Why replace the constant rate?
A constant-rate Poisson process assigns the same expected count to every interval of a fixed length. That is unsuitable for a service whose traffic rises during the morning and falls overnight. We can retain independent counts on disjoint intervals while allowing the expected count to depend on the location of the interval. The replacement for rate times duration is an integral of a deterministic rate.
The central idea is to distinguish calendar time from accumulated arrival opportunity. During a quiet interval calendar time advances but the opportunity clock does not. When traffic is heavy the opportunity clock advances quickly. Running an ordinary rate-one Poisson process on this second clock produces the desired model, and supplies proofs rather than just a mnemonic for its formulas.
The prerequisites are the homogeneous Poisson process, basic integration and conditional probability.
Intensity and cumulative intensity
Definition 1.5.1
(A deterministic intensity on a finite horizon)
Fix an observation horizon . Let be measurable and integrable, and define
A nonhomogeneous Poisson process with this intensity is a counting process with right-continuous, nondecreasing paths with unit jumps, , independent increments and
A Poisson variable of parameter zero is identically zero. Path regularity is part of this definition; the construction below establishes that a process with these properties exists.
Integrability is the finite-horizon assumption that prevents infinitely many arrivals before the endpoint. The cumulative intensity is finite, continuous, absolutely continuous and nondecreasing. It need not be strictly increasing. Changing at a set of Lebesgue measure zero leaves and the process law unchanged; the value assigned at an individual time does not create an arrival atom there.
Intensity has units of events per unit time, whereas is an expected number of events. In particular,
For the covariance, write and use independence of the bracketed increment. Independent increments do not mean that the total counts and are independent.
Constructing the process and proving its laws
Theorem 1.5.2
(Deterministic time change)
Let be a homogeneous Poisson process of rate one. Then
is a nonhomogeneous Poisson process with the specified intensity. Its finite-dimensional distributions are uniquely determined by the increment law and independent increments. It has only finitely many arrivals on almost surely.
Proof
(Counting on disjoint operational intervals)
For ordered calendar times , their cumulative times are ordered, possibly with equalities. The increments
are counts on disjoint rate-one clock intervals, hence independent Poisson variables of parameters . An interval of zero length contributes a constant zero and does not damage independence. Summing these increments recovers every vector , which proves both the claimed laws and their uniqueness.
The path is nondecreasing. Its right continuity follows from continuity and monotonicity of and right continuity of . Continuity of prevents the clock from skipping a nonempty interval of operational time, so it cannot collapse several distinct operational arrivals into one jump. A flat piece introduces no jump: the operational count stays fixed there. Finally is a finite Poisson variable almost surely. This bounds the number of jumps on the entire observation interval.
Proof X-ray
Which assumption performs which job?
Monotonicity puts disjoint calendar intervals into disjoint operational intervals; deterministic time change lets us use the original independent-increment theorem directly. Continuity prevents a skipped operational interval from producing a multiple jump. Finiteness of prevents explosion. Strict positivity of is unnecessary: an interval with zero integrated intensity is simply an interval with no arrivals. Replacing the deterministic clock by a clock depending on the same arrival process would require a new independence argument.
At a point where as , put . The count law gives
The local interpretation therefore holds at every continuity point of , and almost everywhere for an integrable intensity. It need not hold at every prescribed point for an arbitrary measurable representative. The integral increment law is the definition; the informal phrase “probability ” cannot replace its hypotheses or its independence requirement.
The next arrival after an observation time
Fix a deterministic . Let be the information generated by all counts through . Define the next observed waiting time to be the time from to the first arrival in , and set if there is no such arrival. This infinity records censoring at ; it does not assert what an unspecified process does after .
Theorem 1.5.3
(Waiting survival and its possible missing mass)
For ,
On the finite part of its support, has density
Its remaining mass is
Proof
(Zero counts on the future interval)
The event means exactly that . This increment is independent of , and its zero-count probability is the displayed exponential. Independence from the whole count history follows first for each finite collection of past counts from independent increments, then for their generated -field by the usual monotone-class extension. Absolute continuity and the chain rule give the density by differentiating one minus the survival function almost everywhere. Its integral is
The complement is exactly the zero-count event on . There is no extra atom at a finite endpoint, since is continuous.
The future after deterministic is independent of the observed past, but its law depends on . A constant intensity gives an exponential survival function in the physical duration . A varying intensity usually does not. The fixed-time argument above is not, by itself, a theorem about arbitrary random observation times.
On an infinite horizon, assume instead that is integrable on every finite interval. The same construction works at every finite time. If , then the total number of arrivals is Poisson of parameter , and the probability of no arrival after is . Indeed increases to almost surely: a rate-one Poisson process has no jump at the fixed point . If , a next arrival exists almost surely after every deterministic . These are statements about an infinite-horizon model, unlike finite observation censoring.
Worked example 1.5.4
(An intensity with finite total mass)
On the whole half-line let . Find the probability of no arrival ever, the probability of at least two arrivals ever, and the probability of a finite first arrival after given no arrival through .
The total cumulative intensity is one, so the total count is . The first two answers are and . Conditional on the empty past, future counts have integrated intensity , independently of that past. Thus a finite first arrival later has probability . The complementary mass is a genuinely infinite waiting time in this infinite-horizon model.
Inverting the clock without assuming a positive rate
Let be independent rate-one exponentials and the arrival times of . The physical th arrival can be constructed by
with the infimum of the empty set defined as infinity. A nondecreasing continuous clock has a generalized inverse even when it has no ordinary inverse. For this construction, , and hence
In fact is exactly the th arrival of : the event identity also reads . A plateau causes no ambiguity. If an operational arrival equals its constant level, both definitions place the physical arrival at the plateau's left endpoint. Almost surely this equality never occurs: the maximal nontrivial plateau intervals are countable, since their interiors are disjoint and each contains a rational number, and every has a continuous distribution. An ordinary inverse is unnecessary even when the clock pauses.
To generate only the next arrival after deterministic , draw a fresh and solve
using the generalized inverse. If the threshold exceeds , there is no observed arrival before the horizon. The exponential draw is a waiting amount of cumulative intensity, not a waiting amount of calendar time.
Concept lensAnalytic
Two clocks, two kinds of gaps
The operational gaps are independent and identically exponential. The calendar gaps are obtained by applying a nonlinear inverse to their cumulative sums, so they are generally neither identically distributed nor independent. A flat operational clock inserts a definite quiet stretch into calendar time. It does not require an exponential variable to have an atom or a deterministic gap.
Worked calculations with a quiet interval
Worked example 1.5.5
(Traffic rises, stops, then restarts)
Observe through and use
The counts in , and are independent with respective laws , zero and . Consequently and . Equal-length intervals have different laws; independent increments have survived while stationarity has not.
Suppose the process has been observed through . For its next observed wait,
The constant middle piece means there can be no first arrival during the shutdown. The probability of no further observed arrival is . All these conditional formulas are the same whatever counts were recorded before .
For direct simulation of the first arrival, an operational threshold gives . Threshold passes the quiet interval and gives . Threshold gives no arrival in the observed horizon. These are sample threshold calculations, not events of positive probability at a single exact exponential value.
Arrival locations conditional on a count
There is another useful consequence of the increment law. Put and partition into finitely many disjoint time intervals , of integrated intensities . Their independent Poisson counts , conditional on the total being , have mass
To derive it, multiply over the intervals and divide by . Zero-intensity cells must have zero count; their probability factors are interpreted accordingly. This is precisely the allocation law of independent times with density . Because this allocation identity holds for every finite partition into intervals, it determines the law of the finite point configuration. The ordered arrivals conditional on are therefore the order statistics of those times. For , the ordered arrival times themselves are not independent.
In the worked model, conditional on exactly one arrival before time three, its probability of lying in is . Its calendar-time density is on , zero on and on . It is not uniform on the three-unit observation window. Uniformity belongs to cumulative time; physical time receives weight in proportion to intensity.
Worked example 1.5.6
(Conditional locations under a linear intensity)
On take . Given , find the chance that all three arrivals lie in . Give the density of a single unordered arrival location under this conditioning and explain why the three ordered locations are not independent.
The integrated intensities are and . Each conditional unordered location lies in with probability , so all three do with probability . Its density is for . Sorting three independent draws creates the inequalities almost surely, and hence dependence. Explicitly, , and , whereas . The latter value follows by subtracting the binomial probabilities of two or three arrivals in .
Misconceptions to repair
Counterexample mode
A random rate changes the dependence structure
Choose a single random rate , equal to zero or two with equal probabilities, and conditional on run a homogeneous Poisson process of that rate. Counts on two disjoint length-one intervals are conditionally independent, but
Thus their covariance is one. Even its one-interval zero-count probability is , rather than . Averaging a shared random rate does not produce a nonhomogeneous Poisson process with deterministic intensity .
Three further errors come from reading the intensity pointwise or replacing it by a summary.
Common mistake
(Reading a zero rate at a point as a quiet period)
A zero rate at one isolated point says nothing about whether arrivals occur nearby, since changing at a single time changes neither nor the process. A quiet period requires a zero integral: if , there are no arrivals in almost surely.
Common mistake
(Replacing the intensity by its average)
Replacing by its average on preserves the law of alone; it generally changes intermediate counts and arrival locations. In the quiet-interval model the average rate is one, which would give a law instead of the constant zero.
Common mistake
(Renormalizing a censored waiting density)
A finite-horizon waiting density integrates to less than one because “no observed next arrival” is a possible outcome. Dividing it by its integral gives the conditional law of the wait given that an arrival is observed, which answers a different question. In the quiet-interval model observed from , the missing mass is .
Conditionally nothing changes; unconditionally independent increments are lost and counts become overdispersed.
A path to generalization
What changes when the intensity is itself random?
Conditionally nothing changes; unconditionally independent increments are lost and counts become overdispersed.
If the cumulative intensity is a random nondecreasing process independent of the unit clock , then defines a Cox, or doubly stochastic Poisson, process. Conditionally on the intensity path every statement of this note still holds. Unconditionally, independent increments are lost, as the random-rate counterexample showed, and when the law of total variance gives
with equality only when is almost surely constant.
Practice
Quick check
(Let on . Find the law of , the conditional probability of no arrivals in given , and .)
Compute first and use the event identity for .
Solution
Here . The increment is , so the conditional zero-arrival probability is by independence. Also and , giving . The count before time one does not alter the seven units of future cumulative intensity.
Quick check
(For the quiet-interval worked model, an observer starts at . Find the survival of the next observed wait for , and its probability of being infinite. Explain why treating the wait as an exponential of rate two is wrong.)
Find the first time after at which starts to increase.
Solution
There can be no arrival until calendar time two, which is a wait of . The survival is one for and for . The probability of no observed arrival is . A rate-two exponential starting immediately would assign positive probability to a wait shorter than , contradicting the shutdown.
Quick check
(A proposed intensity is for . What does the time-change construction say before time one? Why does it fail the finite-horizon assumptions on ?)
Compute for and let .
Solution
For , , so each such time has a finite Poisson count. But as . Since almost surely as , infinitely many arrivals accumulate before time one. The proposed rate is not integrable on , so it cannot give a finite-valued, non-explosive counting process on that closed horizon. Finite intensity at each time strictly before an endpoint does not imply finite cumulative intensity at the endpoint.
Connections and boundaries
The homogeneous process is recovered by . A deterministic nonlinear clock preserves independent increments but generally removes stationary increments and exponential calendar gaps. The integral formulation accommodates measurable rates and quiet periods without artificial point masses. Random intensities, arbitrary random-time conditioning, estimation from data and renewal residual-life theory require additional arguments and are outside this note.