Jet Bundle Geometry in Variational Integrators

• Jet Bundle Geometry is a framework describing the behavior of field derivatives over a manifold, crucial for formulating variational integrators.
• It enables the derivation of discrete Euler–Lagrange equations that preserve structure, maintaining symplecticity and momentum maps.
• The approach supports high-order discretizations for complex systems, extending to Lie group dynamics, constraints, and stochastic models.

Variational integrators are a class of geometric numerical methods for mechanical systems whose defining property is derivation from a discretization of Hamilton’s principle of stationary action, rather than from direct discretization of the differential equations of motion. This approach yields numerical schemes that preserve key geometric invariants of the original system, such as symplecticity and momentum maps, and exhibit superior long-time energy behavior. Variational integrators have been systematically developed for finite- and infinite-dimensional systems, including general Lagrangian and Hamiltonian dynamics, Lie group systems, systems with constraints, time-dependent and stochastic systems, and applications across areas such as optimal control, plasma physics, electric circuits, and flexible multibody mechanics.

1. Foundations: Discrete Variational Principle and Construction

The fundamental construction is based on the discrete analog of Hamilton’s variational principle. Given a continuous Lagrangian system $(Q, L: TQ \rightarrow \mathbb{R})$, the discrete Lagrangian $L_d: Q \times Q \to \mathbb{R}$ approximates the fixed-time action integral

$$L_d(q_k, q_{k+1}) \approx \int_0^h L(q(t), \dot{q}(t))\,dt$$

along a solution of the Euler–Lagrange equations interpolating $q(0) = q_k$, $q(h) = q_{k+1}$. The discrete action is

$S_d(\{q_k\}) = \sum_{k=0}^{N-1} L_d(q_k, q_{k+1})$

and the discrete variational principle requires stationarity with respect to interior variations, yielding the discrete Euler–Lagrange (DEL) equations:

$D_2 L_d(q_{k-1}, q_k) + D_1 L_d(q_k, q_{k+1}) = 0$

with $D_i$ denoting partial differentiation with respect to the $i$th argument. The resulting two-step map $(q_{k-1}, q_k) \mapsto (q_k, q_{k+1})$ defines the variational integrator (Leok et al., 2011).

Two central strategies for constructing high-order variational integrators are the Galerkin approach—replacing the space of interpolating curves by a finite-dimensional ansatz and then approximating the action by quadrature—and the shooting-based approach—solving the two-point boundary-value problem approximately using a one-step integrator and quadrature. The order of the resulting integrator is governed by the accuracy in matching the exact discrete Lagrangian, as quantified by variational error analysis (Leok et al., 2011, Schmitt et al., 2017).

For systems evolving on Lie groups, the interpolating curves and discretizations are constructed in local charts or using group retractions (exponential or Cayley map), leading to Lie group variational integrators (Hall et al., 2014, Colombo et al., 8 Feb 2025).

2. Structure-Preservation: Symplecticity, Momentum, and Energy Behavior

A distinguishing feature of variational integrators is their exact preservation of a discrete symplectic form. The DEL update map $L_d: Q \times Q \to \mathbb{R}$0 satisfies

$L_d: Q \times Q \to \mathbb{R}$1

where $L_d: Q \times Q \to \mathbb{R}$2 is the discrete symplectic form built from the discrete Poincaré–Cartan one-forms (Hall et al., 2012). This discrete symplecticity leads to bounded numerical energy error over exponentially long times; the numerical energy typically exhibits bounded oscillations rather than systematic drift (Leok et al., 2011, Hall et al., 2012).

If the discrete Lagrangian is invariant under a group action (e.g., for mechanical systems with symmetries), a discrete version of Noether’s theorem guarantees preservation of discrete momentum maps:

$L_d: Q \times Q \to \mathbb{R}$3

which remain constant along the discrete flow (Leok et al., 2011, Hall et al., 2014).

In time-dependent and forced/dissipative systems, backward error analysis of the symplectic discrete flow in extended phase space shows that discrete energies related to the modified Hamiltonian are conserved up to exponentially small errors (Colombo et al., 2022).

3. High-Order and Spectral Variational Integrators

By varying the interpolation trial space and the quadrature scheme, variational integrators admit systematic construction to arbitrarily high order. Spectral (Galerkin) variational integrators use polynomial or spectral approximations in each time step; under appropriate analytic assumptions, this yields geometric (exponential in degree) convergence rates in both trajectory and invariant errors (Hall et al., 2012). For systems on Lie groups, analogous Lie group spectral integrators have been developed (Hall et al., 2014).

In classical (flat) mechanical systems, equivalent high-order variational integrators can also be recast as continuous-stage or partitioned Runge–Kutta (csPRK, SPRK) methods that match the variational structure and order-optimality (Tang, 2018, Leok et al., 2010).

Taylor variational integrators exploit high-order Taylor expansions both in solving the boundary value problem (shooting formulation) and evaluating quadrature nodes—resulting in integrators that, for given computational cost, can achieve one order higher than traditional shooting integrators (Schmitt et al., 2017).

4. Constraints, Extensions, and Generalizations

Variational integrators generalize systematically to systems with holonomic, nonholonomic, or higher-order constraints through the discrete Lagrange multiplier and augmented discrete action principle (Colombo et al., 2013). For higher-order Lagrangian systems, the discrete phase space is $L_d: Q \times Q \to \mathbb{R}$4, and the variational principle yields higher-order discrete Euler–Lagrange equations that are symplectic and momentum-preserving (Colombo et al., 2013).

For nonvariational or degenerate systems (e.g., certain electric circuits, advection equations, or underactuated mechanical systems), the framework extends by constructing a formal (adjoint) Lagrangian, embedding the dynamics into a larger system that admits a Lagrangian structure, then applying the variational discretization to the extended system (Ober-Blöbaum et al., 2011, Kraus et al., 2014, Kraus, 2013).

Non-autonomous (time-dependent) Lagrangian and Hamiltonian systems are accommodated by introducing a family of discrete Lagrangians $L_d: Q \times Q \to \mathbb{R}$5 indexed by time step, with symplecticity demonstrated in the extended (time, configuration) space (Colombo et al., 2022).

5. Stochastic and Dissipative Generalizations

Stochastic variational integrators are constructed by discretizing stochastic action functionals (with Stratonovich integrals) and producing discrete Lagrange–d’Alembert principles. Such integrators preserve symplecticity, momentum, and yield improved long-time statistical behavior compared to non-geometric integrators (Kraus et al., 2019, Holm et al., 2016). For systems with multiplicative noise, the construction yields stochastic symplectic Runge–Kutta methods as special cases (Holm et al., 2016). Applications include Vlasov–Fokker–Planck equations and collisional kinetic plasma models [1904