[Paper Review] Runge-Kutta Theory and Constraint Programming

Julien Alexandre dit Sandretto|arXiv (Cornell University)|Apr 13, 2018

Numerical Methods and Algorithms16 references3 citations

TL;DR

This paper proposes a novel constraint programming approach using interval analysis to compute Runge-Kutta coefficients with guaranteed properties, ensuring stability, symplecticity, and order accuracy. By formulating Butcher tableau conditions as a constraint satisfaction problem and optimizing via a global cost function, the method generates new interval-coefficient Runge-Kutta schemes that outperform traditional floating-point methods in validated simulations with tighter solution enclosures.

ABSTRACT

There exist many Runge-Kutta methods (explicit or implicit), more or less adapted to specific problems. Some of them have interesting properties, such as stability for stiff problems or symplectic capability for problems with energy conservation. Defining a new method suitable to a given problem has become a challenge. The size, the complexity and the order do not stop growing. This informal challenge to implement the best method is interesting but an important unsolved problem persists. Indeed, the coefficients of Runge-Kutta methods are harder and harder to compute, and the result is often expressed in floating-point numbers, which may lead to erroneous integration schemes. Here, we propose to use interval analysis tools to compute Runge-Kutta coefficients. In particular, we use a solver based on guaranteed constraint programming. Moreover, with a global optimization process and a well chosen cost function, we propose a way to define some novel optimal Runge-Kutta methods.

Motivation & Objective

• To address the growing difficulty in computing exact Runge-Kutta coefficients using floating-point arithmetic, which risks introducing errors in numerical integration.
• To preserve key properties like order, stability, and symplecticity in numerical schemes by replacing floating-point coefficients with interval coefficients.
• To develop a systematic, verified method for discovering new Runge-Kutta methods using constraint programming and global optimization.
• To demonstrate the effectiveness of interval-coefficient schemes in validated simulation environments with guaranteed solution enclosures.
• To provide a framework for generating optimal Runge-Kutta methods with interval coefficients that maintain theoretical properties under rigorous computation.

Proposed method

• Formulates the conditions for Runge-Kutta methods (order conditions, stability, symplecticity) as a system of nonlinear equations and inequalities in interval variables.
• Uses a guaranteed constraint programming (CP) solver to find interval boxes satisfying all Butcher tableau conditions, ensuring all solutions are valid and inclusion-wise correct.
• Applies a global optimization procedure with a cost function designed to minimize the local truncation error (LTE) relative to the next higher-order method, favoring optimal schemes.
• Employs interval arithmetic throughout the construction and validation process to maintain rigorous error bounds and property guarantees.
• Implements the resulting schemes in the DynIbex validated simulation tool for experimental evaluation on benchmark problems.
• Validates results by comparing solution enclosures and computational efficiency against classical methods like RadauIIA, Gauss-Legendre, and RK4.

Experimental results

Research questions

• RQ1 Can interval coefficients in Runge-Kutta methods preserve essential properties such as order, stability, and symplecticity, which are often lost with floating-point arithmetic?
• RQ2 Can constraint programming be effectively used to solve the high-dimensional, under-constrained system of equations arising from Butcher tableau conditions with interval coefficients?
• RQ3 Does optimizing the local truncation error (LTE) relative to the next higher-order method yield Runge-Kutta schemes that are competitive in accuracy and efficiency?
• RQ4 Can interval-coefficient schemes produce tighter solution enclosures in validated simulation than their floating-point counterparts, even with small coefficient intervals?
• RQ5 To what extent do interval-based methods improve contractivity and robustness in the presence of parameter uncertainty or stiff dynamics?

Key findings

• The S3O4 method, a third-order scheme with interval coefficients, achieved performance comparable to the fifth-order RadauIIA method but with faster computation, demonstrating high efficiency.
• The S3O5 method, a third-order scheme with interval coefficients, was five times faster than the sixth-order Gauss-Legendre method while producing a significantly narrower final solution enclosure.
• The ERK33 method, an explicit third-order scheme with interval coefficients, consumed similar time to RK4 but produced a solution enclosure only 1.2 times wider than RK4, indicating strong competitiveness.
• Despite interval coefficients with diameters as small as 1×10⁻¹⁰, the final solution enclosures of the new methods were often narrower than those of classical methods with exact coefficients, suggesting improved contractivity.
• The constraint programming approach successfully generated new Runge-Kutta schemes that preserve order and stability properties, even when coefficients are represented as intervals with guaranteed inclusion.
• The optimization process effectively minimized LTE, resulting in schemes that approach the performance of higher-order methods, validating the cost function's effectiveness.