A population selected by survival
Suppose a system can be absorbed into failure. Among all systems started in a given state, look only at those still operating after steps. Their distribution need not be the original distribution: states that tend to fail quickly become underrepresented. A distribution that reproduces itself under this selection is called quasi-stationary. It is not an ordinary stationary distribution, since probability is continuously lost to absorption before normalization.
We work with a finite positive transient block. The strict positivity assumption is stronger than irreducibility, but it allows a direct quantitative proof using a coupling bound. We use the finite-dimensional Brouwer fixed-point theorem to construct positive eigenvectors.
Substochastic evolution
Definition 2.6.1
(Survival and quasi-stationarity)
Let , and let have entries and row sums at most one, with at least one row sum strictly below one. Add a cemetery state with transition probability from and with absorbing. Let . For an initial row probability on ,
A row probability on is quasi-stationary if for every and for which the conditioning is defined.
Before conditioning on survival we check that survival is eventually rare, so that the conditioning is a genuine selection rather than a formality.
Lemma 2.6.2
(Absorption is almost sure)
Under the assumptions of the definition, there is such that for every initial probability and every . In particular , while for every finite .
Proof
(Two steps always carry a fixed chance of absorption)
Set and choose with . From any state , the chain moves to with probability and is then absorbed with probability , so absorption within two steps has probability at least . By the Markov property at times , each further pair of steps is survived with conditional probability at most , which gives ; letting gives almost sure absorption. Every row sum is positive because the entries are positive, so and the conditioning in the definition is always defined.
Almost sure absorption does not prevent the conditional distribution of the survivors from having a meaningful limit; that limit is the subject of the next theorem.
The positive-block theorem
Theorem 2.6.3
(Unique survival-selected law)
Under the assumptions above, there are strictly positive vectors and , a number , and such that
The probability is the unique quasi-stationary distribution. For every initial probability on ,
Starting from , survival is exactly geometric in discrete time: . From an arbitrary start the limiting ratio need not be the exact one-step survival probability at finite times.
Proof
(Eigenvectors, an h-transform and a contraction)
On the probability simplex, the maps and are continuous and map the simplex into itself. Brouwer's theorem gives fixed points and , with positive eigenvalues and . Because ,
so . Rescale to sum to one. Since and at least one inequality is strict, ; positivity gives .
Now define the transformed matrix
Its entries are strictly positive and its rows sum to one. Put and . Any two rows of share a common mass at least . Coupling that common mass shows that applying contracts the total-variation distance between probability rows by at most . More explicitly, subtracting the same uniform subprobability of mass leaves probability kernels of mass , so the total variation of the difference contracts by . Thus has a unique stationary probability and for every probability : successive iterates form a Cauchy sequence by the same contraction.
By induction . Put and . Then
After normalization, the conditional distribution tends to the probability proportional to . But is stationary for , so the limit is . Summing the displayed identity gives the survival ratio limit, since the positive finite sum converges to a positive limit. If a law is quasi-stationary, its conditional law at every equals ; the just-proved limit forces . Finally proves the exact geometric survival formula.
Proof X-ray
Where positivity and the defect in the row sums are used
The proof has three moves, and each hypothesis enters at a definite place.
- Strict positivity of makes the two normalized maps well defined on the simplex (the denominators never vanish) and makes the Brouwer fixed points strictly positive. It is used a second time, more subtly, in : without it the contraction constant could equal one.
- The row-sum defect is used only once, to get . Without it, and would be an ordinary stationary law.
- The key move is the h-transform . It converts the mass-losing matrix into a stochastic matrix, and the identity shows that the survivors are the -chain reweighted by , with the geometric factor canceling on normalization.
The argument closes because convergence for the stochastic matrix is a standard contraction, and the identification turns its stationary law back into the quasi-stationary one. The matrix is a proof device with a different path law; it is not in general the transition matrix obtained by conditioning only on survival for one more step.
Worked examples
Worked example 2.6.4
(The initial survival rate is not the limiting rate)
Take
The first row survives one step with probability and the second with probability . The rows are proportional, and direct multiplication gives
Starting in state 1, the first-step survivor distribution is and remains so. Hence for ,
The initial survival probability differs from the later conditional survival probability . Starting from instead gives survival even at the first step.
The next example is the opposite extreme: equal row sums make survival exactly geometric from every start, while the conditional law still has to converge.
Worked example 2.6.5
(Equal row sums: geometric survival from every start)
Take
Both row sums equal , so . Hence , , and by symmetry ; indeed . Every start then has exactly, so the survival ratio equals at all times. Because is constant, and :
The formula holds at and follows by induction from . So from state 1 the survivors are in state 1 with probability , which tends to geometrically at rate . Here and , so the contraction bound of the proof is attained.
Survival probabilities also determine mean absorption times, through the tail-sum formula .
Worked example 2.6.6
(Mean absorption times for the two operating conditions)
Return to . From state 1, the first example gives and for , so
From state 2 the first survivor law is again , so for and . From itself the absorption time is geometric, with .
Check by first-step analysis: the vector solves , that is . With , , as required.
The positivity hypothesis in the theorem cannot simply be weakened to irreducibility, as the following periodic block shows.
Counterexample mode
Irreducibility without aperiodicity is insufficient
Let . It is irreducible and absorption occurs almost surely. The quasi-stationary law has survival factor , but from state 1 the survivor is certainly in state 1 at even times and state 2 at odd times. The conditional distributions do not converge. Strict positivity rules out this periodic oscillation; it is not a cosmetic hypothesis in the stated proof.
Common mistakes
Common mistake
(Treating the quasi-stationary law as a stationary law)
Ordinary stationarity is the equation for a probability-preserving matrix. Here loses mass at every step, and it is the conditioning on that restores total mass one. The unconditioned law of started from is on plus mass at , which converges to the point mass at .
Common mistake
(Reading survivor convergence as survival of a path)
Convergence of the survivors' distribution says nothing about a typical trajectory surviving forever: under the lemma above every trajectory is absorbed almost surely. The limit describes the shrinking subpopulation that is still alive at time , not any infinite path.
Common mistake
(Using the first-step survival probability as the long-run rate)
From a fixed state the first step is survived with probability , which in general differs from . In the first worked example while . Starting from , or having equal row sums, makes every finite-time survival ratio equal to ; from a general start the ratios only approach .
Exercises
Quick check
(For in the first worked example, compute the first survivor distribution from state 2 and .)
Normalize the second row of .
Solution · Answer
The second row is with sum ; after normalization it is . From then on each step is survived with probability , so .
Exercise 2.6.7
(Suppose with a probability on and . Prove directly that is quasi-stationary, and find for .)
FoundationProof
Solution(3)
- 1
Iterate the eigen-equation
By induction, , so for every .
- 2
Normalize the surviving mass
, and . Thus is quasi-stationary.
Because: and .
- 3
Difference the survival tail
∎For , , a geometric law.
Quick check
(For the periodic block , calculate and the conditional distribution at time .)
Each row has sum .
Solution · Answer
Every step is survived with probability whatever the current state, hence . A surviving path alternates deterministically, so given survival the state is 1 for even and 2 for odd ; the conditional distribution is or and does not converge.
Quick check
(For a one-state block with , find , , , and the expected absorption time.)
In one dimension every eigenvector is a positive scalar.
Solution · Answer
Choose and ; then and . Since , the tail-sum formula gives .