Week 9: Convex conjugate, Fenchel Conjugate
Every function has something known as the convex conjugate, or differential conjugate. And, this is a very important concept in convexity and in duality.
1 Definition
Given function
f
:
R
n
→
R
f: \mathbb{R}^n\rightarrow \mathbb{R}
f:Rn→R, its conjugate
f
∗
f^*
f∗ is:
f
∗
(
y
)
=
max
x
y
T
x
−
f
(
x
)
f^*(y)=\max_x y^Tx-f(x)
f∗(y)=xmaxyTx−f(x)
f
∗
f^*
f∗ is always convex because it is max of affine functions.
2 Properties
Sum of conjugate
f ( x ) + f ∗ ( y ) ≥ y T x f(x)+f^*(y)\geq y^Tx f(x)+f∗(y)≥yTx
Decompose conjugate
If
f
(
u
,
v
)
=
f
1
(
u
)
+
f
2
(
v
)
f(u,v)=f_1(u)+f_2(v)
f(u,v)=f1(u)+f2(v),
then
f
∗
(
w
,
z
)
=
f
1
∗
(
w
)
+
f
2
∗
(
z
)
f^*(w,z)=f_1^*(w)+f_2^*(z)
f∗(w,z)=f1∗(w)+f2∗(z)
Double conjugate
f
∗
∗
(
z
)
≤
f
(
z
)
f^{**}(z)\leq f(z)
f∗∗(z)≤f(z)
Always less than the original function, and is convex. Convex envelope.
If
f
f
f is convex, then
f
∗
∗
(
z
)
=
f
(
z
)
f^{**}(z) = f(z)
f∗∗(z)=f(z).
Convex original
- f ∗ ∗ ( z ) = f ( z ) f^{**}(z) = f(z) f∗∗(z)=f(z).
- f ∗ ( y ) = max x y T x − f ( x ) = y T x ˉ − f ( x ˉ ) f^*(y)=\max_x y^Tx-f(x)=y^T\bar x-f(\bar x) f∗(y)=maxxyTx−f(x)=yTxˉ−f(xˉ), where y ∈ ∂ f ( x ˉ ) y\in \partial f(\bar x) y∈∂f(xˉ)
- The derivatives of f f f and f ∗ f^* f∗ are inverse of each other: y ∈ ∂ f ( x ) ⇔ x ∈ ∂ f ∗ ( y ) y\in \partial f(x)\Leftrightarrow x\in \partial f^*(y) y∈∂f(x)⇔x∈∂f∗(y)
3 Examples-todo
Uses: duality 2