Discretization of polynomial vector fields by polarization

Elena Celledoni, Robert I. McLachlan, David McLaren, Brynjulf Owren, G. Quispel

TL;DR

This paper generalizes Kahan's novel integration method for quadratic vector fields to higher-degree polynomial vector fields. The authors demonstrate that this generalized "polar map" preserves important geometric properties like modified measures and energies, extending the utility of Kahan's approach to a broader class of dynamical systems.

What Problem Does This Paper Address?

Numerical integration of ordinary differential equations (ODEs) has traditionally focused on general-purpose methods like Runge-Kutta. However, there's a growing interest in "geometric integration methods" that preserve specific qualitative features of the underlying dynamical system, such as symmetries, first integrals, or conserved quantities. Kahan's method, introduced in 1993, is a prime example for quadratic vector fields, known for preserving certain conserved quantities and invariant measures, particularly for Hamiltonian systems. The challenge this paper addresses is extending these beneficial geometric properties beyond quadratic systems to higher-degree polynomial vector fields, which are common in various scientific models.

Methods

The core of the paper introduces a generalization of Kahan's method, termed the "polar map." This method leverages the concept of polarization, a mathematical technique that transforms a homogeneous polynomial into a symmetric multilinear form. For a polynomial vector field of degree $k+1$, its polarization results in a multilinear form in $k+1$ variables. The generalized Kahan method (the polar map) then discretizes the ODE by evaluating this multilinear form at $k+1$ consecutive time steps.

Key aspects of the method include:

  • Polarization: This is the central mathematical tool used to generalize Kahan's approach. It allows for the representation of polynomial vector fields in a multilinear form.
  • Polar Map Definition: For a homogeneous polynomial ODE $\dot{x} = f(x)$ of degree $k+1$, the polar map is defined by a relation involving the polarized form $F$ of $f$: $\frac{x_k - x_0}{kh} = F(x_0, \dots, x_k)$. This is a $k$-step numerical integrator.
  • General Linear Methods: The paper shows that the polar map is a specific instance of a general linear method when restricted to homogeneous polynomial vector fields.
  • Geometric Properties: The authors investigate the preservation of modified measures and $k$-integrals (generalized first integrals) under this polar map, analogous to the properties observed for Kahan's original method.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and analysis of methods for ODEs, rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Polarization as a Unifying Principle: The paper elegantly demonstrates how the concept of polarization can be used to generalize Kahan's method from quadratic to any polynomial vector field. This provides a systematic way to construct structure-preserving discretizations.
  • Generalization of Kahan's Method: The "polar map" is a natural extension of Kahan's method, carrying over its core idea of using a specific algebraic structure to define the numerical scheme.
  • Preservation of Geometric Properties: A significant finding is that the polar map preserves modified versions of invariant measures and first integrals, even for higher-degree polynomial systems. This is crucial for long-term simulations where such properties are important.
  • Connection to General Linear Methods: Establishing that the polar map is a type of general linear method provides a broader context and allows for leveraging existing theory on these methods.
  • Integrability for Specific Cases: The paper shows that for certain cases (e.g., $k=2, n=2$ for a specific Hamiltonian), the polar map leads to completely integrable systems, highlighting its potential for analytical and numerical tractability.

Key Results

  • The paper defines the "polar map" as a generalization of Kahan's method for polynomial vector fields of degree $k+1$.
  • It proves that the polar map is a $k$-step numerical integrator and a specific case of general linear methods for homogeneous polynomial vector fields.
  • Proposition 4 establishes that for homogeneous Hamiltonian vector fields, the polar map possesses $k$ independent $k$-integrals, which approximate modified energies of the system.
  • Proposition 5 shows that the polar map preserves a modified invariant measure, which converges to the ODE's invariant measure in the limit of small time steps.
  • Proposition 6 demonstrates a $(k-1)$-dimensional symmetry group for the polar map when $k > 1$, leading to integrability properties.
  • Example 1 explicitly constructs the polar map for $k=2, n=2$, showing it's a birational map of degree 3 and has two 2-integrals and an invariant measure.

Limitations and Future Work

  • Integrability Not Guaranteed: While the paper shows integrability for specific cases (like $k=2, n=2$), it notes that for higher degrees ($k > 2$) or dimensions ($n > 2$), the geometric properties are not always sufficient to guarantee complete integrability, as indicated by the failure of the "entropy test" for a specific quintic Hamiltonian.
  • Focus on Homogeneous Polynomials: The primary focus is on homogeneous polynomial vector fields. The generalization to non-homogeneous polynomials is mentioned as a possibility (e.g., by suspending the problem to a higher dimension), but the detailed analysis is for the homogeneous case.
  • Future Work: The authors suggest exploring these phenomena further for more general classes of vector fields and investigating the implications of the discovered symmetries and integrability for numerical simulations. The paper also hints at the need for further understanding of the "explanations" for why these methods work so well, echoing Kahan's original sentiment.

Summary updated

Abstract

A novel integration method for quadratic vector fields was introduced by Kahan in 1993. Subsequently, it was shown that Kahan's method preserves a (modified) measure and energy when applied to quadratic Hamiltonian vector fields. Here we generalize Kahan's method to cubic resp. higher degree polynomial vector fields and show that the resulting discretization also preserves modified versions of the measure and energy when applied to cubic resp. higher degree polynomial Hamiltonian vector fields.
BibTeX
@inproceedings{celledoni2015discretization,
  author = {Elena Celledoni and Robert I. McLachlan and David McLaren and Brynjulf Owren and G. Quispel},
  title = {{Discretization of polynomial vector fields by polarization}},
  booktitle = {Proceedings of the Royal Society A Mathematical Physical and Engineering Sciences},
  year = {2015},
  doi = {10.1098/rspa.2015.0390},
}
Figure