EC-4.7 Numerical Methods: Roots, Interpolation, and Stepping an Equation Forward
The standard numerical methods treatment as taught in engineering mathematics courses, September 2026
What this is and why it exists
Most equations that describe real hardware have no closed-form solution. That is not a failure of anybody's algebra. It is the normal case, and a numerical answer is the usual way engineering questions get answered.
So the skill here is not implementing the algorithms. Every library implements them better than you will. The skill is knowing where the error comes from, and when a method is about to give you a confident wrong answer. A numerical result carries no warning label.
This topic is deliberately about that judgement.
The vocabulary
- Truncation error — error from the method itself, because it stops an infinite process early.
- Rounding error — error from arithmetic being carried out to finite precision.
- Convergence — whether repeated application of a method approaches the answer.
- Bracketing — knowing two points between which a root must lie.
- Conditioning — how much a problem amplifies small errors in its inputs.
- Condition number — a measure of that amplification for a linear system.
- Interpolation — constructing a function that passes through given points.
- Quadrature — numerical integration.
- Step size — how far a method advances in one move when stepping an equation forward.
The mental model
Begin with the two errors, because everything else is a trade between them.
Truncation error comes from the method. It shrinks as the step size shrinks. Rounding error comes from the arithmetic, and it grows as the step size shrinks, because more steps mean more operations and more accumulated noise. The two run in opposite directions, so there is a best step size and going below it makes the answer worse. People who have never been told this reduce the step until the computer is slow and the answer is wrong.
Root finding. Bisection needs two points where the function has opposite signs, then halves the interval repeatedly. It always converges if you can bracket the root, but slowly. Newton's method uses the derivative to jump toward the root and converges very fast when it works. It can also diverge entirely from a poor starting point, or stall where the derivative is near zero. The choice between them is a judgement about what you know in advance: bracketing buys reliability, a good starting guess buys speed.
Linear systems. Elimination with partial pivoting solves the system, and the pivoting is not an optimisation. It is what stops a small pivot from amplifying rounding error catastrophically. But no amount of careful solving rescues a poorly conditioned matrix. A large condition number means small changes in the input produce large changes in the output, and the answer was uncertain before you started. The condition number is the warning, and it comes from the linear algebra.
Interpolation. The obvious idea is to fit one polynomial through all the points. Do that with many points and the result oscillates wildly near the ends of the interval, sometimes far outside the range of the data. This is why practice uses piecewise fits instead, joining low-order pieces smoothly. Meeting that failure once stops you reaching for the obvious method later.
Quadrature. The trapezium rule joins sampled points with straight lines and adds the areas. Simpson's rule fits a curve through each set of three points instead, and is markedly more accurate for the same samples. The error orders tell you what halving the spacing actually buys, which is the number worth knowing.
Stepping an equation forward. Euler's method takes the derivative at the current point and moves in a straight line. It is the simplest thing that could work and it accumulates error quickly. Runge-Kutta methods sample the derivative at several points within the step and combine them. That buys a great deal of accuracy for a modest amount of extra work. This is why the fourth-order Runge-Kutta method is the default almost everywhere.
What you should now be able to explain or do
- Name the two sources of error and say why reducing the step size does not always help.
- Choose between bisection and Newton's method and defend the choice.
- Say what conditioning is and why a well-implemented solver cannot fix it.
- Explain why fitting one high-order polynomial through many points is a bad idea.
- Compare Euler and Runge-Kutta and say what the extra evaluations buy.
Check yourself
You halve the step size and the answer gets worse. What has happened?
Rounding error now dominates. More steps mean more accumulated arithmetic noise, and that grows as truncation error falls.
When would you choose bisection over Newton's method?
When you can bracket the root but have no good starting guess, or where the derivative is unreliable. Bisection is slower but always converges.
A linear system has a large condition number. What does that mean for your answer?
Small errors in the inputs are amplified into large errors in the output. The problem itself is sensitive, and a better solver does not repair that.
Why does fitting a single high-order polynomial through many data points fail?
It oscillates violently near the ends of the interval, often leaving the range of the data entirely. Piecewise low-order fits are used instead.
What does Runge-Kutta buy over Euler's method?
Much greater accuracy for the same step size, by sampling the derivative at several points within each step rather than only at the start.
Go deeper
We haven't checked most of these for screen reader use yet.