Introduction
Geometric numerical integration is concerned with the construction of numerical methods that reproduce, at the discrete level, geometric properties of the differential equations being approximated. In Hamiltonian mechanics, the central example is provided by symplectic integrators, which preserve the symplectic structure of phase space and typically exhibit favorable long-time behavior that is not shared by generic numerical schemes. We refer to [7 , 22 , 35 , 37 ] for more details about structure-preserving integration techniques. Variational integrators provide a closely related construction based on a discretization of Hamilton’s principle [13 , 18 , 31 ] , while Hamiltonian variational integrators arise from discrete Hamiltonian variational principles and generating functions [14 , 24 , 27 ] .
Locally conformally symplectic geometry provides a natural extension of the symplectic setting. An LCS manifold is a triple ( M , Ω , φ ) (M, \Omega, \varphi) ( M , Ω , φ ) consisting of a smooth manifold M M M , a nondegenerate two-form Ω \Omega Ω , and a closed one-form φ \varphi φ , called the Lee form, satisfying
d Ω = φ ∧ Ω . (1) \mathrm{d}\Omega= \varphi\wedge\Omega. \tag*{(1)} d Ω = φ ∧ Ω. (1)
Since φ \varphi φ is closed, it is locally exact, φ ∣ U α = d σ α \left.\varphi\right|_{U_\alpha} = \mathrm{d}\sigma_\alpha φ ∣ U α = d σ α on a suitable open covering { U α } α ∈ A \{U_\alpha\}_{\alpha\in A} { U α } α ∈ A of M M M , and Ω α = e − σ α Ω ∣ U α \Omega_\alpha= e^{-\sigma_\alpha}\left.\Omega\right|_{U_\alpha} Ω α = e − σ α Ω ∣ U α is closed and nondegenerate, i.e., an ordinary symplectic form on U α U_\alpha U α . For more details on LCS manifolds, see [4 , 5 , 25 , 43 ] . Hamilton–Jacobi theory in this setting has been developed in [15 ] .
This local symplectic description suggests an apparently simple strategy for numerical integration: construct a symplectic integrator in each local symplectic representation and then reconstruct a numerical method for the original LCS system. The essential difficulty, however, is global. A collection of locally defined symplectic numerical maps does not, by itself, determine a global discrete dynamics. The local maps must agree on overlaps after the corresponding changes of representation are taken into account. Indeed, on a connected component of an overlap U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β the potentials differ by a constant, σ α − σ β = C α β \sigma_\alpha- \sigma_\beta= C_{\alpha\beta} σ α − σ β = C α β , so that Ω β = c α β Ω α \Omega_\beta= c_{\alpha\beta}\Omega_\alpha Ω β = c α β Ω α with c α β = e C α β c_{\alpha\beta} = e^{C_{\alpha\beta}} c α β = e C α β , and the numerical constructions performed in the two local symplectic representations must be compatible with this overlap structure.
The purpose of the present work is to formulate this compatibility problem at the level of discrete Hamiltonian mechanics and to provide conditions under which local structure-preserving discretizations globalize to a well-defined LCS numerical dynamics. Our starting point is Hamiltonian rather than Lagrangian. This choice is important because the Hamiltonian formulation does not require the hyperregularity assumptions needed to pass between Hamiltonian and Lagrangian descriptions. We first recall the continuous LCS Hamiltonian theory and formulate its local canonical dynamics through a Type II Hamiltonian variational principle [27 , 39 ] ; this continuous material fixes the geometric and variational notation used in the discrete construction.
The discrete theory begins with the exact discrete right Hamiltonian, the Type II generating function of the exact time-h h h flow of a local Hamiltonian representative H α H_\alpha H α . Replacing it by a computable approximation produces a local discrete Hamiltonian map which, as in ordinary discrete Hamiltonian mechanics, is symplectic in the local canonical variables. Local symplecticity, however, does not solve the globalization problem: the question is whether the discrete constructions obtained in two overlapping local representations describe the same numerical trajectory. In the cotangent-type local representation considered in this paper, we first consider overlap transformations of the form
T α β ( q , r ) = ( q , c α β r ) . (2) T_{\alpha\beta}(q,r) = (q,c_{\alpha\beta}r). \tag*{(2)} T α β ( q , r ) = ( q , c α β r ) . (2)
We show that the exact discrete right Hamiltonians satisfy a definite transformation law under this map. This motivates the corresponding compatibility condition for approximate discrete Hamiltonians,
H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) , (3) H_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}H_{\alpha,h}^{d+}\left(q,c_{\alpha\beta}^{-1}s\right), \tag*{(3)} H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) , (3)
which is a condition on the discrete generating functions themselves, rather than an a priori assumption on the resulting numerical maps. Under (3) the local discrete maps satisfy the equivariance relation F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h , which is precisely the commutativity condition required for the local maps to glue, and after reconstruction in the original LCS variables the resulting global map F h F_h F h satisfies the discrete conformal preservation law
F h ∗ Ω = e σ α ∘ F h − σ α Ω , (4) F_h^*\Omega=e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega, \tag*{(4)} F h ∗ Ω = e σ α ∘ F h − σ α Ω , (4)
which reproduces the characteristic conformal preservation property of the continuous LCS Hamiltonian flow. Symplecticity is thus a local geometric property of the numerical method, whereas equivariance is the compatibility property that allows those local methods to represent a single global dynamics; the globalization statement itself is straightforward once equivariance is known, and the substantive issue is to identify verifiable conditions on the discretization that guarantee it. The Type II framework provides such conditions naturally [27 ] : Galerkin approximations of the Hamiltonian action can be constructed so as to inherit the required scaling law. Moreover, because the transition maps (2) are linear, standard Runge–Kutta discretizations are equivariant when the same tableau is used in every local representation, and if the method is additionally symplectic the resulting global method is an LCS structure-preserving integrator. This includes the Gauss–Legendre family and, as its simplest member, the implicit midpoint rule [22 , 31 ] . The restriction to the cotangent-type setting is relevant here: for general nonlinear coordinate transformations, Runge–Kutta methods are not automatically equivariant, and we therefore formulate the principal results under explicit hypotheses on the local symplectic representations and their transition maps rather than claiming coordinate invariance for arbitrary LCS Darboux charts. Once the Hamiltonian theory has been established, a Lagrangian counterpart can be obtained under appropriate regularity assumptions, see [1 , 30 , 31 ] ; in this sense the Lagrangian formulation is complementary to, rather than logically prior to, the Hamiltonian globalization result.
The framework also extends to affine cotangent transitions
T α β ( q , r ) = ( B q + b , c α β B − T r ) , T_{\alpha\beta}(q,r)=\left(Bq+b,c_{\alpha\beta}B^{-\mathsf T}r\right), T α β ( q , r ) = ( B q + b , c α β B − T r ) ,
where B B B is constant and invertible, b b b is constant, and c α β > 0 c_{\alpha\beta}>0 c α β > 0 is constant on each connected overlap component. We derive the associated Type II transformation law, including the endpoint term arising from the translation, and verify the equivariance of Runge–Kutta methods under these transitions.
We also clarify the distinction between the LCS Hamiltonian framework and the conformal Hamiltonian systems studied in [32 ] , for which conformally symplectic integrators have been developed, for example, in [6 , 19 ] . For Hamiltonian systems with constant damping rate γ \gamma γ , the exact flow satisfies ψ t ∗ ω = e − γ t ω \psi_t^*\omega=e^{-\gamma t}\omega ψ t ∗ ω = e − γ t ω , where ω \omega ω is a closed symplectic form, and conformally symplectic integrators reproduce the corresponding discrete scaling law. In contrast, the flow of an LCS Hamiltonian system satisfies
ψ t ∗ Ω = ρ t Ω , ρ t ( z ) = exp ( ∫ 0 t φ ( ξ H ) ( ψ s ( z ) ) d s ) , \psi_t^*\Omega=\rho_t\Omega,\qquad\rho_t(z)=\exp\left(\int_0^t\varphi(\xi_H)(\psi_s(z))\,\mathrm{d}s\right), ψ t ∗ Ω = ρ t Ω , ρ t ( z ) = exp ( ∫ 0 t φ ( ξ H ) ( ψ s ( z )) d s ) ,
so that the conformal factor generally depends on the initial point, rather than being determined solely by a constant damping rate. Even when the Lee form is exact, φ = d σ \varphi=\mathrm{d}\sigma φ = d σ , one obtains ρ t = e σ ∘ ψ t − σ \rho_t=e^{\sigma\circ\psi_t-\sigma} ρ t = e σ ∘ ψ t − σ , which does not in general equal e − γ t e^{-\gamma t} e − γ t ; in this exact case, the flow preserves the globally defined symplectic form e − σ Ω e^{-\sigma}\Omega e − σ Ω . Our discrete construction preserves the corresponding geometric relation F h ∗ Ω = e σ ∘ F h − σ Ω F_h^*\Omega=e^{\sigma\circ F_h-\sigma}\Omega F h ∗ Ω = e σ ∘ F h − σ Ω for steps represented within a single chart. When the Lee form is non-exact, the conformal potentials are available only locally, and compatibility on chart overlaps is required. Our contribution is to construct compatible local symplectic integrators that define a global LCS structure-preserving method.
The results presented here build on the discrete locally conformal framework developed in [16 ] . While that work derives the discrete LCS Hamiltonian equations, we recover these equations by applying a Type II Hamiltonian variational principle in each local chart and expressing the resulting discrete Hamilton equations in LCS coordinates. Our contributions include this direct Hamiltonian variational derivation and the construction of globally defined integrators through compatibility conditions for local symplectic integrators on chart overlaps. We formulate and verify these conditions and provide numerical examples illustrating convergence, preservation of the LCS conformal scaling law, and consistency under chart switching.
The paper is organized as follows. § 2 recalls the continuous LCS Hamiltonian geometry and develops the Type II variational formulation required in the sequel. § 3 introduces exact and approximate discrete right Hamiltonians and the associated local symplectic maps. § 4 derives the overlap transformation law of the exact discrete right Hamiltonian, introduces the compatibility condition for approximate discrete right Hamiltonians, shows that it implies equivariance of the local discrete maps, and establishes the globalization and discrete LCS preservation results. § 5 identifies concrete classes of compatible methods, including Type II variational constructions and symplectic Runge–Kutta methods, and extends the relevant results to affine base transitions. § 7 recovers the corresponding discrete Lagrangian formulation under regularity assumptions. § 8 investigates the resulting methods numerically through convergence, overlap-consistency, chart-switching, and structure-preservation tests, and § 9 summarizes the results and discusses an open problem.
Continuous LCS Hamiltonian mechanics and variational formulation
In this section, we recall the continuous geometric framework underlying the discrete constructions developed later in the paper. Locally conformally symplectic geometry and the corresponding Hamiltonian and Lagrangian dynamics have been extensively studied in the literature [4 , 5 , 15 , 25 , 28 , 43 ] ; the purpose of this section is not to develop a new continuous theory, but to collect the ingredients that will be needed for the Hamiltonian discretization and globalization problem. The Hamiltonian viewpoint will be primary throughout the paper, and no hyperregularity assumption is required at this stage; the Lagrangian description is developed, under appropriate regularity assumptions, in § 7 .
Definition 2.1. A locally conformally symplectic manifold is a triple ( M , Ω , φ ) (M,\Omega,\varphi) ( M , Ω , φ ) , where M M M is a smooth even-dimensional manifold, Ω ∈ Ω 2 ( M ) \Omega\in\Omega^2(M) Ω ∈ Ω 2 ( M ) is a nondegenerate two-form, and φ ∈ Ω 1 ( M ) \varphi\in\Omega^1(M) φ ∈ Ω 1 ( M ) is a closed one-form such that
d Ω = φ ∧ Ω . (5) \mathrm{d}\Omega=\varphi\wedge\Omega. \tag*{(5)} d Ω = φ ∧ Ω. (5) The one-form φ \varphi φ is called the Lee form .4
For the closed one-form φ \varphi φ , the associated Lichnerowicz–de Rham differential (see [20 ] ) is defined by
d φ : = d − φ ∧ . (6) d_{\varphi} := d-\varphi\wedge. \tag*{(6)} d φ := d − φ ∧ . (6)
Since d φ = 0 d_{\varphi}=0 d φ = 0 , one has d φ 2 = 0 d_{\varphi}^{2}=0 d φ 2 = 0 , and the LCS condition (5) is equivalently written as d φ Ω = 0 d_{\varphi}\Omega=0 d φ Ω = 0 .
The terminology “locally conformally symplectic” follows from the local exactness of the Lee form. Let { U α } α ∈ A \{U_{\alpha}\}_{\alpha\in A} { U α } α ∈ A be an open covering of M M M such that φ ∣ U α = d σ α \left.\varphi\right|_{U_{\alpha}}=d\sigma_{\alpha} φ ∣ U α = d σ α for smooth functions σ α : U α → R \sigma_{\alpha}:U_{\alpha}\to\mathbb{R} σ α : U α → R , and define
Ω α : = e − σ α Ω ∣ U α , equivalently Ω ∣ U α = e σ α Ω α . (7) \Omega_{\alpha}:=e^{-\sigma_{\alpha}}\left.\Omega\right|_{U_{\alpha}},\qquad\text{equivalently}\qquad\left.\Omega\right|_{U_{\alpha}}=e^{\sigma_{\alpha}}\Omega_{\alpha}. \tag*{(7)} Ω α := e − σ α Ω ∣ U α , equivalently Ω ∣ U α = e σ α Ω α . (7)
Then d Ω α = e − σ α ( − d σ α ∧ Ω + d Ω ) = 0 d\Omega_{\alpha}=e^{-\sigma_{\alpha}}(-d\sigma_{\alpha}\wedge\Omega+d\Omega)=0 d Ω α = e − σ α ( − d σ α ∧ Ω + d Ω ) = 0 by (5) , and Ω α \Omega_{\alpha} Ω α is nondegenerate since Ω \Omega Ω is, so that ( U α , Ω α ) (U_{\alpha},\Omega_{\alpha}) ( U α , Ω α ) is a symplectic manifold. If the Lee form is globally exact, φ = d σ \varphi=d\sigma φ = d σ , then e − σ Ω e^{-\sigma}\Omega e − σ Ω is a globally defined symplectic form and the LCS structure is said to be globally conformally symplectic . The distinction between local and global exactness of the Lee form is one of the geometric reasons why compatibility between local descriptions is important.
Overlap relations and the conformal cocycle. On each connected component of an overlap U α ∩ U β U_{\alpha}\cap U_{\beta} U α ∩ U β , the local representatives of the globally defined Lee form agree, d σ α = d σ β d\sigma_{\alpha}=d\sigma_{\beta} d σ α = d σ β , so that
σ α − σ β = C α β , c α β : = e C α β > 0 , Ω β = c α β Ω α , (8) \sigma_{\alpha}-\sigma_{\beta}=C_{\alpha\beta},\qquad c_{\alpha\beta}:=e^{C_{\alpha\beta}}>0,\qquad\Omega_{\beta}=c_{\alpha\beta}\Omega_{\alpha}, \tag*{(8)} σ α − σ β = C α β , c α β := e C α β > 0 , Ω β = c α β Ω α , (8)
where C α β ∈ R C_{\alpha\beta}\in\mathbb{R} C α β ∈ R is constant and the last identity follows from e σ α Ω α = e σ β Ω β e^{\sigma_{\alpha}}\Omega_{\alpha}=e^{\sigma_{\beta}}\Omega_{\beta} e σ α Ω α = e σ β Ω β on the overlap. On a triple overlap the constants satisfy C α β + C β γ = C α γ C_{\alpha\beta}+C_{\beta\gamma}=C_{\alpha\gamma} C α β + C β γ = C α γ , which implies
c α β c β γ = c α γ . (9) c_{\alpha\beta}c_{\beta\gamma}=c_{\alpha\gamma}. \tag*{(9)} c α β c β γ = c α γ . (9)
Thus, the constants c α β c_{\alpha\beta} c α β define a multiplicative conformal cocycle relating the local symplectic representatives.
LCS Hamiltonian dynamics. Consider a smooth Hamiltonian function H : M → R H:M\to\mathbb{R} H : M → R . The LCS Hamiltonian vector field ξ H \xi_H ξ H is defined intrinsically by
ι ξ H Ω = d φ H = d H − H φ , (10) \iota_{\xi_H}\Omega=d_{\varphi}H=dH-H\varphi, \tag*{(10)} ι ξ H Ω = d φ H = d H − H φ , (10)
which determines ξ H \xi_H ξ H uniquely since Ω \Omega Ω is nondegenerate. On the local chart U α U_{\alpha} U α , define the local Hamiltonian function
H α : = e − σ α H ∣ U α . (11) H_{\alpha}:=e^{-\sigma_{\alpha}}\left.H\right|_{U_{\alpha}}. \tag*{(11)} H α := e − σ α H ∣ U α . (11)
Substituting Ω = e σ α Ω α \Omega=e^{\sigma_{\alpha}}\Omega_{\alpha} Ω = e σ α Ω α , φ = d σ α \varphi=d\sigma_{\alpha} φ = d σ α and H = e σ α H α H=e^{\sigma_{\alpha}}H_{\alpha} H = e σ α H α into (10) , the terms containing d σ α d\sigma_{\alpha} d σ α cancel and one obtains ι ξ H Ω α = d H α \iota_{\xi_H}\Omega_{\alpha}=dH_{\alpha} ι ξ H Ω α = d H α . Therefore, the restriction of ξ H \xi_H ξ H to U α U_{\alpha} U α is the ordinary Hamiltonian vector field X H α X_{H_{\alpha}} X H α of H α H_{\alpha} H α with respect to Ω α \Omega_{\alpha} Ω α : on each local symplectic chart, the LCS Hamiltonian dynamics is represented by an ordinary Hamiltonian system.
Thus, Ω \Omega Ω alone does not determine the Lee form; a closed one-form φ \varphi φ is specified as part of the LCS triple ( M , Ω , φ ) (M,\Omega,\varphi) ( M , Ω , φ ) . When dim M = 2 n ≥ 4 \dim M=2n\geq4 dim M = 2 n ≥ 4 , the map α ↦ α ∧ Ω \alpha\mapsto\alpha\wedge\Omega α ↦ α ∧ Ω on 1-forms is injective, so the Lee form, when it exists, is uniquely determined by Ω \Omega Ω .
Local canonical representation
For the Hamiltonian variational construction developed in this paper, we consider local cotangent-type symplectic representations [30 ] . Throughout the paper we take M = T ∗ Q M=T^*Q M = T ∗ Q with bundle projection π Q : T ∗ Q → Q \pi_Q:T^*Q\to Q π Q : T ∗ Q → Q , denote the LCS variables by ( q i , p i ) (q^i,p_i) ( q i , p i ) , write Θ c a n = p i d q i \Theta_{\mathrm{can}}=p_i\,dq^i Θ can = p i d q i for the Liouville one-form, assume that the Lee form is basic , i.e.,
φ = π Q ∗ φ Q (12) \varphi=\pi_Q^*\varphi_Q \tag*{(12)} φ = π Q ∗ φ Q (12)
for a closed one-form φ Q \varphi_Q φ Q on Q Q Q , so that the local potentials depend on the base point only, σ α = σ α ( q ) \sigma_\alpha=\sigma_\alpha(q) σ α = σ α ( q ) , and assume that the LCS two-form is of cotangent type [23 ] , namely
Ω = − d φ Θ c a n = d q i ∧ d p i + φ ∧ Θ c a n . (13) \Omega=-d_\varphi\Theta_{\mathrm{can}}=dq^i\wedge dp_i+\varphi\wedge\Theta_{\mathrm{can}}. \tag*{(13)} Ω = − d φ Θ can = d q i ∧ d p i + φ ∧ Θ can . (13)
In other words, ( T ∗ Q , − d φ Θ c a n ) (T^*Q,-d_\varphi\Theta_{\mathrm{can}}) ( T ∗ Q , − d φ Θ can ) is the cotangent bundle twisted by the basic closed one-form φ \varphi φ [23 , 43 ] . These are standing hypotheses, not consequences of (5) , and they are what makes the local representation below canonical. Note that d φ Ω = − d φ 2 Θ c a n = 0 d_\varphi\Omega=-d_\varphi^2\Theta_{\mathrm{can}}=0 d φ Ω = − d φ 2 Θ can = 0 , so (13) does define an LCS structure for every closed basic φ \varphi φ ; nondegeneracy holds because, φ \varphi φ being basic, the additional term φ ∧ Θ c a n = φ j p i d q j ∧ d q i \varphi\wedge\Theta_{\mathrm{can}}=\varphi_jp_i\,dq^j\wedge dq^i φ ∧ Θ can = φ j p i d q j ∧ d q i is a two-form in the base differentials only, so that the matrix of Ω \Omega Ω in the coordinates ( q i , p i ) (q^i,p_i) ( q i , p i ) is
( A I − I 0 ) \begin{pmatrix} A & I \\ -I & 0 \end{pmatrix} ( A − I I 0 )
with A A A antisymmetric, which has determinant one. The restriction to basic Lee forms is essential here: for a closed one-form with fiber components, such as φ = d p \varphi=dp φ = d p on T ∗ R T^*\mathbb{R} T ∗ R , the two-form (13) equals ( 1 − p ) d q ∧ d p (1-p)\,dq\wedge dp ( 1 − p ) d q ∧ d p and degenerates along { p = 1 } \{p=1\} { p = 1 } .
We further assume that the local representations are adapted to the bundle structure: the chart domains are of the form U α = π Q − 1 ( V α ) U_\alpha=\pi_Q^{-1}(V_\alpha) U α = π Q − 1 ( V α ) , where { V α } \{V_\alpha\} { V α } is an open covering of Q Q Q by domains of the local Lee potentials, φ Q ∣ V α = d σ α \varphi_Q|_{V_\alpha}=d\sigma_\alpha φ Q ∣ V α = d σ α . For the initial construction, we assume that on overlaps the base coordinates q i q^i q i coincide. Under this assumption, the transition maps between local canonical representations only rescale the fiber variables. Under a nonlinear change of base coordinates, the transition maps would be nonlinear cotangent lifts, to which the Runge–Kutta and Galerkin results below do not apply as stated; we extend the construction to affine base transitions, which arise for instance from angular coordinates, in Section 5.4.
In such a representation, the local symplectic form on U α U_\alpha U α is written in canonical coordinates ( q i , r i ) (q^i,r_i) ( q i , r i ) as Ω α = d q i ∧ d r i \Omega_\alpha=dq^i\wedge dr_i Ω α = d q i ∧ d r i , the local Hamilton equations associated with H α H_\alpha H α are q ˙ i = ∂ H α / ∂ r i \dot q^i=\partial H_\alpha/\partial r_i q ˙ i = ∂ H α / ∂ r i and r ˙ i = − ∂ H α / ∂ q i \dot r_i=-\partial H_\alpha/\partial q^i r ˙ i = − ∂ H α / ∂ q i , and the local canonical momenta are related to the global momenta by
r i = e − σ α ( q ) p i , Ψ α ( q , p ) : = ( q , e − σ α ( q ) p ) , (14) r_i=e^{-\sigma_\alpha(q)}p_i,\qquad\Psi_\alpha(q,p):=\left(q,e^{-\sigma_\alpha(q)}p\right), \tag*{(14)} r i = e − σ α ( q ) p i , Ψ α ( q , p ) := ( q , e − σ α ( q ) p ) , (14)
where Ψ α \Psi_\alpha Ψ α denotes this local change of representation. Under the hypothesis (13) the two descriptions are consistent: with φ = d σ α \varphi=d\sigma_\alpha φ = d σ α on U α U_\alpha U α ,
d q i ∧ d r i = e − σ α ( d q i ∧ d p i − p i d q i ∧ d σ α ) = e − σ α ( d q i ∧ d p i + φ ∧ Θ c a n ) = e − σ α Ω , (15) dq^i\wedge dr_i=e^{-\sigma_\alpha}\left(dq^i\wedge dp_i-p_i\,dq^i\wedge d\sigma_\alpha\right)=e^{-\sigma_\alpha}\left(dq^i\wedge dp_i+\varphi\wedge\Theta_{\mathrm{can}}\right)=e^{-\sigma_\alpha}\Omega, \tag*{(15)} d q i ∧ d r i = e − σ α ( d q i ∧ d p i − p i d q i ∧ d σ α ) = e − σ α ( d q i ∧ d p i + φ ∧ Θ can ) = e − σ α Ω , (15)
which is (7) expressed in the canonical coordinates ( q i , r i ) (q^i,r_i) ( q i , r i ) ; equivalently, Ψ α ∗ Ω α = e − σ α Ω \Psi_\alpha^*\Omega_\alpha=e^{-\sigma_\alpha}\Omega Ψ α ∗ Ω α = e − σ α Ω .
Consider now the local coordinates ( q i , r i ) (q^i,r_i) ( q i , r i ) on U α U_\alpha U α and ( q i , s i ) (q^i,s_i) ( q i , s i ) on U β U_\beta U β . On the overlap, r = e − σ α p r=e^{-\sigma_\alpha}p r = e − σ α p and s = e − σ β p s=e^{-\sigma_\beta}p s = e − σ β p , so that s = c α β r s=c_{\alpha\beta}r s = c α β r by (8) , and the transition between the two local canonical representations is
T α β : = Ψ β ∘ Ψ α − 1 , T α β ( q , r ) = ( q , c α β r ) , T α β ∗ Ω β = c α β Ω α . (16) T_{\alpha\beta}:=\Psi_\beta\circ\Psi_\alpha^{-1},\qquad T_{\alpha\beta}(q,r)=(q,c_{\alpha\beta}r),\qquad T_{\alpha\beta}^*\Omega_\beta=c_{\alpha\beta}\Omega_\alpha. \tag*{(16)} T α β := Ψ β ∘ Ψ α − 1 , T α β ( q , r ) = ( q , c α β r ) , T α β ∗ Ω β = c α β Ω α . (16)
Thus T α β T_{\alpha\beta} T α β is conformally symplectic rather than symplectic unless c α β = 1 c_{\alpha\beta}=1 c α β = 1 . The local Hamiltonians satisfy a corresponding transformation law on the overlap: since H H H is globally defined, e σ α H α = e σ β H β e^{\sigma^\alpha}H_\alpha=e^{\sigma^\beta}H_\beta e σ α H α = e σ β H β , and taking into account the change of momentum variable s = c α β r s=c_{\alpha\beta}r s = c α β r we obtain
H β ( q , s ) = c α β H α ( q , c α β − 1 s ) , (17) H_\beta(q,s)=c_{\alpha\beta}H_\alpha\left(q,c_{\alpha\beta}^{-1}s\right), \tag*{(17)} H β ( q , s ) = c α β H α ( q , c α β − 1 s ) , (17)
and, differentiating,
D 1 H β ( q , s ) = c α β D 1 H α ( q , c α β − 1 s ) , D 2 H β ( q , s ) = D 2 H α ( q , c α β − 1 s ) . (18) D_1H_\beta(q,s)=c_{\alpha\beta}D_1H_\alpha\left(q,c_{\alpha\beta}^{-1}s\right),\qquad D_2H_\beta(q,s)=D_2H_\alpha\left(q,c_{\alpha\beta}^{-1}s\right). \tag*{(18)} D 1 H β ( q , s ) = c α β D 1 H α ( q , c α β − 1 s ) , D 2 H β ( q , s ) = D 2 H α ( q , c α β − 1 s ) . (18)
Equivariance of the continuous Hamiltonian dynamics. The local Hamiltonian vector fields X H α X_{H_\alpha} X H α and X H β X_{H_\beta} X H β are related by the transition map. Indeed, by (16) and (17) ,
T α β ∗ ( ι T α β ∗ X H α Ω β ) = ι X H α T α β ∗ Ω β = c α β d H α = d ( H β ∘ T α β ) = T α β ∗ d H β , T_{\alpha\beta}^{*}\left(\iota_{T_{\alpha\beta*}X_{H_\alpha}}\Omega_\beta\right)=\iota_{X_{H_\alpha}}T_{\alpha\beta}^{*}\Omega_\beta=c_{\alpha\beta}\,\mathrm{d}H_\alpha=\mathrm{d}(H_\beta\circ T_{\alpha\beta})=T_{\alpha\beta}^{*}\mathrm{d}H_\beta, T α β ∗ ( ι T α β ∗ X H α Ω β ) = ι X H α T α β ∗ Ω β = c α β d H α = d ( H β ∘ T α β ) = T α β ∗ d H β ,
so that ι T α β ∗ X H α Ω β = d H β \iota_{T_{\alpha\beta*}X_{H_\alpha}}\Omega_\beta=\mathrm{d}H_\beta ι T α β ∗ X H α Ω β = d H β , i.e.,
X H β ∘ T α β = T T α β ∘ X H α ; (19) X_{H_\beta}\circ T_{\alpha\beta}=TT_{\alpha\beta}\circ X_{H_\alpha}; \tag*{(19)} X H β ∘ T α β = T T α β ∘ X H α ; (19)
in coordinates, both sides equal ( D 2 H α ( q , r ) , − c α β D 1 H α ( q , r ) ) (D_2H_\alpha(q,r),-c_{\alpha\beta}D_1H_\alpha(q,r)) ( D 2 H α ( q , r ) , − c α β D 1 H α ( q , r )) by (18) . If Φ α , t \Phi_{\alpha,t} Φ α , t and Φ β , t \Phi_{\beta,t} Φ β , t denote the local Hamiltonian flows of X H α X_{H_\alpha} X H α and X H β X_{H_\beta} X H β , uniqueness of solutions implies
Φ β , t ∘ T α β = T α β ∘ Φ α , t . (20) \Phi_{\beta,t}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ\Phi_{\alpha,t}. \tag*{(20)} Φ β , t ∘ T α β = T α β ∘ Φ α , t . (20)
Thus, the two local Hamiltonian descriptions are related by the transition map and represent the same dynamics on the overlap.
Type II Hamiltonian variational principle
We now recall the Type II Hamiltonian variational principle in the local canonical representation, following [27 ] . This formulation is particularly appropriate for the discrete construction because its discrete counterpart is expressed in terms of right discrete Hamiltonians. Fix a local symplectic representative ( U α , Ω α ) (U_\alpha,\Omega_\alpha) ( U α , Ω α ) with canonical coordinates ( q , r ) (q,r) ( q , r ) and local Hamiltonian H α ( q , r ) H_\alpha(q,r) H α ( q , r ) . For curves ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) , t ∈ [ 0 , T ] t\in[0,T] t ∈ [ 0 , T ] , satisfying the mixed boundary conditions q ( 0 ) = q 0 q(0)=q_0 q ( 0 ) = q 0 and r ( T ) = r T r(T)=r_T r ( T ) = r T , we define the action by
S α I I [ q , r ] = r ( T ) ⋅ q ( T ) − ∫ 0 T [ r ( t ) ⋅ q ˙ ( t ) − H α ( q ( t ) , r ( t ) ) ] d t . (21) \mathcal{S}_{\alpha}^{II}[q,r]=r(T)\cdot q(T)-\int_0^T\left[r(t)\cdot\dot{q}(t)-H_\alpha(q(t),r(t))\right]\mathrm{d}t. \tag*{(21)} S α I I [ q , r ] = r ( T ) ⋅ q ( T ) − ∫ 0 T [ r ( t ) ⋅ q ˙ ( t ) − H α ( q ( t ) , r ( t )) ] d t . (21)
Proposition 2.2. A curve ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) is stationary for (21) with respect to variations satisfying δ q ( 0 ) = 0 \delta q(0)=0 δ q ( 0 ) = 0 and δ r ( T ) = 0 \delta r(T)=0 δ r ( T ) = 0 if and only if it satisfies Hamilton’s equations
q ˙ ( t ) = ∂ H α ∂ r ( q ( t ) , r ( t ) ) , r ˙ ( t ) = − ∂ H α ∂ q ( q ( t ) , r ( t ) ) . (22) \dot{q}(t)=\frac{\partial H_\alpha}{\partial r}(q(t),r(t)),\qquad\dot{r}(t)=-\frac{\partial H_\alpha}{\partial q}(q(t),r(t)). \tag*{(22)} q ˙ ( t ) = ∂ r ∂ H α ( q ( t ) , r ( t )) , r ˙ ( t ) = − ∂ q ∂ H α ( q ( t ) , r ( t )) . (22)
Proof. Integrating the term r ⋅ δ q ˙ r\cdot\delta\dot{q} r ⋅ δ q ˙ by parts and using δ q ( 0 ) = 0 \delta q(0)=0 δ q ( 0 ) = 0 and δ r ( T ) = 0 \delta r(T)=0 δ r ( T ) = 0 , the boundary contributions cancel and
δ S α I I = ∫ 0 T [ ( r ˙ + ∂ H α ∂ q ) ⋅ δ q − ( q ˙ − ∂ H α ∂ r ) ⋅ δ r ] d t , \delta\mathcal{S}_{\alpha}^{II}=\int_0^T\left[\left(\dot{r}+\frac{\partial H_\alpha}{\partial q}\right)\cdot\delta q-\left(\dot{q}-\frac{\partial H_\alpha}{\partial r}\right)\cdot\delta r\right]\mathrm{d}t, δ S α I I = ∫ 0 T [ ( r ˙ + ∂ q ∂ H α ) ⋅ δ q − ( q ˙ − ∂ r ∂ H α ) ⋅ δ r ] d t ,
which vanishes for arbitrary interior variations if and only if (22) holds.
This is the standard Hamiltonian variational principle with Type II boundary conditions. Its importance for the present work is that it uses the initial position and final momentum as boundary data, which are precisely the arguments of a right Hamiltonian generating function.
Recovery of the LCS Hamilton equations. The local canonical Hamilton equations can be expressed in the original LCS variables ( q , p ) (q,p) ( q , p ) through H α ( q , r ) = e − σ α ( q ) H ( q , p ) H_{\alpha}(q,r)=e^{-\sigma_{\alpha}(q)}H(q,p) H α ( q , r ) = e − σ α ( q ) H ( q , p ) and r i = e − σ α ( q ) p i r_i=e^{-\sigma_{\alpha}(q)}p_i r i = e − σ α ( q ) p i . Writing φ = φ i ( q ) d q i \varphi=\varphi_i(q)\,\mathrm{d}q^i φ = φ i ( q ) d q i with φ i = ∂ σ α / ∂ q i \varphi_i=\partial\sigma_{\alpha}/\partial q^i φ i = ∂ σ α / ∂ q i , differentiation of r i = e − σ α p i r_i=e^{-\sigma_{\alpha}}p_i r i = e − σ α p i gives r ˙ i = e − σ α ( p ˙ i − φ j q ˙ j p i ) \dot{r}_i=e^{-\sigma_{\alpha}}(\dot{p}_i-\varphi_j\dot{q}^jp_i) r ˙ i = e − σ α ( p ˙ i − φ j q ˙ j p i ) , and substituting into (22) we obtain
q ˙ i = ∂ H ∂ p i , p ˙ i = − ∂ H ∂ q i − A i j ∂ H ∂ p j + H φ i , A i j : = φ i p j − φ j p i . (23) \begin{aligned}
\dot{q}^i &= \frac{\partial H}{\partial p_i},\\
\dot{p}_i &= -\frac{\partial H}{\partial q^i}-A_{ij}\frac{\partial H}{\partial p_j}+H\varphi_i,\qquad A_{ij}:=\varphi_i p_j-\varphi_j p_i. \tag*{(23)}
\end{aligned} q ˙ i p ˙ i = ∂ p i ∂ H , = − ∂ q i ∂ H − A ij ∂ p j ∂ H + H φ i , A ij := φ i p j − φ j p i . (23)
These equations give the local coordinate expression of the intrinsic LCS Hamiltonian equation (10) . Hence, the Type II variational principle does not introduce a different continuous dynamics; rather, it provides a variational formulation of the LCS Hamiltonian dynamics, previously described in [17 ] , by expressing the local dynamics in canonical Hamiltonian form.
Conformal preservation of the continuous flow. A characteristic property of LCS Hamiltonian dynamics is that its Hamiltonian flow preserves the LCS two-form up to a conformal factor. By Cartan's formula, (5) and (10) ,
L ξ H Ω = d ( d H − H φ ) + ι ξ H ( φ ∧ Ω ) = − d H ∧ φ + φ ( ξ H ) Ω − φ ∧ ( d H − H φ ) = φ ( ξ H ) Ω , (24) \mathcal{L}_{\xi_H}\Omega=\mathrm{d}(\mathrm{d}H-H\varphi)+\iota_{\xi_H}(\varphi\wedge\Omega)=-\mathrm{d}H\wedge\varphi+\varphi(\xi_H)\Omega-\varphi\wedge(\mathrm{d}H-H\varphi)=\varphi(\xi_H)\Omega, \tag*{(24)} L ξ H Ω = d ( d H − H φ ) + ι ξ H ( φ ∧ Ω ) = − d H ∧ φ + φ ( ξ H ) Ω − φ ∧ ( d H − H φ ) = φ ( ξ H ) Ω , (24)
where we have used d φ = 0 \mathrm{d}\varphi=0 d φ = 0 and φ ∧ φ = 0 \varphi\wedge\varphi=0 φ ∧ φ = 0 . Let ψ t : M → M \psi_t:M\to M ψ t : M → M denote the integral flow of ξ H \xi_H ξ H . Since
d d t ψ t ∗ Ω = ψ t ∗ ( L ξ H Ω ) = ( φ ( ξ H ) ∘ ψ t ) ψ t ∗ Ω \frac{\mathrm{d}}{\mathrm{d}t}\psi_t^*\Omega=\psi_t^*(\mathcal{L}_{\xi_H}\Omega)=(\varphi(\xi_H)\circ\psi_t)\psi_t^*\Omega d t d ψ t ∗ Ω = ψ t ∗ ( L ξ H Ω ) = ( φ ( ξ H ) ∘ ψ t ) ψ t ∗ Ω
and ψ 0 = id \psi_0=\operatorname{id} ψ 0 = id ,
ψ t ∗ Ω = exp ( ∫ 0 t φ ( ξ H ) ( ψ s ( x ) ) d s ) Ω . (25) \psi_t^*\Omega=\exp\left(\int_0^t\varphi(\xi_H)(\psi_s(x))\,\mathrm{d}s\right)\Omega. \tag*{(25)} ψ t ∗ Ω = exp ( ∫ 0 t φ ( ξ H ) ( ψ s ( x )) d s ) Ω. (25)
If the flow segment under consideration remains in a local chart U α U_{\alpha} U α , then φ ( ξ H ) = d σ α ( ξ H ) \varphi(\xi_H)=\mathrm{d}\sigma_{\alpha}(\xi_H) φ ( ξ H ) = d σ α ( ξ H ) , so that φ ( ξ H ) ( ψ t ( x ) ) \varphi(\xi_H)(\psi_t(x)) φ ( ξ H ) ( ψ t ( x )) is the time derivative of σ α ( ψ t ( x ) ) \sigma_{\alpha}(\psi_t(x)) σ α ( ψ t ( x )) , and therefore
ψ t ∗ Ω = e σ α ∘ ψ t − σ α Ω . (26) \psi_t^*\Omega=e^{\sigma_{\alpha}\circ\psi_t-\sigma_{\alpha}}\Omega. \tag*{(26)} ψ t ∗ Ω = e σ α ∘ ψ t − σ α Ω. (26)
Although the right-hand side is expressed in terms of the local potential σ α \sigma_{\alpha} σ α , the conformal factor is independent of the choice of local potential: whenever x x x and ψ t ( x ) \psi_t(x) ψ t ( x ) belong to the same connected component of U α ∩ U β U_{\alpha}\cap U_{\beta} U α ∩ U β , the constant C α β C_{\alpha\beta} C α β cancels and σ α ( ψ t ( x ) ) − σ α ( x ) = σ β ( ψ t ( x ) ) − σ β ( x ) \sigma_{\alpha}(\psi_t(x))-\sigma_{\alpha}(x)=\sigma_{\beta}(\psi_t(x))-\sigma_{\beta}(x) σ α ( ψ t ( x )) − σ α ( x ) = σ β ( ψ t ( x )) − σ β ( x ) . Equation (26) is the continuous structure-preservation law that the discrete theory will ultimately seek to reproduce.
Continuous example
We conclude the continuous discussion with the example that will later be used to illustrate the numerical constructions.
Example 2.3. Let Q = R Q=\mathbb{R} Q = R and take Q × R Q\times\mathbb{R} Q × R as the base manifold, that is, the setting of Section 2.1 with Q Q Q replaced by Q × R Q\times\mathbb{R} Q × R , with coordinates ( q , x ) (q,x) ( q , x ) and induced coordinates ( q , p , x , y ) (q,p,x,y) ( q , p , x , y ) on the cotangent bundle T ∗ ( Q × R ) T^*(Q\times\mathbb{R}) T ∗ ( Q × R ) . Consider the Lee form φ = − d x \varphi=-\mathrm{d}x φ = − d x , with local conformal potential σ = − x \sigma=-x σ = − x . Since φ \varphi φ is the pullback of a closed one-form on Q × R Q\times\mathbb{R} Q × R under the cotangent-bundle projection, the cotangent-type hypothesis (13) applies, and in the LCS variables
Ω = d q ∧ d p + d x ∧ d y + p d q ∧ d x . (27) \Omega=\mathrm{d}q\wedge\mathrm{d}p+\mathrm{d}x\wedge\mathrm{d}y+p\,\mathrm{d}q\wedge\mathrm{d}x. \tag*{(27)} Ω = d q ∧ d p + d x ∧ d y + p d q ∧ d x . (27)
Consider the Hamiltonian H ( q , p , x , y ) = 1 2 ( p 2 + y 2 ) + k 2 q 2 + μ x H(q,p,x,y)=\frac{1}{2}(p^{2}+y^{2})+\frac{k}{2}q^{2}+\mu x H ( q , p , x , y ) = 2 1 ( p 2 + y 2 ) + 2 k q 2 + μx , where k k k and μ \mu μ are real parameters. Introducing the local canonical momenta r = e x p r=e^{x}p r = e x p and s = e x y s=e^{x}y s = e x y , the local symplectic form is Ω l o c = d q ∧ d r + d x ∧ d s \Omega_{\mathrm{loc}}=\mathrm{d}q\wedge\mathrm{d}r+\mathrm{d}x\wedge\mathrm{d}s Ω loc = d q ∧ d r + d x ∧ d s , with Ω = e − x Ω l o c \Omega=e^{-x}\Omega_{\mathrm{loc}} Ω = e − x Ω loc consistently with (27) , and the local Hamiltonian is
H l o c ( q , r , x , s ) = 1 2 e − x ( r 2 + s 2 ) + e x ( k 2 q 2 + μ x ) . (28) H_{\mathrm{loc}}(q,r,x,s)=\frac{1}{2}e^{-x}(r^{2}+s^{2})+e^{x}\left(\frac{k}{2}q^{2}+\mu x\right). \tag*{(28)} H loc ( q , r , x , s ) = 2 1 e − x ( r 2 + s 2 ) + e x ( 2 k q 2 + μx ) . (28)
whose canonical Hamilton equations are
q ˙ = e − x r , r ˙ = − k e x q , x ˙ = e − x s , s ˙ = 1 2 e − x ( r 2 + s 2 ) − e x [ k 2 q 2 + μ ( x + 1 ) ] . (29) \dot{q}=e^{-x}r,\qquad\dot{r}=-ke^{x}q,\qquad\dot{x}=e^{-x}s,\qquad\dot{s}=\frac{1}{2}e^{-x}(r^{2}+s^{2})-e^{x}\left[\frac{k}{2}q^{2}+\mu(x+1)\right]. \tag*{(29)} q ˙ = e − x r , r ˙ = − k e x q , x ˙ = e − x s , s ˙ = 2 1 e − x ( r 2 + s 2 ) − e x [ 2 k q 2 + μ ( x + 1 ) ] . (29)
The continuous LCS flow therefore satisfies ψ t ∗ Ω = e − ( x ( t ) − x ( 0 ) ) Ω \psi_{t}^{*}\Omega=e^{-(x(t)-x(0))}\Omega ψ t ∗ Ω = e − ( x ( t ) − x ( 0 )) Ω . Its numerical discretization will be considered only after the general discrete theory has been established.
The continuous framework therefore supplies two pieces of information that must survive discretization: each local representative is an ordinary symplectic Hamiltonian system, and different representatives are related on overlaps by the constant fiber scalings T α β T_{\alpha\beta} T α β . The exact flows respect both structures automatically. We now construct Type II discrete Hamiltonian maps locally and then determine the additional compatibility condition required for those maps to globalize.
Discrete Type II Hamiltonian mechanics
We now turn to the discrete problem. The construction is first carried out on a single local symplectic representative ( U α , Ω α ) (U_{\alpha},\Omega_{\alpha}) ( U α , Ω α ) , with canonical coordinates ( q , r ) (q,r) ( q , r ) and local Hamiltonian H α H_{\alpha} H α ; no compatibility between different charts is imposed at this stage. The basic object is a right discrete Hamiltonian, that is, a Type II generating function depending on an initial position and a final momentum. Such generating-function techniques have also been used to construct symplectic integrators for optimal control problems [12 ] . We first recall the exact discrete right Hamiltonian and then replace it by a computable approximation, which produces a local symplectic one-step map. The behavior of these local constructions on overlaps will be studied in § 4 .
The exact discrete right Hamiltonian
Fix a time step h > 0 h>0 h > 0 and consider the exact solution ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) , t ∈ [ 0 , h ] t\in[0,h] t ∈ [ 0 , h ] , of the local Hamiltonian system q ˙ = D 2 H α ( q , r ) \dot{q}=D_{2}H_{\alpha}(q,r) q ˙ = D 2 H α ( q , r ) , r ˙ = − D 1 H α ( q , r ) \dot{r}=-D_{1}H_{\alpha}(q,r) r ˙ = − D 1 H α ( q , r ) , subject to the Type II boundary conditions
q ( 0 ) = q 0 , r ( h ) = r 1 . (30) q(0)=q_{0},\qquad r(h)=r_{1}. \tag*{(30)} q ( 0 ) = q 0 , r ( h ) = r 1 . (30)
We assume throughout this discussion that the corresponding local two-point boundary-value problem has a unique solution.
Definition 3.1. The exact discrete right Hamiltonian associated with H α H_{\alpha} H α is
H α , h + , E ( q 0 , r 1 ) = r 1 ⋅ q 1 − ∫ 0 h [ r ( t ) ⋅ q ˙ ( t ) − H α ( q ( t ) , r ( t ) ) ] d t , (31) H_{\alpha,h}^{+,E}(q_{0},r_{1})=r_{1}\cdot q_{1}-\int_{0}^{h}\left[r(t)\cdot\dot{q}(t)-H_{\alpha}(q(t),r(t))\right]\mathrm{d}t, \tag*{(31)} H α , h + , E ( q 0 , r 1 ) = r 1 ⋅ q 1 − ∫ 0 h [ r ( t ) ⋅ q ˙ ( t ) − H α ( q ( t ) , r ( t )) ] d t , (31) where q 1 = q ( h ) q_{1}=q(h) q 1 = q ( h ) and ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) is the exact solution satisfying (30) .
The exact discrete right Hamiltonian is the Type II generating function of the exact time-h h h Hamiltonian flow.
Proposition 3.2. The exact discrete right Hamiltonian satisfies
r 0 = D 1 H α , h + , E ( q 0 , r 1 ) , q 1 = D 2 H α , h + , E ( q 0 , r 1 ) , (32) r_0 = D_1 H_{\alpha,h}^{+,E}(q_0,r_1), \qquad q_1 = D_2 H_{\alpha,h}^{+,E}(q_0,r_1), \tag*{(32)} r 0 = D 1 H α , h + , E ( q 0 , r 1 ) , q 1 = D 2 H α , h + , E ( q 0 , r 1 ) , (32) where r 0 = r ( 0 ) r_0=r(0) r 0 = r ( 0 ) and q 1 = q ( h ) q_1=q(h) q 1 = q ( h ) . Consequently, H α , h + , E H_{\alpha,h}^{+,E} H α , h + , E generates the exact time-h h h Hamiltonian map.
Proof. Evaluating the variation of the Type II action (21) on a solution of Hamilton’s equations, the interior terms vanish and only the boundary contribution δ H α , h + , E = r 0 ⋅ δ q 0 + q 1 ⋅ δ r 1 \delta H_{\alpha,h}^{+,E}=r_0\cdot\delta q_0+q_1\cdot\delta r_1 δ H α , h + , E = r 0 ⋅ δ q 0 + q 1 ⋅ δ r 1 remains, which is (32) .
Thus the exact time-h h h Hamiltonian map may be encoded entirely by a function of the mixed endpoint data ( q 0 , r 1 ) (q_0,r_1) ( q 0 , r 1 ) .
Approximate discrete right Hamiltonians. The exact discrete Hamiltonian (31) is generally not available in closed form. A numerical method is obtained by replacing it with an approximation.
Definition 3.3. A discrete right Hamiltonian is a smooth function H α , h d + ( q 0 , r 1 ) H_{\alpha,h}^{d+}(q_0,r_1) H α , h d + ( q 0 , r 1 ) approximating the exact discrete right Hamiltonian H α , h + , E ( q 0 , r 1 ) H_{\alpha,h}^{+,E}(q_0,r_1) H α , h + , E ( q 0 , r 1 ) .
The corresponding right discrete Hamilton equations are
r 0 = D 1 H α , h d + ( q 0 , r 1 ) , q 1 = D 2 H α , h d + ( q 0 , r 1 ) . (33) r_0 = D_1 H_{\alpha,h}^{d+}(q_0,r_1), \qquad q_1 = D_2 H_{\alpha,h}^{d+}(q_0,r_1). \tag*{(33)} r 0 = D 1 H α , h d + ( q 0 , r 1 ) , q 1 = D 2 H α , h d + ( q 0 , r 1 ) . (33)
Given ( q 0 , r 0 ) (q_0,r_0) ( q 0 , r 0 ) , the first equation is solved implicitly for r 1 r_1 r 1 and the second equation then determines q 1 q_1 q 1 . Whenever this procedure is locally well posed, it defines a one-step map F α , h : ( q 0 , r 0 ) ↦ ( q 1 , r 1 ) F_{\alpha,h}:(q_0,r_0)\mapsto(q_1,r_1) F α , h : ( q 0 , r 0 ) ↦ ( q 1 , r 1 ) . A convenient sufficient regularity condition is the nonsingularity of the mixed derivative D 12 H α , h d + ( q 0 , r 1 ) D_{12}H_{\alpha,h}^{d+}(q_0,r_1) D 12 H α , h d + ( q 0 , r 1 ) , under which the first equation can be solved locally for r 1 r_1 r 1 as a function of ( q 0 , r 0 ) (q_0,r_0) ( q 0 , r 0 ) by the implicit function theorem.
Discrete Type II variational formulation
The one-step relations (33) admit a discrete phase-space variational interpretation. Consider a sequence { ( q k , r k ) } k = 0 N \{(q_k,r_k)\}_{k=0}^{N} {( q k , r k ) } k = 0 N with the Type II endpoint conditions q 0 q_0 q 0 and r N r_N r N fixed, and define the discrete Type II action by
S α , d I I = r N ⋅ q N − ∑ k = 0 N − 1 [ r k + 1 ⋅ q k + 1 − H α , h d + ( q k , r k + 1 ) ] . (34) S_{\alpha,d}^{II}=r_N\cdot q_N-\sum_{k=0}^{N-1}\left[r_{k+1}\cdot q_{k+1}-H_{\alpha,h}^{d+}(q_k,r_{k+1})\right]. \tag*{(34)} S α , d I I = r N ⋅ q N − k = 0 ∑ N − 1 [ r k + 1 ⋅ q k + 1 − H α , h d + ( q k , r k + 1 ) ] . (34)
The endpoint term cancels the final contribution r N ⋅ q N r_N\cdot q_N r N ⋅ q N contained in the sum. Taking the variation and using δ q 0 = 0 \delta q_0=0 δ q 0 = 0 , δ r N = 0 \delta r_N=0 δ r N = 0 gives
δ S α , d I I = ∑ k = 1 N − 1 [ − r k + D 1 H α , h d + ( q k , r k + 1 ) ] ⋅ δ q k + ∑ k = 1 N − 1 [ − q k + D 2 H α , h d + ( q k − 1 , r k ) ] ⋅ δ r k , (35) \delta S_{\alpha,d}^{II}=\sum_{k=1}^{N-1}\left[-r_k+D_1H_{\alpha,h}^{d+}(q_k,r_{k+1})\right]\cdot\delta q_k+\sum_{k=1}^{N-1}\left[-q_k+D_2H_{\alpha,h}^{d+}(q_{k-1},r_k)\right]\cdot\delta r_k, \tag*{(35)} δ S α , d I I = k = 1 ∑ N − 1 [ − r k + D 1 H α , h d + ( q k , r k + 1 ) ] ⋅ δ q k + k = 1 ∑ N − 1 [ − q k + D 2 H α , h d + ( q k − 1 , r k ) ] ⋅ δ r k , (35)
and we obtain the following interior stationarity equations.
Proposition 3.4. A sequence is stationary for (34) , with q 0 q_0 q 0 and r N r_N r N fixed, if and only if its interior variables satisfy
r k = D 1 H α , h d + ( q k , r k + 1 ) , q k = D 2 H α , h d + ( q k − 1 , r k ) , k = 1 , … , N − 1. (36) r_k=D_1H_{\alpha,h}^{d+}(q_k,r_{k+1}), \qquad q_k=D_2H_{\alpha,h}^{d+}(q_{k-1},r_k), \qquad k=1,\ldots,N-1. \tag*{(36)} r k = D 1 H α , h d + ( q k , r k + 1 ) , q k = D 2 H α , h d + ( q k − 1 , r k ) , k = 1 , … , N − 1. (36)
Remark 3.5. The stationarity equations are interior equations. Because q 0 q_0 q 0 is fixed, variation of the discrete action does not yield an equation for the initial momentum r 0 r_0 r 0 , and because q N q_N q N cancels from (34) , it does not yield an equation for the final position either. Both are recovered from the Type II generating-function relations r 0 = D 1 H α , h d + ( q 0 , r 1 ) r_0 = D_1H^{d+}_{\alpha,h}(q_0,r_1) r 0 = D 1 H α , h d + ( q 0 , r 1 ) and q N = D 2 H α , h d + ( q N − 1 , r N ) q_N = D_2H^{d+}_{\alpha,h}(q_{N-1},r_N) q N = D 2 H α , h d + ( q N − 1 , r N ) . Likewise, the one-step update is described by the pair of generating relations
r k = D 1 H α , h d + ( q k , r k + 1 ) , q k + 1 = D 2 H α , h d + ( q k , r k + 1 ) . (37) r_k = D_1H^{d+}_{\alpha,h}(q_k,r_{k+1}), \qquad q_{k+1} = D_2H^{d+}_{\alpha,h}(q_k,r_{k+1}). \tag*{(37)} r k = D 1 H α , h d + ( q k , r k + 1 ) , q k + 1 = D 2 H α , h d + ( q k , r k + 1 ) . (37)
This distinction between the interior stationarity conditions and the endpoint generating relations is useful when formulating the discrete Type II principle with mixed boundary data.
Local symplecticity.
Proposition 3.6. Suppose that H α , h d + H^{d+}_{\alpha,h} H α , h d + is regular enough to define the local one-step map F α , h F_{\alpha,h} F α , h . Then F α , h F_{\alpha,h} F α , h is symplectic with respect to the local symplectic form Ω α \Omega_\alpha Ω α .
Proof. By (33) , d H α , h d + = r 0 d q 0 + q 1 d r 1 \mathrm{d}H^{d+}_{\alpha,h} = r_0\,\mathrm{d}q_0 + q_1\,\mathrm{d}r_1 d H α , h d + = r 0 d q 0 + q 1 d r 1 , and taking one more exterior derivative gives 0 = d r 0 ∧ d q 0 + d q 1 ∧ d r 1 0 = \mathrm{d}r_0 \wedge\mathrm{d}q_0 + \mathrm{d}q_1 \wedge\mathrm{d}r_1 0 = d r 0 ∧ d q 0 + d q 1 ∧ d r 1 , i.e.,
d q 0 ∧ d r 0 = d q 1 ∧ d r 1 , (38) \mathrm{d}q_0 \wedge\mathrm{d}r_0 = \mathrm{d}q_1 \wedge\mathrm{d}r_1, \tag*{(38)} d q 0 ∧ d r 0 = d q 1 ∧ d r 1 , (38)
which is F α , h ∗ Ω α = Ω α F_{\alpha,h}^{*}\Omega_\alpha= \Omega_\alpha F α , h ∗ Ω α = Ω α for Ω α = d q i ∧ d r i \Omega_\alpha= \mathrm{d}q^i \wedge\mathrm{d}r_i Ω α = d q i ∧ d r i .
This is a purely local statement. It says that the discrete dynamics is symplectic when expressed in one canonical representative. It does not imply that independently constructed maps on different local representations determine a single global numerical map.
Order of a discrete Hamiltonian method.
The approximation order of the discrete Hamiltonian controls the order of the induced one-step map.
Theorem 3.7 ([39 ] ). Let H α , h + , E H^{+,E}_{\alpha,h} H α , h + , E denote the exact discrete right Hamiltonian and let H α , h d + H^{d+}_{\alpha,h} H α , h d + be an approximate discrete right Hamiltonian. Assume, on a compact set on which the exact and discrete maps are well-defined, that
H α , h d + − H α , h + , E = O ( h p + 1 ) (39) H^{d+}_{\alpha,h} - H^{+,E}_{\alpha,h} = O(h^{p+1}) \tag*{(39)} H α , h d + − H α , h + , E = O ( h p + 1 ) (39) in the C 1 C^1 C 1 sense as h → 0 h \to0 h → 0 . Then, under the usual local regularity assumptions for the corresponding generating relations, the one-step map generated by H α , h d + H^{d+}_{\alpha,h} H α , h d + approximates the exact time-h h h Hamiltonian flow with order p p p .
The C 1 C^1 C 1 requirement is relevant because the numerical map is generated by the derivatives
D 1 H α , h d + , D 2 H α , h d + D_1H^{d+}_{\alpha,h}, \qquad D_2H^{d+}_{\alpha,h} D 1 H α , h d + , D 2 H α , h d +
rather than by the value of the discrete Hamiltonian alone.
Remark 3.8 (Converse of Theorem 3.7). The converse also holds, up to an additive constant. Suppose that the one-step map F α , h F_{\alpha,h} F α , h generated by a regular discrete right Hamiltonian H α , h d + H^{d+}_{\alpha,h} H α , h d + approximates the exact time-h h h flow Φ α , h \Phi_{\alpha,h} Φ α , h with order p p p on a compact set, i.e., F α , h − Φ α , h = O ( h p + 1 ) F_{\alpha,h} - \Phi_{\alpha,h} = O(h^{p+1}) F α , h − Φ α , h = O ( h p + 1 ) there. Let π q \pi_q π q and π r \pi_r π r denote the projections onto the position and momentum, respectively, and let
G α , h ( q 0 , r 0 ) = ( q 0 , π r F α , h ( q 0 , r 0 ) ) = ( q 0 , r 1 ) G_{\alpha,h}(q_0,r_0) = (q_0,\pi_rF_{\alpha,h}(q_0,r_0)) = (q_0,r_1) G α , h ( q 0 , r 0 ) = ( q 0 , π r F α , h ( q 0 , r 0 )) = ( q 0 , r 1 )
denote the map from the initial data to the mixed data. Let G α , h E G^{E}_{\alpha,h} G α , h E denote the corresponding mixed-data map for Φ α , h \Phi_{\alpha,h} Φ α , h . Assume that these maps admit inverse branches on a common mixed-data neighborhood, with uniformly bounded Lipschitz constants for the inverses as h → 0 h \to0 h → 0 . Assume also that both inverse branches take values in the compact set on which F α , h − Φ α , h = O ( h p + 1 ) F_{\alpha,h}-\Phi_{\alpha,h}=O(h^{p+1}) F α , h − Φ α , h = O ( h p + 1 ) holds uniformly. The generating relations (32) and (33) can then be written as
D 1 H α , h d + = π r ∘ G α , h − 1 , D 2 H α , h d + = π q ∘ F α , h ∘ G α , h − 1 , D_1H_{\alpha,h}^{d+}=\pi_r\circ G_{\alpha,h}^{-1},\qquad D_2H_{\alpha,h}^{d+}=\pi_q\circ F_{\alpha,h}\circ G_{\alpha,h}^{-1}, D 1 H α , h d + = π r ∘ G α , h − 1 , D 2 H α , h d + = π q ∘ F α , h ∘ G α , h − 1 ,
and likewise for H α , h + , E H_{\alpha,h}^{+,E} H α , h + , E with Φ α , h \Phi_{\alpha,h} Φ α , h and G α , h E G_{\alpha,h}^{E} G α , h E in place of F α , h F_{\alpha,h} F α , h and G α , h G_{\alpha,h} G α , h . Since G α , h − G α , h E = O ( h p + 1 ) G_{\alpha,h}-G_{\alpha,h}^{E}=O(h^{p+1}) G α , h − G α , h E = O ( h p + 1 ) , the inverses and the compositions differ by O ( h p + 1 ) O(h^{p+1}) O ( h p + 1 ) as well [39 Lemma 2.3] , so that
D i H α , h d + − D i H α , h + , E = O ( h p + 1 ) , i = 1 , 2 , D_iH_{\alpha,h}^{d+}-D_iH_{\alpha,h}^{+,E}=O(h^{p+1}),\qquad i=1,2, D i H α , h d + − D i H α , h + , E = O ( h p + 1 ) , i = 1 , 2 ,
uniformly on compact subsets of the common mixed-data neighborhood. On a fixed convex compact subdomain of this neighborhood, integrating the derivative difference along line segments shows that, up to an additive constant depending only on h h h , H α , h d + − H α , h + , E = O ( h p + 1 ) H_{\alpha,h}^{d+}-H_{\alpha,h}^{+,E}=O(h^{p+1}) H α , h d + − H α , h + , E = O ( h p + 1 ) in the C 1 C^1 C 1 sense. Since additive constants do not affect the generated map, the order of a Hamiltonian variational integrator and the order to which its discrete right Hamiltonian approximates the exact discrete right Hamiltonian coincide.
The local construction and the compatibility problem. Thus, on every local symplectic representative ( U α , Ω α , H α ) (U_\alpha,\Omega_\alpha,H_\alpha) ( U α , Ω α , H α ) , an approximate discrete right Hamiltonian H α , h d + H_{\alpha,h}^{d+} H α , h d + determines, under the regularity assumptions above, a symplectic one-step map F α , h F_{\alpha,h} F α , h . On an overlap, the continuous Hamiltonians are related by (17) and their exact flows are equivariant with respect to T α β T_{\alpha\beta} T α β . There is no analogous automatic conclusion for two independently chosen approximations H α , h d + H_{\alpha,h}^{d+} H α , h d + and H β , h d + H_{\beta,h}^{d+} H β , h d + ; in particular, local symplecticity does not imply F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h . The missing ingredient is a compatibility condition on the discrete generating functions. We derive it next from the transformation law of the exact discrete right Hamiltonian; this keeps the discrete condition tied directly to the underlying continuous LCS geometry.
Globalization of discrete Hamiltonian dynamics
The construction of Section 3 associates a symplectic one-step map with each local canonical representation of the LCS Hamiltonian system. The remaining question is whether these local maps represent a single discrete dynamics on the underlying LCS phase space. The continuous theory provides the model for the required compatibility: on a connected component of an overlap U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β , the local canonical variables are related by the constant fiber scaling T α β T_{\alpha\beta} T α β of (16) and the local Hamiltonians by (17) . We first show that the exact discrete right Hamiltonians inherit precisely the same transformation law. This observation provides the natural compatibility condition for approximate discrete Hamiltonians.
Transformation of the exact discrete right Hamiltonian
Let H α , h + , E ( q 0 , r 1 ) H_{\alpha,h}^{+,E}(q_0,r_1) H α , h + , E ( q 0 , r 1 ) and H β , h + , E ( q 0 , s 1 ) H_{\beta,h}^{+,E}(q_0,s_1) H β , h + , E ( q 0 , s 1 ) denote the exact discrete right Hamiltonians (31) associated with H α H_\alpha H α and H β H_\beta H β , respectively.
Theorem 4.1. For mixed data ( q 0 , r 1 ) (q_0,r_1) ( q 0 , r 1 ) such that the solution ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) of the α \alpha α -Hamiltonian system with q ( 0 ) = q 0 q(0)=q_0 q ( 0 ) = q 0 and r ( h ) = r 1 r(h)=r_1 r ( h ) = r 1 remains in a single connected component of U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β for all t ∈ [ 0 , h ] t\in[0,h] t ∈ [ 0 , h ] , set s 1 = c α β r 1 s_1=c_{\alpha\beta}r_1 s 1 = c α β r 1 . Then the exact discrete right Hamiltonians satisfy
H β , h + , E ( q 0 , s 1 ) = c α β H α , h + , E ( q 0 , c α β − 1 s 1 ) . (40) H_{\beta,h}^{+,E}(q_0,s_1)=c_{\alpha\beta}H_{\alpha,h}^{+,E}\left(q_0,c_{\alpha\beta}^{-1}s_1\right). \tag*{(40)} H β , h + , E ( q 0 , s 1 ) = c α β H α , h + , E ( q 0 , c α β − 1 s 1 ) . (40)
Proof. Let ( q ( t ) , r ( t ) ) (q(t), r(t)) ( q ( t ) , r ( t )) be the solution of the α \alpha α -Hamiltonian system satisfying q ( 0 ) = q 0 q(0)=q_0 q ( 0 ) = q 0 , r ( h ) = r 1 r(h)=r_1 r ( h ) = r 1 , and set s ( t ) = c α β r ( t ) s(t)=c_{\alpha\beta}r(t) s ( t ) = c α β r ( t ) . By the continuous equivariance relation (20) , ( q ( t ) , s ( t ) ) (q(t),s(t)) ( q ( t ) , s ( t )) is the solution of the β \beta β -Hamiltonian system with q ( 0 ) = q 0 q(0)=q_0 q ( 0 ) = q 0 and s ( h ) = s 1 = c α β r 1 s(h)=s_1=c_{\alpha\beta}r_1 s ( h ) = s 1 = c α β r 1 . Along these trajectories H β ( q , s ) = c α β H α ( q , r ) H_\beta(q,s)=c_{\alpha\beta}H_\alpha(q,r) H β ( q , s ) = c α β H α ( q , r ) by (17) and s ⋅ q ˙ = c α β r ⋅ q ˙ s\cdot\dot q=c_{\alpha\beta}r\cdot\dot q s ⋅ q ˙ = c α β r ⋅ q ˙ , so every term of (31) in the β \beta β -representation is c α β c_{\alpha\beta} c α β times the corresponding term in the α \alpha α -representation, i.e., H β , h + , E ( q 0 , s 1 ) = c α β H α , h + , E ( q 0 , r 1 ) H_{\beta,h}^{+,E}(q_0,s_1)=c_{\alpha\beta}H_{\alpha,h}^{+,E}(q_0,r_1) H β , h + , E ( q 0 , s 1 ) = c α β H α , h + , E ( q 0 , r 1 ) with r 1 = c α β − 1 s 1 r_1=c_{\alpha\beta}^{-1}s_1 r 1 = c α β − 1 s 1 .
Equation (40) is not an additional compatibility assumption. It is inherited by the exact Type II generating functions from the overlap relation of the continuous Hamiltonian systems, and it is the basis for the compatibility condition imposed below on numerical approximations.
Compatible discrete right Hamiltonians.
Definition 4.2. A collection of local discrete right Hamiltonians { H α , h d + } α ∈ A \{H_{\alpha,h}^{d+}\}_{\alpha\in A} { H α , h d + } α ∈ A is called LCS-compatible if, on every connected component of an overlap U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β , one has
H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) . (41) H_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}H_{\alpha,h}^{d+}\left(q,c_{\alpha\beta}^{-1}s\right). \tag*{(41)} H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) . (41) This is the discrete analogue of the continuous transformation law (17) . More importantly, it is a condition imposed directly on the discrete generating functions and can therefore be checked at the level at which the numerical method is constructed. Differentiating (41) gives
D 1 H β , h d + ( q , s ) = c α β D 1 H α , h d + ( q , c α β − 1 s ) , D 2 H β , h d + ( q , s ) = D 2 H α , h d + ( q , c α β − 1 s ) ; (42) D_1H_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}D_1H_{\alpha,h}^{d+}\left(q,c_{\alpha\beta}^{-1}s\right), \qquad D_2H_{\beta,h}^{d+}(q,s)=D_2H_{\alpha,h}^{d+}\left(q,c_{\alpha\beta}^{-1}s\right); \tag*{(42)} D 1 H β , h d + ( q , s ) = c α β D 1 H α , h d + ( q , c α β − 1 s ) , D 2 H β , h d + ( q , s ) = D 2 H α , h d + ( q , c α β − 1 s ) ; (42) the cancellation of the factor c α β c_{\alpha\beta} c α β in the second identity is precisely what is needed for the position update to be independent of the local representation.
Compatibility implies discrete equivariance
Let F α , h ( q k , r k ) = ( q k + 1 , r k + 1 ) F_{\alpha,h}(q_k,r_k)=(q_{k+1},r_{k+1}) F α , h ( q k , r k ) = ( q k + 1 , r k + 1 ) be the local discrete map generated by H α , h d + H_{\alpha,h}^{d+} H α , h d + through (37) , and let F β , h F_{\beta,h} F β , h be the map generated by H β , h d + H_{\beta,h}^{d+} H β , h d + .
Theorem 4.3 (Discrete equivariance). Suppose that the local discrete right Hamiltonians are LCS-compatible in the sense of Definition 4.2, and assume that the corresponding generating relations define unique local one-step maps. Then
F β , h ∘ T α β = T α β ∘ F α , h (43) F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} \tag*{(43)} F β , h ∘ T α β = T α β ∘ F α , h (43) on their common domain, i.e., at every point ( q k , r k ) (q_k,r_k) ( q k , r k ) such that ( q k , r k ) (q_k,r_k) ( q k , r k ) and F α , h ( q k , r k ) F_{\alpha,h}(q_k,r_k) F α , h ( q k , r k ) lie in the same connected component of the overlap, so that a single constant c α β c_{\alpha\beta} c α β applies to both endpoints of the step.
Proof. Let F α , h ( q k , r k ) = ( q k + 1 , r k + 1 ) F_{\alpha,h}(q_k,r_k)=(q_{k+1},r_{k+1}) F α , h ( q k , r k ) = ( q k + 1 , r k + 1 ) , so that r k = D 1 H α , h d + ( q k , r k + 1 ) r_k=D_1H_{\alpha,h}^{d+}(q_k,r_{k+1}) r k = D 1 H α , h d + ( q k , r k + 1 ) and q k + 1 = D 2 H α , h d + ( q k , r k + 1 ) q_{k+1}=D_2H_{\alpha,h}^{d+}(q_k,r_{k+1}) q k + 1 = D 2 H α , h d + ( q k , r k + 1 ) , and set s k = c α β r k s_k=c_{\alpha\beta}r_k s k = c α β r k , s k + 1 = c α β r k + 1 s_{k+1}=c_{\alpha\beta}r_{k+1} s k + 1 = c α β r k + 1 . By (42) ,
D 1 H β , h d + ( q k , s k + 1 ) = c α β D 1 H α , h d + ( q k , r k + 1 ) = s k , D 2 H β , h d + ( q k , s k + 1 ) = D 2 H α , h d + ( q k , r k + 1 ) = q k + 1 , D_1H_{\beta,h}^{d+}(q_k,s_{k+1})=c_{\alpha\beta}D_1H_{\alpha,h}^{d+}(q_k,r_{k+1})=s_k,\qquad D_2H_{\beta,h}^{d+}(q_k,s_{k+1})=D_2H_{\alpha,h}^{d+}(q_k,r_{k+1})=q_{k+1}, D 1 H β , h d + ( q k , s k + 1 ) = c α β D 1 H α , h d + ( q k , r k + 1 ) = s k , D 2 H β , h d + ( q k , s k + 1 ) = D 2 H α , h d + ( q k , r k + 1 ) = q k + 1 , so that ( q k , s k ) (q_k,s_k) ( q k , s k ) and ( q k + 1 , s k + 1 ) (q_{k+1},s_{k+1}) ( q k + 1 , s k + 1 ) satisfy the generating relations of H β , h d + H_{\beta,h}^{d+} H β , h d + . By uniqueness,
F β , h ( T α β ( q k , r k ) ) = ( q k + 1 , c α β r k + 1 ) = T α β ( F α , h ( q k , r k ) ) , F_{\beta,h}\left(T_{\alpha\beta}(q_k,r_k)\right)=(q_{k+1},c_{\alpha\beta}r_{k+1})=T_{\alpha\beta}\left(F_{\alpha,h}(q_k,r_k)\right), F β , h ( T α β ( q k , r k ) ) = ( q k + 1 , c α β r k + 1 ) = T α β ( F α , h ( q k , r k ) ) , which proves (43) .
Thus the transformation law for the discrete generating functions is a sufficient condition for compatibility of the resulting numerical maps.
Gluing of the local discrete maps.
Lemma 4.4 (Gluing lemma). Let { F α , h } \{F_{\alpha,h}\} { F α , h } be a collection of local maps satisfying F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h on every overlap. Then the local maps define a unique discrete map on the phase space covered by the local charts, wherever the local one-step maps are defined.
Proof. If a point is represented in two overlapping local descriptions by z β = T α β ( z α ) z_\beta=T_{\alpha\beta}(z_\alpha) z β = T α β ( z α ) , then F β , h ( z β ) = T α β ( F α , h ( z α ) ) F_{\beta,h}(z_\beta)=T_{\alpha\beta}(F_{\alpha,h}(z_\alpha)) F β , h ( z β ) = T α β ( F α , h ( z α )) , so the two local updates represent the same point after one time step; the cocycle relation (9) guarantees consistency on multiple overlaps. □
The gluing statement itself is elementary. The relevant point for the present construction is that Theorem 4.3 provides a concrete sufficient condition for its hypothesis directly in terms of the discrete Hamiltonians.
The requirement that each step remain within a single chart is ensured on compact sets for sufficiently small step sizes; see Remark 5.10 .
Reconstruction in the LCS variables. We now return from the local canonical variables to the original LCS variables through the representation map Ψ α \Psi_\alpha Ψ α of (14) , whose inverse is Ψ α − 1 ( q , r ) = ( q , e σ α ( q ) r ) \Psi_\alpha^{-1}(q,r)=(q,e^{\sigma_\alpha(q)}r) Ψ α − 1 ( q , r ) = ( q , e σ α ( q ) r ) . Given a local discrete map F α , h F_{\alpha,h} F α , h in canonical variables, define the reconstructed map in the LCS variables by
F h = Ψ α − 1 ∘ F α , h ∘ Ψ α . (44) F_h=\Psi_\alpha^{-1}\circ F_{\alpha,h}\circ\Psi_\alpha. \tag*{(44)} F h = Ψ α − 1 ∘ F α , h ∘ Ψ α . (44)
Theorem 4.3 guarantees that this definition is independent of the chosen local representation: on an overlap, T α β = Ψ β ∘ Ψ α − 1 T_{\alpha\beta}=\Psi_\beta\circ\Psi_\alpha^{-1} T α β = Ψ β ∘ Ψ α − 1 , and therefore
Ψ β − 1 ∘ F β , h ∘ Ψ β = Ψ α − 1 ∘ T α β ∘ F α , h ∘ T α β − 1 ∘ Ψ β = Ψ α − 1 ∘ F α , h ∘ Ψ α . (45) \Psi_\beta^{-1}\circ F_{\beta,h}\circ\Psi_\beta=\Psi_\alpha^{-1}\circ T_{\alpha\beta}\circ F_{\alpha,h}\circ T_{\alpha\beta}^{-1}\circ\Psi_\beta=\Psi_\alpha^{-1}\circ F_{\alpha,h}\circ\Psi_\alpha. \tag*{(45)} Ψ β − 1 ∘ F β , h ∘ Ψ β = Ψ α − 1 ∘ T α β ∘ F α , h ∘ T α β − 1 ∘ Ψ β = Ψ α − 1 ∘ F α , h ∘ Ψ α . (45)
Hence the reconstructed local expressions glue to a well-defined map F h : M → M F_h:M\to M F h : M → M on the domain reached by the numerical evolution, characterized on each chart by
Ψ α ∘ F h = F α , h ∘ Ψ α . (46) \Psi_\alpha\circ F_h=F_{\alpha,h}\circ\Psi_\alpha. \tag*{(46)} Ψ α ∘ F h = F α , h ∘ Ψ α . (46)
This identity will be used immediately to transfer local symplecticity to the conformal preservation law for the global LCS form.
The construction of F h F_h F h also admits a symplectic covering interpretation. The following remark explains this connection and the role of our chart-wise Type II formulation.
M ~ \widetilde{M} M . What the Type II formulation adds is a verifiable condition for this equivariance at the level of the discrete generating functions, including the endpoint term under affine transitions, and a computation that avoids the covering, where the canonical variables involve factors e ∓ σ ~ e^{\mp\widetilde{\sigma}} e ∓ σ that are unbounded in the number of windings.
Global discrete LCS preservation
We now establish the structure-preservation property of the global discrete map. Under the cotangent-type hypothesis (13) , the computation (15) shows that the representation map Ψ α \Psi_\alpha Ψ α satisfies
Ψ α ∗ Ω α = e − σ α Ω , (49) \Psi_\alpha^*\Omega_\alpha=e^{-\sigma_\alpha}\Omega, \tag*{(49)} Ψ α ∗ Ω α = e − σ α Ω , (49)
where the left-hand side is expressed in the LCS variables and Ω α \Omega_\alpha Ω α in the local canonical variables.
Theorem 4.7 (Global LCS preservation). Let { F α , h } \{F_{\alpha,h}\} { F α , h } be a family of local one-step maps, each symplectic with respect to Ω α \Omega_\alpha Ω α and equivariant on overlaps in the sense of (43) , for instance the maps generated by an LCS-compatible family of regular discrete right Hamiltonians, and let F h F_h F h denote their global reconstruction. Then
F h ∗ Ω = e σ α ∘ F h − σ α Ω (50) F_h^*\Omega=e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega \tag*{(50)} F h ∗ Ω = e σ α ∘ F h − σ α Ω (50) on every chart on which the one-step map is represented.
Proof. By (49) and (46) ,
e − σ α ∘ F h F h ∗ Ω = F h ∗ ( e − σ α Ω ) = ( Ψ α ∘ F h ) ∗ Ω α = Ψ α ∗ F α , h ∗ Ω α = Ψ α ∗ Ω α = e − σ α Ω , e^{-\sigma_\alpha\circ F_h}F_h^*\Omega=F_h^*(e^{-\sigma_\alpha}\Omega)=(\Psi_\alpha\circ F_h)^*\Omega_\alpha=\Psi_\alpha^*F_{\alpha,h}^*\Omega_\alpha=\Psi_\alpha^*\Omega_\alpha=e^{-\sigma_\alpha}\Omega, e − σ α ∘ F h F h ∗ Ω = F h ∗ ( e − σ α Ω ) = ( Ψ α ∘ F h ) ∗ Ω α = Ψ α ∗ F α , h ∗ Ω α = Ψ α ∗ Ω α = e − σ α Ω ,
where the fourth equality is the local symplecticity of F α , h F_{\alpha,h} F α , h . Hence F h ∗ Ω = e σ α ∘ F h − σ α Ω F_h^*\Omega=e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega F h ∗ Ω = e σ α ∘ F h − σ α Ω . ∎
This is the discrete counterpart of the continuous relation (26) : the numerical map preserves the defining LCS structure in the same conformal sense as the exact flow.
Independence of the local conformal potential. The exponent in (50) is independent of the choice of local potential on a connected component of an overlap: if σ α − σ β = C α β \sigma_\alpha-\sigma_\beta=C_{\alpha\beta} σ α − σ β = C α β on the component containing both z z z and F h ( z ) F_h(z) F h ( z ) , then σ α ( F h ( z ) ) − σ α ( z ) = σ β ( F h ( z ) ) − σ β ( z ) \sigma_\alpha(F_h(z))-\sigma_\alpha(z)=\sigma_\beta(F_h(z))-\sigma_\beta(z) σ α ( F h ( z )) − σ α ( z ) = σ β ( F h ( z )) − σ β ( z ) . Consequently,
ρ h ( z ) : = exp ( σ α ( F h ( z ) ) − σ α ( z ) ) (51) \rho_h(z):=\exp\left(\sigma_\alpha(F_h(z))-\sigma_\alpha(z)\right) \tag*{(51)} ρ h ( z ) := exp ( σ α ( F h ( z )) − σ α ( z ) ) (51)
defines a discrete conformal factor that is independent of the local representation, provided that the initial point z z z and the final point F h ( z ) F_h(z) F h ( z ) lie in the same connected component of the overlap of the representations being compared, and the preservation law may be written intrinsically as F h ∗ Ω = ρ h Ω F_h^*\Omega=\rho_h\Omega F h ∗ Ω = ρ h Ω . This proviso is not vacuous: a discrete map may jump between two connected components of U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β without leaving U α U_\alpha U α , in which case the two local potentials differ by different constants at z z z and at F h ( z ) F_h(z) F h ( z ) . The one-step law then holds with the potential of the representation in which the step is computed, which is the content of Theorem Theorem 4.7 .
Several time steps. The conformal preservation law has a multiplicative cocycle property under iteration: for two steps, ( F h 2 ) ∗ Ω = F h ∗ ( e σ α ∘ F h − σ α Ω ) = e σ α ∘ F h 2 − σ α ∘ F h e σ α ∘ F h − σ α Ω (F_h^2)^*\Omega=F_h^*(e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega)=e^{\sigma_\alpha\circ F_h^2-\sigma_\alpha\circ F_h}e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega ( F h 2 ) ∗ Ω = F h ∗ ( e σ α ∘ F h − σ α Ω ) = e σ α ∘ F h 2 − σ α ∘ F h e σ α ∘ F h − σ α Ω , and inductively
( F h N ) ∗ Ω = e σ α ∘ F h N − σ α Ω (52) (F_h^N)^*\Omega=e^{\sigma_\alpha\circ F_h^N-\sigma_\alpha}\Omega \tag*{(52)} ( F h N ) ∗ Ω = e σ α ∘ F h N − σ α Ω (52)
whenever the trajectory remains in a common representation.
More generally, when chart switching is required, composing the one-step laws, each applied in the local representation in which the corresponding step is computed, gives
( F h N ) ∗ Ω = exp ( ∑ k = 0 N − 1 [ σ γ k ( z k + 1 ) − σ γ k ( z k ) ] ) Ω , (53) (F_h^N)^*\Omega=\exp\left(\sum_{k=0}^{N-1}\left[\sigma_{\gamma_k}(z_{k+1})-\sigma_{\gamma_k}(z_k)\right]\right)\Omega, \tag*{(53)} ( F h N ) ∗ Ω = exp ( k = 0 ∑ N − 1 [ σ γ k ( z k + 1 ) − σ γ k ( z k ) ] ) Ω , (53)
where z k = F h k ( z ) z_k=F_h^k(z) z k = F h k ( z ) and γ k \gamma_k γ k denotes the local representation used for the k k k -th step. The exponent is a discrete line integral of the Lee form along the numerical trajectory, and it is the discrete counterpart of ∫ 0 t φ ( ξ H ) ( ψ s ( x ) ) d s \int_0^t\varphi(\xi_H)(\psi_s(x))\,\mathrm{d}s ∫ 0 t φ ( ξ H ) ( ψ s ( x )) d s in (25) . If the representation is changed from α \alpha α to β \beta β at z m z_m z m and back to α \alpha α at z m ′ z_{m'} z m ′ , the exponent equals
σ α ( z N ) − σ α ( z 0 ) + C α β ( z m ) − C α β ( z m ′ ) , \sigma_\alpha(z_N)-\sigma_\alpha(z_0)+C_{\alpha\beta}(z_m)-C_{\alpha\beta}(z_{m'}), σ α ( z N ) − σ α ( z 0 ) + C α β ( z m ) − C α β ( z m ′ ) ,
where C α β ( z ) C_{\alpha\beta}(z) C α β ( z ) denotes the value of the constant σ α − σ β \sigma_\alpha-\sigma_\beta σ α − σ β on the connected component of U α ∩ U β U_\alpha\cap U_\beta U α ∩ U β containing z z z . Thus (52) remains valid under chart switching precisely when the switching points lie in the same connected component of the overlap. When they do not, the additive constants do not cancel, and the exponent in (53) differs from σ α ( z N ) − σ α ( z 0 ) \sigma_\alpha(z_N)-\sigma_\alpha(z_0) σ α ( z N ) − σ α ( z 0 ) by a period of the Lee form. This is not a deficiency of the discretization but a faithful reproduction of the continuous law (25) , whose exponent along a closed trajectory is ∮ φ \oint\varphi ∮ φ ; the example with a non-exact Lee form in Section § 8 exhibits this monodromy explicitly.
Two logically distinct properties therefore enter the global preservation theorem: local structure preservation, F α , h ∗ Ω α = Ω α F_{\alpha,h}^*\Omega_\alpha=\Omega_\alpha F α , h ∗ Ω α = Ω α , and overlap compatibility, F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h . Neither implies the other: two independently chosen symplectic integrators may each preserve their respective local symplectic forms while failing to describe the same discrete dynamics on an overlap, and equivariance alone does not imply symplecticity. The numerical experiments of Section § 8 measure the two properties separately.
Order under globalization. Compatibility does not alter the local approximation order.
Proposition 4.8. Suppose that, for each α \alpha α , the local discrete right Hamiltonian H α , h d + H_{\alpha,h}^{d+} H α , h d + approximates the corresponding exact discrete right Hamiltonian H α , h + , E H_{\alpha,h}^{+,E} H α , h + , E to order p + 1 p+1 p + 1 in the C 1 C^1 C 1 sense, and suppose that the family { H α , h d + } \{H_{\alpha,h}^{d+}\} { H α , h d + } is LCS-compatible. Then the reconstructed global map F h F_h F h has order p p p with respect to the exact LCS Hamiltonian flow.
Proof. By Theorem 3.7 , each local map F α , h F_{\alpha,h} F α , h has order p p p with respect to the exact local flow Φ α , h \Phi_{\alpha,h} Φ α , h . Since F h = Ψ α − 1 ∘ F α , h ∘ Ψ α F_h=\Psi_\alpha^{-1}\circ F_{\alpha,h}\circ\Psi_\alpha F h = Ψ α − 1 ∘ F α , h ∘ Ψ α and ψ h = Ψ α − 1 ∘ Φ α , h ∘ Ψ α \psi_h=\Psi_\alpha^{-1}\circ\Phi_{\alpha,h}\circ\Psi_\alpha ψ h = Ψ α − 1 ∘ Φ α , h ∘ Ψ α with Ψ α \Psi_\alpha Ψ α smooth and independent of h h h , conjugation preserves the local order, F h − ψ h = O ( h p + 1 ) F_h-\psi_h=O(h^{p+1}) F h − ψ h = O ( h p + 1 ) in local coordinates, and compatibility ensures that these local representations define the same global map on overlaps. □
A practical sufficient-condition principle. The preceding results may be summarized as a construction principle.
Theorem 4.9 (Compatible Hamiltonian discretization principle). Consider a cotangent-type LCS Hamiltonian system admitting local canonical representations related on overlaps by
T α β ( q , r ) = ( q , c α β r ) . T_{\alpha\beta}(q,r)=(q,c_{\alpha\beta}r). T α β ( q , r ) = ( q , c α β r ) . Suppose that, on every chart, a regular discrete right Hamiltonian H α , h d + H_{\alpha,h}^{d+} H α , h d + is constructed such that, on every overlap,
H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) . H_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}H_{\alpha,h}^{d+}\left(q,c_{\alpha\beta}^{-1}s\right). H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) . Then the local maps are equivariant with respect to the transition maps, glue to a well-defined discrete dynamics F h F_h F h , and the resulting global map satisfies
F h ∗ Ω = e σ α ∘ F h − σ α Ω . (54) F_h^*\Omega=e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega. \tag*{(54)} F h ∗ Ω = e σ α ∘ F h − σ α Ω. (54) If, in addition, H α , h d + H_{\alpha,h}^{d+} H α , h d + approximates H α , h + , E H_{\alpha,h}^{+,E} H α , h + , E to order p + 1 p+1 p + 1 in the C 1 C^1 C 1 sense, then F h F_h F h is an order-p p p approximation of the exact LCS Hamiltonian flow.
Proof. The maps generated by the regular discrete right Hamiltonians are locally symplectic by Proposition 3.6 . The conclusions then follow by combining Theorems 4.3 and 4.7 , Lemma 4.4 , and Proposition 4.8 . □
Theorem 4.9 reduces the globalization problem to a concrete question at the level of the numerical construction: how can one build approximate discrete Hamiltonians that inherit the exact overlap transformation law? The next section addresses this question for important classes of geometric integrators.
Classes of compatible structure-preserving integrators
The results of § 4 reduce the construction of global LCS integrators to two requirements. On each local canonical representation, the numerical method must preserve the local symplectic form, and the corresponding local maps must be equivariant with respect to the overlap transformations T α β T_{\alpha\beta} T α β of (16) . In this section, we discuss classes of numerical methods for which these requirements can be verified directly. We first consider Hamiltonian variational constructions based on approximations of the Type II action. We then formulate a map-level criterion that applies to symplectic Runge–Kutta methods: since the overlap transformations are linear with constant coefficients, Runge–Kutta equivariance follows from the corresponding equivariance of the local Hamiltonian vector fields. Runge–Kutta-type methods and their canonical and variational properties have been studied in [41 , 42 ] . This linearity (or, more generally, affinity, see Section 5.4) is essential; the statements below are not intended to assert equivariance of Runge–Kutta methods under arbitrary nonlinear changes of local coordinates.
Compatible Type II Hamiltonian variational constructions
A practical discrete Hamiltonian can be obtained by replacing the exact trajectory and the exact action integral in (31) by finite-dimensional approximations. A general construction proceeds by choosing (i) a finite-dimensional approximation space for the position curve; (ii) a corresponding approximation of the momentum curve; (iii) a quadrature rule for the Hamiltonian phase-space action; (iv) a discrete stationarity condition determining the internal variables. Following standard usage, s s s denotes the number of stages; it should not be confused with the momentum variable of the β \beta β -representation. Let 0 < τ 1 , … , τ s ≤ 1 0 < \tau_1,\ldots,\tau_s \le1 0 < τ 1 , … , τ s ≤ 1 be quadrature nodes with weights b 1 , … , b s b_1,\ldots,b_s b 1 , … , b s , and denote the approximate position and momentum values at the quadrature nodes by Q i ≈ q ( τ i h ) Q_i \approx q(\tau_i h) Q i ≈ q ( τ i h ) and R i ≈ r ( τ i h ) R_i \approx r(\tau_i h) R i ≈ r ( τ i h ) . A generic approximation of the Type II action has the form
A α , h d = r 1 ⋅ q 1 − h ∑ i = 1 s b i [ R i ⋅ V i − H α ( Q i , R i ) ] , (55) \mathcal{A}^{d}_{\alpha,h}=r_1\cdot q_1-h\sum_{i=1}^{s}b_i\left[R_i\cdot V_i-H_\alpha(Q_i,R_i)\right], \tag*{(55)} A α , h d = r 1 ⋅ q 1 − h i = 1 ∑ s b i [ R i ⋅ V i − H α ( Q i , R i ) ] , (55)
where V i V_i V i approximates q ˙ ( τ i h ) \dot{q}(\tau_i h) q ˙ ( τ i h ) . The associated discrete right Hamiltonian is obtained by imposing the chosen finite-dimensional stationarity conditions and eliminating the internal variables, H α , h d + ( q 0 , r 1 ) = ext A α , h d H^{d+}_{\alpha,h}(q_0,r_1)=\operatorname{ext}\mathcal{A}^{d}_{\alpha,h} H α , h d + ( q 0 , r 1 ) = ext A α , h d , where ext \operatorname{ext} ext denotes evaluation at a stationary point of the internal discrete variables, rather than necessarily a minimum. The overlap scaling acts on momentum-type quantities by r ↦ s = c α β r r\mapsto s=c_{\alpha\beta}r r ↦ s = c α β r , while the position variables are unchanged, and the continuous Hamiltonian transforms according to (17) . Consequently, if the discrete approximation space and the stationarity procedure are preserved by the constant scaling of all momentum-type variables, the approximate action inherits the same homogeneity.
Proposition 5.1. Suppose that the discrete Type II construction in each local representation uses the same position approximation space, the same quadrature rule, and momentum variables that transform according to S i = c α β R i S_i=c_{\alpha\beta}R_i S i = c α β R i and s 1 = c α β r 1 s_1=c_{\alpha\beta}r_1 s 1 = c α β r 1 , and that the continuous Hamiltonians satisfy (17) . Assume also that the stationary values defining the discrete Hamiltonians are taken on corresponding regular branches related by the same momentum scaling. Then
H β , h d + ( q 0 , s 1 ) = c α β H α , h d + ( q 0 , c α β − 1 s 1 ) . (56) H^{d+}_{\beta,h}(q_0,s_1)=c_{\alpha\beta}H^{d+}_{\alpha,h}\left(q_0,c_{\alpha\beta}^{-1}s_1\right). \tag*{(56)} H β , h d + ( q 0 , s 1 ) = c α β H α , h d + ( q 0 , c α β − 1 s 1 ) . (56) Hence the resulting discrete Hamiltonian maps are LCS-compatible.
Proof. Since the position variables and their discrete velocities are unchanged, S i ⋅ V i = c α β R i ⋅ V i S_i\cdot V_i=c_{\alpha\beta}R_i\cdot V_i S i ⋅ V i = c α β R i ⋅ V i and H β ( Q i , S i ) = c α β H α ( Q i , R i ) H_\beta(Q_i,S_i)=c_{\alpha\beta}H_\alpha(Q_i,R_i) H β ( Q i , S i ) = c α β H α ( Q i , R i ) , so every term in the discrete phase-space action scales by the same constant;
A β , h d = c α β A α , h d . (57) \mathcal{A}^{d}_{\beta,h}=c_{\alpha\beta}\mathcal{A}^{d}_{\alpha,h}. \tag*{(57)} A β , h d = c α β A α , h d . (57)
Since (57) holds identically in the internal variables, ( Q ∗ , c α β R ∗ ) (Q^*,c_{\alpha\beta}R^*) ( Q ∗ , c α β R ∗ ) is a stationary point of A β , h d \mathcal{A}^{d}_{\beta,h} A β , h d if and only if ( Q ∗ , R ∗ ) (Q^*,R^*) ( Q ∗ , R ∗ ) is a stationary point of A α , h d \mathcal{A}^{d}_{\alpha,h} A α , h d , so that stationary internal variables correspond under the same momentum scaling. Evaluating the two actions at corresponding stationary points gives H β , h d + ( q 0 , c α β r 1 ) = c α β H α , h d + ( q 0 , r 1 ) H^{d+}_{\beta,h}(q_0,c_{\alpha\beta}r_1)=c_{\alpha\beta}H^{d+}_{\alpha,h}(q_0,r_1) H β , h d + ( q 0 , c α β r 1 ) = c α β H α , h d + ( q 0 , r 1 ) , which is equivalent to (56) . □ \square □
By Theorem 4.9, any regular construction satisfying Proposition 5.1 produces a globally defined LCS structure-preserving integrator.
Equivariance of Runge–Kutta discretizations
A second important class of compatible methods is provided by Runge–Kutta discretizations of the local Hamiltonian vector fields. Let z ˙ = f α ( z ) \dot{z}=f_{\alpha}(z) z ˙ = f α ( z ) , z = ( q , r ) z=(q,r) z = ( q , r ) , denote the local Hamiltonian system in the α \alpha α -representation and w ˙ = f β ( w ) \dot{w}=f_{\beta}(w) w ˙ = f β ( w ) , w = T α β z w=T_{\alpha\beta}z w = T α β z , the corresponding system in the β \beta β -representation. By (19) ,
f β ( T α β z ) = T α β f α ( z ) , (58) f_{\beta}(T_{\alpha\beta}z)=T_{\alpha\beta}f_{\alpha}(z), \tag*{(58)} f β ( T α β z ) = T α β f α ( z ) , (58)
where, since T α β T_{\alpha\beta} T α β is linear, the same notation denotes the map and its tangent map. Consider an s s s -stage Runge–Kutta method with coefficients ( a i j , b i ) (a_{ij},b_i) ( a ij , b i ) , whose stages and update in the α \alpha α -representation are
Z i = z n + h ∑ j = 1 s a i j f α ( Z j ) , z n + 1 = z n + h ∑ i = 1 s b i f α ( Z i ) , (59) Z_i=z_n+h\sum_{j=1}^{s}a_{ij}f_{\alpha}(Z_j),\qquad z_{n+1}=z_n+h\sum_{i=1}^{s}b_i f_{\alpha}(Z_i), \tag*{(59)} Z i = z n + h j = 1 ∑ s a ij f α ( Z j ) , z n + 1 = z n + h i = 1 ∑ s b i f α ( Z i ) , (59)
and similarly in the β \beta β -representation with f β f_{\beta} f β .
Proposition 5.2 (Runge–Kutta equivariance). Suppose that the local vector fields satisfy f β ∘ T α β = T α β ∘ f α f_{\beta}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ f_{\alpha} f β ∘ T α β = T α β ∘ f α , where T α β T_{\alpha\beta} T α β is linear, and suppose that the same Runge–Kutta tableau is used in both local representations. Whenever the stage equations determine corresponding local solutions, the Runge–Kutta maps satisfy
F β , h R K ∘ T α β = T α β ∘ F α , h R K . (60) F_{\beta,h}^{RK}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h}^{RK}. \tag*{(60)} F β , h R K ∘ T α β = T α β ∘ F α , h R K . (60)
Proof. Applying the linear map T α β T_{\alpha\beta} T α β to the stage and update equations (59) and using T α β f α ( Z j ) = f β ( T α β Z j ) T_{\alpha\beta}f_{\alpha}(Z_j)=f_{\beta}(T_{\alpha\beta}Z_j) T α β f α ( Z j ) = f β ( T α β Z j ) shows that W i = T α β Z i W_i=T_{\alpha\beta}Z_i W i = T α β Z i and w n + 1 = T α β z n + 1 w_{n+1}=T_{\alpha\beta}z_{n+1} w n + 1 = T α β z n + 1 satisfy the stage and update equations of the same method for f β f_{\beta} f β with initial value w n = T α β z n w_n=T_{\alpha\beta}z_n w n = T α β z n ; by uniqueness of the stage solution, F β , h R K ( T α β z n ) = T α β ( F α , h R K ( z n ) ) F_{\beta,h}^{RK}(T_{\alpha\beta}z_n)=T_{\alpha\beta}(F_{\alpha,h}^{RK}(z_n)) F β , h R K ( T α β z n ) = T α β ( F α , h R K ( z n )) .
Remark 5.3. The linearity assumption in Proposition 5.2 cannot be dropped in favour of general nonlinear transitions. Standard Runge–Kutta methods are affine equivariant, but they are not generally equivariant under arbitrary nonlinear coordinate transformations. The cotangent-type overlap maps considered here are constant linear fiber scalings, and hence lie within the class for which the preceding argument applies directly; the affine case is treated in Proposition 5.15.
Symplectic Runge–Kutta methods. Equivariance provides the overlap compatibility required for globalization. To obtain an LCS structure-preserving method, the local Runge–Kutta discretization must additionally be symplectic. For a canonical Hamiltonian system, an s s s -stage Runge–Kutta method is symplectic if its coefficients satisfy
b i a i j + b j a j i = b i b j , i , j = 1 , … , s , (61) b_i a_{ij}+b_j a_{ji}=b_i b_j,\qquad i,j=1,\ldots,s, \tag*{(61)} b i a ij + b j a j i = b i b j , i , j = 1 , … , s , (61)
in which case ( F α , h R K ) ∗ Ω α = Ω α (F_{\alpha,h}^{RK})^{*}\Omega_{\alpha}=\Omega_{\alpha} ( F α , h R K ) ∗ Ω α = Ω α [22 ] . Combining this fact with Proposition 5.2 immediately gives the following consequence.
Theorem 5.4 (Global LCS integrators from symplectic Runge–Kutta methods). Consider the cotangent-type LCS setting of § 2 , with overlap transformations T α β ( q , r ) = ( q , c α β r ) T_{\alpha\beta}(q,r)=(q,c_{\alpha\beta}r) T α β ( q , r ) = ( q , c α β r ) , and apply the same symplectic Runge–Kutta method to every local canonical Hamiltonian system. Then the resulting local maps are equivariant on overlaps and glue to a global discrete map F h F_h F h satisfying
F h ∗ Ω = e σ ∘ F h − σ Ω . (62) F_h^{*}\Omega=e^{\sigma\circ F_h-\sigma}\Omega. \tag*{(62)} F h ∗ Ω = e σ ∘ F h − σ Ω. (62)
Proof. The local Hamiltonian vector fields are equivariant with respect to the linear maps T α β T_{\alpha\beta} T α β by (58) , so Proposition 5.2 gives equivariance of the local Runge–Kutta maps, and (61) gives local symplecticity. The conclusion follows from Lemma 4.4 and Theorem 4.7.
Collocation and Gauss–Legendre methods
The family of methods that realizes Theorem 5.4 in its sharpest form is the Gauss–Legendre collocation family [22 , 38 ] . Let s ≥ 1 s \ge1 s ≥ 1 and let 0 < τ 1 < ⋯ < τ s < 1 0 < \tau_1 < \cdots< \tau_s < 1 0 < τ 1 < ⋯ < τ s < 1 be the roots of the shifted Legendre polynomial P s ( λ ) = 1 s ! d s d λ s [ λ s ( λ − 1 ) s ] P_s(\lambda) = \frac{1}{s!}\frac{\mathrm{d}^s}{\mathrm{d}\lambda^s}[\lambda^s(\lambda- 1)^s] P s ( λ ) = s ! 1 d λ s d s [ λ s ( λ − 1 ) s ] on [ 0 , 1 ] [0,1] [ 0 , 1 ] . With ℓ 1 , … , ℓ s \ell_1,\ldots,\ell_s ℓ 1 , … , ℓ s the Lagrange interpolation polynomials associated with these nodes, the coefficients
a i j = ∫ 0 τ i ℓ j ( λ ) d λ , b i = ∫ 0 1 ℓ i ( λ ) d λ , (63) a_{ij} = \int_{0}^{\tau_i} \ell_j(\lambda)\,\mathrm{d}\lambda,\qquad b_i = \int_{0}^{1} \ell_i(\lambda)\,\mathrm{d}\lambda, \tag*{(63)} a ij = ∫ 0 τ i ℓ j ( λ ) d λ , b i = ∫ 0 1 ℓ i ( λ ) d λ , (63)
define the s s s -stage Gauss–Legendre method. Equivalently, it is the collocation method that seeks a polynomial u u u of degree s s s with u ( 0 ) = z n u(0) = z_n u ( 0 ) = z n and u ˙ ( τ i h ) = f α ( u ( τ i h ) ) \dot{u}(\tau_i h) = f_\alpha(u(\tau_i h)) u ˙ ( τ i h ) = f α ( u ( τ i h )) , i = 1 , … , s i = 1,\ldots,s i = 1 , … , s , and sets z n + 1 = u ( h ) z_{n+1} = u(h) z n + 1 = u ( h ) . The first three members are
1 2 1 2 1 \begin{array}{c|c}
\frac{1}{2} & \frac{1}{2}\\
\hline
& 1
\end{array} 2 1 2 1 1
1 2 − 3 6 1 4 1 4 − 3 6 1 2 + 3 6 1 4 + 3 6 1 4 1 2 1 2 \begin{array}{c|cc}
\frac{1}{2}-\frac{\sqrt{3}}{6} & \frac{1}{4} & \frac{1}{4}-\frac{\sqrt{3}}{6}\\
\frac{1}{2}+\frac{\sqrt{3}}{6} & \frac{1}{4}+\frac{\sqrt{3}}{6} & \frac{1}{4}\\
\hline
& \frac{1}{2} & \frac{1}{2}
\end{array} 2 1 − 6 3 2 1 + 6 3 4 1 4 1 + 6 3 2 1 4 1 − 6 3 4 1 2 1
1 2 − 15 10 5 36 2 9 − 15 15 5 36 − 15 30 1 2 5 36 + 15 24 2 9 5 36 − 15 24 1 2 + 15 10 5 36 + 15 30 2 9 + 15 15 5 36 5 18 4 9 5 18 \begin{array}{c|ccc}
\frac{1}{2}-\frac{\sqrt{15}}{10} & \frac{5}{36} & \frac{2}{9}-\frac{\sqrt{15}}{15} & \frac{5}{36}-\frac{\sqrt{15}}{30}\\
\frac{1}{2} & \frac{5}{36}+\frac{\sqrt{15}}{24} & \frac{2}{9} & \frac{5}{36}-\frac{\sqrt{15}}{24}\\
\frac{1}{2}+\frac{\sqrt{15}}{10} & \frac{5}{36}+\frac{\sqrt{15}}{30} & \frac{2}{9}+\frac{\sqrt{15}}{15} & \frac{5}{36}\\
\hline
& \frac{5}{18} & \frac{4}{9} & \frac{5}{18}
\end{array} 2 1 − 10 15 2 1 2 1 + 10 15 36 5 36 5 + 24 15 36 5 + 30 15 18 5 9 2 − 15 15 9 2 9 2 + 15 15 9 4 36 5 − 30 15 36 5 − 24 15 36 5 18 5
corresponding to the implicit midpoint rule (s = 1 s = 1 s = 1 ) and to the fourth- and sixth-order Gauss methods. We collect the classical facts that will be used below.
Proposition 5.5. The s s s -stage Gauss–Legendre method has order 2 s 2s 2 s , its underlying quadrature rule being exact for polynomials of degree ≤ 2 s − 1 \le2s - 1 ≤ 2 s − 1 and the collocation polynomial superconverging at the endpoint. It is symmetric, A-stable and algebraically stable, with stability function the ( s , s ) (s,s) ( s , s ) diagonal Padé approximation of the exponential, and it preserves exactly all quadratic first integrals of the differential equation. Its coefficients satisfy
b i a i j + b j a j i − b i b j = 0 , i , j = 1 , … , s , (64) b_i a_{ij} + b_j a_{ji} - b_i b_j = 0,\qquad i,j = 1,\ldots,s, \tag*{(64)} b i a ij + b j a j i − b i b j = 0 , i , j = 1 , … , s , (64) so that (61) holds and the method is symplectic for canonical Hamiltonian systems. Finally, if f α f_\alpha f α is smooth, then for each initial point z ∗ z_\ast z ∗ the stage equations admit a unique local solution branch near Z 1 = ⋯ = Z s = z ∗ Z_1 = \cdots= Z_s = z_\ast Z 1 = ⋯ = Z s = z ∗ for ( z n , h ) (z_n,h) ( z n , h ) sufficiently close to ( z ∗ , 0 ) (z_\ast,0) ( z ∗ , 0 ) . This branch depends smoothly on z n z_n z n and h h h . To see this, write the stage equations as
G i ( Z , z n , h ) = Z i − z n − h ∑ j = 1 s a i j f α ( Z j ) = 0 , G_i(Z,z_n,h) = Z_i - z_n - h \sum_{j=1}^{s} a_{ij}f_\alpha(Z_j) = 0, G i ( Z , z n , h ) = Z i − z n − h j = 1 ∑ s a ij f α ( Z j ) = 0 , where Z = ( Z 1 , … , Z s ) Z = (Z_1,\ldots,Z_s) Z = ( Z 1 , … , Z s ) and G = ( G 1 , … , G s ) G = (G_1,\ldots,G_s) G = ( G 1 , … , G s ) . At h = 0 h = 0 h = 0 , these equations have the solution Z i = z n Z_i = z_n Z i = z n for all i i i , and the Jacobian D Z G D_ZG D Z G at this solution is the identity. The claim then follows from the implicit function theorem.
This local solvability result ensures that the one-step maps F α , h F_{\alpha,h} F α , h are well defined and smooth on the selected stage branches for sufficiently small h h h . By Proposition 5.2 , stage solutions are carried bijectively to stage solutions in corresponding overlap neighborhoods. Consequently, local existence and uniqueness transfer between representations at the same step size, provided the stages remain in these neighborhoods. The same conclusion holds for affine transitions by Proposition 5.15.
Gauss–Legendre as a Type II Galerkin construction
Gauss–Legendre methods occupy a distinguished position in the present framework because they belong simultaneously to both classes of compatible methods identified in this section: they arise from a Type II Galerkin approximation of the Hamiltonian phase-space action, and they are symplectic Runge–Kutta methods.
Fix a local representation and consider the Type II action (55) . Choose
as position approximation space the polynomials q d q_d q d of degree ≤ s \le s ≤ s on [ 0 , h ] [0,h] [ 0 , h ] with q d ( 0 ) = q 0 q_d(0)=q_0 q d ( 0 ) = q 0 ;
as momentum approximation the values R i R_i R i at the quadrature nodes, i.e., an unconstrained polynomial r d r_d r d of degree ≤ s − 1 \le s-1 ≤ s − 1 , the final momentum r 1 r_1 r 1 entering only through the boundary term;
the s s s -point Gauss–Legendre quadrature rule with nodes and weights (63) .
Writing Q i = q d ( τ i h ) Q_i=q_d(\tau_i h) Q i = q d ( τ i h ) , R i = r d ( τ i h ) R_i=r_d(\tau_i h) R i = r d ( τ i h ) and V i = q ˙ d ( τ i h ) V_i=\dot{q}_d(\tau_i h) V i = q ˙ d ( τ i h ) , the discrete right Hamiltonian is
H α , h d + ( q 0 , r 1 ) = ext { r 1 ⋅ q 1 − h ∑ i = 1 s b i [ R i ⋅ V i − H α ( Q i , R i ) ] } , (65) H_{\alpha,h}^{d+}(q_0,r_1)=\operatorname{ext}\left\{r_1\cdot q_1-h\sum_{i=1}^{s}b_i\left[R_i\cdot V_i-H_\alpha(Q_i,R_i)\right]\right\}, \tag*{(65)} H α , h d + ( q 0 , r 1 ) = ext { r 1 ⋅ q 1 − h i = 1 ∑ s b i [ R i ⋅ V i − H α ( Q i , R i ) ] } , (65)
the extremum being taken over the internal variables of q d q_d q d , namely its nodal velocities V j V_j V j , and over the nodal momenta R i R_i R i , with q 1 = q d ( h ) q_1=q_d(h) q 1 = q d ( h ) .
Proposition 5.6. The discrete right Hamiltonian (65) generates, through the right discrete Hamilton equations (33) , the s s s -stage Gauss–Legendre method applied to the local canonical Hamiltonian system associated with H α H_\alpha H α . Consequently, the induced local map has order 2 s 2s 2 s , and (65) approximates the exact discrete right Hamiltonian H α , h + , E H_{\alpha,h}^{+,\mathrm{E}} H α , h + , E to order 2 s + 1 2s+1 2 s + 1 in the C 1 C^1 C 1 sense, up to an additive constant.
Proof. We spell out the correspondence because it is the point at which the Type II variational construction and the Runge–Kutta description meet. Let
Q i = q d ( τ i h ) , R i = r d ( τ i h ) , i = 1 , … , s , Q_i=q_d(\tau_i h),\qquad R_i=r_d(\tau_i h),\qquad i=1,\ldots,s, Q i = q d ( τ i h ) , R i = r d ( τ i h ) , i = 1 , … , s ,
and write the canonical Hamiltonian vector field as
q ˙ = D 2 H α ( q , r ) , r ˙ = − D 1 H α ( q , r ) . \dot{q}=D_2H_\alpha(q,r),\qquad\dot{r}=-D_1H_\alpha(q,r). q ˙ = D 2 H α ( q , r ) , r ˙ = − D 1 H α ( q , r ) .
We use the nodal velocities V j = q ˙ d ( τ j h ) V_j=\dot{q}_d(\tau_j h) V j = q ˙ d ( τ j h ) , j = 1 , … , s j=1,\ldots,s j = 1 , … , s , as internal variables parametrizing q d q_d q d with fixed initial position q 0 q_0 q 0 . Stationarity of (65) with respect to the nodal momenta R i R_i R i gives V i = D 2 H α ( Q i , R i ) V_i=D_2H_\alpha(Q_i,R_i) V i = D 2 H α ( Q i , R i ) , and stationarity with respect to the nodal velocities V j V_j V j gives R j = r 1 + h ∑ i = 1 s ( b i a i j / b j ) D 1 H α ( Q i , R i ) R_j=r_1+h\sum_{i=1}^{s}(b_i a_{ij}/b_j)D_1H_\alpha(Q_i,R_i) R j = r 1 + h ∑ i = 1 s ( b i a ij / b j ) D 1 H α ( Q i , R i ) ; together with r 0 = D 1 H α , h d + ( q 0 , r 1 ) = r 1 + h ∑ i = 1 s b i D 1 H α ( Q i , R i ) r_0=D_1H_{\alpha,h}^{d+}(q_0,r_1)=r_1+h\sum_{i=1}^{s}b_iD_1H_\alpha(Q_i,R_i) r 0 = D 1 H α , h d + ( q 0 , r 1 ) = r 1 + h ∑ i = 1 s b i D 1 H α ( Q i , R i ) , this gives
R j = r 0 − h ∑ i = 1 s ( b i − b i a i j b j ) D 1 H α ( Q i , R i ) , R_j=r_0-h\sum_{i=1}^{s}\left(b_i-\frac{b_i a_{ij}}{b_j}\right)D_1H_\alpha(Q_i,R_i), R j = r 0 − h i = 1 ∑ s ( b i − b j b i a ij ) D 1 H α ( Q i , R i ) ,
and these relations show, as in [27 ] , that the stage values satisfy a symplectic partitioned Runge–Kutta system of the form
Q i = q 0 + h ∑ j = 1 s a i j D 2 H α ( Q j , R j ) , (66) Q_i=q_0+h\sum_{j=1}^{s}a_{ij}D_2H_\alpha(Q_j,R_j), \tag*{(66)} Q i = q 0 + h j = 1 ∑ s a ij D 2 H α ( Q j , R j ) , (66)
R i = r 0 − h ∑ j = 1 s a ~ i j D 1 H α ( Q j , R j ) , (67) R_i=r_0-h\sum_{j=1}^{s}\widetilde{a}_{ij}D_1H_\alpha(Q_j,R_j), \tag*{(67)} R i = r 0 − h j = 1 ∑ s a ij D 1 H α ( Q j , R j ) , (67)
where the coefficients conjugate to a i j a_{ij} a ij are
a ~ i j = b j − b j b i a j i . (68) \widetilde{a}_{ij}=b_j-\frac{b_j}{b_i}a_{ji}. \tag*{(68)} a ij = b j − b i b j a j i . (68)
The endpoint values are
q 1 = q 0 + h ∑ i = 1 s b i D 2 H α ( Q i , R i ) , (69) q_1=q_0+h\sum_{i=1}^{s}b_iD_2H_\alpha(Q_i,R_i), \tag*{(69)} q 1 = q 0 + h i = 1 ∑ s b i D 2 H α ( Q i , R i ) , (69)
r 1 = r 0 − h ∑ i = 1 s b i D 1 H α ( Q i , R i ) . (70) r_1=r_0-h\sum_{i=1}^{s}b_iD_1H_\alpha(Q_i,R_i). \tag*{(70)} r 1 = r 0 − h i = 1 ∑ s b i D 1 H α ( Q i , R i ) . (70)
For the Gauss–Legendre coefficients, (64) gives
b i a i j + b j a j i = b i b j , b_i a_{ij}+b_j a_{ji}=b_i b_j, b i a ij + b j a j i = b i b j ,
and therefore
a ~ i j = b j − b j b i a j i = a i j . \widetilde{a}_{ij}=b_j-\frac{b_j}{b_i}a_{ji}=a_{ij}. a ij = b j − b i b j a j i = a ij .
Thus the partitioned scheme collapses to the ordinary s s s -stage Gauss–Legendre Runge–Kutta method applied to the canonical Hamiltonian system. In particular, the map obtained from the Type II generating function is the same map as the Gauss collocation map considered in Proposition 5.5 .
For the order statement, the induced map is the s s s -stage Gauss collocation map, which has order 2 s 2s 2 s by Proposition 5.5 . By Remark 3.8, (65) therefore approximates H α , h + , E H_{\alpha,h}^{+,E} H α , h + , E to order 2 s + 1 2s+1 2 s + 1 in the C 1 C^1 C 1 sense, up to an additive constant, and Theorem 3.7 with p = 2 s p=2s p = 2 s returns the order 2 s 2s 2 s of the induced map.
Remark 5.7. Proposition 5.6 is the reason why the two routes developed above—the generating-function route and the Runge–Kutta equivariance route—are not merely parallel. For Gauss–Legendre methods the verification of the compatibility condition (41) can be carried out either at the level of the generating function, through Proposition 5.1 , or at the level of the vector field, through Proposition 5.2 , and the two verifications concern the same numerical method.
Compatibility of the Gauss–Legendre family
Proposition 5.8. Let the same s s s -stage Gauss–Legendre tableau be used in every local canonical representation. Then the family of generated discrete right Hamiltonians is LCS-compatible in the sense of Definition 4.2 , and the local stage variables transform as
Q i β = Q i α , S i = c α β R i . (71) Q_i^\beta=Q_i^\alpha,\qquad S_i=c_{\alpha\beta}R_i. \tag*{(71)} Q i β = Q i α , S i = c α β R i . (71)
Proof. The Gauss quadrature and the polynomial approximation spaces are identical in both representations, the position variables are unaffected by T α β T_{\alpha\beta} T α β and all momentum-type variables scale by c α β c_{\alpha\beta} c α β , so Proposition 5.1 applies and yields (41) . The stage scaling (71) is Proposition 5.2 applied to the stages (66) –(67) , which by Proposition 5.6 are those of the Gauss–Legendre method: setting s n = c α β r n s_n=c_{\alpha\beta}r_n s n = c α β r n and S i = c α β R i S_i=c_{\alpha\beta}R_i S i = c α β R i and using (18) , the position stage equations are unchanged and the momentum stage equations are multiplied by c α β c_{\alpha\beta} c α β .
Corollary 5.9. Under the hypotheses of Theorem 4.9 , the s s s -stage Gauss–Legendre method applied in every local canonical representation defines a global discrete map F h F_h F h on the LCS manifold satisfying
F h ∗ Ω = e σ α ∘ F h − σ α Ω , (72) F_h^*\Omega=e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega, \tag*{(72)} F h ∗ Ω = e σ α ∘ F h − σ α Ω , (72) and F h F_h F h is an approximation of order 2 s 2s 2 s of the LCS Hamiltonian flow. In particular the framework contains a hierarchy of arbitrarily high even-order LCS structure-preserving integrators.
The following remark explains why, on compact sets and for sufficiently small step sizes, each numerical step and its stages can be computed within a single chart.
Modified Hamiltonians and long-time behaviour
A further consequence of symplecticity is of direct relevance to the numerical experiments. In each local representation, backward error analysis [22 , 36 ] provides a formal modified Hamiltonian
H ~ α , h = H α + h 2 s H α [ 2 s ] + h 2 s + 2 H α [ 2 s + 2 ] + ⋯ . (73) \widetilde{H}_{\alpha,h}=H_\alpha+h^{2s}H_\alpha^{[2s]}+h^{2s+2}H_\alpha^{[2s+2]}+\cdots. \tag*{(73)} H α , h = H α + h 2 s H α [ 2 s ] + h 2 s + 2 H α [ 2 s + 2 ] + ⋯ . (73)
For analytic H α H_\alpha H α and sufficiently small h h h , assuming that the numerical orbit remains in a fixed compact subset of the chart with uniform analytic control and on a common regular stage branch, a modified Hamiltonian optimally truncated at an order proportional to h − 1 h^{-1} h − 1 has an exponentially small one-step conservation defect [22 Ch. IX] . Consequently, H α H_\alpha H α is conserved up to O ( h 2 s ) O(h^{2s}) O ( h 2 s ) over exponentially long times, as long as these assumptions remain valid.
Two remarks are relevant for the LCS setting.
uniformly in n n n over exponentially long time intervals. Equivalently, whenever H α ( z 0 ) ≠ 0 H_\alpha(z_0) \ne0 H α ( z 0 ) = 0 ,
H ( z n ) = e σ α ( z n ) − σ α ( z 0 ) H ( z 0 ) ( 1 + O ( h 2 s ) ) . H(z_n)=e^{\sigma_\alpha(z_n)-\sigma_\alpha(z_0)}H(z_0)\left(1+O(h^{2s})\right). H ( z n ) = e σ α ( z n ) − σ α ( z 0 ) H ( z 0 ) ( 1 + O ( h 2 s ) ) .
The uniform bound therefore holds for the local Hamiltonian, or in relative form for H H H , but not for the absolute error in H H H itself, which carries the factor e σ α ( z n ) e^{\sigma_\alpha(z_n)} e σ α ( z n ) and may grow along the trajectory. Under chart switching, the accumulated conformal factor is given by (53) ; for closed paths, it records the periods of the Lee form and hence its monodromy. The long-time energy estimate above is restricted to trajectories remaining in a fixed compact subset of a single chart. The weighted relation (75) is the local energy-type prediction tested numerically.
Remark 5.12. The modified Hamiltonians are themselves compatible. In each local representation, the term H α [ k ] H_\alpha^{[k]} H α [ k ] in (73) is a universal linear combination, with coefficients determined by the method alone and with no additive constants, of elementary Hamiltonians built from k + 1 k+1 k + 1 factors given by H α H_\alpha H α and its derivatives, contracted by means of k k k copies of the Poisson tensor Ω α − 1 \Omega_\alpha^{-1} Ω α − 1 . Under the transition map one has H β = c α β H α ∘ T α β − 1 H_\beta=c_{\alpha\beta}H_\alpha\circ T_{\alpha\beta}^{-1} H β = c α β H α ∘ T α β − 1 , while, by T α β ∗ Ω β = c α β Ω α T_{\alpha\beta}^{*}\Omega_\beta=c_{\alpha\beta}\Omega_\alpha T α β ∗ Ω β = c α β Ω α , the Poisson tensor of Ω β \Omega_\beta Ω β is the push-forward by T α β T_{\alpha\beta} T α β of c α β − 1 Ω α − 1 c_{\alpha\beta}^{-1}\Omega_\alpha^{-1} c α β − 1 Ω α − 1 . Each elementary Hamiltonian entering H β [ k ] H_\beta^{[k]} H β [ k ] is therefore c α β k + 1 c α β − k = c α β c_{\alpha\beta}^{k+1}c_{\alpha\beta}^{-k}=c_{\alpha\beta} c α β k + 1 c α β − k = c α β times the corresponding term of H α [ k ] H_\alpha^{[k]} H α [ k ] composed with T α β − 1 T_{\alpha\beta}^{-1} T α β − 1 , and one obtains
H ~ β , h ( q , s ) = c α β H ~ α , h ( q , c α β − 1 s ) . \widetilde{H}_{\beta,h}(q,s)=c_{\alpha\beta}\widetilde{H}_{\alpha,h}\left(q,c_{\alpha\beta}^{-1}s\right). H β , h ( q , s ) = c α β H α , h ( q , c α β − 1 s ) .
Consequently the local functions e σ α H ~ α , h e^{\sigma_\alpha}\widetilde{H}_{\alpha,h} e σ α H α , h agree on overlaps and define a global function H ~ h \widetilde{H}_h H h on M M M . Hence the modified equations also glue, and the numerical trajectory is exponentially close to the exact flow of the nearby globally defined LCS Hamiltonian system associated with H ~ h \widetilde{H}_h H h . This is the structural reason behind the absence of drift observed in § 8 .
The implicit midpoint rule. The simplest Gauss–Legendre method is the one-stage implicit midpoint rule, a second-order symplectic Runge–Kutta method. For the local canonical Hamiltonian system it reads
q n + 1 − q n h = D 2 H α ( q n + q n + 1 2 , r n + r n + 1 2 ) , r n + 1 − r n h = − D 1 H α ( q n + q n + 1 2 , r n + r n + 1 2 ) , (76) \frac{q_{n+1}-q_n}{h}=D_2H_\alpha\left(\frac{q_n+q_{n+1}}{2},\frac{r_n+r_{n+1}}{2}\right),\qquad
\frac{r_{n+1}-r_n}{h}=-D_1H_\alpha\left(\frac{q_n+q_{n+1}}{2},\frac{r_n+r_{n+1}}{2}\right), \tag*{(76)} h q n + 1 − q n = D 2 H α ( 2 q n + q n + 1 , 2 r n + r n + 1 ) , h r n + 1 − r n = − D 1 H α ( 2 q n + q n + 1 , 2 r n + r n + 1 ) , (76)
and its equivariance under T α β T_{\alpha\beta} T α β and the conformal law of its reconstruction are the case s = 1 s=1 s = 1 of Proposition Proposition 5.8 and Corollary Corollary 5.9 ; the general globalization principle thus recovers the implicit midpoint construction as a particular instance rather than as an isolated example. For the running Example 2.3 , with midpoint quantities q ˉ n = 1 2 ( q n + q n + 1 ) \bar{q}_n=\frac{1}{2}(q_n+q_{n+1}) q ˉ n = 2 1 ( q n + q n + 1 ) and similarly for r ˉ n \bar{r}_n r ˉ n , x ˉ n \bar{x}_n x ˉ n and s ˉ n \bar{s}_n s ˉ n , the scheme is
q n + 1 − q n = h e − x ˉ n r ˉ n , r n + 1 − r n = − h k e x ˉ n q ˉ n , x n + 1 − x n = h e − x ˉ n s ˉ n , s n + 1 − s n = h { 1 2 e − x ˉ n ( r ˉ n 2 + s ˉ n 2 ) − e x ˉ n [ k 2 q ˉ n 2 + μ ( x ˉ n + 1 ) ] } . (77) \begin{aligned}
q_{n+1}-q_n&=he^{-\bar{x}_n}\bar{r}_n,\\
r_{n+1}-r_n&=-hke^{\bar{x}_n}\bar{q}_n,\\
x_{n+1}-x_n&=he^{-\bar{x}_n}\bar{s}_n,\\
s_{n+1}-s_n&=h\left\{\frac{1}{2}e^{-\bar{x}_n}\left(\bar{r}_n^2+\bar{s}_n^2\right)-e^{\bar{x}_n}\left[\frac{k}{2}\bar{q}_n^2+\mu(\bar{x}_n+1)\right]\right\}. \tag*{(77)}
\end{aligned} q n + 1 − q n r n + 1 − r n x n + 1 − x n s n + 1 − s n = h e − x ˉ n r ˉ n , = − hk e x ˉ n q ˉ n , = h e − x ˉ n s ˉ n , = h { 2 1 e − x ˉ n ( r ˉ n 2 + s ˉ n 2 ) − e x ˉ n [ 2 k q ˉ n 2 + μ ( x ˉ n + 1 ) ] } . (77)
It preserves the local canonical symplectic form Ω l o c = d q ∧ d r + d x ∧ d s \Omega_{\mathrm{loc}}=\mathrm{d}q\wedge\mathrm{d}r+\mathrm{d}x\wedge\mathrm{d}s Ω loc = d q ∧ d r + d x ∧ d s , and since σ = − x \sigma=-x σ = − x and Ω = e − x Ω l o c \Omega=e^{-x}\Omega_{\mathrm{loc}} Ω = e − x Ω loc , the reconstructed one-step map satisfies ( F h M P ) ∗ Ω = e − ( x n + 1 − x n ) Ω (F_h^{\mathrm{MP}})^*\Omega=e^{-(x_{n+1}-x_n)}\Omega ( F h MP ) ∗ Ω = e − ( x n + 1 − x n ) Ω : the discrete conformal multiplier is determined exactly by the change of the local conformal potential across one numerical time step.
Extension to affine base transitions
The preceding results were formulated for overlap maps that preserve the base coordinates and rescale the momenta. We now extend the relevant transformation and compatibility results to affine base transitions. This extension includes the angular chart transitions used in Example 7.1.
Proposition 5.13 (Exact discrete right Hamiltonians under affine transitions). Consider an affine cotangent transition
T α β ( q , r ) = ( B q + b , c α β B − T r ) , (78) T_{\alpha\beta}(q,r) = \left(Bq+b,c_{\alpha\beta}B^{-T}r\right), \tag*{(78)} T α β ( q , r ) = ( B q + b , c α β B − T r ) , (78) where B B B is a constant invertible matrix, b b b is a constant vector, and c α β > 0 c_{\alpha\beta}>0 c α β > 0 is constant on the connected overlap component under consideration. Suppose that the local Hamiltonians satisfy
H β ∘ T α β = c α β H α . H_\beta\circ T_{\alpha\beta}=c_{\alpha\beta}H_\alpha. H β ∘ T α β = c α β H α . For mixed data ( q 0 , r 1 ) (q_0,r_1) ( q 0 , r 1 ) such that the corresponding solution of the α \alpha α -Hamiltonian system remains in this component throughout [ 0 , h ] [0,h] [ 0 , h ] , set
Q 0 = B q 0 + b , S 1 = c α β B − T r 1 . Q_0=Bq_0+b,\qquad S_1=c_{\alpha\beta}B^{-T}r_1. Q 0 = B q 0 + b , S 1 = c α β B − T r 1 . On corresponding solution branches, the exact discrete right Hamiltonians satisfy
H β , h + , E ( Q 0 , S 1 ) = c α β H α , h + , E ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) + S 1 ⋅ b . (79) H^{+,E}_{\beta,h}(Q_0,S_1)=c_{\alpha\beta}H^{+,E}_{\alpha,h}\left(B^{-1}(Q_0-b),c_{\alpha\beta}^{-1}B^TS_1\right)+S_1\cdot b. \tag*{(79)} H β , h + , E ( Q 0 , S 1 ) = c α β H α , h + , E ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) + S 1 ⋅ b . (79) Proof. The identities T α β ∗ Ω β = c α β Ω α T_{\alpha\beta}^{*}\Omega_\beta=c_{\alpha\beta}\Omega_\alpha T α β ∗ Ω β = c α β Ω α and H β ∘ T α β = c α β H α H_\beta\circ T_{\alpha\beta}=c_{\alpha\beta}H_\alpha H β ∘ T α β = c α β H α imply that T α β T_{\alpha\beta} T α β maps α \alpha α -Hamiltonian trajectories to β \beta β -Hamiltonian trajectories. Thus, a solution ( q ( t ) , r ( t ) ) (q(t),r(t)) ( q ( t ) , r ( t )) with q ( 0 ) = q 0 q(0)=q_0 q ( 0 ) = q 0 and r ( h ) = r 1 r(h)=r_1 r ( h ) = r 1 is transformed into
Q ( t ) = B q ( t ) + b , S ( t ) = c α β B − T r ( t ) , Q(t)=Bq(t)+b,\qquad S(t)=c_{\alpha\beta}B^{-T}r(t), Q ( t ) = B q ( t ) + b , S ( t ) = c α β B − T r ( t ) , with Q ( 0 ) = Q 0 Q(0)=Q_0 Q ( 0 ) = Q 0 and S ( h ) = S 1 S(h)=S_1 S ( h ) = S 1 . Along these corresponding trajectories,
S ⋅ Q ˙ − H β ( Q , S ) = c α β ( r ⋅ q ˙ − H α ( q , r ) ) , S\cdot\dot{Q}-H_\beta(Q,S)=c_{\alpha\beta}\left(r\cdot\dot{q}-H_\alpha(q,r)\right), S ⋅ Q ˙ − H β ( Q , S ) = c α β ( r ⋅ q ˙ − H α ( q , r ) ) , and the endpoint term satisfies
S 1 ⋅ Q ( h ) = c α β r 1 ⋅ q ( h ) + S 1 ⋅ b . S_1\cdot Q(h)=c_{\alpha\beta}r_1\cdot q(h)+S_1\cdot b. S 1 ⋅ Q ( h ) = c α β r 1 ⋅ q ( h ) + S 1 ⋅ b . Hence, by the definition of the exact discrete right Hamiltonian,
H β , h + , E ( Q 0 , S 1 ) = S 1 ⋅ Q ( h ) − ∫ 0 h ( S ⋅ Q ˙ − H β ( Q , S ) ) d t = c α β H α , h + , E ( q 0 , r 1 ) + S 1 ⋅ b . \begin{aligned}
H^{+,E}_{\beta,h}(Q_0,S_1)
&=S_1\cdot Q(h)-\int_0^h\left(S\cdot\dot{Q}-H_\beta(Q,S)\right)\,\mathrm{d}t\\
&=c_{\alpha\beta}H^{+,E}_{\alpha,h}(q_0,r_1)+S_1\cdot b.
\end{aligned} H β , h + , E ( Q 0 , S 1 ) = S 1 ⋅ Q ( h ) − ∫ 0 h ( S ⋅ Q ˙ − H β ( Q , S ) ) d t = c α β H α , h + , E ( q 0 , r 1 ) + S 1 ⋅ b . Substituting q 0 = B − 1 ( Q 0 − b ) q_0=B^{-1}(Q_0-b) q 0 = B − 1 ( Q 0 − b ) and r 1 = c α β − 1 B T S 1 r_1=c_{\alpha\beta}^{-1}B^TS_1 r 1 = c α β − 1 B T S 1 gives (79) .
Corollary 5.14 (Affine compatibility). Suppose that regular discrete right Hamiltonians satisfy
H β , h d + ( Q 0 , S 1 ) = c α β H α , h d + ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) + S 1 ⋅ b . H^{\mathrm{d}+}_{\beta,h}(Q_0,S_1)=c_{\alpha\beta}H^{\mathrm{d}+}_{\alpha,h}\left(B^{-1}(Q_0-b),c_{\alpha\beta}^{-1}B^TS_1\right)+S_1\cdot b. H β , h d + ( Q 0 , S 1 ) = c α β H α , h d + ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) + S 1 ⋅ b . Then their generated maps satisfy
F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h on corresponding local solution branches where both sides are defined.
Proof. Differentiating the transformation law gives
D 1 H β , h d + = c α β B − T D 1 H α , h d + , D 2 H β , h d + = B D 2 H α , h d + + b , D_1H^{d+}_{\beta,h}=c_{\alpha\beta}B^{-T}D_1H^{d+}_{\alpha,h},\qquad D_2H^{d+}_{\beta,h}=BD_2H^{d+}_{\alpha,h}+b, D 1 H β , h d + = c α β B − T D 1 H α , h d + , D 2 H β , h d + = B D 2 H α , h d + + b ,
where the derivatives on the right are evaluated at ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) \left(B^{-1}(Q_0-b),c_{\alpha\beta}^{-1}B^TS_1\right) ( B − 1 ( Q 0 − b ) , c α β − 1 B T S 1 ) . The discrete Hamilton equations therefore yield S 0 = c α β B − T r 0 S_0=c_{\alpha\beta}B^{-T}r_0 S 0 = c α β B − T r 0 and Q 1 = B q 1 + b Q_1=Bq_1+b Q 1 = B q 1 + b , which proves the claim. □
The affine equivariance of Runge–Kutta methods is a classical result; see, [33 ] . For completeness, we give a direct proof in the present setting.
Proposition 5.15 (Runge–Kutta equivariance under affine transitions). Let T α β ( z ) = A α β z + d α β T_{\alpha\beta}(z)=A_{\alpha\beta}z+d_{\alpha\beta} T α β ( z ) = A α β z + d α β , where A α β A_{\alpha\beta} A α β is a constant invertible matrix and d α β d_{\alpha\beta} d α β is a constant vector. Suppose that
f β ( T α β z ) = A α β f α ( z ) , f_\beta(T_{\alpha\beta}z)=A_{\alpha\beta}f_\alpha(z), f β ( T α β z ) = A α β f α ( z ) , and that the same Runge–Kutta tableau is used in both representations. Whenever the stage equations admit unique local solutions on corresponding branches, with all stages lying in the relevant overlap domains, the Runge–Kutta maps satisfy
F β , h R K ∘ T α β = T α β ∘ F α , h R K . F^{RK}_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F^{RK}_{\alpha,h}. F β , h R K ∘ T α β = T α β ∘ F α , h R K .
Proof. Let Z i Z_i Z i be the stages in the α \alpha α -representation, and set w n = T α β z n w_n=T_{\alpha\beta}z_n w n = T α β z n and W i = T α β Z i W_i=T_{\alpha\beta}Z_i W i = T α β Z i . Applying A α β A_{\alpha\beta} A α β to the stage equations and adding d α β d_{\alpha\beta} d α β gives
W i = A α β z n + d α β + h ∑ j = 1 s a i j A α β f α ( Z j ) = w n + h ∑ j = 1 s a i j f β ( W j ) . W_i=A_{\alpha\beta}z_n+d_{\alpha\beta}+h\sum_{j=1}^{s}a_{ij}A_{\alpha\beta}f_\alpha(Z_j)=w_n+h\sum_{j=1}^{s}a_{ij}f_\beta(W_j). W i = A α β z n + d α β + h j = 1 ∑ s a ij A α β f α ( Z j ) = w n + h j = 1 ∑ s a ij f β ( W j ) .
Thus, the W i W_i W i satisfy the stage equations in the β \beta β -representation. Similarly, the update satisfies
T α β z n + 1 = A α β z n + d α β + h ∑ i = 1 s b i A α β f α ( Z i ) = w n + h ∑ i = 1 s b i f β ( W i ) = w n + 1 . T_{\alpha\beta}z_{n+1}=A_{\alpha\beta}z_n+d_{\alpha\beta}+h\sum_{i=1}^{s}b_iA_{\alpha\beta}f_\alpha(Z_i)=w_n+h\sum_{i=1}^{s}b_if_\beta(W_i)=w_{n+1}. T α β z n + 1 = A α β z n + d α β + h i = 1 ∑ s b i A α β f α ( Z i ) = w n + h i = 1 ∑ s b i f β ( W i ) = w n + 1 .
Uniqueness on the corresponding stage branches therefore yields the claimed equivariance. □
The Hamiltonian construction does not require passage through a Lagrangian formulation. Nevertheless, when a regular Lagrangian description is available, the same overlap structure is inherited by the continuous and discrete Legendre transformations. We develop this counterpart in the next section.
The Lagrangian counterpart
The globalization results of § 4 were obtained on the Hamiltonian side and therefore did not require a Legendre transformation. When a regular Lagrangian description is available, the same local-to-global structure admits a variational formulation in terms of discrete Lagrangians; see [21 , 26 , 29 , 31 , 40 , 45 ] for variational integrators for Lagrangian systems and [10 , 11 ] for time-dependent Lagrangian systems. We state the continuous and discrete Lagrangian counterparts in sufficient detail to make the parallel with the Hamiltonian construction explicit, giving complete arguments where the LCS overlap structure enters essentially and indicating the relation with the Hamiltonian proof elsewhere. The logical structure is the following: the overlap scaling of the local Lagrangians induces the momentum scaling T α β T_{\alpha\beta} T α β used throughout, and compatible discrete Lagrangians generate compatible discrete right Hamiltonians, so that the globalization theorem applies verbatim.
Continuous overlap structure. Let { W α } \{W_\alpha\} { W α } be a family of open subsets of T Q TQ T Q , with W α W_\alpha W α contained in the inverse image of V α V_\alpha V α under the tangent-bundle projection, and let L α : W α → R L_\alpha: W_\alpha\to\mathbb{R} L α : W α → R be a local Lagrangian with fiber derivative F L α ( q , q ˙ ) = ( q , D 2 L α ( q , q ˙ ) ) \mathop{\mathrm{FL}}\nolimits_\alpha(q,\dot{q}) = (q,D_2L_\alpha(q,\dot{q})) FL α ( q , q ˙ ) = ( q , D 2 L α ( q , q ˙ )) . We assume that each W α W_\alpha W α is an open Legendre domain, that is, F L α \mathop{\mathrm{FL}}\nolimits_\alpha FL α is a diffeomorphism from W α W_\alpha W α onto an open subset of U α U_\alpha U α , and that the local inverse branches are chosen consistently on overlaps. Regularity of L α L_\alpha L α alone does not guarantee that W α W_\alpha W α can be taken to be the whole velocity fiber; this holds, for instance, when L α L_\alpha L α is hyperregular. A typical choice, when a globally defined Lagrangian L L L is available on compatible domains, is L α = e − σ α L L_\alpha=e^{-\sigma_\alpha}L L α = e − σ α L . Let H α ( q , r ) = r ⋅ q ˙ − L α ( q , q ˙ ) H_\alpha(q,r)=r\cdot\dot{q}-L_\alpha(q,\dot{q}) H α ( q , r ) = r ⋅ q ˙ − L α ( q , q ˙ ) , r = D 2 L α ( q , q ˙ ) r=D_2L_\alpha(q,\dot{q}) r = D 2 L α ( q , q ˙ ) , be the associated local Hamiltonian, whose canonical Hamilton equations are equivalent to the Euler–Lagrange equations of L α L_\alpha L α on such a domain.
Proposition 6.1. Suppose that on a connected component of W α ∩ W β W_\alpha\cap W_\beta W α ∩ W β
L β = c α β L α . (80) L_\beta=c_{\alpha\beta}L_\alpha. \tag*{(80)} L β = c α β L α . (80) Then
F L β = T α β ∘ F L α , H β ( q , s ) = c α β H α ( q , c α β − 1 s ) , (81) \mathop{\mathrm{FL}}\nolimits_\beta=T_{\alpha\beta}\circ\mathop{\mathrm{FL}}\nolimits_\alpha,\qquad H_\beta(q,s)=c_{\alpha\beta}H_\alpha\left(q,c_{\alpha\beta}^{-1}s\right), \tag*{(81)} FL β = T α β ∘ FL α , H β ( q , s ) = c α β H α ( q , c α β − 1 s ) , (81) where the identity on the velocity variables is understood in the first relation.
Proof. Differentiating (80) in q ˙ \dot{q} q ˙ gives D 2 L β = c α β D 2 L α D_2L_\beta=c_{\alpha\beta}D_2L_\alpha D 2 L β = c α β D 2 L α , i.e., s = c α β r s=c_{\alpha\beta}r s = c α β r , which is the first identity, and then H β ( q , s ) = s ⋅ q ˙ − L β ( q , q ˙ ) = c α β [ r ⋅ q ˙ − L α ( q , q ˙ ) ] = c α β H α ( q , r ) H_\beta(q,s)=s\cdot\dot{q}-L_\beta(q,\dot{q})=c_{\alpha\beta}[r\cdot\dot{q}-L_\alpha(q,\dot{q})]=c_{\alpha\beta}H_\alpha(q,r) H β ( q , s ) = s ⋅ q ˙ − L β ( q , q ˙ ) = c α β [ r ⋅ q ˙ − L α ( q , q ˙ )] = c α β H α ( q , r ) , which is (17) after substituting r = c α β − 1 s r=c_{\alpha\beta}^{-1}s r = c α β − 1 s .
Thus the multiplicative scaling of the Lagrangians is the Lagrangian expression of the conformal cocycle (9) , and it reproduces exactly the Hamiltonian overlap relation used in the previous sections.
Compatible discrete Lagrangians. Let L α , h E ( q 0 , q 1 ) = ∫ 0 h L α ( q , q ˙ ) d t L_{\alpha,h}^{E}(q_0,q_1)=\int_0^h L_\alpha(q,\dot{q})\,\mathrm{d}t L α , h E ( q 0 , q 1 ) = ∫ 0 h L α ( q , q ˙ ) d t denote the exact discrete Lagrangian, evaluated on the solution of the Euler–Lagrange boundary-value problem with endpoints q 0 , q 1 q_0,q_1 q 0 , q 1 . Since multiplication of a Lagrangian by a nonzero constant does not alter its Euler–Lagrange equations, the same curve solves both boundary-value problems on an overlap, and L β = c α β L α L_\beta=c_{\alpha\beta}L_\alpha L β = c α β L α implies L β , h E = c α β L α , h E L_{\beta,h}^{E}=c_{\alpha\beta}L_{\alpha,h}^{E} L β , h E = c α β L α , h E , the Lagrangian transcription of Theorem 4.1 .
Definition 6.2. A family of discrete Lagrangians { L α , h d } α ∈ A \{L_{\alpha,h}^{d}\}_{\alpha\in A} { L α , h d } α ∈ A is LCS-compatible if
L β , h d ( q 0 , q 1 ) = c α β L α , h d ( q 0 , q 1 ) (82) L_{\beta,h}^{d}(q_0,q_1)=c_{\alpha\beta}L_{\alpha,h}^{d}(q_0,q_1) \tag*{(82)} L β , h d ( q 0 , q 1 ) = c α β L α , h d ( q 0 , q 1 ) (82) on every connected component of every overlap.
As in the continuous case, multiplication of the discrete action sum ∑ k L α , h d ( q k , q k + 1 ) \sum_k L_{\alpha,h}^{d}(q_k,q_{k+1}) ∑ k L α , h d ( q k , q k + 1 ) by the positive constant c α β c_{\alpha\beta} c α β leaves its stationary sequences unchanged; compatible discrete Lagrangians therefore determine the same discrete configuration trajectories on overlaps. Writing the left and right local discrete Legendre transformations as F − L α , h d ( q k , q k + 1 ) = ( q k , − D 1 L α , h d ( q k , q k + 1 ) ) \mathbb{F}^{-}L_{\alpha,h}^{d}(q_k,q_{k+1})=(q_k,-D_1L_{\alpha,h}^{d}(q_k,q_{k+1})) F − L α , h d ( q k , q k + 1 ) = ( q k , − D 1 L α , h d ( q k , q k + 1 )) and F + L α , h d ( q k , q k + 1 ) = ( q k + 1 , D 2 L α , h d ( q k , q k + 1 ) ) \mathbb{F}^{+}L_{\alpha,h}^{d}(q_k,q_{k+1})=(q_{k+1},D_2L_{\alpha,h}^{d}(q_k,q_{k+1})) F + L α , h d ( q k , q k + 1 ) = ( q k + 1 , D 2 L α , h d ( q k , q k + 1 )) , the associated discrete map is F α , h L = F + L α , h d ∘ ( F − L α , h d ) − 1 F_{\alpha,h}^{L}=\mathbb{F}^{+}L_{\alpha,h}^{d}\circ(\mathbb{F}^{-}L_{\alpha,h}^{d})^{-1} F α , h L = F + L α , h d ∘ ( F − L α , h d ) − 1 , and it satisfies the following properties.
Proposition 6.3. Let { L α , h d } \{L_{\alpha,h}^{d}\} { L α , h d } be regular and LCS-compatible. Then
(i) the discrete Legendre transformations satisfy F ± L β , h d = T α β ∘ F ± L α , h d \mathbb{F}^{\pm}L_{\beta,h}^{d}=T_{\alpha\beta}\circ\mathbb{F}^{\pm}L_{\alpha,h}^{d} F ± L β , h d = T α β ∘ F ± L α , h d ;
(ii) each local map is symplectic, ( F α , h L ) ∗ Ω α = Ω α (F_{\alpha,h}^{L})^{*}\Omega_{\alpha}=\Omega_{\alpha} ( F α , h L ) ∗ Ω α = Ω α ;
(iii) the local maps are equivariant, F β , h L ∘ T α β = T α β ∘ F α , h L F_{\beta,h}^{L}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h}^{L} F β , h L ∘ T α β = T α β ∘ F α , h L , on their common domain in the sense of Theorem 4.3 .
Proof. Assertion (i) is obtained by differentiating (82) in each argument, exactly as (42) was obtained from (41) ; here both derivatives carry the same factor c α β c_{\alpha\beta} c α β , so both discrete momenta scale by it. Assertion (ii) is the transcription of Proposition 3.6 : writing d L α , h d = Θ α , h d , + − Θ α , h d , − \mathrm{d}L_{\alpha,h}^{d}=\Theta_{\alpha,h}^{d,+}-\Theta_{\alpha,h}^{d,-} d L α , h d = Θ α , h d , + − Θ α , h d , − with Θ α , h d , + = D 2 L α , h d ( q 0 , q 1 ) ⋅ d q 1 \Theta_{\alpha,h}^{d,+}=D_2L_{\alpha,h}^{d}(q_0,q_1)\cdot\mathrm{d}q_1 Θ α , h d , + = D 2 L α , h d ( q 0 , q 1 ) ⋅ d q 1 and Θ α , h d , − = − D 1 L α , h d ( q 0 , q 1 ) ⋅ d q 0 \Theta_{\alpha,h}^{d,-}=-D_1L_{\alpha,h}^{d}(q_0,q_1)\cdot\mathrm{d}q_0 Θ α , h d , − = − D 1 L α , h d ( q 0 , q 1 ) ⋅ d q 0 , one more exterior derivative gives d Θ α , h d , + = d Θ α , h d , − \mathrm{d}\Theta_{\alpha,h}^{d,+}=\mathrm{d}\Theta_{\alpha,h}^{d,-} d Θ α , h d , + = d Θ α , h d , − , which under the discrete Legendre transformations is exactly (38) . Assertion (iii) follows from (i) and the definition of F α , h L F_{\alpha,h}^{L} F α , h L by the argument of Theorem 4.3 , with the same connected-component proviso.
Equivalence with the Hamiltonian construction. Assume that r 1 = D 2 L α , h d ( q 0 , q 1 ) r_1=D_2L_{\alpha,h}^{d}(q_0,q_1) r 1 = D 2 L α , h d ( q 0 , q 1 ) can be solved locally for q 1 = q 1 ( q 0 , r 1 ) q_1=q_1(q_0,r_1) q 1 = q 1 ( q 0 , r 1 ) and define the associated discrete right Hamiltonian by the discrete Legendre transform
H α , h d , + ( q 0 , r 1 ) = r 1 ⋅ q 1 − L α , h d ( q 0 , q 1 ) . (83) H_{\alpha,h}^{d,+}(q_0,r_1)=r_1\cdot q_1-L_{\alpha,h}^{d}(q_0,q_1). \tag*{(83)} H α , h d , + ( q 0 , r 1 ) = r 1 ⋅ q 1 − L α , h d ( q 0 , q 1 ) . (83)
Theorem 6.4. The function (83) satisfies
D 1 H α , h d , + ( q 0 , r 1 ) = − D 1 L α , h d ( q 0 , q 1 ) = r 0 , D 2 H α , h d , + ( q 0 , r 1 ) = q 1 , D_1H_{\alpha,h}^{d,+}(q_0,r_1)=-D_1L_{\alpha,h}^{d}(q_0,q_1)=r_0,\qquad D_2H_{\alpha,h}^{d,+}(q_0,r_1)=q_1, D 1 H α , h d , + ( q 0 , r 1 ) = − D 1 L α , h d ( q 0 , q 1 ) = r 0 , D 2 H α , h d , + ( q 0 , r 1 ) = q 1 , so that L α , h d L_{\alpha,h}^{d} L α , h d and H α , h d , + H_{\alpha,h}^{d,+} H α , h d , + generate the same phase-space map. Moreover, if the discrete Lagrangians are LCS-compatible then so are the discrete right Hamiltonians:
H β , h d , + ( q 0 , s 1 ) = c α β H α , h d , + ( q 0 , c α β − 1 s 1 ) . (84) H_{\beta,h}^{d,+}(q_0,s_1)=c_{\alpha\beta}H_{\alpha,h}^{d,+}\left(q_0,c_{\alpha\beta}^{-1}s_1\right). \tag*{(84)} H β , h d , + ( q 0 , s 1 ) = c α β H α , h d , + ( q 0 , c α β − 1 s 1 ) . (84)
Proof. Differentiating (83) and using r 1 = D 2 L α , h d ( q 0 , q 1 ) r_1=D_2L_{\alpha,h}^{d}(q_0,q_1) r 1 = D 2 L α , h d ( q 0 , q 1 ) , the terms multiplying d q 1 \mathrm{d}q_1 d q 1 cancel and d H α , h d , + = − D 1 L α , h d ⋅ d q 0 + q 1 ⋅ d r 1 \mathrm{d}H_{\alpha,h}^{d,+}=-D_1L_{\alpha,h}^{d}\cdot\mathrm{d}q_0+q_1\cdot\mathrm{d}r_1 d H α , h d , + = − D 1 L α , h d ⋅ d q 0 + q 1 ⋅ d r 1 , which yields the two stated identities. For the compatibility statement, (82) gives s 1 = D 2 L β , h d ( q 0 , q 1 ) = c α β r 1 s_1=D_2L_{\beta,h}^{d}(q_0,q_1)=c_{\alpha\beta}r_1 s 1 = D 2 L β , h d ( q 0 , q 1 ) = c α β r 1 , and therefore
H β , h d , + ( q 0 , s 1 ) = s 1 ⋅ q 1 − L β , h d ( q 0 , q 1 ) = c α β [ r 1 ⋅ q 1 − L α , h d ( q 0 , q 1 ) ] = c α β H α , h d , + ( q 0 , r 1 ) , H_{\beta,h}^{d,+}(q_0,s_1)=s_1\cdot q_1-L_{\beta,h}^{d}(q_0,q_1)=c_{\alpha\beta}\left[r_1\cdot q_1-L_{\alpha,h}^{d}(q_0,q_1)\right]=c_{\alpha\beta}H_{\alpha,h}^{d,+}(q_0,r_1), H β , h d , + ( q 0 , s 1 ) = s 1 ⋅ q 1 − L β , h d ( q 0 , q 1 ) = c α β [ r 1 ⋅ q 1 − L α , h d ( q 0 , q 1 ) ] = c α β H α , h d , + ( q 0 , r 1 ) ,
which is (84) after substituting r 1 = c α β − 1 s 1 r_1=c_{\alpha\beta}^{-1}s_1 r 1 = c α β − 1 s 1 .
Corollary 6.5. A regular LCS-compatible family of discrete Lagrangians glues to a global discrete map F h L F_h^L F h L satisfying
( F h L ) ∗ Ω = e σ α ∘ F h L − σ α Ω . (85) \left(F_h^L\right)^*\Omega=e^{\sigma_\alpha\circ F_h^L-\sigma_\alpha}\Omega. \tag*{(85)} ( F h L ) ∗ Ω = e σ α ∘ F h L − σ α Ω. (85) If, in addition, L α , h d − L α , h E = O ( h p + 1 ) L_{\alpha,h}^d-L_{\alpha,h}^E=O(h^{p+1}) L α , h d − L α , h E = O ( h p + 1 ) in the sense of [31 Section 2.3] , with initial data in open subsets whose closures are compact and contained in W α W_\alpha W α , then the resulting global variational integrator has order p p p .
Proof. By Proposition 6.3 , the local maps are symplectic and equivariant, so Theorem 4.7 applies. The local order statement follows from [31 Theorem 2.3.1] . The smooth, h h h -independent representation maps Ψ α \Psi_\alpha Ψ α preserve the local error order, yielding the stated order of the reconstructed map.
Remark 6.6. Galerkin discrete Lagrangians provide a natural class of compatible families: if the same approximation space and quadrature rule are used in every local representation, then L β , h d = c α β L α , h d L_{\beta,h}^d=c_{\alpha\beta}L_{\alpha,h}^d L β , h d = c α β L α , h d as functions of the internal variables, the stationarity equations have the same solutions, and (82) follows. This is the Lagrangian counterpart of Proposition 5.1 .
Summarizing, the scaling L β , h d = c α β L α , h d L_{\beta,h}^d=c_{\alpha\beta}L_{\alpha,h}^d L β , h d = c α β L α , h d and the compatibility relation (84) are intertwined by the discrete Legendre transform. The Hamiltonian route remains the more general starting point, since it requires neither the existence nor the invertibility of a Legendre transformation; the Lagrangian route, when available, provides an equivalent variational interpretation and an alternative practical construction of compatible methods. Having established both formulations, we now turn to numerical experiments designed to test not only convergence, but also the two geometric properties that are central to the globalization framework: overlap compatibility and discrete conformal preservation.
Numerical experiments
The experiments reported here test the three statements on which the construction rests, separately rather than jointly:
(T1) Order. The reconstructed global map F h F_h F h approximates the LCS flow to the order predicted by Proposition 4.8 and Corollary 5.9 .
(T2) Discrete conformal preservation. The map F h F_h F h satisfies F h ∗ Ω = ρ h Ω F_h^*\Omega=\rho_h\Omega F h ∗ Ω = ρ h Ω to round-off, and not merely to the order of the method (Theorem 4.7 ).
(T3) Overlap compatibility. The local maps are equivariant and glue, so that the computed trajectory does not depend on the local representation in which it is computed (Theorem 4.3 and Lemma 4.4 ).
Two further experiments probe the sharpness of the framework: the long-time behaviour of the conformal energy law (75) , and the failure of globalization for families that are locally symplectic but not LCS-compatible. A final example with a non-exact Lee form exhibits the chart switching forced by the topology and the monodromy of the discrete conformal factor.
Test problem and implementation. All experiments of § 8.1 to § 8.3 use the conformal oscillator of Example 2.3 with k = 1 k=1 k = 1 , μ = 1 2 \mu=\frac{1}{2} μ = 2 1 and initial data ( q 0 , p 0 , x 0 , y 0 ) = ( 1 , 0 , 0 , 1 2 ) (q_0,p_0,x_0,y_0)=(1,0,0,\frac{1}{2}) ( q 0 , p 0 , x 0 , y 0 ) = ( 1 , 0 , 0 , 2 1 ) . The conformal potential is σ α = − x \sigma_\alpha=-x σ α = − x , so that c α β = e C α β c_{\alpha\beta}=e^{C_{\alpha\beta}} c α β = e C α β with σ α − σ β = C α β \sigma_\alpha-\sigma_\beta=C_{\alpha\beta} σ α − σ β = C α β constant. In the local chart labelled by c = c α β c=c_{\alpha\beta} c = c α β , the local Hamiltonian is, by (17) ,
H c ( q , r , x , s ) = 1 2 c e − x ( r 2 + s 2 ) + c e x ( k 2 q 2 + μ x ) (86) H_c(q,r,x,s)=\frac{1}{2c}e^{-x}(r^2+s^2)+ce^x\left(\frac{k}{2}q^2+\mu x\right) \tag*{(86)} H c ( q , r , x , s ) = 2 c 1 e − x ( r 2 + s 2 ) + c e x ( 2 k q 2 + μx ) (86)
which reduces to (28) for c = 1 c=1 c = 1 . We emphasize that the vector field integrated in the β \beta β -representation is the one generated by H β H_\beta H β , not by H α H_\alpha H α ; this is exactly the content of the overlap relation, and using the same local Hamiltonian in both charts would destroy the equivariance being tested. In the LCS variables the two-form is (27) , whose matrix in the ordered basis ( ∂ q , ∂ p , ∂ x , ∂ y ) (\partial_q,\partial_p,\partial_x,\partial_y) ( ∂ q , ∂ p , ∂ x , ∂ y ) is
Ω ^ ( z ) = ( 0 1 p 0 − 1 0 0 0 − p 0 0 1 0 0 − 1 0 ) . (87) \widehat{\Omega}(z)=
\begin{pmatrix}
0 & 1 & p & 0 \\
-1 & 0 & 0 & 0 \\
-p & 0 & 0 & 1 \\
0 & 0 & -1 & 0
\end{pmatrix}. \tag*{(87)} Ω ( z ) = 0 − 1 − p 0 1 0 0 0 p 0 0 − 1 0 0 1 0 . (87)
Computations are carried out in double precision. The implicit stage equations of the Gauss–Legendre methods are solved by Newton’s method with the analytic stage Jacobian, iterating until the Newton increment satisfies ∥ Δ Z ∥ ∞ ≤ 10 ε m a c h ( 1 + ∥ Z ∥ ∞ ) \lVert\Delta Z\rVert_\infty\leq10\varepsilon_{\mathrm{mach}}(1+\lVert Z\rVert_\infty) ∥ Δ Z ∥ ∞ ≤ 10 ε mach ( 1 + ∥ Z ∥ ∞ ) , with at most 30 iterations. Since this criterion is relative to the size of the stage variables, it remains attainable in the rescaled representations, where the momenta are of order c α β c_{\alpha\beta} c α β ; an absolute residual tolerance of order 10 − 15 10^{-15} 1 0 − 15 is not attainable in double precision for c α β = e 2 π ≈ 535.5 c_{\alpha\beta}=e^{2\pi}\approx535.5 c α β = e 2 π ≈ 535.5 . In exact arithmetic, Newton’s method is affine covariant: with initial guesses related by T α β T_{\alpha\beta} T α β and compatible stopping criteria, its iterates in the two charts correspond exactly through T α β T_{\alpha\beta} T α β , which is consistent with the round-off-level equivariance defects reported below. One-step Jacobians are computed by differentiating the implicit stage equations analytically, rather than by finite differences. Writing u n = Ψ α ( z n ) u_n=\Psi_\alpha(z_n) u n = Ψ α ( z n ) for the local canonical coordinates and J n l o c = ∂ u n + 1 / ∂ u n J_n^{\mathrm{loc}}=\partial u_{n+1}/\partial u_n J n loc = ∂ u n + 1 / ∂ u n , the Jacobian in the LCS variables is obtained by the chain rule: J n = D ( Ψ α − 1 ) ( u n + 1 ) J n l o c D Ψ α ( z n ) J_n=D(\Psi_\alpha^{-1})(u_{n+1})J_n^{\mathrm{loc}}D\Psi_\alpha(z_n) J n = D ( Ψ α − 1 ) ( u n + 1 ) J n loc D Ψ α ( z n ) . Here the derivative of the inverse representation map is evaluated at the local endpoint u n + 1 = Ψ α ( z n + 1 ) u_{n+1}=\Psi_\alpha(z_{n+1}) u n + 1 = Ψ α ( z n + 1 ) . This avoids finite-difference errors when measuring conformal preservation at the level of round-off. As a non-symplectic control, we use the classical explicit fourth-order Runge–Kutta method applied to the same local canonical system.
Convergence
We integrate to T = 10 T=10 T = 10 and measure the error ∥ z N − z r e f ( T ) ∥ ∞ \lVert z_N-z_{\mathrm{ref}}(T)\rVert_\infty ∥ z N − z ref ( T ) ∥ ∞ in the LCS variables against a reference trajectory computed with the three-stage Gauss–Legendre method and h r e f = T / 2 19 h_{\mathrm{ref}}=T/2^{19} h ref = T / 2 19 . Table 1 and Figure 1 report the results.
Table 1. Global error at T = 10 T=10 T = 10 in the LCS variables. Observed orders are least-squares slopes over the unsaturated range. Entries marked † have reached the accuracy of the reference solution (≈ 5 ⋅ 10 − 12 \approx5\cdot10^{-12} ≈ 5 ⋅ 1 0 − 12 ) and are excluded from the fit.
Figure 1. Convergence of the reconstructed global map, confirming orders 2, 4 and 6 predicted by Corollary 5.9 .
The measured orders are 2, 4 and 6, in agreement with Corollary 5.9 . Since the representation maps Ψ α \Psi_\alpha Ψ α are h h h -independent diffeomorphisms, the order measured after reconstruction coincides with the order in the local canonical variables, as asserted in Proposition 4.8 .
Discrete conformal preservation
This is the central structural test. For each step we evaluate the defect
ε n L C S = ∥ J n T Ω ^ ( z n + 1 ) J n − ρ n Ω ^ ( z n ) ∥ F , ρ n = e σ α ( z n + 1 ) − σ α ( z n ) = e − ( x n + 1 − x n ) , (88) \varepsilon_n^{\mathrm{LCS}} = \left\|J_n^{\mathsf{T}}\widehat{\Omega}(z_{n+1})J_n-\rho_n\widehat{\Omega}(z_n)\right\|_F,\qquad\rho_n=e^{\sigma_\alpha(z_{n+1})-\sigma_\alpha(z_n)}=e^{-(x_{n+1}-x_n)}, \tag*{(88)} ε n LCS = J n T Ω ( z n + 1 ) J n − ρ n Ω ( z n ) F , ρ n = e σ α ( z n + 1 ) − σ α ( z n ) = e − ( x n + 1 − x n ) , (88)
and report max n ε n L C S \max_n\varepsilon_n^{\mathrm{LCS}} max n ε n LCS over [ 0 , 10 ] [0,10] [ 0 , 10 ] in Table 2 and Figure 2 .
Table 2. Maximum conformal defect (88) over [ 0 , 10 ] [0,10] [ 0 , 10 ] . For the Gauss–Legendre methods the defect is at round-off and independent of h h h ; for the non-symplectic RK4 control, the last four values give an observed slope of approximately 4.96, consistent with an O ( h 5 ) O(h^5) O ( h 5 ) one-step defect.
Figure 2. Discrete conformal preservation. The flat curves at ≈ 3 ⋅ 10 − 15 \approx3 \cdot10^{-15} ≈ 3 ⋅ 1 0 − 15 are the signature of Theorem 4.7 ; the curve with slope approximately five in the fine-step range is the non-symplectic control, for which the identity holds only to the order of the method.
The contrast is the point of the experiment. For the compatible symplectic methods the defect is flat in h h h at the level of round-off: the identity F h ∗ Ω = ρ h Ω F_h^*\Omega=\rho_h\Omega F h ∗ Ω = ρ h Ω holds exactly, not to the order of the method. For the explicit Runge–Kutta control the same quantity behaves as O ( h 5 ) O(h^5) O ( h 5 ) in the fine-step range, that is, it is merely a consistency error. It is instructive to report also the purely symplectic defect max n ∥ J n T Ω ^ ( z n + 1 ) J n − Ω ^ ( z n ) ∥ F \max_n\|J_n^{\mathsf{T}}\widehat{\Omega}(z_{n+1})J_n-\widehat{\Omega}(z_n)\|_F max n ∥ J n T Ω ( z n + 1 ) J n − Ω ( z n ) ∥ F , which for the implicit midpoint rule is 5.9, 2.7, 1.28, 0.65, 0.32, 0.16 as h h h is halved. This quantity is not small and decreases like O ( h ) O(h) O ( h ) ; it measures ∣ ρ n − 1 ∣ |\rho_n-1| ∣ ρ n − 1∣ rather than any numerical error, and confirms that the global map is conformally, not canonically, symplectic.
Overlap compatibility and chart switching
We use two local potentials differing by a constant, σ α = − x \sigma_\alpha=-x σ α = − x and σ β = − x − C α β \sigma_\beta=-x-C_{\alpha\beta} σ β = − x − C α β , and the corresponding local Hamiltonians (86) . Since the Lee form of this example is exact, the two representations cover the same domain; the experiment therefore tests the equivariance and gluing statements directly, rather than a situation in which chart switching is forced by the topology.
Equivariance defect. For the two-stage method we evaluate
δ n e q = ∥ F β , h ( T α β ( z n ) ) − T α β ( F α , h ( z n ) ) ∥ ∞ ∥ T α β ( F α , h ( z n ) ) ∥ ∞ (89) \delta_n^{\mathrm{eq}}=\frac{\lVert F_{\beta,h}\left(T_{\alpha\beta}(z_n)\right)-T_{\alpha\beta}\left(F_{\alpha,h}(z_n)\right)\rVert_\infty}{\lVert T_{\alpha\beta}\left(F_{\alpha,h}(z_n)\right)\rVert_\infty} \tag*{(89)} δ n eq = ∥ T α β ( F α , h ( z n ) ) ∥ ∞ ∥ F β , h ( T α β ( z n ) ) − T α β ( F α , h ( z n ) ) ∥ ∞ (89)
along the trajectory. For c α β ∈ { e − 1 , 2 , e 2 π ≈ 535.5 } c_{\alpha\beta}\in\{e^{-1},2,e^{2\pi}\approx535.5\} c α β ∈ { e − 1 , 2 , e 2 π ≈ 535.5 } and h ∈ { 1.56 ⋅ 10 − 1 , 3.91 ⋅ 10 − 2 , 9.77 ⋅ 10 − 3 } h\in\{1.56\cdot10^{-1},3.91\cdot10^{-2},9.77\cdot10^{-3}\} h ∈ { 1.56 ⋅ 1 0 − 1 , 3.91 ⋅ 1 0 − 2 , 9.77 ⋅ 1 0 − 3 } , the largest value observed over all nine runs is 2.9 ⋅ 10 − 16 2.9\cdot10^{-16} 2.9 ⋅ 1 0 − 16 . For the tested transition constants, the normwise relative equivariance defect remains at round-off level, consistently with Theorem 4.3 . This normalization measures the discrepancy relative to the largest state component and does not provide a componentwise relative error.
Chart switching. We integrate with h = T / 512 h=T/512 h = T /512 and switch local representative at step m ∈ { 1 , N / 4 , N / 2 , 3 N / 4 , N − 1 } m\in\{1,N/4,N/2,3N/4,N-1\} m ∈ { 1 , N /4 , N /2 , 3 N /4 , N − 1 } , composing with T α β T_{\alpha\beta} T α β at the switching step, and compare the resulting LCS trajectory with the one computed entirely in the α \alpha α -representation. The largest discrepancy over all switching points is 4.1 ⋅ 10 − 15 4.1\cdot10^{-15} 4.1 ⋅ 1 0 − 15 for c α β = e − 1 c_{\alpha\beta}=e^{-1} c α β = e − 1 , 1.6 ⋅ 10 − 15 1.6\cdot10^{-15} 1.6 ⋅ 1 0 − 15 for c α β = 2 c_{\alpha\beta}=2 c α β = 2 and 3.4 ⋅ 10 − 15 3.4\cdot10^{-15} 3.4 ⋅ 1 0 − 15 for c α β = e 2 π c_{\alpha\beta}=e^{2\pi} c α β = e 2 π . The glued map of Lemma 4.4 is thus independent of where the chart is changed, to round-off.
Long-time conformal energy law. By Remark 5.11 , the local Hamiltonian H α = e − σ α H = e x H H_\alpha=e^{-\sigma_\alpha}H=e^xH H α = e − σ α H = e x H is a first integral of the continuous dynamics. We test the weighted energy relation (75) numerically over a finite time interval by monitoring the relative error
Δ n = ∣ e − σ α ( z n ) H ( z n ) − e − σ α ( z 0 ) H ( z 0 ) ∣ ∣ e − σ α ( z 0 ) H ( z 0 ) ∣ (90) \Delta_n = \frac{\left|e^{-\sigma_\alpha(z_n)}H(z_n)-e^{-\sigma_\alpha(z_0)}H(z_0)\right|}{\left|e^{-\sigma_\alpha(z_0)}H(z_0)\right|} \tag*{(90)} Δ n = e − σ α ( z 0 ) H ( z 0 ) e − σ α ( z n ) H ( z n ) − e − σ α ( z 0 ) H ( z 0 ) (90)
over 2 ⋅ 10 5 2 \cdot10^5 2 ⋅ 1 0 5 steps with h = 0.05 h=0.05 h = 0.05 , that is up to T = 10 4 T=10^4 T = 1 0 4 ; Table 3 and Figure 3 show the result.
Table 3. Relative error (90) in the conformal energy law, h = 0.05 h=0.05 h = 0.05 , T = 10 4 T=10^4 T = 1 0 4 . The Gauss–Legendre errors oscillate without secular growth; the explicit control drifts linearly in t t t .
Figure 3. Conformal energy law over 2 ⋅ 10 5 2 \cdot10^5 2 ⋅ 1 0 5 steps. The Gauss–Legendre curves oscillate within a band fixed by h h h ; the non-symplectic control grows linearly.
The qualitative distinction predicted by Remark 5.12 is clearly visible: the Gauss–Legendre errors remain bounded and oscillatory over the whole interval, whereas the explicit control drifts linearly, its error at t = T / 4 , T / 2 , 3 T / 4 , T t=T/4,T/2,3T/4,T t = T /4 , T /2 , 3 T /4 , T being roughly proportional to 1 : 2 : 3 : 4 1:2:3:4 1 : 2 : 3 : 4 . The size of the bounded band scales with the order of the method: repeating the experiment to T = 10 3 T=10^3 T = 1 0 3 with h = 0.2 , 0.1 , 0.05 , 0.025 h=0.2,0.1,0.05,0.025 h = 0.2 , 0.1 , 0.05 , 0.025 gives observed slopes 2.00 2.00 2.00 , 4.02 4.02 4.02 and 6.04 6.04 6.04 for s = 1 , 2 , 3 s=1,2,3 s = 1 , 2 , 3 , respectively, consistent with an O ( h 2 s ) O(h^{2s}) O ( h 2 s ) energy-error band over the tested interval.
Across the three Gauss–Legendre runs over [ 0 , 10 4 ] [0,10^4] [ 0 , 1 0 4 ] , the observed nodal values satisfy − 3.210 < x n < 0.132 -3.210<x_n<0.132 − 3.210 < x n < 0.132 and ∣ q n ∣ < 3.006 |q_n|<3.006 ∣ q n ∣ < 3.006 , so that the numerical orbits remain in a bounded region over the tested interval, although the energy level set itself is not compact. The maximum energy errors occur near x n = − 3.20 x_n=-3.20 x n = − 3.20 , where the factor e − x e^{-x} e − x in the kinetic term of (28) is largest; the four sampled times do not capture these peaks.
Locally symplectic but incompatible families. This experiment illustrates that local symplecticity alone does not ensure that the local numerical maps define a single global map. We retain the compatible continuous Hamiltonians
H β ( q , s ) = c α β H α ( q , c α β − 1 s ) , H_{\beta}(q,s)=c_{\alpha\beta}H_{\alpha}(q,c_{\alpha\beta}^{-1}s), H β ( q , s ) = c α β H α ( q , c α β − 1 s ) ,
and use the two-stage Gauss–Legendre method as the underlying discretization in both local representations. Its discrete right Hamiltonians satisfy
H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) . H_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}H_{\alpha,h}^{d+}(q,c_{\alpha\beta}^{-1}s). H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) .
We leave H α , h d + H_{\alpha,h}^{d+} H α , h d + unchanged and replace the discrete Hamiltonian in U β U_{\beta} U β by
H ^ β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) + h 5 2 ∥ s ∥ 2 . (91) \widehat{H}_{\beta,h}^{d+}(q,s)=c_{\alpha\beta}H_{\alpha,h}^{d+}(q,c_{\alpha\beta}^{-1}s)+\frac{h^5}{2}\lVert s\rVert^2. \tag*{(91)} H β , h d + ( q , s ) = c α β H α , h d + ( q , c α β − 1 s ) + 2 h 5 ∥ s ∥ 2 . (91)
The power h 5 h^5 h 5 is chosen to match the local error order of the fourth-order Gauss–Legendre method. Both local maps remain symplectic. H ^ β , h d + \widehat{H}_{\beta,h}^{d+} H β , h d + defines a Type II generating function, and the added term does not change its mixed derivative D 12 D_{12} D 12 . Thus the local nondegeneracy condition is inherited from the unmodified discrete Hamiltonian. The coordinate transition itself remains T α β ( q , r ) = ( q , c α β r ) T_{\alpha\beta}(q,r)=(q,c_{\alpha\beta}r) T α β ( q , r ) = ( q , c α β r ) . To compare the two local updates, let
F α , h ( q k , r k ) = ( q k + 1 α , r k + 1 ) . F_{\alpha,h}(q_k,r_k)=(q_{k+1}^{\alpha},r_{k+1}). F α , h ( q k , r k ) = ( q k + 1 α , r k + 1 ) .
Starting in U β U_{\beta} U β from the corresponding point ( q k , s k ) (q_k,s_k) ( q k , s k ) , with s k = c α β r k s_k=c_{\alpha\beta}r_k s k = c α β r k , the generating relations for (91) give
s k = c α β D 1 H α , h d + ( q k , c α β − 1 s k + 1 ) , s_k=c_{\alpha\beta}D_1H_{\alpha,h}^{d+}(q_k,c_{\alpha\beta}^{-1}s_{k+1}), s k = c α β D 1 H α , h d + ( q k , c α β − 1 s k + 1 ) ,
q k + 1 β = D 2 H α , h d + ( q k , c α β − 1 s k + 1 ) + h 5 s k + 1 . q_{k+1}^{\beta}=D_2H_{\alpha,h}^{d+}(q_k,c_{\alpha\beta}^{-1}s_{k+1})+h^5s_{k+1}. q k + 1 β = D 2 H α , h d + ( q k , c α β − 1 s k + 1 ) + h 5 s k + 1 .
On the locally unique solution branch, these imply
s k + 1 = c α β r k + 1 , q k + 1 β = q k + 1 α + h 5 c α β r k + 1 . s_{k+1}=c_{\alpha\beta}r_{k+1},\qquad q_{k+1}^{\beta}=q_{k+1}^{\alpha}+h^5c_{\alpha\beta}r_{k+1}. s k + 1 = c α β r k + 1 , q k + 1 β = q k + 1 α + h 5 c α β r k + 1 .
Consequently, the one-step mismatch in the β \beta β representation is exactly
F ^ β , h ( T α β ( q k , r k ) ) − T α β ( F α , h ( q k , r k ) ) = ( h 5 c α β r k + 1 , 0 ) . (92) \widehat{F}_{\beta,h}(T_{\alpha\beta}(q_k,r_k))-T_{\alpha\beta}(F_{\alpha,h}(q_k,r_k))=(h^5c_{\alpha\beta}r_{k+1},0). \tag*{(92)} F β , h ( T α β ( q k , r k )) − T α β ( F α , h ( q k , r k )) = ( h 5 c α β r k + 1 , 0 ) . (92)
It is therefore nonzero whenever r k + 1 ≠ 0 r_{k+1}\ne0 r k + 1 = 0 , despite symplecticity of both local maps. For the numerical experiment, we use the conformal oscillator with local Hamiltonian (86) , taking c = 1 c=1 c = 1 in U α U_{\alpha} U α and c = 2 c=2 c = 2 in U β U_{\beta} U β , with k = 1 k=1 k = 1 , μ = 1 2 \mu=\frac{1}{2} μ = 2 1 , and initial LCS data ( q 0 , p 0 , x 0 , y 0 ) = ( 1 , 0 , 0 , 1 2 ) (q_0,p_0,x_0,y_0)=(1,0,0,\frac{1}{2}) ( q 0 , p 0 , x 0 , y 0 ) = ( 1 , 0 , 0 , 2 1 ) . In the oscillator coordinates ( q , r , x , s ) (q,r,x,s) ( q , r , x , s ) , the position vector is ( q , x ) (q,x) ( q , x ) and the momentum vector is ( r , s ) (r,s) ( r , s ) ; hence the added term in (91) becomes h 5 2 ( r k + 1 2 + s k + 1 2 ) \frac{h^5}{2}(r_{k+1}^2+s_{k+1}^2) 2 h 5 ( r k + 1 2 + s k + 1 2 ) in the β \beta β representation.
For the numerical comparison, we measure the one-step equivariance defect defined by the norm of (92) and the maximum discrepancy between the two independently computed trajectories, expressed in the same LCS variables, over t ∈ [ 0 , 2 ] t\in[0,2] t ∈ [ 0 , 2 ] ; see Figure 4 . The observed slopes are 4.95 and 4.00, respectively, consistent with the O ( h 5 ) O(h^5) O ( h 5 ) one-step defect in (92) and the O ( h 4 ) O(h^4) O ( h 4 ) accumulated discrepancy under the usual smoothness and stability assumptions. For comparison, the unmodified local discrete maps obtained with the Gauss–Legendre method satisfy the equivariance relation exactly.
Figure 4. Locally symplectic but incompatible discrete Hamiltonian families. The equivariance defect is O ( h 5 ) O(h^5) O ( h 5 ) per step, and the trajectory discrepancy over a fixed interval is O ( h 4 ) O(h^4) O ( h 4 ) . The horizontal line indicates the round-off level attained by the compatible GL2 family.
This construction shows that a family of locally symplectic maps can fail to define a single global numerical map when the local updates disagree on chart overlaps. Here the disagreement is introduced explicitly at the level of the discrete Hamiltonians, while the continuous Hamiltonians remain compatible. Although (92) depends explicitly on c α β c_{\alpha\beta} c α β , the added term also produces a mismatch when c α β = 1 c_{\alpha\beta}=1 c α β = 1 . Thus the experiment illustrates the failure of overlap compatibility, rather than a defect caused solely by a nontrivial conformal factor.
An example with a non-exact Lee form
We now consider an example with a genuinely non-exact Lee form, in order to illustrate the globalization mechanism in the presence of nontrivial monodromy. The purpose is not to compare different integration schemes or to repeat the convergence analysis, but to show how locally defined discrete symplectic maps are related across a nontrivial chart transition; we use the two-stage Gauss–Legendre method throughout.
Example 7.1. Consider the configuration manifold Q = S 1 × R Q = S^1 \times\mathbb{R} Q = S 1 × R and its cotangent bundle M = T ∗ ( S 1 × R ) M = T^*(S^1 \times\mathbb{R}) M = T ∗ ( S 1 × R ) , with points written as ( [ θ ] , p θ , x , p x ) ([\theta], p_\theta, x, p_x) ([ θ ] , p θ , x , p x ) , where [ θ ] ∈ S 1 = R / ( 2 π Z ) [\theta] \in S^1 = \mathbb{R}/(2\pi\mathbb{Z}) [ θ ] ∈ S 1 = R / ( 2 π Z ) denotes the angular equivalence class. Let V α = S 1 ∖ { [ π ] } V_\alpha= S^1 \setminus\{[\pi]\} V α = S 1 ∖ {[ π ]} and V β = S 1 ∖ { [ 0 ] } V_\beta= S^1 \setminus\{[0]\} V β = S 1 ∖ {[ 0 ]} , with angular coordinates θ α ∈ ( − π , π ) \theta_\alpha\in(-\pi,\pi) θ α ∈ ( − π , π ) and θ β ∈ ( 0 , 2 π ) \theta_\beta\in(0,2\pi) θ β ∈ ( 0 , 2 π ) , respectively. Using the cotangent-bundle projection π S 1 × R \pi_{S^1\times\mathbb{R}} π S 1 × R , define U α = π S 1 × R − 1 ( V α × R ) U_\alpha= \pi_{S^1\times\mathbb{R}}^{-1}(V_\alpha\times\mathbb{R}) U α = π S 1 × R − 1 ( V α × R ) and U β = π S 1 × R − 1 ( V β × R ) U_\beta= \pi_{S^1\times\mathbb{R}}^{-1}(V_\beta\times\mathbb{R}) U β = π S 1 × R − 1 ( V β × R ) . Let φ \varphi φ denote the closed angular one-form on S 1 S^1 S 1 , as well as its pullbacks to S 1 × R S^1 \times\mathbb{R} S 1 × R and M M M under the natural projections. Then φ ∣ U α = d θ α \varphi|_{U_\alpha} = d\theta_\alpha φ ∣ U α = d θ α and φ ∣ U β = d θ β \varphi|_{U_\beta} = d\theta_\beta φ ∣ U β = d θ β ; since ∮ S 1 φ = 2 π \oint_{S^1} \varphi= 2\pi ∮ S 1 φ = 2 π , the Lee form is closed but not globally exact, and the local Lee potentials are σ α = θ α \sigma_\alpha= \theta_\alpha σ α = θ α and σ β = θ β \sigma_\beta= \theta_\beta σ β = θ β . In local angular coordinates, the cotangent-type LCS form is represented by
Ω = d θ ∧ d p θ + d x ∧ d p x + p x d θ ∧ d x , \Omega= d\theta\wedge dp_\theta+ dx \wedge dp_x + p_x\,d\theta\wedge dx, Ω = d θ ∧ d p θ + d x ∧ d p x + p x d θ ∧ d x ,
where d θ d\theta d θ denotes the local representative of φ \varphi φ , and it satisfies d Ω = φ ∧ Ω d\Omega= \varphi\wedge\Omega d Ω = φ ∧ Ω . The overlap U α ∩ U β = π S 1 × R − 1 ( ( S 1 ∖ { [ 0 ] , [ π ] } ) × R ) U_\alpha\cap U_\beta= \pi_{S^1\times\mathbb{R}}^{-1}((S^1 \setminus\{[0],[\pi]\}) \times\mathbb{R}) U α ∩ U β = π S 1 × R − 1 (( S 1 ∖ {[ 0 ] , [ π ]}) × R ) has two connected components. On the upper semicircle, 0 < θ α = θ β < π 0 < \theta_\alpha= \theta_\beta< \pi 0 < θ α = θ β < π , so that σ α − σ β = 0 \sigma_\alpha- \sigma_\beta= 0 σ α − σ β = 0 and c α β = 1 c_{\alpha\beta} = 1 c α β = 1 . On the lower semicircle, the same point of S 1 S^1 S 1 is represented by − π < θ α < 0 -\pi< \theta_\alpha< 0 − π < θ α < 0 and π < θ β < 2 π \pi< \theta_\beta< 2\pi π < θ β < 2 π with θ α = θ β − 2 π \theta_\alpha= \theta_\beta- 2\pi θ α = θ β − 2 π , so that σ α − σ β = − 2 π \sigma_\alpha- \sigma_\beta= -2\pi σ α − σ β = − 2 π , i.e., C α β = − 2 π C_{\alpha\beta} = -2\pi C α β = − 2 π and c α β = e − 2 π c_{\alpha\beta} = e^{-2\pi} c α β = e − 2 π . Thus the two local angular coordinates describe the same geometric point, while their local Lee potentials differ by a nonzero constant on the second connected component of the overlap. For γ ∈ { α , β } \gamma\in\{\alpha,\beta\} γ ∈ { α , β } the local symplectic form is Ω γ = e − σ γ Ω = d θ γ ∧ d r γ + d x ∧ d s γ \Omega_\gamma= e^{-\sigma_\gamma}\Omega= d\theta_\gamma\wedge dr_\gamma+ dx \wedge ds_\gamma Ω γ = e − σ γ Ω = d θ γ ∧ d r γ + d x ∧ d s γ in terms of the local canonical momenta r γ = e − θ γ p θ r_\gamma= e^{-\theta_\gamma}p_\theta r γ = e − θ γ p θ and s γ = e − θ γ p x s_\gamma= e^{-\theta_\gamma}p_x s γ = e − θ γ p x , and on the nontrivial component of the overlap the transition between local canonical coordinates is
T α β ( θ α , r α , x , s α ) = ( θ α + 2 π , e − 2 π r α , x , e − 2 π s α ) . T_{\alpha\beta}(\theta_\alpha,r_\alpha,x,s_\alpha) = (\theta_\alpha+ 2\pi,e^{-2\pi}r_\alpha,x,e^{-2\pi}s_\alpha). T α β ( θ α , r α , x , s α ) = ( θ α + 2 π , e − 2 π r α , x , e − 2 π s α ) .
This is an affine transition: the geometric base point is unchanged, but its angular coordinate is translated by 2 π 2\pi 2 π . In the notation of (78) , the base coordinates are ( θ , x ) (\theta,x) ( θ , x ) , B = I 2 B=I_2 B = I 2 , b = ( 2 π , 0 ) ⊤ b=(2\pi,0)^\top b = ( 2 π , 0 ) ⊤ , and c α β = e − 2 π c_{\alpha\beta}=e^{-2\pi} c α β = e − 2 π .
We consider the globally defined Hamiltonian H ( [ θ ] , p θ , x , p x ) = p θ + 1 2 ( p x 2 + x 2 ) H([\theta],p_\theta,x,p_x)=p_\theta+\frac{1}{2}(p_x^2+x^2) H ([ θ ] , p θ , x , p x ) = p θ + 2 1 ( p x 2 + x 2 ) , whose definition is independent of the choice of angular lift. For γ ∈ { α , β } \gamma\in\{\alpha,\beta\} γ ∈ { α , β } , the local Hamiltonian is H γ = e − θ γ H = r γ + 1 2 e θ γ s γ 2 + 1 2 e − θ γ x 2 H_\gamma=e^{-\theta_\gamma}H=r_\gamma+\frac{1}{2}e^{\theta_\gamma}s_\gamma^2+\frac{1}{2}e^{-\theta_\gamma}x^2 H γ = e − θ γ H = r γ + 2 1 e θ γ s γ 2 + 2 1 e − θ γ x 2 , with local canonical Hamilton equations
θ ˙ γ = 1 , r ˙ γ = − 1 2 e θ γ s γ 2 + 1 2 e − θ γ x 2 , x ˙ = e θ γ s γ , s ˙ γ = − e − θ γ x . \dot{\theta}_\gamma=1,\qquad\dot{r}_\gamma=-\frac{1}{2}e^{\theta_\gamma}s_\gamma^2+\frac{1}{2}e^{-\theta_\gamma}x^2,\qquad\dot{x}=e^{\theta_\gamma}s_\gamma,\qquad\dot{s}_\gamma=-e^{-\theta_\gamma}x. θ ˙ γ = 1 , r ˙ γ = − 2 1 e θ γ s γ 2 + 2 1 e − θ γ x 2 , x ˙ = e θ γ s γ , s ˙ γ = − e − θ γ x .
which have the same form in both charts. These Hamiltonians satisfy H β ∘ T α β = e − 2 π H α H_\beta\circ T_{\alpha\beta}=e^{-2\pi}H_\alpha H β ∘ T α β = e − 2 π H α . Consequently, Proposition 5.15 applies to the two-stage Gauss–Legendre method on corresponding stage branches whose stages remain in the overlap.
For corresponding trajectories remaining in the nontrivial component of the overlap, Proposition 5.13 gives
H β , h + E ( θ β , 0 , x 0 ; r β , 1 , s β , 1 ) = e − 2 π H α , h + E ( θ β , 0 − 2 π , x 0 ; e 2 π r β , 1 , e 2 π s β , 1 ) + 2 π r β , 1 . H_{\beta,h}^{+E}\left(\theta_{\beta,0},x_0;r_{\beta,1},s_{\beta,1}\right)=e^{-2\pi}H_{\alpha,h}^{+E}\left(\theta_{\beta,0}-2\pi,x_0;e^{2\pi}r_{\beta,1},e^{2\pi}s_{\beta,1}\right)+2\pi r_{\beta,1}. H β , h + E ( θ β , 0 , x 0 ; r β , 1 , s β , 1 ) = e − 2 π H α , h + E ( θ β , 0 − 2 π , x 0 ; e 2 π r β , 1 , e 2 π s β , 1 ) + 2 π r β , 1 .
For approximate discrete right Hamiltonians, the analogous transformation law, including the same endpoint term, ensures compatibility by Corollary 5.14 .
In the original LCS variables, using a local lift of the angular coordinate, the equations (23) read θ ˙ = 1 \dot{\theta}=1 θ ˙ = 1 , p ˙ θ = p θ − 1 2 p x 2 + 1 2 x 2 \dot{p}_\theta=p_\theta-\frac{1}{2}p_x^2+\frac{1}{2}x^2 p ˙ θ = p θ − 2 1 p x 2 + 2 1 x 2 , x ˙ = p x \dot{x}=p_x x ˙ = p x and p ˙ x = p x − x \dot{p}_x=p_x-x p ˙ x = p x − x . For the initial condition ( θ 0 , p θ , 0 , x 0 , p x , 0 ) = ( 0 , 1 2 , 1 , 0 ) (\theta_0,p_{\theta,0},x_0,p_{x,0})=(0,\frac{1}{2},1,0) ( θ 0 , p θ , 0 , x 0 , p x , 0 ) = ( 0 , 2 1 , 1 , 0 ) the solution of the LCS dynamics is [ θ ( t ) ] = [ t ] [\theta(t)]=[t] [ θ ( t )] = [ t ] ,
x ( t ) = e t / 2 [ cos ( 3 2 t ) − 1 3 sin ( 3 2 t ) ] , x(t)=e^{t/2}\left[\cos\left(\frac{\sqrt{3}}{2}t\right)-\frac{1}{\sqrt{3}}\sin\left(\frac{\sqrt{3}}{2}t\right)\right], x ( t ) = e t /2 [ cos ( 2 3 t ) − 3 1 sin ( 2 3 t ) ] ,
p x ( t ) = − 2 3 e t / 2 sin ( 3 2 t ) , p_x(t)=-\frac{2}{\sqrt{3}}e^{t/2}\sin\left(\frac{\sqrt{3}}{2}t\right), p x ( t ) = − 3 2 e t /2 sin ( 2 3 t ) ,
and
p θ ( t ) = e t − 1 2 ( p x ( t ) 2 + x ( t ) 2 ) . p_\theta(t)=e^t-\frac{1}{2}\left(p_x(t)^2+x(t)^2\right). p θ ( t ) = e t − 2 1 ( p x ( t ) 2 + x ( t ) 2 ) .
Along this solution,
H ( [ θ ( t ) ] , p θ ( t ) , x ( t ) , p x ( t ) ) = p θ ( t ) + 1 2 ( p x ( t ) 2 + x ( t ) 2 ) = e t . H([\theta(t)],p_\theta(t),x(t),p_x(t))=p_\theta(t)+\frac{1}{2}\left(p_x(t)^2+x(t)^2\right)=e^t. H ([ θ ( t )] , p θ ( t ) , x ( t ) , p x ( t )) = p θ ( t ) + 2 1 ( p x ( t ) 2 + x ( t ) 2 ) = e t .
One complete revolution and chart switching. Since θ ˙ = 1 \dot{\theta}=1 θ ˙ = 1 , one revolution around the S 1 S^1 S 1 factor is completed over 0 ≤ t ≤ T = 2 π 0\leq t\leq T=2\pi 0 ≤ t ≤ T = 2 π . We take N = 128 N=128 N = 128 steps, so that h = π / 64 h=\pi/64 h = π /64 , and start from z α 0 = ( 0 , 1 2 , 1 , 0 ) z_\alpha^0=(0,\frac{1}{2},1,0) z α 0 = ( 0 , 2 1 , 1 , 0 ) in the local canonical coordinates of the α \alpha α -chart (at θ = 0 \theta=0 θ = 0 these coincide with the LCS variables). We switch to the β \beta β -chart at θ α = π / 2 \theta_\alpha=\pi/2 θ α = π /2 , where θ β = θ α \theta_\beta=\theta_\alpha θ β = θ α , r β = r α r_\beta=r_\alpha r β = r α and s β = s α s_\beta=s_\alpha s β = s α , continue in the β \beta β -chart until θ β = 3 π / 2 \theta_\beta=3\pi/2 θ β = 3 π /2 , where the nontrivial transition θ α = θ β − 2 π \theta_\alpha=\theta_\beta-2\pi θ α = θ β − 2 π , r α = e 2 π r β r_\alpha=e^{2\pi}r_\beta r α = e 2 π r β , s α = e − 2 π s β s_\alpha=e^{-2\pi}s_\beta s α = e − 2 π s β back to the α \alpha α -chart is applied, and compute the final part of the trajectory again in the α \alpha α -chart, so that [ θ ( T ) ] = [ θ ( 0 ) ] [\theta(T)]=[\theta(0)] [ θ ( T )] = [ θ ( 0 )] . The switching locations π / 2 \pi/2 π /2 and 3 π / 2 3\pi/2 3 π /2 are convenient choices inside the two connected components of the overlap; the necessity of changing local representation during a complete revolution is imposed by the topology of S 1 S^1 S 1 . The global momenta remain unchanged across the nontrivial transition, since e θ α r α = e θ β r β e^{\theta_\alpha}r_\alpha=e^{\theta_\beta}r_\beta e θ α r α = e θ β r β and e θ α s α = e θ β s β e^{\theta_\alpha}s_\alpha=e^{\theta_\beta}s_\beta e θ α s α = e θ β s β .
Monodromy of the discrete conformal factor. The complete revolution is precisely the situation addressed by (53) . Since θ n + 1 = θ n + h \theta_{n+1}=\theta_n+h θ n + 1 = θ n + h in every chart, the chart-wise increments of the local potentials are π / 2 \pi/2 π /2 over the first 32 steps in the α \alpha α -chart, π \pi π over the following 64 steps in the β \beta β -chart, and π / 2 \pi/2 π /2 over the last 32 steps in the α \alpha α -chart, so that
( F h 128 ) ∗ Ω = e 2 π Ω , (F_h^{128})^*\Omega=e^{2\pi}\Omega, ( F h 128 ) ∗ Ω = e 2 π Ω ,
although σ α ( z 128 ) − σ α ( z 0 ) = 0 \sigma_\alpha(z_{128})-\sigma_\alpha(z_0)=0 σ α ( z 128 ) − σ α ( z 0 ) = 0 . This is the discrete counterpart of ψ 2 π ∗ Ω = exp ( φ ) Ω = e 2 π Ω \psi_{2\pi}^*\Omega=\exp(\varphi)\Omega=e^{2\pi}\Omega ψ 2 π ∗ Ω = exp ( φ ) Ω = e 2 π Ω , and of the conformal law H ∘ ψ 2 π = e 2 π H H\circ\psi_{2\pi}=e^{2\pi}H H ∘ ψ 2 π = e 2 π H satisfied by the exact solution above. Both statements are confirmed numerically. The product J J J of the reconstructed one-step Jacobians over the revolution satisfies J ⊤ Ω ^ ( z 128 ) J = e 2 π Ω ^ ( z 0 ) J^\top\widehat{\Omega}(z_{128})J=e^{2\pi}\widehat{\Omega}(z_0) J ⊤ Ω ( z 128 ) J = e 2 π Ω ( z 0 ) , where Ω ^ \widehat{\Omega} Ω now denotes the matrix of Ω \Omega Ω in the LCS variables ( θ , p θ , x , p x ) (\theta,p_\theta,x,p_x) ( θ , p θ , x , p x ) , with a relative defect of 1.7 ⋅ 10 − 14 1.7\cdot10^{-14} 1.7 ⋅ 1 0 − 14 in the Frobenius norm, whereas the relative defect obtained by omitting the conformal factor is ∣ e 2 π − 1 ∣ ≃ 534.492 \lvert e^{2\pi}-1\rvert\simeq534.492 ∣ e 2 π − 1 ∣ ≃ 534.492 . Furthermore, the local energy e − θ γ H e^{-\theta_\gamma}H e − θ γ H is conserved to a relative accuracy of 10 − 8 10^{-8} 1 0 − 8 within each chart, is continuous across the trivial transition at θ = π / 2 \theta=\pi/2 θ = π /2 , and is multiplied by exactly e 2 π e^{2\pi} e 2 π at the nontrivial transition at θ = 3 π / 2 \theta=3\pi/2 θ = 3 π /2 , so that H ( z 128 ) / H ( z 0 ) = e 2 π ( 1 + 1.3 ⋅ 10 − 8 ) H(z_{128})/H(z_0)=e^{2\pi}(1+1.3\cdot10^{-8}) H ( z 128 ) / H ( z 0 ) = e 2 π ( 1 + 1.3 ⋅ 1 0 − 8 ) . The glued discrete map therefore reproduces the monodromy of the Lee form, which no single local potential can express.
One-step equivariance test. We next illustrate the discrete compatibility relation F β , h ∘ T α β = T α β ∘ F α , h F_{\beta,h}\circ T_{\alpha\beta}=T_{\alpha\beta}\circ F_{\alpha,h} F β , h ∘ T α β = T α β ∘ F α , h on the nontrivial component of the overlap. We choose a point in the lower component of the overlap, with local canonical coordinates z α = ( − π / 2 , 1 / 2 , 1 , 1 / 5 ) z_\alpha=(-\pi/2,1/2,1,1/5) z α = ( − π /2 , 1/2 , 1 , 1/5 ) in the α \alpha α -chart, so that the transition z β = T α β ( z α ) z_\beta=T_{\alpha\beta}(z_\alpha) z β = T α β ( z α ) , with θ β = θ α + 2 π \theta_\beta=\theta_\alpha+2\pi θ β = θ α + 2 π , r β = e − 2 π r α r_\beta=e^{-2\pi}r_\alpha r β = e − 2 π r α and s β = e − 2 π s α s_\beta=e^{-2\pi}s_\alpha s β = e − 2 π s α , is nontrivial. We take h = 0.05 h=0.05 h = 0.05 . We compare the state obtained by transforming to the β \beta β -chart and then advancing one step, F β , h ( T α β ( z α ) ) F_{\beta,h}(T_{\alpha\beta}(z_\alpha)) F β , h ( T α β ( z α )) , with the state obtained by advancing in the α \alpha α -chart and then transforming, T α β ( F α , h ( z α ) ) T_{\alpha\beta}(F_{\alpha,h}(z_\alpha)) T α β ( F α , h ( z α )) ; both are expressed in the β \beta β -representation, and we measure their discrepancy by the normalized equivariance defect (89) . For the two-stage Gauss–Legendre scheme with h = 0.05 h=0.05 h = 0.05 , we obtain δ h θ = 3.41 × 10 − 20 \delta_h^\theta=3.41\times10^{-20} δ h θ = 3.41 × 1 0 − 20 . This value lies below the unit round-off of double precision because the normalization is dominated by θ β ≈ 4.76 \theta_\beta\approx4.76 θ β ≈ 4.76 , whereas the momenta in the β \beta β -representation carry the factor e − 2 π e^{-2\pi} e − 2 π . Componentwise, the two routes agree exactly in θ \theta θ , r β r_\beta r β and x x x , and differ by approximately 2.5 × 10 − 15 2.5\times10^{-15} 2.5 × 1 0 − 15 in relative terms in s β s_\beta s β ; the same defect measured in the α \alpha α -representation, where all components are of order one, is 1.5 × 10 − 16 1.5\times10^{-16} 1.5 × 1 0 − 16 . Hence the two discrete routes agree to round-off accuracy, consistently with the theoretical equivariance relation. Together with the computation over one complete revolution, this shows how the local discrete symplectic maps are related across the nontrivial chart transition and illustrates the globalization of the discrete dynamics in the presence of a non-exact Lee form and nontrivial monodromy.
Conclusions
We have developed a Hamiltonian-first construction of structure-preserving integrators for locally conformally symplectic systems. The continuous LCS dynamics was first expressed through local canonical Hamiltonian representatives and a Type II variational principle, without taking hyperregularity as a prerequisite for the Hamiltonian formulation. This representation isolates the geometric data that must be respected by a discretization: ordinary symplectic dynamics in each local representative together with the constant fiber-scaling cocycle on overlaps. At the discrete level, approximate Type II generating functions produce local symplectic maps, and the transformation law of the exact discrete right Hamiltonian leads to a compatibility condition for its approximations which implies equivariance of the local discrete maps. The equivariant maps therefore glue to a globally defined discrete dynamics on the LCS phase space, and after reconstruction in the original variables, local symplectic preservation becomes the discrete conformal law
F h ∗ Ω = e σ α ∘ F h − σ α Ω , F_h^*\Omega= e^{\sigma_\alpha\circ F_h-\sigma_\alpha}\Omega, F h ∗ Ω = e σ α ∘ F h − σ α Ω ,
which is independent of the local conformal representative, and whose composition over several steps reproduces the periods of the Lee form when chart switching is required.
We have exhibited two practical mechanisms for obtaining compatible schemes: Type II Galerkin constructions whose discrete generating functions inherit the overlap scaling, and equivariant symplectic Runge–Kutta methods. The Gauss–Legendre family lies in both classes, with the implicit midpoint rule as its one-stage member. Under regularity, the same structure is recovered from compatible discrete Lagrangians through the discrete Legendre transformation, while the Hamiltonian construction remains available without assuming such a transformation globally. The numerical experiments of § 8 complement these results by testing convergence, chart switching, overlap compatibility, and the discrete conformal preservation law, and they separate clearly the two geometric requirements of an LCS integrator: preservation of the local symplectic structure and compatibility of the local discretizations on overlaps. Numerical integrators have also been developed for other geometric and variational settings, including contact geometry; see, for example, [2 , 8 , 9 , 44 ] , simple thermodynamic systems [3 ] , and non-autonomous Lagrangian systems [10 ] .
The main construction uses shared base coordinates, while affine base transitions are treated separately through the corresponding Type II transformation law, including the endpoint term arising from translations. Affine equivariance also ensures compatibility of symplectic Runge–Kutta methods under these transitions. Extending the construction to general nonlinear base transitions remains an open problem. The relation with conformally symplectic integrators for damped Hamiltonian systems [6 , 19 ] also deserves emphasis: in the present LCS framework, the conformal factor is determined by the Lee geometry rather than prescribed by a constant damping rate. In particular, exactness of the Lee form does not make the two settings equivalent.
Acknowledgments
Ayten Gezici acknowledges support from The Scientific and Technological Research Council of Türkiye (TUBITAK) under the 2219 International Postdoctoral Research Fellowship Program, Grant No. 1059B192500320. Melvin Leok was supported in part by the National Science Foundation under Grants CCF-2112665, DMS-2307801, and by the Air Force Office of Scientific Research under Grant FA9550-23-1-0279. Cristina Sardón acknowledges the Programa Propio of Universidad Politécnica de Madrid for financial support for a research stay at UCSD in the summer of 2025.