{
  "id": "birth-death-process",
  "title": "Birth–Death Process",
  "lang": "en",
  "source_lang": "en",
  "summary": "A birth–death process is a continuous-time Markov model on nonnegative integers whose jumps increase or decrease the state by one, used to describe populations, queues, and other count-valued systems.",
  "category": "mathematics",
  "subcategory": "probability-statistics",
  "type": "concept",
  "aliases": [
    "Birth and death process"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/birth-death-process/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/birth-death-process/",
      "json": "https://aiwiki.page/zh/birth-death-process.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A birth–death process is a continuous-time Markov chain whose states are nonnegative integers and whose only possible jumps are from a state \\(n\\) to \\(n+1\\) or \\(n-1\\). Upward jumps are called births and downward jumps deaths, regardless of whether they represent biological events. The model is used for population sizes, numbers of customers in queues, chemical molecule counts, and other quantities that change one unit at a time. Its nearest-neighbor structure makes many questions about long-run behavior, extinction, and passage times more tractable than in general Markov models."
    },
    {
      "heading": "Definition and transition rates",
      "text": "Let \\(X(t)\\) denote the state at time \\(t\\), with state space either\n\n\\[\\{0,1,2,\\ldots\\}\\]\n\nor a finite set \\(\\{0,1,\\ldots,N\\}\\). In the standard time-homogeneous model, each state \\(n\\) has a birth rate \\(\\lambda_n\\geq 0\\) and a death rate \\(\\mu_n\\geq 0\\). For small \\(h>0\\),\n\n\\[\\Pr\\{X(t+h)=n+1\\mid X(t)=n\\} =\\lambda_n h+o(h),\\]\n\n\\[\\Pr\\{X(t+h)=n-1\\mid X(t)=n\\} =\\mu_n h+o(h),\\]\n\nand\n\n\\[\\Pr\\{X(t+h)=n\\mid X(t)=n\\} =1-(\\lambda_n+\\mu_n)h+o(h).\\]\n\nHere \\(o(h)\\) denotes terms whose ratio to \\(h\\) tends to zero as \\(h\\) tends to zero.\n\nThe Markov property means that, conditional on the present state, the future does not depend on the earlier history. Time homogeneity means that the rates depend on the state but not on calendar time.\n\nBecause the state cannot become negative, \\(\\mu_0=0\\). A finite model also requires \\(\\lambda_N=0\\). If \\(\\lambda_0=0\\), state \\(0\\) is absorbing: once the process reaches zero, it remains there. If \\(\\lambda_0>0\\), upward transitions from zero may represent immigration, arrivals, or spontaneous production.\n\nConditional on entering state \\(n\\), the holding time there has an exponential distribution with parameter \\(\\lambda_n+\\mu_n\\). Provided this parameter is positive, the next jump is upward with probability\n\n\\[\\frac{\\lambda_n}{\\lambda_n+\\mu_n}\\]\n\nand downward with probability\n\n\\[\\frac{\\mu_n}{\\lambda_n+\\mu_n}.\\]\n\nIf both rates vanish, the state is absorbing. The sequence of states observed at jump times is therefore a nearest-neighbor random walk, generally with state-dependent transition probabilities."
    },
    {
      "heading": "Generator and evolution equations",
      "text": "The infinitesimal generator is the tridiagonal matrix \\(Q=(q_{ij})\\) defined by\n\n\\[q_{n,n+1}=\\lambda_n,\\qquad q_{n,n-1}=\\mu_n,\\qquad q_{n,n}=-(\\lambda_n+\\mu_n),\\]\n\nwith all other off-diagonal entries zero.\n\nFor a suitable function \\(f\\), the generator acts as\n\n\\[(Lf)(n) =\\lambda_n\\bigl[f(n+1)-f(n)\\bigr] +\\mu_n\\bigl[f(n-1)-f(n)\\bigr],\\]\n\nwith the death term omitted at zero. This expresses the expected instantaneous change of a function of the state.\n\nForward equations\n\nLet\n\n\\[p_n(t)=\\Pr\\{X(t)=n\\}.\\]\n\nThe Kolmogorov forward equations, also called the master equations in this setting, are\n\n\\[\\frac{dp_n(t)}{dt} =\\lambda_{n-1}p_{n-1}(t) +\\mu_{n+1}p_{n+1}(t) -(\\lambda_n+\\mu_n)p_n(t), \\qquad n\\geq1,\\]\n\nand\n\n\\[\\frac{dp_0(t)}{dt} =\\mu_1p_1(t)-\\lambda_0p_0(t).\\]\n\nEach equation balances probability flowing into a state against probability flowing out. Together with an initial probability distribution, these equations describe the transient law, subject to the usual existence and nonexplosion conditions.\n\nFor a finite state space, the transition matrix is\n\n\\[P(t)=e^{tQ},\\]\n\nwhere the exponential is the matrix exponential. For infinite state spaces, the same notation requires additional care because the generator may be an unbounded operator.\n\nMoment equations\n\nUnder appropriate integrability conditions,\n\n\\[\\frac{d}{dt}\\mathbb E[f(X(t))] =\\mathbb E[(Lf)(X(t))].\\]\n\nTaking \\(f(n)=n\\) gives the evolution of the mean:\n\n\\[\\frac{d}{dt}\\mathbb E[X(t)] =\\mathbb E[\\lambda_{X(t)}-\\mu_{X(t)}].\\]\n\nTaking \\(f(n)=n^2\\) yields\n\n\\[\\frac{d}{dt}\\mathbb E[X(t)^2] = \\mathbb E\\!\\left[ 2X(t)(\\lambda_{X(t)}-\\mu_{X(t)}) +\\lambda_{X(t)}+\\mu_{X(t)} \\right].\\]\n\nThese identities provide equations for the mean and variance. With nonlinear rates, however, lower moments generally depend on higher moments, producing a hierarchy that need not close."
    },
    {
      "heading": "Stationary distributions and reversibility",
      "text": "A stationary distribution is a probability vector \\(\\pi=(\\pi_n)\\) that remains unchanged under the process's evolution. For a nonexplosive, irreducible birth–death process on the nonnegative integers, stationary probabilities satisfy the detailed-balance equations\n\n\\[\\pi_n\\lambda_n=\\pi_{n+1}\\mu_{n+1}.\\]\n\nThus equilibrium flow across each edge is equal in both directions.\n\nAssuming \\(\\lambda_n>0\\) for \\(n\\geq0\\) and \\(\\mu_n>0\\) for \\(n\\geq1\\), define\n\n\\[w_0=1,\\qquad w_n=\\prod_{k=0}^{n-1}\\frac{\\lambda_k}{\\mu_{k+1}}.\\]\n\nIf\n\n\\[Z=\\sum_{n=0}^{\\infty}w_n<\\infty,\\]\n\nthen\n\n\\[\\pi_n=\\frac{w_n}{Z}.\\]\n\nUnder the stated nonexplosion and irreducibility assumptions, this is the unique stationary distribution, and the process is positive recurrent. Detailed balance also makes the stationary process reversible: reversing time leaves its finite-dimensional distributions unchanged.\n\nFor a finite irreducible birth–death process, the corresponding finite normalizing sum always exists. For an infinite process, divergence of \\(Z\\) rules out a stationary probability distribution, but does not by itself distinguish null recurrence from transience.\n\nIf zero is absorbing, the distribution concentrated at zero is stationary. A population may nevertheless remain nonzero for a long time; its distribution conditional on survival can behave very differently from its unconditional stationary law."
    },
    {
      "heading": "Important models",
      "text": "Pure birth and pure death processes\n\nA pure birth process has \\(\\mu_n=0\\) for every state. If \\(\\lambda_n=\\lambda\\) is constant and \\(X(0)=0\\), it is a Poisson process, with\n\n\\[X(t)\\sim\\operatorname{Poisson}(\\lambda t).\\]\n\nA pure death process has \\(\\lambda_n=0\\). When \\(\\mu_n=\\mu n\\), each of the initially present individuals has an independent exponential lifetime with rate \\(\\mu\\). If \\(X(0)=m\\), then\n\n\\[X(t)\\sim\\operatorname{Binomial}(m,e^{-\\mu t}),\\]\n\na binomial distribution.\n\nLinear birth–death branching process\n\nThe linear model has rates\n\n\\[\\lambda_n=\\lambda n,\\qquad \\mu_n=\\mu n.\\]\n\nEach individual independently gives birth at rate \\(\\lambda\\) and dies at rate \\(\\mu\\). It is therefore a continuous-time branching process, with zero absorbing.\n\nStarting from \\(X(0)=m\\),\n\n\\[\\mathbb E[X(t)]=m e^{(\\lambda-\\mu)t}.\\]\n\nThe model is called:\n\n- Subcritical if \\(\\lambda<\\mu\\).\n- Critical if \\(\\lambda=\\mu\\).\n- Supercritical if \\(\\lambda>\\mu\\).\n\nIn the nondegenerate case \\(\\lambda+\\mu>0\\), the eventual extinction probability from one individual is\n\n\\[q= \\begin{cases} 1,&\\lambda\\leq\\mu,\\\\[4pt] \\mu/\\lambda,&\\lambda>\\mu. \\end{cases}\\]\n\nFrom \\(m\\geq1\\) individuals, independence of the descendant families gives extinction probability \\(q^m\\).\n\nWhen \\(\\lambda=\\mu>0\\), extinction occurs with probability one even though the mean population remains \\(m\\) at every finite time. Rare surviving populations compensate for the increasing probability of extinction. A constant mean therefore does not imply that typical sample paths remain near that mean.\n\nSetting \\(\\mu=0\\) and \\(\\lambda>0\\) gives the Yule pure birth model.\n\nImmigration–death process\n\nSuppose\n\n\\[\\lambda_n=\\nu,\\qquad \\mu_n=\\mu n,\\]\n\nwith \\(\\nu,\\mu>0\\). Arrivals occur at a constant rate, while each present individual disappears independently at rate \\(\\mu\\). The stationary law is a Poisson distribution:\n\n\\[\\pi_n=e^{-\\nu/\\mu}\\frac{(\\nu/\\mu)^n}{n!}.\\]\n\nUnlike a population model without immigration, this process can leave state zero.\n\nQueueing models\n\nIn queueing theory, \\(X(t)\\) often counts customers in a service system. An \\(M/M/1\\) queue has Poisson arrivals at rate \\(\\lambda\\), one server with exponential service times of rate \\(\\mu\\), and rates\n\n\\[\\lambda_n=\\lambda,\\qquad \\mu_n= \\begin{cases} 0,&n=0,\\\\ \\mu,&n\\geq1. \\end{cases}\\]\n\nWriting \\(\\rho=\\lambda/\\mu\\), its stationary law exists precisely when \\(\\rho<1\\), and is\n\n\\[\\pi_n=(1-\\rho)\\rho^n.\\]\n\nIn that regime,\n\n\\[\\mathbb E_\\pi[X]=\\frac{\\rho}{1-\\rho}.\\]\n\nWith \\(c\\) identical servers, the total service-completion rate becomes\n\n\\[\\mu_n=\\min(n,c)\\mu,\\]\n\ngiving another birth–death model."
    },
    {
      "heading": "Hitting probabilities and passage times",
      "text": "Nearest-neighbor transitions lead to simple recurrence relations for boundary-hitting questions.\n\nOn a finite interval \\(\\{0,\\ldots,N\\}\\), let \\(u_n\\) be the probability of hitting \\(N\\) before \\(0\\), starting from \\(n\\). For positive interior birth and death rates,\n\n\\[\\lambda_n(u_{n+1}-u_n) +\\mu_n(u_{n-1}-u_n)=0,\\]\n\nwith \\(u_0=0\\) and \\(u_N=1\\). Its solution is\n\n\\[u_n= \\frac{ \\displaystyle\\sum_{k=0}^{n-1} \\prod_{j=1}^{k}\\frac{\\mu_j}{\\lambda_j} }{ \\displaystyle\\sum_{k=0}^{N-1} \\prod_{j=1}^{k}\\frac{\\mu_j}{\\lambda_j} },\\]\n\nwhere an empty product equals one. This generalizes the gambler's ruin probability to state-dependent rates.\n\nSimilarly, if \\(T_n\\) is the expected time to reach a specified absorbing boundary, then at nonabsorbing states,\n\n\\[-1= \\lambda_n(T_{n+1}-T_n) +\\mu_n(T_{n-1}-T_n),\\]\n\nwith zero boundary values at the target. Such equations may also be derived through stopping-time and martingale methods."
    },
    {
      "heading": "Explosion and boundary behavior",
      "text": "Finite rates at every state do not, by themselves, guarantee that the process is defined on ordinary integer states for all time. Explosion occurs when infinitely many jumps take place within a finite interval.\n\nFor a pure birth process with positive rates, the time required to pass through all states is a sum of independent exponential holding times. The standard nonexplosion criterion is\n\n\\[\\sum_{n=0}^{\\infty}\\frac{1}{\\lambda_n}=\\infty.\\]\n\nFor example, rates \\(\\lambda_n=\\lambda(n+1)\\) are nonexplosive, whereas \\(\\lambda_n=(n+1)^2\\) produce explosion in finite time with probability one.\n\nFor general birth–death rates, the possibility of downward jumps makes the criteria more involved. A useful sufficient condition is an at-most-linear bound\n\n\\[\\lambda_n+\\mu_n\\leq C(1+n)\\]\n\nfor some finite constant \\(C\\). Finite-state models are automatically nonexplosive when all rates are finite.\n\nThese distinctions matter because, in explosive cases, specifying local rates alone may not determine a unique conservative process without additional rules for behavior at infinity."
    },
    {
      "heading": "Computation, applications, and limitations",
      "text": "Exact sample paths can be generated by repeatedly drawing an exponential holding time and then choosing the jump direction according to the current rates. Repetition gives a Monte Carlo method for estimating transient probabilities, extinction probabilities, and passage times.\n\nFor finite or truncated systems, forward equations can be solved using methods from numerical analysis. Truncation requires care: an artificial upper boundary can distort results if appreciable probability reaches it.\n\nBirth–death models are useful when events change a count by one and the count alone adequately determines future rates. Applications include population dynamics, queue lengths, reliability systems that count working components, and molecular production and degradation. Their main limitations follow from the same assumptions:\n\n- Batch arrivals, multiple simultaneous losses, and reproduction events producing several offspring require jumps larger than one.\n- Nonexponential waiting times generally destroy the Markov property of the count alone.\n- Age, spatial structure, or hidden environmental conditions may require an enlarged state space.\n- One-dimensional models cannot separately track interacting populations or multiple chemical species.\n\nA birth–death description is therefore not a claim that individual events are intrinsically memoryless. It is a modeling choice whose suitability depends on whether the selected state contains enough information to describe the dynamics."
    }
  ],
  "links": [
    "continuous-time-markov-chain",
    "markov-property",
    "exponential-distribution",
    "random-walk",
    "matrix",
    "probability-distribution",
    "transition-matrix",
    "matrix-exponential",
    "expected-value",
    "variance",
    "stationary-distribution",
    "detailed-balance",
    "poisson-process",
    "binomial-distribution",
    "poisson-distribution",
    "recurrence-relation",
    "martingale",
    "monte-carlo-method",
    "numerical-analysis"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}