Product Integral: Non-trivial example

This text illustrates the behavior of the product (multiplicative) integral using a straightforward, non-trivial example. It demonstrates why the classical integral fails when matrices do not commute at different times.

Problem Formulation

Consider a hypothetical biological model with two populations (e.g., cell types $X$ and $Y$) that transition into each other. However, the intensity of this transition changes abruptly over time. We seek the transition matrix $P(0, 2)$ from time $0$ to time $2$ for a system governed by a piecewise-constant intensity matrix function $A(t)$ defined as:

For time $t \in [0, 1)$: $$A_1 = \begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix}$$ (transformation of cells $Y \to X$)
For time $t \in [1, 2]$: $$A_2 = \begin{pmatrix} 0 & 0 \\ 1 & 0 \end{pmatrix}$$ (transformation of cells $X \to Y$)

Key Observation: The matrices $A_1$ and $A_2$ do not commute. The result of their multiplication depends strictly on their order: $$A_1 A_2 = \begin{pmatrix} 1 & 0 \\ 0 & 0 \end{pmatrix} \quad \neq \quad A_2 A_1 = \begin{pmatrix} 0 & 0 \\ 0 & 1 \end{pmatrix}$$

1. Calculation via Product Integral

Since the matrix $A(t)$ is piecewise constant, the multiplicative property allows the product integral to decompose into a standard matrix product of two separate matrix exponentials. The calculation follows a chronological order, where operators corresponding to later times are ordered to the left:

$$P(0, 2) = \prod_0^2 \big(I + A(t)dt\big) = \left( \prod_1^2 \big(I + A_2 dt\big) \right) \cdot \left( \prod_0^1 \big(I + A_1 dt\big) \right) = e^{A_2 \cdot 1} \cdot e^{A_1 \cdot 1}$$

Because $A_1^2 = 0$ and $A_2^2 = 0$, the Taylor series expansion for both exponentials truncates immediately after the first linear term:

$$e^{A_1} = I + A_1 = \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix}$$ $$e^{A_2} = I + A_2 = \begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix}$$

Multiplying these two matrices chronologically yields the correct final transition matrix:

$$P(0, 2) = e^{A_2} \cdot e^{A_1} = \begin{pmatrix} 1 & 0 \\ 1 & 1 \end{pmatrix} \begin{pmatrix} 1 & 1 \\ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & 1 \\ 1 & 2 \end{pmatrix}$$

2. Calculation via Peano Series

Alternatively, the product integral can be evaluated using the Peano series of nested integrals, which inherently preserves time-ordering ($t_1 \ge t_2 \ge \dots$):

$$P(0, t) = I + \int_0^t A(t_1) dt_1 + \int_0^t \int_0^{t_1} A(t_1) A(t_2) dt_2 dt_1 + \dots$$

Integrating term-by-term for the upper bound $t=2$, we obtain:

0th Term (Identity): $$I = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}$$
1st Term (Single Integral): $$\int_0^2 A(t_1) dt_1 = \int_0^1 A_1 dt_1 + \int_1^2 A_2 dt_1 = A_1 + A_2 = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}$$
2nd Term (Double Integral): The integration region is restricted by $2 \ge t_1 \ge t_2 \ge 0$. A non-zero contribution occurs only within the rectangle where $t_1 \in [1, 2]$ and $t_2 \in [0, 1]$: $$\int_1^2 \int_0^1 A_2 A_1 \, dt_2 dt_1 = A_2 A_1 \int_1^2 \int_0^1 1 \, dt_2 dt_1 = A_2 A_1 = \begin{pmatrix} 0 & 0 \\ 0 & 1 \end{pmatrix}$$
3rd and Higher Terms: All higher-order integrands contain matrix products that vanish (e.g., involving $A_1^2$ or $A_2^2$), resulting in zero matrices.

Summing all the non-zero terms yields the exact same matrix transition operator as obtained via the product integral:

$$P(0, 2) = \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} + \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix} + \begin{pmatrix} 0 & 0 \\ 0 & 1 \end{pmatrix} = \begin{pmatrix} 1 & 1 \\ 1 & 2 \end{pmatrix}$$

Why the Naive Approach Fails

If we attempt to evaluate the system by first using a standard additive integral and then substituting the result into a matrix exponential ($P = e^{\int_0^2 A(t)dt}$), we get: $$P_{naive} = e^{A_1 + A_2} = e^{\begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}} = \begin{pmatrix} \cosh(1) & \sinh(1) \\ \sinh(1) & \cosh(1) \end{pmatrix} \approx \begin{pmatrix} 1.54 & 1.17 \\ 1.17 & 1.54 \end{pmatrix}$$ This result is fundamentally incorrect. Classical additive integration sums up the matrices over time while completely ignoring their chronological order, failing to capture that the system was subjected to $A_1$ before $A_2$.

Generalization to Continuous Functions

While piecewise-constant matrix regimes allow explicit segmentation into distinct matrix exponential components, continuous operations present a deeper analytical layer. If we do not have constant-like functions, the Peano series is necessary. Without a step function to rely on, the infinite system of nested chronological dependencies can only be calculated comprehensively via this expansion framework.

Core Concepts Summary

Time-ordering is critical in systems with non-commuting operators. This is a defining characteristic of both quantum mechanics and the time-dependent transition intensities in Markov processes originally studied by Hostinský.