Is the Euclidean norm self-dual?

Yes, under the usual inner product on real coordinate space.

Why does a coordinatewise budget give a larger value in the example?

Both coordinates can use their full allowance simultaneously.

AI Glossary / Mathematics and optimization / Advanced

Dual norm

Definition

The dual norm associated with a norm measures the largest inner product a vector can achieve against any vector in that norm's unit ball. For real vectors, it is defined by ‖y‖* = sup{yᵀx: ‖x‖ ≤ 1}. The original norm specifies the allowed size of x; the dual norm measures the strongest possible linear response in direction y under that size constraint.

What it means

The phrase dual norm refers to a relationship between two ways of measuring vectors. It is not simply another name for the original norm or its reciprocal. Changing the original norm changes the allowed unit ball and can therefore change the maximum inner product.
This makes dual norms useful for sensitivity analysis. If a linear score has coefficient vector y and an input perturbation δ satisfies ‖δ‖ ≤ ε, then |yᵀδ| ≤ ε‖y‖*. The bound depends on both the coefficients and the geometry used to define the perturbation budget.

How it works

The unit ball is symmetric: if x is allowed, so is −x. Therefore taking the largest signed inner product gives the same value as taking the largest absolute inner product. In finite-dimensional real space, the maximum is attained because the unit ball is compact and the inner product is continuous.
For familiar vector norms, the dual of the ℓ₁ norm is ℓ∞, the dual of ℓ∞ is ℓ₁, and the Euclidean ℓ₂ norm is its own dual. The corresponding inequality |yᵀx| ≤ ‖y‖*‖x‖ generalizes the familiar Euclidean Cauchy - Schwarz bound. The norm on each side must be paired correctly.

Three practical examples

A linear prediction model is evaluated under coordinatewise input tolerances. If every coordinate may change by at most ε, the perturbation uses an ℓ∞ budget and the worst possible score change is ε times the coefficient vector's ℓ₁ norm.
An analyst limits the total absolute input change across coordinates. This is an ℓ₁ budget, whose dual ℓ∞ norm identifies the coefficient magnitude that can produce the largest response per unit of that budget.
An optimization method chooses a direction using a bounded step in a specified norm. The dual norm describes the largest first-order change available from that step budget. For a nonlinear objective, additional reasoning is needed to control the difference between the linear approximation and the actual change.

Think of it this way

Think of having a fixed allowance for changing several inputs. One rule limits each input separately, another limits the total change, and another limits a Euclidean length. The dual norm tells you how strongly a given linear score can react under the selected rule. The same numerical allowance does not describe the same set of possible changes under all three rules.

Worked example

Take the coefficient vector y = (3, −4). First constrain x by ‖x‖₁ = |x₁| + |x₂| ≤ 1. The inner product satisfies 3x₁ − 4x₂ ≤ 3|x₁| + 4|x₂| ≤ 4. Choosing x = (0, −1) attains four. Thus the dual value is ‖y‖∞ = 4.
Next impose ‖x‖₂ ≤ 1. Cauchy - Schwarz gives yᵀx ≤ √(3² + (−4)²) = 5. The choice x = (3/5, −4/5) has Euclidean norm one and inner product 9/5 + 16/5 = 5. The bound is attained, confirming the Euclidean dual value.
Finally impose ‖x‖∞ = max(|x₁|, |x₂|) ≤ 1. Each coordinate can independently use its full allowance. Choosing x = (1, −1) gives 3 + 4 = 7. No allowed vector can exceed |3| + |−4| = 7, so the dual value is ‖y‖₁ = 7.
For a perturbation radius ε = 0.2, the maximum score changes are therefore 0.8 under an ℓ₁ budget, 1 under an ℓ₂ budget and 1.4 under an ℓ∞ budget. Multiplying the maximizing vectors above by 0.2 attains each value. These are exact worst-case changes for the specified linear score and unrestricted norm-ball perturbations.

Illustrative case study

In a fictional measurement system, two input readings each have an allowed error of at most 0.2 in their documented numerical units. A linear score uses coefficients (3, −4). An analyst uses the coefficient vector's Euclidean norm, five, and reports a maximum score error of one.
That calculation used the wrong perturbation geometry. Independent coordinate limits define an ℓ∞ ball, not an ℓ₂ ball of radius 0.2. The permitted error δ = (0.2, −0.2) changes the score by 3 × 0.2 − 4 × (−0.2) = 1.4.
The team corrects the bound using the dual ℓ₁ norm. It also checks whether additional physical restrictions couple the two errors. If so, the full box may be conservative, and a tighter analysis could use the actual allowed set. The dual-norm calculation is exact for the stated norm ball; it does not establish that every mathematically allowed perturbation can occur operationally.

Common misconceptions and mistakes

The dual norm is not generally equal to the original norm. For the same vector (3, −4), the three relevant values are four, five and seven. Nor is it the reciprocal of the original norm.
Do not confuse a dual norm with a Lagrange dual optimization problem, although dual norms can appear in such problems. Also, a first-order bound for a nonlinear model is not automatically a global robustness guarantee. The linearization error and the region over which assumptions hold matter.

Business use

Dual norms help turn clearly defined input tolerances into interpretable sensitivity bounds for linear scores and optimization steps. Specify the feature units, permitted changes and coefficient vector before choosing a norm pair. Explain whether the bound concerns an exact linear model or a local approximation. An apparently small bound is useful only when the allowed perturbation set matches the actual question.

Check your understanding

Which vector attains the ℓ₁-budget maximum?

Answer: (0, −1).

What is the ℓ∞-budget dual value for (3, −4)?

Answer: Seven.

What was wrong with the fictional system's bound of one?

Answer: It replaced independent coordinate limits with a smaller Euclidean perturbation ball.

Related terms

Norm
Inner product
Sensitivity analysis
Dual problem
unit ball
Hölder inequality
Cauchy–Schwarz inequality
perturbation

Learn next

Derive the ℓ₁ - ℓ∞ pairing directly from the unit-ball definition. Then compare exact linear sensitivity bounds with first-order approximations for nonlinear functions.

Sources

Aalto Dictionary of Machine Learning, dual norm: supremum of the inner product over the original unit ball. https://aaltodictionaryofml.github.io/ADictML_Main.pdf Stephen Boyd and Lieven Vandenberghe, Convex Optimization slides, equality-constrained norm minimization: dual-norm definition and its role in a dual constraint. https://web.stanford.edu/~boyd/cvxbook/bv_cvxslides.pdf The three norm-ball calculations and fictional measurement-system case are original teaching illustrations.

Previous: Dual graph Next: Dual problem