Poisson and Birth-and-Death Processes
CSIR NET statistics: Poisson process properties, continuous-time Markov chains, Kolmogorov equations, birth-and-death processes, Yule and M/M/1 queues.
- 19 min
On this page · 12 sections
Prerequisites
- n-Step Transition Limits and Stationary Distributions: stationary distributions and detailed balance carry over to continuous time with \( \pi Q = 0 \).
- Linear ODEs and Variation of Parameters: Kolmogorov’s equations are linear systems of first-order ODEs.
- Counting, Probability and Number Properties: binomial and Poisson probabilities.
The Hook: Why the Bus Is Always Late
The Poisson Process
A counting process \( \{N(t) : t \ge 0\} \) is a Poisson process with rate \( \lambda > 0 \) if
- \( N(0) = 0 \);
- it has independent increments: counts in disjoint intervals are independent;
- for \( s, t \ge 0 \), \( N(s + t) - N(s) \sim \text{Poisson}(\lambda t) \): $$P(N(s + t) - N(s) = k) = e^{-\lambda t}\frac{(\lambda t)^k}{k!}, \quad k = 0, 1, 2, \dots$$
Two equivalent definitions, each useful in its own place:
- Infinitesimal. Independent, stationary increments, and as \( h \to 0 \): \( P(N(h) = 1) = \lambda h + o(h) \), \( P(N(h) \ge 2) = o(h) \). This is the form that turns into differential equations.
- Interarrival. The gaps \( \tau_1, \tau_2, \dots \) between successive events are i.i.d. \( \text{Exp}(\lambda) \), density \( \lambda e^{-\lambda x} \), mean \( \frac1\lambda \); \( N(t) = \max\{n : \tau_1 + \dots + \tau_n \le t\} \).
Properties to know cold
| Quantity | Value |
|---|---|
| \( E[N(t)] \), \( \operatorname{Var}N(t) \) | both \( \lambda t \) |
| \( \operatorname{Cov}(N(s), N(t)) \) | \( \lambda\min(s, t) \) |
| interarrival time \( \tau_k \) | \( \text{Exp}(\lambda) \), memoryless: \( P(\tau > s + t \mid \tau > s) = P(\tau > t) \) |
| \( n \)-th arrival time \( S_n = \tau_1 + \dots + \tau_n \) | \( \text{Gamma}(n, \lambda) \) (Erlang): density \( \frac{\lambda^nt^{n-1}e^{-\lambda t}}{(n-1)!} \), mean \( \frac{n}{\lambda} \), variance \( \frac{n}{\lambda^2} \) |
| duality | \( \{S_n \le t\} = \{N(t) \ge n\} \) |
| given \( N(t) = n \) | \( (S_1, \dots, S_n) \) are the order statistics of \( n \) i.i.d. \( U(0, t) \) |
| given \( N(t) = n \), \( s < t \) | \( N(s) \sim \text{Bin}\left(n, \frac{s}{t}\right) \) |
| strong law | \( \frac{N(t)}{t} \to \lambda \) almost surely |
Superposition. If \( N_1, N_2 \) are independent Poisson processes with rates \( \lambda_1, \lambda_2 \), then \( N_1 + N_2 \) is Poisson with rate \( \lambda_1 + \lambda_2 \), and each event of the sum comes from \( N_1 \) with probability \( \frac{\lambda_1}{\lambda_1 + \lambda_2} \), independently of everything else. Equivalently, for independent exponentials, \( \min(X_1, X_2) \sim \text{Exp}(\lambda_1 + \lambda_2) \) and \( P(X_1 < X_2) = \frac{\lambda_1}{\lambda_1 + \lambda_2} \).
Thinning (splitting). Classify each event of a rate-\( \lambda \) process as type 1 with probability \( p \), independently. The type-1 and type-2 counts are independent Poisson processes with rates \( \lambda p \) and \( \lambda(1 - p) \). Independence is the surprising part: knowing how many type-1 events occurred says nothing about type 2.
Compound Poisson. \( X(t) = \sum_{k=1}^{N(t)} Y_k \), with \( Y_k \) i.i.d. and independent of \( N \): \( E[X(t)] = \lambda tE[Y] \), \( \operatorname{Var}X(t) = \lambda tE[Y^2] \).
Non-homogeneous Poisson process. With a rate function \( \lambda(t) \), \( N(t) - N(s) \sim \text{Poisson}\left(\int_s^t \lambda(u)\,du\right) \), increments still independent but no longer stationary.
Continuous-Time Markov Chains
A process \( \{X(t)\} \) on a countable \( S \) is a continuous-time Markov chain if \( P(X(t + s) = j \mid X(s) = i, \text{history}) = p_{ij}(t) \). Its behaviour is described by:
- Holding times. In state \( i \) the chain waits an \( \text{Exp}(q_i) \) time (exponential is forced by the Markov property).
- Jump chain. It then jumps to \( j \ne i \) with probability \( \frac{q_{ij}}{q_i} \), where \( q_{ij} \ge 0 \) is the transition rate \( i \to j \) and \( q_i = \sum_{j \ne i} q_{ij} \).
- Generator \( Q = (q_{ij}) \) with \( q_{ii} = -q_i \): off-diagonal entries non-negative, rows sum to 0. For small \( h \), \( p_{ij}(h) = q_{ij}h + o(h) \) (\( j \ne i \)) and \( p_{ii}(h) = 1 - q_ih + o(h) \).
With \( P(t) = (p_{ij}(t)) \) and \( P(0) = I \) (for a finite state space, or a well-behaved countable one):
$$\text{backward: } P'(t) = QP(t), \qquad \text{forward: } P'(t) = P(t)Q, \qquad P(t) = e^{tQ}.$$The matrices satisfy Chapman–Kolmogorov: \( P(s + t) = P(s)P(t) \).
A distribution \( \pi \) is stationary iff \( \pi Q = 0 \) (equivalently \( \pi P(t) = \pi \) for all \( t \)). Balance form: for each \( j \), rate out \( \pi_jq_j \) equals rate in \( \sum_{i \ne j}\pi_iq_{ij} \). Detailed balance \( \pi_iq_{ij} = \pi_jq_{ji} \) is again sufficient. For an irreducible, positive recurrent (non-explosive) chain, \( p_{ij}(t) \to \pi_j \) as \( t \to \infty \). There is no periodicity in continuous time.
Birth-and-Death Processes
Forward equations for \( p_n(t) = P(X(t) = n) \):
$$p_n'(t) = \lambda_{n-1}p_{n-1}(t) - (\lambda_n + \mu_n)p_n(t) + \mu_{n+1}p_{n+1}(t), \qquad p_0'(t) = -\lambda_0p_0(t) + \mu_1p_1(t).$$Stationary distribution. Setting the derivatives to 0 and solving from \( n = 0 \) upward gives the cut equations \( \lambda_{n-1}\pi_{n-1} = \mu_n\pi_n \) (flow up across a cut equals flow down):
$$\boxed{\pi_n = \pi_0\prod_{k=1}^{n}\frac{\lambda_{k-1}}{\mu_k}}, \qquad \pi_0 = \left(1 + \sum_{n \ge 1}\prod_{k=1}^{n}\frac{\lambda_{k-1}}{\mu_k}\right)^{-1},$$which is a distribution iff \( \sum_{n \ge 1}\prod_{k=1}^{n}\frac{\lambda_{k-1}}{\mu_k} < \infty \).
The standard models
| Model | Rates | Key result |
|---|---|---|
| Poisson process | \( \lambda_n = \lambda \), \( \mu_n = 0 \) | \( X(t) \sim \text{Poisson}(\lambda t) \) |
| Yule (linear pure birth) | \( \lambda_n = n\lambda \), \( \mu_n = 0 \), \( X(0) = 1 \) | \( P(X(t) = n) = e^{-\lambda t}(1 - e^{-\lambda t})^{n-1} \): geometric, mean \( e^{\lambda t} \) |
| Linear pure death | \( \mu_n = n\mu \), \( X(0) = N \) | \( X(t) \sim \text{Bin}(N, e^{-\mu t}) \) |
| Linear birth–death | \( \lambda_n = n\lambda \), \( \mu_n = n\mu \) | \( E[X(t)] = X(0)e^{(\lambda - \mu)t} \); extinction probability 1 if \( \lambda \le \mu \), \( \left(\frac{\mu}{\lambda}\right)^{X(0)} \) if \( \lambda > \mu \) |
| M/M/1 queue | \( \lambda_n = \lambda \), \( \mu_n = \mu \) (\( n \ge 1 \)) | \( \rho = \frac{\lambda}{\mu} < 1 \): \( \pi_n = (1 - \rho)\rho^n \) |
| M/M/\( \infty \) (immigration–death) | \( \lambda_n = \lambda \), \( \mu_n = n\mu \) | \( \pi = \text{Poisson}\left(\frac{\lambda}{\mu}\right) \), for every \( \lambda, \mu \) |
| M/M/\( c \) | \( \lambda_n = \lambda \), \( \mu_n = \min(n, c)\mu \) | stable iff \( \lambda < c\mu \); \( \pi_n \propto \frac{(\lambda/\mu)^n}{n!} \) for \( n \le c \), \( \propto \frac{(\lambda/\mu)^n}{c!\,c^{n-c}} \) for \( n \ge c \) |
| M/M/1/\( K \) (finite waiting room) | \( \lambda_n = \lambda \) for \( n < K \), \( \mu_n = \mu \) | \( \pi_n \propto \rho^n \), \( 0 \le n \le K \): always stable |
In the linear birth–death process, 0 is absorbing. The extinction probability comes from the jump chain, a gambler’s-ruin walk with up-probability \( \frac{\lambda}{\lambda + \mu} \) and no upper barrier (gambler’s ruin with \( N \to \infty \)); each of the \( X(0) \) initial lines dies out independently.
M/M/1 performance measures
With arrival rate \( \lambda \), service rate \( \mu \) and \( \rho = \frac{\lambda}{\mu} < 1 \):
| Measure | Formula |
|---|---|
| \( P(\text{server idle}) \) | \( \pi_0 = 1 - \rho \) |
| \( P(X \ge n) \) | \( \rho^n \) |
| mean number in system \( L \) | \( \frac{\rho}{1 - \rho} = \frac{\lambda}{\mu - \lambda} \) |
| mean number waiting \( L_q \) | \( \frac{\rho^2}{1 - \rho} \) |
| mean time in system \( W \) | \( \frac{1}{\mu - \lambda} \) |
| mean waiting time in queue \( W_q \) | \( \frac{\rho}{\mu - \lambda} \) |
Little’s law links them: \( L = \lambda W \) and \( L_q = \lambda W_q \); also \( W = W_q + \frac{1}{\mu} \). For \( \rho \ge 1 \) there is no stationary distribution: the queue is null recurrent (\( \rho = 1 \)) or transient (\( \rho > 1 \)), exactly like the reflecting random walk on the previous page.
Explosion
A pure birth process with rates \( \lambda_n \) reaches infinitely many births in finite time with positive probability iff
$$\sum_{n} \frac{1}{\lambda_n} < \infty$$(the expected total of the holding times is finite). Linear growth \( \lambda_n = n\lambda \) does not explode (harmonic series); quadratic growth \( \lambda_n = n^2 \) does. Explosion is why the Kolmogorov equations need care on infinite state spaces.
graph LR
S0((0)) -->|"lambda_0"| S1((1))
S1 -->|"mu_1"| S0
S1 -->|"lambda_1"| S2((2))
S2 -->|"mu_2"| S1
S2 -->|"lambda_2"| S3((3))
S3 -->|"mu_3"| S2
S3 -->|"..."| S4((...))Memory Tricks
- Exponential forgets. The waiting time from now is \( \text{Exp}(\lambda) \) no matter how long you have already waited.
- Up-flow equals down-flow. \( \lambda_{n-1}\pi_{n-1} = \mu_n\pi_n \) across every cut: write the product, normalise.
- Merge adds, split keeps Poisson. Superposition adds rates; thinning gives independent Poisson streams.
- Rows of Q sum to zero, rows of P sum to one. Generator versus transition matrix.
- M/M/1 in one line: \( L = \frac{\lambda}{\mu - \lambda} \), \( W = \frac{1}{\mu - \lambda} \), and \( L = \lambda W \) ties them.
Don’t Confuse
| This | Is not this | The difference |
|---|---|---|
| \( N(t) \sim \text{Poisson}(\lambda t) \) | \( S_n \sim \text{Gamma}(n, \lambda) \) | counts at a fixed time vs time of a fixed count |
| given \( N(t) = n \): arrivals uniform | unconditional arrivals uniform | uniformity holds only after conditioning |
| generator \( Q \) (rows sum to 0) | transition matrix \( P(t) \) (rows sum to 1) | \( P(t) = e^{tQ} \) |
| \( \pi Q = 0 \) | \( \pi P = \pi \) | continuous vs discrete time |
| stationary \( \pi \) of the CTMC | stationary law of the jump chain | the jump chain’s is \( \propto \pi_iq_i \) |
| \( W \) (time in system) | \( W_q \) (time in queue) | differ by the mean service time \( \frac1\mu \) |
| M/M/\( \infty \) | M/M/1 | M/M/\( \infty \) is stable for all rates and has Poisson \( \pi \) |
| \( \operatorname{Var}N(t) = \lambda t \) | \( \operatorname{Var}S_n = \frac{n}{\lambda^2} \) | variance of a count vs of a waiting time |
Common Mistakes
Practice
All questions below are original practice questions written for this page. They are not from any CSIR NET paper. In Level 3, as in Part C of the Mathematical Sciences paper, one or more options may be correct.
Level 1: Foundation
- What is the distribution of the time between consecutive events of a Poisson process with rate \( \lambda \)?
- For a Poisson process with rate 2, what is \( E[N(3)] \)?
- What condition makes the M/M/1 queue have a stationary distribution?
Answers
- Exponential with rate \( \lambda \) (mean \( \frac1\lambda \)), independent from gap to gap.
- 6. \( E[N(t)] = \lambda t \).
- \( \rho = \frac{\lambda}{\mu} < 1 \): arrivals slower than service.
Level 2: Part B Style
Practice Q1. For a Poisson process with rate \( \lambda = 1.5 \), \( P(N(2) = 3) \) is:
(a) \( 4.5e^{-3} \) (b) \( 3e^{-3} \) (c) \( 0.5625e^{-1.5} \) (d) \( 9e^{-3} \)
Answer: (a)
\( N(2) \sim \text{Poisson}(3) \): \( e^{-3}\frac{3^3}{3!} = \frac{27}{6}e^{-3} = 4.5e^{-3} \approx 0.224 \). Option (c) uses \( \lambda \) instead of \( \lambda t \): \( e^{-1.5}\frac{1.5^3}{6} \).
Practice Q2. For a Poisson process, given that \( N(4) = 6 \), the probability that \( N(1) = 2 \) is:
(a) \( \frac{1215}{4096} \) (b) \( \frac{15}{64} \) (c) \( e^{-1}\frac{1}{2} \) (d) depends on the rate \( \lambda \)
Answer: (a)
Given \( N(4) = 6 \), \( N(1) \sim \text{Bin}\left(6, \frac14\right) \): \( \binom62\left(\frac14\right)^2\left(\frac34\right)^4 = \frac{15 \cdot 81}{4096} = \frac{1215}{4096} \approx 0.297 \). Option (d) is the trap: the rate cancels in the conditional probability.
Practice Q3. At a single-server counter, arrivals are Poisson at 4 per hour and service times are exponential with mean 12 minutes. The mean time a customer spends at the counter (waiting plus service) is:
(a) 12 minutes (b) 48 minutes (c) 60 minutes (d) 4 hours
Answer: (c)
\( \lambda = 4 \), \( \mu = 5 \) per hour. \( W = \frac{1}{\mu - \lambda} = 1 \) hour. Option (a) is the service time alone, (b) is \( W_q = \frac{\rho}{\mu - \lambda} = 0.8 \) hour, (d) confuses \( W \) with \( L = 4 \) customers.
Practice Q4. A Yule process with birth rate \( \lambda = 1 \) per individual starts with one individual. At time \( t = \ln 2 \), the probability that the population is exactly 3 is:
(a) \( \frac18 \) (b) \( \frac14 \) (c) \( \frac{3}{8} \) (d) \( \frac{1}{2} \)
Answer: (a)
\( P(X(t) = n) = e^{-\lambda t}(1 - e^{-\lambda t})^{n-1} \) with \( e^{-t} = \frac12 \): \( \frac12 \cdot \left(\frac12\right)^2 = \frac18 \). (Mean population \( e^{t} = 2 \).)
Level 3: Part C Style
Practice Q5. Let \( N(t) \) be a Poisson process with rate 2 and \( S_n \) its \( n \)-th arrival time. Which statements are true?
(a) \( E[S_3] = \frac32 \) and \( \operatorname{Var}(S_3) = \frac34 \). (b) \( \operatorname{Cov}(N(1), N(3)) = 2 \). (c) \( N(3) - N(1) \) is independent of \( N(1) \). (d) \( P(S_2 \le 1) = 3e^{-2} \).
Answer: (a), (b) and (c)
- (a) True. \( S_3 \sim \text{Gamma}(3, 2) \): mean \( \frac32 \), variance \( \frac{3}{4} \).
- (b) True. \( \operatorname{Cov} = \lambda\min(1, 3) = 2 \). (Or: \( \operatorname{Cov}(N(1), N(1) + [N(3) - N(1)]) = \operatorname{Var}N(1) = 2 \).)
- (c) True. Disjoint intervals \( (0, 1] \) and \( (1, 3] \).
- (d) False. \( 3e^{-2} = P(N(1) \le 1) = P(S_2 > 1) \). So \( P(S_2 \le 1) = 1 - 3e^{-2} \approx 0.594 \).
Practice Q6. In a birth-and-death process, arrivals are discouraged: \( \lambda_n = \frac{2}{n + 1} \) for \( n \ge 0 \), and \( \mu_n = 1 \) for \( n \ge 1 \). Which statements are true?
(a) The stationary distribution is Poisson with mean 2. (b) \( \pi_0 = \frac13 \). (c) In the long run, the mean number in the system is 2. (d) A stationary distribution would not exist if the constant in \( \lambda_n \) were 5 instead of 2.
Answer: (a) and (c)
\( \prod_{k=1}^{n}\frac{\lambda_{k-1}}{\mu_k} = \prod_{k=1}^{n}\frac{2}{k} = \frac{2^n}{n!} \), so \( \pi_n = \pi_0\frac{2^n}{n!} \) and \( \pi_0 = e^{-2} \).
- (a) True. \( \pi_n = e^{-2}\frac{2^n}{n!} \).
- (b) False. \( \pi_0 = e^{-2} \approx 0.135 \). (\( \frac13 \) would be an M/M/1 guess.)
- (c) True. Mean of Poisson(2).
- (d) False. With 5, \( \pi_n \propto \frac{5^n}{n!} \), still summable: Poisson(5). The factorial beats any geometric growth, so no stability condition is needed.
Practice Q7. A linear birth-and-death process has birth rate \( \lambda = 2 \) and death rate \( \mu = 1 \) per individual, with \( X(0) = 1 \). Which statements are true?
(a) The probability that the population eventually dies out is \( \frac12 \). (b) \( E[X(t)] = e^{t} \). (c) If \( X(0) = 3 \), the extinction probability is \( \frac18 \). (d) If instead \( \lambda = \mu = 1 \), the population survives forever with positive probability.
Answer: (a), (b) and (c)
- (a) True. \( \lambda > \mu \): extinction probability \( \frac{\mu}{\lambda} = \frac12 \). (Jump chain: up with probability \( \frac23 \), down with \( \frac13 \); ruin against an infinite target from 1 is \( \frac{q}{p} = \frac12 \).)
- (b) True. \( E[X(t)] = X(0)e^{(\lambda - \mu)t} = e^t \).
- (c) True. The three lines die out independently: \( \left(\frac12\right)^3 \).
- (d) False. For \( \lambda \le \mu \) extinction is certain, even though \( E[X(t)] = 1 \) for all \( t \) when \( \lambda = \mu \).
Quick Revision Table
| Idea | Fact |
|---|---|
| Poisson process | independent increments, \( N(s + t) - N(s) \sim \text{Poisson}(\lambda t) \) |
| Gaps | i.i.d. \( \text{Exp}(\lambda) \), memoryless |
| Arrival times | \( S_n \sim \text{Gamma}(n, \lambda) \); \( S_n \le t \iff N(t) \ge n \) |
| Conditioning | given \( N(t) = n \): uniform order statistics; \( N(s) \sim \text{Bin}(n, \frac{s}{t}) \) |
| Covariance | \( \lambda\min(s, t) \) |
| Superposition / thinning | rates add; split streams are independent Poisson |
| CTMC | holding \( \text{Exp}(q_i) \), generator rows sum to 0, \( P(t) = e^{tQ} \) |
| Kolmogorov | backward \( P' = QP \), forward \( P' = PQ \); stationary \( \pi Q = 0 \) |
| Birth–death \( \pi \) | \( \pi_n = \pi_0\prod_{k=1}^{n}\frac{\lambda_{k-1}}{\mu_k} \) |
| Yule from 1 | geometric with \( p = e^{-\lambda t} \), mean \( e^{\lambda t} \) |
| Linear birth–death | mean \( X(0)e^{(\lambda - \mu)t} \); extinction \( \min\left(1, \frac{\mu}{\lambda}\right)^{X(0)} \) |
| M/M/1 | \( \pi_n = (1 - \rho)\rho^n \), \( L = \frac{\rho}{1 - \rho} \), \( W = \frac{1}{\mu - \lambda} \), \( L = \lambda W \) |
| M/M/\( \infty \) | Poisson\( \left(\frac{\lambda}{\mu}\right) \) |
| Explosion (pure birth) | iff \( \sum \frac{1}{\lambda_n} < \infty \) |
Key Takeaways
What’s Next
- Markov Chains Quick Revision: the formulas of all three pages on one sheet.
- n-Step Transition Limits and Stationary Distributions: the discrete-time version of every limit on this page.
- Linear ODEs and Variation of Parameters: solving the linear systems that Kolmogorov’s equations produce.
- Back to the chapter overview or the Mathematical Sciences overview.