Markovian vs Non-Markovian Dynamics
Every open quantum system carries a clock hidden inside its environment. When you trace out the bath, the qubit or spin you are left with evolves under a one-parameter family of completely positive trace-preserving maps, ρ_S(t) = Λ_t[ρ_S(0)], and the entire physics of decoherence turns on one question: does the environment remember? If the bath forgets each interaction before the system has moved appreciably, the future depends only on the present — the evolution is memoryless, or Markovian, and it is generated by a Lindblad/GKSL master equation with non-negative rates. If the bath holds a record and feeds it back, the future depends on the whole history — the evolution is non-Markovian, and coherence that leaked out can flow back in. Turning that intuition into mathematics forced the field to separate two structures classical probability had kept married. One is divisibility: can the map from 0 to t be cut at any intermediate time s into a legitimate channel from s to t? The other is monotone information loss: does the distinguishability of two states only ever decrease? For a classical Markov chain these coincide. Quantumly they split, because "legitimate channel" has two inequivalent meanings — merely positive, or completely positive — and the gap between them is exactly where quantum memory lives. This article builds the modern synthesis: CP-divisibility as the definition of quantum Markovianity, its equivalence with a GKSL generator whose rates never go negative, and the two operational witnesses — trace-distance backflow and Choi-matrix positivity loss — that detect when memory returns.
- The reduced dynamics of an open system is a family of CPTP maps Λ_t; it is Markovian when that family is CP-divisible — every intermediate propagator V_{t,s} in Λ_t = V_{t,s} Λ_s is itself completely positive and trace-preserving.
- CP-divisibility is equivalent to a GKSL master equation ρ̇ = 𝓛_t[ρ] whose canonical rates satisfy γ_k(t) ≥ 0 at every instant — the rates may depend on time, but must not go negative.
- Non-Markovianity is the failure of that condition, seen physically as information backflow: the trace distance between two evolving states, which no channel can increase, temporarily grows (the BLP witness).
- The two witnesses differ: trace-distance backflow detects loss of P-divisibility, while CP-divisibility is strictly stronger, so the Choi-based RHP measure catches negative rates that leave the trace distance untouched.
- Memory appears when the bath correlation time τ_E is no longer tiny against the relaxation time τ_R — structured spectra, strong coupling, or low temperature drive some γ_k(t) negative.
- The damped Jaynes–Cummings model makes the crossover exact: weak coupling (λ > 2γ_0) gives monotone Markovian decay; strong coupling (2γ_0 > λ) gives oscillatory revivals with negative rates.
- Key idea: Quantum Markovianity is CP-divisibility of the dynamical map family, Λ_t = V_{t,s} Λ_s with every propagator completely positive; equivalently a GKSL generator with γ_k(t) ≥ 0 for all t. Non-Markovianity is its violation, witnessed by transient growth of trace distance — information flowing back from the environment.
Quantum Markovianity is CP-divisibility of the dynamical map family, Λ_t = V_{t,s} Λ_s with each propagator V_{t,s} completely positive and trace-preserving — equivalently a time-local GKSL generator with γ_k(t) ≥ 0 for all t. Non-Markovianity is the violation of this, witnessed operationally by transient increases in trace distance (information backflow).
Where this lives today
The distinction has moved from foundational curiosity to a working tool in quantum technology. Superconducting and spin qubits couple to structured environments — two-level-system defects, 1/f charge and flux noise, nearby cavity modes — whose correlation times are not negligible on gate timescales, producing non-exponential T₁/T₂ decay and time-correlated errors; whether a memoryless error model suffices for a fault-tolerance threshold is decided empirically by testing complete positivity of intermediate maps through their Choi matrices in process tomography. The crossover has been engineered and observed directly: photonic experiments by Liu and co-workers realized a tunable open system whose trace distance was measured to switch from monotone decay to reviving, mapping out the transition of equation (7) by hand. Reservoir engineering turns the effect into a resource — photonic band gaps, cavity QED, and dynamical decoupling exploit memory to slow or reverse decoherence, and backflow can revive entanglement after "sudden death" and lift Markovian precision limits in metrology. Where the constant-rate Lindblad equation fails — strong coupling, low temperature, structured spectra — the workhorses are the time-convolutionless master equation and the hierarchical equations of motion (HEOM), which reproduce the negative rates signalling the semigroup's breakdown.
The physical picture
Picture the system and its environment as two accounts between which one quantity — distinguishability, the trace distance between two candidate states — is transferred. A completely positive channel can only move it from system to environment, never the reverse for the system alone; that is the contractivity of trace distance. Markovian dynamics drains the account steadily to zero. Non-Markovian dynamics lets the environment, which has been quietly recording, pay some back, and the visible signature is the trace distance ticking upward.
Divisibility is the algebra behind that cartoon. A divisible family is one you can chop at any instant into pieces V_{t,s} = Λ_t Λ_s^{-1}; Markovianity demands each piece be a legitimate channel on its own — completely positive, valid even acting on half of an entangled pair. When the bath has memory, the piece from s to t is not self-contained: completing it requires reaching into correlations the system already shares with the environment, so its Choi matrix acquires negative eigenvalues and complete positivity fails. The differential face of that failure is a rate γ_k(t) going momentarily negative — a time-reversed jump that partially undoes earlier dissipation, without ever making Λ_t un-physical. The second figure turns this into the quantity an experiment plots: the trace distance of an optimally chosen state pair, monotone in one regime and reviving in the other.
The core equations
The starting object is the reduced dynamical map. Unitary evolution of system-plus-bath, followed by a partial trace over the bath, produces a family of maps on the system alone — completely positive and trace-preserving for each t, provided system and bath start uncorrelated.
where: Λ_t — reduced dynamical map, CPTP for each t; U_t — the unitary e^{−iHt} of system plus environment; ρ_E — fixed initial environment state, uncorrelated at t = 0; Tr_E — partial trace over the environment. Markovianity is a statement not about a single Λ_t but about how the family fits together in time. Divisibility asks whether the evolution from 0 to t factors through any intermediate time s, and the character of the connecting propagator V_{t,s} is what splits the regimes.
where: V_{t,s} — the intermediate propagator carrying the state from time s to t. Reading: if Λ_s is invertible, V_{t,s} = Λ_t Λ_s^{−1} always exists; the physics is whether it is a legitimate channel. If every V_{t,s} is merely positive and trace-preserving the family is P-divisible; if every V_{t,s} is completely positive it is CP-divisible — the definition of quantum Markovianity. Whenever the map family is differentiable it obeys a time-local master equation. Writing that generator in canonical GKSL form isolates a set of instantaneous decay rates whose sign is the whole story.
where: H(t) — effective Hamiltonian, including any Lamb shift; L_k — jump (Lindblad) operators; γ_k(t) — instantaneous canonical rates. Reading: what distinguishes the regimes is not the time dependence but the sign of the rates. The precise theorem, due to Rivas–Huelga–Plenio and to Wolf–Cirac, ties the algebraic property in (2) to the analytic property in (3). It is the load-bearing equivalence of the whole subject.
where: the double arrow is an if-and-only-if. Reading: non-negativity of every canonical rate at every time is necessary and sufficient for Markovianity; a single rate dipping below zero on any interval breaks CP-divisibility — the exact meaning of "the environment has memory". The operational witness needs no access to the generator. It watches two states evolve and measures their distinguishability through the trace distance, whose time derivative defines an information flow rate.
where: ‖X‖_1 = Tr√(X†X) is the trace norm; D is the trace distance, the optimal single-shot probability of telling ρ_1 from ρ_2; σ(t) is the information flow rate. Reading: any CPTP map has σ ≤ 0 (Derivation B), so σ(t) > 0 is a direct sighting of information flowing back. Integrating the positive excursions of σ, maximized over the best input pair, gives the Breuer–Laine–Piilo measure; the divisibility-based Rivas–Huelga–Plenio measure integrates instead the rate at which the propagator's Choi matrix loses positivity.
where: the maximization runs over all initial pairs; the integral collects every interval on which distinguishability grows. Reading: 𝒩_BLP = 0 certifies no trace-distance backflow (P-divisibility); a positive value quantifies how much distinguishability the bath returned. Since CP-divisibility is strictly stronger than P-divisibility, 𝒩_BLP = 0 does not certify Markovianity — a subtlety the warning box returns to. A single exactly solvable model realizes all of this. A two-level atom radiating into a Lorentzian reservoir of spectral width λ and coupling scale γ_0 has an excited-state amplitude whose modulus squared controls both the decay and its reversal.
where: c(t) — excited-state amplitude, |c|² the survival probability; λ — reservoir spectral width, equal to the inverse correlation time 1/τ_E; γ_0 — coupling scale (the Markovian decay rate); d — effective oscillation frequency. Reading: when 2γ_0 > λ the discriminant is positive, d is real, |c(t)|² oscillates through zeros, and the exact rate γ(t) = −(2/|c|) d|c|/dt swings negative — the non-Markovian regime. When λ > 2γ_0, d is imaginary, the trig functions become hyperbolic, |c(t)|² decays monotonically, and γ(t) ≥ 0.
The goal is to show that the strongest form of Markovianity — a time-homogeneous, CP-divisible family — is forced by pure algebra into a one-parameter semigroup, and that demanding complete positivity at every time collapses its generator onto the canonical GKSL form with constant non-negative rates. This anchors equation (4): it identifies the memoryless limit with the textbook Lindblad equation and shows why "constant rates ≥ 0" is not an assumption but a consequence of divisibility plus complete positivity. Assume Λ_t is CP-divisible with time-independent propagators, V_{t,s} = Λ_{t−s}.
Step 1. Time-homogeneity means the propagator depends only on the elapsed interval, V_{t,s} = Λ_{t−s}, so composing evolutions gives Λ_{t+s} = Λ_t Λ_s together with Λ_0 = 𝟙 — the defining law of a one-parameter semigroup. Why: a time-invariant bath supplies no clock, so only elapsed time enters — the semigroup law.
Step 2. Continuity of a semigroup with a bounded generator yields \mathcal{L}=\lim_{\epsilon\to0^+}\frac{\Lambda_\epsilon-\mathbb{1}}{\epsilon},\qquad \Lambda_t=e^{\mathcal{L}t}. The whole family is now fixed by a single time-independent superoperator 𝓛. Why: differentiating the semigroup law at s = 0 gives one master equation ρ̇ = 𝓛[ρ].
Step 3. Requiring e^{𝓛t} to be completely positive and trace-preserving for every t ≥ 0 is a strong constraint. The GKSL theorem states that it holds if and only if 𝓛 has the canonical structure \mathcal{L}[\rho]=-i[H,\rho]+\sum_k\gamma_k\Big(L_k\rho L_k^\dagger-\tfrac12\{L_k^\dagger L_k,\rho\}\Big),\qquad \gamma_k\ge 0. Why: CP of the exponential ⟺ the Kossakowski matrix is positive semidefinite ⟺ every γ_k ≥ 0.
Step 4. With constant, non-negative rates the solution is an exponential relaxation with no revivals — the memoryless Markovian limit. Allowing the rates and jump operators to depend on time, while keeping γ_k(t) ≥ 0, generalizes this to CP-divisible (time-inhomogeneous Markovian) dynamics without spoiling any of the structure. Why: dropping only time-homogeneity, not positivity, gives the full Markovian class of (4). *Verified: reduces to the GKSL/Lindblad generator; consistent with Breuer–Petruccione §3.2 and Wolf–Eisert–Cubitt–Cirac (2008).*
Go deeper · Derivation B · 1 minTrace-distance contraction as the origin of the backflow witness
The aim is to prove that every CPTP map contracts trace distance, D(Φρ_1, Φρ_2) ≤ D(ρ_1, ρ_2), then read the inequality backward: because a genuine channel can only shrink distinguishability, any interval on which the trace distance grows certifies that the intermediate propagator was not even positive. This upgrades equation (5) from a definition into an experimentally accessible witness — it needs only two prepared states and a measurement of their overlap, never the microscopic generator. Let ρ_1, ρ_2 be states and Φ any CPTP map.
Step 1. Write the traceless Hermitian difference and split it into orthogonal positive and negative parts, \Delta=\rho_1-\rho_2=\Delta_+-\Delta_-,\qquad \Delta_\pm\succeq 0, with support orthogonality. Since Tr Δ = 0, the trace distance is D = ½ Tr|Δ| = Tr Δ_+. Why: the Jordan split turns the trace norm into traces of positive operators.
Step 2. Apply the map and use the triangle inequality, ‖Φ(Δ)‖_1 ≤ ‖Φ(Δ_+)‖_1 + ‖Φ(Δ_−)‖_1. Because Φ is positive, Φ(Δ_±) ⪰ 0, so each trace norm equals the bare trace, ‖Φ(Δ_±)‖_1 = Tr Φ(Δ_±) = Tr Δ_± by trace preservation. Why: positivity converts trace norms to traces; only positivity and trace preservation are used, so even a merely positive map suffices.
Step 3. Combining, ‖Φ(Δ)‖_1 ≤ Tr Δ_+ + Tr Δ_− = ‖Δ‖_1, hence D(Φρ_1, Φρ_2) ≤ D(ρ_1, ρ_2). Complete positivity was never needed — positivity and trace preservation alone give contraction. Why: needing no complete positivity is why the witness detects only the weaker P-divisibility.
Step 4. For divisible dynamics ρ_i(t) = V_{t,s} ρ_i(s). If V_{t,s} were positive and trace-preserving the same three steps would force D(t) ≤ D(s). Therefore an observed dD/dt > 0 on some interval proves at least one V_{t,s} fails positivity — the environment has returned distinguishability to the system. Why: contraction is one-way for any channel, so backflow unforgeably signals broken divisibility. *Verified: standard contractivity argument; matches Nielsen–Chuang Thm 9.2 and Wilde ch. 9.*
Go deeper · Derivation C · 1 minExact dephasing rate and its transient negativity
This derivation exhibits non-Markovianity in the simplest exactly solvable setting — a qubit coupled through σ_z to a bosonic bath, so populations are frozen and only coherence decays — and shows how the time-local dephasing rate is a running integral over the bath spectrum that can dip below zero, driving a momentary recoherence. The point is to see, in closed form, that memory is a property of the bath's spectral structure and temperature, not of any coarse approximation. For pure dephasing the off-diagonal element evolves as ρ_{01}(t) = e^{−Γ(t)} ρ_{01}(0) with a decoherence function Γ(t).
Step 1. The model Hamiltonian H=\frac{\omega_0}{2}\sigma_z+\sigma_z\sum_k g_k(b_k+b_k^\dagger)+\sum_k\omega_k b_k^\dagger b_k commutes with σ_z, [σ_z, H] = 0, so populations are conserved and only coherences move — the model is exactly solvable with no approximation. Why: freezing populations isolates dephasing, making the coherence exactly integrable.
Step 2. The bath enters only through its spectral density J(ω) = Σ_k g_k² δ(ω − ω_k), with the Ohmic family J(ω) = α ω e^{−ω/ω_c} as the standard structured example. Integrating the influence functional gives \Gamma(t)=\int_0^\infty \mathrm{d}\omega\, J(\omega)\coth\!\Big(\frac{\omega}{2k_BT}\Big)\frac{1-\cos\omega t}{\omega^2}. Why: each mode adds an independent phase kick; coth carries the thermal occupation.
Step 3. Differentiating gives the time-local rate that appears in ρ̇_{01} = −γ(t) ρ_{01}, \gamma(t)=\dot\Gamma(t)=\int_0^\infty \mathrm{d}\omega\, J(\omega)\coth\!\Big(\frac{\omega}{2k_BT}\Big)\frac{\sin\omega t}{\omega}. Why: the sin ωt kernel is not sign-definite, so a structured spectrum can drive γ(t) < 0.
Step 4. When γ(t) < 0 over an interval, Γ(t) decreases and |ρ_{01}| = e^{−Γ(t)} |ρ_{01}(0)| momentarily grows — partial recoherence, the dephasing face of backflow. A flat, unstructured reservoir gives γ(t) → const ≥ 0 and recovers Markovian exponential dephasing. Why: the sign of γ(t) is the sign of the coherence slope — negative rate means revival. *Verified: matches the exactly solvable dephasing model of Breuer–Petruccione §4.2 and the standard spin–boson literature.*
Limiting cases
The regime is set by the ratio of bath memory time τ_E to system relaxation time τ_R, and by the spectral structure of the reservoir:
| Case | Result | Meaning |
|---|---|---|
| Born–Markov, fast bath | τ_E ≪ τ_R | Flat or smooth spectrum; rates settle to constants γ_k ≥ 0; CP-divisible; exponential relaxation; the standard Lindblad equation applies. |
| Time-dependent Markovian | γ_k(t) ≥ 0 ∀t | Generator depends on time but every rate stays non-negative; CP-divisible; decay non-exponential yet memoryless; no backflow. |
| Non-Markovian, structured bath | τ_E ∼ τ_R, ∃t: γ_k(t) < 0 | Sharp spectral features, strong coupling, or low T; divisibility violated; trace-distance backflow and partial recoherence possible. |
| Strong-coupling coherent limit | 2γ_0 > λ (g ≫ κ) | Damped vacuum Rabi oscillations; repeated exchange of a quantum between system and cavity; strongly non-Markovian with recurring revivals. |
Worked examples
1 · Diagnosing the damped Jaynes–Cummings model (fully worked)
A two-level atom decays into a Lorentzian reservoir of width λ. Take coupling γ_0 = 1 (setting the time unit) and two reservoirs, λ = 5γ_0 and λ = 0.3γ_0. Classify each regime, and for the memory-bearing one find where the excited population first vanishes and confirm that a revival — hence a negative rate — follows.
Step 1. Read off the discriminant d = √(2γ_0λ − λ²) from equation (7). For λ = 5γ_0: 2γ_0λ − λ² = 10 − 25 = −15 < 0, so d is imaginary, the trig functions turn hyperbolic, and |c(t)|² decays monotonically — Markovian, γ(t) ≥ 0. For λ = 0.3γ_0: 2γ_0λ − λ² = 0.6 − 0.09 = 0.51 > 0, so d = √0.51 = 0.7141 is real and the amplitude oscillates — non-Markovian.
Step 2. Confirm the crossover criterion directly: the boundary is 2γ_0 = λ. Here 2γ_0 = 2, and 2 > 0.3 (strong coupling) but 2 < 5 (weak coupling), matching Step 1. The physical reading is τ_E = 1/λ: the strong-coupling bath has τ_E = 3.3/γ_0, comparable to the relaxation time, while the weak-coupling bath forgets in τ_E = 0.2/γ_0, far faster than the system moves.
Step 3. For the non-Markovian reservoir locate the first amplitude zero. The bracket in (7) vanishes when tan(dt/2) = −d/λ = −0.7141/0.3 = −2.381, whose first positive root is dt/2 = π − arctan(2.381) = 1.9686, giving t₁ = 2(1.9686)/0.7141 = 5.51/γ_0. At t₁ the excited population is exactly zero: the quantum has been fully handed to the cavity mode.
Step 4. Check for revival past the zero. At t = 8 the phase is dt/2 = 2.857, the bracket is cos 2.857 + 0.4201 sin 2.857 = −0.843, and e^{−0.3·8} = 0.0907, so |c(8)|² = 0.064 > 0 — the population has climbed back from zero. Since |c|² rose, the exact rate γ(t) = −(2/|c|) d|c|/dt is negative through the revival, the differential signature of backflow. The Markovian reservoir, by contrast, gives |c|² = 0.928, 0.720, 0.138 at t = 0.2, 0.5, 2 — monotone, never reviving.
2 · Trace-distance backflow and a lower bound on 𝒩_BLP
Using the same strong-coupling reservoir (γ_0 = 1, λ = 0.3), take the optimal Breuer–Laine–Piilo pair — two antipodal states on the equator of the Bloch sphere — for which the amplitude-damping dynamics gives D(t) = |c(t)|. Show that D is non-monotone and put a number on the returned distinguishability.
Step 1. For antipodal equatorial states the Bloch-vector difference is purely transverse with length 2, and amplitude damping scales the transverse difference by |c(t)|, so D(t) = ½·|c(t)|·2 = |c(t)|. This pair is the one that maximizes the BLP integral for amplitude damping, so its backflow lower-bounds 𝒩_BLP.
Step 2. Tabulate the distinguishability from the |c(t)|² values of Example 1: D(0) = 1, D(2) = √0.583 = 0.764, D(5.51) = 0 (states become momentarily indistinguishable), and D(8) = √0.064 = 0.253.
Step 3. The trace distance therefore falls from 1 to 0 as information drains into the cavity, then climbs back to 0.253 — an unambiguous σ(t) > 0 interval. A CPTP map can never do this on its own, so the environment returned distinguishability to the system.
Step 4. Bound the measure. Over the revival from t₁ = 5.51 to t = 8 the net rise is ΔD = 0.253, and since 𝒩_BLP integrates every positive excursion, 𝒩_BLP ≥ 0.25 for this reservoir; later oscillations of |c(t)| add smaller revivals. The weak-coupling reservoir gives D(t) = |c(t)| strictly decreasing, so 𝒩_BLP = 0 — no backflow, consistent with its Markovian classification.
Physics of the portrait
Quantum non-Markovianity is a young idea built on an old one. The Markovian master equation was pinned down in 1976; making its failure precise, and turning "the bath remembers" into a measurable number, took another three decades and a deliberate untangling of divisibility from information flow.
Origins · 1 minquantum non-Markovianity and how this was discovered
Non-Markovianity is one of those concepts whose definition was the discovery. For decades "Markovian" meant, loosely, "described by a Lindblad equation" and "non-Markovian" meant "whatever isn't" — a negative, approximation-dependent notion with no agreed witness. The structural foundation was laid in 1976, when Göran Lindblad, and independently Vittorio Gorini, Andrzej Kossakowski and George Sudarshan, derived the most general generator of a completely positive quantum dynamical semigroup — the GKSL form with its non-negative rates. But a semigroup is only the memoryless limit; it cannot grade a process that has memory. The reframing came in two near-simultaneous strokes. In 2009 Heinz-Peter Breuer, Elsi-Mari Laine and Jyrki Piilo gave an operational definition: non-Markovianity is information backflow, present exactly when the trace distance between two evolving states increases — something no completely positive channel can cause. A year later Ángel Rivas, Susana Huelga and Martin Plenio anchored the definition in the algebra of the maps: Markovianity is CP-divisibility, its breakdown measured by the rate at which an intermediate propagator's Choi matrix loses positivity, equivalently a canonical rate γ_k(t) going negative — the criterion Michael Wolf, Jens Eisert, Toby Cubitt and Ignacio Cirac had sharpened in 2008. The two schools, BLP and RHP, do not always agree — trace-distance backflow detects only the weaker P-divisibility — and that disagreement itself clarified the landscape. The 2016 Reviews of Modern Physics colloquium by Breuer, Laine, Piilo and Bassano Vacchini consolidated it into today's framework, in which a folklore constant became a computed, resource-theoretic function of the environment's spectrum.
Going further
Texts and reviews. Breuer & Petruccione, The Theory of Open Quantum Systems (2002), ch. 3–4 and 9–10 · Rivas, Huelga & Plenio, "Quantum non-Markovianity: characterization, quantification and detection", Rep. Prog. Phys. 77, 094001 (2014) · Breuer, Laine, Piilo & Vacchini, "Colloquium: Non-Markovian dynamics in open quantum systems", Rev. Mod. Phys. 88, 021002 (2016) · Nielsen & Chuang, Quantum Computation and Quantum Information, ch. 8 · Wilde, Quantum Information Theory, 2nd ed., ch. 4.
Provenance. A verifier should: (1) confirm CP-divisibility ⟺ γ_k(t) ≥ 0 against Rivas–Huelga–Plenio (2014) and Wolf et al., PRL 101, 150402 (2008); (2) reproduce the trace-distance contraction from Nielsen–Chuang Thm 9.2; (3) check the Jaynes–Cummings amplitude (7) and the exact dephasing rate against Breuer–Petruccione §4.2; (4) verify the worked-example numbers — d = √0.51 = 0.7141, t₁ = 5.51/γ_0, |c(8)|² = 0.064, D-values 1, 0.764, 0, 0.253 — recomputed directly from equation (7). All formulas are textbook-standard; the only numerical inputs are the reservoir parameters λ/γ_0.
Prerequisites. Open Quantum Systems and Reduced Dynamics · CPTP Maps and Quantum Operations · The Lindblad/GKSL Master Equation · Born–Markov and Secular Approximations. Leads to. Decoherence · Relaxation and Dephasing: T₁ and T₂ · Process Tomography and the Choi–Jamiołkowski Isomorphism.
Ask the tutor: Ask me to walk through why a negative decay rate breaks divisibility without ever making the state un-physical, or to compute the trace-distance backflow for a reservoir of your choice.
Is a time-dependent master equation the same thing as non-Markovian dynamics?
No — this is the most common confusion in the subject. A generator 𝓛_t can depend on time and still be perfectly memoryless: what matters is not that the rates vary but whether they go negative. If every canonical rate satisfies γ_k(t) ≥ 0 for all t, the family is CP-divisible and hence Markovian, even though the decay is non-exponential. Non-Markovianity requires at least one rate to dip below zero on some interval, breaking divisibility and letting information flow back. "Time-local" is a property of the equation; Markovian versus non-Markovian is a property of the sign of its rates.
What is the difference between P-divisibility and CP-divisibility, in plain terms?
Both ask that the propagator V_{t,s} from s to t be a legitimate operation, but with different standards. P-divisibility asks only that V_{t,s} be positive and trace-preserving — it maps valid states of the system to valid states. CP-divisibility asks the stronger condition that V_{t,s} be completely positive, remaining valid even when applied to half of an entangled pair, which is the real requirement for a physical channel. CP-divisibility defines Markovianity. The gap between the two is where certain non-Markovian processes hide: a propagator can be positive (trace distance never grows) yet not completely positive (a rate is negative), which is why the trace-distance witness can miss what a full divisibility test catches.
How can a decay rate be negative without violating the laws of physics?
Because the physical constraint is on the full map Λ_t, not the instantaneous rate. Complete positivity and trace preservation of Λ_t hold throughout — the Jaynes–Cummings population stays between 0 and 1 at every instant. A negative γ_k(t) is a property of the generator, not the state: the rate of an effective time-reversed jump that partially undoes earlier dissipation, pulling coherence back from the environment. It breaks divisibility of the family (the intermediate propagator is not a valid channel) but never the physicality of the state. Negative rates diagnose memory, not a breakdown of quantum mechanics.
What sets the boundary between the two regimes microscopically?
The competition of two timescales: the bath correlation time τ_E, over which ⟨B(τ)B(0)⟩ decays, and the relaxation time τ_R, over which the system's state changes. When τ_E ≪ τ_R the bath forgets each interaction before the system moves, so every step sees a fresh environment — Born–Markov holds and rates settle to non-negative constants. When τ_E approaches τ_R, past system states remain imprinted on the bath and re-influence the future, so memory and negative rates appear. Spectrally, τ_E is large when the spectral density is sharply structured — a narrow cavity resonance, a band edge, or low temperature — the natural home of non-Markovian behavior.
Do the BLP and RHP measures ever disagree, and which should I use?
Yes, and the disagreement is informative. RHP (divisibility) is the stricter test: it flags any interval with a negative canonical rate, declaring non-Markovianity whenever CP-divisibility fails. BLP (trace distance) only fires when distinguishability actually flows back, which requires the weaker P-divisibility to fail. So every BLP-non-Markovian process is RHP-non-Markovian, but not conversely — there are processes with negative rates and zero trace-distance backflow. Use RHP when you want the rigorous algebraic classification; use BLP when you care about the operational, directly measurable quantity — whether an experiment can recover distinguishability, which is what matters for metrology or entanglement protection.
When does the ordinary Lindblad equation stop being trustworthy?
When the Born–Markov and secular approximations behind it break down: when τ_E is no longer small against τ_R, when the coupling is strong (2γ_0 > λ), or when the temperature is low enough that the bath spectrum is effectively structured. The tell-tale sign is a correct time-local treatment producing canonical rates γ_k(t) that go negative — which a constant-rate semigroup cannot represent. There you replace Lindblad with memory-keeping methods: the time-convolutionless (TCL) master equation, the Nakajima–Zwanzig equation, or numerically exact schemes like the hierarchical equations of motion (HEOM), which reproduce the revivals and negative rates the GKSL semigroup cannot.