Deriving The Product-Integral Recursion

From a Product Integral to a Recursion

A matrix product integral accumulates transition information over time by multiplying small matrix factors in temporal order. On a finite event-time grid, the continuous-looking operation becomes a finite sequence of ordinary matrix multiplications. The recursion is useful both for understanding the definition and for computing a transition matrix without repeatedly rebuilding the entire product.

Suppose the transition system has $K$ states. Let $A(t)$ be a cumulative transition-intensity matrix, with $A_{ij}(t)$ recording cumulative intensity from state $i$ to state $j$, for $i\ne j$. Its diagonal is defined so that each row sums to zero. At an ordered event time $t_k$, write $\Delta A_k=A(t_k)-A(t_k-)$ for the matrix increment. Between event times, assume there are no increments in the grid representation considered here. The event times satisfy $t_1<\cdots<t_m$.

Definition: For an event-time grid with increments $\Delta A_1,\ldots,\Delta A_m$, define the product-integral factor at time $t_k$ as $I+\Delta A_k$, where $I$ is the $K\times K$ identity matrix. Starting with $P_0=I$, the successive products are $$ P_k=P_{k-1}(I+\Delta A_k),\qquad k=1,\ldots,m. $$ Thus $P_k$ is the product integral accumulated through $t_k$, under the stated right-multiplication convention.

The identity matrix is the correct starting point because before any increment there has been no transition update: a system starting in state $i$ remains represented in state $i$, so the initial transition matrix has ones on the diagonal and zeros elsewhere. At each event time, the current transition matrix is updated by the new factor. The result after $m$ times is the ordered product

P_m=(I+\Delta A_1)(I+\Delta A_2)\cdots(I+\Delta A_m).

This expression is not a sum of increments. Each factor acts on the transition matrix already accumulated, so the recursion retains the sequence of updates.

Why the Recursion Follows

The derivation is a direct consequence of the finite product-integral definition. A product integral over an interval is obtained by multiplying its increment factors in time order. Before the first event time, the product contains no factors and equals the multiplicative identity $I$. Suppose the product through $t_{k-1}$ is $P_{k-1}$. Extending the interval to include $t_k$ appends the new factor $I+\Delta A_k$ on the right. Therefore the updated product is $P_k=P_{k-1}(I+\Delta A_k)$.

P_0=I,\qquad P_k=P_{k-1}(I+\Delta A_k),\qquad P_m=\prod_{k=1}^{m}(I+\Delta A_k),

where the displayed product means factors in increasing event-time order, with the earliest factor on the left. To verify the closed form, substitute the recursion repeatedly: $P_1=I(I+\Delta A_1)=I+\Delta A_1$; then $P_2=(I+\Delta A_1)(I+\Delta A_2)$. If $P_{k-1}=(I+\Delta A_1)\cdots(I+\Delta A_{k-1})$, substitution gives $P_k=(I+\Delta A_1)\cdots(I+\Delta A_k)$. This establishes the expression for every $k$ by induction.

For a matrix derived from transition intensities, an off-diagonal increment $\Delta A_{ij,k}$ is nonnegative and the diagonal increment is the negative sum of the off-diagonal increments in that row. Consequently, each factor has row sums one. For that factor to be a probability transition matrix, its entries must also be nonnegative; in row $i$, this requires the total outgoing increment to be at most one. These conditions are worth checking rather than assuming.

Worked Example: Two Ordered Updates in a Two-State System

Consider two states, 1 and 2, and two hypothetical event times $t_1<t_2$. Suppose the increments are

\Delta A_1= \begin{pmatrix} -0.20&0.20\\ 0.10&-0.10 \end{pmatrix}, \qquad \Delta A_2= \begin{pmatrix} -0.25&0.25\\ 0.05&-0.05 \end{pmatrix}.

Each increment has row sums zero, nonnegative off-diagonal entries, and total outgoing increment no greater than one in either row. The factors are therefore

I+\Delta A_1= \begin{pmatrix} 0.80&0.20\\ 0.10&0.90 \end{pmatrix}, \qquad I+\Delta A_2= \begin{pmatrix} 0.75&0.25\\ 0.05&0.95 \end{pmatrix}.

Start at $P_0=I$. The first update is $P_1=I(I+\Delta A_1)=I+\Delta A_1$. Apply the second increment by multiplying on the right:

\begin{aligned} P_2 &=P_1(I+\Delta A_2)\\ &= \begin{pmatrix} 0.80&0.20\\ 0.10&0.90 \end{pmatrix} \begin{pmatrix} 0.75&0.25\\ 0.05&0.95 \end{pmatrix}\\ &= \begin{pmatrix} (0.80)(0.75)+(0.20)(0.05)&(0.80)(0.25)+(0.20)(0.95)\\ (0.10)(0.75)+(0.90)(0.05)&(0.10)(0.25)+(0.90)(0.95) \end{pmatrix}\\ &= \begin{pmatrix} 0.61&0.39\\ 0.12&0.88 \end{pmatrix}. \end{aligned}

For example, the first row says that after both updates, the probability of being in state 2 when starting in state 1 is $0.39$. Each row sums to one: $0.61+0.39=1$ and $0.12+0.88=1$. The entries are nonnegative, consistent with a transition-probability interpretation.

Three-State Updates and a Direct Check

The same recursion applies without change when there are more than two states. The following hypothetical calculation includes a transient state and two possible destinations. It illustrates how each row of the product is computed from the previous matrix and the new factor.

Worked Example: Updating a Three-State Transition Matrix

Let the states be 1, 2, and 3. Suppose one event-time increment is

\Delta A_1= \begin{pmatrix} -0.30&0.20&0.10\\ 0&-0.25&0.25\\ 0&0&0 \end{pmatrix}.

The first row has total outgoing increment $0.20+0.10=0.30$, the second has $0.25$, and the third has none. All are at most one, so $I+\Delta A_1$ has nonnegative entries and row sums one. Now suppose a second hypothetical increment is

\Delta A_2= \begin{pmatrix} -0.10&0&0.10\\ 0&-0.20&0.20\\ 0&0&0 \end{pmatrix}, \qquad I+\Delta A_2= \begin{pmatrix} 0.90&0&0.10\\ 0&0.80&0.20\\ 0&0&1 \end{pmatrix}.

Again, each row of the factor is nonnegative and sums to one. Starting from $P_0=I$, the first update gives

P_1= \begin{pmatrix} 0.70&0.20&0.10\\ 0&0.75&0.25\\ 0&0&1 \end{pmatrix}.

Apply the second factor:

\begin{aligned} P_2 &=P_1(I+\Delta A_2)\\ &= \begin{pmatrix} (0.70)(0.90)&(0.20)(0.80)&(0.70)(0.10)+(0.20)(0.20)+(0.10)(1)\\ 0&(0.75)(0.80)&(0.75)(0.20)+(0.25)(1)\\ 0&0&1 \end{pmatrix}\\ &= \begin{pmatrix} 0.63&0.16&0.21\\ 0&0.60&0.40\\ 0&0&1 \end{pmatrix}. \end{aligned}

The first row sums to $0.63+0.16+0.21=1$; the second sums to $0.60+0.40=1$, and the third is already a probability row. The contribution $0.70(0.10)+0.20(0.20)+0.10(1)$ to the first row's third entry comes from all routes represented by the intermediate state distribution after the first update, followed by the second factor.

Using and Checking the Recursion

A practical calculation needs only the event-time increments in order. For each time, form $I+\Delta A_k$, multiply the current matrix by that factor, and store the result. The identity start is essential: it handles the case of no prior transition updates and makes the recursion valid for the first event time as well as later ones.

Initialize.
Set $P_0=I$, with the same number of rows and columns as the number of states.

Form the event factor.
At time $t_k$, add the identity to the increment: $F_k=I+\Delta A_k$.

Update in time order.
Compute $P_k=P_{k-1}F_k$ and retain the result for the next event time.

Check the result.
Check dimensions, row sums, and nonnegative entries when the factors are intended to represent transition probabilities.

The recursion is exact for the finite product specified by its increments. It is not, by itself, a claim that an arbitrary collection of matrices is a valid transition model. For example, if an increment has a row with total outgoing amount greater than one, the corresponding diagonal entry of $I+\Delta A_k$ is negative. The matrix product can still be computed algebraically, but that factor cannot be interpreted as a probability transition matrix.

A second diagnostic is the row-sum check. If the rows of $\Delta A_k$ do not sum to zero, then $I+\Delta A_k$ will not have row sums one, and multiplying it into a row-stochastic matrix may destroy the probability interpretation. Such a discrepancy can signal a diagonal constructed inconsistently with the off-diagonal increments, a coding error, or rounding too aggressive for the calculation. Check the increment before blaming the recursion.

What the Recursion Does—and Does Not—Replace

The recursive product is different from simply adding cumulative transition-intensity increments. Addition gives an accumulated intensity matrix; the recursion instead updates a transition matrix, allowing the distribution over states after one event time to be carried into the next. With multiple states, the intermediate state distribution matters: the next factor acts on every row of the current transition matrix.

The product-integral recursion is also distinct from treating the entire interval as if it had one constant intensity matrix and using a matrix exponential. That is a different construction with different assumptions and is not the event-time recursion derived here. For observed or estimated increments at distinct ordered times, the finite sequence of factors is the direct calculation.

Important limitations remain. The result depends on the specified increments, the event-time grid, and the chosen state representation. If increments between listed times are omitted, the calculation cannot account for them. A finite-step factor is a valid probability update only if it has nonnegative entries and row sums one. Finally, the recursion computes a matrix; substantive conclusions about a particular transition require identifying the row for the starting state and the column for the destination state.

Common Mistakes

Starting with a zero matrix. A zero matrix erases the identity action before the first event. Initialize with $I$, not with zeros.
Adding increments to the current transition matrix. The update is $P_k=P_{k-1}(I+\Delta A_k)$, not $P_{k-1}+\Delta A_k$. The increment belongs in a factor.
Skipping the identity in the factor. Multiplying by $\Delta A_k$ alone does not represent “no transition plus the increment.” The factor is $I+\Delta A_k$.
Accepting any computed matrix as a probability matrix. Check the increment structure and resulting row sums and entries. Algebraic computability alone does not guarantee valid probabilities.

For instance, in the two-state example, adding the increments would give an object that is not the calculated transition matrix. The recursion produced first-row destination probability $0.39$; adding the off-diagonal increments $0.20+0.25$ would give $0.45$, incorrectly ignoring the fact that the second update acts on the state distribution already changed at the first time.

Key takeaway: On an ordered event-time grid, the product integral is computed by starting at $P_0=I$ and applying $P_k=P_{k-1}(I+\Delta A_k)$ at each successive time. The recursion is a matrix update, and probability interpretation requires valid factors.

Check Your Understanding

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

For two event times, write the expanded expression for $P_2$ using $\Delta A_1$ and $\Delta A_2$.
A two-state increment is $\Delta A=\begin{pmatrix}-0.4&0.4\\0.2&-0.2\end{pmatrix}$. Write its factor $I+\Delta A$ and check its row sums and entries.
Suppose one row of an increment sums to $0.03$, rather than zero. What happens to the corresponding row sum of its factor, and what should be checked?
Why is adding two increment matrices not generally the same calculation as applying two successive product-integral updates?

Solutions. (1) $P_2=(I+\Delta A_1)(I+\Delta A_2)$, with the earlier factor on the left. (2) The factor is $\begin{pmatrix}0.6&0.4\\0.2&0.8\end{pmatrix}$; its entries are nonnegative and both rows sum to one. (3) The corresponding factor row sums to $1.03$, not one. Check that the diagonal increment is the negative sum of the off-diagonal increments and investigate any data-construction or numerical issue. (4) Successive updates multiply factors and carry the current transition matrix into the next time; adding increments omits this propagation and generally fails to produce the same matrix.