Skip to content

Lesson 22: The Poisson Process and the Exponential Distribution

22.1 Learning Outcomes

By the end of this hour a student will be able to:

  1. State the axioms of a Poisson process and derive its probability mass function.
  2. Compute Poisson probabilities and state the distribution's mean and variance.
  3. Derive the exponential inter-arrival distribution from the Poisson process.
  4. State and prove the memoryless property and explain its consequences.
  5. Test whether real data are consistent with a Poisson process.

22.2 Why These Two Distributions?

The whole of classical queuing theory rests on them.

\[ \boxed{\ \textbf{Poisson counts} \iff \textbf{Exponential times}\ } \]

These are two descriptions of the same process. If arrivals occur as a Poisson process with rate \(\lambda\), then the number in any interval is Poisson and the gaps between arrivals are exponential.

The "M" in \(M/M/1\) stands for Markovian — precisely this pair.

22.3 The Poisson Process

The three axioms

A counting process \(\{N(t),\ t\ge0\}\) is a Poisson process with rate \(\lambda\) if:

(A1) Independence. The numbers of events in disjoint intervals are independent.

(A2) Stationarity. The distribution of the number of events depends only on the length of the interval, not its position.

(A3) Orderliness. For a small interval \(\Delta t\):

\[ \Pr[\text{1 event in } \Delta t] = \lambda\,\Delta t + o(\Delta t) \]
\[ \Pr[\ge2 \text{ events in } \Delta t] = o(\Delta t) \]
\[ \Pr[\text{0 events in } \Delta t] = 1 - \lambda\,\Delta t + o(\Delta t) \]

In words: events occur independently, at a constant average rate, and one at a time.

Deriving the distribution

Let \(P_n(t) = \Pr[N(t) = n]\).

For \(n = 0\). The system is empty at \(t+\Delta t\) only if it was empty at \(t\) and no arrival came:

\[ P_0(t+\Delta t) = P_0(t)\,(1 - \lambda\Delta t) \]
\[ \frac{P_0(t+\Delta t) - P_0(t)}{\Delta t} = -\lambda P_0(t) \]

Letting \(\Delta t\to0\):

\[ \frac{dP_0}{dt} = -\lambda P_0(t), \qquad P_0(0) = 1 \]
\[ \boxed{\ P_0(t) = e^{-\lambda t}\ } \]

For \(n \ge 1\). Either there were \(n\) at time \(t\) and none arrived, or \(n-1\) and one arrived:

\[ P_n(t+\Delta t) = P_n(t)(1-\lambda\Delta t) + P_{n-1}(t)\,\lambda\Delta t \]
\[ \frac{dP_n}{dt} = -\lambda P_n(t) + \lambda P_{n-1}(t) \]

Solving recursively (by induction, using an integrating factor \(e^{\lambda t}\)):

\[ \boxed{\ P_n(t) \ =\ \frac{(\lambda t)^{n}e^{-\lambda t}}{n!}, \qquad n = 0,1,2,\ldots\ } \]

This is the Poisson distribution with parameter \(\lambda t\).

Properties

\[ \boxed{\ E[N(t)] = \lambda t, \qquad \mathrm{Var}[N(t)] = \lambda t\ } \]
\[ \boxed{\ \textbf{MEAN} = \textbf{VARIANCE} \ \textbf{— the signature of a Poisson process.}\ } \]

This is the standard diagnostic. If sample data have variance far from the mean, the process is not Poisson.

Variance vs mean Diagnosis
\(\sigma^2 \approx \mu\) consistent with Poisson
\(\sigma^2 > \mu\) (overdispersed) clustered/batch arrivals; a time-varying rate
\(\sigma^2 < \mu\) (underdispersed) more regular than random; appointments, scheduling

Superposition and thinning

Two further properties, both practically important:

Superposition. The sum of independent Poisson processes with rates \(\lambda_1,\lambda_2\) is Poisson with rate \(\lambda_1+\lambda_2\).

Thinning. If each event of a Poisson process with rate \(\lambda\) is independently retained with probability \(p\), the retained events form a Poisson process with rate \(\lambda p\).

\[ \boxed{\ \textbf{This is why merged customer streams stay Poisson, and why a fraction of traffic is still Poisson.}\ } \]

22.4 The Exponential Distribution

Deriving it from the Poisson process

Let \(T\) be the time until the next arrival. Then

\[ \Pr[T > t] = \Pr[\text{no arrivals in } (0,t]] = P_0(t) = e^{-\lambda t} \]

So the cumulative distribution function is

\[ \boxed{\ F(t) = \Pr[T\le t] = 1 - e^{-\lambda t}, \qquad t \ge 0\ } \]

and differentiating gives the density

\[ \boxed{\ f(t) = \lambda e^{-\lambda t}, \qquad t\ge0\ } \]

Properties

\[ E[T] = \frac{1}{\lambda}, \qquad \mathrm{Var}[T] = \frac{1}{\lambda^{2}}, \qquad \sigma_T = \frac{1}{\lambda} \]
\[ \boxed{\ \textbf{MEAN} = \textbf{STANDARD DEVIATION} \ \text{— so the coefficient of variation is exactly 1.}\ } \]

This is a high-variability distribution. A "10-minute average" service time that is exponential has a standard deviation of 10 minutes: some customers take 30 seconds, others 40 minutes.

The median is \(\frac{\ln 2}{\lambda} = 0.693/\lambda\) — well below the mean, so the distribution is strongly right-skewed. Most service times are shorter than average; a few are far longer.

22.5 The Memoryless Property

Theorem. The exponential is the only continuous distribution with

\[ \boxed{\ \Pr[T > s+t \mid T > s] \ =\ \Pr[T > t] \qquad \text{for all } s,t \ge 0\ } \]

Proof.

\[ \Pr[T>s+t \mid T>s] = \frac{\Pr[T>s+t \ \cap\ T>s]}{\Pr[T>s]} = \frac{\Pr[T>s+t]}{\Pr[T>s]} = \frac{e^{-\lambda(s+t)}}{e^{-\lambda s}} = e^{-\lambda t} = \Pr[T>t] \]

\(\blacksquare\)

What it means

The process has no memory of elapsed time. If a customer has already been in service for 10 minutes, the distribution of the remaining service time is exactly the same as at the start.

Concretely:

  • "This bulb has burned 1000 hours" tells you nothing about its remaining life.
  • "The last bus left 20 minutes ago" tells you nothing about when the next will come.
  • "The machine has run 5 hours since the last breakdown" gives no information about the next one.

Why it is both essential and questionable

Why it is essential. Memorylessness is what makes queuing models Markovian — the future depends only on the current number in the system, not on how long anyone has been waiting. Without it, we would need to track every customer's elapsed time, and the state space would be infinite and continuous.

\[ \boxed{\ \textbf{Memorylessness is what makes the mathematics tractable.}\ } \]

Why it is questionable. Real service times often are not memoryless:

  • A haircut takes 20 ± 3 minutes — nearly deterministic.
  • A machine that has run 5 hours is more likely to fail soon (wear-out).
  • A customer already served 10 minutes is often nearly finished.

When the exponential assumption is reasonable.

Reasonable Questionable
Arrivals from a large independent population Scheduled arrivals (appointments)
Telephone call arrivals Batch arrivals (a tour bus)
Radioactive decay, equipment failures in the useful-life phase Wear-out failures
Highly variable service (hospital A&E, IT helpdesk) Standardised service (car wash, assembly)
\[ \boxed{\ \textbf{The exponential assumption is usually far better for ARRIVALS than for SERVICE times.}\ } \]

The practical consequence of getting it wrong

For \(M/G/1\) (general service), the Pollaczek–Khinchine formula gives

\[ L_q = \frac{\lambda^{2}\sigma_s^{2} + \rho^{2}}{2(1-\rho)} \]

where \(\sigma_s^2\) is the variance of service time. Compare:

Service distribution \(\sigma_s^2\) \(L_q\) at \(\rho = 0.8\), \(\mu = 1\)
Deterministic (D) 0 \(\dfrac{0+0.64}{0.4} = \mathbf{1.60}\)
Exponential (M) \(1/\mu^2 = 1\) \(\dfrac{0.64+0.64}{0.4} = \mathbf{3.20}\)
High variability \(4/\mu^2\) \(\dfrac{2.56+0.64}{0.4} = \mathbf{8.00}\)
\[ \boxed{\ \textbf{Exponential service gives EXACTLY TWICE the queue of deterministic service.}\ } \]

So \(M/M/1\) results are a conservative (pessimistic) estimate when real service is more regular — which is usually the case. This is a useful thing to tell a client.

22.6 Solved Problems

Problem 1 — Poisson probability calculations

Customers arrive at a bank at a Poisson rate of 12 per hour. Find the probability that in a 15-minute period there are (a) exactly 2 arrivals, (b) no arrivals, (c) at least 3 arrivals, (d) at most 1 arrival.

Solution.

Set up. \(\lambda = 12\)/hour and \(t = 15\) minutes \(= 0.25\) hours, so

\[ \lambda t = 12\times0.25 = \mathbf{3} \]

(a) Exactly 2.

\[ P_2 = \frac{3^{2}e^{-3}}{2!} = \frac{9(0.049787)}{2} = \frac{0.448084}{2} = \mathbf{0.2240} \]

(b) No arrivals.

\[ P_0 = \frac{3^{0}e^{-3}}{0!} = e^{-3} = \mathbf{0.0498} \]

(c) At least 3.

\[ \Pr[N\ge3] = 1 - P_0 - P_1 - P_2 \]
\[ P_1 = \frac{3^{1}e^{-3}}{1!} = 3(0.049787) = 0.1494 \]
\[ \Pr[N\ge3] = 1 - 0.0498 - 0.1494 - 0.2240 = \mathbf{0.5768} \]

(d) At most 1.

\[ \Pr[N\le1] = P_0 + P_1 = 0.0498 + 0.1494 = \mathbf{0.1992} \]

Summary table.

\(n\) \(P_n\) Cumulative
0 0.0498 0.0498
1 0.1494 0.1992
2 0.2240 0.4232
3 0.2240 0.6472
4 0.1680 0.8153
5 0.1008 0.9161

Two observations.

  1. \(P_2 = P_3 = 0.2240\). For a Poisson with integer mean \(\lambda t = 3\), the distribution has two modes, at \(n = \lambda t - 1\) and \(n = \lambda t\). (In general the mode is \(\lfloor \lambda t\rfloor\).)

  2. Check the mean and variance.

    \[ E[N] = \lambda t = 3, \qquad \mathrm{Var}[N] = 3, \qquad \sigma = \sqrt3 = 1.732 \]

    So arrivals in a 15-minute window are \(3 \pm 1.73\) — substantial relative variability. With probability 0.58 there are 3 or more; with probability 0.05 there are none at all.

\[ \boxed{\ \textbf{A "12 per hour" bank is empty for 5 \% of its quarter-hours. Averages conceal a great deal.}\ } \]

Problem 2 — Exponential calculations and the memoryless property

Service times at a counter are exponential with mean 6 minutes. Find (a) \(\Pr[T \le 4]\); (b) \(\Pr[T > 10]\); (c) \(\Pr[T > 16 \mid T > 6]\); (d) the median service time.

Solution.

Set up. Mean \(= 1/\mu = 6\) minutes, so \(\mu = \tfrac16\) per minute.

(a) \(\Pr[T\le4]\).

\[ F(4) = 1 - e^{-4/6} = 1 - e^{-0.6667} = 1 - 0.5134 = \mathbf{0.4866} \]

(b) \(\Pr[T>10]\).

\[ 1 - F(10) = e^{-10/6} = e^{-1.6667} = \mathbf{0.1889} \]

(c) \(\Pr[T>16\mid T>6]\) — the memoryless test.

\[ \Pr[T>16\mid T>6] = \frac{\Pr[T>16]}{\Pr[T>6]} = \frac{e^{-16/6}}{e^{-6/6}} = \frac{e^{-2.6667}}{e^{-1}} = e^{-1.6667} = \mathbf{0.1889} \]
\[ \boxed{\ \textbf{IDENTICAL to (b).}\ } \]

The interpretation. A customer who has already been served for 6 minutes has exactly the same probability of needing 10 more minutes as a customer just beginning.

\[ \boxed{\ \textbf{The 6 minutes already spent count for NOTHING. That is memorylessness.}\ } \]

(d) The median.

\[ 1 - e^{-m/6} = 0.5 \ \Longrightarrow\ e^{-m/6} = 0.5 \ \Longrightarrow\ \frac{m}{6} = \ln2 \]
\[ m = 6\ln2 = 6(0.6931) = \mathbf{4.16\ \text{minutes}} \]

The mean is 6 minutes but the median is only 4.16.

\[ \boxed{\ \textbf{More than HALF of all customers are served in under 4.2 minutes, yet the average is 6.}\ } \]

Why. The distribution is strongly right-skewed: a small number of very long services pull the mean up. Specifically,

\[ \Pr[T > 3\times\text{mean}] = \Pr[T>18] = e^{-3} = 0.0498 \]

One customer in twenty takes more than three times the average. Reporting only the mean badly misrepresents customer experience — a lesson that applies to any right-skewed service metric.

Problem 3 — Testing whether data are Poisson

A shop records the number of customers arriving in each of 100 five-minute intervals:

Arrivals \(n\) 0 1 2 3 4 5 ≥6
Frequency 12 27 28 19 9 4 1

Test whether the arrivals are consistent with a Poisson process.

Solution.

Step 1 — estimate the mean.

\[ \bar n = \frac{0(12)+1(27)+2(28)+3(19)+4(9)+5(4)+6(1)}{100} \]
\[ = \frac{0+27+56+57+36+20+6}{100} = \frac{202}{100} = \mathbf{2.02} \]

Step 2 — compute the sample variance.

\[ \overline{n^2} = \frac{0+27+112+171+144+100+36}{100} = \frac{590}{100} = 5.90 \]
\[ s^2 = \overline{n^2} - \bar n^2 = 5.90 - (2.02)^2 = 5.90 - 4.0804 = \mathbf{1.82} \]

Step 3 — the mean/variance test.

\[ \frac{s^2}{\bar n} = \frac{1.82}{2.02} = \mathbf{0.90} \]

For a Poisson process this ratio should be close to 1 ✓ — 0.90 is well within sampling variation for \(n = 100\).

Step 4 — a chi-square goodness-of-fit test. Expected frequencies with \(\lambda t = 2.02\):

\(n\) Observed \(O_i\) \(P_n\) Expected \(E_i = 100P_n\) \(\dfrac{(O_i-E_i)^2}{E_i}\)
0 12 0.1327 13.27 0.122
1 27 0.2680 26.80 0.001
2 28 0.2707 27.07 0.032
3 19 0.1823 18.23 0.033
4 9 0.0920 9.20 0.004
≥5 5 0.0543 5.43 0.034
100 100.00 \(\chi^2 = 0.229\)

(Categories 5 and ≥6 are pooled so that every expected frequency exceeds 5, as the test requires.)

Step 5 — the decision.

Degrees of freedom \(= (\text{6 categories}) - 1 - (\text{1 estimated parameter}) = \mathbf{4}\).

Critical value at 5 %: \(\chi^2_{0.05,4} = 9.488\).

\[ 0.229 \ \ll\ 9.488 \]
\[ \boxed{\ \textbf{Do NOT reject. The data are entirely consistent with a Poisson process with } \lambda t = 2.02.\ } \]

Practical conclusion. With 5-minute intervals and \(\lambda t = 2.02\):

\[ \lambda = \frac{2.02}{5\ \text{min}} = 0.404\ \text{per minute} = \mathbf{24.2\ \text{per hour}} \]

The queuing models of L24–L28 may be applied with \(\lambda = 24.2\)/hour.

What a failure would have looked like.

  • If \(s^2/\bar n \approx 2\), the arrivals are overdispersed — likely clustered (groups arriving together) or a time-varying rate (busier at some times). Remedy: split the data by time period and fit separately, or use a compound Poisson model.
  • If \(s^2/\bar n \approx 0.3\), arrivals are too regular for Poisson — suggesting appointments or scheduling. Remedy: use a \(D/M/1\) or \(E_k/M/1\) model.
\[ \boxed{\ \textbf{Always test the Poisson assumption before applying the formulas. It takes ten minutes and can save a wrong recommendation.}\ } \]

Problem 4 — Superposition and thinning in practice

A call centre receives calls from three regions at Poisson rates 8, 12 and 5 per hour. (a) What is the combined process? (b) 30 % of calls require a specialist — what is the specialist arrival process? (c) If the specialist can handle 6 calls/hour, is one specialist enough?

Solution.

(a) Superposition. The sum of independent Poisson processes is Poisson with the summed rate:

\[ \lambda_{\text{total}} = 8 + 12 + 5 = \mathbf{25\ \text{calls/hour}}, \qquad \text{still Poisson} \]
\[ \boxed{\ \textbf{Merging Poisson streams gives a Poisson stream — the rates simply add.}\ } \]

(b) Thinning. Each call independently requires a specialist with probability \(p = 0.3\):

\[ \lambda_{\text{specialist}} = 25\times0.3 = \mathbf{7.5\ \text{calls/hour}}, \qquad \text{still Poisson} \]

And the complementary stream (general calls) is Poisson at \(25\times0.7 = 17.5\)/hour — and, remarkably, the two thinned streams are independent of each other.

(c) Is one specialist enough?

\[ \rho = \frac{\lambda}{\mu} = \frac{7.5}{6} = \mathbf{1.25} \ >\ 1 \]
\[ \boxed{\ \textbf{NO — the system is UNSTABLE. The specialist queue would grow without bound.}\ } \]

How many specialists are needed?

\[ \rho = \frac{7.5}{6c} < 1 \ \Longrightarrow\ c > 1.25 \ \Longrightarrow\ \boxed{c \ge 2} \]

With 2 specialists (\(M/M/2\), \(r = \lambda/\mu = 1.25\), \(\rho = 0.625\)):

\[ P_0 = \left[1 + 1.25 + \frac{1.25^2}{2(1-0.625)}\right]^{-1} = \left[1+1.25+2.0833\right]^{-1} = \frac{1}{4.3333} = 0.2308 \]
\[ L_q = \frac{P_0\,r^{c}\rho}{c!\,(1-\rho)^2} = \frac{0.2308(1.5625)(0.625)}{2(0.140625)} = \frac{0.2254}{0.28125} = \mathbf{0.801} \]
\[ W_q = \frac{L_q}{\lambda} = \frac{0.801}{7.5} = 0.1068\ \text{hours} = \mathbf{6.4\ \text{minutes}} \]

Recommendation. "Employ 2 specialists. With 1, the queue is mathematically unbounded — callers would wait indefinitely and the system would collapse. With 2, utilisation is 62.5 % and the average specialist wait is 6.4 minutes."

The value of superposition and thinning. They let us decompose a complex system into tractable pieces: we analysed the specialist sub-system in isolation, and were entitled to, because thinning preserves the Poisson property and independence. Without these theorems the decomposition would be invalid.

Problem 5 — Testing the exponential assumption, and what it costs to be wrong

A workshop records 20 service times (minutes): 2, 15, 4, 9, 1, 22, 6, 3, 11, 5, 8, 30, 2, 7, 14, 3, 19, 5, 10, 4. (a) Test whether these are consistent with an exponential distribution. (b) If arrivals are Poisson at 5/hour, compute \(L_q\) under both the exponential assumption and the actual variability, and compare.

Solution.

(a) Testing exponentiality.

Step 1 — the sample mean.

\[ \sum x_i = 2+15+4+9+1+22+6+3+11+5+8+30+2+7+14+3+19+5+10+4 = 180 \]
\[ \bar x = \frac{180}{20} = \mathbf{9.0\ \text{minutes}} \]

Step 2 — the sample standard deviation.

\[ \sum x_i^{2} = 4+225+16+81+1+484+36+9+121+25+64+900+4+49+196+9+361+25+100+16 = 2726 \]
\[ s^{2} = \frac{2726}{20} - 9^{2} = 136.3 - 81 = 55.3 \qquad\Longrightarrow\qquad s = \mathbf{7.44} \]

Step 3 — the coefficient of variation.

\[ \boxed{\ CV = \frac{s}{\bar x} = \frac{7.44}{9.0} = \mathbf{0.827}\ } \]

For an exponential distribution, \(CV = 1\) exactly. Our sample gives 0.827.

Is 0.827 close enough? For \(n = 20\) the sampling standard error of \(CV\) is roughly \(1/\sqrt{2n} = 1/\sqrt{40} = 0.158\). So

\[ z = \frac{0.827 - 1}{0.158} = -1.09 \]
\[ \boxed{\ \textbf{Not significant at 5 \% — the data are CONSISTENT with exponential service.}\ } \]

A supporting check — the median. Sorting the data, the 10th and 11th values are 6 and 7, so the sample median is 6.5. For an exponential with mean 9, the median should be \(9\ln2 = 6.24\). Observed 6.5 versus expected 6.24 ✓ — an excellent match, and the right kind of right-skew (median well below the mean of 9).

(b) The cost of the assumption.

\[ \lambda = 5/\text{hour}, \qquad \bar x = 9 \ \text{min} \Rightarrow \mu = \frac{60}{9} = 6.667/\text{hour} \]
\[ \rho = \frac{5}{6.667} = \mathbf{0.75} \]

Under the exponential assumption (\(M/M/1\)):

\[ L_q = \frac{\rho^2}{1-\rho} = \frac{0.5625}{0.25} = \mathbf{2.25\ \text{jobs}} \]
\[ W_q = \frac{2.25}{5} = 0.45\ \text{hours} = \mathbf{27\ \text{minutes}} \]

Using the ACTUAL variability (\(M/G/1\), Pollaczek–Khinchine). In hours, \(\sigma_s = 7.44/60 = 0.124\) h, so \(\sigma_s^2 = 0.01538\):

\[ L_q = \frac{\lambda^{2}\sigma_s^{2}+\rho^{2}}{2(1-\rho)} = \frac{25(0.01538) + 0.5625}{2(0.25)} = \frac{0.3845+0.5625}{0.5} = \frac{0.947}{0.5} = \mathbf{1.894} \]
\[ W_q = \frac{1.894}{5} = 0.379\ \text{h} = \mathbf{22.7\ \text{minutes}} \]

Comparison.

Model \(L_q\) \(W_q\) Error vs actual
\(M/M/1\) (assumed exponential) 2.25 27.0 min +19 %
\(M/G/1\) (actual \(\sigma\)) 1.89 22.7 min —
\(M/D/1\) (if service were constant) 1.125 13.5 min −41 %
\[ \boxed{\ \textbf{The exponential assumption OVERSTATES the queue by 19 \% here — it is CONSERVATIVE.}\ } \]

Why this direction matters. Because \(CV < 1\) (service is slightly more regular than exponential), \(M/M/1\) predicts a worse queue than reality. Being conservative is the safe direction for a capacity recommendation — you will not under-provision.

The general rule.

\[ \boxed{\ CV < 1 \Rightarrow M/M/1 \text{ is pessimistic (safe)}; \qquad CV > 1 \Rightarrow M/M/1 \text{ is OPTIMISTIC (dangerous)}.\ } \]

When \(CV > 1\) — the real danger. Service times with occasional very long jobs (IT support, medical diagnosis, machine repair) often have \(CV\) of 1.5–3. At \(CV = 2\) and \(\rho = 0.75\):

\[ L_q = \frac{25(4\times0.01538)+0.5625}{0.5} = \frac{1.538+0.5625}{0.5} = \mathbf{4.20} \]

— 87 % worse than the \(M/M/1\) prediction. A capacity plan built on \(M/M/1\) would badly under-provision.

\[ \boxed{\ \textbf{Always compute the sample } CV \textbf{ before trusting an } M/M/1 \textbf{ recommendation.}\ } \]

22.7 Practice Problems

  1. \(\lambda = 6\)/hour. Find \(\Pr[\text{exactly 3 in 30 minutes}]\). [Ans: \(\lambda t=3\), \(P_3 = 0.2240\)]
  2. Exponential with mean 5 minutes. Find \(\Pr[T>8]\). [Ans: \(e^{-1.6}=0.2019\)]
  3. Verify the memoryless property numerically for \(\mu = 0.2\), \(s=3\), \(t=5\).
  4. A sample has mean 10 and variance 25. Is it consistent with Poisson? [Ans: no — overdispersed, ratio 2.5]
  5. Three Poisson streams at 4, 7, 9/hour merge. What is the combined process? [Ans: Poisson, 20/hour]
  6. 20 % of a Poisson stream at 50/hour is diverted. What is the diverted process? [Ans: Poisson, 10/hour]
  7. Service data have \(CV = 1.8\). Is \(M/M/1\) optimistic or pessimistic? [Ans: optimistic — dangerous]

22.8 Summary

  • Poisson counts and exponential times are two views of the same process. The three axioms are independence, stationarity and orderliness.
  • \(P_n(t) = \dfrac{(\lambda t)^n e^{-\lambda t}}{n!}\), with \(E[N] = \mathrm{Var}[N] = \lambda t\). Mean = variance is the Poisson signature, and the ratio \(s^2/\bar n\) is the standard diagnostic.
  • Superposition (rates add) and thinning (rate \(\times p\)) both preserve the Poisson property — these theorems justify decomposing complex systems.
  • Inter-arrival times are exponential: \(f(t) = \lambda e^{-\lambda t}\), with \(E[T] = \sigma_T = 1/\lambda\), so \(CV = 1\) exactly. The median \(= 0.693/\lambda\) is well below the mean — strongly right-skewed.
  • Memorylessness \(\Pr[T>s+t\mid T>s] = \Pr[T>t]\) characterises the exponential uniquely. It is what makes queuing models Markovian and tractable — and it is often unrealistic for service times, though usually fine for arrivals.
  • The Pollaczek–Khinchine formula shows exponential service gives exactly twice the queue of deterministic service. So \(M/M/1\) is conservative when \(CV<1\) and dangerously optimistic when \(CV>1\).
  • Test both assumptions before using the formulas: \(s^2/\bar n \approx 1\) for Poisson arrivals, \(CV\approx1\) for exponential service.

Next lesson: L23 — Birth–death processes and steady-state balance equations.

Practise now: unit3-queue-inventory-lab.html → Distribution Explorer: generate Poisson and exponential samples, see the histograms, and test your own data against both.