Probability that a head eventually turns up (From Grimmett and Stirzaker) This problem is from the exercise section of Grimmett and Stirzaker. I have also listed my proof.
Question.

A fair coin is tossed repeatedly.
(a) Show that a head turns up sooner or later with probability one.
(b) Show similarly that any given finite sequence of heads and tails occurs eventually with probability one.
(c) Explain the connection with Murphy's law.

Proof.

Let $A_{1},A_{2},A_{3},...,A_{n}$ be the events that a head turns up on the $(1,2,3,...,n)$-th toss after a sequence of $n-1$ tails. This can be visualized in the form of binomial tree. The sample space $\Omega=\{H,TH,TTH,TTTH,...\}$ is a countably infinite set.
The probability measure $P$ is countably additive over disjoint sets.
$P(\bigcup\limits_{i=1}^{\infty}{A_{i}})=\sum_{i=1}^{\infty}P(A_{i})$
In this example, all the members of $\mathscr{F}$ are disjoint.
Hence, $P(\bigcup\limits_{i=1}^{\infty}{A_{i}})=P(A_{1})+P(A_{2})+...=(1/2)+(1/2^2)+(1/2^3)+...\infty=1$.
Thus, a head will turn up with a probability 1.

I would like to know, if this proof is correct and sound. Are there alternative approaches to it? Also, I was wondering, how I can extend it to a sequence of heads and tails of length $m$. What is Murphy's law? How do I define the $\sigma$-field $\mathscr{F}$?
 A: Your proof is correct, and extending it is probably possible, although I don't immediately see a way. An alternate way, which is easier to extend to the second case is the following.
The probability that no heads occur in $N$ tosses is $2^{-N}$. As $N \to \infty$, this goes to $0$. Thus, the probability that at least one head occurs in an infinite experiment is $1$.
This is easier to extend this as follows:
Pick any sequence of $m$ digits. The probability that it doesn't occur in $N$ tosses is upper bounded by the probability that it doesn't occur as sequence $(x_{jm - m}, \dots, x_{jm})$ for $1 \le j \le \lfloor N/m \rfloor$. The latter probability is $\left(\frac{2^m - 1}{2^m}\right)^{\lfloor N/m \rfloor} \to 0$ as $N \to \infty$. 
Murphy's law is a tongue in cheek folk-saying that states that anything that can go wrong, will go wrong. Through the above, this is true for, well, coins, assuming some finite sequence makes things wrong for you, and if you're willing to hang around for a while.
A: I don't think your proof is right because of the definitions of $A_n$ and $\mathscr F$. Are the flips independent?

Consider a probability space $(\Omega, \mathscr F, \mathbb P)$ where


*

*$\Omega = \{H,T\}^{\mathbb N}$


So we have $\omega = (\omega_1, \omega_2, ...)$ where $\omega_n \in \{H, T\} \ \forall n \in \mathbb N$


*$\mathscr{F} = \sigma(\omega_n = W | \ W \in \{H, T\})$ like here (because I guess $\mathscr{F} = 2^{\Omega}$ doesn't work)

*$P(\omega_n = H) = P(\omega_n = T) = 1/2$
where
$(\omega_n = H) = (\omega_1, \omega_2, ..., \omega_n = H, ...)$
$(\omega_n = T) = (\omega_1, \omega_2, ..., \omega_n = T, ...)$

In your case, $(\omega_n = H) = A_n$
So $P(A_n) = 1/2$ not $\frac{1}{2^n}$

Let $H_1, H_2, ...$ be events where $H_n$ = {nth flip is heads and 1st, ..., (n-1)th flips are tails}.
Thus, we have $$H_n = A_1^C \cap A_2^C \cap ... \cap A_{n-1}^C \cap A_n$$
$$=\bigcap_{k=1}^{n-1} [A_k^C] \cap A_n$$
Assuming independence of the flips i.e. independence of the $A_n$'s, $$P(H_n) = [\prod_{k=1}^{n-1} P(A_k^C)] \times P(A_n) = \frac{1}{2^n}$$

Now for Q1, we want to show that at least one of the flips will be heads
Or:
$$P(\bigcup_{n=1}^{\infty} A_n) = 1$$
Or:
Almost surely, $\forall \omega \in \Omega$,
$$\omega \in \bigcup_{n=1}^{\infty} A_n$$
Or:
$$\exists z \ge 1 s.t. P(A_z) = 1$$
Or:
Almost surely, $\forall \omega \in \Omega$,
$$\exists z \ge 1 s.t. \omega \in A_z$$



*

*One route: Now we can do what you attempted earlier because the $H_n$'s are pairwise disjoint (*): $$P(\bigcup_{n=1}^{\infty} H_n) = \sum_{n=1}^{\infty} P(H_n) = 1 \ \text{if the flips are independent}$$


Hence almost surely, $\forall \omega \in \Omega$,
$$\omega \in \bigcup_{n=1}^{\infty} H_n$$
$\to \exists! \ q \in \mathbb N$ s.t. $\omega \in H_q$
Observe that $H_q \subseteq A_q$.
Hence, $$\omega \in A_q \subseteq \bigcup_{n=1}^{\infty} A_n \ QED$$
Or prove that $$\bigcup_{n=1}^{\infty} H_n = \bigcup_{n=1}^{\infty} A_n$$

(*) They are pairwise disjoint because:
$\forall m > n$,
$$\omega \in H_n \cap H_m$$
$$\to \omega \in H_n \cap \omega \in H_m$$
So, $$\omega \in H_n \to \omega \in A_n$$
However, $$\omega \in H_m \to \omega \in A_n^c ↯  \ QED$$



*Another route:


$$P(\bigcup_{n=1}^{\infty} A_n) = 1 - P(\bigcap_{n=1}^{\infty} A_n^C)$$
$$= 1 - \prod_{n=1}^{\infty} P(A_n^C) \ \text{if the flips are independent}$$
$$= 1 - \prod_{n=1}^{\infty} (1/2)$$
$$= 1 - \lim_{m \to \infty} \prod_{n=1}^{m} (1/2)$$
$$= 1 - \lim_{m \to \infty} (1/2)^m = 1 - 0 = 1 \ QED$$



*Yet another route:


$$\because \sum_{n=1}^{\infty} P(A_n) = \infty,$$
by Borel-Cantelli Lemma 2, if the flips are independent, we have $P(\limsup A_n) = 1$
Observe that $$\limsup A_n \subseteq \bigcup_{n=1}^{\infty} A_n$$
Hence, by monotonicity of probability, $$P(\bigcup_{n=1}^{\infty} A_n) = 1 \ QED$$
Or:
Hence, almost surely, $\forall \omega \in \Omega, \forall m \ge 1, \exists n \ge m$ s.t.
$$\omega \in A_n \subseteq \bigcup_{n=1}^{\infty} A_n$$

For Q2, let $B_{n,r}$ be a block of length r where
$$B_{n,r} = \bigcap_{i=n}^{n+r-1} A_i*$$
where $A_i* = A_i$ or $A_i^C$
$$\because \sum_{n=1}^{\infty} P(B_{n,r}) = \sum_{n=1}^{\infty} (\frac{1}{2^r}) = \infty,$$
using BCL2 again gives us
$$P(\limsup B_{n,r}) = 1$$
This means that almost surely $\forall \omega \in \Omega, \forall m \ge 1, \exists n \ge m$ s.t. $\omega \in B_{n,r} \ QED$
To see why the statement doesn't hold for a block of infinite length, define $$B_{n, \infty} := \lim_{r \to \infty} B_{n, r}$$
By the continuity of probability, $P(B_{n, \infty}) = \lim_{r \to \infty} P(B_{n, r}) = \lim_{r \to \infty} \frac{1}{2^r} = 0$
$$\because \sum_{n=1}^{\infty} P(B_{n,\infty}) = \sum_{n=1}^{\infty} 0 < \infty,$$
BCL1 gives us
$$P(\limsup B_{n,\infty}) = 0$$

For Q3, Murphy's Law is 'Anything that can go wrong, will go wrong', w/c is technically false:
Flipping 10 coins is a 'thing'. If we define 'go wrong' to be 'at least one head', then we may have 10 tails.
Mathematically,
$P(\bigcup_{n=1}^{10} E_n) \ne 1$ even if $P(E_n) > 0$
or
$P(\bigcup_{n=1}^{\infty} E_n) \ne 1$ where $E_k = \emptyset$ for $k \ge 11$ even if $P(E_n) > 0$ for $n = 1, 2, ..., 10$
Even if we flip infinitely, but $P(E_k) = 0$ for $k \ge 11$, we still may not have $P(\bigcup_{n=1}^{\infty} E_n) = 1$
Let us add the condition that it is not the case that all but a finite number of the $E_n$'s have zero probability (the $E_n$'s have positive probability infinitely often) to have Murphy's Law #2.
To put Murphy's Law #2 mathematically,
$$P(E_n) > 0 \text{i.o} \to P(\bigcup_n E_n) = 1$$
Is Murphy's Law #2 true? If not, what are some sufficient conditions for Murphy's Law #2?
Case 0: $\exists z \in \mathbb N s.t. P(E_z) = 1$
Obviously, Murphy's Law #2 holds.
Case 1: $E_n$'s are independent, $P(E_n) < 1$
$$P(\bigcup_{n=1}^{\infty} E_n) = 1 - P(\bigcap_{n=1}^{\infty} E_n^C)$$
$$= 1 - \prod_{n=1}^{\infty} P(E_n^C) = 1$$
Case 2: $E_n$'s are not independent but disjoint and $\sum_n P(E_n) = 1$, $P(E_n) < 1$
$$P(\bigcup_{n=1}^{\infty} E_n) = \sum_n P(E_n) = 1$$
Case 3: $E_n$'s are disjoint, not independent but $\sum_n P(E_n) = \infty$, $P(E_n) < 1$
Impossible.
Case 4: $E_n$'s are not independent but $\sum_n P(E_n) = \infty$, $P(E_n) < 1$
$$\sum_n P(E_n) = \infty to P(\limsup E_n) = 1 \to P(\bigcup_{n=1}^{\infty} E_n) = 1$$
Just kidding. BCL2 needs independence.
Case 5: $E_n$'s are not independent but $1 < \sum_n P(E_n) < \infty$, $P(E_n) < 1$
Here, we have $P(\liminf E_n^C) = 1$. So for some m, $\omega \in E_m^C, E_{m+1}^C, ...$.
$\omega$ may or may not be in $\bigcup_{n=1}^{m-1} E_n$.
So Murphy's Law #2 does not hold.

To sum up:
Case 0 is obvious. Case 1 corresponds to Q1. Case 2 corresponds to Q2. Case 3 is impossible. Cases 4 and 5 suggest counterexamples.
