Past exam of the probability course of the University of Oxford / 2021 (A8, A13764W1)

Source: Oxford's released examination PDF, pages 164–167, dated 21 June 2021. Questions 1, 2 and 3 occupy bundle pages 165, 166 and 167 respectively, or pages 2, 3 and 4 of the individual paper. All three questions and every part are solved below, including all six revisions requested in 3(d). In 3(c), $c$ is understood to be an integer: the printed transition rule does not define a transition matrix for noninteger $c$.

The cumulative distribution function of the uniform distribution on $[0,1/n]$ isThe constant zero has cumulative distribution function $F(t)=1_{[0,\infty )}(t)$. Its continuity points are precisely $t\ne 0$. For fixed $t<0$, $F_n(t)=0=F(t)$; for fixed $t>0$, $F_n(t)=1=F(t)$ eventually. This proves convergence in distribution to zero directly from the definition.

Also, $\mathbb{P}(0\le U_n\le 1/n)=1$. For every $\epsilon >0$ and $n>1/\epsilon$, $\mathbb{P}(|U_n|>\epsilon )=0$. HenceThe convergence in probability conclusion holds for any realization of these random variables on a common probability space; no independence of random variables is needed.

Fix $\delta >0$. Once $|c_n-c|<\delta$, the following event inclusions hold outside $\{|X_n-X|>\delta \}$:Taking probabilities gives the explicit boundsBy the assumed convergence in probability, the error tends to zero. ThereforeThe continuity of the cumulative distribution function $F$ at $c$ lets $\delta$ decrease to zero, proving $\boxed{F_n(c_n)\to F(c).}$ This proves the needed implication directly, without invoking any theorem comparing convergence of random variables.

The definition of convergence in distribution gives $F_n(t)\to F(t)$ at each continuity point $t$ of $F$. We must handle a moving argument rather than assume this convergence is uniform.

A cumulative distribution function has at most countably many discontinuities: each is a jump of positive size, and there are at most $r$ jumps of size at least $1/r$, since their sizes sum to at most one. Thus we can choose continuity points $u_j<c<v_j$ with $u_j\uparrow c$ and $v_j\downarrow c$. For fixed $j$, eventually $u_j<c_n<v_j$. The monotonicity of each cumulative distribution function impliesFirst let $n\to \infty$, using the definition of convergence in distribution at $u_j,v_j$. Then let $j\to \infty$, using continuity at $c$:No result relating different convergence of random variables has been assumed.

Use $X_n\sim \mathrm{Unif}[0,1/n]$ from part (a). We already proved convergence in distribution to zero. For any prescribed $p\in [0,1]$, put $c_n=p/n$. Then $c_n\to 0$, and the cumulative distribution function satisfiesThis includes $p=0$ and $p=1$, since the uniform distribution assigns zero probability to each endpoint. The limit cumulative distribution function is discontinuous at zero, explaining why the conclusion of (ii) does not apply there.

Write $\Phi$ for the standard normal distribution function. The Poisson process in this part has rate one and starts at $N_0=0$.

Set $T_0=0$ and $E_j=T_j-T_{j-1}$. The Poisson interarrival times $E_j$ are independent and identically distributed random variables with the rate-one exponential distribution. In particular, $\mathbb{E}{E}_j=1$ and $\mathrm{Var}(E_j)=1$, so the finite, positive variance hypothesis of the central limit theorem holds. Since $T_n={\sum }_{j=1}^n{E}_j$,

The unit-time increments $D_j=N_j-N_{j-1}$ are independent and identically distributed random variables, by independent increments and stationary increments of the Poisson process. Each has the Poisson distribution of parameter one, hence expected value one and variance one. The central limit theorem therefore gives

There is a stronger conclusion than the requested limit. For $x>0$, the event $N_n-n\ge x\sqrt{n}$ forces $N_n\ge n+1$, because $N_n$ is an integer. In particular, at least $n$ arrivals have occurred by time $n$, so the $n$-th Poisson arrival time satisfies $T_n\le n$. This contradicts $T_n-n\ge y\sqrt{n}$ for $y>0$. Thus the two events are disjoint:

Separate a long future count from the early Poisson arrival time. Let ${\mathcal{F}}_t=\sigma (N_s:0\le s\le t)$ be the natural filtration, and defineThe truncated Poisson arrival time in $B_n$ is ${\mathcal{F}}_{n^2}$-measurable. The independent increments property makes $A_n$ independent of ${\mathcal{F}}_{n^2}$, and hence of $B_n$.

The numerator of $A_n$ is a centered Poisson random variable of parameter $n^4-n^2$. Apply the central limit theorem to the $n^4-n^2$ unit-time increments, then use $\sqrt{n^4-n^2}/n^2\to 1$ and Slutsky's theorem. This proves $A_n\stackrel{\mathrm{d}}{\to }N(0,1)$. Also, Markov's inequality and $\mathbb{E}{T}_n=n$ giveThus $B_n$ equals $(T_n-n)/\sqrt{n}$ outside an event of vanishing probability, so part (i) and Slutsky's theorem give $B_n\stackrel{\mathrm{d}}{\to }N(0,1)$. Exact independence of random variables for each $n$ now gives the joint limit $(A_n,B_n)\stackrel{\mathrm{d}}{\to }(Z_1,Z_2)$ with independent standard normal random variables.

Finally, $\mathbb{E}{C}_n=0$ and $\mathrm{Var}(C_n)=n^{-2}$, so Chebyshev's inequality gives $C_n\to 0$ in probability. The first coordinate in the question is $A_n+C_n$. The second differs from $B_n$ only on $\{T_n>n^2\}$, and so its difference tends to zero in probability. The vector version of Slutsky's theorem proves the required joint convergence in distribution. The boundary of $(-\infty ,x]\times [y,\infty )$ has zero limiting probability, since both marginal distributions have continuous cumulative distribution functions. The Portmanteau theorem therefore yieldsThis is $\mathbb{P}(Z_1\le x,Z_2\ge y)$ by independence of random variables.

The possible limiting correlation coefficients areThe direction of the dependence comes from inversion of the same counting process: extra arrivals produce a larger count but an earlier $n$-th Poisson arrival time. Shared fluctuations therefore enter the two coordinates with opposite signs. Positive limiting correlation cannot arise; this can be formalized using negative quadrant dependence of a Poisson count and an arrival time.

At the same time scale, $f(n)=n$, part (iii) rules out every positive-positive quadrant. A bivariate normal distribution with standard normal marginal distributions and $\rho >-1$ assigns positive probability to such a quadrant. Thus, under the assumed joint limit, this choice forces $\rho =-1$. At very different time scales, $f(n)=n^4$, part (iv) gives $\rho =0$.

Intermediate time scales retain an intermediate fraction of the same arrival fluctuations. In particular, the Poisson count-arrival joint central limit theorem gives $\rho =-1/\sqrt{\lambda }$ for $f(n)=\lambda n$ with $1\le \lambda <\infty$. These choices realize every $\rho \in [-1,0)$, and $f(n)=n^4$ supplies zero. This identifies the whole interval without requiring an additional limit calculation in this part.

The premise about a bivariate normal distribution is understood to apply when a joint limit exists. Merely requiring $f(n)\to \infty$ cannot guarantee one: taking $f(n)=n$ for even $n$ and $f(n)=n^4$ for odd $n$ gives different limits along the two subsequences just identified.

The uniqueness theorem for moment-generating functions states: if two real random variables $V,W$ have finite moment-generating functions on a common open interval $(-\delta ,\delta )$ with $\delta >0$, andthen $\boxed{V\text{ and }W\text{ have the same probability distribution}.}$ Finiteness throughout a neighborhood of zero is a hypothesis; the identity $M_V(0)=M_W(0)=1$ alone carries no information.

Let $X_f$ have the specified probability density function $f_m$. Splitting the integral into the positive and negative half-lines, and using the Gamma integral ${\int }_0^{\infty }{x}^r{e}^{-sx}dx=r!/s^{r+1}$ for a nonnegative integer $r$ and $s>0$, givesAt $t=2$ or $t=-2$, one term becomes the divergent integral of $x^m$; beyond these endpoints it grows exponentially. Consequently, the full extended-valued moment-generating function isIn particular $M_f(0)=1$, checking normalization of a probability distribution. All absolute moments exist. Direct integration yieldsThe symmetric distribution makes every odd moment zero, so

Now let $X_g$ have probability density function $g_m$. Its moment-generating function would beFor every $t\ne 0$, one summand has exponential growth, which dominates the polynomial decay. HenceThere is no finite moment-generating function on any neighborhood of zero, but some moments still exist. Specifically,ThusFor even $k\ge m+1$, the moment is $+\infty$, so it is not a finite moment. For odd $k\ge m+1$, both the positive and negative contributions have infinite integral, so the expected value is undefined. Cancellation in a Cauchy principal value does not define an expected value. In particular, when $m=0$ there are no finite positive-order moments.

First justify that the moments being used exist. For every $r>0$,The bound $|y{|}^k\le k!r^{-k}{e}^{r|y|}$ gives $\mathbb{E}|Y{|}^k<\infty$ for all $k$.

If $Y$ has a symmetric distribution, then $Y$ and $-Y$ have the same probability distribution, so for odd $k$,which forces $\mathbb{E}{Y}^k=0$.

Conversely, suppose every odd moment vanishes. For fixed $t\in \mathbb{R}$, the absolute values of the partial Taylor series for $e^{tY}$ are bounded by $e^{|t||Y|}$, whose expected value is finite by the preceding inequality. The dominated convergence theorem therefore permits termwise expectation:The uniqueness theorem for moment-generating functions implies equality of the probability distributions of $Y$ and $-Y$. Hence

Let $X$ have the Pareto distribution with probability density function $f_X(x)=3x^{-4}$ for $x\ge 1$, and zero otherwise. Its normalization of a probability distribution and first two moments follow fromSet $Y=X-3/2$. ThenThe probability distribution of $Y$ is not a symmetric distribution: $Y\ge -1/2$ almost surely, whereas $\mathbb{P}(Y>1/2)=\mathbb{P}(X>2)=1/8>0$ and $\mathbb{P}(Y<-1/2)=0$. For $X\ge 3$ we have $|Y|\ge X/2$, soThus $\boxed{Y=X-3/2}$ has every property required.

Take a discrete probability distribution on three points:The probabilities are positive and sum to one. Direct calculation givesandNevertheless it is not a symmetric distribution, since $\mathbb{P}(Y=3)=1/10$ while $\mathbb{P}(Y=-3)=0$. Vanishing of just the first and third moments does not force all odd moments to vanish.

It suffices that the moment-generating function be finite on some open interval containing zero:Equivalently, some two-sided exponential moment exists: $\mathbb{E}{e}^{r|Y|}<\infty$ for some $r>0$. For the forward implication, choose $0<r<\delta$ and use $e^{r|Y|}\le e^{rY}+e^{-rY}$; the converse follows from $e^{tY}\le e^{r|Y|}$ for $|t|\le r$.

Under this weaker hypothesis, all moments exist. For $|t|<r$, dominated convergence justifies the Taylor series argument in (i). Vanishing odd moments gives $M_Y(t)=M_{-Y}(t)$ near zero, and the local uniqueness theorem for moment-generating functions gives the symmetric distribution. The reverse implication follows directly from equality of the two probability distributions.

More generally, it suffices that all moments exist and the probability distribution have moment determinacy: $Y$ and $-Y$ then have identical moments and hence identical probability distributions. Existence of all moments by itself does not provide that uniqueness.

The support condition $-y<x<y$ forces $y>0$. As a check, the joint probability density integrates toFor any real $x$, the marginal density of $X$ isDividing the joint probability density by this marginal density gives the conditional density:Thus, conditionally on $X=x$, $Y-|x|$ has the rate-two exponential distribution.

For $U=X^2$, use the two inverse branches $x=\sqrt{u}$ and $x=-\sqrt{u}$. Each branch has absolute Jacobian determinant $1/(2\sqrt{u})$. Both lie in the original support exactly when $u>0$ and $y>\sqrt{u}$. The change of variables for probability densities therefore yieldsNo atom of a measure is created at $u=0$, since $\mathbb{P}(X=0)=0$. The value of the probability density function at this boundary may be assigned arbitrarily. Checking the integral over $0<u<y^2$ gives $4y{e}^{-2y}$, which integrates to one.

Put $q=1-p=a/(a+b)$. The random walk has independent and identically distributed increments $Y_n$ equal to $a$ with probability $p=b/(a+b)$ and $-b$ with probability $q$. Thus $S_0=0$, $S_n={\sum }_{j=1}^n{Y}_j$, and $\mathbb{E}{Y}_n=0$.

The history $\sigma (S_0,\dots ,S_n)$ is contained in $\sigma (Y_1,\dots ,Y_n)$. The increment $Y_{n+1}$ is independent of this history by the given independence of random variables. ThereforeThis depends only on the current state and not on time. It is precisely the time-homogeneous Markov property, so $S_n$ is a Markov chain with transition probabilitiesThe conditional identity is an identity that holds almost surely; histories of zero probability require no division by their probabilities.

Here $a,b$ are coprime integers.

Fix any two integers $i,j$. By Bézout's identity, there are integers $u,v$ with $au-bv=j-i$. Replacing $(u,v)$ by $(u+kb,v+ka)$ leaves this identity unchanged. Choose $k$ large enough that both entries are nonnegative. A path with $u+kb$ upward steps and $v+ka$ downward steps then takes $i$ to $j$, with strictly positive probability $p^{u+kb}{q}^{v+ka}$. There are no boundary restrictions on this state space.

This applies in both directions, so every pair of states communicates. Hence

A return path with $u$ upward steps and $v$ downward steps satisfies $au=bv$. Since $a,b$ are coprime, this is equivalent to $u=bk$ and $v=ak$ for an integer $k\ge 1$. Its length is therefore $k(a+b)$.

Conversely, for each such $k$, any ordering of $bk$ upward and $ak$ downward steps is an allowed positive-probability return path. The possible strictly positive return times are exactly the positive multiples of $a+b$. By the definition of the period of a state in a Markov chain as their greatest common divisor,In particular, none of the states is aperiodic.

Suppose a stationary distribution $\pi$ exists. Its stationarity equations areSince ${\sum }_j{\pi }_j=1$, its masses tend to zero as $|j|\to \infty$, and a positive maximum $M={\pi }_{j_0}$ is attained. Both terms on the right are at most $M$. Since $p,q>0$ and $p+q=1$, equality forces ${\pi }_{j_0-a}={\pi }_{j_0+b}=M$. Applying the same argument repeatedly at $j_0-a,j_0-2a,\dots$ givesThese are infinitely many distinct states, contradicting ${\sum }_j{\pi }_j=1$. ThusThis maximum principle proof uses neither a finite expected value of the state nor the coprimality assumption.

First identify the recurrence type, rather than infer an occupation limit from a one-time probability. At time $k(a+b)$, return to zero requires exactly $kb$ upward steps. The binomial distribution giveswhere the last step is Stirling's approximation. At all other positive times the return probability is zero. Hence ${\sum }_{n\ge 0}{\mathbb{P}}_0(S_n=0)=\infty$. The recurrence criterion by return probabilities proves the recurrent property; by the irreducible Markov chain property established in (i), every state is recurrent. The equivalence of positive recurrence and existence of a stationary distribution for an irreducible Markov chain on a countable state space, together with (iii), makes this chain null recurrent.

For completeness, the zero occupation proportion follows directly by treating returns as a renewal process. Let $L_1,L_2,\dots$ be successive strictly positive return-cycle lengths to zero. The Strong Markov property makes them independent and identically distributed, the recurrent property makes them finite almost surely, and the null recurrent property gives $\mathbb{E}{L}_1=\infty$. For each positive integer $K$, the strong law of large numbers applied to $\min (L_j,K)$ givesTake a common event of probability one for these countably many $K$. By the monotone convergence theorem, the right side increases to infinity. Thus the number of completed return cycles by time $n$, divided by $n$, tends to zero. Counting the initial visit changes this ratio by at most $1/n$. Consequently,The walk visits zero infinitely often, but the asymptotic fraction of time spent there is zero.

Take an integer $c\ge \max \{a,b\}$. On $I=\{-c,\dots ,c\}$, the modified transition matrix $Q=(q_{i,j})$ agrees with $P$. Above $c$, it stays put with probability $a/(a+b)$ and jumps to each of $c-b+1,\dots ,c$ with probability $1/(a+b)$. Below $-c$, it stays put with probability $b/(a+b)$ and jumps to each of $-c,\dots ,-c+a-1$ with probability $1/(a+b)$. All unspecified entries are zero. These rows sum to one. The integer qualification is necessary: if $c$ were nonintegral, $i=\lfloor c\rfloor +1$ would fit none of the printed row conditions.

DefineAn interior jump cannot go below $-c-b$ or above $c+a$, and every exterior reset lands in $I$. Thus $C$ is closed.

To prove that it is one communicating class, start anywhere in $C$. From $L$ or $U$, a reset can reach $I$ with positive probability. From any state of $I$, repeated $+a$ steps reach some state of $U$, after staying in $I$ up to the last step. A reset from this state can reach every point of the consecutive blockEvery integer $j\in \{-c-b,\dots ,c\}$ has the form $t-kb$ for some $t\in B$ and an integer $k\ge 0$. Repeated $-b$ steps from $t$ reach $j$, with all departure states in $I$ until a possible last step into $L$. This gives access to all of $I\cup L$. Finally, each $j\in U$ can be reached from $j-a\in I$. Every path described has strictly positive probability. Therefore every state in $C$ reaches every other state in $C$.

Each state outside $C$ can only stay put or reset into $I$, and no transition from $C$ returns to it. Different states outside $C$ cannot reach one another. Hence the complete list of communicating classes isThe finite class $C$ is the unique closed communicating class. The exterior singleton classes are transient communicating classes, since their self-transition probabilities are strictly less than one. Notice that this reachability proof does not require $a,b$ to be coprime.

The restriction $Q_C$ is a finite Markov chain that is an irreducible Markov chain. It is aperiodic, because its upper states have self-transition probability $a/(a+b)>0$, and the period is constant on a communicating class. We now find its stationary distribution.

Every column of $Q_C$ sums to one. Here is the full incoming-mass check, with each total measured in units of $1/(a+b)$. For $j\in L$, the incoming contributions are $a$ from $j+b\in I$ and $b$ from the self-transition. For $j\in U$, they are $b$ from $j-a\in I$ and $a$ from the self-transition.

For $j\in I$, the upward predecessor $j-a$ lies in $I$ exactly when $j\ge -c+a$, contributing $b$ units. When $j\le -c+a-1$, that predecessor is replaced by resets from all $b$ states of $L$, again contributing $b$ units. Independently, the downward predecessor $j+b$ contributes $a$ units when $j\le c-b$; when $j\ge c-b+1$, resets from the $a$ states of $U$ contribute those $a$ units instead. Thus every column total is $a+b$ units.

Therefore $Q_C$ is a doubly stochastic matrix, and the uniform distribution on $C$ is stationary:The countable-state Markov chain convergence theorem, applied to the finite closed class $C$, gives $Q_C^n(i,j)\to {\pi }_j$ for every $i\in C$. Its hypotheses hold: $C$ is finite and irreducible, hence positive recurrent, and it is aperiodic.

If $i_0>c+a$, the hitting time $H$ of $C$ has $\mathbb{P}(H>n)=(a/(a+b){)}^n$. If $i_0<-c-b$, the corresponding tail is $(b/(a+b){)}^n$. Thus $H<\infty$ almost surely from every exterior initial state. Conditioning on $H$ and $R_H$, use convergence within $C$ after entrance and the bounded convergence theorem; equivalently, first restrict to $H\le K$ and then let $K\to \infty$. This proves, for every $i_0\in \mathbb{Z}$,The limiting distribution is independent of $i_0$.

Let $d=\gcd (a,b)>1$ and put $a^{\prime }=a/d$, $b^{\prime }=b/d$. Then $a^{\prime },b^{\prime }$ are coprime integers, and $p=b^{\prime }/(a^{\prime }+b^{\prime })$. The six requested revisions are as follows.

(b)(i): communicating classes. Every increment is divisible by $d$, so the residue class modulo $d$ is preserved. Conversely, if $j-i$ is divisible by $d$, Bézout's identity applied to $a^{\prime },b^{\prime }$ gives a path just as in (b)(i). Thus the original Markov chain is reducible on $\mathbb{Z}$, with exactly the $d$ closed communicating classesThe specified walk starting at zero stays in $d\mathbb{Z}$.

(b)(ii): period of a state in a Markov chain. The return equation $au=bv$ reduces to $a^{\prime }u=b^{\prime }v$. Hence $u=b^{\prime }k$, $v=a^{\prime }k$. For every state, the period of a state in a Markov chain is thereforeThis is still at least two, so no state is aperiodic.

(b)(iii): stationary distribution. The conclusion is unchanged: there is no stationary distribution on $\mathbb{Z}$. The maximum-mass proof in (b)(iii) applies verbatim, without using coprimality. Equivalently, each closed communicating class, after translating and dividing its state labels by $d$, is the same centered random walk with step sizes $a^{\prime },-b^{\prime }$ and has no stationary distribution.

(b)(iv): occupation of zero. On $d\mathbb{Z}$, the rescaled walk $S_n/d$ has exactly the transition probabilities of the coprime walk with parameters $a^{\prime },b^{\prime }$. The return-probability argument in (b)(iv) therefore makes this class null recurrent. The renewal process proof of the occupation limit applies unchanged:Starting in another residue class would make every visit to zero impossible, also giving occupation proportion zero.

(c)(i): modified communicating classes. These are unchanged: $C=\{-c-b,\dots ,c+a\}$ is one finite closed communicating class, and all other states are transient singleton classes. The original proof used consecutive reset destinations and never used coprimality. Resetting to the $b$ consecutive integers $c-b+1,\dots ,c$ connects different residue classes, so it would be incorrect to split $C$ according to residue classes modulo $d$.

(c)(ii): limiting distribution. The column-sum calculation, the irreducible Markov chain property on a finite class, positive self-transition probabilities, and geometric entrance argument also required no coprimality. Thus for every initial state the same convergence in distribution holds:In particular, the denominator $2c+a+b+1$ is unchanged; it is not divided by $d$.