Legendre–Fenchel transform

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

A Legendre–Fenchel transform of an extended-real-valued f:Rn(,+]f:\mathbb{R}^n\to(-\infty,+\infty] is the function f:Rn(,+]f^*:\mathbb{R}^n\to(-\infty,+\infty] defined by

f(y)  =  supxRn(y,xf(x)). f^*(y) \;=\; \sup_{x\in\mathbb{R}^n}\big(\langle y,x\rangle - f(x)\big).

This is exactly the ; the “Legendre” terminology is common even when ff is not differentiable, and it underlies and . When ff is smooth and strictly convex, the transform agrees with the classical on the range of f\nabla f.

Examples:

  • If f(x)=xf(x)=|x| on R\mathbb{R}, then f(y)=0f^*(y)=0 for y1|y|\le 1 and f(y)=+f^*(y)=+\infty for y>1|y|>1.
  • If f=δCf=\delta_C is the indicator of a set CC, then f(y)=supxCy,xf^*(y)=\sup_{x\in C}\langle y,x\rangle (the support function of CC).