Master’s Thesis Seminar
VC dimension of bisectors between curves
Carolin Kaffine 3rd April 2020
Overview
• Recap of basic definitions
• Main goals
• Upper bounds:
• Approach 1: via composition lemma for halfspaces
• Approach 2: via VC dimension of function spaces
• Lower bounds
• Summary of results
• Outlook
Basic definitions
Definition:
• curve in Rd: continuous function V: [0,1]→Rd
• polygonal curve: piecewise linear
• Xdk: space of all piecewise linear curves in Rd with k vertices
Basic definitions
Hausdorff and discrete Hausdorff distance:
dH(V,W) := max (
sup
p∈V
q∈Winf d(p,q), sup
q∈W
p∈Vinf d(p,q) )
ddH(V,W) := max
maxv∈V min
w∈Wd(v,w),max
w∈Wmin
v∈Vd(v,w)
Basic definitions
Fréchet distance:
dF(V,W) = inf
f,g max
α∈[0,1]
V(f(α))−W(g(α)) ,
Basic definitions
• Range space (X,R): ground setX, ranges R∈ R ⊆2X
• given Y ⊆X, it is shattered byR if {R∩Y |R∈ R}=2Y
• VC dimension: greatest cardinality of shattered subset
Examples for VC dimension
Ground setX =R2, ranges are disks:
⇒ VCdim≥3 ⇒ VCdim<4
In general:balls and halfspaces inRd have VCdim=d+1.
Basic definitions
Shatter function(or growth function) for a range space (X,R):
π(X,R)(m) = max
Y⊆X,|Y|=m
{R∩Y|R∈ R}
Shatter function lemma (or Sauer’s lemma)
For a range space(X,R) with VC dimension at most δ, we have π(X,R)(m)≤Φδ(m) :=
m 0
+ m
1
+· · ·+ m
δ
⇒polynomial growth inm sinceΦδ(m)≤ emδ δ
∈ O(mδ)
Bisectors between curves
Bisector range space:(Xdm,Bd,k)
• ground setXdm =set of all curves inRd withm vertices
• range setBd,k with ranges
R(V,W)={S ∈Xdm |d(V,S)≤d(W,S)}
for (V,W)∈Xdk ×Xdk.
Bisectors between curves
Bisectors inR2 form=1 (i.e. all distance functions are the same):
X21,Bd,1
X21,Bd,2
Bisectors between curves
Form>1: no graphic representation of ranges anymore, because dimension gets higher than 3
Goal of Master’s thesis
Main goal:
Find upper and lower bounds on VC dimension of the bisector range space, dependent on
• m (complexity of shattered curves)
• k (complexity of curves that define bisectors)
• d (dimension)
• ddH, dH, ddF, or dF (used distance function)
Main paper:
“The VC Dimension of Metric Balls under Fréchet and Hausdorff Distances” by A. Driemel, A. Nusser, J.M. Phillips, and I. Psarros
Upper bound 1: via composition lemma
Composition lemma (simplified):
For a range space(X,R) with VCdim =δ, the range space of all unions/intersections ofn ranges inR has VC dimension
O(nδlogn).
Idea(for m=1):
• Write bisector ranges as unions and intersections of halfspaces
• By the composition lemma, we can bound VC dimension of bisector range space by considering VC dimension of halfspaces (which isd +1)
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace
Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace
Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve
Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve
Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve
Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve
Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
Step 1:For two points: bisector rangeR(v1,w1) is halfspace Step 2:Ranges get intersected when adding a point to one curve Step 3:Final range of two curves is union of intersections
Upper bound 1: via composition lemma
For generald:
• range R(V,W)=S
w∈W
T
v∈V h(v,w), where h(v,w)is halfspace of points that are closer tov than tow
• ifV,W have lengthk, we took(k−1)2 ∈ O(k2) unions and intersections
⇒ VC dimension is in
O (k−1)2(d +1) log((k−1)2)
=O(k2dlogk)
Upper bound 2: via Thm on VC dimension of function spaces
Theorem:
Leth:Ra×Rb→ {0,1}and
H ={x 7→h(α,x)|α∈Ra}.
Supposeh can be computed by an algorithm that takes
(α,x)∈Ra×Rb as input and returnsh(α,x)after no more than t simple operations.
Then, the VC dimension ofH ist≤4a(t+2).
Upper bound 2: via Thm on VC dimension of function spaces
• Write bisector range R(V,W) as functionh (V,W),· that takes a curve S and outputs 1 ifS is closer to V than to W, and 0 else
⇒ Bisector range space can be written as(Xdm,H) for H=
S 7→h (V,W),S S ∈Xdm,(V,W)∈Xdk ×Xdk
• so h: (Xdk ×Xdk
| {z }
∼=R2dk
)×Xdm→ {0,1}, i.e. a=2dk
• it remains to compute t, i.e. check how fasth can be computed
Upper bound 2: via Thm on VC dimension of function spaces Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s)
Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s)
Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s)
Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
Examplefor discrete Hausdorff distance:
Step 1:Calculate d(v,s)2
for allv ∈V,s ∈S
Step 2:Find ddH(V,S) = max maxs∈Sminv∈Vd(v,s),maxv∈Vmins∈Sd(v,s) Step 3:Do same forW and take minimum of ddH(V,S)and ddH(W,S)
Upper bound 2: via Thm on VC dimension of function spaces
• In total: calculation of 2mk squared euclidean distances between vertices, each in O(d)
• O(mk) comparisons to find ddH(V,S)
• All in all: t ∈ O(mkd) simple operations
⇒ VCdim≤4·2dk(c·mkd −1)∈ O(mk2d2)
Lower bounds
Idea:Find lower bounds for k =1 and/or m=1
⇒valid lower bound for all distance functions and all k andm Easy lower bound(for m=k =1:)
Bisector ranges look like halfspaces, so VCdim≥d +1
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈Ω(max(k,d))
Lower bounds
Lower bound(for m=1):
• VC dimension of (open) k-gons is 2k+1
• Bisector ranges can look like open k-gons, so their VC dimension is≥2k+1
⇒Combining the two lower bounds we get VCdim ∈
Summary of results so far:
Upper bounds:
Distance function m arbitrary m=1 discrete Hausdorff O(mk2d2)
O(dk2logk) Hausdorff
– discrete Fréchet
O(k2d2) Fréchet
Lower bound:
Ω(max(k,d))
Outlook
Further goals:
• establish upper bounds for other distance functions than the discrete Hausdorff distance that depend on m
• establish better lower bounds by using geometric properties of bisector range spaces
• reduce gap between upper and lower bounds