The limit of random variable is not defined

Join the discussion
Ask a follow-up here, or get your own question answered by working scientists, mathematicians and engineers — people, not an autocomplete.
Real named experts · corrections over time · the nuance an AI answer skips
25 replies · 4K views
Mike.B
Messages
12
Reaction score
0
Let ##X_i## are i.i.d. and take -1 and +1 with probability 1/2 each. How to prove ##\lim_{n\rightarrow\infty}{\sum_{i=1}^{n}{X_i} }##does not exsits (even infinite limit) almost surely.
My work:
I use cauchy sequence to prove it does not converge to a real number.
But I do not how to prove it does not converge to infinity. Can some one give hints?
 
Last edited:
Physics news on Phys.org
Denote ##\sum_{k=1}^j X_k## by ##S_j##

Then If the sum ##S_j## converges in probability to infinity then
$$\forall M>0\ \forall \epsilon>0:\ \exists N>0 \textrm{ such that } j\geq N\Rightarrow Pr(S_j>M)>1-\epsilon\ \ \ \ \ (1)$$

For this not to be the case we negate the previous line to obtain:

$$\exists M>0\ \exists \epsilon>0:\ \forall N>0: \exists j\geq N\Rightarrow Pr(S_j>M)<1-\epsilon\ \ \ \ \ (2)$$

That's what has to be proved. One way to do that (not necessarily the quickest) is to use the central limit theorem. Note that ##S_j=2B_j-j## where ##B_j## is binomial with parameters ##j,0.5##, which has mean ##\frac{j}{2}## and variance ##\frac{j}{4}##. The Central Limit Theorem tells us that ##B_j## converges in probability to a random variable with distribution ##N(\frac{j}{2},\frac{j}{4})##, so that ##S_j## converges in probability to a RV ##Z_j## with distn ##N(0,j)##.

It's easy enough to prove that the sequence of RVs ##(Z_j)_{j=1}^\infty## satisfies (2). You then just need to put that together with the convergence in probability of ##S_j## to ##Z_j## to get the required result.
 
Last edited:
  • Like
Likes   Reactions: Mike.B
andrewkirk said:
Denote ##\sum_{j=1}^n X_j## by ##S_j##

Then If the sum ##S_j## converges in probability to infinity then
$$\forall M>0\ \forall \epsilon>0:\ \exists N>0 \textrm{ such that } j\geq N\Rightarrow Pr(S_j>M)>1-\epsilon\ \ \ \ \ (1)$$

For this not to be the case we negate the previous line to obtain:

$$\exists M>0\ \exists \epsilon>0:\ \forall N>0: \exists j\geq N\Rightarrow Pr(S_j>M)<1-\epsilon\ \ \ \ \ (2)$$

That's what has to be proved. One way to do that (not necessarily the quickest) is to use the central limit theorem. Note that ##S_j=2B_j-j## where ##B_j## is binomial with parameters ##j,0.5##, which has mean ##\frac{j}{2}## and variance ##\frac{j}{4}##. The Central Limit Theorem tells us that ##B_j## converges in probability to a random variable with distribution ##N(\frac{j}{2},\frac{j}{4})##, so that ##S_j## converges in probability to a RV ##Z_j## with distn ##N(0,j)##.

It's easy enough to prove that the sequence of RVs ##(Z_j)_{j=1}^\infty## satisfies (2). You then just need to put that together with the convergence in probability of ##S_j## to ##Z_j## to get the required result.
##\sum_{j=1}^n X_j## by ##S_j##?
 
What is the relation between almost surely and in probability? How to use here?
 
Last edited:
andrewkirk said:
That's correct. Almost surely is a stronger type of convergence than in probability.
I have difficulty about combining the final results, is there any inequality involved?
 
Last edited:
I got you about (2) [<1/2]
 
Hmm. On reflection, making the connection from the convergence in distribution to a normal, to a conclusion that the set of points ##\omega\in\Omega## for which ##\lim_{j\to\infty}S_j(\omega)=\infty## has zero measure is not as straightforward as I initially thought. I will think more about it when I get a decent slab of free time.
 
WWGD said:
Isn't this a random walk based, say at 0 , in the Real line? I don't know how to describe convergence in a random walk.
Yes. That's how I've been thinking about it.

The formal statement of the problem, as I understand it, is a request to prove that $$Pr(\lim_{j\to\infty} S_j=\infty)=0$$

where we interpret ##\lim_{j\to\infty}S_j=\infty## to mean that ##\forall M\in\mathbb{N}\ \exists n\in\mathbb{N}\ ## such that ##j>n\Rightarrow S_j>M##. All the instances of ##S_j## in that statement should really be written ##S_j(\omega)## where ##\omega## is an element of the sample space ##\Omega## of infinite sequences, in order to clarify that the statements are about specific sequences ##(S_j(\omega))_{j\in\mathbb{N}}##rather than about the random variables ##S_j##.

If we label by ##A## the set of all sequences/walks ##S_j## that have that property, which we call 'diverging to infinity' then we are trying to prove that ##Pr(A)=0##.

I'm starting to think that my Normal Approximation suggestion won't help though.

Another possible avenue of attack would be if we could show that ##A##is countable. Then we would know it must have measure zero, since the set ##\Omega## of all sequences is uncountable. We can also think about the sequences as binary expansions of real numbers in [0,1] and I'm wondering if any known results about the measure of subsets of that interval might help,
 
Mike.B said:
You mean ##\mathbb{P}(V=\infty)=1##?
Yes, that's what I meant. Could that work, as ##(V=\infty) \cap A =\emptyset## (I think)?
 
Samy_A said:
Yes, that's what I meant. Could that work, as ##V \cap A =\emptyset## (I think)?

Define ##B=\{\omega\in \Omega:V(\omega)=\infty\}##? And show ##A\cap B##=0$?
 
Mike.B said:
Define ##B=\{\omega\in \Omega:V(\omega)=\infty\}##? And show ##A\cap B##=0$?
If ##\omega\in A## then ##\forall M\in\mathbb{N}\ \exists n\in\mathbb{N}\ ## such that ##\forall j>n\Rightarrow S_j(\omega)>M##.
Take ##M>0##. Doesn't that imply that ##\omega## can return to 0 at most ##n## times, so ##\omega \not\in B##?
 
Last edited:
  • Like
Likes   Reactions: Mike.B and WWGD
WWGD said:
Isn't it also possible to use a recursion in this case , whose solution gives a closed form for the general probability?
What do you mean by "recursion"?
 
WWGD said:
Isn't it also possible to use a recursion in this case , whose solution gives a closed form for the general probability?

Can you tell me more?
 
Consider the recursion ## P(n,x) =[P(n-1,x-1)+P(n-1, x+1)]/2 , P(0,0)=1 , P(0,x)=0 ; x \neq 0 ## , where ##n## is the step number in the walk and ##P(n,x)## is the probability of being at ##x## at step ##n## on the Real line at step ##n##, and ##x## is any Natural number.
 
Last edited:
I tried the link given in post 12 but it was full of broken latex links and hence unusable on the computer I'm at now.
However I found the following http://www.dartmouth.edu/~chance/teaching_aids/books_articles/probability_book/Chapter12.pdf
It proves a number of results, and in Exercise 12.1 on page 6 it proves that ##w_*=1## where
$$w_*\equiv \lim_{n\to\infty} w_n$$
and ##w_n## is the probability that the random walk has returned to the origin no later than step ##n##.

We can then proceed as follows
[Edit: I discovered subsequently that this argument is flawed. See post 26 below for corrected version]

##A\subseteq B## where ##B## is the set of paths that do not return to the origin infinitely often.

##B\subseteq J_n## where ##J_n## is the set of paths that do not return to the origin in the first ##n## steps. Note that ##Pr(J_n)=1-w_n##.

Hence
$$Pr(A)\leq Pr(B)\leq \inf\{Pr(J_n)\ |\ n\in\mathbb{N}\}\leq \inf\{1-w_n\ |\ n\in\mathbb{N}\}=1-\sup\{w_n\ |\ n\in\mathbb{N}\}=1-w_*=0$$

The proof in the link that ##w_*=1## is far from trivial though, involving a generating function and requiring a proof of an interim result, at the bottom of p4 (attributed to Wilf), which is set as an exercise for the reader in Exercise 1. I have not done that exercise.
 
Last edited:
  • Like
Likes   Reactions: Mike.B
andrewkirk said:
##B\subseteq J_n## where ##J_n## is the set of paths that do not return to the origin in the first ##n## steps.
I don't understand this step: how do you know that ##B\subseteq J_n##?
 
@Samy_A You're right, that's not valid. Let me fix that.
First, change the definition of ##A## slightly to be the union of the sets of paths that diverge to positive and negative infinity respectively.

##B=\bigcup_{j=0}^\infty C_j## where ##C_j## is the set of paths that are zero at time ##k## and never thereafter. Note that this is a disjoint union.
Then

$$Pr(A)\leq Pr(B)=\sum_{k=0}^\infty Pr(C_k)
=\sum_{k=0}^\infty Pr(C_k)Pr(C_k|D_k)$$

where ##D_k## is the set of paths that is zero at time ##k##.

By the independence of the ##X_j##s, we have that ##\forall k\geq 0\ Pr(C_k)=Pr(C_0)##.
And if, as before we define ##J_k## to be the set of paths that do not return to the origin in the first ##k## steps, then

$$Pr(C_0)=Pr(\bigcup_{k=1}^\infty
J_k)
\leq
\lim_{k\to\infty}Pr(J_k)=\lim_{n\to\infty} (1-w_n)
=1-\lim_{n\to\infty} w_n
=1-w_*
=1-1
=0
$$

So $$Pr(A)\leq
\sum_{k=0}^\infty (Pr(D_k)\cdot 0)=0$$
 
Last edited:
  • Like
Likes   Reactions: Samy_A