Is a geometric distribution a special case?

Yes. One transient phase with a fixed probability of exit at each step gives a geometric duration on positive integers.

Must phases be directly observed?

No. They can be latent mathematical states, although that limits how they should be interpreted.

Can duration be zero?

Only if the chosen convention allows starting in the absorbing state. State that convention explicitly when comparing formulas or software.

AI Glossary / Mathematics and optimization / Advanced

Discrete phase-type distribution

Definition

A discrete phase-type distribution describes the number of steps until a finite-state, discrete-time Markov chain reaches an absorbing state. Before absorption, the process moves among transient states called phases. The distribution combines these transitions into probabilities for completion after one step, two steps, and so on.

What it means

The phases can represent stages in a process or simply provide a flexible mathematical representation of a duration distribution. Absorption marks an event after which the model stops counting, such as completion of a task.

“Discrete” means the duration is counted in steps rather than continuously measured time. A step might be a service attempt or a fixed observation interval, depending on the model. The starting convention matters: if the process begins in a transient phase, completion takes at least one step. Some formulations also allow an initial probability of immediate absorption at step zero.

How it works

Let α be the row vector of initial probabilities over transient phases, and T the matrix of probabilities of moving between those phases in one step. Its entries are nonnegative, and each row sum is at most one. The missing probability in each row gives the chance of absorption from that phase.

Writing 1 for a column of ones, the exit vector is t=1−T1. Under the usual eventual-absorption assumption, the probability of completion at positive step n is αTⁿ⁻¹t, while the probability of still being transient after n steps is αTⁿ1.

The matrix records transition probabilities, not continuous-time rates. A valid representation must ensure the relevant transient process eventually absorbs. Estimation and goodness-of-fit checks are separate from calculating probabilities from an already specified representation.

Three practical examples

1. Workflow analysis: model the number of review rounds before an item reaches an accepted completion state, including possible repeat stages.

2. Queueing research: represent a service duration measured in discrete slots using several transient phases.

3. Reliability modeling: describe the number of operating cycles before an absorbing failure event, when a suitable finite Markov-state model is defensible.

Think of it this way

Think of moving through rooms with probabilistic doors, where one exit leads permanently outside. The distribution records how many moves occur before that exit is reached.

The rooms need not correspond to observable real-world departments. They may be mathematical phases introduced to fit durations. Also, the doors' probabilities must follow the model's Markov assumptions; arbitrary history-dependent behavior cannot be ignored without consequences.

Worked example

Consider an original two-phase process starting in phase A. From A, the next step stays in A with probability 0.5 and moves to B with probability 0.5. From B, it stays in B with probability 0.25 and completes with probability 0.75.

Thus α=[1,0], T=[[0.5,0.5],[0,0.25]], and t=[0,0.75]ᵀ. Completion in one step is impossible because A has no direct exit, so P(N=1)=0.

For two steps, the only completion path is A→B→exit, with probability 0.5×0.75=0.375. For three steps, completion can follow A→A→B→exit or A→B→B→exit. Their probabilities are 0.5×0.5×0.75=0.1875 and 0.5×0.25×0.75=0.09375. Therefore P(N=3)=0.28125.

The probability of completion by step 3 is 0.65625, leaving survival probability 0.34375. The expected time in A before moving to B is 1/0.5=2 steps; the expected time from B to exit is 1/0.75=4/3. Total expected duration is 10/3 steps. These hand calculations describe an assumed process, not fitted operational data.

Illustrative case study

Fictional case: Maple Review studies how many weekly processing rounds applications require. Analysts consider two latent phases representing an initial stage and a final-check stage, each with possible repeats.

They compare the model's duration probabilities with held-out observations and inspect long-running cases separately. Because the phases are partly statistical constructs, they avoid claiming that every application physically follows two identifiable departments. They also account for incomplete cases rather than recording them as completed durations. The scenario illustrates modeling discipline without inventing a measured fit or service improvement.

Common misconceptions and mistakes

Do not confuse a discrete transition matrix with the generator matrix of a continuous phase-type model. Their entries and formulas have different meanings.

Having nonnegative entries and row sums at most one is not sufficient by itself to guarantee eventual absorption from every initial phase. A closed nonabsorbing class would violate the intended proper-duration model.

Different phase representations can produce the same duration distribution, so fitted phases need not be uniquely identifiable. A flexible fit also does not prove the real process is exactly Markovian or that its phases have causal interpretations.

Business use

This distribution can connect duration estimates to queueing and workflow models while retaining a matrix-based calculation. It is useful when completion timing matters more than only the average duration.

Choose the step definition and absorption event carefully. Examine tail probabilities, censoring, parameter uncertainty, and changes in process behavior. Validate whether a simpler distribution would suffice before adding many phases. When the model supports staffing or capacity decisions, test the downstream decision's sensitivity to plausible duration distributions rather than relying on one fitted representation.

Check your understanding

Why is the probability of completion in two steps 0.375 in the worked process?

Answer: The process must move from A to B on the first step, probability 0.5, and exit from B on the second, probability 0.75. Multiplying gives 0.375; no other two-step completion path exists.

Related terms

Markov Chain
Censoring
absorbing state
transient state
geometric distribution
phase-type distribution
survival probability
queueing model
latent state

Learn next

Draw the worked transition diagram and calculate survival probabilities using successive row-vector multiplications. Check that completion probabilities plus remaining survival sum to one. Then compare the discrete construction with a continuous-time phase-type model, keeping probabilities, rates, and time units separate.

Sources

Primary source: Luz Judith Rodriguez Esparza, “Maximum likelihood estimation of phase-type distributions”, Section 2.3, https://backend.orbit.dtu.dk/ws/portalfiles/portal/5512970/phd245_ljre.pdf . The discrete definition, exit vector, probability mass and survival formulas were read. The two-phase example and expected-duration check are independently derived; no model fitting or simulation is claimed.

Previous: Discrete optimization Next: Discrete random variable