Sauer's lemma

cosmos 15th May 2019 at 7:29pm
VC dimension

Bounding the size of families of sets of nn distinct elemtns with bounded VC dimension. In particular, it can be used to bound the Growth function of sets with bounded VC dimension.

See here

Bounding the size of families of sets with bounded VC dimension.

ΠC(m)=max{Π(s)S=m}\Pi_C(m)=\max \{|\Pi(s)| \mid |S|=m\}

Clearly if mdm \leq d (d=VC dim of C), then Πc(m)=2m\Pi_c(m) = 2^m

For mdm\geq d, ΠC(m)=O(md)\Pi_C(m) = O(m^d)

Def: Φd(m)=Φd(m1)+Φd1(m1)\Phi_d(m) = \Phi_d(m-1)+\Phi_{d-1}(m-1) if m1m \geq 1 and d1d \geq 1

Φ0(m)=1,m\Phi_0(m)=1, \forall m, Φd(0)=1,d\Phi_d(0)=1, \forall d

Lemma: ΠC(m)Φd(m)\Pi_C(m) \leq \Phi_d(m) if VC dim(C) = d.

Φd(m)=i=0d(mi)\Phi_d(m) = \sum\limits_{i=0}^d \binom{m}{i}

m=0m=0, then d,Φd(0)=1\forall d, \Phi_d(0)=1, i=0d(0i)=1 \sum\limits_{i=0}^d \binom{0}{i} =1

d=0d=0, then mΦ0(m)=1\forall m \Phi_0(m)=1, i=00(mi)=1 \sum\limits_{i=0}^0 \binom{m}{i} =1

Φd(m)=Φd(m1)+Φd1(m1)=i=0d(m1i)+i=0d1(m1i)\Phi_d(m) = \Phi_d(m-1)+\Phi_{d-1}(m-1) = \sum\limits_{i=0}^d \binom{m-1}{i} + \sum\limits_{i=0}^{d-1} \binom{m-1}{i}

=(m10)+i=1d((m1i)+(m1i1)) = \binom{m-1}{0}+ \sum\limits_{i=1}^d (\binom{m-1}{i} + \binom{m-1}{i-1})

=(m0)+i=1d((mi))=i=0d(mi) = \binom{m}{0}+ \sum\limits_{i=1}^d (\binom{m}{i} ) = \sum\limits_{i=0}^d \binom{m}{i}

Φd(m)=(md)di=0d(mi)(dm)d\Phi_d(m) = (\frac{m}{d})^d\sum\limits_{i=0}^d \binom{m}{i} (\frac{d}{m})^d (md)di=0d(mi)(dm)i\leq (\frac{m}{d})^d\sum\limits_{i=0}^d \binom{m}{i} (\frac{d}{m})^i (md)di=0m(mi)(dm)i\leq (\frac{m}{d})^d\sum\limits_{i=0}^m \binom{m}{i} (\frac{d}{m})^i =(md)d(1+dm)m(md)ded=(med)d=(\frac{m}{d})^d(1+\frac{d}{m})^m \leq (\frac{m}{d})^d e^d = (\frac{me}{d})^d

Proof of lemma: ΠC(m)Φd(m)\Pi_C(m) \leq \Phi_d(m) if VC dim(C) = d.

Φd(m)=Φd(m1)+Φd1(m1)\Phi_d(m) = \Phi_d(m-1)+\Phi_{d-1}(m-1)

Let S be some set of size m.

Πc(m)=1\Pi_c(m)=1, if VC-dim(C)=0 (can't satisfy all asignemtns of one point -> must have only one hypothesis).

Πc(0)=1\Pi_c(0)=1. For no points.

xSx \in S, consider S{x}S \setminus \{x\}

ΠC(S{x})Πc(m1)Φd(m1)|\Pi_C(S \setminus \{x\})| \leq \Pi_c (m-1) \leq \Phi_d(m-1)

Πc(s)Πc(Sz)|\Pi_c(s)|-|\Pi_c(S\setminus z)|: How many dichotomies over S{x}S \setminus \{x\} using C that can be extended to S in 2 ways

C={cΠC(s)c(x)=0 and c~Π(S) s.t. c~(x)=1 and for all other points z not equal to x, c(z)=tildec(z)}C'=\{c \in \Pi_C(s) \mid c(x)=0 \text{ and } \exists \tilde{c} \in \Pi(S) \text{ s.t. } \tilde{c}(x)=1 \text{ and for all other points z not equal to x, c(z)=tildec(z)}\}

C' counts the number of concepts in C restricted to S{x}S \setminus \{x\}, which can be extended to SS in two possible ways (with x either 0 or 1).

... etc see photo. We need to show that c' has VC-dimension dimension at most d-1 (proved here)

Let T be a set TST \subset S

This shows that if the VC dimension of a concept class is finite, then the growth on the size of the concept class is polynomial with the dimension of the input space (number of points, m). This restriction is what allows generalization.