Let X,Y be a dual pair with bilinear pairing ⟨x,y⟩.
Let C⊂X be a nonempty convex set containing 0.
We define two main objects:
-
The gauge function (Minkowski functional):
γC(x):=inf{t>0:x∈tC}.
-
The indicator function:
δC∘(y)={0,+∞,y∈C∘,y∈/C∘.
We also define three fundamental transformations:
-
The polar transform of a set:
C∘:={y∈Y:⟨x,y⟩≤1;∀x∈C}.
-
The polar transform of a gauge:
k∘(y):=inf{μ≥0:⟨x,y⟩≤μk(x);∀x∈X}.
-
The Fenchel transform (Fenchel conjugate) of a convex function:
f∗(y):=sup{⟨x,y⟩−f(x):x∈X}.
In this note, we prove the following results:
-
Polar transform of a gauge: (γC)∘=γC∘.
-
Fenchel transform of a gauge: (γC)∗=δC∘.
NOTE. There is a more general definition of the polar transform on the class of nonnegative convex functions vanishing at 0:
k∘(y):=inf{μ≥0:⟨x,y⟩≤1+μk(x);∀x∈X}.
On this class of functions, the following diagram commutes (see Convex Analysis - Rockafellar - Corollary 15.5.1 - P.138):
γC⇢γC∘↓↓δC∘⇢δC
where ⇢ denotes the Polar transform and ↓ denotes the Fenchel transform.
1. Polar of a Gauge
Fix μ>0. The condition
⟨x,y⟩≤μ,γC(x)∀x
is equivalent to checking it only on C.
If x∈C, then γC(x)≤1, hence
⟨x,y⟩≤μ∀x∈C.
This is equivalent to
⟨x,μ−1y⟩≤1∀x∈C,
i.e. μ−1y∈C∘, or equivalently y∈μC∘.
Therefore
(γC)∘(y)=inf{μ≥0:y∈μC∘}=γC∘(y).
2. Fenchel Conjugate of a Gauge
Normalize any x as
x=tc,t:=γC(x),c∈C.
Then
⟨x,y⟩−γC(x)t(⟨c,y⟩−1).
Hence
(γC)∗(y)=t≥0,;c∈Csupt(⟨c,y⟩−1).
Case analysis
Case 1: y∈/C∘
Then there exists c∈C such that ⟨c,y⟩>1.
Letting t→+∞ gives +∞.
Case 2: y∈C∘
Then
⟨c,y⟩≤1∀c∈C.
Hence t(⟨c,y⟩−1)≤0 for all t≥0.
Taking t=0 yields value 0, so the supremum equals 0.
Conclusion
We obtain the compact formula
(γC)∗(y)={0,+∞,y∈C∘,y∈/C∘,
or equivalently
(γC)∗=δC∘.