Borel-Cantelli Lemma: Difference between revisions

From Optimal Transport Wiki
Jump to navigation Jump to search
(Add links to related articles)
 
(33 intermediate revisions by one other user not shown)
Line 1: Line 1:


Many interesting theorem in probability utilizes measure theory, as it often provides some form of one-line argument that allows the proof to go through. Among one of them is the law of large numbers, which informally states that the average of iid samples approaches the mean of the distribution as the sample size grows large.  
Modern probability is built on the foundation of measure theory. One of its most basic and key results, the strong law of large numbers, demonstrates the usefulness and power of this foundation. Informally, this law states that the sample mean gets close to the true mean as sample size grows large.  


We would like to be more precise about what the words "approaches" mean. Clearly we cannot replace the word "approaches" with "converges to", as one could be unlucky and consistently draw below or above the mean so that the sample average may never even be close to the true mean. Despite of this, one thing that we do know is that consistently drawing below or above the mean, while not impossible, becomes increasingly unlikely as the sample grows large. This naturally leads to the idea of convergence in probability.  
We would like to be more precise about the meaning of the phrase "get close to". Certainly we cannot replace it by "converges to", as one could be unlucky and consistently draw below or above the mean so that the sample mean is no where near the true mean. However, just by intuition we know that consistently drawing below or above the mean, while not impossible, becomes increasingly unlikely as the sample size grows. This naturally leads to the idea of convergence in probability.  


==Convergence in probability==
==Convergence in probability==
Line 10: Line 10:
<math display="block">\lim_{n\to \infty}\mathbb{P}(|X_n - X|>\epsilon) = 0.</math>
<math display="block">\lim_{n\to \infty}\mathbb{P}(|X_n - X|>\epsilon) = 0.</math>


Notice that this is simply convergence in measure in probability space. Equipped with this definition, we can now state the weak law of large numbers: the sample average ''converges in probability'' to the mean. Bernoulli first proved this theorem for Bernoulli random variables in 1713, when tools like Chebyshev's inequality wasn't discovered. Later on, mathmaticians proved more general cases that does not the require the assumption of finite variance or iid random variables.  
This is simply [[Convergence_in_Measure|convergence in measure]] for the probability space. Equipped with this definition, we can now state a version of the weak law of large numbers: the sample average ''converges in probability'' to the mean, if the iid random variables have finite variance. In history, Bernoulli first proved this theorem for Bernoulli random variables in 1713, when tools like Chebyshev's inequality wasn't even discovered. Later on, mathmaticians proved more general cases that does not the require the assumption of finite variance or strict independence of random variables.


Proof. Without loss of generality assume that the random variables are centered, and let <math> \overline X_n</math> denote the sample average. By Chebyshev's inequality,  
Proof. Without loss of generality assume that the random variables are centered, and let <math> \overline X_n</math> denote the sample average. By Chebyshev's inequality,  
<math>  P(|\overline X_n | > \epsilon) \leq \frac{\sigma^2}{n\epsilon} \to 0</math> as <math> n  </math> approaches infinity.
<math>  P(|\overline X_n | > \epsilon) \leq \frac{\sigma^2}{n\epsilon} \to 0</math> as <math> n  </math> approaches infinity. Note that for Chebyshev's inequality to hold we do need finite variance.  


It is natural to wonder the possibility of obtaining a stronger notion of convergence, perhaps under stronger condition than finite variance. To do so we need to formally define the stronger type of convergence that we need, which in study of probability we often refer to as "almost sure convergence".
It is natural to wonder the possibility of obtaining a stronger notion of convergence, perhaps under stronger assumptions. To do so we need to formally define the stronger type of convergence that we need, which in study of probability is often refered to as "almost sure convergence".


==Almost sure convergence==
==Almost sure convergence==
Line 22: Line 22:


<math display="block">\mathbb{P}(\lim_{n\to \infty}X_n = X ) = 1.</math>
<math display="block">\mathbb{P}(\lim_{n\to \infty}X_n = X ) = 1.</math>
Note that this is the almost everywhere convergence in probability space, which is a stronger version of convergence. This is due to the fact that in a finite measure space, convergence almost everywhere implies convergence in measure.  
Note that this is the almost everywhere convergence for probability space, which is a stronger notion of convergence than convergence in probability. This is due to the fact that in a finite measure space, convergence almost everywhere implies convergence in measure.  


The strong law of large number carries this stronger notion of converges and states: the sample average ''converges'' ''almost surely'' to the mean. Before the proof, we first introduce a lemma.
The strong law of large number carries this stronger notion of converges and states: under suitable assumptions, the sample average ''converges'' ''almost surely'' to the mean. Before the presentation of the proof, we first introduce a lemma.


==Borel Catelli Lemma==
==Borel Catelli Lemma==
Line 36: Line 36:
Fix <math display="inline">\epsilon > 0</math>, there exists <math display="inline">n</math> sufficiently large such that <math display="inline">\sum_{i=n}^\infty \mathbb{P}(E_i) < \epsilon</math>. By union bound we obtain <math display="inline">\mathbb{P}(\cup_{k\geq n} E_k) < \epsilon</math>. So <math display="inline">\lim_{n\to \infty} \mathbb{P}(\cup_{k\geq n} E_k) = 0</math>. By continuity of measure , we have <math display="inline">\mathbb{P}(\cap_{n=1}^\infty \cup_{k\geq n} E_k) = \lim_{n\to \infty} \mathbb{P}(\cup_{k\geq n} E_k) = 0</math>.  
Fix <math display="inline">\epsilon > 0</math>, there exists <math display="inline">n</math> sufficiently large such that <math display="inline">\sum_{i=n}^\infty \mathbb{P}(E_i) < \epsilon</math>. By union bound we obtain <math display="inline">\mathbb{P}(\cup_{k\geq n} E_k) < \epsilon</math>. So <math display="inline">\lim_{n\to \infty} \mathbb{P}(\cup_{k\geq n} E_k) = 0</math>. By continuity of measure , we have <math display="inline">\mathbb{P}(\cap_{n=1}^\infty \cup_{k\geq n} E_k) = \lim_{n\to \infty} \mathbb{P}(\cup_{k\geq n} E_k) = 0</math>.  


The other statement follows easily from the similar technique.  
The other statement follows easily from similar technique.


==Strong law of large numbers==
==Strong law of large numbers==
To be finished...
Instead of directly stating the unspecified "suitable assumptions", we work backward to see what assumptions is needed to achieve this "almost sure convergence" to the mean. Here, we are interested in the set where <math>\overline X_n </math> converges to <math> \mu </math>, and we can rewrite this set as <math> \{x:  \cap_{k}\cup_n \cap_{m\geq n} |\overline X_m| < \frac{1}{k}\} </math>. To relate it to Borel-Cantelli lemma, we instead study its complement
<math> \{x:  \cup_{k}\cap_n \cup_{m\geq n} |\overline X_m| \geq \frac{1}{k}\} </math>. By continuity of measure, we have <math> P(\{x:  \cup_{k}\cap_n \cup_{m\geq n} |\overline X_m| \geq \frac{1}{k}\}) =  \lim_{k \to \infty} P \{x: \cap_n \cup_{m\geq n} |\overline X_m| \geq \frac{1}{k}\} </math>.
 
It suffices to show that for any fixed <math display="inline">k</math>, <math display="inline">P(\cap_{n}\cup_{k\geq n} |\overline X_n|> \frac{1}{k}) = 0</math>. This closely resembles the conclusion of the Borel-Cantelli lemma, so we might be tempted to show that <math display="inline">\sum_{n=1}^\infty P(|\overline X_n| > \frac{1}{k}) < \infty</math>. Unfortunately, a simple application of Chebyshev's inequality shows that this fall short, as we get
 
<math display="block">\sum_{n=1}^\infty P(|\overline X_n|> \frac{1}{k}) \leq \sum_{n=1}^\infty \frac{k^2\sigma^2}{n} = \infty.</math>
Therefore, stronger assumptions must be made if we are to use the same proof technique. Since finiteness of higher moments implies that of the the lower ones, it is natural for us to assume finite fourth moment, which gives us
<math display="block">\begin{align}
 
\mathbb{P}(|\overline X_n | > \frac{1}{k}) &\leq k^4\mathbb{E}(|\overline X_n|^4) \\
 
&= k^4 \frac{n\mathbb{E}(X_i^4) + 6n(n-1)\mathbb{E}(X_i^2)}{n^4}\\
 
&= k^4 O(\frac{1}{n^2}).
 
\end{align}</math>
Then by Borel-Cantelli lemma, <math display="inline">\overline X_n</math> converges a.s to <math display="inline">\mu</math>.
<ref name="Klenke">, ''Probability Theory: A Comprehensive Course'', §1.0</ref>.
 
==reference==

Latest revision as of 03:54, 19 December 2020

Modern probability is built on the foundation of measure theory. One of its most basic and key results, the strong law of large numbers, demonstrates the usefulness and power of this foundation. Informally, this law states that the sample mean gets close to the true mean as sample size grows large.

We would like to be more precise about the meaning of the phrase "get close to". Certainly we cannot replace it by "converges to", as one could be unlucky and consistently draw below or above the mean so that the sample mean is no where near the true mean. However, just by intuition we know that consistently drawing below or above the mean, while not impossible, becomes increasingly unlikely as the sample size grows. This naturally leads to the idea of convergence in probability.

Convergence in probability

A sequence of random variables converges in probability to if for any ,

This is simply convergence in measure for the probability space. Equipped with this definition, we can now state a version of the weak law of large numbers: the sample average converges in probability to the mean, if the iid random variables have finite variance. In history, Bernoulli first proved this theorem for Bernoulli random variables in 1713, when tools like Chebyshev's inequality wasn't even discovered. Later on, mathmaticians proved more general cases that does not the require the assumption of finite variance or strict independence of random variables.

Proof. Without loss of generality assume that the random variables are centered, and let denote the sample average. By Chebyshev's inequality, as approaches infinity. Note that for Chebyshev's inequality to hold we do need finite variance.

It is natural to wonder the possibility of obtaining a stronger notion of convergence, perhaps under stronger assumptions. To do so we need to formally define the stronger type of convergence that we need, which in study of probability is often refered to as "almost sure convergence".

Almost sure convergence

A sequence of random variables converges almost surely to if

Note that this is the almost everywhere convergence for probability space, which is a stronger notion of convergence than convergence in probability. This is due to the fact that in a finite measure space, convergence almost everywhere implies convergence in measure.

The strong law of large number carries this stronger notion of converges and states: under suitable assumptions, the sample average converges almost surely to the mean. Before the presentation of the proof, we first introduce a lemma.

Borel Catelli Lemma

Let be a probability space, and any sequence of events in . If , then .

On the other hand, if , then .

Proof.

Fix , there exists sufficiently large such that . By union bound we obtain . So . By continuity of measure , we have .

The other statement follows easily from similar technique.

Strong law of large numbers

Instead of directly stating the unspecified "suitable assumptions", we work backward to see what assumptions is needed to achieve this "almost sure convergence" to the mean. Here, we are interested in the set where converges to , and we can rewrite this set as . To relate it to Borel-Cantelli lemma, we instead study its complement . By continuity of measure, we have .

It suffices to show that for any fixed , . This closely resembles the conclusion of the Borel-Cantelli lemma, so we might be tempted to show that . Unfortunately, a simple application of Chebyshev's inequality shows that this fall short, as we get

Therefore, stronger assumptions must be made if we are to use the same proof technique. Since finiteness of higher moments implies that of the the lower ones, it is natural for us to assume finite fourth moment, which gives us
Then by Borel-Cantelli lemma, converges a.s to . [1].

reference

  1. , Probability Theory: A Comprehensive Course, §1.0