I encountered the following unsubstantiated fact in a proof:

$f$ is differentiable, strictly convex, and coercive. Thus we may conclude $$(\nabla f)^{-1} = \nabla f^*$$ where $f^*$ is the Legendre transformation of $f$

Loosely speaking, I see how this might work: if a function is differentiable then Legendre trasformation will return the slope at some point $x \in \text{dom} f$. If $f$ is strictly coercive, then $\nabla f$ is monotonic, in the appropriate sense. So an inverse is defined. Loosely, I think the role of coercivity is to make the map between $x \in \text{dom} f$ and its slope at $x$ bijective. I can't understand how coercivity is causing this though.

My questions:

  1. Can someone provide a heuristic description what coercivity is giving us?
  2. Provide a proof of the fact above (or reference)? Assume $f: \mathbb{R}^d \rightarrow \mathbb{R}$.

Recall that $f^*(y) = \sup \langle x,y\rangle -f(x)$. Assuming there is a minimizer $x^*$ we have, $f^*(y) = \langle x^*,y\rangle -f(x^*)$. At the same time, $x^*$ satisfies the optimality conditions $0=\nabla_x(\langle x,y\rangle -f(x))|_{x=x^*} = y-\nabla_x f(x^*)\implies y=\nabla_xf(x^*)$. If we take the gradient of $f^*(y)$ w.r.t. $y$ to get $\nabla_y f^*(y) = x^*(y)$. So we know that $\nabla f^*(y) = (\nabla f)^{-1}(x^*(y))$. But $x^*$ might not be unique.

So you want,

$$\arg\sup_x \langle x,y\rangle - f(x)=\{x^*\}\neq\emptyset$$

a singleton for all $y$. Coercivity in $x$ will ensure this since $f$ is also strictly convex (and thus $-f$ is strictly concave).

The injectivity of $\nabla f$ needed to define an inverse is a consequence of strict convexity and coercivity. Since $f$ is strictly convex, $\nabla f$ is strictly monotone (and thus injective). Notice that

$$\arg\sup_x \langle x,y\rangle - f(x) = \arg\min_x \langle x,-y\rangle + f(x)$$

Fix $y$ and let $\phi(x) = \langle x,-y \rangle + f(x)$. The sublevel sets of $\phi$ are bounded by coercivity and closed since $\phi$ is continuous. Taking any sublevel set (e.g. the sublevel set for $\phi(0)$), we have that $\phi$ is strictly convex and differentiable and thus admits a unique minimum on a compact set. Then, for any $y$, $x^*$ exists and is unique so that $\nabla f$ is surjective.

  1. You don't need coercivity. The result is true, e.g., for the exponential function which is not coercive.

  2. Check Chapter 26 in Rockafellar's Convex Analysis about the Legendre transformation.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.