Canonical: https://knowlpedia.com/convex-analysis/legendre-fenchel-transform/

# Legendre–Fenchel transform

The general convex-conjugation transform defined by a supremum pairing, without smoothness assumptions.

A **Legendre–Fenchel transform** of an extended-real-valued [function](https://knowlpedia.com/shared-foundations/function/) \(f:\mathbb{R}^n\to(-\infty,+\infty]\) is the function \(f^*:\mathbb{R}^n\to[-\infty,+\infty]\) defined by
\[
f^*(y) \;=\; \sup_{x\in\mathbb{R}^n}\big(\langle y,x\rangle - f(x)\big).
\]
where \(\langle y,x\rangle=\sum_{i=1}^n y_i x_i\) is the Euclidean [inner product](https://knowlpedia.com/linear-algebra/inner-product/).

## Properties and limiting cases

If \(f\equiv+\infty\), then \(f^*\equiv-\infty\); otherwise \(f\) is finite somewhere and \(f^*\) never takes the value \(-\infty\).

This is exactly the [Fenchel conjugate](https://knowlpedia.com/convex-analysis/convex-conjugate-fenchel/); the “Legendre” terminology is common even when \(f\) is not differentiable, and it underlies [Fenchel–Young inequality](https://knowlpedia.com/convex-analysis/fenchel-young-inequality/) and [biconjugation](https://knowlpedia.com/convex-analysis/biconjugate/). When \(f\) is smooth and strictly convex, the transform agrees with the classical [Legendre transform](https://knowlpedia.com/convex-analysis/legendre-transform/) on the range of \(\nabla f\).

## Examples

- If \(f(x)=|x|\) on \(\mathbb{R}\), then \(f^*(y)=0\) for \(|y|\le 1\) and \(f^*(y)=+\infty\) for \(|y|>1\).
- If \(f=\delta_C\) is the indicator of a set \(C\), then \(f^*(y)=\sup_{x\in C}\langle y,x\rangle\) (the support function of \(C\)).
