Abstract
It is well known that Lawson methods suffer from a severe order reduction when integrating initial boundary value problems where the solutions are not periodic in space or do not satisfy enough conditions of annihilation on the boundary. However, in a previous paper, a modification of Lawson quadrature rules has been suggested so that no order reduction turns up when integrating linear problems subject to time-dependent boundary conditions. In this paper, we describe and thoroughly analyse a technique to avoid also order reduction when integrating nonlinear problems. This is very useful because, given any Runge–Kutta method of any classical order, a Lawson method can be constructed associated to it for which the order is conserved.
Similar content being viewed by others
References
Alonso-Mallo, I., Cano, B., Reguera, N.: Avoiding order reduction when integrating linear initial boundary value problems with exponential splitting methods. IMA J. Numer. Anal. 38, 1294–1323 (2018)
Alonso-Mallo, I., Cano, B., Reguera, N.: Analysis of order reduction when integrating linear initial boundary value problems with Lawson methods. Appl. Numer. Math. 118, 64–74 (2017)
Alonso-Mallo, I., Cano, B., Reguera, N.: Avoiding order reduction when integrating linear initial boundary value problems with Lawson methods. IMA J. Numer. Anal. 37, 2091–2119 (2017)
Alonso-Mallo, I., Cano, B., Reguera, N.: Avoiding order reduction when integrating reaction–diffusion boundary value problems with exponential splitting methods. J. Comput. Appl. Math. 357, 228–250 (2019)
Cano, B., Moreta, M.J.: Exponential quadrature rules without order reduction for integrating linear initial boundary value problems. SIAM J. Numer. Anal. 56–3, 1187–1209 (2018)
Cano, B., Reguera, N.: Avoiding order reduction when integrating nonlinear Schrödinger equation with Strang method. J. Comp. Appl. Math. 316, 86–99 (2017)
Cano, B., Reguera, N.: Order reduction and how to avoid it when Lawson methods integrate reaction-diffusion boundary value problems, http://arxiv.org/abs/1909.12659
Cano, B., Reguera, N.: Why improving the accuracy of exponential integrators can decrease their computational cost? Mathematics 9, 1008 (2021). https://doi.org/10.3390/math909100
Einkemmer, L., Ostermann, A.: Overcoming order reduction in diffusion-reaction splitting. Part 1: Dirichlet boundary conditions. SIAM J. Sci. Comput. 37(3), A1577–A1592 (2015)
Einkemmer, L., Ostermann, A.: Overcoming order reduction in diffusion-reaction splitting. Part 2: Oblique boundary conditions. SIAM J. Sci. Comput. 38, A3741–A3757 (2016)
Faou, E., Ostermann, A., Schratz, K.: Analysis of exponential splitting methods for inhomogeneous parabolic equations. IMA J. Numer. Anal. 35(1), 161–178 (2015)
Hairer, E., Lubich, C.: Solving Ordinary Differential Equations II, Stiff and Differential Algebraic Problems. Second Revised Ediyion, Springer (1996)
Hochbruck, M., Ostermann, A.: Explicit exponential Runge–Kutta methods for semilinear parabolic problems. SIAM J. Numer. Anal. 43, 1069–1090 (2005)
Hochbruck, M., Ostermann, A.: Exponential Runge-Kutta methods for parabolic problems. Appl. Numer. Math. 53(2–4), 323–339 (2005)
Hochbruck, M., Ostermann, A.: Exponential integrators. Acta Numer. 209–286 (2010)
Lawson, J.D.: Generalized Runge–Kutta processes for stable systems with large Lipschitz constants. SIAM J. Numer. Anal. 4, 372–380 (1967)
Luan, V.T., Ostermann, A.: Explicit exponential Runge-Kutta methods of high order for parabolic problems. J. Comput. Appl. Math. 262, 361–372 (2014)
Niesen, J., Wright, W.M.: Algorithm 919: a Krylov subspace algorithm for evaluating the \(\varphi \)-functions appearing in exponential integrators, ACM Trans. Math. Software 38, no. 3, Art. 22 (2012)
Sanz-Serna, J.M.: Diez lecciones de Cálculo Numérico. Secretariado de Publicaciones e Intercambio Científico, Universidad de Valladolid (1998)
Strehmel, K., Weiner, R.: Linear-implizite Runge-Kutta Methoden und ihre Anwendungen. Teubner, Leipzig (1992)
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Mechthild Thalhammer.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was funded by Ministerio de Ciencia e Innovación and Regional Development European Funds through project PGC2018-101443-B-I00 and by Junta de Castilla y León and Feder through projects VA169P20.
Appendix
Appendix
1.1 Proof of Theorem 3.1
Using Lemma 3.1 in [3],
Then,
where, for the last line, (3.9), (A4) together with (7.3) and the first condition of (2.11) have been used. From this, the first result on the local error follows.
As for the second result, looking at the term in \(k^2\) in \({\bar{u}}_{n+1}\) and using that \(u\in C^3([0,T],X)\), we can notice that
Using (2.7), (7.3), (A4) and (3.9), the first, third and fifth terms are \(O(k^3)\). As for the fourth one, considering (7.2), it can be written as
where, for the last equality, we have used (2.7) again, (3.10), the second condition in (2.11) and the fact that \(\sum _{i=1}^s b_i c_i=1/2\) due to the second order of the Butcher tableau. Inserting this in (7.5) and simplifying notation,
where the differentiation of (2.1) with respect to time shows that the term in bracket in the previous expression vanishes.
1.2 Proof of Theorem 3.3
Firstly notice that
Then, using (3.6), when considering Dirichlet boundary conditions, in which case \(\partial Au(t_n)\) and \(\partial f(t_n,u(t_n))\) are calculated exactly in terms of data, as \({\bar{U}}_{n+1,h}\) is the same as \(U_{n+1,h}\) but starting from \(P_h u(t_n)\) instead of \(U_{n,h}\),
where, recursively, for \(i=1,\dots ,s\),
In such a way, it is inductively proved that \(K_{n+1,h,i}-{\bar{K}}_{n+1,h,i}=O(e_{n,h})\) and finally, using (7.7) and (7.6),
from what the result follows from Theorem 3.2 by a summation-by-parts argument and a discrete Gronwall lemma in the same way than the proof of Theorem 22 in [4] for Strang method.
On the other hand, when considering Robin/Neumann boundary conditions, as, according to Remark 3.1, \(\partial f(t_n,u(t_n))\) is just calculated approximately with an error which is \(O(e_{n,h})\), using (3.6) again,
where \(K_{n+1,h,i}-{\bar{K}}_{n+1,h,i}\) is the same as in (7.8) because \(\partial u(t_n)\) is given exactly in terms of data with this type of boundary conditions. Then, using (2.7),
Using now (H2c), it follows that \(U_{n+1,h}-{\bar{U}}_{n+1,h}=e^{ k A_{h,0}} e_{n,h}+O(k e_{n,h})\), from what (7.9) applies again and the result follows in the same way as above.
1.3 Proof of Theorem 4.3
For the proof, as in Theorem 3.3, we must consider the decomposition (7.6) where \({\bar{U}}_{n+1,h}\) is calculated as \(U_{n+1,h}\) but starting from \(P_h u(t_n)\) and calculating the boundaries in (4.2) in an exact way. In contrast, according to Table 1, when considering \(U_{n+1,h}\), the boundaries in (4.2) can just be calculated approximately.
More precisely, with Dirichlet boundary conditions, the terms on the boundary for the stages in (4.2) can be calculated exactly. However, when calculating \(U_{n+1,h}\), \(\partial A^2 u(t_n)\) and \(\partial A f(t_n,u(t_n))\) can just be calculated except for \(O(\nu _h+\frac{e_{n,h}}{h^\gamma })\). Because of this,
where \(K_{n,h,i}-{\bar{K}}_{n,h,i}=O(e_{n,h})\) as in the proof of Theorem 3.3. Therefore, using (2.7) and (H2c),
from what, using condition (4.4),
The classical argument of convergence and the first part of Theorem 4.2 leads then to the first result of this theorem for Dirichlet boundary conditions. For the second part, the second part of Theorem 4.2 must be used, apart from (3.13) and the additional regularity (4.5).
On the other hand, with Robin/Neumann boundary conditions, there is some error when approximating the boundaries for both the stages and the numerical solution. More precisely, using Table 1 and (4.2),
and then
From this, under condition (4.4),
so that, using the first part of Theorem 4.2 and the classical argument of convergence, \(e_{n,h}=O(k^2+\varepsilon _h+k \nu _h+k \mu _{k,1})\). Again, under the second set of hypotheses in Theorem 4.2 and using (3.13) and the regularity (4.5), the finer result \(e_{n,h}=O(k^3+k \varepsilon _h+\eta _h+k \mu _{k,1}+k \nu _h)\) can be achieved.
1.4 Proof of Theorem 5.3
As in the proof of Theorem 4.3, we must consider the decomposition (7.6) and then study the difference \(U_{h,n+1}-{\bar{U}}_{n+1}\) taking into account that the boundaries for \(U_{h,n+1}\) in (5.1) are just calculated approximately with an error which is given through Table 1.
More precisely, with Dirichlet boundary conditions,
where
and, for the last equality, (2.7), (H2c) and (4.4) have been used. Inserting this in (7.10) and using again (2.7), (H2c) and (4.4), it follows that
From here,
and using a discrete Gronwall Lemma and the first part of Theorem 5.2, the first part of the theorem follows for Dirichlet boundary conditions. For the second part, the second part of Theorem 5.2 must be used, apart from (3.13) and the additional regularity (5.7).
As for Robin/Neumann boundary conditions, with similar arguments,
from what
From this,
so that, using the first part of Theorem 5.2 and the classical argument of convergence, \(e_{n,h}=O(k^3+\varepsilon _h+k \nu _h+k \mu _{k,1}+k^2 \mu _{k,2})\). Again, under the second set of hypotheses in Theorem 5.2 and using (3.13) and the regularity (5.7), the finer result \(e_{n,h}=O(k^4+k \varepsilon _h+\eta _h+k \mu _{k,1}+k^2 \mu _{k,2}+k \nu _h)\) is achieved.
Rights and permissions
About this article
Cite this article
Cano, B., Reguera, N. How to avoid order reduction when Lawson methods integrate nonlinear initial boundary value problems. Bit Numer Math 62, 431–463 (2022). https://doi.org/10.1007/s10543-021-00879-8
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10543-021-00879-8