Non-Markovian Dynamics (Overview)

When the bath remembers

The Lindblad equation rests on the Markov assumption: the bath forgets the system far faster than the system evolves, so the future depends only on the present. Non-Markovian dynamics arise whenever this fails — when the bath has a memory time comparable to the system's, and information that leaked into the bath can flow back into the system. This happens with structured baths (a qubit in a cavity, an emitter near a photonic band edge), strong coupling, or low temperature, where the bath correlation time $\tau_B$ is no longer negligible.

Divisibility: the precise dividing line

A clean characterization uses the family of maps $\Lambda_{t}$ taking $\rho(0)\mapsto\rho(t)$ . The process is CP-divisible if it can be split into independent completely-positive intermediate maps for any intermediate time $s$ ,

\Lambda_t = V_{t,s}\,\Lambda_s, \qquad V_{t,s}\ \text{completely positive for all } t\ge s\ge 0.

Markovian (Lindblad) evolution is exactly the CP-divisible case with a time-independent generator. When CP-divisibility breaks — when some $V_{t,s}$ fails to be completely positive — the dynamics are non-Markovian. Operationally this can be detected through a time-dependent master equation whose Lindblad-like rates $\gamma_k(t)$ become temporarily negative:

\dot\rho = -\tfrac{i}{\hbar}[H(t),\rho] + \sum_k \gamma_k(t)\!\left(L_k\rho L_k^\dagger - \tfrac12\{L_k^\dagger L_k,\rho\}\right),\qquad \gamma_k(t) < 0 \ \text{possible.}

A momentarily negative rate signals memory: the channel is briefly undoing decoherence.

Information backflow

A complementary, measurable signature is the behaviour of distinguishability between two initial states, quantified by the trace distance $D(\rho_1,\rho_2)=\tfrac12\lVert\rho_1-\rho_2 \rVert_1$ . Under any Markovian (CP-divisible) evolution, distinguishability can only decrease — information flows monotonically from system to bath. A temporary increase of the trace distance means information is flowing back from the bath to the system, the hallmark of non-Markovianity in the Breuer–Laine–Piilo measure:

\mathcal{N} = \max_{\rho_{1},\rho_{2}} \int_{\frac{dD}{dt}>0}\frac{dD}{dt}\,dt .

Models and methods

Because there is no universal time-local generator, non-Markovian dynamics use heavier machinery: the exact Nakajima–Zwanzig equation with a memory kernel $\mathcal{K}(t-s)$ convolved with the history, the time-convolutionless (TCL) expansion that restores time-locality order by order, pseudomode and reaction-coordinate mappings that enlarge the system to re-Markovianize it, and numerically exact tensor-network/HEOM methods. The exactly solvable damped Jaynes–Cummings model (a qubit in a leaky cavity) is the standard testbed, showing oscillatory coherence revival when the qubit reabsorbs its own emitted photon.

Why it matters

Far from a curiosity, engineered memory effects can help: information backflow can partially restore lost coherence, and non-Markovian baths are a resource in some quantum-control and quantum-metrology schemes. Recognizing when the Markov approximation breaks — and reaching for the right tool — is essential whenever a real device couples to a structured environment rather than the idealized flat continuum assumed by the Lindblad equation.