Optimization Reminders Part I
Jean-Christophe Pesquet Center for Visual Computing
Inria, CentraleSup´elec Institut Universitaire de France
jean-christophe@pesquet.eu
Reference books
◮ D. Bertsekas, Nonlinear programming, Athena Scientic, Belmont, Massachussets, 1996.
◮ Y. Nesterov, Introductory Lectures on Convex Optimization : A Basic Course, Springer, 2004.
◮ S. Boyd and L. Vandenberghe, Convex optimization, Cambridge University Press, 2004.
◮ H. H. Bauschke and P. L. Combettes, Convex Analysis and Monotone Operator Theory in Hilbert Spaces, Springer, New York, 2017.
A few mathematical background
Hilbert spaces
A (real) pre-Hilbertian spaceH is a real vector space endowed with a function h· | ·i: H2→R, called inner product such that
➀ (∀(x,y)∈ H2)hx |yi=hy |xi
➁ (∀(x,y,z)∈ H3) hx +y |zi=hx |zi+hy |zi
➂ (∀(x,y)∈ H2) (∀α∈R)hαx |yi=αhy |xi
➃ (∀x ∈ H)hx |xi ≥0 and hx |xi= 0 ⇔ x = 0.
◮ The associated norm is
(∀x ∈ H) kxk=p hx |xi.
Hilbert spaces
A (real) Hilbert spaceH is a complete pre-Hilbertian space.
◮ Particular case :H=RN (Euclidean space with dimensionN).
Hilbert spaces
A (real) Hilbert spaceH is a complete pre-Hilbertian space.
◮ Particular case :H=RN (Euclidean space with dimensionN).
2H is the power set of H, i.e. the family of all subsets of H.
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
kxkH≤1
kLxkG <+∞
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
kxk≤1
kLxk<+∞
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
kxk=1
kLxk<+∞
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
x∈H\{0}
kLxk
kxk <+∞
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
x∈H\{0}
kLxk
kxk <+∞
◮ A linear operator fromHto G is continuous if and only if it is bounded.
◮ In finite dimension, every linear operator is bounded.
◮ In the following, it will be assumed that all the underlying Hilbert spaces are finite dimensional.
Hilbert spaces
Let HandG be two Hilbert spaces.
A linear operator L:H → G is bounded if kLk= sup
x∈H\{0}
kLxk
kxk <+∞
◮ A linear operator fromHto G is continuous if and only if it is bounded.
◮ In finite dimension, every linear operator is bounded.
◮ In the following, it will be assumed that all the underlying Hilbert spaces are finite dimensional.
B(H,G) : Banach space of (bounded) linear operators fromH to G.
Hilbert spaces
Let HandG be two Hilbert spaces.
Let L∈ B(H,G). Its adjointL∗ is the operator inB(G,H) defined as (∀(x,y) ∈ H × G) hy |LxiG =hL∗y |xiH.
Hilbert spaces
Let HandG be two Hilbert spaces.
Let L∈ B(H,G). Its adjointL∗ is the operator inB(G,H) defined as (∀(x,y)∈ H × G) hy |Lxi=hL∗y |xi.
Hilbert spaces
Let HandG be two Hilbert spaces.
Let L∈ B(H,G). Its adjointL∗ is the operator inB(G,H) defined as (∀(x,y)∈ H × G) hLx |yi=hx |L∗yi.
Hilbert spaces
Let HandG be two Hilbert spaces.
Let L∈ B(H,G). Its adjointL∗ is the operator inB(G,H) defined as (∀(x,y)∈ H × G) hLx |yi=hx |L∗yi.
Example :
If L:H → Hn:x 7→(x, . . . ,x) then L∗:Hn→ H:y = (y1, . . . ,yn)7→
Xn i=1
yi
Proof :
hLx |yi=h(x, . . . ,x)|(y1, . . . ,yn)i= Xn
i=1
hx |yii=
* x|
Xn
i=1
yi
+
Hilbert spaces
Let HandG be two Hilbert spaces.
Let L∈ B(H,G). Its adjointL∗ is the operator inB(G,H) defined as (∀(x,y)∈ H × G) hLx |yi=hx |L∗yi.
◮ We havekL∗k=kLk.
◮ If Lis bijective (i.e. an isomorphism ) thenL−1 ∈ B(G,H) and (L−1)∗ = (L∗)−1.
◮ If H=RN andG=RM then L∗=L⊤.
Infinite values functions
Functional analysis : definitions
LetS be a nonempty set of a Hilbert space H.
Letf : S →]− ∞, +∞] .
◮ The domain off isdomf ={x ∈S|f(x)<+∞}.
◮ The functionf is proper ifdomf 6=∅.
Domains of the functions ?
x f(x)
x f(x)
δ
Functional analysis : definitions
LetS be a nonempty set of a Hilbert space H.
Letf : S →]− ∞, +∞] .
◮ The domain off isdomf ={x ∈S|f(x)<+∞}.
◮ The functionf is proper ifdomf 6=∅.
Domains of the functions ?
x f(x)
domf =R (proper)
x f(x)
δ
Functional analysis : definitions
LetS be a nonempty set of a Hilbert space H.
Letf : S →]− ∞, +∞] .
◮ The domain off isdomf ={x ∈S|f(x)<+∞}.
◮ The functionf is proper ifdomf 6=∅.
Domains of the functions ?
x f(x)
domf =R (proper)
x f(x)
δ
domf =]0, δ]
(proper)
Functional analysis : definitions
Let C ⊂ H.
The indicator function of C is (∀x∈ H) ιC(x) =
(0 ifx ∈C +∞ otherwise.
Example :C = [δ1, δ2]
f(x) =ι[δ1,δ2](x)
δ1 δ2 x
Limits inf and sup
Let (ξn)n∈N be a sequence of elements in [−∞,+∞].
Its infimum limit is lim infξn= limn→+∞inf ξk
k ≥n ∈[−∞,+∞]
and its supremum limit is lim supξn = limn→+∞sup
ξk k ≥n ∈ [−∞,+∞].
Limits inf and sup
Let (ξn)n∈N be a sequence of elements in [−∞,+∞].
Its infimum limit is lim infξn= limn→+∞inf ξk
k ≥n ∈[−∞,+∞]
and its supremum limit is lim supξn = limn→+∞sup
ξk k ≥n ∈ [−∞,+∞].
◮ lim supξn=−lim inf(−ξn)
◮ lim infξn≤lim supξn
◮ limn→+∞ξn=ξ∈[−∞,+∞] if and only if lim infξn= lim supξn=ξ.
Epigraph
Let f :H → ]−∞,+∞]. The epigraph of f is epif =
(x, ζ)∈domf ×R f(x)≤ζ
Epigraph
Let f :H → ]−∞,+∞]. The epigraph of f is epif =
(x, ζ)∈domf ×R f(x)≤ζ
x f(x) =|x|
epif
x
−δ δ
f(x) =ι[−δ,δ](x) epif
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous (l.s.c.) function at x ∈ H if, for every se- quence (xn)n∈Nof H,
xn→x ⇒ lim inff(xn)≥f(x).
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous function on Hif and only if epif is closed
◮ l.s.c. functions ?
x f(x)
x f(x)
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous function on Hif and only if epif is closed
◮ l.s.c. functions ?
f(x)
x x
f(x)
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous function on Hif and only if epif is closed
◮ l.s.c. functions ?
x f(x)
x f(x)
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous function on Hif and only if epif is closed
◮ l.s.c. functions ?
x f(x)
x f(x)
Lower semi-continuity
Let f :H → ]−∞,+∞].
f is a lower semi-continuous function on Hif and only if epif is closed
◮ l.s.c. functions ?
x f(x)
x f(x)
Lower semi-continuity
◮ Every continuous function onH is l.s.c.
◮ Every finite sum of l.s.c. functions is l.s.c.
◮ Let (fi)i∈I be a family of l.s.c functions.
Then, supi∈Ifi is l.s.c.
Does a minimum exist ?
Minimizers
Let S be a nonempty set of a Hilbert spaceH.
Letf :S → ]−∞,+∞] be a proper function and let xb∈S.
◮ bx is a local minimizer of f ifxb∈domf and there exists an open neigborhood O of xbsuch that
(∀x∈O∩S) f(x)b ≤f(x).
◮ bx is a (global) minimizer of f if
(∀x∈S) f(x)b ≤f(x).
Minimizers
Let S be a nonempty set of a Hilbert spaceH.
Letf :S → ]−∞,+∞] be a proper function and let xb∈S.
◮ bx is a strict local minimizer of f if there exists an open neigborhood O ofxbsuch that
(∀x ∈(O∩S)\ {xb}) f(bx)<f(x).
◮ bx is a strict (global and unique) minimizer of f if (∀x∈S\ {x})b f(xb)<f(x).
Existence of a minimizer
Weierstrass theorem
Let S be a nonempty compact set of a Hilbert spaceH.
Letf :S → ]−∞,+∞] be a proper l.s.c function.
Then, there exists xb∈S such that f(xb) = inf
x∈Sf(x).
Existence of a minimizer
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is coercive if limkxk→+∞f(x) = +∞.
Existence of a minimizer
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is coercive if limkxk→+∞f(x) = +∞.
Coercive functions ?
x f(x) +∞
x f(x) +∞
x f(x) +∞
Existence of a minimizer
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is coercive if limkxk→+∞f(x) = +∞.
Coercive functions ?
x f(x) +∞
x f(x) +∞
x f(x) +∞
Existence of a minimizer
Theorem
Let Hbe a (finite dimensional) Hilbert space.
Let f :H → ]−∞,+∞] be a proper l.s.c. coercive function.
Then, the set of minimizers of f is a nonempty compact set.
Existence of a minimizer
Theorem
Let Hbe a (finite dimensional) Hilbert space.
Let f :H → ]−∞,+∞] be a proper l.s.c. coercive function.
Then, the set of minimizers of f is a nonempty compact set.
Proof : Sincef is proper, there exists x0 ∈ H such thatf(x0)∈R. The coercivity off implies that there existsη∈ ]0,+∞[ such that, for every x ∈ H satisfyingkx−x0k> η,f(x)>f(x0).
LetS =
x ∈ H kx−x0k ≤η ,S∩domf 6=∅andS is compact.
Then, there existsxb∈S such that f(x) = infb x∈Sf(x)≤f(x0). Thus, f(x) = infb x∈Hf(x).
Argminf ⊂S is bounded. In addition, if (xn)n∈N is a sequence of
minimizers converging tobx∈ H. Then,f(x)b ≤lim inff(xn) = infx∈Hf(x) and, consequently, bx∈Argminf. Therefore,Argminf is closed.
Exercise 1
Letf be the Shannon entropy function defined as
f(x) =
XN
i=1
x(i)lnx(i) ifx = (x(i))1≤i≤N ∈ ]0,+∞[N +∞ if (∃j ∈ {1, . . . ,N}) x(j) <0.
1. How can we extend the definition of functionf so that it is l.s.c. on RN?
2. What can be said about the existence of a minimizer of this function on a nonempty closed subset of the set
C =
(x(i))1≤i≤N ∈[0,+∞[N PNi=1x(i)= 1 ?
Hints from differential calculus
Necessary condition for the existence of a minimizer (Euler’s inequality)
Theorem
LetD be an open subset of a Hilbert space Hand letC ⊂D. Let f: D→ ]−∞,+∞] be differentiable at xb∈ C. If bx is a local minimizer off on C then, for every y ∈ H such that [xb,y]⊂C,
h∇f(xb)|y −xbi ≥0.
If xb∈int(C), then the condition reduces to
∇f(xb) = 0.
Necessary condition for the existence of a minimizer (Euler’s inequality)
Theorem
LetD be an open subset of a Hilbert space Hand letC ⊂D. Let f: D→ ]−∞,+∞] be differentiable at xb∈ C. If bx is a local minimizer off on C then, for every y ∈ H such that [xb,y]⊂C,
h∇f(xb)|y −xbi ≥0.
If xb∈int(C), then the condition reduces to
∇f(xb) = 0.
Remark : A zero of the gradient∇f is called a critical point of f.
Necessary condition for the existence of a minimizer
Theorem
Let C be an open subset of a Hilbert space H. Let f : C → R be diffe- rentiable on C. Let f be twice differentiable at xb ∈ C. If If bx is a local minimizer off on C, then
(i) ∇f(bx) = 0 ;
(ii) the Hessian∇2f(xb) off atxbis positive semi-definite, i.e.
(∀z ∈ H)
z | ∇2f(xb)z
≥0.
Sufficient conditions for the existence of a minimizer
Theorem
Let C be an open subset of a Hilbert space H. Let f: C →R be differen- tiable onC.
(i) Iff is twice differentiable atxb∈C, ∇f(bx) = 0 and the Hessian
∇2f(bx) off atbx is positive definite, i.e.
(∀z ∈ H \ {0})
z | ∇2f(xb)z
>0.
thenf has a strict local minimum atbx.
Sufficient conditions for the existence of a minimizer
Theorem
Let C be an open subset of a Hilbert space H. Let f: C →R be differen- tiable onC.
(i) Iff is twice differentiable atxb∈C, ∇f(bx) = 0 and the Hessian
∇2f(bx) off atbx is positive definite, i.e.
(∀z ∈ H \ {0})
z | ∇2f(xb)z
>0.
thenf has a strict local minimum atbx.
(ii) Iff is twice differentiable on an open neighborhood D⊂C of bx,
∇f(bx) = 0 and the Hessian off is positive semi-definite on D, i.e.
(∀x ∈D)(∀z ∈ H)
z| ∇2f(x)z
≥0, thenf has a local minimum atxb.
Magic of convexity
Convex set : definition
Let Hbe a Hilbert space. C ⊂ His a convex set if
(∀(x,y)∈C2)(∀α∈]0,1[) αx+ (1−α)y ∈C
Convex sets ?
C C C
Convex set : definition
Let Hbe a Hilbert space. C ⊂ His a convex set if
(∀(x,y)∈C2)(∀α∈]0,1[) αx+ (1−α)y ∈C
Convex sets ?
C C C
Convex set : properties
◮ ∅is considered as a convex set.
◮ If C is a convex set, then (∀n∈N∗) (∀(x1, . . . ,xn)∈Cn) (∀(α1, . . . , αn)∈[0,+∞[n) with
Xn i=1
αi = 1, Xn
i=1
αixi ∈C.
◮ Every vector (affine) space is convex.
◮ If C is a convex set, thenint(C) andC are convex sets.
Convex set : properties
◮ If C is a convex set then, for every α∈R,
αC =
αx x ∈C is a convex set.
◮ If C1 andC2 are convex sets, then C1×C2
C1+C2=
x1+x2 (x1,x2)∈C1×C2 are convex sets.
◮ If (Ci)i∈I is a family of convex sets ofH, then \
i∈I
Ci is convex.
Convex hull
LetHbe a Hilbert space andC ⊂ H. The convex hull ofC is the smallest convex set includingC. It is denoted by conv(C).
◮ conv(C) is the intersection of all the convex sets including C.
◮ Letx ∈ H.x ∈conv(C) if and only if (∃n∈N∗) (∃(x1, . . . ,xn)∈Cn) (∃(α1, . . . , αn)∈ ]0,+∞[n with
Xn i=1
αi = 1 such that
x= Xn i=1
αixi.
Convex function : definitions
f :H → ]−∞,+∞] is a convex function if
∀(x,y)∈ H2
(∀α∈]0,1[)
f(αx + (1−α)y)≤αf(x) + (1−α)f(y)
Convex function : definitions
f :H → ]−∞,+∞] is a convex function if
∀(x,y)∈(domf)2
(∀α∈]0,1[)
f(αx+ (1−α)y)≤αf(x) + (1−α)f(y)
Convex function : definitions
f :H → ]−∞,+∞] is a convex function if
∀(x,y)∈(domf)2
(∀α∈]0,1[)
f(αx+ (1−α)y)≤αf(x) + (1−α)f(y)
Convex functions ?
x f(x) =|x|
x f(x) =p
|x|
+∞sinon)
x
−δ δ
f(x) =ι[−δ,δ](x) (0 six ∈[−δ, δ]
Convex function : definitions
f :H → ]−∞,+∞] is a convex function if
∀(x,y)∈(domf)2
(∀α∈]0,1[)
f(αx+ (1−α)y)≤αf(x) + (1−α)f(y)
Convex functions ?
x f(x) =|x|
x f(x) =p
|x|
+∞sinon)
x
−δ δ
f(x) =ι[−δ,δ](x) (0 six ∈[−δ, δ]
Convex function : definitions
f :H → ]−∞,+∞] is a convex function if
∀(x,y)∈(domf)2
(∀α∈]0,1[)
f(αx+ (1−α)y)≤αf(x) + (1−α)f(y)
◮ f : H →[−∞,+∞[ is concave if −f is convex.
Convex functions : definition
f :H → ]−∞,+∞] is convex⇔ its epigraph is a convex set.
Convex functions : definition
f :H → ]−∞,+∞] is convex⇔ its epigraph is a convex set.
x f(x) =|x|
epif
x f(x) =p
|x|
epif
x
−δ δ
f(x) =ι[−δ,δ](x) epif
Convex functions : definition
f :H → ]−∞,+∞] is convex⇔ its epigraph is a convex set.
x f(x) =|x|
epif
x f(x) =p
|x|
epif
x
−δ δ
f(x) =ι[−δ,δ](x epif
Convex functions : properties
◮ If f:H → ]−∞,+∞] is convex, then domf is convex and its lower level set at height η∈R
lev≤η f =
x∈ H f(x)≤η is a convex set.
◮ f:H → ]−∞,+∞] is convex if and only if (∀(x,y)∈(domf)2) ϕx,y: [0,1]→ ]−∞,+∞] :α7→f(αx+ (1−α)y) is convex.
Convex functions : properties
◮ Every finite sum of convex functions is convex.
◮ Let (fi)i∈I be a family of convex functions. Then, supi∈Ifi is convex.
◮ Γ0(H) : class of convex, l.s.c., and proper functions from Hto ]−∞,+∞].
◮ LetC ⊂ H.
ιC ∈Γ0(H) ⇔ C is a nonempty closed convex set.
Proof : epiιC =C ×[0,+∞[.
Strictly convex functions
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is strictly convex if
(∀x∈domf)(∀y ∈domf)(∀α∈]0,1[)
x6=y ⇒ f(αx+ (1−α)y)< αf(x) + (1−α)f(y).
Strictly convex functions
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is strictly convex if
(∀x∈domf)(∀y ∈domf)(∀α∈]0,1[)
x6=y ⇒ f(αx+ (1−α)y)< αf(x) + (1−α)f(y).
Strictly convex functions ?
x f(x)
x f(x)
x f(x)
Strictly convex functions
Let Hbe a Hilbert space. Letf :H → ]−∞,+∞].
f is strictly convex if
(∀x∈domf)(∀y ∈domf)(∀α∈]0,1[)
x6=y ⇒ f(αx+ (1−α)y)< αf(x) + (1−α)f(y).
Strictly convex functions ?
x f(x)
x f(x)
x f(x)
Minimizers of a convex function
Theorem
Let H be a Hilbert space. Let f: H → ]−∞,+∞] be a proper convex function such that µ= inff >−∞.
◮
x ∈ H f(x) =µ is convex.
◮ Every local minimizer off is a global minimizer.
◮ If f is strictly convex, then there exists at most one minimizer.
Minimizers of a convex function
Theorem
Let H be a Hilbert space. Let f: H → ]−∞,+∞] be a proper convex function such that µ= inff >−∞.
◮
x ∈ H f(x) =µ is convex.
◮ Every local minimizer off is a global minimizer.
◮ If f is strictly convex, then there exists at most one minimizer.
Proof : Let Ω =
x∈ H f(x) =µ . Let (x,y)∈Ω2 and letα∈[0,1].
We have
f(αx+ (1−α)y)≤αf(x) + (1−α)f(y) =µ which shows that αx+ (1−α)y ∈Ω.
Minimizers of a convex function
Theorem
Let H be a Hilbert space. Let f: H → ]−∞,+∞] be a proper convex function such that µ= inff >−∞.
◮
x ∈ H f(x) =µ is convex.
◮ Every local minimizer off is a global minimizer.
◮ If f is strictly convex, then there exists at most one minimizer.
Proof : Letxbbe a local minimizer off. For everyy ∈ H \ {x}, thereb exists α∈]0,1[ such that
f(x)b ≤f xb+α(y−bx)
≤(1−α)f(bx) +αf(y)
⇒ f(x)b ≤f(y)
If f is strictly convex, the inequality is strict.
Existence and uniqueness of a minimizer
Theorem
LetHbe a Hilbert space andC a closed convex subset ofH. Letf ∈Γ0(H) such that domf ∩C 6=∅.
If f is coercive or C is bounded, then there existsxb∈C such that f(bx) = inf
x∈Cf(x).
If, moreover, f is strictly convex, this minimizerxbis unique.
Exercise 2
LetH be a Hilbert space.
1. Show that the functionx 7→ kxk2 is strictly convex.
2. A functionf :H → ]−∞,+∞] is strongly convexwith modulus β∈ ]0,+∞[ if there exists a convex functiong:H → ]−∞,+∞]
such that
f =g +β 2k · k2.
Show that that every strongly convex function is strictly convex.
3. Show that a functionf:H → ]−∞,+∞] is strongly convex with modulusβ ∈ ]0,+∞[ if and only if
(∀(x,y)∈ H2)(∀α∈]0,1[) f(αx+ (1−α)y) +α(1−α)β
2kx−yk2 ≤αf(x) + (1−α)f(y).
Convex + smooth
Characterization of differentiable convex functions
Let f: H → ]−∞,+∞] be differentiable on domf, which is a nonempty open convex set.
Then, f is convex if and only if
(∀(x,y)∈(domf)2) f(y)≥f(x) +h∇f(x)|y −xi.
Characterization of differentiable strictly convex functions
Let f: H → ]−∞,+∞] be differentiable on domf, which is a nonempty open convex set.
Then, f is strictly convex if and only if, for every (x,y) ∈ (domf)2 with x 6=y,
f(y)>f(x) +h∇f(x)|y −xi.
Characterization of differentiable convex functions
Let f :H → ]−∞,+∞] be differentiable on domf, which is a nonempty open convex set.
Then,f is convex if and only if∇f is monotone on domf, i.e.
(∀(x,y)∈(domf)2) h∇f(y)− ∇f(x)|y−xi ≥0.
Characterization of strictly differentiable convex functions
Let f :H → ]−∞,+∞] be differentiable on domf, which is a nonempty open convex set.
Then, f is strictly convex if and only if∇f is strictly monotone on domf, i.e. for every (x,y)∈(domf)2 withx 6=y,
h∇f(y)− ∇f(x)|y−xi>0.
Characterization of twice differentiable convex functions
LetHbe a Hilbert space.
Letf: H → ]−∞,+∞] be a twice differentiable function ondomf, which is a nonempty open convex set.
◮ f is convex if and only if, for every x ∈domf, ∇2f(x) is positive semi-definite.
◮ If, for everyx ∈domf, ∇2f(x) is positive definite, thenf is strictly convex.
Condition for the existence of a minimizer
Theorem
LetHbe Hilbert space.
Let f: H → ]−∞,+∞] be a differentiable convex function on domf, which is an open set. LetC ⊂domf be a nonempty convex set. xb∈C is a (global) minimizer of f on C if and only if
(∀y ∈C) h∇f(bx)|y −xbi ≥0.
If xb∈int(C), then the condition reduces to
∇f(xb) = 0.
Condition for the existence of a minimizer
Theorem
LetHbe Hilbert space.
Let f: H → ]−∞,+∞] be a differentiable convex function on domf, which is an open set. LetC ⊂domf be a nonempty convex set. xb∈C is a (global) minimizer of f on C if and only if
(∀y ∈C) h∇f(bx)|y −xbi ≥0.
If xb∈int(C), then the condition reduces to
∇f(xb) = 0.
Proof : We have already seen that the inequality is a necessary condition for b
x to be a local minimizer of f on C and that it reduces to the vanishing condition on the gradient ifxb∈int(C).
Condition for the existence of a minimizer
Theorem
LetH be Hilbert space.
Let f: H → ]−∞,+∞] be a differentiable convex function on domf, which is an open set. Let C ⊂domf be a nonempty convex set.xb∈C is a (global) minimizer off onC if and only if
(∀y ∈C) h∇f(x)b |y−bxi ≥0.
Ifbx ∈int(C), then the condition reduces to
∇f(x) = 0.b
Proof : Conversely, assume that the inequality holds. Lety ∈C. Sincef is convex and Gˆateaux differentiable,
f(y)≥f(bx) +h∇f(xb)|y−xbi ≥f(xb).
Hence, bx is a minimizer off onC.
Exercice 3
Letf :H →R be differentiable.
Show thatf isβ-strongly convex if and only if
(∀(x,y)∈ H2) f(y)≥f(x) +h∇f(x)|y −xi+β
2ky −xk2.
Exercice 4
Let
f: RN →R (x(i))1≤i≤N 7→ln
XN i=1
exp(x(i))
! .
Show that f is convex. Is it strictly convex ?
Projections
Projection onto a closed convex set
Theorem
LetC be a nonempty closed convex set of a Hilbert space H.
(i) For everyx ∈ H, there exists a unique pointxbinC which lies at minimum distance ofx. The applicationPC: H →C which maps everyx ∈ Hto its associated pointxbis called the projection ontoC . (ii) For everyx ∈ H, bx =PC(x) if and only ifxb∈C and
(∀y ∈C) hx −xb|y−bxi ≤0.
Geometrical interpretation
LetC be a nonempty subset of H.
For everyx ∈ H, the normal cone to C atx is defined as NC(x) =
(u∈ H (∀y∈C) hu|y−xi ≤0 ifx ∈C
∅ otherwise.
C
NC(x)
x u
C
NC(x)
x u
Geometrical interpretation
Let C be a nonempty subset ofH.
For everyx ∈ H, the normal cone toC atx is defined as NC(x) =
(u∈ H (∀y ∈C) hu |y−xi ≤0 if x∈C
∅ otherwise.
◮ If x∈intC, thenNC(x) ={0}.
◮ If C is a vector space, then for everyx ∈C,NC(x) =C⊥.
◮ LetC be nonempty closed convex set of a Hilbert space H. For every x∈ H,
b
x=PC(x) ⇔ x−xb∈NC(bx).
Examples of projections
◮ If C is a (closed) vector space of a Hilbert space H, thenPC is the (linear) orthogonal projection ontoC. Then, for every x ∈ H,
b
x =PC(x) ⇔
(xb∈C x−xb∈C⊥.
Properties of the projection
◮ LetC be a nonempty closed convex set of a Hilbert space H. The projection onto C is a firmly nonexpansive operator , i.e.
(∀(x,y)∈ H2) kPC(x)−PC(y)k2 ≤ hx−y|PC(x)−PC(y)i.
Properties of the projection
◮ LetC be a nonempty closed convex set of a Hilbert space H. The projection onto C is a firmly nonexpansive operator , i.e.
(∀(x,y)∈ H2) kPC(x)−PC(y)k2 ≤ hx−y|PC(x)−PC(y)i.
Properties of the projection
◮ LetC be a nonempty closed convex set of a Hilbert space H. The projection onto C is a firmly nonexpansive operator , i.e.
(∀(x,y)∈ H2) kPC(x)−PC(y)k2 ≤ hx−y|PC(x)−PC(y)i.
◮ The projection ontoC is a nonexpansive operator , i.e.
(∀(x,y)∈ H2) kPC(x)−PC(y)k ≤ kx−yk.
◮ The projection ontoC is uniformly continuous.
◮ The distance to C defined as (∀x∈ H) dC(x) = inf
y∈Ckx−yk=kx−PC(x)k is continuous.
Exercise 5
LetH andG be Hilbert spaces.
LetD be a nonempty closed convex set ofG.
LetL∈ B(H,G) be a bijective isometry and let C =
x∈ H Lx ∈D . 1. Show that the projection ontoC is well-defined.
2. Show thatPC =L∗◦PD◦L.
3. ExpressPC when H=G=RN andD = [0,+∞[N.