The Poisson Process

What the hypothesis "events form a Poisson process" asserts, in three equivalent forms, and how each form becomes a test. Proves that independent Poisson increments, independent exponential gaps and the infinitesimal definition describe the same process; that the exponential is the only memoryless waiting time; that given the count the arrival times are uniform order statistics; that thinning splits a Poisson process into independent ones and superposition merges them; and the mean, variance and generating function of a compound Poisson count. Tests the hypothesis on 6,992 earthquakes of magnitude 6 and over from 1975 to 2024, where every test fails because earthquakes cluster and the yearly rate drifts, and the compound Poisson model predicts the excess variance of monthly counts; derives the Fisher information of a page's change rate estimated from visits that only see whether it changed, with the best visit interval. Worked problems on uniform times given a count, thinning by magnitude, clusters as a compound Poisson count and a change rate with its interval.

From 1975 to 2024 the catalog of the U.S. Geological Survey recorded 6,992 earthquakes of magnitude 6 or more somewhere on Earth, 139.84 a year on average (U.S. Geological Survey, 2017). The catalog gives the origin time of each one. The standard model for times of events that occur independently at a steady rate is the Poisson process. What exactly does that model assert about a sequence of times, and how can each part of the assertion be checked against a catalog like this one?

A Poisson process with rate $\lambda$ is a random set of times on $[0, \infty)$ whose counts in disjoint intervals are independent and Poisson distributed, with mean $\lambda$ times the length of the interval. The course article The Poisson Distribution derives the Poisson count of such a process from three assumptions about short intervals, shows that the waiting time to the first event is exponential, and extends the counts to a rate that varies in time. This article starts where that one stops. It assumes Conditional Expectation and the joint densities of Random Variables and Distribution Functions, and it is part of Stochastic Processes and Time Series.

First we state three definitions and prove that they describe the same process, so that a test of any one of them tests the model. Then we prove three properties that turn into tests or tools: the exponential is the only waiting time without memory, the arrival times are uniform once their number is known, and a Poisson process can be split and merged without leaving the family. Next come sums of random amounts at Poisson times. Then the earthquakes are tested and fail, in an informative way. Finally we return to a question from web crawling: how precisely can a page's rate of change be estimated from visits that see only whether it changed?

Three definitions of one process

A counting process $N(t)$, $t \ge 0$, counts the events in $[0, t]$. It starts at $N(0) = 0$, takes whole-number values, never decreases, is continuous from the right, and is finite for every $t$. Its arrival times are $S_n = \min\{t : N(t) \ge n\}$ for $n = 1, 2, \ldots$, and its interarrival times (gaps) are $X_1 = S_1$ and $X_n = S_n - S_{n-1}$. Right continuity gives the identity that ties times to counts:

$$\{S_n \le t\} = \{N(t) \ge n\}.$$

The arrival times are a random walk with positive steps, $S_n = X_1 + \cdots + X_n$, the object that Random Walks and Martingales studies with steps of either sign. Three descriptions of the Poisson process are in use, each natural in a different setting. Write $N(s, t] = N(t) - N(s)$ for the count in $(s, t]$.

  • (A) Increments. For every $0 = t_0 < t_1 < \cdots < t_m$, the counts $N(t_{i-1}, t_i]$ are mutually independent, and $N(s, t]$ has the Poisson distribution with mean $\lambda(t - s)$.
  • (B) Gaps. The interarrival times $X_1, X_2, \ldots$ are independent and identically distributed (i.i.d.), each exponential with rate $\lambda$, that is with $P(X_k > x) = e^{-\lambda x}$ for $x \ge 0$; and $N(t)$ is the number of $n \ge 1$ with $S_n \le t$.
  • (C) Short intervals. For any finitely many disjoint intervals, the counts are mutually independent, and in every interval of length $h$ the probability of exactly one event is $\lambda h + o(h)$ and of two or more is $o(h)$, where each $o(h)$ is a function of $h$ alone whose ratio to $h$ tends to 0.

Definition (C) is the one that The Poisson Distribution starts from. Definition (B) is how a Poisson process is simulated. Definition (A) is the one that statistical tests of counts use. Durrett (2019, Section 3.7) gives (A) as the definition, constructs the process from (B), and derives the Poisson count from a form of (C) (his Theorem 3.7.2).

Theorem 1 (the definitions agree). A counting process satisfies (A) if and only if it satisfies (C). A process built by (B) satisfies (A), and every process satisfying (A) has interarrival times that are i.i.d. exponential with rate $\lambda$, so it satisfies (B).

Here is the plan. The step from gaps to counts is a calculation with the joint density of the arrival times, and two lemmas prepare it. The step from counts back to gaps uses the identity $\{S_n \le t\} = \{N(t) \ge n\}$: the counts determine the distribution of the times. The equivalence of (A) and (C) is a short computation one way and the course article's differential equations the other.

Lemma 1 (the volume of an ordered region). For $a < b$ and $k \ge 1$, the set of points $(u_1, \ldots, u_k)$ with $a < u_1 < u_2 < \cdots < u_k \le b$ has volume $(b - a)^k/k!$.

Proof. For $k = 1$ the set is the interval $(a, b]$. Suppose the claim holds for $k - 1$. Fix the last coordinate $u_k = u$. The remaining coordinates form an ordered set in $(a, u)$, of volume $(u - a)^{k-1}/(k-1)!$ by the hypothesis (the endpoint $u$ has no volume). Integrating over $u$ from $a$ to $b$ gives $(b - a)^k/k!$. $\square$

The cube $(a, b]^k$ splits into $k!$ such regions, one for each order of the coordinates, apart from ties of volume zero. Lemma 1 says the regions have equal volume, which is what symmetry suggests.

Lemma 2 (the arrival times of exponential gaps). If $X_1, \ldots, X_n$ are i.i.d. exponential with rate $\lambda$, the arrival times $(S_1, \ldots, S_n)$ have the joint density

$$g_n(s_1, \ldots, s_n) = \lambda^n e^{-\lambda s_n} \qquad \text{on } 0 < s_1 < s_2 < \cdots < s_n,$$

and zero elsewhere. In particular $S_n$ has the Erlang (gamma) density $\lambda^n s^{n-1}e^{-\lambda s}/(n-1)!$ for $s > 0$.

Proof. The gaps have the joint density $\prod_k \lambda e^{-\lambda x_k} = \lambda^n e^{-\lambda(x_1 + \cdots + x_n)}$ on $x_k > 0$, the product of their densities (part 2 of Theorem 6 of Random Variables and Distribution Functions, with $n$ integrals). For a region $B$ of vectors $s$, $P((S_1, \ldots, S_n) \in B)$ is the integral of that density over the $x$ whose partial sums lie in $B$. Change variables to the partial sums $s_k = x_1 + \cdots + x_k$. The map is linear, and its matrix has ones on and below the diagonal and zeros above, so its Jacobian determinant is 1 and volumes are unchanged (the change-of-variables formula for multiple integrals, taken as given here). The exponent becomes $-\lambda s_n$, and the conditions $x_k > 0$ become $s_{k-1} < s_k$, with $s_0 = 0$. So the probability is the integral of $\lambda^n e^{-\lambda s_n}$ over $B$ intersected with the ordered set. For the density of $S_n$ alone, integrate out $s_1, \ldots, s_{n-1}$ for fixed $s_n = s$: they range over an ordered set in $(0, s)$ of volume $s^{n-1}/(n-1)!$ by Lemma 1. $\square$

Durrett (2019, Theorem 2.1.18) obtains the Erlang density by repeated convolution instead.

Proof of Theorem 1. (B) gives (A). First, $N(t)$ is finite. The events $\{S_n \le t\}$ decrease in $n$, and by the Erlang density and $e^{-\lambda s} \le 1$,

$$P(S_n \le t) = \int_0^t \frac{\lambda^n s^{n-1}e^{-\lambda s}}{(n-1)!}\,ds \le \frac{(\lambda t)^n}{n!} \to 0,$$

so $P(N(t) = \infty) = 0$ by the continuity of probability (the lemma of Random Variables and Distribution Functions). Now fix $0 = t_0 < t_1 < \cdots < t_m$ and whole numbers $k_1, \ldots, k_m \ge 0$ with sum $K$, and write $\Delta_i = t_i - t_{i-1}$. The event that $N(t_{i-1}, t_i] = k_i$ for every $i$ says that the first $k_1$ arrivals fall in $(t_0, t_1]$, the next $k_2$ in $(t_1, t_2]$, and so on, and that arrival $K + 1$ comes after $t_m$. Because the arrival times are increasing, the first part says that $(S_1, \ldots, S_K)$ lies in the region $R$ of increasing vectors whose first $k_1$ coordinates lie in $(t_0, t_1]$, the next $k_2$ in $(t_1, t_2]$, and so on. By Lemma 2 with $n = K + 1$,

$$P\big(N(t_{i-1}, t_i] = k_i \text{ for all } i\big) = \int_R \int_{t_m}^\infty \lambda^{K+1}e^{-\lambda s_{K+1}}\,ds_{K+1}\,ds = \lambda^K e^{-\lambda t_m}\,\operatorname{vol}(R).$$

The coordinates in different intervals are automatically in order, so $R$ is a product of ordered regions, one in each interval. By Lemma 1 its volume is $\prod_i \Delta_i^{k_i}/k_i!$, and since $t_m = \Delta_1 + \cdots + \Delta_m$,

$$P\big(N(t_{i-1}, t_i] = k_i \text{ for all } i\big) = \prod_{i=1}^m \frac{e^{-\lambda\Delta_i}(\lambda\Delta_i)^{k_i}}{k_i!}.$$

(When $K = 0$ the event is $\{S_1 > t_m\}$, of probability $e^{-\lambda t_m}$, which agrees.) The joint probability mass function of the counts is a product of Poisson mass functions, so the counts are independent and Poisson with means $\lambda\Delta_i$ (part 1 of Theorem 6 of the same article). Counts over disjoint intervals that do not start at 0 are a sub-collection of such counts, so (A) holds.

(A) gives (B). Take any process satisfying (A), and times $t_1, \ldots, t_n \ge 0$. By the identity $\{S_k \le t_k\} = \{N(t_k) \ge k\}$,

$$P(S_1 \le t_1, \ldots, S_n \le t_n) = P\big(N(t_1) \ge 1, \ldots, N(t_n) \ge n\big).$$

The right side depends only on the joint distribution of $N(t_1), \ldots, N(t_n)$. Sort the times; each $N(t_k)$ is a sum of the independent Poisson increments between successive sorted times, so that joint distribution is fixed by (A) and is the same for every process satisfying (A). The process built by (B) is one such process. So the arrival times of the two processes have the same joint distribution function, hence the same joint distribution (a joint distribution function determines the distribution; this step is cited, as Random Variables and Distribution Functions explains). The gaps are differences of arrival times, so they too have the same joint distribution as in (B): i.i.d. exponential with rate $\lambda$.

(A) gives (C). Independence of disjoint counts is part of (A). An interval of length $h$ holds a Poisson count with mean $\lambda h$, so

$$P(\text{one}) - \lambda h = \lambda h\,(e^{-\lambda h} - 1), \qquad P(\text{two or more}) = 1 - e^{-\lambda h} - \lambda h e^{-\lambda h}.$$

The first, divided by $h$, is $\lambda(e^{-\lambda h} - 1) \to 0$. For the second, $g(x) = 1 - e^{-x}(1 + x)$ has $g(0) = 0$ and $g'(x) = xe^{-x} \le x$, so $0 \le g(\lambda h) \le (\lambda h)^2/2$, which is $o(h)$. Both functions depend on $h$ alone.

(C) gives (A). The Poisson Distribution proves, in its section "The Poisson process", that (C) makes $N(t)$ Poisson with mean $\lambda t$; its differential equations use only the independence of the counts in $[0, t]$ and $(t, t + h]$ and the two short-interval probabilities. For $s \ge 0$ the shifted process $N(s, s + u]$, $u \ge 0$, satisfies the same assumptions, since its intervals are intervals of the original process with the same lengths. So $N(s, s + u]$ is Poisson with mean $\lambda u$. The differential equations give only these marginal distributions; the mutual independence that (A) requires is the independence assumed in (C). $\square$

The theorem makes the model testable in three ways. Counts can be compared with the Poisson distribution (A), gaps with the exponential (B), and the placement of the times with the uniform distribution, by Theorem 3 below. A failure of any one refutes all three, since they are one hypothesis.

Waiting without memory

The gaps of a Poisson process are exponential. Is that a coincidence of the construction, or is the exponential forced by the idea of events that "do not remember" how long it has been? It is forced, under one mild condition.

A positive waiting time $T$ is memoryless if $P(T > s + t \mid T > s) = P(T > t)$ for all $s, t \ge 0$ with $P(T > s) > 0$: having waited $s$, the remaining wait has the original distribution. The Poisson Distribution shows that the exponential has this property.

Theorem 2 (the exponential is the only memoryless waiting time). Let $T$ be a random variable with $P(T > 0) = 1$. If $T$ is memoryless, then $P(T > t) = e^{-\lambda t}$ for some $\lambda > 0$ and all $t \ge 0$.

Here is the idea. Memorylessness turns the survival function into one that changes addition into multiplication, and the only such functions that are monotone are exponentials.

Proof. Write $G(t) = P(T > t)$. Then $G(s + t) = G(s)\,G(t)$ for all $s, t \ge 0$: when $G(s) > 0$ this is the definition multiplied out, and when $G(s) = 0$ both sides are 0, because $G(s + t) \le G(s)$. Applying the equation repeatedly, $G(mt) = G(t)^m$ for whole numbers $m$. With $t = 1/n$ and $m = n$ this gives $G(1/n) = G(1)^{1/n}$, and then $G(m/n) = G(1)^{m/n}$ for every positive rational $m/n$.

$G(1)$ cannot be 0. Otherwise $G(1/n) = 0$ for every $n$, and since $G = 1 - F$ is continuous from the right (Theorem 1 of Random Variables and Distribution Functions), $G(0) = \lim_n G(1/n) = 0$, against $P(T > 0) = 1$. $G(1)$ cannot be 1 either. Otherwise $G = 1$ at every positive rational, and since $G$ never increases, $G(t) = 1$ for every $t$, which contradicts $G(t) \to 0$ as $t \to \infty$. So $G(1) = e^{-\lambda}$ with $\lambda = -\ln G(1)$ in $(0, \infty)$. For any $t \ge 0$, take rationals $q_k$ decreasing to $t$; right continuity gives $G(t) = \lim_k e^{-\lambda q_k} = e^{-\lambda t}$. $\square$

Durrett (2019, Exercise 3.7.1) poses this result with the hint of the dyadic rationals. Its content for modeling is that the exponential gap is not an extra assumption. Any process in which the wait for the next event, from any moment, ignores the past has exponential gaps, and Theorem 1 then makes it a Poisson process provided the gaps are also independent. In the language of survival analysis, a memoryless lifetime is one whose hazard is constant (Censoring, Survival Functions and Hazards, the example after Theorem 1). Parametric Survival Models starts from the exponential and moves to families such as the Weibull, whose hazard changes with time.

Example (the gap that covers a moment is twice as long). "The time between the event before noon and the event after noon is a typical gap, with mean $1/\lambda$." It is not. For a Poisson process with rate $\lambda$ and a fixed time $t > 0$, let $A_t$ be the time back to the last event before $t$ (or back to 0 if there was none) and $R_t$ the time forward to the next event. The forward time $R_t$ exceeds $r$ exactly when $(t, t + r]$ holds no event, so $R_t$ is exponential with mean $1/\lambda$. For $0 \le a < t$, $A_t$ exceeds $a$ exactly when $(t - a, t]$ holds no event, which has probability $e^{-\lambda a}$, and $A_t \le t$ always. By the tail formula for the mean of a nonnegative variable (Expectation, Variance and Covariance),

$$E[A_t] + E[R_t] = \int_0^t e^{-\lambda a}\,da + \frac{1}{\lambda} = \frac{1 - e^{-\lambda t}}{\lambda} + \frac{1}{\lambda},$$

which tends to $2/\lambda$. A fixed moment is more likely to fall in a long gap than in a short one, so the gap that covers it is longer on average. For $\lambda = 1$ and $t = 10$ the formula gives 2.0000. In 30 simulations (seeds 64,000,000 to 64,000,029) of 20,000 processes each, the mean length of the gap that covers $t = 10$ had median 2.001 (interquartile range 1.994 to 2.006). A crawler that visits a page at a fixed time, as in Freshness and Recrawling, meets the page in such a covering gap: the version it fetches lives, on average, about twice as long as a typical version of the page.

Given the count, the times are uniform

Suppose we know that exactly $n$ events fell in $[0, t]$. Where did they fall? Under the Poisson model, anywhere, independently and uniformly. This is the property that makes tests of the model free of the unknown rate.

Theorem 3 (conditional uniformity). For a Poisson process with rate $\lambda$, $t > 0$ and $n \ge 1$, the conditional distribution of $(S_1, \ldots, S_n)$ given $N(t) = n$ has the density $n!/t^n$ on $0 < s_1 < \cdots < s_n < t$. That is the joint density of the order statistics $U_{(1)} < \cdots < U_{(n)}$ of $n$ i.i.d. uniform variables on $(0, t)$. It does not involve $\lambda$.

Proof. Let $B$ be a region of increasing vectors in $(0, t)^n$. The event $\{(S_1, \ldots, S_n) \in B,\ N(t) = n\}$ is $\{(S_1, \ldots, S_n) \in B,\ S_{n+1} > t\}$. By Theorem 1 the arrival times have the distribution of Lemma 2, so with $n + 1$ arrivals its probability is

$$\int_B \int_t^\infty \lambda^{n+1}e^{-\lambda s_{n+1}}\,ds_{n+1}\,ds = \int_B \lambda^n e^{-\lambda t}\,ds.$$

Dividing by $P(N(t) = n) = e^{-\lambda t}(\lambda t)^n/n!$ leaves $\int_B n!/t^n\,ds$. For the uniform variables, the event that their sorted values lie in $B$ is the union, over the $n!$ orders $\sigma$, of the events $(U_{\sigma(1)}, \ldots, U_{\sigma(n)}) \in B$. Apart from ties, which have probability zero, these events are disjoint, and each has probability $\int_B t^{-n}\,ds$ because the uniform variables have the joint density $t^{-n}$. So the sorted values lie in $B$ with probability $n!\int_B t^{-n}\,ds$, the same. $\square$

This is Durrett's (2019) Theorem 3.7.9. Two consequences follow at once.

Corollary 1 (counts given the total). Given $N(t) = n$, the count $N(s)$ for $0 < s < t$ has the binomial distribution with $n$ trials and success probability $s/t$. More generally, the counts in the cells of a partition of $(0, t]$ have the multinomial distribution with probabilities proportional to the cell lengths.

Proof. By Theorem 3 the times behave as the sorted values of $n$ i.i.d. uniform variables, and $N(s)$ is the number of those values at most $s$. Each uniform variable falls in $(0, s]$ with probability $s/t$, independently of the others. The same argument with several cells gives the multinomial. $\square$

The second consequence is a principle of testing. The total $N(t)$ is a sufficient statistic for $\lambda$ (the sense of Sufficient Statistics: given it, the data no longer depend on $\lambda$). So any statistic of the times, however complicated, has a null distribution given $N(t) = n$ that can be computed exactly, or simulated, without estimating the rate. A test of the gaps against an exponential whose rate is estimated from the same data would need its null distribution worked out afresh, since the estimate depends on the data; conditioning on the count avoids that. The earthquake tests below use this.

Splitting and merging

Give each event a type at random: a magnitude class, a kind of page change, a customer's sex. Are the events of one type again a Poisson process, and does what happens in one type say anything about another? Yes, and no.

Theorem 4 (thinning). Let $N$ be a Poisson process with rate $\lambda$. Assign each arrival, independently of the process and of the other arrivals, type $j$ with probability $p_j$, for $j = 1, \ldots, r$ with $p_1 + \cdots + p_r = 1$. Let $N_j(t)$ count the arrivals of type $j$ in $[0, t]$. Then $N_1, \ldots, N_r$ are independent Poisson processes with rates $\lambda p_1, \ldots, \lambda p_r$.

Here is the idea. Given $n$ arrivals in an interval, their types are multinomial. Averaging over a Poisson $n$ turns the multinomial into a product of independent Poisson counts, because the factorials and powers split exactly.

Proof. Fix $0 = t_0 < \cdots < t_m$ and whole numbers $k_{ij} \ge 0$, and let $n_i = \sum_j k_{ij}$. The arrivals in $(t_{i-1}, t_i]$ are those with indices $N(t_{i-1}) + 1$ to $N(t_i)$. Given the counts $N(t_{i-1}, t_i] = n_i$, the types of different intervals belong to disjoint sets of indices, so they are independent groups of i.i.d. types, and the type counts within interval $i$ are multinomial. Hence, with $\Delta_i = t_i - t_{i-1}$,

$$P\big(N_j(t_{i-1}, t_i] = k_{ij} \text{ for all } i, j\big) = \prod_{i=1}^m \frac{e^{-\lambda\Delta_i}(\lambda\Delta_i)^{n_i}}{n_i!}\cdot\frac{n_i!}{\prod_j k_{ij}!}\prod_j p_j^{k_{ij}} = \prod_{i=1}^m\prod_{j=1}^r \frac{e^{-\lambda p_j\Delta_i}(\lambda p_j\Delta_i)^{k_{ij}}}{k_{ij}!},$$

using $e^{-\lambda\Delta_i} = \prod_j e^{-\lambda p_j\Delta_i}$ and $(\lambda\Delta_i)^{n_i}\prod_j p_j^{k_{ij}} = \prod_j(\lambda p_j\Delta_i)^{k_{ij}}$. The joint mass function of all the type counts in all the intervals is a product of Poisson mass functions. So all these counts are mutually independent, and each is Poisson with mean $\lambda p_j\Delta_i$. Each $N_j$ therefore satisfies (A) with rate $\lambda p_j$, and the processes are independent: their counts over any finite set of disjoint intervals are, and this fixes the joint distribution of the processes by the extension theorem cited after Theorem 5. $\square$

This is Durrett's (2019) Theorem 3.7.4, which he introduces as "the somewhat surprising fact", for two reasons he gives: the parts are Poisson, and they are independent. Seeing many events of one type in an hour says nothing about the number of the other type in that hour, although both come from one stream. The script checked it on a simulated process with rate 2, built from exponential gaps over 1,000 time units and thinned with $p = 0.3$: the correlation of the two thinned counts per unit time had median $-0.002$ over 30 seeds (interquartile range $-0.030$ to 0.018). In the same simulations, counts in the disjoint intervals $(3k, 3k + 1]$ and $(3k + 1, 3k + 3]$ had means with medians 2.024 and 4.051 (against 2 and 4), the first had variance with median 2.023, and the correlation between them had median $-0.020$ (interquartile range $-0.051$ to 0.022).

Theorem 5 (superposition). Let $N_1, \ldots, N_r$ be independent Poisson processes with rates $\lambda_1, \ldots, \lambda_r$, and $N = N_1 + \cdots + N_r$, with $\lambda = \lambda_1 + \cdots + \lambda_r$. Then $N$ is a Poisson process with rate $\lambda$. With probability 1 no two of the processes have an event at the same time, and if each event of $N$ is labeled by the process it came from, the labels are i.i.d. with $P(\text{label} = j) = \lambda_j/\lambda$ and independent of $N$.

Proof. The count of $N$ in an interval is a sum of independent Poisson counts, hence Poisson with the sum of the means (The Poisson Distribution, "Sums of independent Poisson counts"). Counts of $N$ over disjoint intervals are functions of disjoint groups of the mutually independent counts of the $N_j$, so they are independent (the grouping fact of Random Variables and Distribution Functions). So $N$ satisfies (A). For coincidences, divide $[0, t]$ into $k$ equal pieces. Two simultaneous events put at least two events of $N$ in one piece, which has probability at most $k\,P(N(t/k) \ge 2)$. That is $k\,o(t/k) \to 0$ by (C), so coincidences in $[0, t]$ have probability zero, for every $t$.

For the labels, take a Poisson process $M$ with rate $\lambda$ and give its arrivals i.i.d. labels with probabilities $\lambda_j/\lambda$, independent of $M$. By Theorem 4 its labeled components are independent Poisson processes with rates $\lambda_j$, which is exactly the joint distribution of $(N_1, \ldots, N_r)$. The merged process and its sequence of labels are the same function of the components in both cases, so they have the same joint distribution in both cases. In the second case the labels are i.i.d. and independent of the merged process by construction, so in the first case too. $\square$

The proof's last step treats the components as one random object, the sequences of their arrival times, whose distribution is fixed by the joint distributions of finitely many of those times, and so by finitely many counts. That a distribution on sequences is fixed by its finite-dimensional distributions is Kolmogorov's extension theorem (Durrett, 2019, Theorem 2.1.21 and Appendix A.3), cited. A special case can be checked by hand: the first event of $N$ is the earliest of the first events of the $N_j$, and the minimum of independent exponential variables with rates $\lambda_j$ is exponential with rate $\lambda$, by Theorem 8 of Random Variables and Distribution Functions.

Thinning that depends on time

A rate that varies in time can be produced by thinning a constant one, which is also how such a process is simulated.

Theorem 6 (thinning with a probability that depends on time). Let $N$ be a Poisson process with rate $\lambda$, and keep an arrival at time $s$ with probability $p(s)$, where $p$ is a continuous function with values in $[0, 1]$, independently given the arrival times. Then the counts of kept arrivals in disjoint intervals are independent, and the count in $(a, b]$ is Poisson with mean $\int_a^b \lambda p(s)\,ds$.

Proof. Fix $t$ and a partition $0 = t_0 < \cdots < t_m = t$, and condition on $N(t) = n$. By Theorem 3 the arrival times are, as a set, $n$ i.i.d. uniform points on $(0, t)$, and each point is kept with probability $p$ at its own location, independently. So each point independently falls in one of $2m$ categories, "in $(t_{i-1}, t_i]$ and kept" with probability $q_i = \frac1t\int_{t_{i-1}}^{t_i}p(s)\,ds$ or "in $(t_{i-1}, t_i]$ and dropped" with probability $\Delta_i/t - q_i$. Given $N(t) = n$ the category counts are multinomial, and $N(t)$ is Poisson with mean $\lambda t$. The calculation in the proof of Theorem 4, with one interval of length $t$ and the $2m$ categories as types, shows that the category counts are independent Poisson variables with means $\lambda t$ times the category probabilities. The kept counts have means $\lambda t q_i = \int_{t_{i-1}}^{t_i}\lambda p(s)\,ds$. $\square$

Durrett (2019, Theorem 3.7.5) states this result. The counts of the kept arrivals are the counts of the inhomogeneous process of The Poisson Distribution, with intensity $\lambda(s) = \lambda p(s)$ and mean function $m(t) = \int_0^t \lambda(s)\,ds$. To simulate a process with intensity at most $\lambda_{\max}$, generate exponential gaps at rate $\lambda_{\max}$ and keep each arrival at $s$ with probability $\lambda(s)/\lambda_{\max}$.

Random amounts at Poisson times

Each event may carry an amount: the number of people in a car, the bytes of a page change, the earthquakes in a cluster. The total up to time $t$ is the compound Poisson variable

$$C(t) = Y_1 + Y_2 + \cdots + Y_{N(t)}, \qquad C(t) = 0 \text{ when } N(t) = 0,$$

where $Y_1, Y_2, \ldots$ are i.i.d. and independent of the process. Its mean and variance are a case of the random sums of Conditional Expectation (section "Random sums"), with a Poisson number of terms of mean $\lambda t$:

$$E[C(t)] = \lambda t\,E[Y], \qquad \operatorname{Var}(C(t)) = \lambda t\,E[Y^2],$$

for $Y$ with a finite variance. Durrett (2019, Theorem 3.7.3) proves the same formulas.

Proposition 1 (compound Poisson counts). Let $C(t)$ be as above.

  1. If the moment generating function $M_Y(\theta) = E[e^{\theta Y}]$ is finite, then $E[e^{\theta C(t)}] = \exp\{\lambda t\,(M_Y(\theta) - 1)\}$.
  2. Totals over disjoint intervals are independent, and the total over $(s, s + u]$ has the distribution of $C(u)$.
  3. If the $Y_k$ are whole numbers at least 1 with $E[Y^2] < \infty$, the index of dispersion of $C(t)$ is $\operatorname{Var}(C(t))/E[C(t)] = E[Y^2]/E[Y] \ge 1$, with equality only when every $Y_k = 1$.

Proof. (1) Given $N(t) = n$, the total is a sum of $n$ independent terms, and the mean of a product of independent variables is the product of their means, so the conditional moment generating function is $M_Y(\theta)^n$. By the law of total expectation (setting D of Conditional Expectation),

$$E[e^{\theta C(t)}] = \sum_{n=0}^\infty \frac{e^{-\lambda t}(\lambda t)^n}{n!}M_Y(\theta)^n = e^{-\lambda t}e^{\lambda t M_Y(\theta)}.$$

(2) The total over $(t_{i-1}, t_i]$ adds the amounts of the arrivals with indices $N(t_{i-1}) + 1$ to $N(t_i)$. Given the counts, these are disjoint groups of i.i.d. amounts. So for sets $B_i$, $P(\text{total}_i \in B_i \text{ for all } i \mid \text{counts } n_1, \ldots, n_m) = \prod_i P(Y_1 + \cdots + Y_{n_i} \in B_i)$. Averaging over the independent Poisson counts gives a product over $i$ of $\sum_{n}P(N(\Delta_i) = n)\,P(Y_1 + \cdots + Y_n \in B_i)$, and the $i$th factor is $P(C(\Delta_i) \in B_i)$.

(3) The two displayed formulas give the ratio. For a whole number $y \ge 1$, $y^2 - y = y(y - 1) \ge 0$, with equality only at $y = 1$. So $E[Y^2] - E[Y] \ge 0$, and it is 0 only if the nonnegative variable $Y(Y - 1)$ is 0 with probability 1 (the lemma of Probability Inequalities). $\square$

Part 3 is a mechanism for overdispersion different from the varying rate of Overdispersion in Count Data. The rate can be constant, but if events arrive in clusters and the clusters form a Poisson process, the counts vary more than their mean by the factor $E[Y^2]/E[Y]$, where $Y$ is the size of a cluster. The earthquakes show both mechanisms at once.

The earthquake catalog, tested

The Advanced National Seismic System Comprehensive Catalog (ComCat) of the U.S. Geological Survey lists earthquakes worldwide (U.S. Geological Survey, 2017). Its public query service returned 7,054 events of magnitude 6 or more from 1 January 1975 to 31 December 2024. Of them 6,992 are earthquakes; the other 62 are explosions, 61 of them nuclear, and are set aside. The period is 18,263 days. Of the earthquakes, 693 have magnitude 7 or more. The file kept for the scripts has each event's identifier, origin time, magnitude and location, and nothing else.

Left: the number of earthquakes in each calendar year, for magnitude 6 and over and for magnitude 7 and over, with the band where a Poisson count with the observed mean falls with probability 0.95 (log scale). Notice that the magnitude 6 counts leave their band in many years, and that the counts before 1983 sit below the rest. Right: the share of gaps between successive magnitude 7 earthquakes shorter than x days, against the exponential distribution with the same rate. Notice the step at the left: 7.7% of the gaps are under a day, where the exponential allows 3.7%.
Left: the number of earthquakes in each calendar year, for magnitude 6 and over and for magnitude 7 and over, with the band where a Poisson count with the observed mean falls with probability 0.95 (log scale). Notice that the magnitude 6 counts leave their band in many years, and that the counts before 1983 sit below the rest. Right: the share of gaps between successive magnitude 7 earthquakes shorter than x days, against the exponential distribution with the same rate. Notice the step at the left: 7.7% of the gaps are under a day, where the exponential allows 3.7%.

Each definition gives a test.

Counts (A). Under the model, the counts in the 50 calendar years are independent Poisson counts with a common mean, so Fisher's index of dispersion (defined in The Poisson Distribution) should be near its 49 degrees of freedom. Calendar years differ by a day, which is negligible here. For magnitude 6 and over the yearly counts have mean 139.84 and variance 866.2, from 83 (the fewest, in 1981) to 207 (2011). The index is 303.5 on 49 degrees of freedom, with an upper-tail p-value below $10^{-30}$. For magnitude 7 and over the mean is 13.86 and the variance 15.10, from 6 (the fewest, in 1980) to 24 (2010); the index is 53.4 on 49 degrees of freedom, $p = 0.31$. Over 600 equal windows of 30.44 days the ratio of variance to mean is 2.53 for magnitude 6 (mean 11.65 per window) and 1.13 for magnitude 7.

Gaps (B), tested given the count. Aftershocks follow large earthquakes within hours or days, so the model's weakest point should be an excess of short gaps. The statistic is the share of the gaps under a day. For magnitude 6 it is 0.388, against 0.318 for an exponential with the observed rate $n/T$. For magnitude 7, 53 of 692 gaps are under a day, a share of 0.077 against 0.037. Is that excess larger than chance allows? By Theorem 3, given the number of events the times are uniform, so the null distribution of the statistic can be simulated with no rate at all: place $n$ uniform times in the 18,263 days and compute the share. In a Monte Carlo of 30 seeds (61,000,000 to 61,000,029) with 2,000 sets of times each, the simulated shares averaged 0.318 for magnitude 6 and 0.037 for magnitude 7, and none of the 60,000 sets reached the observed share for either magnitude. The short-gap excess is real at both thresholds.

Uniformity given the count (Theorem 3). Given $n$ events in $T$ days, the times divided by $T$ should look like $n$ uniform values. The Kolmogorov–Smirnov statistic, the largest distance between their empirical distribution function and the uniform one (computed with SciPy's kstest), is 0.057 for magnitude 6, with a p-value below $10^{-15}$, and 0.061 for magnitude 7, with $p = 0.011$. The magnitude 6 catalog fails because the yearly rate drifts: the counts averaged 92.6 a year from 1975 to 1982 and 148.8 from 1983 to 2024, and the largest departure from the uniform distribution function lies at the fraction 0.194 of the period, in September 1984, after the low early years (1983 and 1984 were also slightly below the 50-year mean, so the deficit kept growing until then) (for the cluster starts below, at 0.174, September 1983). Why is not something this analysis can tell; it may lie in the catalog or in the Earth, and this article does not investigate it.

Clusters as a compound Poisson count. Proposition 1 suggests reading the short gaps as clusters. To see whether it accounts for the window variance, join two earthquakes when they occur within 30 days and 100 km of each other, and take the groups that the joins connect as clusters. The rule is a crude choice made for this article, not a published declustering method. Seismologists remove aftershocks before testing whether earthquakes are Poissonian; the title of Gardner and Knopoff (1974) asks whether the earthquakes of southern California, "with aftershocks removed", are Poissonian, and the abstract answers in one word: "Yes." The rule gives 5,617 clusters, of which a share of 0.876 hold a single earthquake; the largest holds 62. The mean cluster size is $E[Y] = 1.245$ and $E[Y^2] = 3.07$, so Proposition 1 predicts an index of dispersion of $E[Y^2]/E[Y] = 2.47$ for counts in windows long enough to hold whole clusters, if the clusters themselves arrived as a Poisson process. The observed ratio over the 30.44-day windows is 2.53. Clusters that straddle two windows and the drift of the yearly rate both disturb the comparison, so the agreement is an observation, not a test.

The cluster starts themselves are closer to a Poisson process but still not one. Their ratio of variance to mean over the 600 windows is 1.21, and their yearly index is 168.5 on 49 degrees of freedom. Their share of gaps under a day is 0.281 against 0.265 for the exponential, with a Monte Carlo p-value of median 0.0010 over the 30 seeds (interquartile range 0.0005 to 0.0015; seeds 66,000,000 to 66,000,029). The Kolmogorov–Smirnov statistic is 0.052. Removing the clusters removes most of the short-gap excess and lowers the ratio of variance to mean over the windows from 2.53 to 1.21, removing 86% of its excess over 1. It does not remove the drift.

What the tests say. No form of the model fits earthquakes of magnitude 6 and over: they cluster in time and space, and their yearly rate changes. For magnitude 7 and over the yearly counts pass the dispersion test, yet the gaps fail decisively and the uniformity test fails at the 5% level. For magnitude 7 the yearly counts pass because the clusters of large earthquakes are small: the ratio of variance to mean is 1.09 over the years and 1.13 over the windows, about the same in windows of either length, as Proposition 1 predicts for windows that hold whole clusters. With 50 yearly counts the dispersion test detects an excess that small rarely: if the index were 1.134 times (the unrounded window ratio) a chi-square variable with 49 degrees of freedom, it would exceed the 5% critical value with probability 0.17. Theorem 1 says that the three tests examine one hypothesis; passing one of them does not make the hypothesis true. A model that keeps the Poisson process for the cluster starts, lets the rate vary in time (Theorem 6) and attaches a random cluster to each start (Proposition 1) is the natural next model; this article does not fit it.

Estimating a change rate from visits

Freshness and Recrawling models the changes of a web page as a Poisson process with rate $\lambda$ and estimates $\lambda$ from $n$ visits at a fixed interval $I$, each of which sees only whether the page changed since the previous visit. If $X$ of the $n$ visits detect a change, its estimate is

$$\hat\lambda = -\frac{1}{I}\ln\Big(1 - \frac{X}{n}\Big).$$

That article treats the detections as independent Bernoulli trials with success probability $p = 1 - e^{-\lambda I}$. Definition (A) is what justifies that: the change counts in the $n$ disjoint intervals between visits are independent Poisson counts with mean $\lambda I$, and a detection is the event that a count is at least 1. How precise is the estimate, and which interval makes it most precise for a given number of visits?

Proposition 2 (information in a "changed or not" visit). With $r = \lambda I$, the expected number of changes between visits:

  1. The Fisher information about $\lambda$ in one visit is $I^2e^{-\lambda I}/(1 - e^{-\lambda I})$.
  2. $\hat\lambda$ is the maximum likelihood estimate, and for large $n$ it is approximately normal with mean $\lambda$ and variance $(e^{\lambda I} - 1)/(nI^2)$; its relative variance is $\operatorname{Var}(\hat\lambda)/\lambda^2 \approx f(r)/n$ with $f(r) = (e^r - 1)/r^2$.
  3. $f$ has a single minimum, at the root $r^* = 1.594$ of $r = 2(1 - e^{-r})$, where $f(r^*) = 1.544$.
  4. If $X = n$ has positive probability, which it has for every $\lambda > 0$, then $\hat\lambda = \infty$ with positive probability and $E[\hat\lambda] = \infty$.

Proof. (1) A Bernoulli trial with success probability $p$ carries information $1/(p(1 - p))$ about $p$. With $p = 1 - e^{-\lambda I}$, $dp/d\lambda = Ie^{-\lambda I}$, and the reparametrization rule of Fisher Information and the Cramér–Rao Bound gives

$$\frac{(dp/d\lambda)^2}{p(1 - p)} = \frac{I^2e^{-2\lambda I}}{(1 - e^{-\lambda I})\,e^{-\lambda I}} = \frac{I^2e^{-\lambda I}}{1 - e^{-\lambda I}}.$$

(2) The maximum likelihood estimate of $p$ is $X/n$, and $\lambda$ is a monotone function of $p$, so by invariance (Maximum Likelihood Estimation) $\hat\lambda$ is the maximum likelihood estimate of $\lambda$. The normal approximation with the inverse of $n$ times the information as variance is Asymptotics of the Maximum Likelihood Estimator, Theorem 2 (equivalently, the delta method applied to $X/n$). The inverse information is $(1 - e^{-\lambda I})/(I^2e^{-\lambda I}) = (e^{\lambda I} - 1)/I^2$, and dividing by $\lambda^2 = r^2/I^2$ gives $f(r)/n$.

(3) $f'(r) = h(r)/r^3$ with $h(r) = re^r - 2e^r + 2$. Then $h(0) = 0$ and $h'(r) = (r - 1)e^r$, so $h$ falls on $(0, 1)$, is negative there, and rises from $h(1) < 0$ to infinity on $(1, \infty)$. So $h$ has exactly one positive root $r^*$, with $f$ decreasing before it and increasing after. Setting $h(r) = 0$ and dividing by $e^r$ gives $r = 2(1 - e^{-r})$, which the script solves.

(4) $P(X = n) = p^n > 0$, and on that event the logarithm is of zero. $\square$

The optimum says to visit about once per 1.6 expected changes. That is less often than once per expected change ($r = 1$), the rate of the course article's running example: a visit wasted on an unchanged page and a visit that misses several changes both carry little information, and $r^*$ balances the two. Daily visits are close to optimal for a page that changes twice a day, $f(2) = 1.597$, and a little worse for one that changes once a day, $f(1) = 1.718$, at $r = 1$. If the crawler could count every change instead of seeing a yes or no, each visit would see a Poisson count with mean $m = \lambda I$, which carries information $(dm/d\lambda)^2/m = I^2/(\lambda I) = I/\lambda$ about $\lambda$, growing without limit with the interval. The yes-or-no record is what caps it.

Example (the course's 12 detections in 30 visits). Freshness and Recrawling (Problem 4) finds $\hat\lambda = -\ln 0.6 = 0.511$ changes a day from 12 detections in 30 daily visits. Here $e^{\hat\lambda} - 1 = 1/0.6 - 1 = 2/3$, so the estimated standard error is $\sqrt{(2/3)/30} = 0.149$, and the large-sample 95% interval is $0.511 \pm 1.96 \times 0.149 = (0.219, 0.803)$. An interval built on the scale of $p$ and mapped through $\lambda = -\ln(1 - p)$, which stays inside $(0, \infty)$ when the interval for $p$ lies inside $(0, 1)$, as it does here (otherwise clip the lower end at 0; if the upper end reaches 1, the interval for $\lambda$ is unbounded above), is $(0.254, 0.856)$. To estimate a rate near 0.5 a day most precisely per visit, the interval should be $r^*/0.5 = 3.19$ days.

Example (when every visit sees a change). For a page with $\lambda = 2$ visited daily for 30 days, every one of the 30 visits detects a change with probability $(1 - e^{-2})^{30} = 0.0127$, and then $\hat\lambda$ is infinite. For $\lambda = 0.5$ the same probability is about $7.0 \times 10^{-13}$, but it is still positive, so $\hat\lambda$ has no finite mean for any rate. The form with 0.5 added inside the logarithm, $-\frac1I\ln\frac{n - X + 0.5}{n + 0.5}$, which the course article attributes to Cho and Garcia-Molina (2003), is always finite. The script simulated 20,000 pages per seed for 30 seeds (seeds 65,000,000 + s for $\lambda = 0.5$, 65,000,100 + s for $\lambda = 2$), with 30 daily visits each:

$\lambda = 0.5$ $\lambda = 2$
Pages with every visit a detection 0.0000 0.0128 (IQR 0.0123 to 0.0133)
Relative bias of $\hat\lambda$, finite cases 2.2% (2.1 to 2.4) 5.1% (5.0 to 5.2)
Relative bias with 0.5 added 0.0% ($-0.1$ to 0.1) 0.1% (0.0 to 0.2)
95% Wald intervals that cover $\lambda$ 0.942 (0.941 to 0.944) 0.945 (0.944 to 0.946)

Entries are medians over the 30 seeds with interquartile ranges. The Wald coverage counts a page whose estimate is infinite as a miss. In this simulation, adding 0.5 removed nearly all of the bias at both rates.

What is proved here, what is cited, and what the experiments show

Proved on this page: the equivalence of the three definitions (Theorem 1, with Lemmas 1 and 2, given the course article's differential equations for (C) to (A)); the exponential as the only memoryless waiting time (Theorem 2); the mean length of the gap that covers a fixed time; conditional uniformity and the binomial and multinomial counts given the total (Theorem 3, Corollary 1); thinning, superposition and thinning that depends on time (Theorems 4 to 6); the generating function, independent increments and index of dispersion of a compound Poisson count (Proposition 1); the information in a yes-or-no visit, the best visit interval and the infinite mean of the maximum likelihood estimate (Proposition 2).

Cited: that a joint distribution function determines a joint distribution, the grouping fact for independent variables and the facts about multiple integrals, as Random Variables and Distribution Functions cites them, and the change-of-variables formula for multiple integrals (used in Lemma 2); Kolmogorov's extension theorem, used in the last step of Theorem 5 (Durrett, 2019); the asymptotic normality of the maximum likelihood estimate, proved in Asymptotics of the Maximum Likelihood Estimator. The measure-theoretic stance is that of the Probability Theory unit, set out in Random Variables and Distribution Functions (section "What measure theory would add").

What the experiments show: on simulated processes, the counts, the thinned counts and the covering gap behave as proved. On the earthquake catalog, every form of the Poisson hypothesis fails for magnitude 6 and over, and the gap and uniformity tests fail for magnitude 7 and over while the yearly counts do not. The compound Poisson reading of the clusters reproduces the window variance to within a few percent. These are observations about one catalog, one clustering rule and one period.

Worked problems

The four problems are posed on this article's earthquake catalog and on the change-rate example of the crawling course; they have no course assignment behind them. The script tools/experiments/statistics/poisson-processes.py checks every number.

Problem 1 (where the events of a year fall). Suppose the earthquakes of magnitude 7 and over form a Poisson process, and a year has 14 of them. Find the probability that none fell in the first quarter of the year, and the expected number in that quarter.

Solution. By Corollary 1, given 14 events in the year, the number in the first quarter is binomial with 14 trials and success probability 1/4, whatever the rate. So $P(\text{none}) = (3/4)^{14} = 0.0178$ and the expected number is $14/4 = 3.5$. The rate of 13.86 a year plays no part: conditioning on the count removes it. That is Theorem 3 at work, and it is why the uniformity test needs no estimate of the rate.

Problem 2 (thinning by magnitude). Of the 6,992 earthquakes of magnitude 6 and over, 693 reached magnitude 7. Treat the magnitude 6 earthquakes as a Poisson process with rate 139.84 a year and their magnitudes as independent marks, each at least 7 with probability $q = 693/6{,}992$. Find (a) the rates of the two processes of magnitude 7 and over and of magnitude 6 to 7, (b) the probability that the next three earthquakes of magnitude 6 and over are all below 7, (c) the probability that a month (one twelfth of a year) passes with no earthquake of magnitude 7 or more. Say what the catalog says about the assumption.

Solution. (a) $q = 0.0991$. By Theorem 4 the two processes are independent Poisson processes with rates $139.84\,q = 13.86$ and $139.84\,(1 - q) = 125.98$ a year. (b) The marks of the next three events are independent, so the probability is $(1 - q)^3 = 0.7312$. (c) The count of magnitude 7 events in a twelfth of a year is Poisson with mean $13.86/12$, so the probability of none is $e^{-13.86/12} = 0.3151$. The catalog contradicts the assumptions first in the parent process: Theorem 4 needs a Poisson stream of magnitude 6 earthquakes, and the catalog's is clustered, so even independent marks would give clustered magnitude 7 events. To separate the two, the script shuffled the magnitudes over the observed magnitude 6 times (marks independent by construction; 30 seeds, 67,000,000 to 67,000,029, 200 shuffles each). The magnitude 7 share of gaps under a day was then 0.060 on average, against 0.037 for a Poisson stream, so most of the excess comes from the clustered parent. A shuffled share reached the observed 0.077 in a share of shuffles with median 0.038 (interquartile range 0.030 to 0.049), a Monte Carlo p-value near 0.04: some evidence, not strong, that large earthquakes follow one another more often than independent marks allow. Thinning describes a stream whose marks ignore one another, and these may not quite.

Problem 3 (clusters as a compound Poisson count). The clustering rule gave 5,617 clusters in 18,263 days, with $E[Y] = 1.245$ and $E[Y^2] = 3.07$ earthquakes per cluster. If the cluster starts formed a Poisson process and each cluster fell inside one window, find the mean and variance of the number of earthquakes in a window of $18{,}263/600 = 30.44$ days, and compare them with the 600 windows of the catalog.

Solution. Clusters start at $5{,}617/18{,}263$ per day, so a window holds a Poisson number of clusters with mean $5{,}617 \times 30.44/18{,}263 = 9.362$. By the compound Poisson formulas, the number of earthquakes has mean $9.362 \times 1.245 = 11.65$ and variance $9.362 \times 3.07 = 28.73$. The mean agrees exactly with the catalog's 11.65 a window, because the cluster sizes add up to the 6,992 earthquakes. The catalog's variance is 29.52. The compound Poisson model accounts for almost all of the excess over the plain Poisson variance of 11.65. The remainder is consistent with the drift of the yearly rate, though this calculation does not show that.

Problem 4 (a change rate and how often to look). A page is visited daily for 30 days and a change is detected on 12 visits. Estimate its change rate with a standard error and a 95% interval, and find the visit interval that would estimate a rate of this size most precisely for the same number of visits.

Solution. By Proposition 2, $\hat\lambda = -\ln(1 - 12/30) = -\ln 0.6 = 0.511$ changes a day. The estimated variance is $(e^{\hat\lambda} - 1)/n = (2/3)/30$, a standard error of 0.149, and the Wald interval is $0.511 \pm 1.96 \times 0.149 = (0.219, 0.803)$. The interval mapped from the scale of $p$, $X/n \pm 1.96\sqrt{(X/n)(1 - X/n)/n}$ transformed by $-\ln(1 - p)$, is $(0.254, 0.856)$; it is shifted to the right because the transformation is convex. The relative variance is smallest at $r^* = 1.594$ expected changes per interval, so for a rate near 0.5 a day the best interval is $1.594/0.5 = 3.19$ days. Thirty visits spread over about 96 days would then estimate the rate more precisely than thirty daily visits: $f(0.511) = 2.555$ against $f(r^*) = 1.544$, a variance smaller by the factor 1.65. Daily visits waste precision here because most of them find nothing new.

Summary

Result Statement
Three definitions independent Poisson increments; i.i.d. exponential gaps; short-interval probabilities $\lambda h + o(h)$ and $o(h)$ (Theorem 1)
Arrival times joint density $\lambda^n e^{-\lambda s_n}$ on $0 < s_1 < \cdots < s_n$; $S_n$ is Erlang
Memoryless only the exponential (Theorem 2)
Covering gap mean $(1 - e^{-\lambda t})/\lambda + 1/\lambda \to 2/\lambda$
Given $N(t) = n$ times are uniform order statistics; $N(s) \sim$ binomial$(n, s/t)$ (Theorem 3)
Thinning, superposition independent Poisson processes with rates $\lambda p_j$; merged rates add, labels i.i.d. (Theorems 4 to 6)
Compound Poisson mean $\lambda tE[Y]$, variance $\lambda tE[Y^2]$, generating function $e^{\lambda t(M_Y(\theta) - 1)}$, dispersion $E[Y^2]/E[Y]$
Yes-or-no visits information $I^2e^{-\lambda I}/(1 - e^{-\lambda I})$ per visit; best $r = \lambda I = 1.594$

What we have learned

The Poisson process is one hypothesis with three faces: independent Poisson counts, independent exponential gaps, and a constant small chance of one event in every short interval. Its gaps are exponential because waiting without memory forces them to be. Given their number, its times are uniform, which turns every test of the model into a test free of the rate. It splits into independent Poisson processes and merges back into one. The earthquake catalog fails every face of the hypothesis, through clusters that a compound Poisson count describes well and a yearly rate that drifts. A page's change rate, seen only through yes-or-no visits, is estimated best by visiting about once per 1.6 expected changes.

The arrival times $S_n$ of the process form a random walk with positive steps. The next article, Random Walks and Martingales, studies walks whose steps can go either way, the fair games called martingales, and what they say about stopping a statistical test whenever the data look good.

References

  • Cho, J., and Garcia-Molina, H. (2003). Estimating frequency of change. ACM Transactions on Internet Technology, 3(3), 256–290. https://doi.org/10.1145/857166.857170 record · 2 pages
  • Durrett, R. (2019). Probability: Theory and Examples (5th ed.). Cambridge University Press. Read in the author's copy: Section 3.7 (Theorem 3.7.2, the Poisson count from conditions on short intervals; the construction from exponential gaps; Theorem 3.7.3, compound Poisson; Theorem 3.7.4, thinning; Theorem 3.7.5, thinning with a probability depending on time; Theorem 3.7.9 and Corollary 3.7.10, conditioning on the count; Exercise 3.7.1, the memoryless property), Theorem 2.1.18 (sums of exponentials are gamma), Theorem 2.1.21 and Appendix A.3 (Kolmogorov's extension theorem). record · 15 pages
  • Gardner, J. K., and Knopoff, L. (1974). Is the sequence of earthquakes in Southern California, with aftershocks removed, Poissonian? Bulletin of the Seismological Society of America, 64(5), 1363–1367. https://doi.org/10.1785/BSSA0640051363 record
  • U.S. Geological Survey (2017). Advanced National Seismic System (ANSS) Comprehensive Catalog. U.S. Geological Survey. https://doi.org/10.5066/F7MS3QZH. Queried on 19 November 2026 for events of magnitude 6 and over from 1975 to 2024. record