Motivation
A finite number of states supplies a common upper bound on jump rates. An infinite model may have no such bound: as a population grows, its event clock can run faster. The question is no longer only where the next jump goes, but whether the successive holding times still add up to an infinite lifetime. A matrix with finite entries does not answer that question.
The minimal construction follows the specified clocks until they accumulate. It makes no attempt to decide what happens after infinitely many jumps. This is a useful discipline: an extra state at infinity, a restart rule or an entrance boundary would be additional modeling information. We first build the process actually determined by the local rates, and then prove when that process has no finite accumulation time.
We use independent exponential variables, the strong law, nonnegative convergence theorems and elementary finite-state jump conditioning.
Definitions
Definition 3.2.1
(Stable conservative rate matrix and minimal lifetime)
Let be countable. A rate matrix is stable and conservative here when
At a nonabsorbing state , its embedded chain chooses with probability . With independent unit exponentials , its holding time on visit to state is . The jump times are and the successive holding-time sums. Write for the explosion time. An absorbing state has infinite remaining lifetime.
Before , set on . After explosion, send the minimal process to a cemetery state outside . Its transition probabilities within are
They can have row sums below one. Non-explosion means for every initial state. Conservation of the rate matrix concerns its rows; conservation of the transition probabilities concerns survival of the constructed paths. These are different assertions.
A proper Lyapunov function is a map with finite sublevel sets for every finite . Where , define
Properness makes escaping every bounded-value region equivalent to escaping every fixed finite collection of states along a suitable exhaustion. A function that is constant on an infinite state space does not satisfy it.
Theorem/proposition
Theorem 3.2.2
(Minimal first-jump equation and uniqueness before explosion)
For a stable conservative matrix, the construction above is well-defined before explosion. Its transitions are the minimal nonnegative solution of
If the constructed process is non-explosive, it is the unique non-explosive pure-jump Markov chain with these holding rates and jump destinations. This is a path-law statement; it does not assert uniqueness of every unrestricted formal infinite differential system.
Theorem 3.2.3
(A proper-function non-explosion criterion)
Suppose a proper satisfies for each , and for a constant . Then the minimal chain is non-explosive. More precisely, for and ,
Theorem 3.2.4
(Exact pure-birth explosion test)
For a pure-birth chain on the nonnegative integers with , starting from zero,
It is finite almost surely exactly when ; otherwise it is infinite almost surely. Removing a finite initial segment gives the same classification from any finite starting state.
Proof sketch or proof idea
What the local clocks determine
Proof
(First-jump conditioning and minimality)
Every finite initial segment of the construction consists of finitely many valid exponential draws and destination choices. Independent fresh choices and exponential memorylessness give the Markov restart property before explosion. Condition on the first holding time: either it lasts beyond , or it ends at time with destination . Its destination-weighted density is . This gives the first-jump equation, including absorbing rows where the integral is zero.
Iterate the equation starting with the no-jump term. The successive nonnegative partial sums describe paths ending at by time with at most a prescribed number of jumps. Their increasing limit is exactly the minimal probability, because a surviving path has finitely many jumps by that time. Any other nonnegative solution is at least every such partial sum, proving minimality. A pure-jump chain with the specified clocks has the same law for every finite jump sequence. If explosion has probability zero, these finite sequences exhaust every bounded observation interval and determine the whole path law. When explosion can occur, those sequences leave the post-explosion behavior unspecified.
Stopping before the dangerous region
Proof
(Localized drift bound and exhaustion)
Inside there are only finitely many states, so their departure rates are bounded. Infinitely many jumps cannot occur while the path stays there: one may dominate their jump count by a Poisson clock with a finite common rate. Thus the process is unambiguously defined through its first exit, even if its unrestricted lifetime has not yet been established.
Stopped finite-region jump conditioning gives
Here the exit state is included, so overshoots of are not silently replaced by . To justify this identity, sum the changes of at jumps before and including exit. The exponential density at each pre-exit state converts its expected jump increments into the time integral of . The absolute expected jump-change rate is bounded on the finite pre-exit region, because each row's -weighted sum is finite. Integrating over a bounded horizon therefore justifies summation and expectation, even when the exit value has no uniform bound.
Use and nonnegativity to bound the stopped expectation by . Gronwall's inequality gives . On the stopped value is at least , proving the probability bound. Gronwall follows, for example, by iterating the integral inequality into its exponential series; no probabilistic theorem is hidden there.
Almost surely, explosion in a finite interval requires leaving every finite sublevel region: otherwise infinitely many jumps would occur among finitely many bounded-rate states. Hence, up to a null set, is contained in every exit event . Sending to infinity makes its probability zero. Taking a countable union over positive integer horizons proves non-explosion for all finite times.
Proof X-ray
Localize before using the generator
Each hypothesis of the criterion enters at one point of the argument.
- Properness makes every region finite. Rates are bounded there, so the path is well defined up to the exit time , and in the last step explosion forces exit from every such region.
- Finiteness of the weighted row sums makes finite and controls the expected jump increments, including the overshoot at the exit jump.
- The drift bound feeds Gronwall's inequality, giving .
The key move is to prove the generator identity only for the stopped process; using it for the unrestricted process first would assume the survival property under investigation. The argument closes with Markov's inequality: the stopped value is at least on , and for every , so .
Why divergent mean holding sum is sufficient for pure birth
Proof
(Expectation in one direction, a Laplace product in the other)
If , monotone convergence gives , hence almost surely. The converse is not justified by infinite expectation alone: a finite random variable can have infinite expectation. Independence supplies the needed extra argument.
For , bounded convergence of and exponential integration give
where . If diverges, then diverges: either infinitely often, giving infinitely many terms greater than , or eventually , where applies. The product is therefore zero. A nonnegative variable has expectation zero only if it is zero almost surely, so almost surely. This proves the exact dichotomy.
Concept lensAnalytic
Explosion as a solution of a linear equation
Fix and put . Conditioning on the first jump gives , that is, with . Any other solution satisfies for every , hence . So the chain is non-explosive exactly when has no nonzero solution with values in . For pure birth the equation reads , and a bounded nonzero solution exists exactly when , recovering the test above.
What a finite approximation can legitimately promise
The exit bound supplies a precise use for truncation. Fix a finite sublevel region and construct a finite model that follows the original rates while inside it, sending every outward jump to an absorbing cemetery. The outward rate from each retained state is the sum of its rates to omitted states. This aggregate is finite, so the stopped model is a legitimate finite-state chain. It preserves survival inside the region instead of reflecting omitted transitions back into the model.
Couple the finite model and the original construction by using identical clocks and destinations until the first exit. At a specified observation time, their states agree unless an exit has occurred. For any event determined by the state at that time, their probability difference is therefore at most the exit probability. The Lyapunov bound gives the explicit estimate
for sets of retained states. Here denotes the killed finite approximation, and the original process is now known to be non-explosive by the theorem. A return from outside the region can contribute to the original probability but not the killed approximation; the same exit bound covers that possibility. There is no need to pretend the two models have identical boundary behavior after separation.
This estimate is a finite-horizon guarantee. For a fixed truncation level, its upper bound grows with the observation time and eventually becomes uninformative. It does not justify passing directly to a stationary distribution or interchanging a large-time limit with a large-region limit. Choosing the horizon first and then the region is part of the argument. This is the practical meaning of localizing a generator: compute within a controlled region, and account explicitly for the probability of leaving it.
Worked examples
Worked example 3.2.5
(Linear birth and death rates are safe)
Let and , with and . Rates are unbounded, so one global finite uniformization clock is unavailable. Take . Nearest-neighbor jumps make every weighted row sum finite, and
The criterion proves non-explosion from every finite state, including the pure-birth case . Birth can exceed death without causing explosion; exponential growth over long times is different from infinitely many jumps before a finite time.
Worked example 3.2.6
(Polynomial birth clocks and the threshold)
For with , the holding-mean series is . The process explodes almost surely for and is non-explosive for . At the boundary the harmonic series diverges. For , the expected explosion time is the finite square-reciprocal sum, but identifying its full distribution is unnecessary for the classification.
Worked example 3.2.7
(Why infinite expectation is not a proof)
A positive integer variable with probabilities proportional to is finite almost surely but has infinite expectation, since diverges. Thus the sentence “the total expected waiting time is infinite, so explosion is impossible” is invalid without further reasoning. For independent pure-birth holding times the Laplace-product proof supplies that reasoning; the elementary implication from finite expectation only works in the opposite direction.
Common mistakes
Common mistake
(Reading conservative rows as conservation of probability)
The first-jump equation is not an instruction to invert an arbitrary infinite matrix numerically. Infinite sums and boundary behavior carry mathematical content. Finite truncations provide a different model until their boundary error has been controlled. A killed minimal process can lose mass through explosion even though every row of sums to zero.
Common mistake
(Treating the Lyapunov test as a characterization)
The Lyapunov test is sufficient, not necessary, and failure to find a suitable function does not prove explosion. Its properness and weighted-row integrability assumptions are part of the theorem.
Common mistake
(Applying the pure-birth series to birth–death chains)
The pure-birth series criterion cannot be applied unchanged to an arbitrary birth–death chain: revisits and downward moves change the sequence of clocks actually used. The later birth–death unit separates the embedded path classification from lifetime and equilibrium.
Properness is the hypothesis most easily overlooked, because the drift inequality alone looks like the whole condition.
Counterexample mode
A drift bound without properness proves nothing
Claim: any with finite weighted row sums and certifies non-explosion. Take the pure-birth chain with and . Every weighted row sum equals , and . Yet , so the chain explodes almost surely. The proof breaks at localization: for the region is the whole infinite space, its rates are unbounded, and the exit event never occurs, so the bound says nothing about . A proper choice such as gives , which no bound dominates; the criterion then correctly gives no conclusion.
Summary
Finite local rates construct each finite segment but do not guarantee infinite lifetime. The minimal first-jump equation records the law before explosion. Proper-function bounds rule out escape through faster and faster states in finite time. Pure birth admits an exact holding-mean-series test, whose divergent direction requires independence rather than an expectation slogan.
Exercises
Quick check
(Starting from zero, classify the pure-birth rates and . Find the expected explosion time in the explosive case.)
Apply the exact pure-birth test to .
Solution · Answer
diverges, so the first process is non-explosive. is finite, so the second explodes almost surely, with expected lifetime .
Exercise 3.2.8
(A nearest-neighbor chain has and arbitrary finite death rates , with . Prove non-explosion using a proper function.)
CoreProof
Solution(3)
- 1
Choose the function
, and each sublevel set is finite, so is proper. Each row has at most two neighbors, so the weighted row sums are finite.
- 2
Compute the drift
Key step
, and at zero .
- 3
Apply the criterion
∎The proper-function theorem applies with . Large downward rates only lower the drift, so they cannot spoil the bound.
Exercise 3.2.9
(For a pure-birth process of constant rate , derive and from the first-jump equation.)
Foundation
Solution(3)
- 1
Use the no-jump term at zero
Pure birth never returns to zero, so the integral term for vanishes and . In the same way .
- 2
Condition on the first jump
Because: The only destination from zero is one.
- 3
Evaluate the integral
The integrand is the constant , so .
Quick check
(A suitable function has and . Give an upper bound for exit from by time one, and say whether the bound alone shows that exit is impossible.)
Substitute into .
Solution · Answer
The bound is . It bounds a genuine finite-level exit probability and is not zero. Non-explosion follows only after sending the level to infinity, for each fixed time horizon.