\rho}_n)e^{sA^\dagger}L_k^\dagger\,\mathrm{d}s\right\|_1\\ &\leq\left\|\rho_n-\widetilde{\rho}_n\right\|_1\leq\left\|\rho_0-\widetilde{\rho}_0\right\|_1=\left\|\rho_0-\varrho_0\right\|_1=\delta. \tag*{(42)} \end{aligned}

The third error component ρ~n+1−ϱn+1\widetilde{\rho}_{n+1}-\varrho_{n+1} can be decomposed as

ρ~n+1−ϱn+1=(ρ~n+1−ϱ~n+1)+(ϱ~n+1−ϱ^n+1)+(ϱ^n+1−ϱn+1).(43)\widetilde{\rho}_{n+1}-\varrho_{n+1}=(\widetilde{\rho}_{n+1}-\widetilde{\varrho}_{n+1})+(\widetilde{\varrho}_{n+1}-\widehat{\varrho}_{n+1})+(\widehat{\varrho}_{n+1}-\varrho_{n+1}). \tag*{(43)}

Subtracting (5.6a) from (5.13) and applying Lemma 4.3, we have

∥ρ~n+1−ϱ~n+1∥1≤∥ρ~n−ϱn∥1.\left\|\widetilde{\rho}_{n+1}-\widetilde{\varrho}_{n+1}\right\|_1\leq\left\|\widetilde{\rho}_n-\varrho_n\right\|_1.

It then follows from (5.5), (5.6), (5.15), Lemma 3.2 and Theorem 5.2 that

∥ρ~n+1−ϱn+1∥1≤∥ρ~n+1−ϱ~n+1∥1+∥ϱ~n+1−ϱ^n+1∥1+∥ϱ^n+1−ϱn+1∥1≤∥ρ~n−ϱn∥1+∥ϑn+1∥1+∥(Tr⁡(ϱ^n+1)−1)ϱn+1∥1=∥ρ~n−ϱn∥1+∥ϑn+1∥1+∣Tr⁡(ϱ~n+1)−Tr⁡(ϑn+1)−1∣⋅∥ϱn+1∥1=∥ρ~n−ϱn∥1+∥ϑn+1∥1+∣Tr⁡(ϱn)−Tr⁡(ϑn+1)−1∣=∥ρ~n−ϱn∥1+∥ϑn+1∥1+∣Tr⁡(ϑn+1)∣≤∥ρ~n−ϱn∥1+2∥ϑn+1∥1≤∥ρ~n−ϱn∥1+2τ(C1C2τ+c2ϵ1+ϵ2)≤∥ρ~0−ϱ0∥1+2(n+1)τ(C1C2τ+c2ϵ1+ϵ2)=2tn+1(C1C2τ+c2ϵ1+ϵ2).(44)\begin{aligned} \left\|\widetilde{\rho}_{n+1}-\varrho_{n+1}\right\|_1 &\leq\left\|\widetilde{\rho}_{n+1}-\widetilde{\varrho}_{n+1}\right\|_1+\left\|\widetilde{\varrho}_{n+1}-\widehat{\varrho}_{n+1}\right\|_1+\left\|\widehat{\varrho}_{n+1}-\varrho_{n+1}\right\|_1\\ &\leq\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+\left\|\vartheta_{n+1}\right\|_1+\left\|\left(\operatorname{Tr}(\widehat{\varrho}_{n+1})-1\right)\varrho_{n+1}\right\|_1\\ &=\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+\left\|\vartheta_{n+1}\right\|_1+\left|\operatorname{Tr}(\widetilde{\varrho}_{n+1})-\operatorname{Tr}(\vartheta_{n+1})-1\right|\cdot\left\|\varrho_{n+1}\right\|_1\\ &=\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+\left\|\vartheta_{n+1}\right\|_1+\left|\operatorname{Tr}(\varrho_n)-\operatorname{Tr}(\vartheta_{n+1})-1\right|\\ &=\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+\left\|\vartheta_{n+1}\right\|_1+\left|\operatorname{Tr}(\vartheta_{n+1})\right|\\ &\leq\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+2\left\|\vartheta_{n+1}\right\|_1\\ &\leq\left\|\widetilde{\rho}_n-\varrho_n\right\|_1+2\tau\left(C_1C_2\tau+c_2\epsilon_1+\epsilon_2\right)\\ &\leq\left\|\widetilde{\rho}_0-\varrho_0\right\|_1+2(n+1)\tau\left(C_1C_2\tau+c_2\epsilon_1+\epsilon_2\right)\\ &=2t_{n+1}\left(C_1C_2\tau+c_2\epsilon_1+\epsilon_2\right). \tag*{(44)} \end{aligned}

The desired result follows from (5.12), Theorem 4.4, (5.14) and (5.16). □\square

Numerical experiments

In this section, we present results of numerical experiments that demonstrate the accuracy and efficiency of our FREE integrator (2.9) and LREE integrator (5.4). All simulations are carried out using Python 3.12.4 on a laptop with Intel(R) Core(TM) i7-8565U [email protected] and 16GB RAM. We use the codes in the Python package scipy to compute the matrix exponential and the products of matrix exponential with vectors. The codes are based on algorithms proposed in [34, 35]. We also use the code in scipy to solve the algebraic Lyapunov equations in the FREE scheme (2.9), which is based on the Bartels-Stewart algorithm; see, e.g., [36, 37].

To show the accuracy of the proposed scheme, we display the relative errors defined as follows:

Error=∥ρN−ρ(T)∥1∥ρ(T)∥1,Error = \frac{\lVert\rho_N-\rho(T)\rVert_1}{\lVert\rho(T)\rVert_1},

and test its convergence order. The exact solution ρ(T)\rho(T) is computed through the vectorized Lindblad equation (1.2) using matrix exponential.

The initial condition of the Lindblad equation is mostly chosen to be the Greenberger-Horne-Zeilinger (GHZ) state

ρ(0)=12(∣0⟩⊗K⟨0∣⊗K+∣0⟩⊗K⟨d−1∣⊗K+∣d−1⟩⊗K⟨0∣⊗K+∣d−1⟩⊗K⟨d−1∣⊗K)=12[10⋯0100⋯00⋮⋮⋱⋮⋮00⋯0010⋯01]∈RdK×dK.\begin{aligned} \rho(0)=\frac{1}{2}\left(|0\rangle^{\otimes K}\langle0|^{\otimes K}+|0\rangle^{\otimes K}\langle d-1|^{\otimes K}+|d-1\rangle^{\otimes K}\langle0|^{\otimes K}+|d-1\rangle^{\otimes K}\langle d-1|^{\otimes K}\right) \\ &=\frac{1}{2} \begin{bmatrix} 1 & 0 & \cdots& 0 & 1\\ 0 & 0 & \cdots& 0 & 0\\ \vdots& \vdots& \ddots& \vdots& \vdots\\ 0 & 0 & \cdots& 0 & 0\\ 1 & 0 & \cdots& 0 & 1 \end{bmatrix}\in\mathbb{R}^{d^K\times d^K}. \end{aligned}

The GHZ state for KK qudits is a highly entangled quantum state that is a generalization of the KK-qubit GHZ state, which is a superposition of all qubits being in state ∣0⟩|0\rangle and all qubits being in state ∣1⟩|1\rangle. GHZ states are a valuable resource in quantum information processing due to their strong entanglement properties, which can be utilized for various tasks such as quantum teleportation, quantum error correction, and quantum cryptography.

We consider the Lindblad equation with the X-X Ising Chain Hamiltonian given by

H(t)=∑k=1K(aJz(k)+b(Jz(k))2)+∑k=1K−1∑l=k+1Kgkl(t)Jx(k)Jx(l),(45)H(t)=\sum_{k=1}^{K}\left(aJ_z^{(k)}+b(J_z^{(k)})^2\right)+\sum_{k=1}^{K-1}\sum_{l=k+1}^{K}g_{kl}(t)J_x^{(k)}J_x^{(l)}, \tag*{(45)}

where

Jw(k)=Id⊗⋯⊗Id⊗Jw⊗Id⊗⋯⊗Id,w=x,z,k=1,…,K,J_w^{(k)}=I_d\otimes\cdots\otimes I_d\otimes J_w\otimes I_d\otimes\cdots\otimes I_d,\qquad w=x,z,\qquad k=1,\ldots,K,

with k−1k-1 identity matrices before JwJ_w and K−kK-k identity matrices after it.

IdI_d is the d×dd\times d identity matrix and the Jx,JzJ_x,J_z operators are defined as follows

Jx=12[0d−10⋯0d−102(d−2)0⋮⋮⋱⋱⋱⋮⋮⋱(d−2)20d−10⋯0d−10]∈Rd×d,J_x=\frac{1}{2} \begin{bmatrix} 0 & \sqrt{d-1} & 0 & \cdots& 0\\ \sqrt{d-1} & 0 & \sqrt{2(d-2)} & 0 & \vdots\\ \vdots& \ddots& \ddots& \ddots& \vdots\\ \vdots& \ddots& \sqrt{(d-2)2} & 0 & \sqrt{d-1}\\ 0 & \cdots& 0 & \sqrt{d-1} & 0 \end{bmatrix}\in\mathbb{R}^{d\times d},

and

Jz=[d−120⋯00d−320⋮⋮⋱⋱00⋯01−d2]∈Rd×d.J_z= \begin{bmatrix} \frac{d-1}{2} & 0 & \cdots& 0\\ 0 & \frac{d-3}{2} & 0 & \vdots\\ \vdots& \ddots& \ddots& 0\\ 0 & \cdots& 0 & \frac{1-d}{2} \end{bmatrix}\in\mathbb{R}^{d\times d}.

The first sum in (45) is the so-called drift-Hamiltonian, which characterizes the inner dynamics of the qudit. The term aJz(k)aJ_z^{(k)} allows for an even spacing of eigenvalues of amplitude aa between the levels, while the term b(Jz(k))2b(J_z^{(k)})^2 allows for an uneven spacing of amplitude bb and therefore, the dd levels become individually addressable. The second sum is the interaction Hamiltonians, where gkl(t)g_{kl}(t) is the coupling strength. If gklg_{kl} depends on time, we have a time-dependent Hamiltonian. The Ising-type Hamiltonian describes the interaction between two qudits as a product of their respective local operators. The jump operators LkL_k are always set to be Lk=Jz(k)L_k=J_z^{(k)}, unless stated otherwise.

Tests on the FREE integrator

In Figure 6.1, we present numerical results of the proposed FREE integrator for the Lindblad equation with time-independent Hamiltonian (gkl=g=1g_{kl}=g=1). The solvers for matrix exponential and Lyapunov equation used in FREE integrator were set with default tolerance (machine precision 10−1610^{-16}). We see that the expected first-order accuracy is achieved for the FREE scheme, which is in agreement with the convergence estimate stated in Theorem 4.4. It can also be seen that Tr⁡(ρn)−1\operatorname{Tr}(\rho_n)-1 is close to machine precision for different time step sizes, which demonstrates the preservation of unit trace unconditionally. We also see that the populations ρ1,1\rho_{1,1}, ρ3,3\rho_{3,3}, ρ8,8\rho_{8,8} and ρ256,256\rho_{256,256} stay nonnegative in the interval [0,1][0, 1], which is consistent with the positive property of the density matrix. Note that although we display evolutions of populations ρ1,1\rho_{1,1}, ρ3,3\rho_{3,3}, ρ8,8\rho_{8,8} and ρ256,256\rho_{256,256}, the positive features are visible in all populations.

Numerical results of the FREE integrator for the Lindblad equation with Hamiltonian (6.1)

Figure 6.1. Numerical results of the FREE integrator for the Lindblad equation with Hamiltonian (45) (d=4d=4, K=4K=4, a=1.5a=1.5, b=0.5b=0.5, γk=γ\gamma_k=\gamma, gkl(t)≡gg_{kl}(t)\equiv g). Top left: relative errors at T=1T=1 vs step sizes for different gg and γ\gamma. Top right: evolutions of Tr⁡(ρn)−1\operatorname{Tr}(\rho_n)-1 vs time for different τ\tau with fixed g=1g=1, γ=0.01\gamma=0.01 and T=20T=20. Bottom left: evolutions of the populations ρ1,1\rho_{1,1} and ρ256,256\rho_{256,256} with τ=0.1\tau=0.1, g=1g=1, γ=0.01\gamma=0.01 and T=20T=20. Bottom right: evolutions of the populations ρ3,3\rho_{3,3} and ρ8,8\rho_{8,8} with τ=0.1\tau=0.1, g=1g=1, γ=0.01\gamma=0.01 and T=20T=20.

In Figure 6.2, we display numerical results of the FREE scheme for the Lindblad equation with time-dependent Hamiltonian. Here the coupling strength was set to

Numerical results of the FREE integrator for the Lindblad equation with time-dependent Hamiltonian

Figure 6.2. Numerical results of the FREE integrator for the Lindblad equation with time-dependent Hamiltonian (6.1) (d=6d = 6, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ\gamma_k = \gamma, gkl(t)=δk,l−1⋅g(t)g_{kl}(t) = \delta_{k,l-1} \cdot g(t)). Top left: relative errors at T=1T = 1 vs step sizes for different g(t)g(t) and γ\gamma. Top right: evolutions of Tr⁡(ρn)−1\operatorname{Tr}(\rho_n) - 1 vs time for different τ\tau with g(t)=(1+t)14g(t) = (1 + t)^{\frac{1}{4}}, γ=0.05\gamma= 0.05 and T=20T = 20. Bottom left: evolutions of the populations ρ1,1\rho_{1,1} and ρ216,216\rho_{216,216} with τ=0.1\tau= 0.1, g(t)=(1+t)14g(t) = (1 + t)^{\frac{1}{4}}, γ=0.05\gamma= 0.05 and T=20T = 20. Bottom right: evolutions of the populations ρ4,4\rho_{4,4} and ρ11,11\rho_{11,11} with τ=0.1\tau= 0.1, g(t)=(1+t)14g(t) = (1 + t)^{\frac{1}{4}}, γ=0.05\gamma= 0.05 and T=20T = 20.

gkl(t)=δk,l−1⋅g(t)={g(t)if k=l−1,0else,g_{kl}(t) = \delta_{k,l-1} \cdot g(t) = \begin{cases} g(t) & \text{if } k = l - 1,\\ 0 & \text{else}, \end{cases}

and we chose g(t)g(t) as (1+t)14(1 + t)^{\frac{1}{4}}, e−te^{-t}, sin⁡(2πt)\sin(2\pi t) and 10sin⁡(10πt)10\sin(10\pi t), respectively. Again, we see that the FREE scheme achieves first-order accuracy in time. The numerical solutions obtained by our FREE scheme preserve positivity and unit trace unconditionally.

Tests on the LREE integrator

Now, we test the performance of the LREE scheme for the Lindblad equation with Hamiltonian (6.1). First, we test the error behavior of the LREE scheme with respect to time discretization and the initial low-rank error δ\delta. In this case, the initial condition of the Lindblad equation is set to be

ρ(0)=(1−δ2)q1q1⊤+δ2q2q2⊤,\rho(0) = \left(1 - \frac{\delta}{2}\right)q_1q_1^\top+ \frac{\delta}{2}q_2q_2^\top,

where q1q_1 and q2q_2 are orthonormal vectors and we obtain them from SVD of a random m×2m \times2 real matrix. The low-rank initial factor is set to be Z0=q1Z_0 = q_1 and clearly we have ∥ρ(0)−Z0Z0⊤∥1=δ\lVert\rho(0) - Z_0Z_0^\top\rVert_1 = \delta. The error tolerances of the matrix exponential and column compression are set to be tol1=10−10tol_1 = 10^{-10} and tol2=10−10tol_2 = 10^{-10}, respectively. In Figure 6.3, we present the error behavior of the LREE scheme with various choices of δ\delta. As we can see, the LREE scheme has convergence of order one when the initial low-rank error is small enough. When the initial low-rank error is dominant, decreasing the step size does not lead to more accurate solution. These numerical results are in agreement with the error estimate stated in Theorem 5.3 and we note that the estimate with respect to δ\delta can not be made sharper.

Numerical results of the LREE scheme

Figure 6.3. Numerical results of the LREE scheme with tol1=10−10tol_1 = 10^{-10} and tol2=10−10tol_2 = 10^{-10} for the Lindblad equation with Hamiltonian (6.1). Left: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=4d = 4, K=4K = 4, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, gkl(t)≡g=1g_{kl}(t) \equiv g = 1. Right: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=6d = 6, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅(1+t)14g_{kl}(t) = \delta_{k,l-1} \cdot(1 + t)^{\frac{1}{4}}.

Second, we focus on error behavior with respect to the time discretization and low-rank approximation. The initial condition is chosen to be the GHZ state. The initial low-rank factor is set to be Z0=12[1,0,…,0,1]⊤Z_0 = \frac{1}{\sqrt{2}}[1, 0, \ldots, 0, 1]^\top. In this case, ρ(0)=Z0Z0⊤\rho(0) = Z_0Z_0^\top and there is no initial low-rank approximation error (δ=0\delta= 0). The error tolerance of the code for matrix exponential is set to be tol1=10−10tol_1 = 10^{-10}. The error tolerance of the column compression algorithm is set to be tol2=ϵ2τtol_2 = \epsilon_2\tau. In Figure 6.4, we present the error behavior of the LREE scheme with various ϵ2\epsilon_2. We see that the LREE scheme achieves first-order accuracy when the low-rank approximation error (due to the column compression) is small enough. Decreasing the step size cannot lead to more accurate solution when the low-rank approximation error is dominant. These results are also in agreement with the error estimate stated in Theorem 5.3.

Numerical results of the LREE scheme

Figure 6.4. Numerical results of the LREE scheme with δ=0\delta= 0 and tol1=10−10tol_1 = 10^{-10} for the Lindblad equation with Hamiltonian (6.1). Left: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=4d = 4, K=4K = 4, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, gkl(t)≡g=1g_{kl}(t) \equiv g = 1. Right: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=6d = 6, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅(1+t)14g_{kl}(t) = \delta_{k,l-1} \cdot(1 + t)^{\frac{1}{4}}.

Finally, we test the performance of the LREE scheme in terms of errors due to time discretization and the approximation of the products of matrix exponential with vectors. We set Z0=12[1,0,…,0,1]TZ_0 = \frac{1}{\sqrt{2}}[1,0,\ldots,0,1]^{\mathsf{T}} and in this case δ=0\delta= 0. The error tolerance of the column compression is set to tol2=10−10tol_2 = 10^{-10}. The error tolerance of the algorithm for matrix exponential is set to be tol1=ϵ1τtol_1 = \epsilon_1\tau. In Figure 6.5, we show the error behavior of the LREE scheme with various ϵ1\epsilon_1. We see that the LREE scheme achieves first-order accuracy even when ϵ1\epsilon_1 is much bigger than τ\tau. This error behavior is different from those observed in Figure 6.3 and Figure 6.4. We remark that this performance may be caused by the matrix exponential algorithm [35] that we used. In the Theorem 5.3, we assume that the matrix exponential algorithm satisfies

Numerical results of the LREE scheme

Figure 6.5. Numerical results of the LREE scheme with δ=0\delta= 0 and tol2=10−10\mathit{tol}_2 = 10^{-10} for the Lindblad equation with Hamiltonian (45). Left: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=4d = 4, K=4K = 4, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, gkl(t)≡g=1g_{kl}(t) \equiv g = 1. Right: relative errors at T=1T = 1 vs step sizes for the Lindblad problem with d=6d = 6, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅(1+t)14g_{kl}(t) = \delta_{k,l-1} \cdot(1 + t)^{\frac{1}{4}}.

∥eτAσeτA†−eτAσeτA†∥1≤C3 tol1∥σ∥1,\left\lVert e^{\tau A}\sigma e^{\tau A^\dagger} - e^{\tau A}\sigma e^{\tau A^\dagger} \right\rVert_1 \le C_3\,tol_1\lVert\sigma\rVert_1,

where tol1tol_1 is the input error tolerance. It seems that the error constant C3C_3 depends on τ\tau, which was verified in Figure 6.6, in which we show the values of Ce:=∥eτAnσeτAn†−eτAnσeτAn†∥1/tol1C_e := \left\lVert e^{\tau A_n}\sigma e^{\tau A_n^\dagger} - e^{\tau A_n}\sigma e^{\tau A_n^\dagger} \right\rVert_1/tol_1 for various τ\tau and tol1tol_1, where σ>0\sigma> 0 is a random matrix with ∥σ∥1=1\lVert\sigma\rVert_1 = 1 and we set An=AA_n = A for equation with time-independent Hamiltonian and An=A10A_n = A_{10} for problem with time-dependent Hamiltonian.

$C_e$ vs $\tau$

Figure 6.6. CeC_e vs τ\tau. Left: time-independent case with d=4d = 4, K=4K = 4, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, gkl(t)≡g=1g_{kl}(t) \equiv g = 1. Right: time-dependent case with d=6d = 6, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅(1+t)14g_{kl}(t) = \delta_{k,l-1} \cdot(1 + t)^{\frac{1}{4}}.

Numerical comparison

Now, we compare integration algorithms with the Lindblad equation suite mesolve developed in QuTip [31], a widely used Python package for the simulation of open quantum systems. We consider five ODE solvers in mesolve: adams, bdf, lsoda, dop853 and vern9, which are based on twelfth-order Adams method, fifth-order backward differentiation formula, adaptive method chosen between adams and bdf, Dormand and Prince’s eighth-order Runge-Kutta method and Verner’s ninth-order Runge-Kutta method (see [38, 39]), respectively. Note that these ODE solvers consider the Lindblad equation in the vectorized form (1.2).

In Figure 6.7, we compare the performance of our FREE and LREE schemes and the solvers used in QuTip with respect to accuracy and computational times measured in seconds. Note that we only report the running time of the ODE solvers. The initial condition is set to be the GHZ state (δ=0\delta= 0 for the LREE scheme) and relative errors are reported at T=1T = 1. The error tolerances of the matrix exponential algorithm and column compression method used in LREE scheme are set to be tol1=τ20tol_1 = \frac{\tau}{20} and tol2=τ220tol_2 = \frac{\tau^2}{20} for equation with time-independent Hamiltonian and tol1=τ10tol_1 = \frac{\tau}{10} and tol2=τ210tol_2 = \frac{\tau^2}{10} for equation with time-dependent Hamiltonian, respectively. We observe that the LREE scheme performs faster than the FREE scheme for a fixed level of accuracy. For a fixed level of moderate accuracy (i.e., bigger than 10−310^{-3} in the time-independent case), our LREE scheme is more efficient than the solvers in QuTip. However, for the given problems (with m=256m = 256), the solvers in QuTip performs faster than our exponential integrators in the high accuracy case. After all, our methods are only first-order accurate and much more steps are required to reach similar accuracy.

Numerical comparison between the proposed exponential schemes and the solvers in QuTip

Figure 6.7. Numerical comparison between the proposed exponential schemes and the solvers in QuTip for the Lindblad equation with Hamiltonian (45). Left: results for problem with d=4d=4, K=4K=4, a=1.5a=1.5, b=0.5b=0.5, γk=γ=0.01\gamma_k=\gamma=0.01, gkl(t)≡g=1g_{kl}(t)\equiv g=1. Right: results for problem with d=4d=4, K=4K=4, a=1a=1, b=1b=1, γk=γ=0.05\gamma_k=\gamma=0.05, gkl(t)=δk,l−1⋅sin⁡(2πt)g_{kl}(t)=\delta_{k,l-1}\cdot\sin(2\pi t).

In Figure 6.8, we compare the computational times of our FREE and LREE schemes with those of the Lindblad solvers used in QuTip as the size m=dKm = d^K of the Lindblad equation increases. We fix K=1K = 1 and let dd increase. The jump operators LkL_k are set to be Lk=Jx(k)L_k = J_x^{(k)}. The error tolerances of the matrix exponential algorithm and column compression method used in the LREE scheme are set to be tol1=τ10tol_1 = \frac{\tau}{10} and tol2=τ210tol_2 = \frac{\tau^2}{10}, respectively. In this experiment, the absolute and relative tolerances of the QuTip solvers and the step sizes of the FREE and LREE schemes are chosen such that the relative errors at T=0.1T = 0.1 are approximately 10−310^{-3}, see the right-hand plot in Figure 6.8. We see that our LREE scheme performs much faster than the QuTip Lindblad solvers for problems with high dimensions. In particular, we notice a strikingly better computational time complexity of our LREE scheme, with much smaller change in the runtime of our algorithm with respect to increasing mm. However, we see that the FREE scheme is less efficient than the other solvers. This is mainly due to the computation of the associated algebraic Lyapunov equations.

Numerical comparison between the proposed exponential schemes and the solvers in QuTiP for the Lindblad equation with Hamiltonian (6.1) and jump operators

Figure 6.8. Numerical comparison between the proposed exponential schemes and the solvers in QuTiP for the Lindblad equation with Hamiltonian (45) and jump operators Lk=Jx(k)L_k = J_x^{(k)} (K=1K = 1, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, T=0.1T = 0.1). Left: CPU times vs. mm. Right: relative errors vs mm.

Note that the numerical results presented in Figure 6.7 and Figure 6.8 are with respect to problems with sparse Hamiltonian (6.1) and sparse jump operators. In Figure 6.9, we compare the computational times of our exponential schemes with those of the five Lindblad solvers used in QuTip in the case of a Lindblad equation with dense jump operator. We consider the Lindblad equation with Hamiltonian (6.1) and K=1K = 1, and the jump operator L1L_1 is chosen as a random and dense matrix. We observe that now both of our FREE and

Numerical comparison between the proposed exponential schemes and the solvers in QuTiP for the Lindblad equation with Hamiltonian (6.1) and random jump operators

Figure 6.9. Numerical comparison between the proposed exponential schemes and the solvers in QuTiP for the Lindblad equation with Hamiltonian (45) and random jump operators LkL_k (K=1K = 1, a=1.5a = 1.5, b=0.5b = 0.5, γk=γ=0.01\gamma_k = \gamma= 0.01, T=0.1T = 0.1). Left: CPU times vs. mm. Right: relative errors vs mm.

LREE schemes outperform the five QuTip Lindblad solvers in terms of computational times. We also see that the computational times of the QuTip Lindblad solvers increase quickly as the dimension mm increases.

We remark that, in the dense case, the Lindblad solvers used in QuTip require to store a dense and complex m2×m2m^2\times m^2 matrix L\mathcal{L}, which requires memory of 16m416m^4 Bytes. On the other hand, our FREE and LREE schemes only require to store a matrix of size m×mm\times m, which needs memory of 16m216m^2 Bytes. For example, if m=120m=120, QuTip needs nearly 3000 MB memory while our algorithms only require 0.2 MB memory. In the sparse case, the memory requirements of the QuTip and our exponential schemes on the coefficient matrix reduce to O(m2)O(m^2) and O(m)O(m), respectively. It seems that the QuTip Lindblad lsoda solver requires dense representation of the data matrices. So the lsoda method was absent from the numerical comparison presented in Figure 6.7 and Figure 6.8 since memory error appears for the considered dimensions.

In Figure 6.10-Figure 6.11, we compare the performance of our exponential schemes with the five solvers in QuTip concerning the preservation of positivity and unit trace. We set tol1=τ10tol_1=\frac{\tau}{10} and tol2=τ210tol_2=\frac{\tau^2}{10} for the LREE scheme and τ=0.1\tau=0.1 for both the FREE and LREE schemes. The absolute and relative tolerances of the adams, bdf, lsoda and dop853 solvers are all set to be 10−310^{-3}. The absolute and relative tolerances of the vern9 solver are set to be 10−210^{-2}. We see that all the solvers can preserve unit trace. However, we also see that, while both our FREE and LREE schemes are positivity preserving, the solvers in QuTip do not guarantee the positive semidefinite property of the density matrix, which is in agreement with the first-order barrier for positivity of (standard) Runge–Kutta and multistep methods as discussed in [8, 9].

Numerical comparison between the proposed exponential schemes and adams, bdf, lsoda solvers

Figure 6.10. Numerical comparison between the proposed exponential schemes and adams, bdf, lsoda solvers in QuTip for the Lindblad equation with Hamiltonian (45) and d=4d = 4, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅sin⁡(2πt)g_{kl}(t) = \delta_{k,l-1} \cdot\sin(2\pi t). Top left: evolutions of the populations ρ46,46\rho_{46,46}. Top right: evolutions of the populations ρ45,45\rho_{45,45}. Bottom left: evolutions of the populations ρ38,38\rho_{38,38}. Bottom right: evolutions of Tr⁡(ρn)−1\operatorname{Tr}(\rho_n) - 1 vs time.

Numerical comparison between the proposed exponential schemes and dop853, vern9 solvers

Figure 6.11. Numerical comparison between the proposed exponential schemes and dop853, vern9 solvers in QuTip for the Lindblad equation with Hamiltonian (45) and d=4d = 4, K=3K = 3, a=1a = 1, b=1b = 1, γk=γ=0.05\gamma_k = \gamma= 0.05, gkl(t)=δk,l−1⋅10sin⁡(10πt)g_{kl}(t) = \delta_{k,l-1} \cdot10\sin(10\pi t). Left: evolutions of the populations ρ17,17\rho_{17,17}. Right: evolutions of Tr⁡(ρn)−1\operatorname{Tr}(\rho_n) - 1 vs time.

Conclusion

In this paper, full- and low-rank exponential Euler integrators for solving the Lindblad equation were developed and analyzed. It was proven and numerically validated that these schemes preserve positivity and unit trace unconditionally. Moreover, sharp accuracy estimates were presented. The computational performance of the new exponential Euler integrators were successfully validated by results of numerical simulations with the Lindblad equation with various Hamiltonians, and by comparison with the Lindblad solvers available in the well-known QuTip package for simulation of open quantum systems.

Acknowledgements

A.B. would like to thank Guofeng Zhang (Hong Kong Polytechnic University), Xiu-Hao Deng (Shenzhen Institute of Quantum Science and Engineering, SUSTech) and Song Zhang (Shenzhen International Quantum Academy) for inspiring discussions on open quantum systems and very kind hospitality.

Funding The work of Hao Chen was partly supported by the Natural Science Foundation Project of CQ CSTC (No. cstc2021jcyj-msxmX0034).

References

  1. [1]H.-P. Breuer and F. Petruccione. The Theory of Open Quantum Systems. Oxford University Press, 2007.DOI
  2. [2]E.B. Davies. Quantum Theory of Open Systems. Academic Press, 1976.
  3. [3]V. Gorini, A. Kossakowski, and E.C.G. Sudarshan. Completely positive dynamical semigroups of N-level systems. Journal of Mathematical Physics, 17(5):821–825, 05 1976.DOI
  4. [4]G. Lindblad. On the generators of quantum dynamical semigroups. Communications in Mathematical Physics, 48(2):119–130, 1976.DOI
  5. [5]M. Riesch and C. Jirauschek. Analyzing the positivity preservation of numerical methods for the Liouville-von Neumann equation. Journal of Computational Physics, 390:290–296, 2019.DOI
  6. [6]R.W. Złokowski, J.M. Arnold, and D.M. Gogny. Ultrafast pulse interactions with two-level atoms. Phys. Rev. A, 52:3082–3094, Oct 1995.
  7. [7]B. Bidégaray, A. Bourgeade, and D. Reignier. Introducing physical relaxation terms in Bloch equations. Journal of Computational Physics, 170(2):603–613, 2001.DOI
  8. [8]S. Blanes, A. Iserles, and S. Macnamara. Positivity-preserving methods for ordinary differential equations. ESAIM: M2AN, 56(6):1843–1870, 2022.DOI
  9. [9]A Martiradonna, G. Colonna, and F. Diele. Geco: Geometric conservative nonstandard schemes for biochemical systems. Applied Numerical Mathematics, 155:38–57, 2020. Structural Dynamical Systems: Computational Aspects held in Monopoli (Italy) on June 12-15, 2018.DOI
  10. [10]M. Riesch, A. Pikl, and C. Jirauschek. Completely positive trace preserving methods for the Lindblad equation. In 2020 International Conference on Numerical Simulation of Optoelectronic Devices (NUSOD), pages 109–110, 2020.DOI
  11. [11]A. Bourgeade and O. Saut. Numerical methods for the bidimensional Maxwell-Bloch equations in nonlinear crystals. Journal of Computational Physics, 213(2):823–843, 2006.DOI
  12. [12]M.E. Songolo and B. Bidégaray-Fesquet. Strang splitting schemes for N-level Bloch models. International Journal of Modeling, Simulation, and Scientific Computing, 14(03):2350044, 2023.arxiv.org/abs/1909.10932
  13. [13]M.E. Songolo and B. Bidégaray-Fesquet. Nonstandard finite-difference schemes for the two-level Bloch model. *International Journal of Modeling, Simulation, and Scientific Computing*, 09(04):1850033, 2018.DOI
  14. [14]C. Le Bris and P. Rouchon. Low-rank numerical approximations for high-dimensional Lindblad equations. *Phys. Rev. A*, 87:022125, Feb 2013.DOI
  15. [15]C. Le Bris, P. Rouchon, and J. Roussel. Adaptive low-rank approximation and denoised Monte Carlo approach for high-dimensional Lindblad equations. *Phys. Rev. A*, 92:062126, Dec 2015.DOI
  16. [16]Y. Cao and J. Lu. Structure-preserving numerical schemes for Lindblad equations, 2024.DOI
  17. [17]A.W. Schlimgen, K. Head-Marsden, L.M. Sager, P. Narang, and D.A. Mazziotti. Quantum simulation of the Lindblad equation using a unitary decomposition of operators. *Phys. Rev. Res.*, 4:023216, Jun 2022.DOI
  18. [18]W. Weimer, A. Kshetrimayum, and R. Orús. Simulation methods for open quantum many-body systems. *Rev. Mod. Phys.*, 93:015008, Mar 2021.DOI
  19. [19]A.H. Werner, D. Jaschke, P. Silvi, M. Kliesch, T. Calarco, J. Eisert, and S. Montangero. Positive tensor network approach for simulating open quantum many-body systems. *Phys. Rev. Lett.*, 116:237201, Jun 2016.DOI
  20. [20]R. Orús. A practical introduction to tensor networks: Matrix product states and projected entangled pair states. *Annals of Physics*, 349:117–158, 2014.DOI
  21. [21]U. Schollwöck. The density-matrix renormalization group. *Rev. Mod. Phys.*, 77:259–315, Apr 2005.DOI
  22. [22]G. Vidal. Efficient classical simulation of slightly entangled quantum computations. *Phys. Rev. Lett.*, 91:147902, Oct 2003.DOI
  23. [23]Guifré Vidal. Efficient simulation of one-dimensional quantum many-body systems. *Phys. Rev. Lett.*, 93:040502, Jul 2004.pubmed.ncbi.nlm.nih.gov/15323740
  24. [24]M.J. Hartmann and G. Carleo. Neural-network approach to dissipative quantum many-body dynamics. *Phys. Rev. Lett.*, 122:250502, Jun 2019.DOI
  25. [25]A. Nagy and V. Savona. Variational quantum Monte Carlo method with a neural-network ansatz for open quantum systems. *Phys. Rev. Lett.*, 122:250501, Jun 2019.DOI
  26. [26]M. Reh, M. Schmitt, and M. Gärtner. Time-dependent variational principle for open quantum systems with artificial neural networks. *Phys. Rev. Lett.*, 127:230501, Dec 2021.DOI
  27. [27]L. Arceci, P. Silvi, and S. Montangero. Entanglement of formation of mixed many-body quantum states via tree tensor operators. *Phys. Rev. Lett.*, 128:040501, Jan 2022.DOI
  28. [28]D. Sulz, C. Lubich, G. Ceruti, I. Lesanovsky, and F. Carollo. Numerical simulation of long-range open quantum many-body dynamics with tree tensor networks. *Phys. Rev. A*, 109:022420, Feb 2024.DOI
  29. [29]H. Chen and A. Borzì. Positivity preserving exponential integrators for differential Riccati equations. *Journal of Scientific Computing*, 96(2):50, 2023.DOI
  30. [30]H. Chen and A. Borzì. Low-rank exponential integrators for stiff differential Riccati equations. *Submitted*, 2024.DOI
  31. [31]J.R Johansson, P.D. Nation, and F. Nori. QuTiP 2: A Python framework for the dynamics of open quantum systems. *Computer Physics Communications*, 184(4):1234–1240, 2013.
  32. [32]M. Hochbruck and A. Ostermann. Exponential integrators. *Acta Numerica*, 19:209–286, 2010.DOI
  33. [33]N. Lang, H. Mena, and J. Saak. On the benefits of the \(ldl^t\) factorization for large-scale differential matrix equation solvers. *Linear Algebra and its Applications*, 480:44–71, 2015.
  34. [34]A.H. Al-Mohy and N.J. Higham. A new scaling and squaring algorithm for the matrix exponential. *SIAM Journal on Matrix Analysis and Applications*, 31(3):970–989, 2010.DOI
  35. [35]A.H. Al-Mohy and N.J. Higham. Computing the action of the matrix exponential, with an application to exponential integrators. *SIAM Journal on Scientific Computing*, 33(2):488–511, 2011.DOI
  36. [36]Z. Gajic and M.T.J. Qureshi. *Lyapunov Matrix Equation in System Stability and Control*. Dover Books on Engineering. Dover Publications, 2008.
  37. [37]J.D. Hamilton. *Time Series Analysis*. Princeton University Press, 2020.
  38. [38]E. Hairer, S.P. Nørsett, and G. Wanner. *Solving Ordinary Differential Equations I: Nonstiff Problems*. Springer Series in Computational Mathematics. Springer Berlin Heidelberg, 2008.
  39. [39]J. H. Verner. Explicit Runge-Kutta methods with estimates of the local truncation error. *SIAM Journal on Numerical Analysis, 15(4):772–790, 1978.DOI

Paper details

vlnc:W_b7bxn9hz
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 10

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