Matrix Norms and Condition Numbers

College Depth 61 in the knowledge graph I know this Set as goal
Unlocks 1 downstream topic
matrix-norms conditioning numerical-stability

Core Idea

Matrix norms measure matrix size; common ones are the operator norm ||A||₂ = σ_max(A), Frobenius norm ||A||_F = √(Σᵢⱼ Aᵢⱼ²), and max norm ||A||_∞. The condition number κ(A) = ||A|| ||A⁻¹|| quantifies sensitivity of Ax = b to perturbations: small relative errors in b lead to large relative errors in x when κ is large. κ = σ_max/σ_min for SVD.

Explainer

You already know that a vector norm measures the size of a vector — it gives you a single number capturing how "big" a vector is. A matrix norm extends this idea to linear transformations. The most geometrically meaningful one is the operator norm (or spectral norm), ||A||₂, which asks: over all unit vectors u, what is the largest ||Au|| can be? In other words, what is the maximum factor by which the matrix stretches any input? From your study of the SVD, you know this answer immediately — it is σ_max, the largest singular value. The matrix is at most that many times bigger than any input it acts on.

The Frobenius norm takes a different approach: it treats the matrix as a long vector of all its entries and computes the ordinary Euclidean length. It is computationally simpler and appears often in optimization and statistics, but it does not have a clean geometric interpretation as "maximum stretch." The connection to the SVD is still elegant: ||A||_F = √(σ₁² + σ₂² + ··· + σₙ²), the square root of the sum of squared singular values.

Now for the central concept: the condition number κ(A) = ||A|| · ||A⁻¹||. To understand what it measures, consider solving Ax = b. Suppose b is perturbed slightly — say by measurement noise — giving you b̃ = b + δb. The solution shifts to x̃ = A⁻¹b̃. How large can the relative error ||δx||/||x|| be relative to the relative perturbation ||δb||/||b||? The answer is bounded by κ(A). A condition number of 10 means errors in b can be amplified by at most a factor of 10. A condition number of 10⁸ means tiny relative errors in b can become enormous relative errors in x — the system is numerically ill-conditioned.

Using the SVD, the condition number has a beautiful form: κ₂(A) = σ_max/σ_min. Think about what this means geometrically. The SVD shows that A stretches space by σ_max in one direction and σ_min in another. A⁻¹ must "undo" those stretches, so it compresses by σ_max and stretches by 1/σ_min. A matrix with very unequal singular values — one enormous direction and one nearly-zero direction — has a huge condition number. Geometrically, this means the matrix nearly collapses space in some direction; recovering the original vector from the output requires extreme amplification, making the problem numerically fragile. When σ_min is nearly zero, the matrix is nearly singular and κ → ∞.

Practice Questions 5 questions

Prerequisite Chain

Counting to 10Counting to 20Understanding ZeroThe Number ZeroCounting to FiveOne-to-One CorrespondenceCombining Small Groups Within 5Addition Within 10Addition Within 20Two-Digit Addition Without RegroupingTwo-Digit Addition with RegroupingAddition Within 100Repeated Addition as MultiplicationMultiplication Facts Within 100Division as Equal SharingDivision as Grouping (Measurement Division)Division: Grouping (Repeated Subtraction) ModelDivision: Fair Sharing ModelDivision as Equal SharingDivision as GroupingBasic Division FactsDivision Facts Within 100Two-Digit by One-Digit DivisionDivision with RemaindersRemainders and Quotients in DivisionDivision Word ProblemsIntroduction to Long DivisionFactors and MultiplesPrime and Composite NumbersEquivalent FractionsRelating Fractions and DecimalsDecimal Place ValueIntegers and the Number LineOpposites and Additive InversesAbsolute ValueAdding IntegersSubtracting IntegersMultiplying IntegersDividing IntegersUnit RatesProportionsPercent ConceptConverting Between Fractions, Decimals, and PercentsOperations with Rational NumbersTwo-Step EquationsSolving Multi-Step EquationsEquations with Variables on Both SidesLiteral EquationsSlope-Intercept FormPoint-Slope FormWriting Linear EquationsParallel and Perpendicular Line SlopesGraphing Linear EquationsSystems of Equations — Graphing MethodSystems of Equations — Elimination MethodSystems of Three VariablesMatrices IntroductionLinear TransformationsEigenvalues and EigenvectorsSymmetric Matrices and Their PropertiesSingular Value Decomposition (SVD)Matrix Norms and Condition Numbers

Longest path: 62 steps · 247 total prerequisite topics

Prerequisites (2)

Leads To (1)