Why the Order of Intensities Matters
For a time-homogeneous Markov model, a constant generator \(Q\) produces the transition matrix \(P(0,t)=e^{tQ}\). When the generator changes with time, it is tempting to replace \(tQ\) by the integrated generator \(\int_0^t Q(u)\,du\) and exponentiate. That shortcut is valid when the generators at different times commute, but it can fail when their order matters. In a multiple-state model, the sequence of possible state changes can affect the final distribution even if the total accumulated intensities are unchanged.
Throughout, use the row-vector convention: probability distributions are row vectors, \(Q(t)\) acts on the right, and \(P(s,t)\) is the transition matrix from time \(s\) to time \(t\). The forward equation is \(\frac{\partial}{\partial t}P(s,t)=P(s,t)Q(t)\), with \(P(s,s)=I\). Each \(Q(t)\) is a generator: its off-diagonal entries are transition intensities and each row sums to zero.
If all the generators commute pairwise, then the solution is \(P(s,t)=\exp\left(\int_s^t Q(u)\,du\right)\). Without that commutation condition, the integral does not retain the order in which the generators acted. A time-ordered solution, or an approximation that preserves chronological order, is needed instead.
A Two-Interval Calculation Shows the Failure
Consider a hypothetical two-state model with states 0 and 1. During the first interval, transitions occur only from 0 to 1; during the second, they occur only from 1 to 0. In each interval the active transition intensity is \(\ln 2\) per time unit and the interval has length one. The state not listed as the origin has no outgoing transition during that interval. These are piecewise-constant Markov generators, so each interval transition matrix is the matrix exponential of its generator times the interval length.
Worked Example: Compare the Chronological Product with One Exponential
Method. Exponentiate each constant generator, multiply the interval matrices in chronological order, and compare the result with the exponential of the integrated generator. The common mistake to avoid is treating a sum of noncommuting generators as though it were their chronological action.
The generators are \(Q_1=(\ln 2)\begin{pmatrix}-1&1\\0&0\end{pmatrix}\) and \(Q_2=(\ln 2)\begin{pmatrix}0&0\\1&-1\end{pmatrix}\). For a one-way transition at intensity \(\lambda\) over duration \(h\), the probability of no transition is \(e^{-\lambda h}\), and the probability of a transition is \(1-e^{-\lambda h}\). Here \(e^{-(\ln 2)(1)}=1/2\). Therefore,
The first interval acts before the second, so under the row-vector convention \(P(0,2)=P_1P_2\), not \(P_2P_1\). Multiplication gives
Now add the generators, since each interval lasts one unit. The integrated generator is \(Q_1+Q_2=(\ln 2)\begin{pmatrix}-1&1\\1&-1\end{pmatrix}\). The symmetric matrix inside this expression has eigenvalues 0 and \(-2\); hence its exponential over the stated scale is
The matrices differ: for example, the probability of ending in state 1 from state 0 is \(1/4\) from the chronological product but \(3/8\) from the integrated-generator exponential. The commutator is nonzero. Writing \(c=\ln 2\), direct multiplication gives \(Q_1Q_2=c^2\begin{pmatrix}1&-1\\0&0\end{pmatrix}\) and \(Q_2Q_1=c^2\begin{pmatrix}0&0\\-1&1\end{pmatrix}\), so \(Q_1Q_2\ne Q_2Q_1\). The product is the correct transition matrix for the stated order of intervals.
Chronological Products for Piecewise-Constant Intensities
Suppose \(Q(t)=Q_k\) on each successive interval \([t_{k-1},t_k]\), with length \(\Delta_k=t_k-t_{k-1}\). The interval transition matrix is \(e^{Q_k\Delta_k}\). Applying the row-vector forward equation interval by interval gives
The order follows from composition: a distribution at \(t_0\) is first multiplied by the first interval matrix, then by the second, and so on. Reversing the matrices describes a different ordering of intensities and may produce a different transition matrix. If the generators commute pairwise, the product can be combined into the exponential of their sum; otherwise, keep the product in chronological order.
Worked Example: Multiply Three Different Intensity Intervals
Method. Find each one-way interval transition matrix, then multiply left to right in time order. This calculation avoids the trap of combining all three generators before checking commutation.
In a hypothetical two-state model, let the first interval have length one and only a \(0\to1\) intensity of \(\ln 3\). The second has length one and only a \(1\to0\) intensity of \(\ln 2\). The third has length one and only a \(0\to1\) intensity of \(\ln 2\). All other outgoing intensities are zero in the relevant intervals. Since \(e^{-\ln 3}=1/3\) and \(e^{-\ln 2}=1/2\), the interval matrices are
First multiply the first two intervals:
Then apply the third interval:
Thus, starting in state 0, the probability of ending in state 1 after all three intervals is \(2/3\). As an arithmetic check, each row of the final matrix sums to one. That check confirms the row totals but does not make the integrated-generator shortcut valid.
Why Time Ordering Appears in the Solution
The reason for the product order can be seen by expanding the forward equation. For a short time step \(dt\), the transition matrix over that step is approximately \(I+Q(t)\,dt\). Over two successive small steps, the earlier factor is on the left: \((I+Q(t_1)dt)(I+Q(t_2)dt)\), where \(t_1<t_2\). In the continuous limit, the second-order term therefore preserves the chronological order of the matrices.
For \(s\le t\), repeated integration of the forward equation gives the beginning of the time-ordered expansion:
By contrast, expanding \(\exp(\int_s^t Q(u)\,du)\) gives a second-order term \(\frac12(\int_s^t Q(u)\,du)^2\). Splitting that square into the two possible orders of each pair of times shows that the exponential averages the products \(Q(u_1)Q(u_2)\) and \(Q(u_2)Q(u_1)\). The forward equation requires the first product when \(u_1<u_2\). The two expressions agree at this order if the commutators vanish; when they do not, their difference is
This calculation justifies the commuting-generator result: when every pair commutes, chronological ordering no longer changes the products, and the ordinary exponential expansion matches the time-ordered expansion. Nonzero commutators identify the leading source of discrepancy; they do not, by themselves, give the full transition matrix.
Worked Example: Calculate a Leading Commutator Correction
Method. For a smoothly varying generator, use the second term in the time-ordered logarithm to measure the leading effect of ordering. This is a correction to an approximation, not an exact transition matrix unless higher terms are also accounted for.
Consider a hypothetical model on \(0\le t\le1\) with \(Q(t)=A+tB\), where \(A=\begin{pmatrix}-2&2\\1&-1\end{pmatrix}\) and \(B=\begin{pmatrix}1&-1\\-1&1\end{pmatrix}\). The transition intensities are \(2-t\) from state 0 to state 1 and \(1-t\) from state 1 to state 0. Both are nonnegative throughout the interval, so these are valid generators.
The integral of the generator is
Next compute the commutator. Direct multiplication gives \(AB=\begin{pmatrix}-4&4\\2&-2\end{pmatrix}\) and \(BA=\begin{pmatrix}-3&3\\3&-3\end{pmatrix}\), so \([A,B]=\begin{pmatrix}-1&1\\-1&1\end{pmatrix}\). At times \(u_1<u_2\),
The leading commutator contribution to the logarithm of the transition matrix is therefore
Thus the first two terms in the time-ordered logarithm are \(\int_0^1 Q(t)\,dt+\frac1{12}[A,B]\). The correction is nonzero, so exponentiating only the integrated generator omits an ordering effect. Further terms may also matter for an accurate numerical transition matrix; the calculation here identifies the leading correction rather than claiming an exact result.
Practical Choices and Common Traps
For piecewise-constant intensities, the chronological matrix product is direct and exact under the stated model. For a continuously varying generator, one may approximate the solution by dividing time into short intervals, selecting a representative generator on each interval, exponentiating it over that interval, and multiplying the resulting matrices in time order. Refining the intervals improves the approximation when the generator varies regularly. Numerical integration of the Kolmogorov equations is another way to calculate the transition matrix.
The central exam trap is to integrate every intensity and then exponentiate as if all state-flow mechanisms acted in an interchangeable order. A correct calculation first specifies the convention and the time sequence. It then checks whether the generators commute before using a single exponential. If they do not, use time ordering or a chronological product. This distinction matters because changing the order can change which states are reached early enough to experience later transitions.
- Do not reverse the product. With row distributions and \(P'=PQ\), the earliest interval matrix is the leftmost factor.
- Do not treat equal total intensity as equal transition behavior. Different state-to-state routes can interact through the changing state distribution.
- Do not mistake a small commutator for a zero commutator. It may justify an approximation, but it does not make the integrated exponential exact.
Check Your Understanding
For each problem, use row-vector convention and preserve chronological order where needed.
- A two-state model has a one-unit interval with only \(0\to1\) intensity \(\ln 2\), followed by a one-unit interval with only \(1\to0\) intensity \(\ln 2\). Find the full transition matrix. (Answer: \(\begin{pmatrix}3/4&1/4\\1/2&1/2\end{pmatrix}\))
- Reverse the order in question 1: first only \(1\to0\), then only \(0\to1\), with intensity \(\ln 2\) in each one-unit interval. Find the full transition matrix. (Answer: \(\begin{pmatrix}1/2&1/2\\1/4&3/4\end{pmatrix}\))
- A model has a half-unit interval with only \(0\to1\) intensity \(\ln 4\), followed by a half-unit interval with only \(1\to0\) intensity \(\ln 9\). Find the transition matrix. (Answer: \(\begin{pmatrix}5/6&1/6\\2/3&1/3\end{pmatrix}\))
- For \(Q(t)=A+tB\) on \(0\le t\le1\), let \(A=\begin{pmatrix}-1&1\\0&0\end{pmatrix}\) and \(B=\begin{pmatrix}0&0\\1&-1\end{pmatrix}\). Find \([A,B]\) and the leading commutator correction \(\frac1{12}[A,B]\). (Answer: \([A,B]=\begin{pmatrix}1&-1\\-1&1\end{pmatrix}\), correction \(=\begin{pmatrix}1/12&-1/12\\-1/12&1/12\end{pmatrix}\))
- Two successive intervals have generators \(Q_1=\begin{pmatrix}-1&1\\0&0\end{pmatrix}\) and \(Q_2=\begin{pmatrix}-2&2\\0&0\end{pmatrix}\), each of length one. Do these generators commute, and may their interval exponentials be combined into one exponential of \(Q_1+Q_2\)? (Answer: They commute; yes, \(e^{Q_1}e^{Q_2}=e^{Q_1+Q_2}\).)