Dissipative State Preparation

Dissipative state preparation uses irreversible dynamics to drive a quantum system toward a desired state or subspace. Instead of trying to keep the environment out, one engineers selected loss, pumping, reset, or auxiliary damping so that unwanted components are removed and the target remains invariant.

The central design statement is:

make the target stationary, make all unwanted directions flow toward it,
and verify that competing noise does not dominate the preparation

This page is the focused preparation workflow. The broader design language for engineered baths and auxiliary modes is Reservoir Engineering. The fixed-point and gap terminology is owned by Steady States and Relaxation.

Preparation Target

Let the designed Markovian dynamics be

$$\dot\rho = \mathcal L(\rho) = - \frac{i}{\hbar}[H,\rho] + \sum_\alpha \mathcal D[L_\alpha]\rho,$$

where

$$\mathcal D[L]\rho = L\rho L^\dagger - \frac12 \{L^\dagger L,\rho\}.$$

For a target state $\rho_\star$, stationarity means

$$\mathcal L(\rho_\star)=0.$$

For preparation from arbitrary initial states in the modeled state space, stationarity is not enough. One wants $\rho_\star$ to be the unique attractive steady state:

$$\rho(t) \longrightarrow \rho_\star \quad \text{as }t\to\infty$$

for all allowed initial states, or at least for a clearly specified basin of attraction.

For subspace preparation, the target is a projector $P_\star$ or code space rather than one density matrix. Then the goal is usually invariance plus attraction into the subspace, not collapse to one vector inside it.

Dark Is Not Enough

For a pure target $\lvert\psi_\star\rangle$, a useful sufficient dark-state condition is

$$L_\alpha\lvert\psi_\star\rangle=0 \quad \text{for all }\alpha, \qquad H\lvert\psi_\star\rangle = E_\star\lvert\psi_\star\rangle.$$

Then

$$\rho_\star = \lvert\psi_\star\rangle\langle\psi_\star\rvert$$

is stationary. But a dark state may be only one member of a large dark subspace. Preparation requires an attracting flow:

target invariant
unwanted states connected to lossy channels
no other stable traps

The practical question is therefore not only “is the target dark?” but “what happens to every orthogonal component?”

Uniqueness and the Liouvillian Gap

In finite dimensions, the clean algebraic condition for unique preparation is

$$\ker\mathcal L = \operatorname{span}\{\rho_\star\}$$

within the physical sector being modeled. This statement should be checked at the superoperator level, not guessed from one jump operator.

If the steady state is unique, the relaxation gap

$$\Delta_{\mathcal L} = \min_{\lambda_a\ne0} \left[ - \operatorname{Re}\lambda_a \right]$$

sets the leading preparation timescale,

$$\tau_{\mathrm{prep}} \sim \frac{1}{\Delta_{\mathcal L}},$$

up to prefactors, nonnormality, and initial-state overlap with slow modes. A very small gap means preparation may be formally correct but experimentally useless.

If $\ker\mathcal L$ contains more than one physical state, the dynamics prepares a manifold, not a unique state. That may be desired for a protected subspace, but it is a failure for single-state preparation.

Engineering the Flow

A dissipative preparation scheme usually combines three elements.

First, the target is invariant. The Hamiltonian and jump operators do not push it away.

Second, unwanted components are pumped into sectors that decay. This often requires coherent drives, auxiliary levels, selection rules, or multi-step transitions.

Third, the reset channel removes entropy. A photon, phonon, quasiparticle, cavity excitation, or auxiliary-state population must be dumped into degrees of freedom that do not return the same entropy to the system.

A useful schematic is

$$\lvert e\rangle \xrightarrow{\text{engineered loss}} \lvert\psi_\star\rangle, \qquad \lvert\psi_\star\rangle \xrightarrow{\text{loss}} \text{nothing}.$$

The first arrow should exist for every unwanted sector. The second arrow should be absent or strongly suppressed for the target.

Optical-Pumping Prototype

Optical pumping gives the simplest intuition. A drive excites unwanted ground states to an excited manifold; spontaneous emission returns population to lower states; a chosen dark state stops absorbing. After many cycles, population accumulates in the dark state.

In an idealized three-level picture, the target $\lvert d\rangle$ is dark:

$$H_{\mathrm{drive}}\lvert d\rangle=0, \qquad L_\alpha\lvert d\rangle=0,$$

while an unwanted state $\lvert u\rangle$ is connected to an excited state that decays with some branching probability into $\lvert d\rangle$.

The preparation rate is not just the spontaneous-emission rate. It depends on drive strength, detuning, saturation, branching ratios, off-resonant scattering, and leakage out of the target. Stronger driving can increase the pump rate but also spoil dark-state selectivity.

Dissipative Entanglement

Dissipative preparation becomes especially useful when the target is entangled. The idea is to make the entangled state dark while all other states are actively drained toward it.

For two qubits, the singlet

$$\lvert\psi_-\rangle = \frac{ \lvert ge\rangle-\lvert eg\rangle } {\sqrt2}$$

is dark under collective lowering

$$J_-=\sigma_-^{(1)}+\sigma_-^{(2)}.$$

But collective lowering alone does not prepare the singlet uniquely, because $\lvert gg\rangle$ is also dark. A dissipative Bell-state preparation protocol must add drives, auxiliary levels, measurement-conditioned steps, or extra dissipators so that $\lvert gg\rangle$ and the triplet sector are not stable traps.

This is the canonical lesson: a pretty dark vector is not a preparation protocol.

Subspace Preparation

Sometimes the target is not one state but a subspace. For a code space $\mathcal C$ with projector $P_{\mathcal C}$, the requirement is

$$\rho_{\mathcal C} = P_{\mathcal C}\rho P_{\mathcal C} \quad \Longrightarrow \quad \mathcal L(\rho_{\mathcal C}) \text{ remains in }\mathcal C.$$

For autonomous error correction, one wants error sectors to be drained back into $\mathcal C$ without revealing logical information. A schematic recovery jump is

$$L_\mu = \sqrt{\Gamma_\mu}\, R_\mu P_\mu,$$

where $P_\mu$ projects onto an error syndrome sector and $R_\mu$ maps that sector back to the code space.

The important distinction is that state preparation intentionally erases information to reach one state, while subspace stabilization should preserve arbitrary logical superpositions inside the target subspace.

Fidelity and Error Budget

A common preparation score for a pure target is

$$F(t) = \langle\psi_\star\rvert \rho(t) \lvert\psi_\star\rangle.$$

For a steady-state protocol, the asymptotic score is

$$F_{\mathrm{ss}} = \langle\psi_\star\rvert \rho_{\mathrm{ss}} \lvert\psi_\star\rangle.$$

If unwanted natural noise competes with the engineered preparation, the infidelity often scales like a ratio of harmful to useful rates:

$$1-F_{\mathrm{ss}} \sim \frac{\Gamma_{\mathrm{bad}}} {\Gamma_{\mathrm{prep}}}$$

in a simple rate-limited regime. This is only a scaling estimate. Saturation, leakage, coherent errors, thermal excitations, and metastable traps can change the dependence.

Good reports separate:

• target fidelity;
• preparation time;
• Liouvillian gap or slowest observed decay mode;
• leakage outside the intended Hilbert space;
• sensitivity to detuning, drive amplitude, and bath temperature;
• comparison with passive relaxation or coherent pulse preparation.

Numerical Verification

For a finite model, the most direct verification is a Liouvillian audit.

1. Build the full generator with the engineered and natural channels.
2. Compute the steady-state null space.
3. Check whether the target is stationary by evaluating $\mathcal L(\rho_\star)$.
4. Check whether the physical steady state is unique.
5. Compute the steady-state fidelity and trace leakage.
6. Inspect nonzero eigenvalues and the relaxation gap.
7. Time-evolve representative initial states.
8. Sweep uncertain parameters and unwanted noise rates.
9. Increase Hilbert-space cutoffs for oscillator or multilevel models.

The notebook workflow in Solving Lindblad Equations is the local computational starting point.

Relation to Other Control Methods

Dissipative preparation can be autonomous: after the engineered drives and losses are turned on, no real-time measurement record is required. That distinguishes it from Measurement-Based Feedback.

It can be implemented through routed fields or auxiliary quantum systems, in which case Coherent Feedback may be the more detailed network description before eliminating fast degrees of freedom. If strong measurement or strong dissipation primarily confines dynamics to a subspace, Quantum Zeno Dynamics gives the control language.

It can also be combined with Optimal Control: one can optimize drives or couplings to maximize the preparation gap, suppress leakage, or improve robustness. The objective should still include the dissipative steady state, not only a short-time state-transfer score.

Common Mistakes

• Treating $\mathcal L(\rho_\star)=0$ as proof of preparation.
• Checking one jump operator for darkness while ignoring the Hamiltonian.
• Forgetting that Lindblad operators are representation dependent.
• Reporting a target dark state without checking the full Liouvillian null space.
• Ignoring a second dark state or symmetry sector that traps population.
• Optimizing the preparation rate while increasing leakage or heating.
• Using an oscillator cutoff that artificially creates a steady state.
• Calling single-state pumping “autonomous error correction” when logical information is not preserved.
• Forgetting to include natural dephasing, relaxation, and thermal excitation in the final fidelity.

Exercises

Stationary versus prepared

Suppose $\mathcal L(\rho_\star)=0$. What additional property is needed before saying that $\rho_\star$ is prepared from arbitrary initial states?

Solution

One needs attractivity, usually uniqueness of the physical steady state within the modeled sector and decay of all other modes. Algebraically, in a finite model one wants the physical kernel of $\mathcal L$ to contain only $\rho_\star$, with all nonzero modes having negative real parts. Otherwise the target may be stationary but not reached from generic initial states.

Two dark states

A dissipator has jump operator $L=\lvert d\rangle\langle u\rvert$ and no Hamiltonian. Which states are stationary?

Solution

Population in $\lvert u\rangle$ decays into $\lvert d\rangle$, but any state orthogonal to both $\lvert u\rangle$ and the coupled coherences can be unaffected if the Hilbert space has other levels. At minimum, $\lvert d\rangle\langle d\rvert$ is stationary. If there is another state $\lvert s\rangle$ with $L\lvert s\rangle=0$ and no Hamiltonian coupling out of it, then $\lvert s\rangle\langle s\rvert$ is also stationary. Thus the jump alone prepares $\lvert d\rangle$ uniquely only in a two-level space where all non-target population is connected to $\lvert u\rangle$.

Gap and preparation time

If the smallest nonzero decay rate of a diagonalizable Liouvillian is $\Delta_{\mathcal L}$, what preparation time scale should you expect?

Solution

The leading estimate is $\tau_{\mathrm{prep}}\sim1/\Delta_{\mathcal L}$. This estimate can be modified by prefactors, nonnormality, and weak overlap of the initial state or observable with the slow mode, but a small gap still signals slow convergence.

State versus subspace

Why is pumping every state into $\lvert0_L\rangle$ not autonomous error correction for a logical qubit?

Solution

A logical qubit must preserve arbitrary superpositions $\alpha\lvert0_L\rangle+\beta\lvert1_L\rangle$. Pumping everything into $\lvert0_L\rangle$ erases the coefficient $\beta$ and destroys the encoded information. Error correction should remove error syndromes while preserving the logical state inside the code space.

Cross-Links

• Reservoir Engineering
• Steady States and Relaxation
• Lindblad–GKSL Equation
• Lindblad Operators
• Quantum Dynamical Semigroups
• Channel Composition and Fixed Points
• Coherent Feedback
• Measurement-Based Feedback
• Optimal Control
• Quantum Zeno Dynamics
• Solving Lindblad Equations
• Approximation Checklist
• Reading List

References

• J. F. Poyatos, J. I. Cirac, and P. Zoller, “Quantum reservoir engineering with laser cooled trapped ions,” Physical Review Letters 77, 4728-4731 (1996).
• B. Kraus, H. P. Buchler, S. Diehl, A. Kantian, A. Micheli, and P. Zoller, “Preparation of entangled states by quantum Markov processes,” Physical Review A 78, 042307 (2008).
• F. Verstraete, M. M. Wolf, and J. I. Cirac, “Quantum computation and quantum-state engineering driven by dissipation,” Nature Physics 5, 633-636 (2009).
• S. Diehl, A. Micheli, A. Kantian, B. Kraus, H. P. Buchler, and P. Zoller, “Quantum states and phases in driven open quantum systems with cold atoms,” Nature Physics 4, 878-883 (2008).
• A. Reiter and A. S. Sorensen, “Effective operator formalism for open quantum systems,” Physical Review A 85, 032111 (2012).
• J. T. Barreiro, M. Muller, P. Schindler, D. Nigg, T. Monz, M. Chwalla, M. Hennrich, C. F. Roos, P. Zoller, and R. Blatt, “An open-system quantum simulator with trapped ions,” Nature 470, 486-491 (2011).
• Y. Lin, J. P. Gaebler, F. Reiter, T. R. Tan, R. Bowler, A. S. Sorensen, D. Leibfried, and D. J. Wineland, “Dissipative production of a maximally entangled steady state of two quantum bits,” Nature 504, 415-418 (2013).