0% found this document useful (0 votes)
12 views10 pages

Homework4 2

The document discusses the properties of a class of functions S, including its ability to shatter one-point, two-point, and three-point sets, while noting that it cannot shatter any five-point set. It also explores the VC index of S, concluding that it can shatter any non-increasingly placed set of points and has an infinite VC index. Additionally, the document outlines the operations needed for various functions and provides upper bounds for certain function classes.

Uploaded by

zwu363
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
12 views10 pages

Homework4 2

The document discusses the properties of a class of functions S, including its ability to shatter one-point, two-point, and three-point sets, while noting that it cannot shatter any five-point set. It also explores the VC index of S, concluding that it can shatter any non-increasingly placed set of points and has an infinite VC index. Additionally, the document outlines the operations needed for various functions and provides upper bounds for certain function classes.

Uploaded by

zwu363
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 10

Define class S : 3 (a ,, b .

] x (ac , bat :
(9 ., ac) , (bridal @R *

S can shatter one-point two-point .


any
or set

also shatter the


& can
following 3-point set ,
where each point is denoted as

(xc ya) and


(Xi
Y:) i 1/ 2 , 3 X ,< X2
Y
<
· ,
/ ,
=
X5 , ,
<
YS
<
Y

(x 1, y . )0 · (Xs ,
ys)
s can also shatter the
following 4-point set, where each point is denoted as

(x2 ,
·

ya) and
(Xi
Y:) i 1/ 2 , 3 , 4 X ,< X4 *2 <3 ,4 YS <
% Ye
= >
& / , ,

(x 16
Y)
·
(Xs .
Ys)

(4)

S cannot shatter
any 5-point set since any arbitrary rectangle cannot

cover all
of the two points combination without
covering the other 3 points

Thus UC (5) = 4 and the UC index is 5 .


Consider the function f(x) =
g(x-0) · Denote S as the collection of all such functions
Since under
gls is monotone , f() is also monotone
shifting

&
g(x)
f(x)

i
·
*

we can see thatS can shatter any set of points that are
non-increasingly placed.
Thus , VC index S is infinite
of
consider the class S : &XIR" :
11x-all : b3 .

Then it's thatS shatter two-point


easy to see can
any one-point , set.

also
S can shatter the
following 3-point set.

In fact, I can shatter


any subset of 3 non-collinear points .
Since every s non-collinear points can be shattered by half spaces ,
and a half
space can be
approximately be a
sphere by making the radius
sufficiently large .

Next , we want to show no subsets of 4 points can be shattered .

Consider the convex hull of 3 points 1,


X X, XS which is a
triangle
The
remaining one point y
is either inside or outside the triangle.
If it's inside ,
thenI cannot shatter.

If it's outside we can draw line between and X X2 , XS.


, a
y ,

There is no circle that can cover the two


points with the
longest connecting line

without the points. For example There


covering other two
,
is no circle to cover

· Y
only (c , y) in a) and (X ,, X3) in b) . X
X-Y
X ·
XS
Xs
as b)

·

Since BEST , diE&, it b -
s , there are
p = (s + 1) + ed total number of parameters .

Next , count the number of operations.


ha ; (x) =
max [0 , <;ix] needs & multiplications ,
dr additions ,
I
comparison

BiTh] needs sel multiplications s additions a

And the indicator function needs one


comparison .
Thus ,
in total , the number of operations is

E= (d + d- + 1)s + (S + 1) + s + 1 = 2ds + 2(S + 1)

By theorem 4 3 . . 2
, VC(f) = 4p(t+1)

4p(t+) = 4 (St + sd)(2ds + 25 + 1 + 2) : Cdst

Regrading [ (x y) #-
,
1 [y + f(x)3 :
fe 53 , it
requires one additional comparison
operation .

Thus UC of this class is also C52d


Rn(f) = [23f(Xi) + Y: 3

R(f) =
P(f(x) + Y) =
E2(f(x) + Y] .

Rn(f) =
infree Ralf) + An up . in

near minimizer

W .
P.

We aim to find an upper bound on


RLF) .
R(F) =
M(F) -

Ra(f) +
Rn(f)
=
MIF)-Rn(f) +
inffej Ralf) + On

>
- RIFS-Rn(F) + Ra > fol + On where fot j
= RIFS-Rn(F) infftf R(f)-infftFR(f)
+
Ra(fol + On +

=
[R(F)-Rn(F(3 [Ralfol -inftF RIfIS On infff R(f)
+ + +

[2sup/R(q)-Rn(f))
JEj
An infff R(f) + +

By Thm 42 2
.
. , since J is a collection of [0 , 1 -

valued functions ,

P) I/IRn-RIIF-EllRa-RIIg 1st) =
dexpl-Ined(
Take to blogn , the

PL/Ra-RIIj -
EllRu-RI/q <

By Samer's Lemma , Since VC(F) = C52d[max[12]5tdn

EllRa-RIE Galogd + 2 Ve
log(en /VC

- Edelogd + 2
(5dlog(n/ Cs2dd) (2)

have
plug (2) into 12 , we

PL/IEn-Bly-dalogd + 2 (5 d loghn/sidd) d) a
probability 1- in
With

BLF) Adaloga +
2Csdloghm/sidis id + An +
infff M(f)
= 10 sdflogi +
inffte Rif)
+ En

Here K = 16 Jo
By Thm 5 3 . . 11 , we know that Nis (3 , F , <(p)) : NLZ , F ,
1 .
110) .

Next , I want to construct functions belong to F .


Split the x-axis and y-axis into small intervals of length .
E

Then we obtain break points (Xi %: ) Where Xi 20 2 22 , k2] With KE = 1.


y: <
...
, , , ,

2
. Next , we construct all
possible piecewise linear functions.

Start at the breakpoints on the


y-axis .
Move breakpoint
on a
straight path to the next at X:= &
by either
going up or down.

Report this until reaches X := KE .

constructed
↑U functions this
way follow the Lipschitz condition,

1 f(z) ·

flos/E

Additionally ,
any constructed function is within 2K of one
of those functions .

Thus ,
we know about the cardinality of the set , INLEK J , ,
11 110) 1 -2
.
2

Since at each break point , there are two directions to choose and each line needs to make

& decisions Also each line has from


. ,
starting points to choose

Therefore , we conclude that

log(Nez (2)) =
log 1 2 12) = -

Loga + logt = Ega

I & ⑧

i
-------
"
-
O
a I E=
Let Eg 3 : be the E-cover of the set 9 where go :
fil-fiz , fil-fin ej
For any ge G , we have

19 :
-glyp
=
lific -

fic-fo + fo llp

all fis -foll +


1182-firlr by Minkoski's
inequality
choose the radius
coverings of F with &K , then

11
f. llp
fil -

11f2-fir llp-
Then 119 : -

g 1125 =E

Thus , to cover 6 ,
we need to consider all possible pairs (fir fiz)
·

The number of possible pairs is

N(2 , g) = N(Sk , 5) N(EK F( ,

logN(2 , 3) =
2 log N(E/2 , 5)
Consider a square grid .

Split the intervals with


equal probability mass under Q
.

consider & distribution and this ,


For simplicity , as the uniform on [0 , 1) , set the

breaking points as multiples of .


E

For simplicity is
, assume an
integer . To MJ,

Define F as the collection of monotonically increasing functions mapping from [0 , 1 to

Define mapping from


[0
F to be the collection of functions 13 to [0 , MJ S . t
.
, .

1) montonically increasing ,
pointwise constant,
right-continuous functions.

2) only change values at multiples of a

3) only take values that are


multiples of M2 .

Show ES :1 , F , is a 2ME cover of F

we want to show If EF ,
. s .
t .
Ilf-fillies = E
.

jt 0 o koh] , let f (ja)


, =
min[U(Kj-1) E, M3

where
kj is the smallest natural number S.t .
f(jg) >
min[MKj2 ,
M3 .

It's easy to see that If .


poinecisely .

3"Ilf -

fill +(a)
=
gizS gminLjls If(x) -
f , (x))dRx)

[2- [S SminElj9)f-minM(j-)2 , M3dQ(x) t is monotonically increasing


a

· [ (Mn9lljtE3) -

Min[M(kj-1)E , M3 fl is monotonically increasing

= f(k -
Min9M(ko-18sM3 +36 f(j) -

minSM(k -
1)3 ,
M]

jo term

CM + ME
That this proves that llf-fillies Ma + MEL 2ME .

Next , we need to find the


Cardinality in F .

log N(dME , 5 , L'(9)) =


Log (
= Ca
For
sufficiently small 130 ,
LogN(1 , F , (19)) : IMC5

You might also like