Motivation
A nearest-neighbor process carries two kinds of information. The ratios of upward to downward moves determine whether its path returns or escapes. The amount of time spent on each visit determines whether returning paths possess a finite mean cycle and an equilibrium probability law. Keeping these questions separate is essential when jump rates grow with the state.
The stationary product weights encountered in bounded-rate examples are only half of the story. A second series, built from edge resistances, controls escape. These two series can behave differently, and a formal summable invariant vector can even accompany an explosive process. The point of this unit is to derive the series tests and identify exactly when they describe a genuine non-explosive chain.
The resistance terminology is a mathematical interpretation of explicit positive quantities. General networks, diffusions and entrance-boundary classifications are outside this unit.
Definitions
Consider states with finite rates for all , for , and . State zero reflects upward rather than absorbing. The embedded chain moves upward with probability and downward with the complementary probability. Define
The weights measure prospective time occupation, is an edge conductance and its resistance. These are mathematical names for explicit positive quantities; no physical unit convention is needed. The relation is built into the definitions.
Recurrence here refers to returning after a genuine departure, equivalently recurrence of the embedded chain. A recurrent continuous-time chain is positive recurrent if its mean cycle from entry into zero through the next entry into zero is finite, including the initial holding time at zero. It is null recurrent if that mean is infinite. This cycle convention avoids the meaningless “first positive return time” for a path that stays at zero throughout a holding interval.
For finite boundaries let denote the first time state is reached. Before hitting either of two finite boundaries, only finitely many states are visited, so rates there are bounded and these hitting calculations require no global non-explosion assumption. They can equally be performed on the embedded chain.
Theorem/proposition
Theorem 3.3.1
(Resistance formula and recurrence)
For integers and ,
with boundary values zero at and one at , and empty numerator zero. The embedded chain is recurrent exactly when . If , its probability of escaping from one without hitting zero is . If , the minimal continuous-time process is automatically non-explosive.
Theorem 3.3.2
(Stationary and mean-cycle classification)
Assume the minimal process is non-explosive. It has a stationary probability exactly when , in which case it is unique and
Consequently: gives positive recurrence; gives null recurrence; gives transience and, under non-explosion, necessarily . Without non-explosion, summability alone is insufficient for stationary probability.
Proof sketch or proof idea
Finite boundaries and escape to infinity
Proof
(Finite-boundary differences, escape and non-explosion)
Let . First-jump conditioning at an interior state gives . Since , successive differences have the form . Sum the differences and impose to obtain the formula. The finite boundary problem terminates almost surely: on its finite interior there is a common positive probability of reaching a boundary within a fixed number of embedded steps. A geometric survival bound proves termination and uniqueness of the recurrence solution.
Starting from one, the events of hitting before zero decrease as increases. Their intersection is the event of never hitting zero, up to a null set: a path avoiding zero cannot stay forever in a fixed finite interval, by the same geometric argument. The formula therefore gives escape probability , interpreted as zero when . Departure from zero is always to one, so eventual return to zero is certain exactly when . Irreducibility spreads recurrence to every state.
In the recurrent case the embedded chain returns to zero infinitely many times. Every visit supplies an independent exponential holding time of mean . Their sum diverges almost surely by the strong law. Thus the full sum of holding times cannot accumulate in finite time. From another state, the chain first hits zero in finitely many embedded steps almost surely, establishing non-explosion from every state. This is a useful case where a path classification itself settles the lifetime question.
Reversal proves invariance without bounded rates
Proof
(Path reversal gives invariance)
The product weights satisfy detailed balance at the rate level. To pass from rates to transition probabilities with unbounded rates, examine a path with finitely many jumps within time . Its density is the product of the forward jump rates and the exponential survival factors for the successive holding durations, including the final no-jump duration. Multiplying by and applying edgewise detailed balance telescopes to times the product of reverse rates. Reverse both the order of visited states and the order of durations. The exponential factors are identical, and the time-simplex transformation preserves volume. Integrating and summing over finite paths gives
All summands are nonnegative, so no absolute-convergence assumption is being concealed. If and the process is non-explosive, divide by and sum over . The right side becomes , since row has total mass one. This proves stationarity for the actual process. In an explosive minimal process that row sum can be smaller than one; the missing mass is exactly the obstruction to this argument.
Return-cycle occupation and necessity
Proof
(Stationarity forces a finite mean cycle)
A stationary non-explosive law makes the discrete skeleton stationary for every fixed . This skeleton is irreducible: any finite nearest-neighbor path has positive probability of being completed within , followed by no further jump before . The countable-chain stationary theorem therefore makes it recurrent. In particular the continuous-time path returns to zero after departures; otherwise it could not revisit zero at arbitrarily large skeleton times. Thus a stationary law implies recurrence of the embedded chain.
In a recurrent chain let be expected total time in state during one complete zero-to-zero cycle. There is exactly one initial holding at zero, so . For each other fixed state, the number of visits before returning to zero has a geometric upper tail: on each visit there is a fixed positive chance of following a descending path to zero before revisiting that state. Its expected occupation is therefore finite. Every completed excursion crosses each edge equally often in both directions. Since a visit to lasts on average and departs upward with probability , expected upward crossings equal ; downward crossings equal . Hence
To finish necessity without assuming a finite mean cycle, suppose that mean were infinite. The number of complete cycles per unit time would tend to zero: truncate their independent identically distributed lengths at any fixed level, apply the strong law, then let the level increase. Their cumulative time at zero, including an upper bound for any unfinished zero holding, is bounded by the sum of the first one-more-than-completed independent exponential holds at zero. Its time fraction therefore tends to zero, because the exponential average tends to while the number of those holds divided by elapsed time tends to zero. A finite initial segment from another state changes no limit.
But for a stationary initial law , Fubini gives
The bracketed variables lie between zero and one. Bounded convergence would make their expectations tend to zero if the cycle mean were infinite, a contradiction. Thus the mean is finite and .
For finite mean cycles, the continuous renewal-reward argument gives occupation fractions . Applied to a stationary initial law, bounded convergence identifies that law with these fractions, proving uniqueness and the cycle identity. This also establishes the three classification cases and the impossibility of finite with finite in a non-explosive model.
Proof X-ray
Where non-explosion and recurrence enter the cycle argument
The proof of the stationary classification uses its hypotheses in a fixed order.
- Non-explosion is used twice: in the reversal identity, to replace the row sum by one, and in the skeleton argument, where the stationary law must belong to a process whose rows have total mass one at every time.
- Recurrence is not assumed. It is derived from stationarity through the irreducible skeleton, and only then are complete zero-to-zero cycles available.
- The key move compares two descriptions of the fraction of time spent at zero. Along paths it is controlled by the cycle lengths and tends to zero when the mean cycle is infinite; under a stationary start its expectation equals for every . Bounded convergence makes the two descriptions incompatible.
- Edge-crossing balance then identifies the occupation vector as , so a finite mean cycle is the same condition as .
The argument closes with renewal-reward: the occupation fractions are , and any stationary law must coincide with them, which gives uniqueness.
The three proofs answer three separate questions, and the independence of those questions is worth isolating.
Concept lensStructural
Paths and clocks answer different questions
The resistance series decides whether the embedded path returns. The occupation-weight series decides whether a recurrent path has a finite mean cycle. A lifetime check links these formal series to a process defined at every finite time. Solving one of these problems does not automatically solve the other two.
Worked examples
Worked example 3.3.3
(Constant rates and the critical boundary)
Let for and for . Rates are bounded, hence non-explosive. Put . Then and . For , and , giving geometric equilibrium . At , both series diverge and the chain is null recurrent. For , is finite and the chain is transient. Its escape probability from one is , not one: a transient chain can still return with substantial probability.
Worked example 3.3.4
(Linear unbounded rates with the same equilibrium weights)
Take and with . The preceding unit's criterion proves non-explosion. Again , now with , but
For the chain is positive recurrent with the same geometric stationary law as a constant-rate model. At equality the resistance series is harmonic and diverges, giving null recurrence. For it converges, giving transience. Equal stationary weights in the stable case do not make the transition speeds or finite-time laws equal.
Worked example 3.3.5
(Summable weights in an explosive model)
Let and for . Then and , but , so . The embedded chain has upward probability at positive states, and reflects at zero. It is transient. In fact it is explosive, as the following finite expected-lifetime computation proves. Starting from zero it reaches each finite level almost surely, allowing the visit counts to be computed. Expected total visits are two at zero and three at each positive state. For a positive state, the return probability is ; from below it is reached surely, and from above the biased walk returns with probability . Thus
The lifetime is finite almost surely. Normalized weights satisfy formal rate balance but are not stationary for the killed minimal process, which loses mass. This is why non-explosion cannot be dropped from the stationary theorem.
Common mistakes
Common mistake
(Changing the boundary changes the equilibrium question)
At zero the death rate is zero, and this reflecting boundary determines the zero-current product weights. An absorbing zero would instead support a point-mass equilibrium and lead to a different extinction problem. Finite upper truncation introduces another boundary; it can normalize weights even when the infinite model has no stationary probability.
Common mistake
(Using the discrete Kac formula in continuous time)
Continuous-time Kac factors include the departure rate. The mean zero-to-zero cycle is , not time units. The latter reciprocal is a discrete-step return formula.
Common mistake
(Reading the recurrence class from a finite simulation)
A null-recurrent chain returns almost surely but has an infinite mean continuous cycle; a transient chain can make finitely many returns before escaping. Neither classification can be read off from a numerical simulation over a finite horizon.
The resistance theorem proves non-explosion only in the recurrent direction, and the converse fails even for bounded rates.
Counterexample mode
Transience does not imply explosion
Claim: if , the chain escapes to infinity, so it must explode. Take constant rates for and for . Then , and , so the chain is transient and from one it avoids zero forever with probability . Yet every holding rate is at most , so the jumps can be obtained by thinning a single Poisson clock of rate , which has finitely many marks in every bounded interval: the process is non-explosive. Escape concerns the order of visited states; explosion concerns the sum of holding times. The correct statement is one-directional: forces non-explosion, while for the lifetime must be examined separately, as in the third worked example.
Summary
Finite-boundary differences produce the resistance hitting formula. Infinite resistance forces recurrence and, through repeated positive holding times at zero, non-explosion. Under non-explosion, finite total occupation weight is equivalent to stationary probability and finite mean cycle. Path reversal justifies invariance even with unbounded rates; explosive examples show why formal rate balance alone is insufficient.
Exercises
Exercise 3.3.6
(For constant rates and , find the stationary law, the mean zero-to-zero cycle and the probability of hitting four before zero from one.)
Foundation
Solution(3)
- 1
Compute the weights and the stationary law
and , so and .
Because: Rates are bounded, so the process is non-explosive and the stationary theorem applies.
- 2
Apply the continuous-time cycle formula
The mean cycle is .
- 3
Apply the resistance formula with ,
Resistances are proportional to , so .
Quick check
(For constant rates and , find the chance of ever hitting zero from one. Is a stationary probability available?)
Compute the escape probability first.
Solution · Answer
Here , and the escape probability is , so the hit-zero probability is . The chain is transient; rates are bounded and , so no stationary probability exists.
Quick check
(With and , classify the chain. Explain why a finite uniform bound on the departure rates is not required.)
Compute and explicitly.
Solution · Answer
Each factor equals one, so and ; both series diverge and the chain is null recurrent. The proper-function test gives for , and the infinite-resistance theorem gives non-explosion directly, so no uniform rate bound is needed.
Exercise 3.3.7
(Multiply every rate of a non-explosive positive recurrent birth–death model by . Describe the changes in the stationary law, the recurrence classification, the mean cycle and the finite-boundary hitting probabilities.)
Core
Solution(3)
- 1
Weights are ratios of rates
Each factor is unchanged, so , and the stationary law are unchanged.
- 2
Resistances scale uniformly
Conductances multiply by and resistances divide by . The series diverges or converges exactly as before, and the hitting ratios in the resistance formula are unchanged.
- 3
Only the clock changes
The mean cycle divides by . The scaled process is the original one run at speed , so non-explosion and positive recurrence are preserved.