Why Zero Elapsed Time Gives a Boundary Condition
A transition matrix describes how a model's state can change between two times. Its value when those times coincide is not an arbitrary special case: it is fixed by what “zero elapsed time” means. If a person is in a particular state at time \(t\), then at that same time the person is still in that state. This gives a useful boundary condition for constructing or checking transition matrices.
Throughout this tutorial, consider a finite-state continuous-time model with a well-defined state at each time and no instantaneous transitions. The last condition excludes models in which a state change is deliberately imposed at a particular instant. No mortality assumption, interest rate, or numerical transition intensity is needed to establish the zero-duration result; it follows from the definition of the state at the same instant.
The row identifies the state at the starting time; the column identifies the state at the ending time. In particular, \(P(t,t)\) describes the state at time \(t\) conditional on the state at that same time. Once the row and column states are specified in the same order, each entry has an immediate interpretation: a diagonal entry asks for the probability of remaining in the known state at the same instant, while an off-diagonal entry asks for the probability of being in a different state at that instant.
Deriving the Identity Matrix
Suppose the process is known to be in state \(i\) at time \(t\). The event “the state at time \(t\) is \(i\)” then has conditional probability 1. For any different state \(j\), the event “the state at time \(t\) is \(j\)” has conditional probability 0. Thus the probability in row \(i\), column \(j\) is 1 exactly when \(i=j\), and 0 otherwise.
Here \(I_n\) is the \(n\times n\) identity matrix. This result is a boundary condition: it specifies the transition matrix at the instant the elapsed time is zero. It does not say that the matrix at a positive elapsed time is the identity; with time to evolve, transitions may occur.
Write the same ordered list of states along the rows and the columns.
In row \(i\), condition on being in state \(i\) at time \(t\).
The entry in column \(i\) is 1; entries in every other column are 0.
The resulting matrix has ones on its main diagonal and zeros everywhere else.
The key point in this argument is that both the condition and the event refer to the same time. A transition from one state to another would require the state to differ between the starting and ending times, but here there is no interval between them. This argument relies on the ordinary interpretation of a process that has a single well-defined state at each instant. If a model explicitly includes instantaneous state changes, its event-time convention must be specified separately rather than silently applying this boundary condition.
Worked Checks
Worked Example: Constructing a Four-State Boundary Matrix
A hypothetical care model has states \(A\) (active), \(H\) (receiving home support), \(F\) (receiving facility care), and \(D\) (dead). Assume the state at each time is well-defined and there are no instantaneous transitions. Construct \(P(12,12)\), using the stated order for both rows and columns, and explain two selected entries.
Method. Apply the equal-time definition entry by entry. The row and column order is \(A,H,F,D\). There is no elapsed time between month 12 and month 12, so a person known to be in any one of these states remains in that same state at that instant.
In row \(A\), the conditional probability of being in \(A\) at month 12, given that the person is in \(A\) at month 12, is 1. The probabilities in columns \(H,F,D\) are each 0. The same reasoning gives a 1 in the matching column for each of rows \(H,F,D\), and zeros in the other columns.
For instance, the entry in row \(H\), column \(F\) is 0: conditional on being in home support at month 12, the person is not in facility care at that same instant. The entry in row \(D\), column \(D\) is 1: conditional on being dead at month 12, the state at month 12 is dead.
Check. The state labels on the rows and columns are identical and in the same order. Each diagonal entry is 1, and each off-diagonal entry is 0. The matrix is \(I_4\), as required. The common mistake this avoids is putting a 1 wherever a transition seems clinically plausible; at zero duration, plausibility over a positive interval is irrelevant.
Worked Example: Finding Errors in a Proposed Zero-Duration Matrix
A hypothetical three-state model uses the state order \(W,C,D\): working, on claim, and dead. An analyst proposes the following matrix for \(P(5,5)\). Determine whether it satisfies the equal-time boundary condition and correct it.
Method. Compare each entry with the required value: a diagonal entry must be 1, and an off-diagonal entry must be 0. The order \(W,C,D\) is used for both rows and columns, so the matching-state entries are \(M_{WW}\), \(M_{CC}\), and \(M_{DD}\).
The diagonal entries are \(M_{WW}=0.96\), \(M_{CC}=1\), and \(M_{DD}=0.98\). The required values are \(1,1,1\), respectively. The off-diagonal entries \(M_{WC}=0.04\) and \(M_{DW}=0.02\) are also nonzero, although each is required to be 0. Therefore, the proposed matrix fails the boundary condition.
At \(P(5,5)\), no probability is assigned to a change from working to on claim, or from dead to working. Those entries must be zero, regardless of whether such transitions might be possible over a positive interval. Replacing the proposed matrix with the identity matrix gives
Check. The corrected diagonal entries are all 1; its six off-diagonal entries are all 0. The common mistake avoided here is treating plausible positive-time transition probabilities as if they belonged in a matrix with equal start and end times. A check based on other matrix properties cannot replace this entry-by-entry boundary check.
Worked Example: Applying the Boundary Matrix to an Uncertain Starting State
A hypothetical administrator has an uncertain description of a member's state at month 8: the probabilities of being active, on claim, or dead are \(0.2, 0.5,\) and \(0.3\), respectively. Use the order \(A,C,D\) for rows, columns, and the distribution vector. Find the state distribution at month 8 after applying the zero-duration transition matrix.
Assumptions and method. Treat the stated values as the initial row distribution at month 8. They sum to \(0.2+0.5+0.3=1.0\), so they specify a probability distribution. With no elapsed time and no instantaneous transitions, the transition matrix is \(I_3\). Multiply the row distribution by this matrix.
Compute each component of the product. The probability assigned to active is \(0.2(1)+0.5(0)+0.3(0)=0.2\). The probability assigned to on claim is \(0.2(0)+0.5(1)+0.3(0)=0.5\). The probability assigned to dead is \(0.2(0)+0.5(0)+0.3(1)=0.3\). Therefore,
Check. The resulting distribution is unchanged, and its components still sum to \(1.0\). This confirms that multiplying by the identity preserves the vector. The common mistake avoided is interpreting uncertainty about the member's state as time passing: uncertainty in the starting distribution does not create transitions during a zero-length interval.
State Ordering and Interpretation
The identity-matrix statement depends on a consistent state order. If the states are listed as \(A,H,F\) on the rows, they must be listed as \(A,H,F\) on the columns for the ordinary identity matrix to appear. Each 1 must connect a state to the same state, not merely occupy a position that happens to be on a diagonal under mismatched labels.
Changing the labels' order is allowed, but the row and column orders must change together. For example, if both are reordered from \(A,H,F\) to \(F,A,H\), the entries still put a 1 where each row's state matches the same column state. The displayed matrix remains an identity matrix. Using one order for rows and a different order for columns changes how the entries are displayed and is not the convention used in the definition here.
| Entry type at equal times | Meaning | Required value |
|---|---|---|
| Diagonal, \(p_{ii}(t,t)\) | Same state at the same instant, conditional on state \(i\) | 1 |
| Off-diagonal, \(p_{ij}(t,t)\), \(i\ne j\) | Different state at the same instant, conditional on state \(i\) | 0 |
Common Exam Traps and Why the Check Matters
- Using a positive-time transition entry at equal times. A nonzero chance of moving between states over an interval does not make the corresponding entry nonzero when the interval has length zero.
- Checking only diagonal entries. A proposed matrix with diagonal ones but a nonzero off-diagonal entry still fails the boundary condition. Every entry is determined.
- Mixing row and column state orders. The diagonal matches states only when both axes use the same order. Verify labels, not just the visual location of a 1.
- Confusing an uncertain starting state with a transition. An initial distribution may assign positive probabilities to several states, but multiplication by \(P(t,t)\) leaves that distribution unchanged.
- Assuming the result for an instantaneously forced model. The identity boundary condition uses a process with a well-defined state and no instantaneous state changes. A model that imposes a change at a specified instant needs its own event-time convention.
The boundary condition is useful because it gives a direct diagnostic for formulas and numerical calculations: as the two times coincide, the diagonal entries must equal 1 and the off-diagonal entries must equal 0. It is a specific equal-time check, not a substitute for checking other properties that transition matrices may need to satisfy over positive intervals.
Check Your Understanding
For each problem, use the same state order for rows and columns unless stated otherwise. The answer is included so you can check the boundary condition and any arithmetic.
- A two-state model has states active and dead. Write \(P(3,3)\) in that order. (Answer: \(\begin{pmatrix}1&0\\0&1\end{pmatrix}\).)
- For states \(R,S,T\) in that order, a proposed \(P(9,9)\) has entries \(p_{RR}=1\), \(p_{SS}=0.9\), \(p_{TT}=1\), and \(p_{ST}=0.1\), with all other off-diagonal entries zero. Identify the incorrect entries and give their required values. (Answer: \(p_{SS}\) must be 1 and \(p_{ST}\) must be 0.)
- A four-state model is ordered \(A,B,C,D\). What are \(p_{CC}(t,t)\), \(p_{CA}(t,t)\), and \(p_{AD}(t,t)\)? (Answer: \(1,0,0\), respectively.)
- An uncertain state distribution in order \(L,H,D\) is \((0.35,0.45,0.20)\). Find the distribution after multiplication by \(P(6,6)\), and check its total. (Answer: \((0.35,0.45,0.20)\); total \(1.00\).)
- For states \(X,Y,Z\), both axes are reordered from \(X,Y,Z\) to \(Z,X,Y\). Write the zero-duration matrix in the new order, then state \(p_{ZX}(t,t)\) using that order. (Answer: \(\begin{pmatrix}1&0&0\\0&1&0\\0&0&1\end{pmatrix}\); \(p_{ZX}(t,t)=0\).)