Why is it called the infinity norm?

For finite-dimensional vectors, it is the limiting value of the Lp norm as p grows without bound.

No. It is the maximum of absolute coordinate values.

Is a matrix infinity norm the largest entry?

The induced infinity norm is the largest absolute row sum. An entrywise maximum is a different convention.

AI Glossary / Mathematics and optimization / Intermediate

Max norm

Definition

The max norm of a finite-dimensional vector is the largest absolute value of any of its coordinates. It is also called the infinity norm or L-infinity norm and is written ‖x‖∞=max_i |x_i|. It measures the size of the vector by its largest individual coordinate magnitude.

What it means

Unlike a sum or an average, the max norm focuses on the most extreme coordinate. A vector with many moderate entries can have the same max norm as one with just one entry of that size.

For the difference between two vectors, the max norm measures their largest coordinate-wise discrepancy. Requiring ‖x−y‖∞≤epsilon means every corresponding coordinate differs by at most epsilon. This is useful when an error tolerance applies to all components individually. Coordinates still need compatible scales or an explicit normalization; otherwise the largest number may mainly reflect the units chosen.

How it works

Take the absolute value of every coordinate and select the largest. The result is nonnegative, equals zero only for the zero vector, scales by the absolute value of a scalar multiplier, and satisfies the triangle inequality.

The last property follows coordinate by coordinate: |x_i+y_i|≤|x_i|+|y_i|≤‖x‖∞+‖y‖∞. Taking the largest left-hand side gives the required bound.

The associated distance between vectors is ‖x−y‖∞, sometimes called Chebyshev distance. In two dimensions, points at distance at most one from the origin form a square aligned with the coordinate axes. For finite vectors, the max norm is also the limit of the usual Lp norm as p increases without bound. The infinity symbol does not mean that a finite vector's norm is infinite.

Three practical examples

1. Numerical validation: require every component of a computed residual vector to remain below a specified absolute tolerance.

2. Model comparison: measure the largest prediction difference across a fixed collection of cases, alongside average-difference measures.

3. Robustness analysis: describe perturbations that limit the change to each normalized input coordinate, with the coordinate bounds clearly specified.

Think of it this way

Think of inspecting a row of measuring instruments and recording the largest absolute deviation from their targets. One extreme reading determines the summary, even if the others are small.

The analogy assumes the readings are comparable. Combining temperatures, currency amounts, and durations without scaling would make the largest numerical reading a poor overall summary. The max norm also omits how many instruments are close to the limit.

Worked example

Let x=(−3,4,1). Its max norm is max(3,4,1)=4. Its L1 norm is 3+4+1=8, and its Euclidean norm is √(9+16+1)=√26, approximately 5.099. These measures answer different questions about the same vector.

Suppose prediction vectors are a=(10,20,30) and b=(10.2,19.7,30.1). Their difference is (−0.2,0.3,−0.1), giving max-norm distance 0.3. A requirement that every absolute difference be at most 0.25 fails because the second coordinate differs by 0.3. The average absolute difference is only 0.2, so an average-based check would answer a different question.

For a matrix distinction, consider A with rows (2,−3) and (1,4). Its largest absolute entry is 4, but its induced infinity norm is the largest absolute row sum: max(2+3,1+4)=5. Applying A to z=(1,−1) gives (5,−3), whose max norm is 5 while ‖z‖∞=1. This confirms why the induced matrix norm is not simply the largest entry.

Illustrative case study

Fictional case: a team replaces a numerical component in a scoring pipeline. The average change in scores is small, so the first comparison appears satisfactory.

A reviewer also examines the largest absolute score change and finds one case beyond the agreed tolerance. The team investigates that case, compares relevant subgroups, and confirms the measurement units. It uses the max norm as a precise all-coordinate check on the tested cases. It does not claim that passing a finite test establishes the same bound for every future input.

Common misconceptions and mistakes

Do not take the largest signed entry and ignore negative magnitude. For (−9,2), the max norm is 9, not 2. Do not confuse it with an average error or assume it reveals the number of large errors.

Matrix terminology needs particular care. The induced matrix infinity norm uses absolute row sums, while an entrywise maximum is another quantity. In other machine-learning contexts, max-norm regularization can refer to constraints on groups of weights under a different specified norm. This entry addresses the vector max norm, so such uses need their own definitions.

Business use

The max norm is useful when a service requirement applies to every component, such as bounded numerical deviations across a defined validation set. It makes the worst observed coordinate discrepancy visible rather than hiding it inside an average.

Use it together with the coordinate location, distribution of errors, and sample coverage. A single outlier may reveal an important problem or a measurement issue; the norm alone cannot distinguish them. When features have different units, choose defensible scaling before treating their maximum as a shared tolerance measure.

Check your understanding

A residual vector is (−0.04,0.02,0.03). Does it satisfy a max-norm tolerance of 0.035?

Answer: No. Its max norm is 0.04, exceeding the tolerance even though the other components are smaller.

Related terms

Norm
Euclidean norm
Residual
infinity norm
Chebyshev distance
L1 norm
induced matrix norm
error tolerance

Learn next

compare L1, L2, and max norms on the same residual vectors. Then examine induced matrix norms and why a coordinate-wise bound differs from a bound on total or average error.

Residual

Sources

NumPy, vector and matrix norm definitions: https://numpy.org/doc/stable/reference/generated/numpy.linalg.norm.html . The infinity-norm table was read. Editorial scope is the vector maximum absolute coordinate, with the matrix row-sum distinction explicit. Calculations and the fictional validation scenario are original.

Previous: Max error Next: Max pooling