General techniques for constructing variational integrators

Melvin Leok, Tatiana Shingel

TL;DR

This paper provides a systematic framework for building high-order variational integrators—numerical solvers that preserve the fundamental geometric properties of physical systems. By leveraging standard tools like numerical quadrature and one-step methods, the authors demonstrate how to construct integrators that are not only highly accurate but also guarantee the conservation of momentum and near-conservation of energy over long simulation times.

What Problem Does This Paper Address?

In classical physics, systems are often described by a Lagrangian, and their motion follows the "Principle of Least Action." Standard numerical integrators (like the common Runge-Kutta methods) often ignore this underlying geometry. Over long simulations, these standard methods tend to "drift," artificially adding or removing energy and failing to preserve physical invariants like angular momentum.

Variational integrators solve this by discretizing the action principle itself rather than the differential equations. However, constructing high-order variational integrators has historically been seen as a specialized, somewhat abstract task. This paper aims to bridge the gap between traditional numerical analysis and variational mechanics, providing "recipes" to build these integrators using familiar components like polynomial interpolation and standard one-step solvers.

Methods

The authors propose two primary "construction kits" for discrete Lagrangians:

  1. The Galerkin Approach: This method replaces the infinite-dimensional space of possible paths with a finite-dimensional subspace (like polynomials). To compute the "action" along these paths, the authors use numerical quadrature (weighted sums of function values). A key contribution here is the extension to Lie Groups (like the space of rotations), where they use "natural charts" and the exponential map to ensure the simulation stays on the correct geometric manifold (e.g., keeping a rotation matrix perfectly orthogonal).

  2. The Shooting-Based Approach: This is a more "plug-and-play" method. It takes an existing, non-variational one-step method (like a standard 4th-order Runge-Kutta) and uses it to approximate the path between two points. By then applying a quadrature rule to the action along that path, they "wrap" the standard method in a variational framework. This allows the resulting integrator to inherit the computational efficiency of the underlying solver while gaining the stability of a variational approach.

Datasets

As a theoretical and numerical analysis paper, it does not use external datasets. Instead, the authors evaluate their methods on classic benchmark problems in mechanics:

  • The Free Rigid Body: Testing motion on Lie groups (SO(3)).
  • The Planar Pendulum: A standard nonlinear Hamiltonian system.
  • The Simple Harmonic Oscillator: A basic test for energy stability.

Key Insights and Ideas

  • Inherited Accuracy: The authors prove that the order of accuracy of a shooting-based variational integrator is simply the minimum of the accuracy of the underlying solver ($p$) and the quadrature rule ($q$). This makes the design process highly predictable.
  • Symmetry Matters: If you use a self-adjoint (symmetric) one-step method and a symmetric quadrature rule, the resulting variational integrator is also symmetric, which automatically grants it an even order of accuracy and better long-term stability.
  • Efficiency Gains: They show that a Lie Group Velocity Verlet (LGVI) method can be constructed that requires only one force evaluation per time-step, making it significantly faster than other Lie group methods that require multiple evaluations.
  • The "Aha" Moment: The realization that one doesn't need to reinvent the wheel to get high-order variational properties; one can "recycle" existing high-quality numerical solvers by using them as the basis for a shooting-based discrete Lagrangian.

Key Results

  • Geometric Fidelity: In the rigid body example, the Lie Group Variational Integrator (LGVI) maintained the orthogonality of rotation matrices to machine precision ($10^{-15}$), whereas standard methods showed significant drift as the step size decreased due to round-off accumulation.
  • Energy Stability: While variational integrators don't preserve energy exactly, they preserve a "modified Hamiltonian." The paper demonstrates that for the pendulum and oscillator, the energy error remains bounded and oscillates near zero for long durations, whereas non-variational methods (like explicit Midpoint) show a clear trend of energy error growth.
  • Computational Speed: The LGVI outperformed the Crouch-Grossman and Implicit Midpoint methods in CPU time vs. error benchmarks, largely due to its efficient handling of the Lie group structure.

Limitations and Future Work

  • Boundary Value Problems: The shooting-based approach requires solving a small boundary value problem at each step to find the "initial velocity" that connects two points. While the authors provide efficient iterative solvers (like Newton's method using Cayley transforms), this is still more complex than a simple explicit step.
  • Complexity of Implementation: While the framework is systematic, implementing the higher-order derivatives required for some quadrature rules (like Euler-Maclaurin) can be mathematically tedious.
  • Future Directions: The authors suggest these techniques could be extended to more complex scenarios, such as multiphysics systems, stochastic variational integrators, or systems with non-smooth collisions.

Summary updated

Abstract

The numerical analysis of variational integrators relies on variational error analysis, which relates the order of accuracy of a variational integrator with the order of approximation of the exact discrete Lagrangian by a computable discrete Lagrangian. The exact discrete Lagrangian can either be characterized variationally, or in terms of Jacobi's solution of the Hamilton-Jacobi equation. These two characterizations lead to the Galerkin and shooting-based constructions for discrete Lagrangians, which depend on a choice of a numerical quadrature formula, together with either a finite-dimensional function space or a one-step method. We prove that the properties of the quadrature formula, finite-dimensional function space, and underlying one-step method determine the order of accuracy and momentum-conservation properties of the associated variational integrators. We also illustrate these systematic methods for constructing variational integrators with numerical examples.
BibTeX
@inproceedings{leok2012general,
  author = {Melvin Leok and Tatiana Shingel},
  title = {{General techniques for constructing variational integrators}},
  booktitle = {Frontiers of Mathematics in China},
  year = {2012},
  doi = {10.1007/s11464-012-0190-9},
}

Loading related content...

Figure