How this instrument works
A condition number measures how much a matrix stretches errors rather than how large the matrix itself is. This sheet uses the infinity-norm version, built from the largest absolute row sum: ‖A‖∞ takes the two row sums |a|+|b| and |c|+|d| and keeps the bigger one, while ‖A⁻¹‖∞ applies the same largest-row-sum idea to the inverse matrix, worked out here as max(|d|+|b|, |c|+|a|) divided by the determinant's absolute value rather than by inverting the matrix entry by entry first. Multiplying those two figures together gives κ, a single score summarizing how the matrix behaves under both itself and its reverse.
A score close to 1 describes a well-conditioned matrix — one where a small change in an input barely moves the output of whatever calculation uses it, such as solving a linear system Ax = b. A score far above 1 flags an ill-conditioned matrix instead, where a tiny rounding error or measurement slip in b can swing the solved x wildly, even though the arithmetic itself was carried out correctly at every step. The identity matrix sits at the best possible floor, κ = 1, since it distorts nothing at all; a nearly singular matrix, by contrast, can push κ toward enormous values as its determinant shrinks toward zero.
The practical stakes show up the moment this matrix represents real measurements rather than clean textbook entries. Feed a poorly conditioned matrix into an engineering or financial model built on solving Ax = b, and ordinary sensor noise or rounding in the sixth decimal place can produce wildly different answers depending on which numerically equivalent path the calculation happens to take. A well-conditioned matrix offers no such trapdoor — small input wobble stays small in the output, which is exactly why this score is checked before trusting a solved system, not after.
- Enter the top row of the matrix into a (row 1, col 1) and b (row 1, col 2).
- Enter the bottom row into c (row 2, col 1) and d (row 2, col 2).
- Read Condition number (∞-norm) for κ, the product of the matrix's infinity-norm and its inverse's infinity-norm.
- Treat a result near 1 as well-conditioned and a result climbing into the hundreds or higher as a warning sign for numerical instability.
- Avoid a determinant of exactly zero — a singular matrix has no inverse and therefore no finite condition number at all.
Worked example — the matrix [[2, 1], [3, 4]]
Set a to 2, b to 1, c to 3, and d to 4 — the matrix [[2, 1], [3, 4]]. First the determinant: ad − bc = 2×4 − 1×3 = 8 − 3 = 5. The infinity-norm of A itself compares the two row sums, |2|+|1| = 3 and |3|+|4| = 7, and keeps the larger: ‖A‖∞ = 7.
The inverse's infinity-norm swaps in |d|+|b| and |c|+|a|, giving |4|+|1| = 5 and |3|+|2| = 5 — a tie at 5 — divided by the determinant's absolute value, 5, so ‖A⁻¹‖∞ = 5 ⁄ 5 = 1. Multiplying the two pieces together, κ = 7 × 1 = 7, exactly the figure Condition number (∞-norm) reports for these four entries. A κ of 7 sits well toward the well-conditioned end: solving Ax = b with this matrix will not turn a small input wobble into a large output swing.
Questions
What does a matrix condition number actually measure?
It measures how much a matrix can amplify small errors when it is used to solve a linear system, expressed as one number: κ = ‖A‖∞ × ‖A⁻¹‖∞. A low κ, close to 1, means tiny changes in input stay tiny in output; a high κ means the opposite — small measurement noise or rounding can blow up into large, unreliable swings in the answer.
What counts as a well-conditioned matrix versus an ill-conditioned one?
A condition number near 1 is well-conditioned: the identity matrix scores exactly 1, the best possible floor, since it distorts nothing at all. A condition number in the hundreds, thousands, or beyond is ill-conditioned, meaning the matrix sits close to singular and even minor input errors can produce wildly different solved results, despite every calculation step being carried out correctly.
Why does the formula use the infinity norm instead of another norm?
The infinity norm is simply the largest absolute row sum, which is quick to compute directly from a matrix's entries without eigenvalues, singular values, or any iterative step. Other norms, such as the Frobenius or spectral norm, produce a related but not identical condition-number figure, so a κ value is always understood alongside the specific norm used to build it, infinity-norm here.
Why can't a matrix with a zero determinant have a condition number?
Because ‖A⁻¹‖∞ divides by the determinant's absolute value, and a determinant of exactly zero makes that division undefined, the same way dividing an ordinary number by zero is undefined. A zero determinant means the matrix is singular and has no inverse at all, so there is no ‖A⁻¹‖∞ to multiply into the formula in the first place.
What should I actually do if my matrix has a large condition number?
Treat any solved system built on it with real caution rather than trusting the numbers at face value: a large κ means the same matrix, fed slightly different input due to measurement noise or rounding, can hand back a substantially different answer. Rescaling the problem, gathering more precise measurements, or reformulating the system to avoid the near-singular matrix are the usual practical fixes.