A matrix's condition number measures how much errors in inputs get amplified
In short: Solving a system of linear equations can magnify tiny input errors into large output errors. The condition number, a single figure computed from a matrix, tells you how much that amplification to expect. Small values mean stable computation; large values warn that the answer may be unreliable.
When a system of linear equations is solved, the input is the right-hand side vector and the matrix, and the output is the solution. If the input changes slightly, the solution changes too. The condition number quantifies how strongly that change is amplified. For a square nonsingular matrix, it is defined as the product of the norm of the matrix and the norm of its inverse. The relative error in the computed solution can then be bounded by the condition number multiplied by the relative error in the input. A matrix with a large condition number is called ill-conditioned, and one with a small condition number is called well-conditioned. The condition number can never be less than 1, and the identity matrix achieves that minimum. Orthogonal matrices also have optimal conditioning in the 2-norm. The condition number is closely tied to how near a matrix is to being singular: a large value means the matrix is nearly singular. The determinant, by contrast, is not a reliable indicator of near-singularity.
The value depends on the chosen norm. For the 1-norm, it uses maximum absolute column sums; for the infinity-norm, maximum absolute row sums; and for the 2-norm, the ratio of the largest to smallest singular values. Because the norm must be specified, the number is always stated with its norm. A related but distinct idea is the relative residual, which measures how well an approximate solution satisfies the system. A small relative residual implies a small relative error only when the matrix is well-conditioned. The condition number is a property of the problem itself rather than of the algorithm used to solve it. Algorithms with a property called backward stability can be expected to solve well-conditioned problems accurately. Gaussian elimination with partial pivoting produces a small residual in practice, although that guarantee depends on the growth of intermediate values. An ill-conditioned problem does not make the solution impossible to find, but it means that small errors in the data can strongly affect the answer.
Practitioners therefore check the condition number before trusting a computed solution, especially when the data come from measurements with uncertainty.
Opens the Intellect+ feed with this card first. Swipe up for the next one.
Sources
- Summary by Intellect · source: Condition Numbers · CS 357 Textbook
- Also: Condition number
To cite this card, link to https://intellectplus.co.uk/app/science-condition-number.
How we source, check and correct cards: Editorial standards.
General information, not professional advice. Spotted an error? Flag the card or see corrections and complaints.
Intellect Plus is a visual educational discovery platform for exploring fascinating facts and ideas. Brain Food, Not Brain Rot. What is Intellect Plus?
Related
More in Algebra · More in Science · Latest · A–Z · All topics