1、Introduction
Moreau -Yosida regularitztion is a preconditioner of a convex function f.
And this preconditioner has second-order properties
共轭: conjugate of a closed convex function
closed function =lower semi-continuous function
具体表示: We denote by F the Moreau-Yosida regularization of a given closed convex function f associated to the metric de ned by M :
Denote
p
(
x
)
p(x)
p(x) is the unique minimizer in
F
(
x
)
F(x)
F(x)
2、Properities of the moreau-yosida regularization
Definition1:
The recession (or asymptotic) function of a closed convex function C is defined by
(a limit which does not depend on
x
∈
d
o
m
ϕ
x\in dom\phi
x∈domϕ). This function is useful because
ϕ
\phi
ϕ has a nonempty bounded set of minima if and only if
ϕ
∞
′
(
d
)
>
0
\phi^{'}_{\infty}(d)>0
ϕ∞′(d)>0 for all
d
≠
0
d\not=0
d=0.
Definition2: Proximal Point: We will extensively use the following system of notation:
p
M
(
x
)
=
a
r
g
m
i
n
y
f
(
y
)
+
1
2
M
T
(
y
−
x
)
T
(
y
−
x
)
p_{M}(x)=argmin_{y} f(y)+\frac{1}{2}M^{T}(y-x)^{T}(y-x)
pM(x)=argminyf(y)+21MT(y−x)T(y−x)
is called the proximal point of x.
Lemma1:
The minimization problem of
F
(
x
)
F(x)
F(x) has a unique solution, characterized as the unique point
y
∈
R
n
y\in R^{n}
y∈Rn satisfying
M
(
x
−
y
)
∈
∂
f
(
y
)
M(x-y)\in\partial f(y)
M(x−y)∈∂f(y)即
M
(
x
−
p
M
(
x
)
)
∈
∂
f
(
p
M
(
x
)
)
M(x-p_{M}(x))\in\partial f(p_{M}(x))
M(x−pM(x))∈∂f(pM(x))
proof is on J.-B. Hiriart-Urruty and C. Lemar echal, Convex Analysis and Minimization Algorithms, Springer-Verlag, Berlin, New York, 1993;XV,Lemma.4.1.1
Theorem1:
The recession functions of
f
f
f and
F
F
F are identical.
proof : Ref:Rockafellar,Corollary 9.2.1
Theorem 2.
For a finite-valued convex function
f
f
f , the following statements are equivalent:
(i)
f
f
f is strongly convex with modulus
1
l
\frac{1}{l}
l1;
(ii)
f
f
f has a Lipschitzian gradient with Lipschitz constant
l
l
l
(iii)
F
F
F has a Lipschitzian gradient with Lipschitz constant
L
L
L;
(iv)
F
F
F is strongly convex with modulus
1
L
\frac{1}{L}
L1.
Furthermore, we have the inequalities
l
−
1
λ
<
L
<
l
−
1
λ
\frac{l-1}{\lambda}<L<\frac{l-1}{\lambda}
λl−1<L<λl−1.
proof:[REF:Lemar C , chal, Sagastiz C A . Practical Aspects of the Moreau–Yosida Regularization[J]. SIAM Journal on Optimization, 1997.]
Theorem3:
The function
F
(
y
)
F(y)
F(y) is finite everywhere,convex and differentiable ;its gradient is
G
(
y
)
=
∇
F
(
y
)
=
s
M
(
y
)
=
M
(
y
−
p
(
y
)
)
G(y)=\nabla F(y)=s_{M}(y)=M(y-p(y))
G(y)=∇F(y)=sM(y)=M(y−p(y)).
Furthermore,there holds for all
y
y
y and
y
′
y^{'}
y′ in
R
n
R^{n}
Rn,
⟨
s
M
(
y
)
−
s
M
(
y
′
)
,
W
(
s
M
(
y
)
−
s
M
(
y
′
)
)
⟩
≤
⟨
s
M
(
y
)
−
s
M
(
y
′
)
,
y
−
y
′
⟩
\langle s_{M}(y)-s_{M}(y^{'}),W(s_{M}(y)-s_{M}(y^{'}))\rangle \leq \langle s_{M}(y)-s_{M}(y^{'}),y-y^{'}\rangle
⟨sM(y)−sM(y′),W(sM(y)−sM(y′))⟩≤⟨sM(y)−sM(y′),y−y′⟩
and
∥
∇
F
(
y
)
−
∇
F
(
y
′
)
∥
≤
∥
M
∥
∥
y
−
y
′
∥
\parallel \nabla F(y)-\nabla F(y^{'})\parallel\leq \parallel M\parallel \parallel y-y^{'} \parallel
∥∇F(y)−∇F(y′)∥≤∥M∥∥y−y′∥
proof
Theorem4:
Minimizing
f
f
f and
F
F
F are equivalent problems,in the sense
inf
x
∈
R
n
F
(
x
)
=
inf
x
∈
R
n
f
(
x
)
\inf_{x\in R^{n}} F(x) = \inf_{x\in R^{n}} f(x)
x∈RninfF(x)=x∈Rninff(x)
an equality in
R
∪
{
−
∞
}
R\cup\lbrace -\infty\rbrace
R∪{−∞},and the following statements are equvalent.
(i)
x
x
x minimizes
f
f
f
(ii)
p
M
(
x
)
=
x
p_{M}(x)=x
pM(x)=x
(iii)
s
M
(
x
)
=
0
s_{M}(x)=0
sM(x)=0
(iv)
x
x
x minimizes
f
M
(
x
)
f_{M}(x)
fM(x)
(v)
f
(
p
M
(
x
)
)
=
f
(
x
)
f(p_{M}(x))=f(x)
f(pM(x))=f(x)
(vi)
f
M
(
x
)
=
f
(
x
)
f_{M}(x)=f(x)
fM(x)=f(x)
We can use the corollary 2.1.4 to proov the equivalent of
inf
F
(
x
)
\inf F(x)
infF(x) and
inf
(
x
)
\inf (x)
inf(x)
proof: On J.-B. Hiriart-Urruty and C. Lemar echal, Convex Analysis and Minimization Algorithms, Springer-Verlag, Berlin, New York, 1993.XV,TH4.1.7
Proposition 1
Assume
f
f
f is a closed convex function. Then
∇
F
(
)
\nabla F ( )
∇F() has directional derivatives if and only if
p
(
)
p( )
p() has directional derivatives. The Hessian
∇
2
F
(
x
)
\nabla^{2}F (x)
∇2F(x) exists if and only if the Jacobian
∇
p
(
x
)
\nabla p(x)
∇p(x) exists:
∇
2
F
(
x
)
=
M
(
I
−
∇
p
(
x
)
)
,
f
o
r
a
l
l
x
∈
R
n
\nabla ^{2}F(x)=M(I-\nabla p(x)),for\quad all \quad x\in R^{n}
∇2F(x)=M(I−∇p(x)),forallx∈Rn
Reference:
[1] Lemar C , chal, Sagastiz C A . Practical Aspects of the Moreau–Yosida Regularization[J]. SIAM Journal on Optimization, 1997.
[2]J.-B. Hiriart-Urruty and C. Lemar echal, Convex Analysis and Minimization Algorithms, Springer-Verlag, Berlin, New York, 1993