Interpreting the Poisson Rate

What a Poisson Rate Measures

A rate describes how quickly events accumulate on a chosen time scale. If a process has rate \(6\) events per hour, its rate is not the probability of an event in an hour; it is the expected number of events accumulated per hour. The same physical process can be described by different numerical rates when the time unit changes. Correct interpretation therefore requires keeping track of both the number and the time units.

Work on a probability space \((\Omega,\mathcal{F},\mathbb{P})\). Let \((N_t)_{t\ge0}\) be a counting process, where \(N_t\) is the number of events observed by time \(t\), and take \(N_0=0\). Its count on the interval \((a,b]\), for \(0\le a\le b\), is \(N_b-N_a\). We use this endpoint convention throughout. Assume the counts being discussed have finite expectations.

For example, if time is measured in hours and \(\lambda=6\) events per hour, an interval lasting \(2\) hours has expected count \(12\). The observed count is an integer, but its expectation need not be: an expected count of \(\tfrac32\) is a valid average, not a prediction that a particular interval contains a fractional event. Also, a rate is not a probability. In particular, a rate of \(6\) events per hour is not a probability of \(6\) for an event during an hour.

Why Expected Counts Scale with Duration

The linear relationship between expected count and duration is consistent with two structural properties of homogeneous counting: increments over intervals of equal length have the same distribution, and counts are additive across adjacent intervals. The following argument shows that these properties, together with finite expected counts, force the mean count to be linear in time. Independence is not needed for this mean calculation.

Proof. Define \(g(t)=\mathbb{E}[N_t]\), which is finite by assumption. For \(s,t\ge0\), pathwise additivity gives \(N_{s+t}=N_s+(N_{s+t}-N_s)\). Taking expectations and using stationary increments, the increment \(N_{s+t}-N_s\) has the same distribution as \(N_t-N_0=N_t\). Therefore \(g(s+t)=g(s)+g(t)\). Since counts are nondecreasing in time, \(g\) is nondecreasing. Set \(c=g(1)\), so \(c\ge0\). Additivity gives \(g(n)=nc\) for every nonnegative integer \(n\), and \(n g(1/n)=g(1)=c\), so \(g(1/n)=c/n\) for positive integers \(n\). It follows by additivity that \(g(r)=cr\) for every nonnegative rational \(r\). For any real \(t\ge0\), choose rational sequences \(r_m\uparrow t\) and \(q_m\downarrow t\). Monotonicity gives \(c r_m=g(r_m)\le g(t)\le g(q_m)=c q_m\). Taking limits yields \(g(t)=ct\), including \(t=0\). Finally, \(N_b=N_a+(N_b-N_a)\), and stationary increments give \(\mathbb{E}[N_b-N_a]=g(b-a)=c(b-a)\). \(\square\)

For a homogeneous Poisson process, the parameter \(\lambda\) is this mean rate \(c\). Hence, in a fixed time unit, the working formula is

The duration \(b-a\) and the units attached to \(\lambda\) must use the same time unit. If the rate is given per hour, the duration in the formula must be in hours. The formula applies to any interval length, not just intervals starting at zero. When \(a=b\), the interval has zero duration and the expected count is zero.

Changing the Time Unit

A unit conversion changes the numerical value of a rate but not the underlying process or its expected count. Suppose one new time unit contains \(c\) old time units. A duration of \(u\) new units is \(cu\) old units. If the old rate is \(\lambda_{\mathrm{old}}\) events per old unit, the expected count over that duration is \(\lambda_{\mathrm{old}}cu\). To write the same expectation as a rate per new unit times \(u\), the new rate must be \(c\lambda_{\mathrm{old}}\).

Proof. Let \(u\ge0\) be an interval duration measured in new units. It is a duration of \(cu\) old units. By the expected-count formula in old units, its expected count is \(\lambda_{\mathrm{old}}cu\). By the same formula in new units, that expected count is \(\lambda_{\mathrm{new}}u\). These expressions must agree for every \(u\ge0\). Taking \(u=1\) gives \(\lambda_{\mathrm{new}}=c\lambda_{\mathrm{old}}\). Conversely, substituting this rate gives \(\lambda_{\mathrm{new}}u=c\lambda_{\mathrm{old}}u\), which is exactly the old-unit expectation for \(cu\) old units. The conversion therefore preserves the expected count. \(\square\)

For instance, a rate per minute converts to a rate per hour by multiplying by \(60\), because one hour contains \(60\) minutes. A rate per hour converts to a rate per minute by dividing by \(60\). These are inverse conversions, not competing rules.

Worked Rate Calculations

Common Unit Errors and What the Rate Does Not Say

Common mistake: For a rate of \(18\) events per hour, someone might calculate the expected count in \(25\) minutes as \(18\cdot25=450\). That multiplication pairs an hourly rate with a duration measured in minutes, so its units do not cancel to give a count. The correction is either to convert the rate to \(18/60=3/10\) events per minute and calculate \((3/10)\cdot25=15/2\), or to convert the duration to \(25/60=5/12\) hour and calculate \(18\cdot(5/12)=15/2\).

A second error is to treat \(\lambda h\) as the probability of at least one event. The formula established here gives an expected count. It does not, by itself, give the probability of zero events or of one or more events. Those are questions about the count distribution, not just its mean. Similarly, a mean rate describes average accumulation and does not assert that every interval of a given duration will contain exactly its expected count.

When applying the formula, first identify the interval \((a,b]\), then calculate its elapsed duration \(b-a\), and finally express that duration in the rate’s time unit. A rate of zero gives expected count zero for every finite interval. For a positive rate and zero-length interval, the expected count is still zero. For positive duration, the expected count can be nonintegral; that is consistent with averaging integer-valued counts across outcomes.