Condition number

The condition number of a matrix is a measure of how sensitive the solution of a linear system is to errors in the data. For a matrix $, the condition number with respect to inversion is $\kappa(A) = \|A\| \|A^{-1}\|$, which in the spectral norm equals the ratio of largest to smallest singular value: $\kappa(A) = \sigma_{\max}/\sigma_{\min}$. A condition number near 1 indicates a well-conditioned problem; a very large condition number signals an ill-posed problem where small perturbations in the input can produce catastrophic errors in the output.

In control theory, the condition number of a plant's transfer function matrix at a given frequency measures how differently the system responds to inputs in different directions. A high condition number means that some input directions are strongly amplified while others are nearly blocked, making control design difficult because no single controller can simultaneously handle both. The condition number is thus a diagnostic tool for identifying the structural limits of what control can achieve.

The condition number is numerical analysis's way of admitting that some problems are inherently harder than others — not because we lack clever algorithms, but because the problem's own geometry amplifies uncertainty. It is a humility metric. A model with a condition number of 0^6$ is not merely computationally difficult; it is epistemologically fragile, because the data required to solve it to three significant figures would need to be known to nine.