Approximating Symplectic Realizations:
A General Framework for the Construction of Poisson Integrators

Alejandro Cabrera UFRJ, Rio de Janeiro (alejandro@matematica.ufrj.br) David Martín de Diego ICMAT (david.martin@icmat.es) Miguel Vaquero IE University (mvaquero@faculty.ie.edu)

Abstract

While the construction of symplectic integrators for Hamiltonian dynamics is well understood, an analogous general theory for Poisson integrators is still lacking. The main challenge lies in overcoming the singular and non-linear geometric behavior of Poisson structures, such as the presence of symplectic leaves with varying dimensions. In this paper, we propose a general approach for the construction of geometric integrators on any Poisson manifold based on independent geometric and dynamic sources of approximation. The novel geometric approximation is obtained by adapting structural results about symplectic realizations of general Poisson manifolds. We also provide an error analysis for the resulting methods and illustrative applications.

1 Introduction

Hamiltonian dynamical systems play a crucial role in many branches of science, including theoretical and applied physics, differential geometry, optimal control theory, economics, biology, robotics, and computer graphics. See [18, 23, 25] and the references therein. The versatility and robustness of Hamiltonian systems make them powerful tools across diverse disciplines, enhancing our ability to model, analyze, and predict the behavior of complex systems. In this context, it becomes evident the need for the development of geometric numerical integrator methods which are adapted to the underlying geometry. In the case of symplectic Hamiltonian system, such numerical methods have already been developed and applied successfully, see [17] for a general discussion.

On the other hand, it remains an open problem to extend these geometric methods to the case of general Poisson Hamiltonian systems ([8, 10, 13, 16, 17, 21, 24, 22]). In this paper, we present a general and practical approach for the construction of such Poisson integrators.

Challenges: The main obstacles in the development of Poisson integrators arise from the singular and non-linear properties of Poisson geometry, as discussed in [12]. Unlike symplectic geometry, where the symplectic two-form is non-degenerate, the Poisson tensor can abruptly change rank from one point to another. This phenomenon results in symplectic leaves, where solutions must lie, of varying dimensions and non-trivially glued together. Moreover, the general non-linear nature of the Poisson structure makes it impossible to rely on an explicit parametrization of the symplectic leaves, so that the method must handle them all at once. Consequently, designing numerical schemes for Poisson manifolds differs significantly from symplectic integration and requires more sophisticated techniques. Furthermore, even in the symplectic case, most methods rely on explicit Darboux coordinates, which are generally unavailable for non-canonical symplectic manifolds [3, 4, 14, 17]. Thus, the particular case of designing general methods for non-canonical symplectic forms is already a challenging task.

Setting of the problem: We consider a general Poisson manifold $(M,\pi)$ together with a smooth Hamiltonian function $H:M\rightarrow\mathbb{R}$ . Our main objective is to provide an approximation method for the flow of the corresponding Hamilton’s equation,

\dot{x}=\pi^{\sharp}(dH)|_{x},\ x(t)\in M.

(1)

We shall denote $X_{H}:=\pi^{\sharp}(dH)\in\mathfrak{X}(M)$ the corresponding Hamiltonian vector field. Thus, our goal is to develop a scheme that approximates the dynamics of $X_{H}$ while preserving the same geometric properties as the original flow:

•

The approximation should conserve the Hamiltonian to a certain degree, corresponding to $X_{H}(H)=\pi(dH,dH)=0$ .
•

The approximation scheme should preserve the underlying Poisson geometry, corresponding to $\mathcal{L}_{X_{H}}\pi=0$ (a known consequence of Jacobi identity for $\pi$ ). Additionally, we highlight the property $\mathcal{L}_{X_{H}}C=\pi(dH,dC)=0$ for any Casimir function $C$ .

The condition regarding the preservation of Casimirs is recalled since it allows to have a direct control on how $x(t)$ restricts to the symplectic leaves of $M$ in concrete examples.

The realization approach: One interesting general approach to addressing the approximation problem highlighted above is through the introduction of an auxiliary regularizing structure: a (strict¹¹1This is the nomenclature of [11], but since it will be the only type of realization considered, we do not specify it in the sequel.) symplectic realization of $(M,\pi)$ , in the sense of [11]. Such a structure comes with realization data $R=(S,\omega,\alpha,\beta,\sigma)$ (see Sec. 2.2 below) which can be used to produce Poisson diffeomorphisms $\varphi_{L}:M\to M$ for each choice of so-called Lagrangian bisection $L\hookrightarrow(S,\omega)$ . The realization approach thus consists on observing that the desired dynamic approximation can be translated into the matter of choosing a Lagrangian bisection $\hat{L}$ in $R$ such that $\varphi_{\hat{L}}$ approximates the flow of our ODE (1). Notice that, by construction, the approximation preserves the Poisson geometry of $(M,\pi)$ exactly.

Such symplectic realizations are intimately related to (local) symplectic groupoid structures and these concepts have already been successfully exploited in references [9, 10, 15] to develop a theoretical framework for Poisson integrators, as well as to study the dynamics of Poisson systems [26]. Moreover, this approach can also be seen as “universal” for Poisson integrators: since any Poisson diffeomorphism can be constructed using this scheme, virtually any Poisson integrator could be designed following a general procedure outlined in [15].

On the other hand, to transform this theoretical approach into concrete integration methods one is strongly limited by the need to know explicitly the symplectic realization data $R$ for the given $(M,\pi)$ ; it is well-known that such situations are rare in practice.

The new method: In this paper, we propose to combine the above realization approach with an initial geometric approximation step which produces approximate realization data $\hat{R}$ for $(M,\pi)$ . We implement this step by adapting known structural results about symplectic realizations. The resulting complete method, combining $\hat{R}$ with a dynamic approximation $\hat{L}$ , yields a general practical integration method for every $(M,\pi)$ . This methodology incorporates a decoupling between the geometric and dynamic approximations, which, to the best of the authors’ knowledge, is novel and yet to be fully explored. We also provide detailed analysis of the impact of the approximation on the numerical schemes and illustrate the method in concrete examples.

In more detail, the complete method consists of two stages:

•

Geometric Approximation, $\hat{R}$ : Theoretically, exact symplectic realization data $R$ always exists ([5, 6, 20, 19]). Nonetheless, we need an explicit description of it to incorporate into our numerical schemes. Here, we propose to approximate the symplectic realization through explicit power series expansions (building on [6]). We show that, by constructing an approximation $\hat{R}$ of the geometric structures involved in the description of the realization, we indeed obtain methods that preserve the underlying Poisson geometry approximately to a certain order.
•

Dynamic Approximation, $\hat{L}$ : Theoretically, the Lagrangian bisection $L$ in $R$ inducing the exact flow of (1) always exists (see a review in Section 2.2 below). In this step, we provide an approximation $\hat{L}$ which combines with the approximate geometric data $\hat{R}$ into a complete approximation method for the dynamics defined by $(M,\pi,H)$ while also preserving the Poisson geometry of $(M,\pi)$ to a certain order.

Results: We explore the properties of the resulting complete approximation methods which are a combination of both approximations $\hat{R}$ and $\hat{L}$ above, leading to the following main contributions:

Main Results 1: (Geometric approximation) We present a general scheme for obtaining order $n$ (for arbitrary $n$ ) approximations $\hat{R}$ of the mapping structures defining the symplectic realization $R$ that integrates locally any Poisson manifold. We show that any Lagrangian bisection $\hat{L}$ for $\hat{R}$ induces a diffeomorphism $M\to M$ which preserves the Poisson structure $\pi$ and Casimirs $C$ up to order $n$ . See Section 3 and Theorem 7.

Main Results 2: (Orders of the general combined method) We show that the combination of geometric and dynamic approximations, $\hat{R}$ and $\hat{L}$ , results in explicit numerical integrator method for (1) and describe its total order of approximation both of the dynamics and of the underlying Poisson geometry. See Section 4.1 and Theorem 8.

Main Results 3: (Implementing dynamic approximation) Building on our geometric approximation scheme $\hat{R}$ , we provide two concrete approaches for the dynamic approximation $\hat{L}$ : one based on the Hamilton-Jacobi theory and another based on collective integrators. The theoretical properties of both approaches are carefully discussed. See Sections 4.2 and 4.3.

As mentioned before, these results can also be applied to symplectic Hamiltonian systems, where $\pi=\omega^{-1}$ for a non-canonical symplectic form $\omega$ . We also note that, since we aim at local numerical methods which can be applied computationally, most of the relevant constructions are specialized to the case where $M$ is replaced by a coordinate chart endowed with an arbitrary Poisson structure $\pi$ . Nonetheless, we will describe most of our results and constructions in a global setting, with the local versions following from the obvious restrictions.

Outline of the paper: This paper is organized as follows. In Section 2, we introduce basic concepts and notation, and we review the ingredients behind the general realization approach to Poisson integrators. In Section 3, we present our construction of approximate realization data based on a truncation of Karasev’s symplectic realization and study its consequences in Theorem 7. In Section 4, we describe the novel class of complete methods for Poisson integrators, proving their orders of approximation in Theorem 8, and describing concrete versions of them in subsections 4.2 and 4.3. In Section 5, we provide computational illustrations by applying the methods in concrete problems. Finally, in Section 6, we discuss conclusions and directions of future work.

2 Notation and preliminaries

In this section, we introduce the definitions and notations of the main geometric objects to be used. We also describe the general realization approach to Poisson integrators.

2.1 Basic Poisson and symplectic geometry

We recall here the basic geometric structures used along the paper. See also [2, 23] for a complete description of these topics.

A Poisson structure on a differentiable manifold $M$ is given by a bilinear map

\begin{array}[]{rcc}C^{\infty}(M)\times C^{\infty}(M)&\longrightarrow&C^{% \infty}(M)\\ (f,g)&\longmapsto&\{f,g\}\end{array}

called the Poisson bracket, satisfying the following properties: for all $f,g,h\in C^{\infty}(M)$ ,

(i)

Skew-symmetry, $\{g,f\}=-\{f,g\}$ ;
(ii)

Leibniz rule, $\{fg,h\}=f\{g,h\}+g\{f,h\}$ ;
(iii)

Jacobi identity, $\{\{f,g\},h\}+\{\{h,f\},g\}+\{\{g,h\},f\}=0$ .

We call the pair $(M,\{\,,\,\})$ a Poisson manifold. A Poisson bracket is equivalent to a Poisson tensor field $\pi\in\mathfrak{X}^{2}(M)$ by

\pi(df,dg)=\{f,g\}.

We will usually refer to a Poisson manifold as $(M,\pi)$ .

Example 1 (Dual of a Lie Algebra, $\mathfrak{g}^{*}$ ).

If ${\mathfrak{g}}$ is a Lie algebra with Lie bracket $[\;,\;]$ , then it is defined a Poisson bracket on ${\mathfrak{g}}^{*}$ by

\{\xi,\eta\}(\alpha)=-\langle\alpha,[\xi,\eta]\rangle\;,

where $\xi$ and $\eta\in{\mathfrak{g}}$ are equivalently considered as linear forms on ${\mathfrak{g}}^{*}$ , and $\alpha\in{\mathfrak{g}}^{*}$ . This linear Poisson structure on ${\mathfrak{g}}^{*}$ is called the Kirillov-Kostant-Souriau Poisson structure.

A Poisson map between two Poisson manifolds $(M,\{\cdot,\cdot\}_{M})$ and $(N,\{\cdot,\cdot\}_{N})$ is a smooth map $\phi:M\to N$ that preserves the Poisson brackets. This means that for any pair of smooth functions $f,g$ on $N$ , the map $\phi$ satisfies:

\{\phi^{*}f,\phi^{*}g\}_{M}=\phi^{*}\{f,g\}_{N}.

Given a function $H\in C^{\infty}(M)$ , its associated Hamiltonian vector field $X_{H}\in\mathfrak{X}(M)$ is defined by

\{H,f\}=X_{H}(f),\ \forall f\in C^{\infty}(M).

The ODE on $M$ associated to the vector field $X_{H}$ is exactly our main equation (1).

We now turn to symplectic geometry. A symplectic structure on $M$ is a Poisson structure in which the bracket is non-degenerate,

\{f,g\}=0\,,\ \forall g\in C^{\infty}(M)\Rightarrow df=0.

This is equivalent to the tensor $\pi_{x}\in\Lambda^{2}T_{x}M$ being invertible at each $x\in M$ . In this case, the inverse (as a bilinear form)

\omega_{x}=\pi_{x}^{-1}\in\Lambda^{2}T^{*}_{x}M

defines a $2$ -form, $\omega\in\Omega^{2}(M)$ , which is closed, that is $d\omega=0$ . In this way, we arrive to the familiar description of a symplectic manifold as a pair $(M,\omega)$ where $M$ is a manifold and $\omega$ is a closed, non-degenerate, $2$ -form $\omega$ . (In this paper, the conventional signs when prescribing the inverse and Hamiltonian vector fields for $\omega$ are set so that $\{H,f\}_{\omega}=L_{X_{H}}f$ .)

Example 2 (Canonical Symplectic Form in the Cotangent Bundle).

The cotangent bundle of a manifold, say $T^{*}M$ , is endowed with a canonical symplectic form by taking²²2Notice the convention $\omega_{M}=dp_{i}\wedge dx^{i}$ intead of $dx^{i}\wedge dp_{i}$ , which we adopt to have a direct relation to the conventions of [5, 6] in the context of symplectic realization constructions. $\omega_{M}=d\theta$ , where $\theta$ is the Liouville one form. In canonical coordinates $(x^{i},p_{i})$ , these forms take the familiar expressions $\omega_{M}=dp_{i}\wedge dx^{i}$ and $\theta=p_{i}dx^{i}$ ; the corresponding Poisson brackets have $\{x^{i},p_{j}\}_{M}=\delta_{ij}$ and the Hamiltonian vector field reads $X_{H}=-\partial_{p_{i}}H\ \partial_{x^{i}}+\partial_{x^{i}}H\ \partial_{p_{i}}$ .

As well known, Lagrangian submanifolds play a crucial role in symplectic geometry and will also be key ingredients of this paper. A Lagrangian submanifold $i:L\hookrightarrow M$ of a symplectic manifold $(M,\omega)$ is an embedded submanifold which is maximal among embedded isotropic submanifolds, namely, among submanifolds such that $i^{*}\omega=0$ . We recall that $dim(M)$ is even and that $dim(L)=dim(M)/2$ for any Lagrangian submanifold.

Example 3 (Type I Generating Functions).

Given a manifold $M$ and its cotangent bundle $T^{*}M$ , any differentiable function $S:M\rightarrow\mathbb{R}$ produces a Lagrangian submanifold in $(T^{*}M,\omega_{M})$ by just taking

L:=graph(dS)=\{dS|_{x}:\ x\in M\}\subset T^{*}M.

In canonical coordinates, $(x^{i},p_{i})$ , this submanifold is determined by the equations $p_{i}=\partial_{x^{i}}S(x)$ with $x\in M$ .

2.2 Symplectic realization data, bisections and induced Poisson diffeomorphisms

The next definitions will become instrumental throughout the paper, see [11, 29] for more details.

Definition 1 (Symplectic Realization).

A symplectic realization $(S,\omega,\alpha,\sigma)$ of a Poisson manifold $(M,\{\cdot,\cdot\})$ consists of a symplectic manifold $(S,\omega)$ , a Poisson map $\alpha:(S,\{\cdot,\cdot\}_{\omega})\to(M,\{\cdot,\cdot\})$ , where $\{\cdot,\cdot\}_{\omega}$ are the symplectic brackets on $S$ , and a section $\sigma:M\to S$ of $\alpha$ , $\alpha\circ\sigma=id_{M}$ , such that $\sigma(M)\hookrightarrow(S,\omega)$ is a Lagrangian submanifold.

As mentioned in the Introduction, this is the type of realization called “strict” in [11] and is the only type that we shall consider in this paper.

Following that reference further, we now describe the dual map $\beta$ associated with a given symplectic realization $(S,\omega,\alpha,\sigma)$ . This map $\beta:U\subset S\to M$ is a submersion defined on a neighborhood $U$ of $\sigma(M)$ and is characterized by the properties:

•

$\sigma$ is also a section of $\beta$ , $\beta\circ\sigma=id_{M}$ ,

•

the $\beta$ -fibers and the $\alpha$ -fibers are symplectic orthogonal,

Ker(D_{y}\beta)^{\omega}=Ker(D_{y}\alpha),\ y\in U.

In [11], it is shown that, given a symplectic realization, such a dual map $\beta$ exists, that it is an anti-Poisson morphism (that is, a Poisson map onto $(M,-\pi)$ ), and that its germ around $\sigma(M)$ is uniquely determined by the realization.

By replacing $S$ with a suitable small enough neighborhood of $\sigma(M)$ , we can assume that both maps are defined on the entire $S$ . Our constructions will indeed be only based on local considerations around $\sigma(M)\subset S$ . Moreover, independently of how the maps are obtained, our constructions will only depend on having such maps $\alpha,\beta,\sigma$ with the properties highlighted above. We resume this discussion with the following definition.

Definition 2.

We say that $R:=(S,\omega,\alpha,\beta,\sigma)$ defines realization data for $(M,\pi)$ when:

•

$(S,\omega)$ is a symplectic manifold,
•

$\alpha,\beta:(S,\omega)\to(M,\pi)$ are Poisson and anti-Poisson maps, respectivelly, having $\omega$ -orthogonal fibers,
•

$\sigma$ is a section for both $\alpha$ and $\beta$ , and $\sigma(M)\hookrightarrow(S,\omega)$ is Lagrangian.

Remark 1 (Dual pairs).

The data given by the two maps $\alpha,\beta:(S,\omega)\to(M,\pi)$ , with $\alpha$ being Poisson and $\beta$ being anti-Poisson, with domain a symplectic manifold and with symplectically orthogonal fibers, is known as a dual pair. All the constructions in this paper can be formulated as originating from such a dual pair with the extra data of a map $\sigma:M\to S$ which is a section for both $\alpha$ and $\beta$ , and whose image is Lagrangian in $S$ . Such a section for both $\alpha$ and $\beta$ determines $\sigma(M)\subset S$ a type of submanifold called bisection which will play a key role below.

Remark 2 (Symplectic groupoids).

Another important result from [11] is that a symplectic realization $(S,\omega,\alpha,\sigma)$ as above determines a unique germ of local symplectic groupoid structure $(S,\omega)\rightrightarrows M$ around $\sigma(M)$ . In this local groupoid, $S$ is the set of arrows, $M$ is the set of objects, $\alpha$ is the source map, and $\sigma$ is the identities map. Moreover, the dual map $\beta$ is the target map, there is an inversion map $inv:U_{inv}\subset S\to S$ defined around $\sigma(M)$ , and there is a multiplication map

m:U_{(2)}\subset\{(z_{1},z_{2})\in S\times S:\alpha(z_{1})=\beta(z_{2})\}\to S

defined on a neighborhood of $\{(\sigma(x),\sigma(x)):x\in M\}$ . All the structure maps satisfy local versions of the algebraic groupoid axioms around the units and $\omega$ becomes multiplicative. Conversely, any local (or global) symplectic groupoid determines a symplectic realization with $(S,\omega)$ the space of arrows, $\alpha$ the source map, and $\sigma$ the identity map.

We thus see that (local or global) symplectic groupoids provide examples of the symplectic realization data needed for our method. On the other hand, we emphasize that the multiplication map does not enter the construction of this paper. (The incorporation of the multiplication map into the methods will be explored elsewhere, [7].)

We now describe how to use realization data for $(M,\pi)$ in order to produce Poisson diffeomorphisms. The idea is that this approach can be used to provide Poisson maps which approximate the flow of our ODE (1).

Definition 3 (Lagrangian Bisection and the induced mapping).

Given realization data $(S,\omega,\alpha,\beta,\sigma)$ for $(M,\pi)$ , a bisection is a submanifold $L\subset S$ such that both restrictions $\alpha|_{L},\beta|_{L}:L\to M$ are diffeormofisms. When, additionally, $L\hookrightarrow(S,\omega)$ is Lagrangian, we say that it determines a Lagrangian bisection. A bisection $L$ determines an induced diffeormorphism by the rule

\varphi_{L}=\alpha\circ(\beta_{|L})^{-1}:M\rightarrow M.

Note that, replacing $M$ with a suitable open, one obtains an analogous notion of local (Lagrangian) bisection and its induced locally defined diffeomorphism of $M$ . Also note that $\sigma(M)$ is, by definition, a Lagrangian bisection and that the induced map is the identity,

\varphi_{\sigma(M)}=id_{M}.

The following compilation of results is the heart of the realization approach for Poisson integrators.

Theorem 4.

([11]) Let $R=(S,\omega,\alpha,\beta,\sigma)$ be realization data for $(M,\pi)$ . Then,

1.

when $L\hookrightarrow(S,\omega)$ is a Lagrangian bisection for $R$ , the induced map $\varphi_{L}$ defines a Poisson diffeomorphism $(M,\pi)\to(M,\pi)$ .

if $\phi^{H}_{t}:M\to M$ is the Hamiltonian flow on $(M,\pi)$ defined by $H\in C^{\infty}(M)$ , then

\phi^{H}_{t}=\alpha\circ\phi^{\alpha^{*}H}_{t}\circ\sigma,

with $\phi^{\alpha^{*}H}_{t}:S\to S$ the Hamiltonian flow of $\alpha^{*}H$ in $(S,\omega)$ .

In the setting of the previous item,

\phi_{t}^{H}=\varphi_{L_{t}}\text{ for the Lagrangian bisection }L_{t}=\phi^{% \alpha^{*}H}_{t}(\sigma(M))\,.

We observe that, since $\alpha$ and $\beta$ have symplectically orthogonal fibers, it follows that

\beta

is conserved along the flow

\phi^{\alpha^{*}H}_{t}

(S,\omega)

This fact explains the last item. We also observe that, while the former are Poisson objects (singular), the latter are symplectic constructions (regular); this is at the heart of the use of symplectic realizations in numerical methods for $(M,\pi,H)$ .

2.3 The general realization approach to Poisson integrators

We can now describe in detail the general realization approach to approximately integrate the flow of (1) on $(M,\pi,H)$ , see [9, 10, 15]. The key assumption is that we have at hand realization data $R=(S,\omega,\alpha,\beta,\sigma)$ for $(M,\pi)$ .

From Thm. 4, we know that the exact flow $\phi^{H}_{t}:M\to M$ of eq. (1), for $t$ small enough, can be presented as an induced map $\varphi_{L_{t}}$ for the Lagrangian bisection $L_{t}=\phi^{\alpha^{*}H}_{t}(\sigma(M))$ in $R$ . The key idea is then to replace $L_{t}$ by an approximation $\hat{L}_{t}$ , which is still a Lagrangian bisection for $R$ , and then define the approximate flow $\hat{\phi}^{H}_{t}:M\to M$ of our ODE (1) as

\hat{\phi}^{H}_{t}:=\varphi_{\hat{L}_{t}}=\alpha\circ\beta|_{\hat{L}_{t}}^{-1}% :M\to M.

The upshot of the method is that, by construction, $\hat{\phi}^{H}_{t}$ is a family of Poisson diffeomorphisms in $(M,\pi)$ , thus preserving the ambient Poisson geometry. The degree of approximation of $\hat{\phi}^{H}_{t}$ to $\phi^{H}_{t}$ will depend on how well $\hat{L}_{t}$ approximates $L_{t}$ , see a discussion in [9].

Remark 3 (Rescaling $\pi$ ).

Let us denote $\Phi^{\pi,H}_{t}:M\to M$ the Hamiltonian flow associated to $(M,\pi,H)$ , putting in relevance in the notation the dependence on the Poisson tensor $\pi$ . From the structure of the underlying ODE (1), it follows directly that rescaling time is equivalent to rescaling $\pi$ or $H$ ,

\Phi_{\lambda t}^{\pi,H}=\Phi_{t}^{\lambda\pi,H}=\Phi_{t}^{\pi,\lambda H}.

Having the above Remark in mind, let us assume that $R_{\epsilon}=(S,\omega,\alpha_{\epsilon},\beta_{\epsilon},\sigma)$ is a family of realization data for the $\epsilon$ -family $(M,\epsilon\pi),\ \epsilon\geq 0$ . Fixing an additional “time ste” parameter $h>0$ , we can then conclude that

\phi_{t=\epsilon h}^{H}=\Phi^{\pi,H}_{\epsilon h}=\Phi^{\epsilon\pi,H}_{h}=% \alpha_{\epsilon}\circ\beta_{\epsilon}|_{L_{\epsilon,h}}^{-1}\text{ with }L_{% \epsilon,h}=\phi^{\alpha_{\epsilon}^{*}H}_{h}(\sigma(M)).

(2)

This is the precise approach we shall take in the sequel, in which we shall introduce both: geometric approximations of order $O(\epsilon^{n+1})$ for the data $\alpha_{\epsilon},\beta_{\epsilon}$ and dynamic approximations $\hat{L}_{\epsilon,h}$ for $L_{\epsilon,h}$ and each $\epsilon$ , as before. (See Section 4.1 where this discussion resumes.)

3 Geometric approximation: the realization data

In this section, we first review the construction of the Karasev symplectic realization ([5, 6]) and then present an approximation scheme for this symplectic realization. The main objective is to demonstrate that if the realization data is approximated to order $n$ , then the diffeomorphism $\varphi_{L}$ induced by a Lagrangian bisection $L$ is a Poisson diffeomorphisms up to order $n$ . This will ensure that the error of the induced Poisson mappings can be controlled to a desired accuracy.

As explained in the Introduction, with an eye on computer-implementable methods, we mostly restrict to the case in which $M\simeq\mathbb{R}^{n}$ endowed with an arbitrary Poisson structure $\pi$ .

3.1 Review of the Karasev realization

There are two local symplectic realizations that have been extensively studied in the literature: the Weinstein symplectic realization ([29]) and the Karasev approach ([19]). In both cases, the realization space is given by $S\subset T^{*}M$ a neighborhood of the zero section. In the Weinstein case, the symplectic structure $\omega$ is a deformation of the canonical symplectic structure $\omega_{M}$ while the realization map is the projection $S\subset T^{*}M\to M$ . In the Karasev case, the symplectic structure is canonical $\omega_{M}$ and the realization map $\alpha:S\to M$ is a deformation of the projection. The connection between the two can be seen in [6].

In this paper, we shall work with the Karasev realization which we now review. Let $M=\mathbb{R}^{n}$ with coordinates $x^{i}$ and consider a general Poisson tensor $\pi=\pi^{ij}(x)\partial x_{i}\wedge\partial x_{j}$ . We want to construct realization data $(S,\omega,\alpha_{\epsilon},\beta_{\epsilon},\sigma)$ for $(M,\epsilon\pi)$ , for any $\epsilon\geq 0$ .

To this end, we follow the presentation of [6] and introduce the following auxiliary objects. On $T^{*}M=\mathbb{R}^{n}\times\mathbb{R}^{n}$ , we consider canonical coordinates $(x^{i},p_{j})$ and the “flat Poisson spray” vector field

\overline{V}(x,p)=-\pi(x)^{ij}p_{i}\partial x^{j}\in T_{(x,p)}(T^{*}M).

(

\overline{V}

)

Notice that the $p$ -variables do not evolve. We denote the corresponding flow at time $t$ by $\varphi^{\overline{V}}_{t}$ . Next, we consider the corresponding “ $x$ -average” mapping $\phi_{\epsilon}:T^{*}M\to M$ is defined by

\phi^{i}_{\epsilon}(x,p)=\frac{1}{\epsilon}\int_{0}^{\epsilon}\left((\varphi^{% \overline{V}}_{s})^{*}x^{i}\right)\big{|}_{(x,p)}\,ds.

(

\phi_{\epsilon}

)

Then, following Karasev [20], we define $\alpha_{\epsilon}:U_{\epsilon}\subset T^{*}M\to M$ on a neighborhood of $p=0$ , via the implicit relation

\phi_{\epsilon}(\alpha_{\epsilon}(x,p),p)=x.

(3)

By the implicit function theorem, it follows that the above formula indeed defines a smooth map on a neighborhood $U_{\epsilon}$ of the zero section

0_{M}:=\{(x,p)\in T^{*}M:p=0\}\subset T^{*}M.

Note that $U_{\epsilon}$ grows as $\epsilon\to 0$ and that

\alpha_{\epsilon}(x,0)=\alpha_{0}(x,p)=x.

We thus obtain the following realization data.

Definition 5.

Let $(M\simeq\mathbb{R}^{n},\pi)$ be Poisson and $\alpha_{\epsilon}$ the map defined by (3). Then,

K_{\epsilon}:=(S,\omega,\alpha_{\epsilon},\beta_{\epsilon},\sigma)

defines the Karasev realization data for $(M,\pi)$ , where $S\subset T^{*}M$ is a suitable small enough neighborhood of $0_{M}\subset T^{*}M$ , $\omega=\omega_{M}$ is the canonical symplectic structure,

\beta_{\epsilon}(x,p)=\alpha_{\epsilon}(x,-p)

and $\sigma(x)=(x,0)$ .

Details about why these maps indeed define realization data for $(M,\pi)$ can be found in [5, §3.3].

Remark 4 (Associated local Symplectic Groupoid).

Following [5] further, the realization data $K_{\epsilon}$ can be enriched to a local symplectic groupoid structure $G_{\epsilon}\rightrightarrows M$ for each $\epsilon\geq 0$ . For latter use, in this local groupoid $G_{\epsilon}$ , the inversion map is given by $inv(x,p)=(x,-p)$ . As remarked before, the multiplication map will not be used in this paper.

Remark 5 (Rescaling properties of $\alpha_{\epsilon}$ ).

We recall from [5, Lemma 3.22] that the Karasev realization $\alpha_{\epsilon}$ has special rescaling properties:

\alpha_{\lambda\epsilon}(x,p)=\alpha_{\epsilon}(x,\lambda p).

It then follows that

\phi^{\alpha_{\epsilon}^{*}H}_{\lambda t}\circ\mu_{\lambda}=\mu_{\lambda}\circ% \phi^{\alpha_{\lambda\epsilon}^{*}H}_{t},

where $\mu_{\lambda}(x,p)=(x,\lambda p)$ and $\lambda\in\mathbb{R}$ .

3.2 Approximate realization data through truncation

The idea in this subsection is to produce order $n$ approximations for the maps $\alpha_{\epsilon}$ and $\beta_{\epsilon}$ in the Karasev realization data $K_{\epsilon}$ . Before doing that, we define more generally such approximations.

To this end, we first recall that, for smooth maps $f,\hat{f}:(D\subset\mathbb{R}^{n})\times I\to\mathbb{R}^{m}$ with $I\subset\mathbb{R}$ an interval containing $0$ and $D$ an open domain, we say that $\hat{f}$ is an approximation of order $n$ of $f$ , denoted $f=\hat{f}\ mod\ \mathcal{O}(\epsilon^{n+1})$ , when

\left\lVert f(x,\epsilon)-\hat{f}(x,\epsilon)\right\rVert=\mathcal{O}(\epsilon% ^{n+1}),

uniformly for $x$ varying in any compact in $D$ , and for $\left\lVert\cdot{}\right\rVert$ any norm in $\mathbb{R}^{m}$ . We also recall the notation $F(\epsilon)=\mathcal{O}(\epsilon^{k})$ if $|F(\epsilon)|\leq C\epsilon^{k}$ for some $C>0$ and for all $\epsilon$ in a small enough neighborhood of $\epsilon=0$ . Notice that the definition of $f=\hat{f}\ mod\ \mathcal{O}(\epsilon^{n+1})$ can be extended to $\epsilon$ -families defined on smooth manifolds $f,\hat{f}:M\times I\to N$ . Nevertheless, as mentioned in the Introduction, we will be mostly interested on the cases $M\simeq\mathbb{R}^{n},N\simeq\mathbb{R}^{m}$ , so we do not review the details of the general case.

Definition 6.

Let $R_{\epsilon}=(S,\omega,\alpha_{\epsilon},\beta_{\epsilon},\sigma)$ be realization data for $(M,\epsilon\pi),\ \epsilon\geq 0$ . We say that $\hat{R}_{\epsilon}=(S,\omega,\hat{\alpha}_{\epsilon},\hat{\beta}_{\epsilon},\sigma)$ defines approximate realization data of order $n$ when the maps $\hat{\alpha}_{\epsilon}$ and $\hat{\beta}_{\epsilon}$ are approximations of order $n$ of $\alpha_{\epsilon}$ and $\beta_{\epsilon}$ , respectively, and with the additional condition that $\sigma$ is a bisection for all $\epsilon$ : $\hat{\alpha}_{\epsilon}\circ\sigma=\hat{\beta}_{\epsilon}\circ\sigma=id_{M}$ .

Let us note that, for any submanifold $L\subset S$ which is a (possibly local) bisection for both $\hat{\alpha}_{\epsilon}$ and $\hat{\beta}_{\epsilon}$ in the obvious sense, the corresponding induced map

\hat{\varphi}_{L}=\hat{\alpha}_{\epsilon}\circ\hat{\beta}_{\epsilon}|_{L}^{-1}% :M\to M

is a well defined (possibly local) diffeomorphism.

Next, we observe that, clearly, order $n$ approximations can be obtained for a smooth $\epsilon$ -family of functions $f_{\epsilon}(x)=f(x,\epsilon)$ by truncating its Taylor expansion around $\epsilon=0$ ,

\tau^{n}[f_{\epsilon}](x):=\sum_{k=0}^{n}\frac{\epsilon^{k}}{k!}\partial^{k}_{% \epsilon}f|_{(x,0)}.

We thus come back to the Karasev realization data $K_{\epsilon}=(S\subset T^{*}M,\omega_{M},\alpha_{\epsilon},\beta_{\epsilon},\sigma)$ of $(M=\mathbb{R}^{n},\epsilon\pi)$ and consider the corresponding approximate realization data of order $n$ ,

\hat{K}_{\epsilon}=(S\subset T^{*}M,\omega_{M},\hat{\alpha}_{\epsilon}:=\tau^{% n}[\alpha_{\epsilon}],\hat{\beta}_{\epsilon}:=\tau^{n}[\beta_{\epsilon}],% \sigma).

We call this the approximate Karasev realization data of order $n$ . Note that the condition that $\sigma(x)=(x,0)$ is a bisection for $\hat{\alpha}_{\epsilon},\hat{\beta}_{\epsilon}$ follows directly from the rescaling property in Remark 5.

Finally, we observe that the Taylor expansion of $\alpha_{\epsilon}$ around $\epsilon=0$ (and, hence, of $\beta_{\epsilon}$ ) in the Karasev realization was explicitly computed in [6], by means of the defining relation (3). In that reference, the coefficients in the expansion

\tau^{n}[\alpha^{i}_{\epsilon}](x,p)=\sum_{k=0}^{n}\alpha_{(k)}^{i}(x,p)\ % \epsilon^{k}

where explicitly given in alternative ways: by an explicit recursion [6, eq. (3.5)]; by an explicit formula in terms of rooted trees and elementary differentials of the vector field $\overline{V}$ in [6, eq. (3.13)] with weights defined by [6, eq. (3.12)] or by the iterated integrals [6, eq. (3.15)].

We shall not be needing the precise formulas here, but is important to have in mind that these explicit formulas for $\tau^{n}[\alpha_{\epsilon}](x,p)$ and $\tau^{n}[\beta_{\epsilon}](x,p)=\tau^{n}[\alpha_{\epsilon}](x,-p)$ , which can be found in the references above, are such that our overall method can be indeed implemented practically in a computer. For completeness, we write the first terms of the general expansion [6]:

\alpha_{\epsilon}^{i}(x,p)=x^{i}+\frac{\epsilon}{2}\pi^{vi}p_{v}+\frac{% \epsilon^{2}}{12}\partial_{u}\pi^{vi}\pi^{wu}p_{v}p_{w}+\frac{\epsilon^{3}}{48% }\partial_{u}\partial_{w}\pi^{vi}\pi^{ku}\pi^{lw}p_{v}p_{k}p_{w}+\mathcal{O}(% \epsilon^{4}).

Remark 6 (Rescaling properties of $\hat{K}_{\epsilon}$ ).

The truncated $\hat{\alpha}_{\epsilon}$ has the same rescaling property as $\alpha_{\epsilon}$ recalled in Remark 5. It analogously follows that

\phi^{\hat{\alpha}_{\epsilon}^{*}H}_{\lambda t}\circ\mu_{\lambda}=\mu_{\lambda% }\circ\phi^{\hat{\alpha}_{\lambda\epsilon}^{*}H}_{t},

for any Hamiltonian $H\in C^{\infty}(M)$ .

Remark 7 (Relation to Kontsevich’s quantization).

In [6], it was further shown that the Taylor expansion of Karasev’s realization $\alpha_{\epsilon}$ is directly related to (the tree level part of) Kontsevich’s quantization formula for $(M=\mathbb{R}^{n},\pi)$ . This thus establishes a non-trivial connection between that formula and the present methods for Poisson integrators.

Remark 8 (Other approximations of $K_{\epsilon}$ ).

We observe that other approximation approaches can be taken. In particular, one can implement a computational method to approximate the Karasev map $\alpha_{\epsilon}$ by solving for the defining relation (3) in an approximate way (e.g. replacing the integral by an approximating sum and the flow $\varphi^{\overline{V}}_{s}$ by a numerical approximation). We shall explore these practical possibilities elsewhere.

3.3 Order of geometry preservation under approximate realizations

We now explore the degree of approximate preservation of $\pi$ for a map $\hat{\varphi}_{\epsilon,L}$ induced by a Lagrangian bisection $L\subset S$ with respect to approximations $\hat{\alpha}_{\epsilon}$ and $\hat{\beta}_{\epsilon}$ of the realization data.

First, observe that when $L\hookrightarrow(S,\omega)$ is a Lagrangian submanifold which is close enough to the Lagrangian $\sigma(M)\subset S$ , then $L$ is a bisection relative to any approximate realization data $\hat{R}_{\epsilon}$ . This follows by transversality since, by definition, $\sigma(M)$ defines a bisection for the underlying maps $\hat{\alpha}_{\epsilon},\hat{\beta}_{\epsilon}$ for all $\epsilon$ . Similarly, $L$ is also a bisection for the exact $R_{\epsilon}$ when is close enough to $\sigma(M)$ .

Theorem 7 (First main result).

Let $(M,\pi)$ be Poisson and $\hat{R}_{\epsilon}=(S,\omega,\hat{\alpha}_{\epsilon},\hat{\beta}_{\epsilon},\sigma)$ be approximate realization data of order $n$ for $(M,\epsilon\pi)$ . Consider $L\hookrightarrow(S,\omega)$ a Lagrangian which is close enough to $\sigma(M)$ . Then, $L$ is a bisection for $\hat{\alpha}_{\epsilon}$ and $\hat{\beta}_{\epsilon}$ , and the induced map

\hat{\varphi}_{\epsilon,L}=\hat{\alpha}_{\epsilon}\circ\hat{\beta}_{\epsilon}|% _{L}^{-1}:M\to M

preserves the Poisson tensor $\pi$ and any Casimir $C\in C^{\infty}(M)$ up to order $n$ :

\hat{\varphi}_{\epsilon,L}^{*}\pi-\pi=\mathcal{O}(\epsilon^{n+1}),\ \ \hat{% \varphi}_{\epsilon,L}^{*}C-C=\mathcal{O}(\epsilon^{n+1}),

(4)

uniformly on any compact in $M$ .

Proof.

It is clear that we can reduce the proof to the case $M=\mathbb{R}^{n}$ and, by the Lagrangian tubular neighborhood theorem applied to $\sigma(M)\hookrightarrow(S,\omega)$ , to $S\simeq\mathbb{R}^{2n}$ .

We thus begin the proof with the following general facts:

1.

if $f_{\epsilon}:\mathbb{R}^{n}\to\mathbb{R}^{n}$ is a smooth family of diffeomorphisms and $\hat{f}_{\epsilon}$ is a family of diffeomorphism which is an order $n$ approximation of $f_{\epsilon}$ , then $\hat{f}_{\epsilon}^{-1}$ is an order $n$ approximation of $f_{\epsilon}^{-1}$ ;
2.

if $f_{\epsilon},g_{\epsilon}$ are composable $\epsilon$ -families of maps and $\hat{f}_{\epsilon},\hat{g}_{\epsilon}$ are corresponding order $n$ approximations, then $\hat{f}_{\epsilon}\circ\hat{g}_{\epsilon}$ is an order $n$ approximation of $f_{\epsilon}\circ g_{\epsilon}$ .

The first fact (1) follows by observing that the $n$ -first Taylor coefficients of the inverse $f^{-1}_{\epsilon}$ of any family $f_{\epsilon}$ are completely determined the $n$ -first coefficients of $f_{\epsilon}$ by a recursion following from expanding $f_{\epsilon}\circ f_{\epsilon}^{-1}=id$ . Then, since $f_{\epsilon}$ and $\hat{f}_{\epsilon}$ share the same $n$ -first coefficients by hypothesis, the result follows. Similarly, the second fact (2) follows from the fact that the $n$ -first Taylor coefficients of the composition $f_{\epsilon}\circ g_{\epsilon}$ are completely determined by the $n$ -first coefficients of $f_{\epsilon}$ and $g_{\epsilon}$ .

Coming back to the proof of the theorem, let $R_{\epsilon}=(S,\omega,\alpha_{\epsilon},\beta_{\epsilon},\sigma)$ be the realization data of which $\hat{R}_{\epsilon}$ is an order $n$ approximation. Since $L$ is close to $L_{0}:=\sigma(M)$ and $L_{0}$ is a bisection for $\alpha_{\epsilon}$ and $\beta_{\epsilon}$ , then $L$ is also a bisection for $\alpha_{\epsilon},\beta_{\epsilon}$ . From the general results (see Thm. 4), since $L\hookrightarrow(S,\omega)$ is Lagrangian, we thus know that

(a):\text{$\varphi_{\epsilon,L}=\alpha_{\epsilon}\circ\beta_{\epsilon}|_{L}^{-% 1}$ is a Poisson diffeomorphism on $(M,\pi)$. }

Note that, a priori, it is a Poisson map for $(M,\epsilon\pi)$ but this implies that it must preserve $\pi$ by linearity of $\varphi_{\epsilon,L}^{*}$ and for $\epsilon\neq 0$ .

Next, we use (1.) above to conclude that $\hat{\beta}_{\epsilon}|_{L}^{-1}$ is an order $n$ approximation of $\beta_{\epsilon}|_{L}^{-1}$

Approximating Symplectic Realizations: A General Framework for the Construction of Poisson Integrators