An environment can remember what a quantum system did and later influence it again. The reduced system may then evolve differently even when its current density matrix looks the same. Non-Markovian quantum dynamics is the mathematics needed to describe and test that retained influence.
Non-Markovian dynamics connects memory kernels, Nakajima–Zwanzig projection equations, time-convolutionless generators, complete-positive divisibility, information backflow and multitime quantum processes. These are related but not identical descriptions. A master equation that refers only to the present state can still describe memory. A negative instantaneous rate can coexist with a completely positive evolution from the initial time. A revival of distinguishability can occur under classical coloured noise as well as through a quantum environment.
The central proposition of this guide is: define which memory question is being asked before choosing a mathematical test. We will derive an exact memory-kernel equation, solve two finite-environment models, construct a channel that separates two popular definitions, and explain what experiments can and cannot infer from a limited set of time traces. The goal is a calculation that keeps the initial conditions, measurement protocol and physical meaning visible.
Your 50-second route
Need the physical idea? Read what the system boundary removes and the exact one-qubit environment. Need the formal derivation? Go to Nakajima–Zwanzig projection and time-local generators. Testing a memory claim? Compare trace-distance backflow, the Pauli counterexample and multitime interventions. Learning by doing? Finish the worked laboratory.
Open the reading map and conventions
Part I starts from unitary evolution of system plus environment and derives exact reduced equations. Part II solves dephasing and excitation exchange with a finite environment. Part III separates CP divisibility, distinguishability revival and operational multitime memory. Part IV covers approximations, collision models, numerical checks and experimental design. All examples use ℏ=1. X,Y,Z are Pauli matrices. A dynamical map acts on density operators; a superoperator such as a memory kernel acts on operators rather than directly on state vectors.
Guide 6 supplies channels and Kraus operators. Guide 52 supplies Markovian semigroups. Guide 54 supplies conditional measurements and filtering. Here we ask what changes when the environment cannot be replaced by a memoryless reservoir.
A complete positive map from the beginning does not guarantee complete positive maps between every pair of later times. A present-time differential equation does not prove that the past is irrelevant.
1. The system boundary determines what is hidden
Write the total Hamiltonian as Htot=HS+HE+Hint. The joint density operator ϱ evolves unitarily when this enlarged model is closed: ϱ(t)=U(t)ϱ(0)U(t)†. The observed system state is ρS(t)=TrEϱ(t).
Taking a partial trace removes access to environmental variables and system-environment correlations. Those variables do not cease to exist. If they later affect the system, a description based only on its current density matrix can be incomplete for predicting future experiments.
Memory is therefore relative to the chosen reduced description. A resonator treated as part of the environment may store information and return it to an atom. If the resonator is included in the system, the enlarged model may admit a simpler Markovian description with respect to a still larger external bath.
This does not make memory arbitrary or unreal. It says exactly which degrees of freedom the reduced model has eliminated. A useful experiment specifies that boundary, the accessible controls and the observations to be predicted. Reviews of non-Markovian open-system methods organise their techniques around precisely this distinction between reduced and enlarged descriptions. [1]
2. A fixed initial environment gives a channel from the initial time
Assume ϱ(0)=ρS(0)⊗σE, where σE is fixed independently of the input state. Then Φt(ρ)=TrE[U(t)(ρ⊗σE)U(t)†] is completely positive and trace preserving for every t.
This follows from preparation of a fixed ancillary state, unitary conjugation and partial trace, each a physical operation. It remains true when the environment is small, strongly coupled or capable of returning information. Complete positivity from time zero is not the same statement as memorylessness.
At a later time s, the environment generally depends on the initial preparation and may be correlated with the system. It cannot automatically be replaced by the same σE while keeping the actual experiment unchanged. That is the obstruction behind many failures of a simple composition law.
When comparing two input states to test distinguishability, use the same initial environmental preparation and the same controls. Otherwise the experiment changes more than the system input, and a revival can be attributed to an uncontrolled change of the overall process.
3. Initial correlations require a different statement
If the initial system and environment are already correlated, a system density matrix by itself does not specify their joint state. Two joint preparations can have the same ρS and different correlations, leading to different later reduced states under the same total unitary.
One must therefore specify a preparation procedure, a compatible family of joint states or an assignment map. An unrestricted completely positive map acting on every possible system state is not guaranteed by the same argument used for a fixed product preparation.
This is not a failure of quantum mechanics. The larger system evolves normally; the reduced input description omitted information. If a reconstructed map seems nonphysical, check whether the data came from preparations with a common environment or from different correlated preparation histories.
For this guide’s exact channel examples, the initial environment is fixed and uncorrelated so that complete positivity can be verified explicitly. We will introduce multitime interventions later rather than silently changing this assumption mid-calculation.
4. Three increasingly different uses of the word Markovian
A time-homogeneous semigroup obeys Φt+s=ΦtΦs and Φt=exp(tL), with a fixed generator L. This is the setting developed in Guide 52. It is a strong, convenient model of stationary memoryless evolution.
A CP-divisible family instead requires that for t≥s there exist completely positive trace-preserving maps Vt,s with Φt=Vt,sΦs. The propagators may depend on both times. A driven or time-dependent memoryless model need not be a stationary semigroup.
A multitime operational condition asks whether future statistics become independent of earlier interventions once the present is appropriately fixed. This includes information that a family of unperturbed one-time maps may not reveal. Pollock and colleagues formulated such an operational Markov condition using the full process structure. [2]
These distinctions prevent false equivalences. A time-dependent generator need not imply memory. CP divisibility is a property of a map family, not a complete description of every intervention experiment. A single revival is one witness, not a universal definition covering all operational memory.
5. Nakajima–Zwanzig projection: define what is retained
Let the total Liouvillian be LX=−i[Htot,X]. Choose a reference environment state σE and define the projection PX=TrE(X)⊗σE. Then P²=P. Set Q=I−P.
The retained part is p(t)=Pϱ(t); the discarded part is q(t)=Qϱ(t). The discarded part includes environmental changes and correlations, not just an independent environment density matrix.
Projecting the full equation ϱ̇=Lϱ gives ṗ=PLPp+PLQq and q̇=QLPp+QLQq. These equations are exact identities for the chosen projection. No weak-coupling or short-memory approximation has been used.
The purpose of the construction is now visible: solve q in terms of its past forcing by p, then substitute it back. The resulting equation for p contains a memory integral because the eliminated variables have their own dynamics.
6. Derive the exact memory kernel
For a time-independent total Liouvillian, solve the second projected equation by variation of constants:
q(t)=exp(QLQt)q(0)+∫₀ᵗ exp[QLQ(t−s)]QLPp(s)ds.
Substituting into the first equation yields
ṗ(t)=PLPp(t)+I(t)+∫₀ᵗ K(t−s)p(s)ds,
where I(t)=PLQ exp(QLQt)q(0) and K(τ)=PLQ exp(QLQτ)QLP. Taking the environment trace returns an equation for ρS.
The kernel describes a round trip: retained information enters the Q sector, evolves there for a time τ, and influences the P sector again. The inhomogeneous term carries the initial discarded component. For a product preparation compatible with P, q(0)=0, so I(t)=0; the memory kernel generally remains.
If the total Hamiltonian is time dependent, the Q-sector evolution requires a time-ordered propagator with two time arguments rather than the simple exponential used here. The exact projection method and its relation to alternative master equations are treated in the open-system literature. [1,3]
7. What a memory integral does and does not prove
A convolution ∫₀ᵗK(t−s)ρ(s)ds explicitly depends on earlier reduced states. It provides a natural way to model retained environmental influence. But the appearance of a kernel is a representation of dynamics, not a stand-alone experimental certificate.
The same map can sometimes be represented both by a memory equation and by a time-local equation with time-dependent coefficients. Conversely, one can write mathematically arbitrary kernels that do not generate positive density matrices.
An exact kernel derived from a valid joint model with a fixed product preparation inherits physical reduced dynamics. An approximate or phenomenological kernel must be checked separately for trace, Hermiticity, positivity and complete positivity where claimed.
The safe question is therefore not whether a formula contains an integral. It is whether the formula reproduces the appropriate map or multitime statistics and which physical assumptions justify it.
8. A time-local generator can still contain memory
If Φt is differentiable and invertible as a linear map on operators over a time interval, define Kt=Φ̇tΦt⁻¹. Then ρ̇(t)=Ktρ(t). This is time local: only the current ρ appears explicitly.
The coefficients of Kt can nevertheless encode the entire preparation-to-time-t evolution and can include negative canonical decay rates. Time locality alone therefore does not establish CP divisibility or operational memorylessness.
If Φt becomes noninvertible, Kt may diverge even while the physical state remains smooth. The inverse-map representation has become singular. It is incorrect to interpret every pole of a time-local rate as a divergence of the underlying experiment.
Smirne and Vacchini compared exact Nakajima–Zwanzig and time-convolutionless descriptions for two-level open systems, showing that their operator structures and coefficient functions can differ. One should derive the relevant form rather than replace a constant rate by an arbitrary kernel inside a familiar equation. [3]
9. Exact laboratory A: one system qubit and one environment qubit
Take Htot=g ZSZE, no additional free Hamiltonians, and prepare the environment in |+⟩E. The initial system state ρ is arbitrary and uncorrelated with it.
Because (ZSZE)²=I, the joint unitary is U(t)=cos(gt)I−i sin(gt)ZSZE. Tracing out E leaves system populations unchanged and multiplies its off-diagonal element by
G(t)=cos(2gt).
Thus Φt([[a,c],[c*,1−a]])=[[a,Gc],[Gc*,1−a]]. Coherence vanishes at t=π/(4g), returns with reversed sign at t=π/(2g), and returns with its original sign at t=π/g.
No probability has been lost from the joint system. The environment and correlations temporarily make the system’s local coherence inaccessible. Later the joint unitary restores it. This is an exact finite model of reduced-state revival, not an approximation to exponential decay.
10. Verify complete positivity at every time
The dephasing map has Kraus operators K₀=√[(1+G)/2] I and K₁=√[(1−G)/2] Z. Since −1≤G≤1, both coefficients are real and nonnegative, and K₀†K₀+K₁†K₁=I.
Therefore the map from the original preparation time is CPTP for every t, including times when G is negative or its magnitude is increasing. Negative coherence is not a negative probability; it is a phase reversal of an off-diagonal matrix element.
At G=−1, the map is simply ρ→ZρZ, a unitary phase flip. At G=0 it is complete dephasing. These two limiting maps are physical, but no CPTP map acting only on the completely dephased output can reconstruct every possible initial coherence. The later revival must use information outside that reduced output.
This example already separates two claims: every Φt is a channel, yet the family is not CP divisible over every interval.
11. The time-local dephasing rate and its poles
Where G(t) is nonzero, write ρ̇=γ(t)[ZρZ−ρ]/2. The off-diagonal equation is ċ=−γ(t)c, so
γ(t)=−Ġ/G=2g tan(2gt).
The rate is positive while |G| decreases and negative while |G| increases. At zeros of G, it diverges. The exact map remains finite there because the original expression c(t)=G(t)c(0) is perfectly regular.
At a zero of G, many distinct initial system states produce the same reduced state. The map is noninvertible. A time-local equation constructed with Φt−1 cannot pass through that point as an ordinary finite-rate equation valid for arbitrary system inputs.
The pole is therefore telling us about the representation’s lost inverse. Removing it by clipping the rate changes the physical map and can erase the revival we were trying to describe.
12. The same exact model has a regular memory kernel
The coherence function obeys G̈=−4g²G, with G(0)=1 and Ġ(0)=0. Integrating once gives Ġ(t)=−4g²∫₀ᵗG(s)ds.
Accordingly, the full reduced state satisfies
ρ̇(t)=2g²∫₀ᵗ[Zρ(s)Z−ρ(s)]ds.
Diagonal entries are unchanged because the bracket vanishes on them. Off-diagonal entries acquire −4g² times the integral, reproducing G(t)=cos(2gt). The kernel is constant and nonsingular even when the time-local rate has a pole.
This is a complete example of two exact representations with very different appearance. The integral equation remembers the past explicitly. The time-local equation encodes it through a coefficient that becomes singular where the map loses invertibility. Neither can be judged merely by which one looks more like an ordinary differential equation.
13. Recover the kernel with a Laplace transform
Suppose a scalar component obeys Ġ(t)=∫₀ᵗk(t−s)G(s)ds and G(0)=1. Taking Laplace transforms gives zG̃(z)−1=k̃(z)G̃(z), hence k̃(z)=z−1/G̃(z).
For G(t)=cos(2gt), G̃(z)=z/(z²+4g²). Therefore k̃(z)=−4g²/z, which inverts to the constant scalar kernel k(t)=−4g².
This derivation is useful for exactly solvable maps. It is much less stable when G is known only through noisy data: division by a small transform amplifies error. An algebraic inverse formula does not automatically provide a reliable experimental kernel estimator.
For multicomponent channels, kernels are superoperators and different operator sectors can require different coefficient functions. A scalar fit to one coherence does not reconstruct the full process.
14. Trace distance makes one kind of information return measurable
The trace distance of two density matrices is D(ρ,σ)=½||ρ−σ||₁. It quantifies their distinguishability in an optimal single-shot discrimination experiment with equal prior probabilities. A CPTP map cannot increase it.
Prepare |+⟩ and |−⟩ as the two system inputs for our dephasing model. Their Bloch vectors are opposite along X. After time t, their difference is multiplied by G(t), so
D(t)=|cos(2gt)|.
At t=π/(4g), the states are identical maximally mixed states in this pair experiment, so D=0. At t=π/(2g), they are orthogonal again, so D=1. No CPTP propagator acting only on the identical intermediate system states could produce different outputs. The map family therefore fails divisibility over that interval.
Breuer, Laine and Piilo used positive rates of change of trace distance to define a measure of non-Markovian information backflow. The exact revival here is a simple realisation of that diagnostic. [4]
15. Backflow is not proof that the environment’s memory is quantum
Consider a classical random field that is chosen once at the beginning: with probability one half the system evolves under +gZ, and with probability one half under −gZ. The choice is held fixed during a run.
Averaging the two unitary phase factors multiplies coherence by [exp(−2igt)+exp(2igt)]/2=cos(2gt). This is the same single-time dephasing map as the quantum-environment model.
The random field has perfect temporal correlation within each run. It can therefore produce the same trace-distance revival without requiring an initially coherent quantum environment. The observed map diagnoses a particular failure of divisibility; it does not uniquely identify the microscopic nature of the stored information.
To distinguish environmental mechanisms, use additional interventions, environment measurements or a constrained physical model. An inverse problem cannot be solved uniquely from one observable when several different mechanisms give exactly the same function.
16. Exact laboratory B: exchange an excitation with the environment
Take H=g(σ+Sσ−E+σ−Sσ+E), with the environment initially in |0⟩. The state |00⟩ is unchanged, while |10⟩→cos(gt)|10⟩−i sin(gt)|01⟩.
The system’s reduced map is an amplitude-damping family with G(t)=cos(gt). In the basis |0⟩,|1⟩, Kraus operators are K₀=diag(1,G) and K₁=√(1−G²)|0⟩⟨1|.
Thus ρ11(t)=G²ρ11(0) and ρ10(t)=Gρ10(0). At t=π/(2g), the excitation has moved completely into the environment. At t=π/g, it has returned to the system, with the phase specified by the joint evolution.
The generator, where G≠0, has the form ρ̇=γ(t)D[σ−]ρ with γ(t)=−2Ġ/G=2g tan(gt). Again, negative instantaneous rates occur while the complete map from the initial time remains CPTP. The environment is returning an excitation rather than acting as a one-way sink.
17. Identical present system states can have different futures
Run the exchange model twice. In the first run, start the system in |0⟩; in the second, start it in |1⟩. In both runs the environment starts in |0⟩. At time s=π/(2g), the system is |0⟩ in both experiments.
The environments differ. In the first run the environment remains |0⟩. In the second it carries the excitation. Evolve for another s. The first system stays in |0⟩; the second returns to |1⟩.
Even explicitly resetting the system to |0⟩ at the intermediate time would not erase the stored environmental excitation. The same present system state and the same future Hamiltonian therefore lead to different futures because the preparation history changed hidden variables.
This is a particularly direct memory example: no ambiguous fitted decay law is needed. It also shows why a reduced state alone need not be a sufficient statistic for an intervention-based prediction.
18. A damped reservoir replaces perfect revivals by a competition of scales
A common exactly solvable resonant amplitude model uses a reservoir correlation f(t)=(γ₀λ/2)exp(−λt). Its survival amplitude satisfies Ġ(t)=−∫₀ᵗf(t−s)G(s)ds.
Differentiating gives G̈+λĠ+(γ₀λ/2)G=0, with G(0)=1 and Ġ(0)=0. Hence, for d nonzero,
G(t)=exp(−λt/2)[cosh(dt/2)+(λ/d)sinh(dt/2)], where d=√(λ²−2γ₀λ). At d=0, take the continuous limit G(t)=exp(−λt/2)(1+λt/2); the apparent division by d is removable.
For 2γ₀>λ, d becomes imaginary and damped oscillations can appear. In the broad-reservoir limit λ much larger than γ₀, the slowly varying amplitude approaches exp(−γ₀t/2) after a short transient. The time-local population decay rate is −2Re(Ġ/G), wherever the inverse exists.
The factors here follow the explicitly stated correlation convention. Different definitions of γ₀ or λ change a displayed threshold by factors of two. Exact damped and undamped two-level reservoir models are analysed by Smirne and Vacchini. [3]
19. CP divisibility is an intermediate-map question
When Φs is invertible, the only candidate linear intermediate propagator is Vt,s=ΦtΦs⁻¹. CP divisibility asks whether that map is completely positive and trace preserving for all t≥s.
For dephasing, V multiplies coherence by G(t)/G(s). It is CPTP when the magnitude of this ratio does not exceed one. A revival of |G| therefore directly violates CP divisibility.
At a noninvertible time, one must ask whether any valid V exists on the whole operator space, not plug a singular inverse into the formula. In the complete-dephasing-and-revival example, identical intermediate states later separate, proving that no state-only propagator can reproduce the family across that interval.
Rivas, Huelga and Plenio developed a divisibility-based approach related to entanglement with an ancillary system. The role of complete positivity is precisely to preserve positivity when the system is part of a larger entangled state. [5]
20. A Choi test makes complete positivity concrete
Let |Ω⟩=Σj|jj⟩/√d be a normalised maximally entangled state. The Choi matrix of a candidate V is JV=(I⊗V)(|Ω⟩⟨Ω|). Complete positivity is equivalent to JV being positive semidefinite.
With this normalised convention, trace preservation requires tracing the output subsystem of JV to give I/d. Some references use an unnormalised maximally entangled vector and obtain I instead. Mixing the conventions creates false normalisation errors.
For a qubit map that merely seems positive on a few tested input states, a negative Choi eigenvalue can reveal that it would fail on an entangled input. That is why positivity on the Bloch ball and complete positivity are distinct conditions.
In experimental reconstruction, small negative eigenvalues may also reflect finite-shot estimation or an inappropriate fitted inverse. Report uncertainty and conditioning before assigning physical significance to a barely negative value.
21. Canonical decay rates are not arbitrary coefficients
A regular time-local generator can be expressed using an orthonormal traceless operator basis and a Hermitian decoherence matrix. Diagonalising that matrix gives canonical decay rates.
For finite-dimensional differentiable invertible map families, nonnegative canonical rates characterise the usual time-local CP-divisibility condition. A negative coefficient in an arbitrary, redundant or nonorthogonal dissipator expansion does not by itself establish the same conclusion.
Hall, Cresser, Li and Andersson developed the canonical-rate formulation and examples that distinguish different non-Markovianity measures. Their work is an important guard against treating all visible negative coefficients as equivalent or all memory witnesses as equally sensitive. [6]
When comparing rates between papers, match operator normalisation, the factor multiplying each dissipator and the time unit. A rate reported for D[Z] differs by a factor of two from a rate defined directly as the decay of an off-diagonal element.
22. Exact laboratory C: CP indivisibility without trace-distance revival
Use dimensionless time and the Pauli generator ρ̇=½Σi γi(t)[σiρσi−ρ], with γx=γy=1 and γz=−tanh t.
For t>0 one canonical rate is negative. Nevertheless, the map from time zero is a valid Pauli channel. Its Bloch-vector contraction factors are
λx(t)=λy(t)=(1+e⁻²ᵗ)/2, λz(t)=e⁻²ᵗ.
To check them, note that ẋ=−(γy+γz)x. Integrating 1−tanh t gives e⁻ᵗcosh t=(1+e⁻²ᵗ)/2. The z component decays under γx+γy=2.
This example, associated with the canonical-rate analysis of Hall and colleagues, lets us verify the distinction directly rather than merely state that definitions differ. [6]
23. Verify that the Pauli channel remains physical
Write Φt(ρ)=p₀ρ+pxXρX+pyYρY+pzZρZ. For the contraction factors above, the probabilities are p₀=(1+e⁻²ᵗ)/2, px=py=(1−e⁻²ᵗ)/4 and pz=0.
They are all nonnegative and sum to one. Thus √piσi, with σ₀=I, form a valid Kraus representation. There is no conflict between a negative instantaneous canonical rate and a physical map from the original time.
The negative rate says something different: dividing the process into arbitrary intermediate CPTP steps is impossible. It does not say that the integrated channel contains a negative probability.
This distinction is easy to lose when a master equation is inspected locally but the integrated dynamics is never calculated. Always test both objects before declaring the model unphysical.
24. Why every pairwise trace distance decreases in that example
For qubits, the trace distance between states with Bloch-vector difference Δr is D=|Δr|/2. Under the Pauli channel,
D(t)=½√[λx²Δx²+λy²Δy²+λz²Δz²].
All three λ factors are positive and nonincreasing. Therefore D(t) is nonincreasing for every fixed pair of initial system states. There is no BLP-style trace-distance revival, despite the negative canonical rate and CP indivisibility.
The example is not a paradox. A witness based only on pairwise distinguishability of the system can miss a failure that appears when ancillary entanglement and complete positivity are tested. Different operational resources expose different aspects of the dynamics.
A failed search for backflow is therefore not a proof of CP divisibility. Conversely, a clear backflow event is a sufficient witness of a failure of CPTP intermediate propagation for that map family. The logical directions should be stated explicitly.
25. The backflow measure needs a time window and an optimisation
The BLP construction maximises, over pairs of initial states, the total positive increase of trace distance over the observation interval. Symbolically, N= maxρ₁,ρ₂ ∫Ḋ>0 Ḋ(t)dt.
If only one pair is tested, its integrated revival is a lower bound on that optimisation, not automatically the full measure. If the experiment stops at a finite time, the result belongs to that finite window.
For the undamped cosine dephasing model, repeated perfect revivals can make the accumulated measure grow without bound as the observation interval is extended. That does not imply infinite distinguishability at one time: D always remains between zero and one. The integral counts repeated returns.
State the pair, interval and numerical treatment of derivatives. Otherwise two reported backflow numbers may differ simply because one experiment ran longer or used a better probe pair.
26. Multitime memory asks what interventions reveal
A family Φt tells us the reduced state after unperturbed evolution from a fixed starting time. It does not automatically tell us what happens when we measure, reset or control the system halfway through.
A process tensor, or equivalent multitime description, takes a sequence of operations as inputs and returns later outcome statistics. It retains correlations between different times that may not be reconstructed from one-time maps alone.
An operational memory test introduces a causal break: measure or discard the system and prepare a specified new state. If future statistics still depend on earlier interventions or records after the appropriate conditioning, the inaccessible degrees of freedom retained relevant information. This is the perspective formalised by Pollock and colleagues. [2,7]
The exchange example gives a concrete version. Resetting the system to ground does not erase an excitation already stored in the environment. The future reveals which past preparation occurred.
27. Not every reduced-state revival is an intrinsic memory witness
A driven system can have an oscillating population under a perfectly memoryless unitary Hamiltonian. Population revival alone is therefore not the same as trace-distance backflow between two preparations under a common process.
Likewise, postselection can increase distinguishability in a selected subensemble while reducing its probability. If only accepted runs are plotted, a conditional revival must not be passed off as a property of the unconditional channel. Guide 74 explains the related distinction for non-Hermitian no-jump evolution.
A common unitary control applied to both states preserves their trace distance. A common CPTP operation cannot increase it. But different controls, different environment preparations or state-dependent data selection change the comparison itself.
Good experimental design specifies which interventions are part of the process and which are part of the test. Without that separation, a curve can be mathematically correct and still answer the wrong question.
28. How a Markov approximation removes the kernel
A microscopic weak-coupling derivation often produces an integral weighted by a reservoir correlation function. If that correlation decays on a time τE much shorter than the relevant reduced-state evolution time, the system changes little over the kernel’s support.
One may then approximate the slowly varying state inside the integral by its current value, extend an upper integration limit when justified, and derive a time-local generator. Additional secular or coarse-graining steps may be needed to obtain a controlled GKSL form.
Each step is an approximation with conditions. Weak coupling alone does not imply that every environment has a short correlation time. A narrow reservoir resonance can retain information for a long time even when the coupling is modest.
The derivation should compare bath correlation time, system transition times, relaxation rates and drive times rather than use the word Markov as a general synonym for simple. The range of available non-Markovian methods is reviewed by de Vega and Alonso. [1]
29. Fresh ancillas and reused ancillas
In a collision model, the system interacts with one ancillary unit at each step. If every ancilla is fresh, identically prepared and initially uncorrelated with the others, discarding each after its interaction gives a repeated composition of the same channel.
If the same ancilla returns, or neighbouring ancillas interact, or the initial ancillas are correlated, information can persist between collisions. The system’s next step can then depend on more than its own current state.
The construction is useful pedagogically because it makes the memory carrier explicit. A small reused register can store a past excitation or phase. A large chain of fresh registers can approximate a one-way reservoir.
But the full multitime process depends on the exact collision protocol. Two arrangements can agree on a limited set of unperturbed state trajectories and differ when intermediate resets or controls are inserted. Specify the circuit, not only a fitted decay curve.
30. Enlarging the system can simplify the memory problem
Suppose a qubit couples strongly to one narrow resonator, while the resonator leaks into a broad external reservoir. Treating only the qubit as the system can produce a long memory kernel. Treating qubit plus resonator as the system may permit a Markovian equation for the enlarged density matrix.
Pseudomode, reaction-coordinate and chain-mapping strategies use related ideas: move the important structured environmental degrees of freedom across the system boundary, then approximate the remaining environment more simply.
This trades memory complexity for state-space complexity. An enlarged system requires more amplitudes or density-matrix entries, but can avoid storing the entire history explicitly. The choice should be guided by the structure of the reservoir correlation, not by a rule that one representation is always best.
A successful embedding is also a physical explanation. It identifies which modes store the information and on what timescale they release it.
31. Numerical memory integration has two error budgets
A discretised convolution evaluates a sum over past states. With N time steps, a direct method can require order N² history operations, although structured kernels and fast convolution can reduce the cost.
There is a time-step error and a memory-truncation error. Reducing the step size while keeping an unjustifiably short memory cutoff does not solve the second problem. Increasing the cutoff without resolving rapid oscillations does not solve the first.
Use an exact toy model as a reference. For cosine dephasing, compare the convolution result with G(t)=cos(2gt), including the first zero and revival. For excitation exchange, verify both the population G² and coherence G, not only one curve.
Check trace, Hermiticity and eigenvalues of the density matrix. If a full map is claimed, test its Choi matrix. Passing positivity for one chosen initial state does not establish complete positivity for every input.
32. A negative numerical eigenvalue is a diagnostic, not a conclusion
An approximate master equation may violate positivity because its approximation was used outside its regime, a coarse-graining step was omitted, or numerical integration is inaccurate. A tiny negative eigenvalue can also arise from floating-point or statistical reconstruction error.
Do not repair every negative eigenvalue by clipping and then declare the original equation physical. Clipping changes the dynamics. First refine the solver, check the derivation and compare with a larger exact model or a known physical embedding.
Conversely, negative canonical rates are not the same as negative density-matrix eigenvalues. Our exact models have negative rates over some intervals while remaining CPTP from the initial time. One is a statement about intermediate divisibility; the other concerns the validity of the state or map itself.
The distinction should be maintained in code variable names and reports. Calling both quantities negative probability obscures the mathematics.
33. Experimental backflow needs uncertainty analysis
Reconstructing trace distance at neighbouring times and differentiating it amplifies shot noise. Random fluctuations can create positive slopes even when the true distance is monotonic.
Before integrating positive derivative segments, estimate confidence intervals from the underlying measurement counts, account for shared calibration uncertainty and test a monotonic null model. Smoothing should be specified because it can either remove a genuine narrow revival or manufacture one through an inappropriate fit.
A convincing witness can instead compare two times with a clearly resolved positive difference in D, avoiding a noisy pointwise derivative where possible. Optimising the initial state pair on the same noisy data used to claim significance requires care; held-out validation reduces selection bias.
In a fictional study group, Clara notices a small bump and labels it backflow. Ben asks whether the bump survives repeated data sets and a fixed analysis rule. The constructive response is to quantify that uncertainty, not to choose between a dramatic claim and abandoning the experiment.
34. Memory can help a task, but it is not automatically a resource advantage
A returning excitation or revived coherence can be useful when a protocol is designed to exploit it. It can also disrupt a control sequence that assumed independent errors or make calibration history dependent.
A resource claim must name the task: state transfer, sensing, error recovery, cooling or communication. It must compare the final performance under matched controls, environment access, observation time and success probability.
One non-Markovianity measure need not predict every task’s advantage. The Pauli example already shows that different memory tests order processes differently. A scalar backflow score is not a universal performance index.
For system identification and feedback, connect this guide to Guide 70 and Guide 53. A controller should model the memory carrier or validate its approximation rather than assume that yesterday’s calibrated single-step channel composes unchanged today.
35. A practical decision sequence
First identify the observed system and the environmental degrees of freedom omitted. Next specify the initial preparation and whether system-environment correlations are present. Then choose the question: predict one-time states, test CP divisibility, witness distinguishability return or describe multitime interventions.
Choose the mathematical representation only after those decisions. A memory kernel may be convenient for a known correlation function. A time-local generator may be convenient on an invertible interval. An enlarged-system simulation may be best when a few modes dominate. A process tensor may be necessary when interventions expose temporal correlations.
Finally test the physical map and the experiment separately. Trace preservation and a positive Choi matrix are mathematical checks. Reproducible revivals and causal-break dependence are experimental observations. A good conclusion states both what was calculated and what was observed, without treating either as a substitute for the other.
36. Practice laboratory with worked answers
Problem 1: initial conditions. Why does Φt(ρ)=TrE[U(t)(ρ⊗σE)U†(t)] define a CPTP map even when the environment remembers the past?
Answer. It is a composition of appending a fixed ancillary state, unitary evolution and partial trace. These operations are CPTP. Memory concerns the relation between different times or interventions, not the validity of each preparation-to-time-t map.
Problem 2: projection. For PX=TrE(X)⊗σE with TrσE=1, show that P²=P.
Answer. Applying P twice gives TrE[TrE(X)⊗σE]⊗σE=TrE(X)Tr(σE)⊗σE=PX. The normalisation of the reference environment state makes it a projection.
Problem 3: the inhomogeneous term. If the initial state is ρS⊗σE, does the Nakajima–Zwanzig memory integral vanish?
Answer. No. q(0)=Qϱ(0)=0 removes the initial-correlation source term I(t). The kernel remains because the interaction can generate discarded correlations after the initial time and return their influence later.
Problem 4: finite-bath dephasing. For G(t)=cos(2gt), find G at t=0, π/(4g), π/(2g) and π/g.
Answer. The values are 1, 0, −1 and 1. Coherence is first preserved, then erased locally, then restored with a phase flip, then restored with its original sign. Every map remains CPTP.
Problem 5: Kraus probabilities. At G=−1/2, what are the weights of I and Z in the dephasing channel?
Answer. They are (1+G)/2=1/4 and (1−G)/2=3/4. Both are positive. A negative coherence multiplier is not a negative channel probability.
Problem 6: a singular rate. Why does γ(t)=2g tan(2gt) diverge at t=π/(4g), while the state stays finite?
Answer. The generator uses the ratio −Ġ/G. At G=0 the reduced map loses invertibility, so this representation becomes singular. The direct map c(t)=G(t)c(0) and the exact memory-kernel equation remain regular.
Problem 7: distinguishability. For initial |+⟩ and |−⟩ in the finite-bath dephasing model, what is D(t), and why does its revival exclude a state-only CPTP propagator across the dephasing time?
Answer. D(t)=|G(t)|. At complete dephasing the two reduced states are identical, but they later differ. A fixed map acting on identical inputs must return identical outputs. The later distinction must use information not contained in the intermediate system state.
Problem 8: excitation exchange. In the exchange model, begin with the system excited and the environment in its ground state. Where is the excitation at t=π/(2g)?
Answer. It is entirely in the environment. The system is in its ground state. Another equal evolution interval returns the excitation to the system. This provides an exact memory carrier that can survive resetting the system at the intermediate time.
Problem 9: the Pauli counterexample. Let e⁻²ᵗ=1/2. Find p₀, px, py, pz for the channel in Sections 22–24.
Answer. They are 3/4, 1/8, 1/8 and 0. The channel is a valid mixture of Pauli unitaries. Its negative canonical rate concerns intermediate CP divisibility rather than these integrated probabilities.
Problem 10: absent backflow. Why does finding no positive derivative of trace distance for any system-state pair still not prove CP divisibility in the Pauli example?
Answer. Every Bloch contraction factor decreases, so every such distance decreases. Complete positivity of intermediate maps is a stronger condition involving possible ancillary entanglement. The negative canonical rate exposes a failure that the system-only pair witness misses.
Problem 11: a classical alternative. What classical noise model produces the same G(t)=cos(2gt) as the one-qubit environment?
Answer. Choose a static Hamiltonian +gZ or −gZ with equal probability at the start of each run. Averaging their opposite phase factors gives the cosine. The field is temporally correlated, so the revival does not uniquely certify a quantum memory carrier.
Problem 12: a bath limit. For the exponential correlation model with λ much larger than γ₀, why does approximately exponential decay emerge?
Answer. The reservoir correlation decays much faster than the system’s slow amplitude. After a short transient, the slow root of r²+λr+γ₀λ/2=0 approaches −γ₀/2. Thus G(t) approaches e⁻ᵞ⁰ᵗ/² in that scale-separated regime.
Problem 13: numerical proof. A solver produces positive states for the initial input |0⟩. Has it proved the full map is CPTP?
Answer. No. That tests one input only. Complete positivity requires the map to remain positive under arbitrary ancillary extension; in finite dimensions a Choi-matrix test provides the corresponding condition. Trace preservation must also be checked.
Problem 14: an experimental bump. A reconstructed trace-distance curve increases slightly between two neighbouring noisy time points. What must be checked before interpreting it as backflow?
Answer. The increase must exceed statistical and calibration uncertainty under a specified analysis rule. Check common preparation and controls, postselection, reconstruction bias and the effect of smoothing or pair optimisation. A positive finite-difference value alone is not a controlled witness.
37. A study sequence from exact dynamics to experimental inference
First derive the cosine dephasing channel directly from a two-qubit unitary. Verify its Kraus representation and the |+⟩,|−⟩ trace distance. Then derive both its time-local rate and its constant memory kernel. This one example proves that representation, map validity and divisibility are different questions.
Next derive the excitation-exchange channel and carry two initial preparations through the complete swap interval. Insert a reset at the intermediate time. Identify exactly where the information resides when the system states are identical.
Then solve the Pauli counterexample. Compute its integrated probabilities and every Bloch contraction factor rather than stopping at the negative rate. The result teaches the logical difference between a sufficient witness and a complete characterisation.
Finally add simulated finite-shot noise to a known monotonic or reviving signal. Test whether the chosen analysis would falsely detect a revival. Separate the quality of the dynamics model from the quality of the statistical procedure. A good theory-to-experiment bridge requires both.
Frequently asked questions
Does a time-local equation imply Markovian dynamics? No. An exact time-local generator can contain negative canonical rates or singular coefficients inherited from the full map. The equation’s syntax does not determine its divisibility or multitime memory.
Does a negative rate make a channel unphysical? Not necessarily. The exact dephasing, exchange and Pauli examples all give physical maps from their initial time. Check the integrated map, canonical representation and the particular intermediate-map claim.
Does information backflow mean energy flows back? Not always. Pure dephasing can revive distinguishability without changing populations. Energy exchange and information distinguishability are different observables.
Can classical noise be non-Markovian? Yes. Temporally correlated classical fields can produce reduced quantum dynamics with revivals. A memory witness does not automatically certify that the environment’s stored information is intrinsically quantum.
What should be reported? Specify the system boundary, initial environmental preparation, allowed interventions, map or kernel convention, memory criterion, observation window, numerical convergence and statistical uncertainty. Then the result can be reproduced and compared without confusing distinct definitions.
Sources and further study
[1] Inés de Vega and Daniel Alonso, Dynamics of non-Markovian open quantum systems, Reviews of Modern Physics 89, 015001 (2017). Reduced and enlarged-system methods and their physical assumptions.
[2] Felix A. Pollock, César Rodríguez-Rosario, Thomas Frauenheim, Mauro Paternostro and Kavan Modi, Operational Markov condition for quantum processes, Physical Review Letters 120, 040405 (2018). Multitime operational memory and interventions.
[3] Andrea Smirne and Bassano Vacchini, Nakajima–Zwanzig versus time-convolutionless master equation for the non-Markovian dynamics of a two-level system, Physical Review A 82, 022110 (2010). Exact map, time-local and kernel representations.
[4] Heinz-Peter Breuer, Elsi-Mari Laine and Jyrki Piilo, Measure for the Degree of Non-Markovian Behavior of Quantum Processes in Open Systems, Physical Review Letters 103, 210401 (2009). Trace-distance information-backflow measure.
[5] Ángel Rivas, Susana F. Huelga and Martin B. Plenio, Entanglement and non-Markovianity of quantum evolutions, Physical Review Letters 105, 050403 (2010). Divisibility and ancillary-entanglement perspective.
[6] Michael J. W. Hall, James D. Cresser, Li Li and Erika Andersson, Canonical form of master equations and characterization of non-Markovianity, Physical Review A 89, 042120 (2014). Canonical rates and examples separating memory witnesses.
[7] Felix A. Pollock and colleagues, Non-Markovian quantum processes: Complete framework and efficient characterization, Physical Review A 97, 012127 (2018). Process-tensor description beyond one-time maps.
Continue through Quantum Mathematics — Batch 19
Guide 73: Quantum Floquet Systems, Periodic Driving, Quasienergies, Magnus Expansion and Floquet Engineering describes periodic driving and the distinction between stroboscopic and within-cycle dynamics. Guide 74: Non-Hermitian Quantum Mechanics, PT Symmetry, Exceptional Points and Biorthogonal Spectra separates conditional effective evolution from full channels. Guide 75: Many-Body Localization, Anderson Localization, l-Bits, Entanglement Growth and Thermalisation Breakdown examines local memory in interacting isolated systems.