WorksMathematics
Computation of some dispersive equations through their iterated linearisation
Abstract
It is often the case that, while the numerical solution of the non-linear dispersive equation represents a formidable challenge, it is fairly easy and cheap to solve closely related linear equations of the form , where . In that case we advocate an iterative linearisation procedure that involves fixed-point iteration of the latter equation to solve the former. A typical case is when the original problem is a nonlinear Schrödinger or Gross–Pitaevskii equation, while the `easy' equation is linear Schrödinger with time-dependent potential. We analyse in detail the iterative scheme and its practical implementation, prove that each iteration increases the order, derive upper bounds on the speed of convergence and discuss in the case of nonlinear Schrödinger equation with cubic potential the preservation of structural features of the underlying equation: the norm, momentum and Hamiltonian energy. A key ingredient in our approach is the use of the Magnus expansion in conjunction with Hermite quadratures, which allows effective solutions of the linearised but non-autonomous equations in an iterative fashion. The resulting Magnus–Hermite methods can be combined with a wide range of numerical approximations to the matrix exponential. The paper concludes with a number of numerical experiments, demonstrating the power of the proposed approach.
Introduction
In this paper we concern ourselves with nonlinear dispersive equations in a Hamiltonian form,
evolving on a complex-valued Hilbert space . In many cases of interest the time-dependent and nonlinear Hamiltonian ,
comprises of a linear self-adjoint operator on , , and a nonlinear component, , which is a real-valued function. We assume throughout that is a Sobolev space determined by the nature of the Hamiltonian but, to make sense of concepts like the order of a method we require additional regularity, as appropriate.
The main idea underlying the proposed scheme is that it is often considerably easier to solve linear dispersive equations with time-variable potential and to leverage this to the solution of their nonlinear counterparts. We consider the equation (1) written in the form
such that the equation
is easy to solve for a reasonable linear Hamiltonian . (While this often corresponds to the separation into potential and kinetic parts, this is not necessary for our argument.) In that case we propose the iterative scheme
where the linear equations are equipped with the same initial and boundary side conditions as (1). Note that, even if , the linearised equation has a time-dependent potential. This is not a significant issue, provided that we have effective means for a numerical treatment of this case (as we do in the sequel). Note that in the case (2) the iterative scheme simplifies to
We terminate the iteration after a suitable number of steps. Needless to say, even taking for granted that (3) is easy to solve repeatedly, the first relevant issue is the convergence of (4) to the exact solution of (1) and its speed. The second issue, in the spirit of geometric numerical integration [7], is whether important geometric features of (1), e.g. Hamiltonicity or conservation, are preserved by the scheme (4).
It is important to mention that our method is not an operatorial splitting [3] (McLachlan & Quispel 2002), an approach ubiquitous in modern numerical analysis of PDEs, since we do not compose the solution from a split form of (3), and . It is perhaps more illuminating to consider our approach as an iterative linearisation of the PDE.
The proposed schemes will be illustrated in this paper with the example of the cubic nonlinear Schrödinger (NLS) equation,
as well as the Gross–Pitaevskii (GP) equation,
where it is typical to consider and either or . (In reality we implicitly assume higher regularity, otherwise our discussion of order and speed of convergence in § 3 need be amended in a transparent manner.) A negative value of corresponds to a focusing equation, while to a defocusing one. In the case of the cubic NLS example, , and . The linearisation (3) splits the kinetic part and the potential part . In the case of the Gross–Pitaevskii (GP) equation, for some real-valued functions (typically a confining potential) and (typically an external time-dependent potential).
The iteratively linearised schemes developed in this work can be combined with a wide range of methods for approximating the matrix exponential. In § 5 we present numerical examples to demonstrate the combination of our schemes with splitting methods and Krylov subspace methods.
The algorithm
Iterated linearisation
It is a good idea to recall that we are solving the dispersive PDE
by the iterative procedure
with the initial condition and the same boundary conditions as the original PDE. Formally, the later defines functional iteration,
in the ambient Hilbert space and our first concern is with the convergence of this procedure from the initial guess . In other words, we wish to ascertain whether a fixed point of (7), which clearly is a solution of (1), is attractive and whether belongs to its basin of attraction. In practical terms, we wish to demonstrate that is a contraction mapping in a closed neighbourhood of , i.e. for some , where is a closed ball of radius about . Note that this might require a restriction upon the time step so that .
In practice the PDE (4) is computed by a sequence of existing solvers for linear dispersive equations, ,
either for a fixed number of iterations or terminating when the update in step , , is smaller than some threshold .
In other words, is a discretisation of the solution map in step . We allow the dependence of the discrete solution operator upon because for different iterations we might use, for example, solvers of different accuracy.
For the cubic NLS (5), for instance, we solve a sequence of linear Schrödinger equations,
where
acts as a time-dependent potential.
The Magnus expansion
The solution of linear dispersive equations with time-dependent Hamiltonians can be approximated effectively using the Magnus expansion [2, 16, 8] (Kormann, Holmgren & Karlsson 2008), Specifically, where, for example, a fourth-order Magnus expansion for the linear dispersive equation
with time-dependent vector field is
Supposing (2), the PDE (4) is linear,
and we may set . Using the Magnus expansion, the solution can be approximated as
where indicates explicitly the dependence of the vector field on the previous iterate.
The expression (10) is not a ‘real’ numerical method because in practice it requires a semidiscretisation of , computation of the integrals in (9) and the computation (or approximation) of matrix exponential. Each of these steps might be costly – cf. [1] for a more detailed discussion – and ideally we need additional ideas to bring down the expense of the computation.
If the linear self-adjoint operator can be decomposed into time-independent and time-dependent components, , the commutator in (9) can be simplified as
The Magnus expansion (9) reduces to
where are the line integrals defined in Iserles, Kropielnicka & Singh [12]; Iserles, Kropielnicka & Singh [14],
and is the rescaled th Bernoulli polynomial. It is known that and .
In the context of the linearised Schrödinger equations (8) that appear in the iterative solution of the cubic NLS (5), and . In the context of the Gross–Pitaevskii equation, and is the external time-dependent potential.
Computing with information at the endpoints
The integrals in (9) can be computed to fourth order using just two Gaussian points in (Iserles et al. [16]) but this takes no account of the iterative nature of (4). To explain why, consider the second iteration : its computation means that we need to evaluate at the two Gaussian points, , where .
This means, though, that we need three values of : and to approximate , while (again) and are required to approximate . In the general case, it is easy to ascertain that we need function evaluations of to approximate to fourth order. This clearly is suboptimal and we wish to avoid this unwelcome linear growth.1
The remedy to this state of affairs is to use Hermite quadrature (Micchelli & Rivlin 1973), employing just the values of the integrand and its derivatives at the endpoints. To attain order 4 in both integrals, we need to know just the integrand and its derivative at and : altogether evaluations for iterations.
The fourth-order Magnus expansion (9) requires a fourth-order approximation of the integrals. This can be achieved via Hermite quadrature utilising only the Hamiltonian and its time-derivative, , , and , at the two end points of the time interval . The main idea, following (Iserles & Nørsett (1999); Iserles et al. [16]; Zanna (1998)), is to form a Hermite interpolation polynomial (in the present case a cubic) , say, replace by in (9) and integrate exactly the ensuing polynomial expressions.
More concretely, this can be directly applied to the approximation of the integrals in (11). In particular, since is expected to be known for all , a variety of quadrature approximations can be utilised for computing up to the required accuracy of . In contrast, is not known at all and we use the fourth-order Hermite approximations,
for approximating and .
General dispersive PDE. For a general dispersive PDE (1.1) the nonlinear function is assumed to be of the form and the derivative can be computed as
The cubic nonlinear Schrödinger (NLS) and Gross–Pitaevskii (GP) equations. In the case of cubic nonlinearity in the NLS or GP equations, , , and , where is the (conjugate) linear functional , while is the linear functional . Consequently,
which can be readily evaluated at and . The distinction between the NLS and GP equations is that in the case of the NLS, whereas in the case of GP.
Pseudocode for cubic NLS and GP
A schematic pseudocode for our algorithm, as applied to the cubic NLS and Gross–Pitaevskii equation, is described in Algorithm 2.1.
Algorithm 2.1 (Iterative solver for cubic NLS and GP).
Input: $\lambda$, $T$, $N$ and $u_0 \in \mathcal{H}$
1: $h=T/N$
2: for $n=0,1,2,\ldots,N$ do
3: $V_n^{\mathrm{nl}}=\lambda|u_n|^2$
$v_n=(-\Delta+V^0+V^e(t_n)+V_n^{\mathrm{nl}})u_n$
$dV_n^{\mathrm{nl}}=-2i\lambda\operatorname{Re}\langle u_n,v_n\rangle$
4: $\mu^{\mathcal{P}}_{0,0},\mu^e_{1,1}=\operatorname{Quadrature}(V^e(t),t_n,h)$
5:
6: for do
7:
8:
9:
10:
11:
12: end for
13:
14: end for
For most problems of interest suffices for fourth-order convergence, cf. Theorem 3. Note that any reasonable quadrature can be used to compute and .
Approximation of the exponential
The exponential of the Magnus expansion , (2.5), in Algorithm 2.1 only needs to be computed to a fourth-order accuracy. This can be achieved using a wide range of methods for matrix exponential (Moler & Van Loan 1978). Among these, some of the most widely utilised methods are splitting methods [3] (McLachlan & Quispel 2002) and Krylov subspace methods (Saad 2003).
Splitting methods are particularly useful when the exponent can be split into two parts, which are easy to exponentiate separately. In the context of the cubic NLS and GP equations, the exponent is a sum of a Laplacian, a multiplication operator, and a commutator involving a Laplacian and a multiplicative operator.
In the absence of the commutator, splitting methods could be effectively used to compute the exponential of the Magnus expansion (2.5). However, the presence of the commutator makes this infeasible. The presence of the commutator can also make Krylov iterations more expensive. However, if required, this commutator can be eliminated with minimal additional cost.
We note that if and commute, a simple application of the Baker–Campbell–Hausdorff formula shows that
Consequently,
where
The central exponential in (12) is now free of commutators and can be approximated by using any method for computing the matrix exponential. In particular, any fourth-order splitting can be employed, provided exponentials of , and are readily computable: this is contingent upon their structure following discretisation. Due to linearity of the integral, and should have the same structure as and , respectively.
In the context of cubic NLS and Gross–Pitaevskii equations, where is a function, we may also approximate the central exponential by utilising compact fourth-order splittings such as Chin & Chen [5] and Omelyan, Mryglod & Folk 2003 since the first component is a Laplacian and the second component is a multiplication operator.
Convergence
It is convenient to assume in this section that the underlying domain is , noting that a generalisation to many other boundary conditions (zero Dirichlet or Neumann conditions in a compact domain, a Cauchy problem in ) is straightforward. We further assume in this section that the Hamiltonian can be written in the form , understanding that, for example, needs to be taken for the NLS (3).
Our starting point is the equation (1), where we assume that for some and acts weakly on . We note that the self-adjointness of the Hamiltonian implies that
where is the standard inner product. An immediate consequence is that
and the norm is preserved for all .
The concern of this section is with the convergence of the iterative algorithm (8). We let
Lemma 1 For every it is true that
where .
Proof It is an immediate consequence of (1) and (8) that
Similarly to [9], we observe that
on the other hand
Comparing the two expressions,
However, (15) implies that , therefore
Moreover, and integration yields
for . Simple application of the Cauchy–Schwarz inequality results in (16) and concludes the proof. □
Our next goal is to identify the order of each iteration, i.e. the number (if it exists) such that
for some . Were we to prove that the s form a strictly increasing sequence and the s form at most geometrically increasing sequence (i.e., ), the convergence of the iterative scheme (2.2) for sufficiently small will be assured.
We assume that is Fréchet differentiable and denote its Fréchet derivative by . Note that, being a differential operator, this is a somewhat delicate assumption. For example, for the Schrödinger operator (1.3), namely , the potential part presents no problem but the kinetic part, , requires more consideration. The Fréchet derivative of is the least number such that : this is a consequence of being a linear operator. Finding is easy, provided that everything lives in – in that case
and maximising over results in . Taking , which is typically sufficient for the consideration of the equation (1.3) in its weak form, falls short of the requirements of our convergence analysis.
Since
we deduce that
It now follows from (16) by induction that
Theorem 2 Given that is Fréchet differentiable in , it is true that
and the method (8) converges for .
Proof Since , the s form a geometric sequence and convergence follows.
The inductive argument in the theorem can be commenced from , whereby implies immediately that . In other words, each iteration increases the order by at least one unit! The clear practical implication is that, once the goal is to attain specific temporal order, we need to iterate (8) certain number of times, thereby stoping, without any convergence tests.
Note four important points:
The notion of order in Theorem 2 is different from the standard concept of order in time-stepping schemes, which makes sense only in the presence of adequate regularity because the error of time-stepping schemes (at any rate, those that fit into B-series formalism [7]) is expressed in terms of high derivatives and elementary differentials. Order in the sense of Theorem 2, though, is independent of regularity, as long as the ambient space is smooth enough for to be bounded.
The operator need not be bounded: this is precisely the reason why we employ Fréchet derivatives (which remain bounded locally, for specific function ), rather than operator norms.
The statement of Theorem 2 is about an exact solution of linear equations in (8), while practical computation requires the use of numerical methods, of the kind to which we have alluded in Section 1. The order of the method, however, remains the same as long as the order of the ‘linear’ method used in solving for , is at least .
Since in general the higher the order of a time-stepping method, the greater its computational cost, we conclude that it is considerably more efficient to vary the order of the ‘linear’ method in the course of the iteration: done right, this reduces the cost without affecting the rate of convergence.
However, for a large set of problems (1.1) we can do better than that! Suppose thus that is a function of only – with a minor abuse of notation we write . As in (1.2), , except that now, abusing notation again, . For a want of a better name, we might term such problems as being of Schrödinger type.
We solve the equation
using the Strang splitting [7],
The second step in the Strang splitting is an ODE. Consider though the solution of the ODE
– this is a linear equation whose exact solution is . In particular, , therefore and the solution of the linear ODE is also the solution of
In other words, the solution of (3.5) is identical to that of
The system (3.6), however, is exactly what we obtain by applying Strang splitting in the first step of the iterative algorithm (2.2), applied to the PDE (3.4). To rephrase, is nothing else but the Strang-splitting solution of (3.4). Since Strang splitting is a second-order method, we deduce that . This explains why in the proof of Theorem 2 we commenced the iterative argument from and it is worthwhile to formulate the new result as a theorem.
Theorem 3 Let , where , is a self-adjoint linear differential operator and a real-valued scalar function. Provided that the first step of the algorithm (2.2) is solved by the Strang splitting, it is true that , and in place of (3.3) we have the error bound
As before, convergence is assured provided that .
Preservation of the norm, momentum and Hamiltonian energy in cubic Schrödinger equation
As in the heart of geometric numerical integration lays preservation of qualitative behaviour of solutions, in this section we are concerned with the preservation of major qualitative features of the nonlinear Schrödinger equation. Specifically, we wish to ascertain the preservation of , momentum and Hamiltonian energy under iterated linearization, as described in subsection 2, once they are applied to equation (5). Throughout this section we assume that the NLS (5) is solved in a cube in with periodic boundary conditions. A generalisation to zero Dirichlet or Neumann boundary conditions is straightforward.
L2 energy preservation
The energy of the cubic NLS equation (1) with periodic boundary conditions is conserved and a trivial – yet pleasing – feature of our method is that so does its numerical solution by means of the method (4).
Theorem 4 Subject to periodic boundary conditions it is true that and the method (4) preserves energy.
Proof Each step of (4) involves a solution of a linear Schrödinger equation (8), which itself preserves energy subject to the above boundary conditions. This remains true once the solution of (8) is itself computed by the methods of Section 2, because they also preserve energy.
Realistic application of Theorem 4 requires for the linear equation to be discretised and, to retain the spirit of the theorem, the linearised equation must discretised by a time-stepping method that respects norm conservation, e.g. the methods from [10, 14]. One crucial consequence of the theorem is that the iterated linearisation (4) is stable, provided that the linear equation is solved stably.
Conservation of momentum
Given boundary conditions either in an interval with periodic boundary conditions or in , the NLS (1) conserves the momentum
(Killip, Tao & Vișan 2009). Therefore, in a periodic setting in ,
and, integrating by parts,
This can be easily generalised componentwise to or and we deduce that
Theorem 5 The method (1.4) conserves momentum.
Hamiltonian energy
Hamiltonian energy of cubic Schrödinger equation in our setting is given by
and is preserved in time [6]. For completeness, we prove this statement: it is enough to show that its time derivative equals zero. Subject to the notational convention that we observe that
It is evident from (22) and (23) that so the Hamiltonian energy of cubic Schrödinger equation is preserved. The same cannot be concluded for , the solution of kth linearised iteration
Indeed, given that the Hamiltonian of the linear Schrödinger equation is
all we can deduce from (24) is that
is conserved under the flow and, at a first glance, a factor of is missing. Fortunately, it is recovered once tends to infinity and the following theorem establishes the rate of convergence.
Theorem 6 Once iterated linearisation is applied to the NLS (1.5), it is true that
Proof We commence by recalling from Theorem 3 that and observing that (21) is equivalent to
Moreover,
Since , we have
Therefore
since
because of the energy conservation. Thus, implies (25) and we are done.
Numerical examples
In this section we present numerical evidence to support the theoretical developments. Our approach of iterative linearisation using Magnus expansions in conjunction with Hermite quadratures is denoted by MH throughout this section. We combine it with two different methods for computing the matrix exponential: splitting methods and Krylov subspace methods.
MH with splitting methods
In this section, we consider splitting methods for the exponentiation of the linearised Magnus expansion. We note that classical splitting methods such as the Strang splitting Strang (1968) and Blanes–Moan splittings Blanes & Moan [4], referred to as BM in this section, can be applied directly to nonlinear and non-autonomous equations.
It is possible to utilise such splittings for the exponentiation of the central term in (12), the commutator-eliminated form of our iteratively linearised Magnus-Hermite (MH) scheme. The combination of MH and BM is termed MHBM in the sequel. When splittings methods can be directly applied to the nonlinear equations, they typically perform better, as seen by comparing BM vs MHBM in Figure 5.2.
Figure 5.2. error in for Example 1
However, a large class of splittings that feature commutators, such as compact splittings Chin & Chen [5], Omelyan et al. (2003), asymptotic splittings such as symmetric Zassenhaus splittings Bader et al. [1], Magnus–Zassenhaus splitting Iserles, Kropielnicka & Singh [15], and specialised splittings for laser-matter interaction Iserles, Kropielnicka & Singh [13] have been developed in recent years, and there is no straightforward way of extending their application to nonlinear equations. We demonstrate that our iterative linearisation procedure allows the use of such splittings for nonlinear equations. In particular, we combine our Magnus-Hermite (MH) schemes with the Chin–Chen splitting Chin & Chen [5], after the elimination of the commutators (12). This method is referred to as MHC in the sequel.
We consider the following examples, where the Hamiltonian is given by
where the total potential is a function of , , and , containing a static potential , an external field , and a nonlinear potential :
where
We chose the static potential
the external field
and the initial condition as the as wave packet
where and . We consider the solution on a periodic spatial grid , discretised with 1000 grid points, and use a normalised version of as an initial condition .
Example 1 (GP, defocusing, with external time-dependent field). In this example, we consider the potential
with static potential , external field , and nonlinear potential . The evolution of solution is shown in Figure 5.1, while the convergence of the error of at computed by different methods is shown in Figure 5.2. Observables over time are shown in Figure 5.3, and the change in observables with time is shown in Figure 5.4.
Figure 5.1. The linear part of the potential, (scaled down by a factor of 100 for ease of visualisation), real and imaginary parts and absolute value of the wave function at initial time (left), intermediate time (middle) and final time (right), where .
Figure 5.3. Observables over time for Examples 1, 2, 3 and 4.
Figure 5.4. Change in observables for Examples 1, 2, 3 and 4 over time, where is the number of time steps.
Example 2 (GP, defocusing, without external time-dependent field). In this example, we consider the potential
while the external field is absent.
Example 3 (GP, focusing, without external time-dependent field). In this example, we consider the potential
while the external field is absent.
Example 4 (NLS, defocusing). In this example, we consider the potential
i.e., with the nonlinear potential only.
Observables over time for all the examples computed with the MHC method are shown in Figure 5.3. The energy is always preserved. The momentum is preserved only in case of the NLS, i.e, Example 4. The Hamiltonian energy is preserved up to except in the scenario presented in Example 1, which involves an external time-independent field. These behaviours are in line with the discussion in § 4. The changes in observables with time are shown in Figure 5.4.
MH with a Krylov subspace method
In this subsection, we consider an example where a splitting approach is not feasible. For example, when the Hamiltonian is a large sparse matrix, the Krylov method is particularly useful. Consider
where the linear self-adjoint operator can be decomposed into time-independent and time-dependent components, , where , and , with and being matrices for discretised on points.
We combine our iteratively linearised Magnus-Hermite expansion (2.5) without elimination of commutators with a Lanczos approximation of the matrix exponential [11] (Park & Light 1986; Saad 2003). The Lanczos method is a Krylov subspace method that is well suited to Hermitian and skew-Hermitian matrices and is one of the most commonly applied methods in quantum dynamics. We use the mkprop package (Jawecki 2024) for the implementation of the Lanczos method, with an accuracy . The overall scheme is denoted MHK.
We use randomly generated Hermitian matrix and , and defocusing cubic nonlinear potential
Example 5 (nonlinear, with external time-dependent field). We consider external time-dependent field
Example 6 (nonlinear, without external time-dependent field). We consider no external time-dependent field
In Examples 5 and 6, the Hamiltonian energy is preserved as desired, as can be seen in Figure 5.5 (left). The convergence of the error of at computed by Krylov method for Examples 5 and 6 is displayed in Figure 5.5 (right).
Figure 5.5. Left: change of observables over time for Example 6. Right: convergence of the error of at computed by different methods for Examples 5, 6.
Conclusions
The main theme of this paper is an iterated linearisation of dispersive equations (1), converting them into the solution of simpler systems (1.4. In particular, in the case of NLS (5), we convert a nonlinear problem into an iterated solution of significantly simpler linear Schrödinger equations, for which very effective and cheap methods are available. We proved that, in the case of NLS, this approach conserves energy and approximates well the Hamiltonian.
However, there is nothing special in dispersive equations or in the form (1). Our approach applies to numerous PDEs of evolution. An example of a dispersive equation that does not fit into the (1) pattern is the Korteveg–de Vries (KdV) equation
with an initial condition and suitable boundary conditions. An iterated linearisation is
and it relies on the fact that the linear equation
where is known, is considerably easier to solve. Iterated linearisation can be also applied to non-dispersive PDEs. Consider, for example, the parabolic Fitzhugh–Nagumo system
from mathematical biology ((Izhikevich 2007)), given with an initial condition , and Dirichlet boundary conditions. and are given functions, typically low-degree polynomials, while are given constants. Provided that and , we may consider the iterated linearisation
The computation of a diffusion equation is an easy numerical endeavour and this can be leveraged toward an effective iteration of the form (31).
Note that both (29) and (31) are merely at the level of preliminary ideas, without any stability or convergence analysis. Moreover, both our examples, KdV (28) and the Fitzhugh–Nagumo system (30), possess a wealth of qualitative features which remain unexplored in the context of iterated linearisation. We mention the two equations mostly to illustrate the potential scope of our approach beyond the realm of dispersive equations.
Acknowledgements
The work of GC in this project has been supported by EPSRC (EP/S022945/1). The work of KK in this project has been supported by The National Center for Science (NCN), based on Grant No. 2019/34/E/ST1/00390.
References
- [1]Bader, P., Iserles, A., Kropielnicka, K. & Singh, P. (2014), ‘Effective approximation for the semiclassical Schrödinger equation’, Found. Comput. Math. 14(4), 689–720. URL: https://doi.org/10.1007/s10208-013-9182-8
- [2]Blanes, S. & Casas, F. (2016), A concise introduction to geometric numerical integration, Monographs and Research Notes in Mathematics, CRC Press, Boca Raton, FL.
- [3]Blanes, S., Casas, F. & Murua, A. (2024), ‘Splitting methods for differential equations’, Acta Numerica. To appear.arxiv.org/abs/2401.01722
- [4]Blanes, S. & Moan, P. C. (2002), ‘Practical symplectic partitioned Runge–Kutta and Runge–Kutta–Nyström methods’, Journal of Computational and Applied Mathematics 142(2), 313–330.
- [5]Chin, S. A. & Chen, C. R. (2005), ‘Forward symplectic integrators for solving gravitational few-body problems’, Celestial Mech. Dynam. Astron. 91(3-4), 301–322. URL: https://doi.org/10.1007/s10569-004-4622-z
- [6]Faou, E. (2012), Geometric numerical integration and Schrödinger equations, Zurich Lectures in Advanced Mathematics, European Mathematical Society (EMS), Zürich. URL: https://doi.org/10.4171/100
- [7]Hairer, E., Lubich, C. & Wanner, G. (2006), Geometric numerical integration, Vol. 31 of Springer Series in Computational Mathematics, second edn, Springer-Verlag, Berlin.
- [8]Hochbruck, M. & Lubich, C. (2003a), ‘On magnus integrators for time-dependent schrödinger equations’, SIAM journal on numerical analysis 41(3), 945–963.DOI
- [9]Hochbruck, M. & Lubich, C. (2003b), ‘On Magnus integrators for time-dependent Schrödinger equations’, SIAM J. Numer. Anal. 41(3), 945–963. URL: https://doi.org/10.1137/S0036142902403875
- [10]Iserles, A., Kropielnicka, K. & Singh, P. (2018a), ‘Magnus–Lanczos methods with simplified commutators for the Schrödinger equation with a time-dependent potential’, SIAM J. Numer. Anal. 56(3), 1547–1569. URL: https://doi.org/10.1137/17M1149833
- [11]Iserles, A., Kropielnicka, K. & Singh, P. (2018b), ‘Magnus–Lanczos methods with simplified commutators for the Schrödinger equation with a time-dependent potential’, SIAM Journal on Numerical Analysis 56(3), 1547–1567.
- [12]Iserles, A., Kropielnicka, K. & Singh, P. (2019a), ‘Compact schemes for laser–matter interaction in schrödinger equation based on effective splittings of magnus expansion’, Comput. Phys. Comm. 234, 195–201.arxiv.org/abs/1805.09163
- [13]Iserles, A., Kropielnicka, K. & Singh, P. (2019b), ‘Compact schemes for laser–matter interaction in schrödinger equation based on effective splittings of magnus expansion’, Computer Physics Communications 234, 195–201.arxiv.org/abs/1805.09163
- [14]Iserles, A., Kropielnicka, K. & Singh, P. (2019c), ‘Solving Schrödinger equation in semiclassical regime with highly oscillatory time-dependent potentials’, J. Comput. Phys. 376, 564–584. URL: https://doi.org/10.1016/j.jcp.2018.09.047
- [15]Iserles, A., Kropielnicka, K. & Singh, P. (2019d), ‘Solving schrödinger equation in semiclassical regime with highly oscillatory time-dependent potentials’, Journal of Computational Physics 376, 564–584.
- [16]Iserles, A., Munthe-Kaas, H. Z., Nørsett, S. P. & Zanna, A. (2000), ‘Lie-group methods’, Acta Numerica 9, 215–365. URL: https://doi.org/10.1017/S0962492900002154