Clarification about the definition of subgradient

64 Views Asked by At

I have a basic clarificatory question about the definition of a subgradient from the book Infinite Dimensional Analysis by Aliprantis and Border. Here is the definition.

Given a dual pair $\langle X,X' \rangle$ and a real-valued convex function $f$ on $X$, we say that $x' \in X'$ is a subgradient of $f$ at $x$ if $f(y) \geq f(x) + x'(y-x)$ for all $y \in X$.

And here is the definition of dual pair.

A dual pair is a pair $\langle X,X' \rangle$ of vector spaces with a bilinear functional $(x,x') \mapsto \langle x,x'\rangle$ from $X \times X'$ to $\mathbb R$ that separates the points of $X$ and $X'$. That is, $(x,x') \mapsto \langle x,x'\rangle$ is linear in both of its arguments and satisfies (1) if $\langle x,x' \rangle = 0$ for each $x' \in X'$, then $x = 0$, and (2) if $\langle x,x' \rangle = 0$ for each $x \in X$, then $x'=0$.

My question (again, very basic) is how to understand the expression $x'(y-x)$ in the definition of subgradient. Is this just an alternative notation for $\langle y-x, x' \rangle$? If not, then I am unsure how to connect the definition of the subgradient to the definition of dual pairs.

1

There are 1 best solutions below

1
On BEST ANSWER

Yes, this is just an alternative notation. Since the dual space is frequently viewed as all bounded linear functionals from $X$ to $\mathbb{R}$, sometimes the pairing $\langle x, x'\rangle$ is defined by the "evaluation" of $x'$ at $x$, written $\langle x, x'\rangle = x'(x)$.

The definition from the book uses an alternative (albeit equivalent) definition, where $x'\in X'$ is a vector. In this case, the bounded linear functional that $x'$ "represents" is $x\mapsto\langle x, x'\rangle$. However, folks frequently abuse notation and do not distinguish between the vector $x'\in X'$ and its functional form, which can lead to the (somewhat confusing) convention that $x'(x)=\langle x,x'\rangle$.