Why a system with memory can drift away from equilibrium ✎

Original: "Relative Entropy Revivals Caused by Memory-Induced Loss of Divisibility in Classical Non-Markovian Dynamics"
· Koichi Nakagawa
arXiv:2602.09054 · 2026-02-07 · arXiv non-exclusive · independent analysis · 1 min
If a system remembers its past, it can temporarily drift away from equilibrium before returning to it again.

The orchestra plays in sync, but a hidden conductor changes the beat for a second. In the same way, a system moving toward equilibrium suddenly takes a step back.

Usually, if a system does not remember its past, it approaches equilibrium monotonically, like sand in an hourglass. But if it has memory, it can temporarily drift away from equilibrium. People had noticed this before, but they did not understand why it happens.

Memory works like a hidden conductor: it does not change the music, but it delays some of the beats. Because of this, the orchestra momentarily loses its rhythm.

The author built simple models and found that if the memory is long, delayed feedback temporarily reverses the motion toward equilibrium. Then the usual proof of monotonicity stops working, and the distance to equilibrium can temporarily grow.

The orchestra is back in unison. Now it is clear that the stumble was not an accident but the work of memory, which briefly returned the system to its past.

In the Markovian case, relative entropy acts as a Lyapunov function that guarantees an irreversible approach to equilibrium. A loss of divisibility means the evolution cannot be split into successive Markovian steps, and then the familiar arrow of time can briefly turn backwards.

\frac{dP(t)}{dt} = \int_0^t K(t-\tau) P(\tau) d\tau
The rate of change of the probabilities at time t depends on the entire previous history through the kernel K.
K(t) = \gamma e^{-\gamma t} R
Memory decays exponentially at rate γ, while R is the Markovian generator.
F(t) = D_{KL}(P(t) \| \pi) = \sum_i P_i(t) \ln \frac{P_i(t)}{\pi_i}
A measure of the information distance of the current distribution from the stationary one, π.
Scientists
Ludwig BoltzmannRudolf ClausiusJacob BekensteinStephen HawkingJames Clerk MaxwellEdward Witten
Concepts
boltzmann entropy entropy entropic gravity arrow of time Markov chain relaxation damping rate energy gap relaxation time logarithmic correction irreversibility entropic shadow