Oxfam lunch time talk 25.07.2014 Inequality Trends in Russia
Centerpoints Theory Lunch Talk
-
Upload
don-sheehy -
Category
Technology
-
view
217 -
download
1
description
Transcript of Centerpoints Theory Lunch Talk
Searching for the center.
Don SheehyCMU Theory Lunch
October 8, 2008
1
2
It’s a fine line between stupid and
clever.
3
The Divide and Conquer Game
4
The Divide and Conquer Game
How to win
5
The Divide and Conquer Game
How to win
Pick a center point.
6
The Divide and Conquer Game
How to win
Pick a center point.
Given a set S ! Rd, a center point p is a pointsuch that every closed halfspace with p on itsboundary contains at least n
d+1points of S.
7
Some definitions you probably already know.
8
!
pi!P
cipiLinear:
9
!
pi!P
cipiLinear:
Nonnegative: ci ! 0
10
!
pi!P
cipiLinear:
Nonnegative:
Affine:
ci ! 0
!ci = 1
11
!
pi!P
cipiLinear:
Nonnegative:
Affine:
Convex: Affine and Nonnegative
ci ! 0
!ci = 1
12
Radon => Helly => Center Points Exist.
13
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
14
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
15
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
16
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
17
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
18
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
19
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
20
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
21
Radon’s TheoremIf P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)
such that conv(U) " conv(U) is nonempty.
22
Radon’s Theorem
d+2!
i=1
cipi = 0
d+2!
i=1
ci = 0
If P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)such that conv(U) " conv(U) is nonempty.
23
Radon’s Theorem
d+2!
i=1
cipi = 0
d+2!
i=1
ci = 0
I+
= {i : ci > 0}
I!
= {i : ci < 0}
!
i!I+
cipi =!
i!I!
(!ci)pi
If P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)such that conv(U) " conv(U) is nonempty.
24
Radon’s Theorem
d+2!
i=1
cipi = 0
d+2!
i=1
ci = 0
I+
= {i : ci > 0}
I!
= {i : ci < 0}
!
i!I+
cipi =!
i!I!
(!ci)pi
!
i!I+
ci =!
i!I!
(!ci)
If P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)such that conv(U) " conv(U) is nonempty.
25
Radon’s Theorem
d+2!
i=1
cipi = 0
d+2!
i=1
ci = 0
I+
= {i : ci > 0}
I!
= {i : ci < 0}
!
i!I+
cipi =!
i!I!
(!ci)pi
!
i!I+
ci =!
i!I!
(!ci)
If P ! Rd has d+2 (or more) points then there is a partition of P into (U,U)such that conv(U) " conv(U) is nonempty.
x =
!
i!I+
"
ci#
j!I+ cj
$
pi =
!
i!I!
"
!ci#
j!I!!cj
$
pi
26
Helly’s TheoremGiven some convex sets in Rd such that every d + 1 sets
have common intersection, then the whole collection of setshas a common intersection.
27
Helly’s TheoremGiven some convex sets in Rd such that every d + 1 sets
have common intersection, then the whole collection of setshas a common intersection.
28
Helly’s TheoremGiven some convex sets in Rd such that every d + 1 sets
have common intersection, then the whole collection of setshas a common intersection.
29
Proof Hint:Use Radon’s Theorem!
30
31
32
33
Fun Exercise:Show that the Radon Point is in every set.
34
More than d+2 sets?
35
The Center Point Theorem
Consider the set of all minimal halfspaces containingat least dn
d+1+ 1 points.
Helly’s Theorem implies that all the halfspaces havea common intersection. The intersection is the setof center points.
Observe that every d + 1 have a common intersection.
36
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
37
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
38
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
39
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
40
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
41
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
42
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
43
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
44
Tverberg’s Theorem
Let S be a set of at least (d + 1)(r ! 1) + 1 points in Rd.There exists a partition of S into r subsets X1, . . . , Xr
such that!r
i=1conv(Xi) "= #
Choose r = n/(d+1)It’s a center point!
45
Proof via Helly’s Theorem
Proof via Tverberg’sTheorem
46
Proof via Helly’s Theorem
Proof via Tverberg’sTheorem
coNP NP
47
An Algorithm
48
Approximating Center Points with Iterated Radon Points[ Clarkson, Eppstein, Miller, Sturtivant, Teng, 1993 ]
49
Approximating Center Points with Iterated Radon Points[ Clarkson, Eppstein, Miller, Sturtivant, Teng, 1993 ]
1. Randomly sample points into sets of d+2.2. Compute the Radon point for each set. 3. Compute the Radon points of the Radon points 4. Continue until only one point remains.5. Return that point.
50
Approximating Center Points with Iterated Radon Points[ Clarkson, Eppstein, Miller, Sturtivant, Teng, 1993 ]
1. Randomly sample points into sets of d+2.2. Compute the Radon point for each set. 3. Compute the Radon points of the Radon points 4. Continue until only one point remains.5. Return that point.
O!
n
d2
"
-center with high probability.
51
52
53
54
Analysis looks like Helly-type proof.
Look at all projections to one dimensionat the same time.
55
Let’s build an algorithm so that the analysis will look less like Helly and more like Tverberg.
56
57
58
59
60
61
62
63
64
This almost works.
65
66
67
68
69
70
71
72
73
74
75
76
77
Analysis: How good is the resulting center point?
78
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
79
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Suppose for contradiction that gn < n
(d+1)2 .
80
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Then, gn/2 < n2(d+1)2 and the corresponding partition
uses less than n2(d+1) points.
Suppose for contradiction that gn < n
(d+1)2 .
81
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Then, gn/2 < n2(d+1)2 and the corresponding partition
uses less than n2(d+1) points.
So, with n points, we can construct d + 2 pointswith partitions of size gn/2.
Suppose for contradiction that gn < n
(d+1)2 .
82
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Then, gn/2 < n2(d+1)2 and the corresponding partition
uses less than n2(d+1) points.
So, with n points, we can construct d + 2 pointswith partitions of size gn/2.
This means we can iterate the algorithm, andgn ! 2gn/2
Suppose for contradiction that gn < n
(d+1)2 .
83
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Then, gn/2 < n2(d+1)2 and the corresponding partition
uses less than n2(d+1) points.
So, with n points, we can construct d + 2 pointswith partitions of size gn/2.
This means we can iterate the algorithm, andgn ! 2gn/2
Base case: gd+2 = 2.
Suppose for contradiction that gn < n
(d+1)2 .
84
Analysis: How good is the resulting center point?
Say gn is the minimum guaranteed partition size on n points
Then, gn/2 < n2(d+1)2 and the corresponding partition
uses less than n2(d+1) points.
So, with n points, we can construct d + 2 pointswith partitions of size gn/2.
This means we can iterate the algorithm, andgn ! 2gn/2
Base case: gd+2 = 2. =! gn " 2log n
d+2 =n
d + 2
Suppose for contradiction that gn < n
(d+1)2 .
85
Thank you.
86