Abstract
We introduce some general tools to design exact splitting methods to compute numerically semigroups generated by inhomogeneous quadratic differential operators. More precisely, we factorize these semigroups as products of semigroups that can be approximated efficiently, using, for example, pseudo-spectral methods. We highlight the efficiency of these new methods on the examples of the magnetic linear Schrödinger equations with quadratic potentials, some transport equations and some Fokker–Planck equations.
Similar content being viewed by others
Notes
I.e., a set equipped with an associative binary operation and an identity element.
Note that as usual this formalism include nonlinear equations, since it is enough to consider the transport equations they generate.
In some specific sense depending on the problem.
Note that here it could be proven more elementarily, applying Lemma 2.
Also called transvection matrices.
An inverse Fourier transform has the same cost as a direct one. For example, to compute the solution of the semigroup generated by the harmonic oscillator with the factorization (2), we need 2 1d-FFTs, one to go on the Fourier side and one other to come back.
The reader could refer, for example, to Lemma 3.10 in [2] for a detailed proof.
References
P. Alphonse, J. Bernier, Smoothing properties of fractional Ornstein-Uhlenbeck semigroups and null-controllability, Bull. Sci. math., 102914 (2020)
P. Alphonse, J. Bernier, Polar decomposition of semigroups generated by non-selfadjoint quadratic differential operators and regularizing effects, preprint, arxiv:1909.03662
J. Ameres, Splitting methods for Fourier spectral discretizations of the strongly magnetized Vlasov-Poisson and the Vlasov-Maxwell system, preprint, arxiv:1907.05319
V. I. Arnold, S. P. Novikov, Dynamical Systems IV, Symplectic Geometry and its Applications. Springer-Verlag, Berlin Heidelberg (2001)
P. Bader, Fourier-splitting methods for the dynamics of rotating Bose–Einstein condensates, Journal of Computational and Applied Mathematics, 336, 268–280 (2018)
P. Bader, S. Blanes, Fourier methods for the perturbed harmonic oscillator in linear and nonlinear Schrödinger equations, Phys. Rev. E, 83(4), 046711 (2011)
D. Bambusi, B. Grébert, A. Maspero, D. Robert, Reducibility of the quantum harmonic oscillator in \(d\)-dimensions with polynomial time-dependent perturbation, Anal. PDE, 11 no. 3, 775–799 (2018)
W. Bao ,Q. Du, Y. Zhang, Dynamics of rotating Bose–Einstein condensates and its efficient and accurate numerical computation, SIAM Journal on Applied Mathematics, 66(3), 758–786 (2006)
J. Bernier, N. Crouseilles, F. Casas, Splitting methods for rotations: application to Vlasov equations, SIAM Journal on Scientific Computing, 42(2), A666-A697 (2020)
J. Bernier, N. Crouseilles, Y. Li, Exact splitting methods for kinetic and Schrödinger equations, preprint, arXiv:1912.13221, to appear in Journal of Scientific Computing
S. Blanes, F. Casas, P. Chartier, A. Murua, Optimized high-order splitting methods for some classes of parabolic equations, Math. Comp., 82, 1559–1576 (2013)
F. Castella, P. Chartier, S. Descombes and G. Vilmart, Splitting methods with complex times for parabolic equations, Bit Numer Math, 49, 487 (2009)
N. Crouseilles, L. Einkemmer, E. Faou, Hamiltonian splitting for the Vlasov-Maxwell equations, J. Comput. Phys., 283, 224-240 (2015)
B. Chen, A. Kaufman, 3D volume rotation using shear transformation, Graphical Models, 62, 308–322 (2000)
S.A. Chin, E. Krotscheck, Fourth-order algorithms for solving the imaginary-time Gross-Pitaevskii equation in a rotating anisotropic trap, Phys. Rev. E, 72(3), 036705 (2005)
G. Dujardin, F. Hérau, P. Lafitte, Coercivity, hypocoercivity, exponential time decay and simulations for discrete Fokker-Planck equations, Numer. Math., 144, 615–697 (2020)
N. Dunford, J. T. Schwartz, Linear operators. Part I. Wiley Classics Library, John Wiley & Sons, Inc., New York, (1988)
E. Hairer, C. Lubich, G. Wanner, Geometric numerical integration: Structure-Preserving Algorithms for Ordinary Differential Equations. Springer Series in Computational Mathematics (2006)
F. Hérau, J. Sjöstrand, M. Hitrik, Tunnel effect for the Kramers-Fokker-Planck type operators, Ann. Henri Poincaré, 9, 209–274 (2008)
L. Hörmander, Symplectic classification of quadratic forms, and general Mehler formulas, Math. Z., 219, 413 (1995)
L. Hörmander, The analysis of linear partial differential operators. III, Pseudo-differential operators. Classics in Mathematics Springer, Berlin (2007)
R. A. Horn, C. R. Johnson, Matrix Analysis. Cambridge University Press (1985)
A. Kolmogorov, Zufällige Bewegungen (zur Theorie der Brownschen Bewegung), Ann. of Math., 35, 116–117 (1934)
F. Nicola, L. Rodino, Global Pseudo-Differential Calculus on Euclidean Spaces. Birkhäuser Basel (2010)
A. W. Paeth, A fast Algorithm for General Raster Rotation, Proc. Graphics Interface 36, Vancouver (Canada), 77–81 (1986)
K. Pravda-Starov, Generalized Mehler formula for time-dependent non-selfadjoint quadratic operators and propagation of singularities, Math. Ann., 372, 1335–1382, (2018)
J.J. Rotman, An introduction to the Theory of Groups. Springer-Verlag, New York (1995)
J. Viola, The elliptic evolution of non-self-adjoint degree-2 Hamiltonians, preprint, arxiv:1701.00801v1
J. Welling, W. Eddy, T. Young, Rotation of 3D volumes by Fourier-interpolated shears, Graphical Models, 68, 356–370 (2006)
Acknowledgements
The author thanks P. Alphonse, N. Crouseilles and Y. Li for many enthusiastic discussions about this work.
Author information
Authors and Affiliations
Corresponding author
Additional information
Communicated by Arieh Iserles.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Research of the author was supported by ANR project NABUCO, ANR-17-CE40-0025.
Appendix
Appendix
1.1 Proof of Lemma 2
Here, the proof relies essentially on computations by block requiring to introduce a more convenient basis than the canonical basis of \(\mathbb {C}^{2n+2}\) denoted by \(e_1,\dots ,e_{2n+2}\). This basis, denoted \(\mathscr {B}\), is just a permutation of the canonical basis and is defined by
In this basis the matrix of \(J_{2n+2}\) is
and the matrix of \(\mathbb {P}p_j\) is
Consequently, in this basis the matrix of the Hamiltonian map is
Here, it is really relevant to observe the double triangular structure of this matrix. The four blocks on the top left corner define an upper triangular matrix by blocks, whereas, considering these four blocks as a single block, the global matrix is lower triangular by blocks.
Now observe, through the power expansion series, that if \(\varPsi \) is an entire function and
is an upper triangular matrix by blocks, then
Consequently, we have
where
At the end of the proof, we will check that \(\widetilde{\kappa _j} = \kappa _j\), so for the moment assume that this relation holds.
Thus, realizing a product by block we get by a straightforward induction
Identifying the blocks (1, 1), (3, 1), (3, 2) with those of \(\mathrm {mat}_{\mathscr {B}} \ \varPhi _1^{\mathbb {P}p_{m+1}}\), we get the system (24). Conversely, we have to check that if (24) is satisfied, then the blocks (1, 2) are the same. Indeed, consider a complex symplectic matrix \(M\in \mathrm {Sp}_{2n}(\mathbb {C})\) with a block structure of the form
Note that since M is symplectic, M is invertible and consequently A is also invertible. Since M is symplectic, then is satisfies
However, due to the double triangular nature of M, its invert can be computed easily and we have
Consequently, a straightforward block product leads to
Thus, since M is symplectic, we have
A fortiori, if two symplectic matrices have this block structure and the same top left corner blocks, if their blocks (3, 1) are equal, then their blocks (1, 2) are equal. Consequently, applying this to the symplectic matrices \(\varPhi _1^{\mathbb {P}p_{1}} \dots \varPhi _1^{\mathbb {P}p_{m}}\) and \(\varPhi _1^{\mathbb {P}p_{m+1}}\), we deduce that if the system (24) is satisfied, then we have the factorization (23).
Finally, we just have to check that \(\widetilde{\kappa _j} = \kappa _j\). For this computation, we omit the indices j since they are clearly irrelevant. First, we split the even indices from the odd indices in the power expansion of \(\widetilde{\kappa }\) :
Observing that
since \(J_{2n}\) is skew-symmetric, we have
Consequently, \(\varSigma _{even}\) vanishes. Similarly, since we have
we get
1.2 Proof of Theorem 1
Before proving Theorem 1, let us to prove some preparatory lemmas.
Lemma 5
If L is a bounded operator on \(L^2(\mathbb {R}^n)\) such that there exists \(T\in \mathrm {Sp}_{2n}(\mathbb {R})\) satisfying \(L=\pm \mathscr {K}(T)\), then L can be computed by an exact splitting.
Proof
Let G be the group generated by \(\varPhi _t^{i x_j^2},\varPhi _t^{i \xi _j^2},\varPhi _t^{i x_j \xi _k}\) for \(t\in \mathbb {R}, j,k\in \llbracket 1,n \rrbracket \). Applying Theorem 3, if we prove that \(G= \mathrm {Sp}_{2n}(\mathbb {R})\), we deduce that L is a product of operators of the form \(e^{it x_j^2 },e^{it \partial _{x_j}^2},e^{t x_j \partial _{x_k}}\) (up to the sign). Thus, since Proposition 6 states that dilatations (i.e. operators of the form \(e^{t x_k \partial _k}\)) can be factorized similarly, we would deduce that L can be computed by an exact splitting.
Consequently, we aim at proving that \(G= \mathrm {Sp}_{2n}(\mathbb {R})\). First, let us prove that G contains a neighborhood of the identity in \(\mathrm {Sp}_{2n}(\mathbb {R})\).
Indeed, consider the map \(\varPsi : \mathscr {N} \rightarrow \mathfrak {sp}_{2n}(\mathbb {R})\), where \(\mathscr {N}\) is a neighborhood of the origin in \(\mathfrak {sp}_{2n}(\mathbb {R})\), defined for \(Q\in S_{2n}(\mathbb {R})\) such that \(J_{2n}Q \in \mathscr {N}\) by
where the natural block decomposition of Q is
with \(A,B\in S_{n}(\mathbb {R})\) and \(C\in M_n(\mathbb {R})\). Note that to prove that \(\varPsi \) takes its values in \(\mathfrak {sp}_{2n}(\mathbb {R})\), it is enough to apply the Baker–Campbell–Hausdorff formula.
Since the differential of the exponential in the origin and the differential of the logarithm in the identity are equal to the identity, we deduce by composition that the differential of \(\varPsi \) in the origin is also the identity. Thus, since \(\varPsi \) vanishes in the origin, we deduce of the Inverse Function Theorem that \(\varPsi \) defines a local homeomorphism around the origin. Furthermore, we recallFootnote 7 that the exponential is an homeomorphism between a neighborhood of the origin in \(\mathfrak {sp}_{2n}(\mathbb {R})\) and a neighborhood of the identity in \(\mathrm {Sp}_{2n}(\mathbb {R})\). Consequently, we deduce that each matrix in \(\mathrm {Sp}_{2n}(\mathbb {R})\) close enough to the identity can be written as a product of the form of the product in the logarithm in (39). A fortiori, we have proved that G contains a neighborhood of the identity in \(\mathrm {Sp}_{2n}(\mathbb {R})\). Let V denote this neighborhood.
Since \(\mathrm {Sp}_{2n}(\mathbb {R})\) is connected (see, e.g., the subsection 4.4 of [4]), to prove that \(G= \mathrm {Sp}_{2n}(\mathbb {R})\), we just have to prove that G is closed and open in \(\mathrm {Sp}_{2n}(\mathbb {R})\). Indeed, if \(g\in G\), then gV is a neighborhood of g in \(\mathrm {Sp}_{2n}(\mathbb {R})\) and since V is included in G, then gV is also included in G. Thus, G is open in \(\mathrm {Sp}_{2n}(\mathbb {R})\). Conversely, if \(g\notin G\), then gV is also a neighborhood of g in \(\mathrm {Sp}_{2n}(\mathbb {R})\), but since G is a group we have \(gV \cap G = \emptyset \). Consequently, the complementary of G in \(\mathrm {Sp}_{2n}(\mathbb {R})\) is open, i.e., G is closed in \(\mathrm {Sp}_{2n}(\mathbb {R})\), which conclude the proof. \(\square \)
Lemma 6
If \(\ell :\mathbb {R}^{2n}\rightarrow \mathbb {R}\) is a real linear form, then \(\exp (i \ell ^w)\) can be computed by an exact splitting.
Proof
Let \(L\in \mathbb {R}^{2n}\) be the matrix of \(\ell \) and let \(c=c_1\) where
Applying Lemma 2, we have
Consequently, applying Proposition 4 at \(t=1\), we get
\(\square \)
Lemma 7
If \(p:\mathbb {R}^{2n}\rightarrow \mathbb {R}\) is a real-valued polynomial of degree 2 or less, then there exists a real linear form \(\ell :\mathbb {R}^{2n}\rightarrow \mathbb {R}\) and \(c\in \mathbb {R}\), such that
where q is the quadratic part of p.
Proof
Let l be the linear part of p. Considering the natural action of the entire functions on \(M_{2n}(\mathbb {C})\), let \(\ell =\ell _1\) and \(c=c_1\) where
with \(Q\in S_{2n}(\mathbb {R})\), the matrix of q and L the matrix of l. Applying Lemma 2, we have
Consequently, applying Proposition 4 at \(t=1\), we get (40). \(\square \)
Lemma 8
If \(p:\mathbb {R}^{2n}\rightarrow \mathbb {R}\) is a real-valued polynomial of degree 2 or less bounded below, then there exists a real-valued linear form \(\ell :\mathbb {R}^{2n}\rightarrow \mathbb {R}\) and \(c\in \mathbb {R}\) such that
where q is the quadratic part of p.
Proof
Applying Lemma 3, we get \(Y\in \mathbb {R}^{2n}\) such that \(p=p_1\) where
and \(X=(x_1,\dots ,x_n,\xi _1,\dots ,\xi _n)\).
Let and \(c=p(Y)\). Let \(\mathscr {B}\) be the basis introduce in the proof of Lemma 2 and defined by (37). Observing that
we deduce that
Consequently, we have
However, \(\varPhi ^{-i \mathbb {P}\ell }_t\) is a symplectic map, so we have
Now observing that the Hamiltonian \(\mathbb {P} c\) commutes (i.e., for the canonical Poisson bracket) with all the other Hamiltonians, we deduce of the Noether theorem (or of Lemma 2) that
Consequently, applying Proposition 4 at \(t=1\), we get (41). \(\square \)
Lemma 9
If \(q:\mathbb {R}^{2n}\rightarrow \mathbb {R}\) is a nonnegative real quadratic form, then \(\exp (-q^w)\) can be computed by an exact splitting.
Proof
Applying Theorem 21.5.3 of [21], we get a symplectic change of coordinates \(T\in \mathrm {Sp}_{2n}(\mathbb {R})\) such that there exists \(m\in \llbracket 1,n \rrbracket \) and \(0<\lambda _1\le \dots \le \lambda _m\) some real numbers such that
Consequently, since T is symplectic, applying Noether theorem, we have
Applying Theorem 3, we get, at the level of the Fourier Integral Operators,
Recalling that, as a consequence of the formula (2), the semigroups \(e^{\lambda _j (x_j^2-\partial _{x_j}^2) }\) can be computed by exact splittings, we deduce of Lemma 5 that the same applies for \(e^{-tq^w}\).
\(\square \)
Now, we can prove Theorem 1.
Proof of Theorem 1
Let p be a polynomial of degree 2 or less on \(\mathbb {C}^{2n}\) whose real part is bounded below on \(\mathbb {R}^{2n}\). We aim at proving that \(e^{-p^w}\) can be computed by an exact splitting in the sense of Definition 1.
Applying Corollary 1, we get a constant \(c\in \mathbb {R}\) such that \(\mathbb {P}(p-c)\ge 0\). Then, applying Theorem 2.1 and Lemma 3.10 of [2], we get \(t_0>0\) and two real quadratic forms \(a_t,b_t:\mathbb {R}^{2n+2} \rightarrow \mathbb {R}\) depending analytically on \(t\in (-t_0,t_0)\), \(a_t\) being nonnegative, and such that if \(|t| < t_0\), then
Furthermore, it follows of the Baker–Campbell–Hausdorff formulas and formulas of (3.22) and (3.41) of [2] defining \(a_t,b_t\) that these quadratic forms belong to the complex Lie algebra generated by \(\mathbb {P}(p-c)\) and \(\mathbb {P}(\overline{p}-c)\). Observing that for all polynomials \(p_1,p_2\) of degree 2 or less on \(\mathbb {C}^{2n}\) we have
the image of \(\mathbb {P}\) is a Lie algebra. Consequently, \(a_t,b_t\) belong to the image of \(\mathbb {P}\), i.e., there exist two real polynomials of degree 2 or less on \(\mathbb {R}^{2n}\) depending analytically on t and denoted \(p_{t}^{(r)}\) and \(p_{t}^{(i)}\) such that
Note that since \(p_{t}^{(r)}\) is the restriction of \(a_t \) on the affine subspace \(\{(x_{n+1},\xi _{n+1})=(1,0)\}\), it is also nonnegative.
Now applying Proposition 4, we deduce that if \(0\le t<t_0\) we have
Let \(t_\star \in (0,t_0)\) be such that there exists \(n\in \mathbb {N}\) satisfying \(t_\star =n^{-1}\). Since \((e^{-t p^w})_{t\ge 0}\) is a semigroup, we have
Consequently, if \(e^{-t_{\star } p^w}\) can be computed by an exact splitting, then the same applies for \(e^{-p^w}\). Furthermore, from the factorization (42), we deduce that if \(e^{-t_{\star } (p_{t_{\star }}^{(r)})^w}\) and \(e^{-it_{\star } (p_{t_{\star }}^{(i)})^w}\) can be computed by an exact splitting, then the same applies for \(e^{-t_{\star } p^w}\). Consequently, we just have to focus on these two semigroups.
On the one hand, applying Lemmas 7, 6, 5 and Theorem 3 (to justify that the semigroup generated by a quadratic differential operator is a Fourier Integral Operator), we deduce that \(\exp (-it_{\star } (p_{t_\star }^{(i)})^w)\) can be computed by an exact splitting.
On the other hand, applying Lemmas 8, 6 and 9, we deduce that \(e^{-t_{\star } (p_{t_{\star }}^{(r)})^w}\) can be computed by an exact splitting. \(\square \)
1.3 Proof of Lemma 1
We aim at proving that if \(\mathfrak {R}q\ge 0\) on \(\mathbb {R}^{2n}\), then \(\varPhi _t^q\equiv e^{-2itJ_{2n}Q} \in \mathrm {Sp}_{2n}^+(\mathbb {C})\), where Q is the matrix of q. First, we recall that from Proposition 3 we know that \(\varPhi _t^q\) is a symplectic transformation (i.e., \(\varPhi _t^q \in \mathrm {Sp}_{2n}(\mathbb {C})\)).
So we aim at proving that \(\varPhi _t^q\) is nonnegative, i.e.,
Since \(\varPhi _0^q=I_{2n}\), we have
Since \(\mathfrak {R}Q\) is a real symmetric nonnegative matrix, is a Hermitian nonnegative matrix, and thus, is also a Hermitian nonnegative matrix which proves (43).
Rights and permissions
About this article
Cite this article
Bernier, J. Exact Splitting Methods for Semigroups Generated by Inhomogeneous Quadratic Differential Operators. Found Comput Math 21, 1401–1439 (2021). https://doi.org/10.1007/s10208-020-09487-4
Received:
Revised:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s10208-020-09487-4
Keywords
- Splitting methods
- Quadratic differential operators
- Fokker–Planck equations
- Schrödinger equations
- Baker–Campbell–Hausdorff formula