Which interval does an observer see?
Suppose a device is replaced after each failure, and its successive lifetimes are independent copies of a positive random variable . An engineer who records every replacement sees the lifetime law of . An engineer who visits at an unrelated time is more likely to find the device inside a long lifetime. Moreover, the visitor sees only the unexpired part of that lifetime. Neither operation is the same as drawing a fresh .
This distinction matters even when the mean replacement rate is already known. The renewal strong law gives one replacement per units of time on average. It does not say how long a visitor must wait. We shall derive that waiting law and build a process for which it holds at every deterministic observation time. The construction will also explain why an ordinary process started with a replacement at time zero need not approach equilibrium at each fixed time.
We use the earlier renewal strong law, the strong law for independent identically distributed integrable variables, Tonelli's theorem and dominated convergence. No general key renewal theorem is assumed or proved.
Age, residual life and the interval containing time
Definition 1.6.1
(Ordinary renewal process and its two clocks)
Let be independent copies of , where almost surely and . Write , and
is the age, is the residual life until the strictly next renewal, and is the containing interval's length. At a renewal time, our convention gives age zero and a new positive residual life.
The epochs diverge almost surely, so these variables are defined at every finite time on an event of probability one. On the interior of a cycle of length , age increases from zero to , residual life decreases from to zero, and their sum stays . The countable set of renewal endpoints has Lebesgue measure zero, so our endpoint convention does not affect an integral over observation time.
Put for . The tail-integral identity gives . We write .
Completing the unfinished-cycle argument
The completed-cycle reward theorem needs one extra argument before it can be used for a quantity observed continuously within a cycle. That argument is precisely where integrability belongs.
Theorem 1.6.2
(A cycle formula for continuous-time averages)
Let be measurable and suppose
Then, for this fixed , almost surely
Every bounded measurable satisfies the assumption. No nonarithmetic assumption on is required. For an unbounded , this pathwise conclusion alone asserts no convergence of expectations.
Proof
(Complete cycles plus one controlled remainder)
Define the cycle rewards and the absolute bounds . These are independent identically distributed pairs, with . Decompose the time integral into the first complete cycles and one unfinished cycle. If its signed contribution is , then
The reward strong law and give the desired limit for the completed sum. To control the last term, put . The strong law yields , and hence
The expression is used for . Since ,
No independence between the random index and its reward is needed: the sequence limit holds for all sufficiently large indices on the same probability-one event. Finally, if , then is integrable.
Proof X-ray
A small boundary in time can carry a large reward
The unfinished cycle is a single interval, but its reward is not automatically negligible. The proof bounds it by a whole-cycle absolute reward and uses its finite expectation to obtain . For residual life itself, the cycle reward is , so a second moment is needed for the finite reward formula. Calling the unfinished cycle “only one term” would conceal the essential hypothesis.
Apply the theorem to , for fixed . During a cycle of length , this indicator is one for a set of times of length . Therefore
almost surely. This is a statement about the fraction of observed time, not the limit of the indicator at one moving endpoint .
The probability law behind the cycle formula
Definition 1.6.3
(Length bias and equilibrium age–residual law)
Define a length-biased random variable by
for Borel sets . Let be uniform on and independent of . Set
This is a probability law because the length-bias numerator equals when . Conditional on the containing length, the observation location is uniform within that interval. For every bounded measurable , conditioning first on gives
The factor from length bias cancels the factor from the uniform location. In particular, the joint survival and marginal survival are
Even if has atoms, has an absolutely continuous distribution with density almost everywhere on . The continuous observation location smooths the discrete interval choices. Equal marginal laws for age and residual life do not imply independence.
There is also an expectation statement about an explicitly randomized observation. Let be uniform on and independent of the entire ordinary renewal path. For bounded , conditional expectation given that path is exactly its time average. Dominated convergence, using the fixed bound on , consequently proves
This supplies a justified expectation limit; it is not an interchange of a strong law with expectation for an arbitrary unbounded reward. Each assertion concerns the specified test function .
Constructing equilibrium at every observation time
The equilibrium pair can be made the origin of a whole process. Sample , put renewal points at and , and extend in both directions using fresh independent gaps with law . The left gaps, right gaps and are mutually independent. Positive independent identically distributed gaps have sums diverging almost surely in both directions, so this yields a locally finite set of renewal points on the whole real line.
Theorem 1.6.4
(The length-biased construction is stationary)
The law of this whole-line renewal point set is unchanged when every point is shifted by any fixed real amount. Thus its age–residual pair has law at every deterministic time. Restricted to future points, its first waiting time has law ; all subsequent gaps are independent with law and independent of that first waiting time.
Proof
(Shift invariance by reindexing the gaps)
We first describe a different, auxiliary point set with a renewal point pinned at zero. On both sides use independent gaps with law . Denote expectation for this auxiliary construction by , and let be its gap immediately to the right of zero. For a point set , define : the observer moves forward by .
For a bounded measurable function of the whole point set, the length-biased/uniform construction has expectation
The identity follows from the same cancellation of length bias and uniform location used above; the other gaps keep their independent original laws. Its right side is finite and sends the constant function one to one.
Moving to its next point, , merely reindexes a two-sided independent gap sequence. It therefore has the same auxiliary law as . Fix . Subtracting the integrals over and gives
The integral identity holds even when ; it is just subtraction of endpoint integrals. We use the usual -field on locally finite point sets, generated by their counts in bounded Borel sets. The shift map is jointly measurable on this space; together with boundedness this permits Fubini over the fixed interval . Reindexing shows that the two integrands have equal expectations for each . Invariance for negative follows from positive-shift invariance by applying it to when . Taking to be event indicators proves equality of the point-set laws under all deterministic shifts.
At the origin the containing pair is the one sampled in the construction; stationarity gives that same distribution at any deterministic observation time. Future gaps after were sampled independently from the start, which proves the last assertion.
Concept lensStructural
A renewal origin and an observation origin
Pinning a renewal at zero produces the original law for the next gap, but gives a distinguished time origin. Choosing a typical time instead produces for the first wait and length bias for its containing interval. The stationary construction randomizes this origin correctly. It does not turn all future gaps into copies of : only the initial wait is exceptional.
Calculations: long intervals are overrepresented
Worked example 1.6.5
(Rare long lifetimes dominate what the visitor sees)
Let and . Then and . The containing interval at equilibrium has probabilities
Only one tenth of replacements begin a long lifetime, but more than half of observation time lies in long lifetimes. The equilibrium residual survival is
Thus , while a fresh lifetime exceeds four with probability . Also
For the last calculation, the allowed part of a length-eleven interval has length , so its time contribution divided by is . Multiplying the marginal tails would instead give , an incorrect independence answer.
For general , Tonelli and the conditional uniform law give the extended nonnegative identities
These quantities may be infinite. If , the cycle theorem applied to also gives the almost-sure ordinary-process time average . If the second moment is infinite, apply the bounded theorem simultaneously for at positive integers . Their limiting constants increase to infinity by monotone convergence. The untruncated average dominates each one, so its lower limit is infinite almost surely. This conclusion concerns a time average, not .
Worked example 1.6.6
(Continuous lifetimes, continuous observation)
Let be uniform on . Find the equilibrium residual survival, its mean and .
Here and for . For ,
The tail is zero after two. Since , the mean is . The joint event uses the integrated tail beginning at , giving . Equal age and residual marginals alone do not permit multiplying marginal probabilities.
Worked example 1.6.7
(Why the exponential case looks different)
Let be exponential of rate . Derive the equilibrium joint survival and decide whether age and residual life are independent.
With ,
Both marginals are exponential of rate , and the factored joint survival for all nonnegative proves independence. Meanwhile has density , not the original exponential density. Memorylessness preserves the residual law; it does not remove length bias from the whole containing interval.
Common mistake
(Expecting a visitor to wait half a mean lifetime)
The tempting answer ignores length bias. When the variance is finite, , which equals only for almost-surely constant lifetimes; in the two-lifetime example it is , not . Length bias never decreases the containing interval's mean: , with equality exactly for almost-surely constant lifetimes, and this mean may be infinite. The remaining part need not have mean exceeding ; for constant lifetimes it is only .
Two limits that must stay separate
Counterexample mode
A time-average law does not remove a deterministic phase
Take almost surely and start the ordinary process with a renewal at zero. At times , ; at times , . Thus the law of has no limit as real , even though its time-averaged law is uniform on .
The stationary construction instead uses a uniform phase: with each marginal uniform on . It is stationary despite the lattice gap law. Here , whereas the product of the marginal probabilities is .
A nonarithmetic gap law means that there is no with . Such hypotheses enter separate renewal theorems about ordinary fixed-time convergence. None is needed for the time averages, randomized-observation expectation limit or stationary construction proved here. Those proofs do not establish the nonarithmetic fixed-time theorem.
There is a second integrability trap.
Common mistake
(A finite mean does not give a finite equilibrium wait)
For a lifetime with survival , we have , but . Its equilibrium residual survival is , whose integral diverges. The stationary probability law exists and the mean replacement rate is finite, yet the expected equilibrium wait is infinite. If instead , the length-bias normalization used in this unit is not a probability construction at all; our finite-mean theorems cannot be applied.
Practice with complete solutions
Quick check
(In the two-lifetime example, let be independent uniform observation time on . Find the limits of the time fraction with residual life exceeding four and of . Does this calculation prove a limit for ?)
Find the reward of each cycle type first.
Solution · Guided solution
In a length-one cycle the reward is zero; in a length-eleven cycle it is seven. Hence both requested limits are : the first follows from the cycle theorem and the second from bounded dominated convergence after conditioning on the path. It proves no ordinary fixed-time limit. The integer gaps identify a lattice setting, but that fact alone does not decide whether a limit exists at this particular threshold. A separate fixed-time renewal argument would be required.
Quick check
(For with a constant , determine when the equilibrium probability construction exists and when its residual mean is finite.)
Integrate the survival function twice.
Solution · Guided solution
Integrating the survival gives exactly when ; for it is infinite. In the finite-mean case,
Its mean is for and infinity for . A finite mean gap ensures normalization of the probability law, while a finite second gap moment ensures a finite residual mean. These are different requirements.
Quick check
(A simulator samples , then a uniform point inside that sampled interval, and places fresh independent gaps on either side. Does it produce the stationary renewal construction? Explain using the two-lifetime example.)
Compare the law of the sampled containing length with .
Solution · Guided solution
The uniform location is correct conditional on the containing interval, but the containing length must first have law , not . The proposed simulator chooses length eleven with probability instead of . Its mean residual is instead of . Repair it by length-biased sampling before drawing the uniform location. Future gaps beyond the containing interval retain the original law .
Where to continue
The earlier renewal-reward note supplies the event-rate law and completed-cycle rewards; this note adds within-cycle rewards, an explicit observation-time expectation limit and a stationary construction. Ordinary fixed-time convergence for nonarithmetic renewals requires a further renewal theorem. Infinite-mean equilibrium alternatives, quantitative convergence rates and stopping-time observation bias are separate topics.