Upload
others
View
10
Download
0
Embed Size (px)
Citation preview
IE 521 Convex Optimization
Lecture 2: Convex Geometry
Niao He
4th February 2019
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Outline
Warm-upQuick ReviewQuestions
Convex GeometryRadon’s TheoremHelley’s TheoremSeparation Theorem
1 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Quick Review
I Convex setI X is convex iff λx + (1− λ)y ∈ X ,∀x , y ∈ X , λ ∈ [0, 1]
I Convex hullI Conv(X ) ={∑k
i=1 λixi : k ∈ N, λi ≥ 0,∑k
i=1 λi = 1, xi ∈ X ,∀i}
I Convexity-preserving operatorsI Taking intersection, Cartesian product, summationI Taking affine mapping, inverse affine mapping
I Topological propertiesI For convex sets, rint(X ) is dense in cl(X )
I Representation theoremI Any point in the convex hull of set X with dimension d
can be written as the convex combination of at mostd + 1 points in X .
2 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Question 1
Can you find a partition of the sets whose convex hullsintersect?
Figure: Four sets
3 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Question 2
If I1, I2 and I3 are intervals on the real line such that any twohave a point in common, do all three have a point incommon?
4 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Question 3
Which group is different from others?
Figure: Four groups of disjoint sets
5 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
The Mathematicians
Figure: JohannRadon (1887–1956)
Figure: EduardHelley (1884–1943)
Figure: HermannMinkowski(1864–1909)
6 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Radon’s Theorem (J. Radon, 1921)
Theorem. Let S be a collection of N points in Rd withN ≥ d + 2. Then we can write S = S1 ∪ S2 s.t.
S1 ∩ S2 = ∅, and Conv(S1) ∩ Conv(S2) 6= ∅.
Remark.I Any set of d + 2 points in Rd can be partitioned into
two disjoint sets whose convex hulls intersect.I Can be used to show the VC-dimension of the class of
halfspaces (linear separators) in d-dimensions is d + 1.
Figure: 3 points separable vs 4 points nonseparable
7 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Proof of Radon’s Theorem
I Let S = {x1, ..., xN} with N ≥ d + 2.
I Consider the linear system{∑Ni=1 γixi = 0∑Ni=1 γi = 0
⇒ (d + 1) equationsN ≥ (d + 2) unknowns
So there exists a non-zero solution γ1, ..., γN .
I Let I = {i : γi ≥ 0}, J = {j : γj < 0} anda =
∑i∈I γi = −
∑j∈J γj , then∑
i∈Iγixi =
∑j∈J
(−γj)xj ⇒∑i∈I
γiaxi =
∑j∈J
−γja
xj
I The partition S1 = {xi , i ∈ I} and S2 = {xj : j ∈ J}gives the desired result.
8 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Back to the Question
If I1, I2 and I3 are intervals on the real line such that any twohave a point in common, do all three have a point incommon?
9 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Helley’s Theorem (E. Helly, 1923)
Theorem. Let S1, ...,SN be a collection of convex sets in Rd
with N > d . Assume every (d + 1) sets of them have apoint in common, then all the sets have a point in common.
Figure: Four convex sets in R2
Q. Does the theorem still hold if we relax N =∞?Q. Does the theorem still hold if we relax (d + 1) sets to dsets?
10 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Helley’s Theorem (E. Helly, 1923)
Theorem. Let S1, ...,SN be a collection of convex sets in Rd
with N > d . Assume every (d + 1) sets of them have apoint in common, then all the sets have a point in common.
Remark.I Not true for infinite collection:
I E.g. Si = [i ,∞),∩+∞i=1 Si = ∅
I Not true if reduce (d + 1) sets to d sets.
Corollary. Let {Sα} be any collection of compact convex setsin Rd . If every (d + 1) sets have a point in common, then allsets have a points in common.
11 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Proof of Helley’s Theorem
Figure: Illustration of N = 4, d = 2
12 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Proof of Helley’s Theorem
By induction on N.
I Base case: N = d + 1, obviously true.
I Induction step: Assume the collection of N(≥ d + 1)sets have common point if every (d + 1) of them havecommon point. Show this holds for N + 1 sets.
I From the assumption, ∃ {x1, x2, ..., xN+1} such thatxi ∈ S1 ∩ ... ∩ Si−1 ∩ Si+1 ∩ ... ∩ SN+1 6= ∅.
I By Radon’s theorem, we can split it into two disjointsets, {x1, . . . , xk} and {xk+1, . . . , xN}, and
Conv({x1, ..., xk}) ∩ Conv({xk+1, ..., xN+1}) 6= ∅.
I Let z ∈ Conv({x1, ..., xk}) ∩ Conv({xk+1, ..., xN+1}). Itcan be shown that z ∈ S1 ∩ ... ∩ SN+1 (why?).
13 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Application of Helley’s Theorem
Baby Theorem Let X contain a finite set of points in theplane, such that every three of them are contained in a diskof radius 1. Then X is contained in a disk of radius 1.
Jung’s Theorem. Let X contain a finite set of points in theplane, such that any two of them has distance no greaterthan 1. Then X is contained in a disk of radius 1/
√3.
Jung’s Theorem. Let X ⊂ Rn be a compact set such thatany two of them has Euclidean distance no greater than 1.
Then X is contained in a ball with radius√
n2(n+1) .
14 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Application of Helley’s Theorem
Question. Consider the optimization problem
p∗ = minx∈R10
g0(x), s.t. gi (x) ≤ 0, i = 1, ..., 521.
I Suppose ∀t ∈ R, X0 ={x ∈ R10 : g0(x) ≤ t
}is convex,
Xi ={x ∈ R10 : gi (x) ≤ 0
}is convex.
I How many constraints can you drop without affectingthe optimal value?
15 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Other Applications of Helley’s Theorem
Helleys theorem is a very fundamental result in convexgeometry and can be applied to show many results.
I The centerpoint theorem
I Farkas Lemma
I Sion-Kakutani Theorem
I Chebyshev approximation
16 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Back to the Question
When can we separate two sets?
Figure: Four groups of disjoint sets
17 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Separation of Sets
Definition. Let S and T be two nonempty convex sets in Rn.A hyperplane H =
{x ∈ Rn : aT x = b
}with a 6= 0 is said to
separate S and T if S ∪ T 6⊂ H and
S ⊂ H− ={x ∈ Rn : aT x ≤ b
}T ⊂ H+ =
{x ∈ Rn : aT x ≥ b
}
Figure: Separation of two sets
18 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Strict Separation of Sets
Definition. Let S and T be two nonempty convex sets in Rn.A hyperplane H =
{x ∈ Rn : aT x = b
}with a 6= 0 is said to
strictly separate S and T if
S ⊂ H−− ={x ∈ Rn : aT x < b
}T ⊂ H++ =
{x ∈ Rn : aT x > b
}
Figure: Strict Separation of two sets
19 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Strong Separation of Sets
Definition. Let S and T be two nonempty convex sets in Rn.A hyperplane H =
{x ∈ Rn : aT x = b
}with a 6= 0 is said to
strongly separate S and T if there exits b′ < b < b′′ suchthat
S ⊂{x ∈ Rn : aT x ≤ b′
}T ⊂
{x ∈ Rn : aT x ≤ b′′
}Remark.
I Strict separation does not necessarily imply strongseparation.
I Strong separation is equivalent to say
supx∈S
aT x < infx∈T
aT x .
20 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Separation Hyperplane Theorem
Theorem. Let S and T be two nonempty convex sets. ThenS and T can be separated if and only if
rint(S) ∩ rint(T ) = ∅.
21 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Supporting Hyperplane Theorem
Theorem. Let S be a nonempty convex set and x0 ∈ ∂S .Then there exists a hyperplane H =
{x : aT x = aT x0
}with
a 6= 0 such that
S ⊂{x : aT x ≤ aT x0
}, and x0 ∈ H.
Figure: Supporting hyperplane
I This follows directly from the previous theorem.
I Such a hyperplane is called a supporting hyperplane.
22 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Strict Separation Hyperplane Theorem I
Theorem. Let S be closed and convex and x0 6∈ S , Thenthere exists a hyperplane that strictly separates x0 and S .
Figure: Strict separation
I Closedness of the set is crucial here.
I Separating hyperplane can be constructed based on theprojection.
23 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
Strict Separation Hyperplane Theorem II
Theorem. Let S and T be two nonempty convex sets andS ∩ T = ∅. If S − T is closed, then S and T can be strictlyseparated.
Remark.
I Even if both S and T are closed convex, S − T mightnot be closed, and they might not be strictly separated.
I When both S and T are closed convex, S ∩ T = ∅ andat least one of them is bounded, then S − T is closed,and S and T can be strictly separated
24 / 25
IE 521 ConvexOptimization
Niao He
Warm-up
Quick Review
Questions
Convex Geometry
Radon’s Theorem
Helley’s Theorem
Separation Theorem
References
I Boyd & Vandenberghe, Chapter 2.5
I Ben-Tal & Nemirovski, Chapter 1.2.2-1.2.6
25 / 25