A stationary reading does not mean that nothing happens. Molecules keep moving and microscopic transitions continue at equilibrium. The useful distinction is between balanced activity and a steady state maintained by directed flows.
A population can settle while transitions continue
Consider a system with two possible states, A and B. Transitions from A to B occur at rate a, and transitions from B to A at rate b. Let p(t) be the probability of B. Its rate of change is dp/dt = a(1−p) − bp: gains from A minus losses from B.
The stationary probability is p* = a/(a+b). Solving the equation gives p(t) = p* + [p(0)−p*]e⁻⁽ᵃ⁺ᵇ⁾ᵗ. The relaxation time is 1/(a+b). Stationarity describes a probability distribution that no longer changes, while relaxation describes the approach to it.
Choose a = 1 s⁻¹ and b = 3 s⁻¹. Then p* = 1/4 and the relaxation time is 0.25 seconds. At stationarity, the forward probability flux is (3/4) × 1 = 0.75 s⁻¹ and the backward flux is (1/4) × 3 = 0.75 s⁻¹. Their difference vanishes, although their sum does not.
A detector following one realization would still observe jumps. A detector averaging many independently prepared systems could show an essentially constant fraction in B. Those are compatible observations at different descriptive levels.
Detailed balance adds a condition
In a finite Markov jump description with no variables odd under time reversal, detailed balance means that each stationary transition flux is matched by its reverse: πiwij = πjwji. Stationarity alone requires only that the total inflow and outflow at each state agree.
The difference becomes visible with three states arranged in a cycle. Give each clockwise transition rate 2 s⁻¹ and each counterclockwise transition rate 1 s⁻¹. The uniform distribution π = (1/3, 1/3, 1/3) is stationary: every state receives and loses probability at the same total rate.
Nevertheless, each edge carries net clockwise current (1/3) × 2 − (1/3) × 1 = 1/3 s⁻¹. There is continuing circulation. The three-state Shannon entropy remains ln 3, so monitoring that state entropy alone would miss the directed current.
A steady distribution can sustain production
For this simple jump model, the stationary path-irreversibility rate is the sum over the three undirected edges of current times the logarithm of the forward-to-backward stationary flux ratio. It is 3 × (1/3) ln 2 = ln 2 per second. Multiplying by kB supplies entropy units.
Calling that quantity thermodynamic entropy production requires a physical realization with the appropriate relation between transition rates and exchanged energy or other driving. The cyclic rate bias cannot describe an undriven, equal-energy system in equilibrium with one bath. A reservoir or external drive must sustain the bias in such a realization.
This distinction prevents two errors: equating constant state entropy with zero entropy production, and identifying every fitted probability current with a measured heat flow. Transition statistics and energetic accounting must be compared, not conflated.
Equilibrium is assessed within a regime
A finite observation window may hide slow relaxation. If an instrument resolves changes only larger than one percent, a trajectory can look stationary before it has reached its limiting distribution exactly. Conversely, equilibrium fluctuations can produce individual departures from an average without signalling a drift in the distribution.
A practical test combines repeated preparations, response to a small perturbation and measurements sensitive to currents. A stable mean alone does not distinguish equilibrium from a driven stationary process. The relevant timescale and the variables measured should be stated alongside the conclusion.
The supplied rates in our example already use seconds. A microscopic theory must explain their relation to a physical clock. Relabelling every transition as one completion establishes an event count, but waiting times and rate ratios remain additional observables.
What should a completion model reproduce?
The expanded DoT book asks for a physical reduction with a stationary reference state. This example adds a useful diagnostic: determine whether that state is an equilibrium distribution or a driven stationary distribution, and calculate the currents left invisible by the state probabilities.
A source-level construction could predict both relaxation rates and stationary circulation under specified controls. Agreement with only the final population would leave those dynamical tests open. Relating clocks, records and exchange processes would make the connection to inner generation more restrictive.
Sources and further reading
- David Tong, Kinetic Theory, for equilibrium and detailed balance. The finite-rate examples are calculated here.
- Continue with exchanged resources and information and thermodynamic entropy.

