Matrix Norm Calculator — 1, 2, infinity, Frobenius and condition number

The matrix norm calculator above computes the four norms in everyday use — the 1-norm, the 2-norm or spectral norm, the infinity-norm and the Frobenius norm — together with the full set of singular values and the condition number. Presenting them side by side is deliberate, because the useful information is usually in how far apart they are rather than in any one of them.

Arb Digital publishes it because a norm is how a matrix gets a single number describing its size, and different norms answer different questions about it. Our matrix calculator performs the arithmetic operations; this page measures the result. Our matrix rank calculator answers the related question of how many independent directions the matrix actually uses.

What This Matrix Norm Calculator Does

Three of the four norms are direct sums over the entries and are computed as such. The 1-norm is the largest absolute column sum, the infinity-norm the largest absolute row sum, and the Frobenius norm the square root of the sum of every squared entry. None of them requires anything beyond addition and one square root.

The spectral norm is different. It is the largest singular value, which means it needs the eigenvalues of A transpose times A. The tool builds that symmetric matrix and runs a Jacobi rotation sweep on it, which converges reliably for symmetric input and gives every eigenvalue rather than just the largest. Taking square roots produces the complete list of singular values, so the smallest comes free and the condition number follows.

Rectangular matrices are handled throughout. All four norms are defined for any shape, and so is the condition number, although for a non-square matrix it describes the least-squares problem rather than a solve. A rank-deficient matrix has a zero smallest singular value and therefore an infinite condition number, and the tool reports that in words rather than dividing by zero.

How to Use It

1. Paste the matrix one row per line. Every row must have the same number of entries, but the number of rows and columns need not match.
2. Read the four norms together. If they are close, the matrix is well balanced; if the spectral norm is far below the Frobenius, the matrix has several comparable directions rather than one dominant one.
3. Check the condition number before trusting any solve involving this matrix. It is the multiplier on your input error.
4. Look at the singular value list. A sharp drop between consecutive values is where the effective rank sits.
5. Try the presets to see a well-conditioned matrix, a singular one, and the Hilbert matrix, which is the standard example of severe ill-conditioning.

The Formulas and How They Are Calculated

Wolfram MathWorld's page on matrix norms defines the three induced norms directly: the 1-norm as the maximum over columns of the sum of absolute entries in that column, the infinity-norm as the same maximum taken over rows, and the 2-norm as the square root of the largest eigenvalue of A conjugate-transpose times A. The Frobenius norm, √(∑ a_ij²), is not induced by a vector norm but is used constantly because it is cheap and differentiable.

The first three are called induced or operator norms because each one answers the same question with a different vector norm: what is the largest factor by which A can stretch a vector? Measure the vector with the taxicab norm and you get the 1-norm; with the Euclidean norm you get the spectral norm; with the maximum-component norm you get the infinity-norm. That is why they differ — they are measuring stretch with different rulers.

Work the default 3×3 by hand. A has rows (1, 2, 3), (4, 5, 6) and (7, 8, 10). The column sums are 12, 15 and 19, so the 1-norm is 19. The row sums are 6, 15 and 25, so the infinity-norm is 25. Every squared entry adds to 304, so the Frobenius norm is √304 = 17.4355958. The spectral norm comes out at 17.4125052, just below the Frobenius as it always must be, and the smallest singular value is 0.1968665, giving a condition number of about 88.4. As a check, the product of the three singular values is 3, which is the absolute value of the determinant.

The Inequalities Between Them

Several relationships always hold and they are worth using as sanity checks. The spectral norm never exceeds the Frobenius norm, and equals it only when the matrix has rank one. The Frobenius norm never exceeds √r times the spectral norm, where r is the rank. And the spectral norm is bounded by the geometric mean of the 1-norm and the infinity-norm, which gives a quick estimate without computing any eigenvalues at all.

The gap between the spectral and Frobenius norms is informative in itself. Since the Frobenius norm is the square root of the sum of all squared singular values and the spectral norm is the largest one alone, the two coincide when a single direction carries everything. A wide gap means the matrix spreads its action across several comparable directions, which is precisely what low-rank approximation methods exploit.

The maximum absolute entry is sometimes called the max norm. It is a norm on the entries but it is not submultiplicative — the max norm of a product can exceed the product of the max norms — which makes it unsuitable for the error analysis that norms are usually wanted for. That is why it does not appear among the four headline figures here.

The Condition Number and What It Costs You

MathWorld's page on the condition number defines it as the ratio of the largest to the smallest singular value in the singular value decomposition of a matrix. That is exactly what this page reports as κ₂. The equivalent norm-based definition, ‖A‖ times ‖A⁻¹‖, coincides with it for the 2-norm and gives slightly different values for the 1- and infinity-norms.

Its meaning is a bound on error amplification. If you solve A x = b and the right-hand side carries a relative error, the relative error in the solution can be up to κ times as large. A condition number of one is perfect, achieved by orthogonal matrices, which is why numerical algorithms are built out of them wherever possible. A condition number of 10⁸ in double precision leaves roughly eight of the sixteen available digits, which is usually still workable; 10¹⁶ leaves none.

Press the Hilbert preset to see this at its most extreme. The 4×4 Hilbert matrix, whose entry in row i column j is 1 divided by i + j − 1, has a condition number above fifteen thousand, and the 10×10 version exceeds 10¹³. Its entries are innocuous fractions, which is exactly the point: ill-conditioning is not visible by inspection and has to be measured. Our matrix inverse calculator and determinant calculator both become unreliable long before they fail outright on such a matrix.

Why a Small Determinant Is Not the Same as Ill-Conditioning

This is the misconception worth spending a paragraph on. A near-zero determinant is often taken as the sign of a badly behaved matrix, and it is not reliable. Take the identity matrix scaled by 0.001 in ten dimensions: its determinant is 10⁻³⁰, vanishingly small, yet its condition number is exactly 1 and solving with it is perfectly stable. Every direction is scaled identically, so nothing is lost.

Conversely a matrix can have a determinant of exactly 1 and a condition number in the millions, if it stretches enormously in one direction and shrinks correspondingly in another. The determinant is the product of the singular values and tells you about volume; the condition number is their ratio and tells you about distortion. Those are different questions, and only the second one predicts numerical trouble.

The right diagnostic is always the singular value spectrum, which is why this page prints all of them rather than just the extremes. A clean gap in that list identifies the effective rank; a gradual decay says the matrix has no clean rank at all. Our eigenvalue calculator covers the related but distinct spectrum of a square matrix, and our matrix diagonalization calculator builds the eigenvector factorisation where one exists.

Common Mistakes to Avoid

• Swapping the 1-norm and the infinity-norm — the 1-norm sums down columns, the infinity-norm across rows. They coincide only for symmetric matrices.
• Calling the Frobenius norm the 2-norm — the 2-norm is the largest singular value. The Frobenius norm is the root of their sum of squares, and it is always the larger of the two.
• Reading a small determinant as ill-conditioning — a uniformly scaled matrix has a tiny determinant and a condition number of exactly one.
• Using the max-entry norm for error analysis — it is not submultiplicative, so the bounds that make norms useful do not hold for it.
• Ignoring the condition number before a solve — it is the multiplier on your input error, and it is far more informative than any single norm on its own.

Related Free Tools From Arb Digital

Multiply and transpose with the matrix calculator, invert with the matrix inverse calculator, check independence with the matrix rank calculator, find the spectrum with the eigenvalue calculator, or factor a symmetric positive definite matrix with the Cholesky decomposition calculator. The full free online tools hub lists every linear algebra tool we publish.

Frequently Asked Questions

This page explains a topic in numerical linear algebra for educational purposes. Singular values are computed in double-precision floating point, so a value reported as very small may be a genuine zero obscured by rounding.