Abstract
We consider commutator-free exponential integrators as put forward in Alverman and Fehske (J Comput Phys 230:5930–5956, 2011). For parabolic problems, it is important for the well-definedness that such an integrator satisfies a positivity condition such that essentially it only proceeds forward in time. We prove that this requirement implies maximal convergence order of four for real coefficients, which has been conjectured earlier by other authors.
Similar content being viewed by others
Notes
Note that the situation is similar to that encountered for exponential splitting methods. For this class of time integrators, no methods of order greater than two exist with only positive real coefficients. This was first shown in [22], see also [13]. The ensuing instability can be avoided by splitting with complex coefficients [14], see also [10]. Splitting methods of high order with complex coefficients have been constructed for example in [5]. Similarly, for exponential commutator-free Magnus-type methods, stable high-order schemes with complex coefficients have been derived in [7]. Moreover, unconventional schemes involving additionally evaluation of some commutators are introduced there which are stable for parabolic problems. An error analysis of high-order commutator-free exponential integrators applied to semi-discretizations of parabolic problems is given in [8].
References
Alverman, A., Fehske, H.: High-order commutator-free exponential time-propagation of driven quantum systems. J. Comput. Phys. 230, 5930–5956 (2011)
Alverman, A., Fehske, H., Littlewood, P.: Numerical time propagation of quantum systems in radiation fields. New J. Phys. 14, 105008 (2012)
Bader, P., Iserles, A., Kropielnicka, K., Singh, P.: Effective approximation for the linear time-dependent Schrödinger equation. Found. Comput. Math. 14, 689–720 (2014)
Bader, P., Iserles, A., Kropielnicka, K., Singh, P.: Efficient methods for linear Schrödinger equation in the semiclassical regime with time-dependent potential. Proc. R. Soc. A 472, 20150733 (2016)
Blanes, S., Casas, F., Chartier, P., Murua, A.: Optimized high-order splitting methods for some classes of parabolic equations. Math. Comp. 82, 1559–1576 (2013)
Blanes, S., Casas, F., Oteo, J., Ros, J.: The Magnus expansion and some of its applications. Phys. Rep. 470, 151–238 (2008)
Blanes, S., Casas, F., Thalhammer, M.: High-order commutator-free quasi-Magnus integrators for non-autonomous linear evolution equations. Comput. Phys. Commun. 220, 243–262 (2017)
Blanes, S., Casas, F., Thalhammer, M.: Convergence analysis of high-order commutator-free quasi-Magnus exponential integrators for nonautonomous linear evolution equations of parabolic type. IMA J. Numer. Anal. 38, 743–778 (2018)
Blanes, S., Moan, P.: Fourth- and sixth-order commutator-free Magnus integrators for linear and non-linear dynamical systems. Appl. Numer. Math. 56, 1519–1537 (2005)
Castella, F., Chartier, P., Descombes, S., Vilmart, G.: Splitting methods with complex times for parabolic equations. BIT Numer. Math. 49, 487–508 (2009)
Celledoni, E.: Eulerian and semi-Lagrangian schemes based on commutator-free exponential integrators. Group Theory Numer. Anal. CRM Proc. Lecture Notes 39, 77–90 (2005)
Celledoni, E., Marthinsen, A., Owren, B.: Commutator-free Lie-group methods. Future Gen. Comput. Syst. 19(3), 341–352 (2003)
Goldman, D., Kaper, T.: \(n\)th-order operator splitting schemes and nonreversible systems. SIAM J. Numer. Anal. 33(1), 349–367 (1996)
Hansen, E., Ostermann, A.: High order splitting methods for analytic semigroups exist. BIT Numer. Math. 49, 527–542 (2009)
Horn, R., Johnson, C.: Matrix Analysis. Cambridge University Press, Cambridge (1985)
Iserles, A., Nørsett, S.: On the solution of linear differential equations on Lie groups. Phil. Trans. R. Soc. Lond. A 357, 983–1019 (1999)
Magnus, W.: On the exponential solution of differential equations for a linear operator. Commun. Pure Appl. Math. 7, 649–673 (1954)
Moler, C., Van Loan, C.: Nineteen dubious ways to compute the exponential of a matrix, twenty-five years later. SIAM Rev. 45(1), 3–000 (2003)
Owren, B.: Order conditions for commutator-free Lie group methods. J. Phys. A: Math. Gen. 39, 5585–5599 (2006)
Park, T., Light, J.: Unitary quantum time evolution by iterative Lanczos reduction. J. Chem. Phys. 85, 5870–5876 (1986)
Saad, Y.: Analysis of some Krylov subspace approximations to the matrix exponential operator. SIAM J. Numer. Anal. 29(1), 209–228 (1992)
Sheng, Q.: Solving linear partial differential equations by exponential splittings. IMA J. Numer. Anal. 9(2), 199–212 (1989)
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work has been supported in part by the Austrian Science Fund (FWF) under Grant P30819-N32 and the Vienna Science and Technology Fund (WWTF) under Grant MA14-002.
Appendices
Proof of the order conditions (7)–(10)
For the global error to have order \(p=5\) it is required that the local error have convergence order \(p+1=6\). If, without restriction of generality, we consider only the first integration step for the special problem (5), this condition for the local error is written as
A Taylor expansion of the left-hand side leads to (let \({\mathbf {k}}=(k_1,\ldots ,k_m), \ |{\mathbf {k}}| =\sum _{l=1}^m k_l\))
Here for \(J=1\) we have [note that already a subset of coefficients suffices to derive the order conditions (7)–(10)],
and for \(J\ge 2\) the coefficients can be computed recursively,
An inductive argument involving straightforward but laborious calculations gives
Repeated differentiation of the differential equation (5) yields
Thus for the Taylor expansion of the right-hand side of (18) we obtain
with coefficients [only those corresponding to the subset of coefficients as in (19)]
Equating corresponding coefficients in (19) and (20) leads to the order conditions (7)–(10).
A geometric lemma
Lemma 1
Let \(\{a_1,\ldots ,a_m\}\) be a linearly independent set of vectors in \({\mathbb {R}}^n\) and \(S\in {\mathbb {R}}^{n\times n}\) symmetric positive definite. Further let \(c=(\gamma _1,\ldots ,\gamma _m)^T\in {\mathbb {R}}^m\) and \(\delta \in {\mathbb {R}}\). Then the intersection \({\mathcal {I}}\) of the m hyperplanes in \({\mathbb {R}}^n\) given by the equations \(a_1^Tx=\gamma _1,\ldots ,a_m^Tx=\gamma _m\) intersects the hyper-ellipsoid \({\mathcal {Q}}\) given by the equation \(x^TSx=\delta \) if and only if it holds
where \(\varGamma = (a_i^TS^{-1}a_j)_{i,j=1}^m\) denotes the Gram matrix of the vectors \(a_1,\ldots ,a_m\) with respect to the scalar product \(x^TS^{-1}y\).
Proof
First we consider the special case \(S=I_n\) (identity matrix), where \({\mathcal {Q}}\) is a hyper-sphere. In this case \({\mathcal {I}}\) intersects \({\mathcal {Q}}\) if and only if the point \(x_*\in {\mathcal {I}}\) of minimal norm satisfies
It is easy to see that this point \(x_*\) lies in the linear subspace of \({\mathbb {R}}^n\) spanned by \(a_1, \ldots a_m\) (the normal vectors to the given hyperplanes), i.e., there exists \(b=(\beta _1,\ldots ,\beta _m)^T\in {\mathbb {R}}^m\) such that
where \(A=[a_1\ \cdots \ a_m]\in {\mathbb {R}}^{n\times m}\). Because \(x_*\in {\mathcal {I}}\) it holds
and thus
which shows that (22) is equivalent to (21). This completes the proof for the special case \(S=I_n\).
For the general case, the symmetric positive definite matrix S can be written as
with \(\varLambda =\mathrm {diag}(\lambda _1,\ldots ,\lambda _n)\), where \(\lambda _j>0\) are the eigenvalues of S, and X orthogonal. We define \({\widetilde{a}}_j=\varLambda ^{-1/2}X^Ta_j\), \(j=1,\ldots ,m\). Then under the transformation of variables \(\widetilde{x}=\varLambda ^{1/2}X^Tx\), the equation \({\widetilde{a}}_j^T{\widetilde{x}}=\gamma _j\) is equivalent to \(a_j^Tx=\gamma _j\) and \({\widetilde{x}}^T{\widetilde{x}}=\delta \) is equivalent to \(x^TSx=\delta \). For these transformed equations the special case from above is applicable. Using \({\widetilde{A}} = [{\widetilde{a}}_1\ \cdots \ {\widetilde{a}}_m] = \varLambda ^{-1/2}X^TA\) with \(A=[a_1\ \cdots \ a_m]\) it follows that for the transformed equations the corresponding Gram matrix satisfies \({\widetilde{\varGamma }}={\widetilde{A}}^T{\widetilde{A}}=A^TX\varLambda ^{-1}X^TA=A^TS^{-1}A = \varGamma \) as claimed. \(\square \)
Maple code for checking (16)=(17)
Here, the Maple identifiers s, s1, bb, dd, eSe, eSd, dSd correspond to \(\sigma \), \({\widetilde{\sigma }}\), \(b_{J+1}\), \(d_{J+1}\), \(e^TS^{-1}e\), \(e^TS^{-1}d\), \(d^TS^{-1}d\), respectively.
Rights and permissions
About this article
Cite this article
Hofstätter, H., Koch, O. Non-satisfiability of a positivity condition for commutator-free exponential integrators of order higher than four. Numer. Math. 141, 681–691 (2019). https://doi.org/10.1007/s00211-018-1015-x
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00211-018-1015-x