Defining The Matrix Product Integral

Why a Matrix Product Integral Is Needed

A transition-intensity matrix describes local movement among states. A transition-probability matrix describes movement over a whole time interval. For a single state, accumulating a rate over time leads to a scalar survival calculation. With several states, the state reached at one time affects which transitions are possible later, so the increments must be combined as matrices, in time order. The matrix product integral is the operation that performs this accumulation.

The definition below uses the row-transition convention: rows of a transition matrix correspond to the state at the start of an interval, and columns correspond to the state at its end. A row probability vector is therefore multiplied by the transition matrix on the right.

Definition: Consider a finite-state, time-inhomogeneous Markov process with state space \(\{1,\ldots,K\}\). Let \(Q(u)\) be its transition-intensity matrix at time \(u\): for \(i\ne j\), \(q_{ij}(u)\ge 0\) is the instantaneous intensity of a transition from state \(i\) to state \(j\), and \(q_{ii}(u)=-\sum_{j\ne i}q_{ij}(u)\). Define the cumulative intensity matrix \(A(t)=\int_0^t Q(u)\,du\), or more generally use a right-continuous cumulative intensity matrix of finite variation. The product integral from \(s\) to \(t\), written \(\prod_{(s,t]}(I+dA(u))\), is the limit of ordered products of identity matrices plus increments of \(A\). It represents the transition-probability matrix \(P(s,t)\), with \(P_{ij}(s,t)=P(X(t)=j\mid X(s)=i)\).

Definition by Ordered Increments

Take a partition \(s=t_0<t_1<\cdots<t_m=t\), and write \(\Delta A_k=A(t_k)-A(t_{k-1})\). The corresponding finite product is

\prod_{k=1}^{m}(I+\Delta A_k) =(I+\Delta A_1)(I+\Delta A_2)\cdots(I+\Delta A_m).

The factors appear from earliest interval to latest interval, left to right. As the partition is refined, the product approaches the product integral:

P(s,t) = \prod_{(s,t]}(I+dA(u)) = \lim_{\|\Pi\|\to 0} \prod_{k=1}^{m}\{I+A(t_k)-A(t_{k-1})\}.

Here \(\|\Pi\|\) is the largest partition-interval length. For a cumulative intensity with jumps, the partitioning convention must retain the jump increments as whole increments; the product-integral definition includes their factors as well as the limiting contribution from the continuous part. In applications, the cumulative intensity and its increments must satisfy conditions that make the resulting matrices valid probabilities.

For a continuous cumulative intensity, the increments become small as the partition is refined. The factors \(I+\Delta A_k\) are local approximations to transition over short intervals. They are not generally the transition matrices over the entire interval \((s,t]\); the ordered product combines local movements to represent the whole interval.

How the Factors Become Probabilities

Each row of an intensity matrix sums to zero. Consequently each increment \(\Delta A_k\) has row sums zero, and \(I+\Delta A_k\) has row sums one. Its off-diagonal entries are nonnegative because they are cumulative intensities for transitions between distinct states. For the factor to be a probability matrix, its diagonal entries must also be nonnegative. For continuous, locally bounded intensities, sufficiently fine intervals make the diagonal entries nonnegative: the total outgoing increment from any state on a short interval is then at most one.

At an actual jump of the cumulative intensity, refining the partition does not make that jump smaller. The jump increment must itself yield a valid factor \(I+\Delta A(u)\). In particular, for each state, its total outgoing jump increment cannot exceed one. These conditions are sufficient to ensure that the factors are row-stochastic: their entries are nonnegative and each row sums to one. A product of row-stochastic matrices is also row-stochastic, and a limit of such matrices remains row-stochastic. Thus the product integral has the defining algebraic properties required of a transition-probability matrix.

The probability interpretation also depends on the model. The process must have the stated state space and transition intensities, and the state and time information in the model must be sufficient to describe future transitions. If a transition intensity depends on unrecorded history, a product integral based only on the current state and time need not give the desired transition probabilities. Adding states to represent relevant history can sometimes restore a suitable Markov description.

From Local Movement to Interval Probabilities

The product-integral definition is motivated by composing short-interval transition probabilities. For a small interval of length \(h\), the transition matrix under a locally bounded intensity matrix has the first-order form \(I+Q(u)h+o(h)\). The diagonal term accounts for remaining in the current state; an off-diagonal term accounts for a transition to another state. Over successive short intervals, the Markov property lets us multiply these local matrices in chronological order. Since \(Q(u)h\) is the corresponding small increment of \(A\), the limiting ordered product is precisely the product integral.

This justification explains both ingredients of the definition. The identity matrix represents no change over an infinitesimal interval, and \(dA(u)\) represents the first-order transition changes. Ordering is needed because a transition in an earlier interval can change the state from which later transitions occur. The full evolution can therefore not, in general, be found by adding all the increments of \(A\) and using one matrix factor.

Worked Example: Combining Two Ordered Matrix Increments

Suppose a hypothetical three-state process has two successive cumulative-intensity increments. In the first interval, state 1 can move toward state 2 and state 2 toward state 3. In the later interval, movement is possible from state 2 to state 1 and from state 3 to state 2. The increments are

\Delta A_1= \begin{pmatrix} -0.10&0.10&0\\ 0&-0.20&0.20\\ 0&0&0 \end{pmatrix}, \qquad \Delta A_2= \begin{pmatrix} 0&0&0\\ 0.15&-0.15&0\\ 0&0.10&-0.10 \end{pmatrix}.

Each increment has zero row sums. The factors are

F_1=I+\Delta A_1= \begin{pmatrix} 0.90&0.10&0\\ 0&0.80&0.20\\ 0&0&1 \end{pmatrix}, \qquad F_2=I+\Delta A_2= \begin{pmatrix} 1&0&0\\ 0.15&0.85&0\\ 0&0.10&0.90 \end{pmatrix}.

Both have nonnegative entries and rows summing to one. Since the first increment occurs earlier, the ordered product is \(F_1F_2\). Multiplication gives

F_1F_2= \begin{pmatrix} 0.90(1)+0.10(0.15)&0.90(0)+0.10(0.85)&0\\ 0(1)+0.80(0.15)+0.20(0)&0(0)+0.80(0.85)+0.20(0.10)&0.20(0.90)\\ 0&0.10(1)+0.90(0.10)&0.90(0.90) \end{pmatrix} = \begin{pmatrix} 0.915&0.085&0\\ 0.120&0.700&0.180\\ 0&0.190&0.810 \end{pmatrix}.

For example, the entry 0.120 in row 2, column 1 is the probability contribution \(0.80\times0.15\): a process starting in state 2 follows the first factor to state 2 with probability 0.80, then the later factor moves it to state 1 with probability 0.15. Each row of the product sums to one, and all entries are nonnegative. The product is therefore a valid transition-probability matrix for these two increments.

Worked Example: Approximating a Continuous Product Integral

Consider a hypothetical two-state process with state 1 able to move to absorbing state 2 at a constant intensity of \(0.20\) per time unit. Its intensity matrix is \(Q=\begin{pmatrix}-0.20&0.20\\0&0\end{pmatrix}\). Over a unit of time, divide the interval into four equal pieces. Each increment is \(\Delta A=Q/4\), so the local factor is

I+\Delta A= \begin{pmatrix} 0.95&0.05\\ 0&1 \end{pmatrix}.

The four-interval approximation is the product of four identical factors. The probability of remaining in state 1 in this approximation is \(0.95^4=0.814506\), and the probability of being in state 2 is \(1-0.814506=0.185494\). Thus the approximating matrix is

\begin{pmatrix} 0.8145&0.1855\\ 0&1 \end{pmatrix}.

This is a partition-based approximation, not the exact product integral. Refining the partition makes the local increments smaller and the ordered product converge to the continuous-time transition matrix. The example illustrates the definition: a whole-interval transition matrix is built from local factors, rather than treating the cumulative intensity increment itself as a transition probability.

Worked Example: Detecting an Invalid Jump Increment

Suppose a proposed jump increment for a hypothetical two-state system is

\Delta A= \begin{pmatrix} -1.20&1.20\\ 0&0 \end{pmatrix}.

The factor would be

I+\Delta A= \begin{pmatrix} -0.20&1.20\\ 0&1 \end{pmatrix}.

Although its rows sum to one, the first row has a negative entry and an entry greater than one. It cannot be a probability distribution. If this increment is intended to represent one jump of the cumulative intensity, it is inadmissible as a probability factor: its outgoing jump increment from state 1 is 1.20, exceeding one. If instead it represents a continuous increase over an interval, it should be resolved into sufficiently small increments before forming factors. Splitting an actual jump into artificial pieces would change the model, so that is not a remedy for an invalid jump.

Assumptions, Checks, and Limits

Before interpreting a product integral as transition probabilities, check the model and the factors. A useful set of checks is:

State and time definition: the states are mutually exclusive and cover the modeled possibilities, and the intensity matrix refers to transitions among those states.
Intensity structure: off-diagonal intensities are nonnegative and each diagonal is the negative sum of its row’s off-diagonal intensities.
Existence and regularity: cumulative intensities have finite variation on the interval, with locally bounded rates being a convenient sufficient condition for the continuous-time construction.
Probability factors: increments yield nonnegative factors with rows summing to one, including at jumps.
Model sufficiency: the stated intensities adequately describe transition behavior given the modeled state and time.

Negative entries in a factor, rows that fail to sum to one, or a factor with an outgoing jump increment exceeding one are direct warnings that the proposed increment cannot be used as a probability step. A product that is not row-stochastic cannot be interpreted as a transition-probability matrix under this convention. Other failures are less visible: an apparently valid matrix calculation does not establish that the chosen states or intensity model describe the actual process appropriately.

The product integral is a definition and a way to express accumulation, not a guarantee that the intensities are known correctly or that their estimates have desirable statistical properties. Its transition-probability interpretation relies on the process model and admissible increments. The exact way to estimate transition probabilities from observed data, and the consequences of particular event-time structures, require additional statistical arguments.

Common Mistake: Adding the Increments

A tempting but generally incorrect shortcut is to add all cumulative-intensity increments and use \(I+\sum_k\Delta A_k\) as the transition matrix. That expression retains only a single first-order change. It does not account for paths in which a process moves to another state early and then moves again later. The product integral includes such sequential possibilities through its ordered factors. It also preserves the probability-matrix structure when its factors are valid. The sum of increments alone need not have nonnegative entries after adding the identity, and even when it does, it generally omits the effect of successive transitions.

Key takeaway: Under the row-transition convention, the matrix product integral combines identity-plus-increment factors in chronological order. With suitable intensity and increment conditions, its limit is a row-stochastic matrix whose entry in row \(i\), column \(j\) is the probability of being in state \(j\) at time \(t\), given state \(i\) at time \(s\).

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

Under the row-transition convention, what do the row and column of \(P_{ij}(s,t)\) represent?
Why are the increment factors multiplied in chronological order rather than replaced by one factor involving the sum of the increments?
A proposed increment has zero row sums and nonnegative off-diagonal entries. What additional check is needed before \(I+\Delta A\) can be used as a probability factor?
For a continuous intensity, what happens to the increment factors as the time partition is refined?

Solutions. (1) Row \(i\) is the starting state and column \(j\) is the state at the end of the interval, so \(P_{ij}(s,t)=P(X(t)=j\mid X(s)=i)\). (2) Earlier movement changes the state from which later movement can occur; ordered multiplication combines these sequential transitions. (3) Check that every diagonal entry of \(I+\Delta A\) is nonnegative, as well as verifying that its rows sum to one. (4) The continuous increments become small, and the ordered products converge to the product integral under the stated regularity conditions.