Growth function

The growth function (also called: shatter coefficient or shattering number) measures the richness of a set family. It is especially used in the context of statistical learning theory, where it measures the complexity of a hypothesis class. The term 'growth function' was coined by Vapnik and Chervonenkis in their 1968 paper, where they also proved many of its properties.

Definitions

Set-family definition

Let $H$ be a set family (a set of sets) and $C$ a set. Their intersection is defined as the following set-family:

$H\cap C:=\{h\cap C\mid h\in H\}$

The index of $H$ with respect to $C$ is the size of this intersection:

$\operatorname {Index} (H,C):=|H\cap C|$

Obviously, if a set $C_{m}$ has $m$ elements then the index is at most $2^{m}$.

The growth function measures the size of $H\cap C$ as a function of $|C|$. Formally:

$\operatorname {Growth} (H,m):=\max _{C:|C|=m}|H\cap C|$

Hypothesis-class definition

Equivalently, let $H$ be a hypothesis-class (a set of binary functions) and $C$ a set with $m$ elements. The restriction of $H$ to $C$ is the set of binary functions on $C$ that can be derived from $H$:

$H_{C}:=\{(h(x_{1}),\ldots ,h(x_{m}))\mid h\in H,x_{i}\in C\}$

The growth function measures the size of $H_{C}$ as a function of $|C|$:

$\operatorname {Growth} (H,m):=\max _{C:|C|=m}|H\cap C|(H,m):=\max _{C:|C|=m}|H_{C}|$

Examples

1. The domain is the real line $\mathbb {R}$. The set-family $H$ contains all the half-lines (rays) from a given number to positive infinity, i.e., all sets of the form $\{x>x_{0}\mid x\in \mathbb {R} \}$ for some $x_{0}\in \mathbb {R}$. For any set $C$ of $m$ real numbers, the intersection $H\cap C$ contains $m+1$ sets: the empty set, the set containing the largest element of $C$, the set containing the two largest elements of $C$, and so on. Therefore: $\operatorname {Growth} (H,m)=m+1$. The same is true whether $H$ contains open half-lines, closed half-lines, or both.

2. The domain is the segment $[0,1]$. The set-family $H$ contains all the open sets. For any set $C$ of $m$ real numbers, the intersection $H\cap C$ contains all possible subsets of $C$. There are $2^{m}$ such subsets, so $\operatorname {Growth} (H,m)=2^{m}$.

3. The domain is the Euclidean space $\mathbb {R} ^{n}$. The set-family $H$ contains all the half-spaces of the form: $x\cdot \phi \geq 1$, where $\phi$ is a fixed vector. Then $\operatorname {Growth} (H,m)=\operatorname {Comp} (n,m)$, where Comp is the number of number of components in a partitioning of an n-dimensional space by m hyperplanes.

4. The domain is the real line $\mathbb {R}$. The set-family $H$ contains all the real intervals, i.e., all sets of the form $\{x\in [x_{0},x_{1}]|x\in \mathbb {R} \}$ for some $x_{0},x_{1}\in \mathbb {R}$. For any set $C$ of $m$ real numbers, the intersection $H\cap C$ contains all runs of between 0 and $m$ consecutive elements of $C$. The number of such runs is ${m+1 \choose 2}+1$, so $\operatorname {Growth} (H,m)={m+1 \choose 2}+1$.

Properties

The growth function has two trivial bounds.

1. For any finite $H$:

$\operatorname {Growth} (H,m)\leq |H|$

since for every $C$, the number of elements in $H\cap C$ is at most $|H|$. Therefore, the growth function is mainly interesting when $H$ is infinite.

2. For any nonempty $H$:

$\operatorname {Growth} (H,m)\leq 2^{m}$

I.e, the growth function has an exponential upper-bound.

We say that a set-family $H$ shatters a set $C$ if their intersection contains all possible subsets of $C$, i.e. $H\cap C=2^{C}$. If $H$ shatters $C$ of size $m$, then $\operatorname {Growth} (H,C)=2^{m}$, which is the upper bound.

3. The following is a property of the Index function:

• If, for some set $C_{m}$ of size $m$, and for some number $n\leq m$, $|H\cap C_{m}|\geq \operatorname {Comp} (n,m)$ -
• then, there exists a subset $C_{n}\subseteq C_{m}$ of size $n$ such that $|C_{n}\cap H|$ = $2^{n}$.

This implies the following property of the Growth function. For every family $H$ there are two cases:

• The exponential case: $\operatorname {Growth} (H,m)=2^{m}$ identically.
• The polynomial case: $\operatorname {Growth} (H,m)$ is majorized by $\operatorname {Comp} (n,m)\leq m^{n}+1$, where $n$ is the smallest integer for which $\operatorname {Growth} (H,n)<2^{n}$.

The VC dimension of $H$ is defined according to these two cases:

• In the polynomial case, $\operatorname {VCDim} (H)=n-1$ = the largest integer $d$ for which $\operatorname {Growth} (H,d)=2^{d}$.
• In the exponential case $\operatorname {VCDim} (H)=\infty$.

So $VCDim(H)\geq d$ if-and-only-if $Growth(H,d)=2^{d}$.

T‎he growth function can be regarded as a refinement of the concept of VC dimension. The VC dimension only tells us whether $\operatorname {Growth} (H,d)$ is equal to or smaller than $2^{d}$, while the growth function tells us exactly how $Growth(H,m)$ changes as a function of $m$.

4. Another connection between the growth function and the VC dimension is given by the Sauer–Shelah lemma:

If $VCDim(H)=d$, then:

for all $m$: $\operatorname {Growth} (H,m)\leq \sum _{i=0}^{d}{m \choose i}$

In particular,

for all $m>d+1$: $\operatorname {Growth} (H,m)\leq (em/d)^{d}=O(m^{d})$

so when the VC dimension is finite, the growth function grows polynomially with $m$.

Applications in probability theory

Let $\Omega$ be a set on which a probability measure $Pr$ is defined. Let $H$ be family of subsets of $\Omega$ (= a family of events).

Suppose we choose a set $C_{m}$ that contains $m$ elements of $\Omega$, where each element is chosen at random according to the probability measure $P$, independently of the others (i.e, with replacements). For each event $h\in H$, we compare the following two quantities:

• Its relative frequency in $C_{m}$, i.e, $|h\cap C_{m}|/m$;
• Its probability $Pr[h]$.

We are interested in the difference, $\delta (h,C_{m}):={\big |}|h\cap C_{m}|/m-Pr[h]{\big |}$. This difference satisfies the following upper bound:

$Pr{\big [}\forall h\in H:\delta (h,C_{m})\leq \epsilon {\big ]}~~~~>~~~~1-4\cdot \operatorname {Growth} (H,2m)\cdot \exp(-\epsilon ^{2}\cdot m/8)$

In words: the probability that for all events in $H$, the relative-frequency is near the probability, is lower-bounded by an expression that depends on the growth-function of $H$.

A corollary of this is that, iff the growth function is polynomial in $m$ (i.e, there exists some $n$ such that $\operatorname {Growth} (H,m)\leq m^{n}+1$), then the above probability approaches 1 as $m\to \infty$. I.e, the family $H$ enjoys uniform convergence to probability.