Condition Number¶

Quantify the worst local amplification of admissible relative input perturbations into relative output changes for a specified mathematical problem, point, and choice of norms.

Version: v2 · 2026-09-06 · History
Domain-specific #: 1530
Origin domain: mathematics
Subdomain: numerical analysis
Aliases: Conditioning, Problem condition number, Relative condition number

Core Idea¶

A condition number is a numerical-analysis measure of how strongly a mathematical problem amplifies small changes in its data. The problem is represented by a solution map that sends admissible input data to an answer. At a specified input, one compares the size of a small perturbation of the data with the resulting change in the answer, then takes a local worst case over permitted perturbation directions. A relative condition number compares relative changes; an absolute condition number compares absolute changes. John Rice's general theory made the problem-and-data formulation explicit and separated the sensitivity of the mathematical task from the behavior of any particular implementation.^[1]

That separation is constitutive. Conditioning belongs to the posed problem, including its representation, scaling, metric, and evaluation point. Stability belongs to an algorithm: a backward-stable method returns the exact answer to a nearby problem. A stable algorithm can still produce a large forward error on an ill-conditioned problem because a small backward perturbation is then strongly amplified. Conversely, a well-conditioned problem does not rescue an unstable implementation. A complete accuracy analysis connects both layers rather than using condition number as a synonym for numerical failure.^[2]

For a differentiable map between normed spaces, a local condition number is governed by the derivative, with scaling factors determined by whether absolute or relative changes are measured. For the nonsingular linear system A x = b, when A is fixed and perturbations are normwise, the familiar matrix condition number is κ(A) = ||A|| ||A⁻¹||. Under the Euclidean norm it is the ratio of largest to smallest singular value. The value depends on the selected norm and on what part of the data is perturbed; componentwise, structured, right-hand-side, eigenvalue, and root-finding condition numbers answer different questions.

A large value is a warning about attainable accuracy, not a deterministic count of lost digits and not a certificate that every perturbation is amplified maximally. The definition is asymptotic and worst-case. A singular square matrix is conventionally assigned infinite inverse-problem condition because arbitrarily small changes can destroy or create solutions, but rectangular or rank-deficient problems require an explicitly stated generalized convention. Condition estimation is itself an algorithmic task, and a computed estimate must not be mistaken for the exact mathematical quantity.

Structural Signature¶

The mathematical problem. A solution map identifies which output is sought from which data.
The base input. Conditioning is evaluated at a point, matrix, or problem instance rather than floating without a reference.
The admissible perturbations. A declared class specifies which data components or structures may change.
The input size rule. A norm, metric, scaling, or componentwise convention measures the perturbation.
The output size rule. A compatible norm or metric measures the response in the answer.
The local limiting regime. Perturbation magnitude tends to zero so the measure captures first-order sensitivity.
The worst-direction operator. A supremum selects the largest response ratio over allowed directions.
The absolute-or-relative frame. Scaling by the base data and answer determines which interpretation is licensed.
The well/ill-conditioned classification. Magnitude summarizes whether data uncertainty is mildly or strongly amplified.
The algorithm boundary. Conditioning is combined with, but never substituted for, a separate stability analysis.

What It Is Not¶

Not numerical stability. Stability describes how an algorithm responds to rounding and computational error.
Not an algorithm's convergence rate. A slowly convergent method may solve a well-conditioned problem and vice versa.
Not a universal matrix attribute without a norm. Matrix condition numbers change with norm and perturbation model.
Not exact forward error. It is normally a local worst-case amplification factor used in a bound.
Not merely proximity to singularity in every problem. That interpretation is exact only under a specified distance, representation, and task.
Not preconditioning. Preconditioning changes the represented computational problem or coordinates to improve behavior.
Not uncertainty propagation for a full probability distribution. Standard conditioning is deterministic and worst-case unless a probabilistic model is added.

Scope of Application¶

Condition numbers are literal wherever a mathematical answer depends on uncertain or finite-precision input and the local amplification can be specified independently of an algorithm.

Linear systems. Bounding sensitivity of a solution to changes in a matrix or right-hand side.
Least squares. Distinguishing residual, solution, normwise, and componentwise sensitivity.
Eigenvalue problems. Measuring how eigenvalues or invariant subspaces respond to perturbations.
Polynomial roots. Explaining why nearby coefficients can produce widely displaced roots.
Function evaluation. Using derivatives and scaling to assess local relative sensitivity.
Optimization. Measuring sensitivity of solutions or values under a declared regularity regime.
Geometric numerical problems. Defining condition relative to distance from ill-posed instances.
Scientific computing. Combining data uncertainty, condition estimation, and backward-error analysis.

Clarity¶

State the solution map, base point, perturbed data, admissible perturbation set, input and output norms, and whether the quantity is absolute or relative. Write the limiting supremum or the theorem that reduces it to a derivative or operator norm. If the output can vanish, explain how relative scaling is handled. For matrices, name the norm and whether A, b, or both vary. Distinguish the exact condition number from an estimate. Do not infer algorithm quality from conditioning alone; publish a separate backward- or forward-stability statement. When using the rule of thumb about lost digits, label it approximate and state the floating-point and error-model assumptions.

Manages Complexity¶

The abstraction compresses infinitely many nearby problem instances into one local amplification scale. That scale lets analysts decide whether observed error is explained by the problem itself, by the algorithm, or by both. It guides precision selection, data-quality requirements, preconditioning, and reformulation. The compression discards direction-specific behavior, finite-perturbation nonlinearities, probability structure, and correlations among data errors. A single scalar may therefore be pessimistic, and structured or componentwise condition numbers can be much more informative than an unstructured normwise value. Condition estimates reduce computational cost but introduce estimator uncertainty. The method manages complexity only when its perturbation model matches how errors actually enter the problem.

Abstract Reasoning¶

Specify the mathematical solution map rather than naming only a matrix or algorithm.
Fix the data point and identify which components are uncertain.
Choose admissible perturbations that preserve any relevant structure.
Choose input and output size rules and declare absolute or relative scaling.
Form the ratio of output change to input change.
Take a local limit and worst case over allowed directions.
Reduce the result to a derivative, operator norm, or distance-to-ill-posedness formula when justified.
Interpret magnitude against data precision and desired output accuracy.
Analyze algorithmic backward stability separately.
Test whether componentwise, structured, or finite-perturbation analysis changes the conclusion.

Knowledge Transfer¶

The strict parent is Perturbation. A condition number begins with a reference instance, applies a small admissible disturbance, observes response, and extracts the leading local amplification. Perturbation applies to physical, statistical, and dynamical systems without producing a condition number. Condition Number adds a solution map, paired size measures, a limiting worst-direction ratio, and the numerical-analysis distinction between problem conditioning and algorithm stability. Sensitivity Analysis in Operations Research is a close surface but its accepted identity is post-optimality analysis of solved optimization models, so it is not the literal parent of a general numerical condition number.

Examples¶

Canonical¶

Consider A x = b with nonsingular A, and vary only b. The exact solution is x = A⁻¹ b; a perturbation δb produces δx = A⁻¹ δb. Taking the worst normwise relative amplification over nonzero directions yields a bound involving ||A|| ||A⁻¹||. Under the Euclidean norm, a diagonal matrix with entries 1 and 10⁻⁶ has condition number 10⁶: perturbations aligned with the small-singular-value direction can be amplified about a millionfold relative to the best-scaled direction. The value describes the linear-system problem, not whether Gaussian elimination was implemented stably.^[2]

Mapped back: solution map plus base matrix → admissible right-hand-side perturbation → relative response ratio → worst direction → matrix condition number → separate stability analysis.

Applied / In Practice¶

A scientific model computes f(x) = sqrt(x) at a positive input. Its local relative condition is one half, so small relative input errors are locally halved in the output. By contrast, for g(x) = 1/(1-x) near one, the same relative input uncertainty can create a much larger relative output change. An engineer uses these values to decide data precision, then tests the chosen evaluation algorithms for backward error. The comparison is meaningful only because the functions, points, and relative scaling are all explicit; saying that square root is always well-conditioned or reciprocal evaluation is always ill-conditioned would erase the point dependence established by the general theory.^[1]

Mapped back: function and evaluation point → derivative-scaled local response → required input precision → algorithm check → defensible output accuracy.

Structural Tensions¶

Problem sensitivity vs. algorithm stability. Both contribute to forward error but answer different questions. Diagnostic: Would the same condition number apply if the algorithm changed?
Worst case vs. typical case. A supremum can be dominated by a rare direction. Diagnostic: Are actual data errors capable of aligning with that direction?
Normwise simplicity vs. componentwise realism. One norm is compact but may hide scaling disparities. Diagnostic: Does the perturbation model match component-level uncertainty?
Local linearization vs. finite changes. Derivative-based condition can miss nonlinear behavior away from the base point. Diagnostic: Are perturbations small enough for the first-order bound?
Representation dependence vs. intrinsic difficulty. Rescaling or reformulating data can alter condition. Diagnostic: Is the chosen representation part of the problem users actually solve?
Exact quantity vs. estimator cost. Computing condition exactly may rival solving the problem. Diagnostic: What guarantee accompanies the estimate?
Autonomous abstraction vs. generic sensitivity. Many systems are sensitive. Diagnostic: Is there a declared solution map and a local worst-case amplification ratio?

Structural–Framed Character¶

The perturbation-response ratio and limiting supremum are structural once a problem representation and metrics are fixed. The choice of representation, norm, data components, scaling, and evaluation point is framed and can materially change the number. The well-conditioned/ill-conditioned label is relative to desired accuracy and input precision. Condition Number is domain-specific because numerical analysis supplies the exact solution-map, normed-perturbation, and algorithm-boundary apparatus rather than merely observing that a system is sensitive.

Structural Core vs. Domain Accent¶

The transferable skeleton is reference instance + small disturbance → response amplification. The domain accent is a mathematical solution map, normed input and output spaces, absolute or relative scaling, an asymptotic supremum, derivative or singular-value formulas, and explicit separation of problem conditioning from algorithmic stability. Removing those yields generic sensitivity or perturbation analysis.

Perturbation is the strict parent by composition: a condition number presupposes a specified small departure from a reference input and evaluates the leading response. It does not specialize the entire Perturbation prime because it is a numerical scalar diagnostic rather than every small-disturbance analysis. The proposed edge records that constitutive dependency.

The prospective workspace queue contains one strict upward edge to prime:perturbation. No live DAG mutation is authorized.

Relationships to Other Abstractions¶

Current abstraction Condition Number Domain-specific

Parents (1) — more general patterns this builds on

Condition Number is a kind of Perturbation Prime

Perturbation is the strict parent by composition: a condition number presupposes a specified small departure from a reference input and evaluates the leading response.

Hierarchy paths (2) — routes to 2 parentless roots

Condition Number → Perturbation → Observability

Show alternative path (1)

Neighborhood in Abstraction Space¶

Condition Number sits in a sparse region of the domain-specific corpus (85^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Numerical Stability. Property of an algorithm under computational error.
Backward Error. Smallest data perturbation for which a computed answer is exact.
Sensitivity Analysis in Operations Research. Post-optimality exploration of model parameters and solutions.
Preconditioner. Transformation intended to improve computational conditioning or solver behavior.
Error Bound. Inequality on actual error, often combining condition and backward error.
Singular Value. Matrix quantity used to compute one norm-specific condition number.
Elasticity. A relative derivative that coincides with a scalar relative condition in certain settings.

References¶

[1] John R. Rice, A Theory of Condition, SIAM Journal on Numerical Analysis 3, no. 2 (1966): 287–310, https://doi.org/10.1137/0703023. registry ↩a ↩b

[2] Nicholas J. Higham, Accuracy and Stability of Numerical Algorithms, 2^nd ed. (SIAM, 2002), especially chapter 7, https://doi.org/10.1137/1.9780898718027. registry ↩a ↩b