A numerical measure of how sensitively the output of a function or solution of a linear system responds to small relative perturbations of the input; for invertible matrices under a chosen norm it is typically ||A|| ||A^{-1}||, and for matrices equals the ratio of largest to smallest singular value in the spectral norm.

Principle

The condition number quantifies worst-case relative amplification of input error into output error for the problem and norm specified; it separates problem conditioning from algorithmic stability and depends on both operator and norm choice.

Demonstration

For a 2×2 diagonal matrix diag(100,1) under spectral norm the condition number is 100/1=100, meaning a 1% relative perturbation in input can cause up to ~100% relative change in the solution. For a singular matrix the condition number is infinite, signaling no stable inverse.

Misapplication

Omitting to specify the norm when quoting a condition number, or conflating a high condition number with algorithmic failure regardless of solver design. Using the matrix 2-norm condition number to predict errors for an algorithm whose stability is governed by a different norm may mislead.

Consequence

A large condition number indicates ill-conditioning: solutions are highly sensitive and require regularization, higher precision, or reformulation; a small condition number (near 1) indicates a well-conditioned problem where forward errors remain commensurate with input errors.

Reversal

Interpreting a low condition number as guarantee of accurate computed solution regardless of algorithmic behavior; the inverse statement is that good conditioning does not replace the need for numerically stable algorithms.

Boundary

Defined relative to a specified norm and problem (matrix inversion, linear solve, function evaluation). For nonlinear problems a local condition number uses the Jacobian; for noninvertible operators the classical condition number is infinite and must be replaced by condition measures on subspaces or regularized variants.

Semantic Tension

Tension arises between problem conditioning (intrinsic sensitivity measured by condition number) and numerical stability (algorithmic error amplification): a well-conditioned problem can be solved badly, and an ill-conditioned problem may be handled by specialized regularization or stabilized algorithms.

Synthesis

The condition number is a norm-dependent scalar that measures worst-case relative amplification of input perturbations by a problem; for matrices it equals ||A||·||A^{-1}|| and, in the spectral norm, the ratio of largest to smallest singular value, guiding expectations about sensitivity and the need for stabilization.