What does it mean for a method to be convergent?

A method is convergent if its computed solution approaches the exact solution as the step size goes to zero. By the fundamental equivalence theorem this happens precisely when the method is both consistent (locally accurate) and stable (errors do not blow up).

Why are there so many different ODE methods?

Different problems prioritize different things: high accuracy, low cost per step, low memory, or robustness to stiffness. Runge-Kutta, multistep, explicit, and implicit families each occupy a different point in these trade-offs, so no single method is best for all problems.

Numerical Solution of Ordinary Differential Equations

This area develops and analyses time-stepping methods that approximate the solution of ordinary differential equations, advancing an initial state forward step by step while controlling accuracy and stability.

Definition

The numerical solution of ordinary differential equations is the construction and analysis of algorithms that produce approximate solutions to differential equations with given initial (or boundary) conditions by discretizing the independent variable.

Scope

It covers initial value problems for systems of ODEs solved by one-step (Runge-Kutta) and multistep methods, the concepts of consistency, stability, and convergence (the Dahlquist theory), error control through adaptive step-size selection, and the special treatment required for stiff problems; boundary value problems and geometric integrators are treated as extensions.

Sub-topics

Core questions

How is a continuous differential equation discretized into a stable, convergent time-stepping scheme?
What is the relationship between consistency, stability, and convergence for these methods?
How is the step size chosen adaptively to meet an accuracy requirement efficiently?
Why do stiff problems demand implicit methods, and how is stiffness characterized?

Key theories

Consistency, stability, and convergence: A method converges to the true solution as the step size tends to zero if and only if it is consistent (accurate to leading order) and stable (does not amplify errors uncontrollably); this Lax-type equivalence, made precise for multistep methods by Dahlquist, is the organizing principle of the field.
One-step versus multistep methods: One-step (Runge-Kutta) methods use only the current state but several internal stages, whereas multistep methods reuse several past values; each family trades implementation complexity, memory, and stability differently.
Adaptive error control: Embedded method pairs provide an estimate of the local truncation error at each step, which is used to accept or reject the step and to adjust the step size so that a prescribed tolerance is met with minimal work.

Clinical relevance

ODE solvers are fundamental modelling tools across science and engineering: they integrate the equations of motion in mechanics and astronomy, reaction kinetics in chemistry and systems biology, circuit and control-system dynamics, and population and epidemiological models; the reliability of such simulations depends directly on the accuracy and stability of the time-integration method chosen.

History

Classical one-step methods were developed by Runge and Kutta around 1900 and multistep methods by Adams, Bashforth, and Moulton; the modern theory was unified by Germund Dahlquist's mid-twentieth-century results on stability and order barriers and by John Butcher's algebraic theory of Runge-Kutta methods, with stiff-problem solvers following in the 1960s and 1970s.

Key figures

Carl Runge
Wilhelm Kutta
Germund Dahlquist
John C. Butcher

Seminal works

hairer1993
iserles2008
butcher2016

Frequently asked questions

What does it mean for a method to be convergent?: A method is convergent if its computed solution approaches the exact solution as the step size goes to zero. By the fundamental equivalence theorem this happens precisely when the method is both consistent (locally accurate) and stable (errors do not blow up).
Why are there so many different ODE methods?: Different problems prioritize different things: high accuracy, low cost per step, low memory, or robustness to stiffness. Runge-Kutta, multistep, explicit, and implicit families each occupy a different point in these trade-offs, so no single method is best for all problems.