Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
MotivationPrevious ResultsRecent Results
Convex functions and convex cones on theEuclidean Space
Yu-Lin Chang
Department of MathematicsNational Taiwan Normal University
11 19-21, 2016
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Outline
1 Motivation
2 Previous Results
3 Recent Results
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Patr I
Motivation
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Monotonicity, Convexity and Cones
Some inequalities are meaningful.For example, arithmetic-geometric inequality is meaningful,Holder’s inequality is meaningful. But why are they meaningful? Iwould say that an inequality is meaningful if it comes from aconvex function. And where does convexity comes from? I wouldguess cones!!!
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Second Order Cone
The second-order cone (SOC) in Rn, also called Lorentz cone, isthe set defined as
Kn :={
(x1, x2) ∈ R× Rn−1 | x1 ≥ ‖x2‖}, (1)
where ‖ · ‖ denotes the Euclidean norm. When n = 1, Kn reducesto the set of nonnegative real numbers R+. Kn is also a setcomposed of the squared elements from Jordan algebra (Rn, ◦),where the Jordan product “◦” is a binary operation defined by
x ◦ y := (〈x , y〉, x1y2 + y1x2) (2)
for any x = (x1, x2), y = (y1, y2) ∈ R× Rn−1.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Spectral Decomposition
We recall that each x ∈ Rn admits a spectral decompositionassociated with Kn of the following form
x = λ1(x)u(1)x + λ2(x)u
(2)x , (3)
where λi (x) and u(i)x for i = 1, 2 are the spectral values and the
associated spectral vectors of x , respectively, defined by
λi (x) = x1 + (−1)i‖x2‖, u(i)x =
1
2
(1, (−1)i x2
), (4)
with x2 = x2‖x2‖ if x2 6= 0, and otherwise x2 being any vector in Rn−1
such that ‖x2‖ = 1. The determinant and trace of x are defined asdet(x) := λ1(x)λ2(x) and tr(x) := λ1(x) + λ2(x), respectively.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities I on SOC
Theorem
The following functions associated with Kn ⊆ Rn are all strictlyconvex with e = (1, 0).
(a) F1(x) = − ln(det(x)) for x ∈ intKn.
(b) F2(x) = tr(x−1) for x ∈ intKn.
(c) F3(x) = tr(h(x)) for x ∈ intKn, where
h(x) =
{xp+1−ep+1 + x1−q−e
q−1 if p ∈ [0, 1], q > 1;xp+1−ep+1 − ln x if p ∈ [0, 1], q = 1.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities II on SOC
Theorem (Cont.)
The following functions associated with Kn ⊆ Rn are all strictlyconvex with e = (1, 0).
(d) F4(x) = − ln(det(e − x)) for x ≺Kn e.
(e) F5(x) = tr((e − x)−1 ◦ x) for x ≺Kn e.
(f) F6(x) = tr(exp(x)) for x ∈ Rn.
(g) F7(x) = ln(det(e + exp(x))) for x ∈ Rn.
(h) F8(x) = tr
(x + (x2 + 4e)1/2
2
)for x ∈ Rn.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities III on SOC
Theorem (Cont.)
For any x ∈ K, let F9(x) := −[det(x)]p with 0 < p < 1. Then,
(a) F9 is twice continuously differentiable in intK.
(b) F9 is convex when p ≤ 12 , and moreover, it is strictly
convex when p < 12 in intK.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities IV on SOC
Theorem (Cont.)
Suppose x ∈ Kn and y ∈ Kn. Then the following inequalities hold.
(a) det(αx + (1− α)y) ≥ (det(x))α(det(y))1−α for any0 < α < 1.
(b) det(x + y)1/p ≥ 22p−1 (
det(x)1/p + det(y)1/p)
forany p ≥ 2.
(c) det(x + y)1/2 ≥ det(x)1/2 + det(y)1/2 anddet(x + y) ≥ det(x) + det(y).
(d) det(αx + (1− α)y) ≥ α2det(x) + (1− α)2det(y) forany 0 < α < 1.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities V on SOC
Theorem (Cont.)
Suppose x ∈ Kn and y ∈ Kn. Then the following inequalities hold.
(e) [det(e + x)]1/2 ≥ 1 + det(x)1/2.
(f) If x �K y , then det(x) > det(y).
(g) tr(x ◦ y) ≥ 2 det(x)1/2det(y)1/2.
(h) det(x)1/2 =
min
{1
2tr(x ◦ y) : det(y) = 1, y �K 0
}.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Question
How can we prove this? Before that let us take a look atsymmetric cone.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Jordan Algebra
Let V be an n-dimensional vector space over the real field R,endowed with a bilinear mapping (x , y) 7→ x ◦ y from V× V intoV. The pair (V, ◦) is called a Jordan algebra if
(i) x ◦ y = y ◦ x for all x , y ∈ V,
(ii) x ◦ (x2 ◦ y) = x2 ◦ (x ◦ y) for all x , y ∈ V.
Note that a Jordan algebra is not necessarily associative, i.e.,x ◦ (y ◦ z) = (x ◦ y) ◦ z may not hold for all x , y , z ∈ V. We call anelement e ∈ V the identity element if x ◦ e = e ◦ x = x for allx ∈ V. A Jordan algebra (V, ◦) with an identity element e is calleda Euclidean Jordan algebra if there is an inner product 〈·, ·〉Vsuch that
(iii) 〈x ◦ y , z〉V = 〈y , x ◦ z〉V for all x , y , z ∈ V.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Symmetric Cones
Given a Euclidean Jordan algebra A = (V, ◦, 〈·, ·〉V), we denote theset of squares as
K :={x2 | x ∈ V
}.
K is a symmetric cone which means that K is a self-dual closedconvex cone with nonempty interior and for any two elementsx , y ∈ intK, there exists an invertible linear transformationT : V→ V such that T (K) = K and T (x) = y .
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Rank
For any given x ∈ A, let ζ(x) be the degree of the minimalpolynomial of x , i.e.,
ζ(x) := min{k : {e, x , x2, · · · , xk} are linearly dependent
}.
Then the rank of A is defined as max{ζ(x) : x ∈ V}. In this paper,we use r to denote the rank of the underlying Euclidean Jordanalgebra.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Jordan Frame
Recall that an element c ∈ V is idempotent if c2 = c . Twoidempotents ci and cj are said to be orthogonal if ci ◦ cj = 0. Onesays that {c1, c2, . . . , ck} is a complete system of orthogonalidempotents if
c2j = cj , cj◦ci = 0 if j 6= i for all j , i = 1, 2, · · · , k and∑k
j=1 cj = e.
An idempotent is primitive if it is nonzero and cannot be writtenas the sum of two other nonzero idempotents. We call a completesystem of orthogonal primitive idempotents a Jordan frame. Nowwe state the second version of the spectral decomposition theorem.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Decomposition Theorem
Theorem
Suppose that A is a Euclidean Jordan algebra with rank r . Thenfor any x ∈ V, there exists a Jordan frame {c1, . . . , cr} and realnumbers λ1(x), . . . , λr (x), arranged in the decreasing orderλ1(x) ≥ λ2(x) ≥ · · · ≥ λr (x), such that
x = λ1(x)c1 + λ2(x)c2 + · · ·+ λr (x)cr .
The numbers λj(x) (counting multiplicities), which are uniquelydetermined by x , are called the eigenvalues.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Trace Function and Euclidean Structure
We can define trace function as the following
tr(x) =r∑
j=1
λj(x).
It is shown that 〈x , y〉 = tr(x ◦ y) defines a Euclidean structure onA. We also define determinant function det(x) = λ1λ2 · · ·λr .
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Lowner Operator
Suppose φ : R→ R be a scalar valued function and we define theLowner operator associated with φ as
φscV (x) =:r∑
j=1
φ(λj(x))cj ,
where x ∈ V has the spectral decomposition x =∑r
j=1 λj(x)cj .The associated trace function on Lowner Operator is defined asφtrV :=
∑rj=1 φ(λj(x)).
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Koranyi’s Results
Theorem
Let x =∑r
j=1 λj(x)cj . If φ is continuously differentiable, then φscVis differentiable at x and its derivative, for any h ∈ V, is given by
(∇φscV
)(x)(h) =
r∑i=1
(φ[1](λ(x))
)iiCii (x)h (5)
+∑
1≤j<l≤r
(φ[1](λ(x))
)jlCjl(x)h
where the coefficient is defined as
φ[1](λ(x))jl :=
{φ′(λj) if λj = λl ,
φ(λj )−φ(λl )λj−λl if λj 6= λl .
(6)
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Part II
Previous Results
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Differential Properties
Lemma
For any given scalar function φ : J ⊆ R→ R, let φscV : S → V andφtrV : S → R be defined previously. Assume that J is an openinterval in R. Then, the following results hold.
(a) The domain S of φscV and φtrV is open and convex.
(b) If φ is (continuously) differentiable, then φtrV is(continuously) differentiable on S with∇φtrV (x)(h) = 〈h, (φ′)scV (x)〉 for all h ∈ V.
(c) If φ is twice (continuously) differentiable, then φtrV istwice (continuously) differentiable on S with∇2φtrV (x)(h, k) = 〈h,∇(φ
′)scV (x)k〉 for all h, k ∈ V.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Convexity Preservation
Theorem
For any given φ : J ⊆ R→ R, let φscV : S ⊆ V→ Rn andφtrV : S ⊆ V→ R be given previously. Assume that J is an openinterval in R. If φ is twice differentiable on J, then
(a) φ′′
(t) ≥ 0 for any t ∈ J ⇐⇒ ∇2φtrV (x) � 0 for anyx ∈ S ⇐⇒ φtrV is convex in S .
(b) φ′′
(t) > 0 for any t ∈ J ⇐⇒ ∇2φtrV (x) � 0∀x ∈ S =⇒ φtrV is strictly convex in S .
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities on SC I
Theorem
The following functions associated with K are all strictly convex.
(a) F1(x) = − ln(det(x)) for x ∈ intK.
(b) F2(x) = tr(x−1) for x ∈ intK.
(c) F3(x) = tr(h(x)) for x ∈ intK, where
h(x) =
{xp+1−ep+1 + x1−q−e
q−1 if p ∈ [0, 1], q > 1;xp+1−ep+1 − ln x if p ∈ [0, 1], q = 1.
(d) F4(x) = − ln(det(e − x)) for x ≺K e.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities on SC II
Theorem (Cont.)
The following functions associated with K are all strictly convex.
(e) F5(x) = tr((e − x)−1 ◦ x) for x ≺K e.
(f) F6(x) = tr(exp(x)) for x ∈ V.
(g) F7(x) = ln(det(e + exp(x))) for x ∈ V.
(h) F8(x) = tr
(x + (x2 + 4e)1/2
2
)for x ∈ V.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities on SC III
Theorem (Cont.)
For any x ∈ K, let F9(x) := −[det(x)]p with 0 < p < 1. Then,
(a) F9 is twice continuously differentiable in intK.
(b) F9 is convex when p ≤ 1r , and moreover, it is strictly
convex when p < 1r in intK where r is the rank of
the Euclidean Jordan Algebra.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities on SC VI
Theorem (Cont.)
Suppose r is the rank of the Euclidean Jordan Algebra. For anyx �K 0 and y �K 0, the following inequalities hold.
(a) det(αx + (1− α)y) ≥ (det(x))α(det(y))1−α for any0 < α < 1.
(b) det(x + y)1/p ≥ 2rp−1 (
det(x)1/p + det(y)1/p)
forany p ≥ r .
(c) det(x + y)1/r ≥ det(x)1/r + det(y)1/r anddet(x + y) ≥ det(x) + det(y).
(d) det(αx + (1− α)y) ≥ αrdet(x) + (1− α)rdet(y) forany 0 < α < 1.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Inequalities on SC V
Theorem (Cont.)
Suppose r is the rank of the Euclidean Jordan Algebra. For anyx �K 0 and y �K 0, the following inequalities hold.
(e) [det(e + x)]1/r ≥ 1 + det(x)1/r .
(f) If x �K y , then det(x) > det(y).
(g) tr(x ◦ y) ≥ r det(x)1/rdet(y)1/r .
(h) det(x)1/r =
min
{1
rtr(x ◦ y) : det(y) = 1, y �K 0
}.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Proof of III
Proof. (a) Note that det(x) = exp(tr(ln x)) for any x ∈ intK,and tr(ln(x)) = φtrV (x) with φ(t) = ln(t) for t > 0. By Lemmapart(c), tr(ln(x)) is twice continuously differentiable in intK, andhence det(x) is twice continuously differentiable in intK. Sincet1/p with p > 1 is infinite continuously differentiable on (0,+∞),the desired result then follows.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Proof of III
(b) We now prove the convexity over intK. From our constructionwe have that F9(x) = − exp(pφtrV (x)) with φ(t) = ln(t). By directcomputation we get
∇2F9(x)(h, k) = − exp(pφtrV (x))[p2∇φtrV (x)(h)∇φtrV (x)(k)
+p∇2φtrV (x)(h, k)]
From Lemma and Theorem of Koranyi we obtain
∇2φtrV (x) =r∑
i=1
−1
λiλiCii (x) +
∑1≤j<l≤r
−1
λjλlCjl(x)
where λi (1 ≤ i ≤ r) are the eigenvalues of x .
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Cases of Symmetric Cones
Theorem
Every simple Euclidean Jordan algebra is isomorphic to one of thefollowing
(i) The Jordan spin algebra Kn.
(ii) The algebra Sn of n × n real symmetric matrices.
(iii) The algebra Hn of all n × n complex Hermitianmatrices.
(iv) The algebra Qn of all n × n quaternion Hermitianmatrices.
(v) The algebra O3 of all 3× 3 octonion Hermitianmatrices.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Part III
Recent Results
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
What have we just done?(i) Given any one variable convex function f : R→ R.(ii) We construct a convex function F : Rn → R on Rn.(iii) And this construction heavily depends on the symmetric conestructure on the Euclidean spaces. So it is natural to ask that
What would happen tonon-symmetric cones?
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
An application
Define a function d(x , y) = tr(x + y)− 2tr(P(x1/2)y)1/2. WhereP(x) = 2L2(x)− L(x2) is the quadratic representation. We canshow that
Theorem
For any x , y ∈ K, assume that x and y operator commute. Then(a) d(x , y) ≥ 0.(b) d(x , y) = 0⇐⇒ x = y .(c) d(x , y) = d(y , x).
Moreover d(x , y) is convex.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Circular Cone
An extension of second-order cone (SOC) in Rn, which is calledcircular cone, is the set defined as
Lθ :={
(x1, x2) ∈ R× Rn−1 | x1 ≥ tan θ‖x2‖}, (7)
where ‖ · ‖ denotes the Euclidean norm.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Spectral Decomposition
Similar to second order cone, each x ∈ Rn admits a spectraldecomposition associated with Lθ of the following form
x = λ1(x)u(1)x + λ2(x)u
(2)x , (8)
where λi (x) and u(i)x for i = 1, 2 are the spectral values and the
associated spectral vectors of x . The determinant and trace of xare defined as det(x) := λ1(x)λ2(x) and tr(x) := λ1(x) + λ2(x),respectively.
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Convexity Preservation(Cont.)
Theorem
For any given φ : J ⊆ R→ R, let φtrV : S ⊆ V→ R be givenpreviously. Assume that J is an open interval in R. If φ is twicedifferentiable on J, then
(a) φ′′
(t) ≥ 0 and φ′(t) ≥ 0 for any t ∈ J whenπ/4 ≤ θ ≤ pi/2 =⇒ φtrV is convex in S .
(b) φ′′
(t) ≥ 0 and φ′(t) ≤ 0 for any t ∈ J when0 ≤ θ ≤ pi/4 =⇒ φtrV is convex in S .
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space
MotivationPrevious ResultsRecent Results
Thank You!
Yu-Lin Chang Convex functions and convex cones on the Euclidean Space