Dissipative Thermal State Preparation

• Dissipative thermal state preparation is a technique that uses engineered Lindblad dynamics to steer quantum systems toward prescribed Gibbs or low-entropy states.
• It employs mechanisms such as driven ancilla pseudospins, continuous-time Gibbs samplers, and weak-coupling collision models to enable energy-selective transitions.
• The approach provides rigorous convergence guarantees and practical pathways while balancing detailed-balance enforcement with nonthermal state engineering.

Dissipative thermal state preparation is the use of engineered open-system dynamics to drive a quantum system toward a prescribed thermal or low-entropy steady state, typically a Gibbs state

$$\rho_\beta=\frac{e^{-\beta H}}{Z},\qquad Z=\operatorname{Tr}(e^{-\beta H}),$$

by designing Lindblad generators, ancillary baths, or collision channels that mediate energy-selective transitions (Chen et al., 2023). In current usage, the term has both a narrow and a broad meaning. In the narrow sense, it denotes genuine Gibbs-state preparation with detailed-balance structure and a thermal fixed point (Metcalf et al., 2019). In a broader engineered sense, it also includes dissipative preparation of pure energy eigenstates, narrow energy-window states, and basis-tracking cooling channels whose behavior is thermal-state-like without enforcing equilibration to a Boltzmann ensemble (Liu et al., 2023, Tian et al., 20 Apr 2026).

1. Formal framework and target states

The common mathematical framework is the GKSL or Lindblad master equation,

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$

or, in system Hamiltonian notation,

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$

Within this framework, the Lindbladian is treated not only as a model of decoherence but as an algorithmic or engineered channel generator whose fixed point is the desired state (Lin, 27 May 2025, Liu et al., 2023).

For genuine thermal preparation, the target is the Gibbs state of a Hamiltonian $H$, and the design criterion is $\mathcal L(\sigma_\beta)=0$. For dissipative state preparation in the broader sense, the target may instead be a pure energy eigenstate

$$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$

selected by

$$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$

The latter case includes excited-state stabilization and reservoir-engineered entanglement, and the underlying bath need not be thermal or satisfy detailed balance (Liu et al., 2023).

This distinction is structurally important. A Lindbladian with KMS or quantum detailed balance is designed to reproduce thermal equilibration, whereas many dissipative preparation schemes deliberately violate detailed balance to stabilize nonthermal steady states or excited eigenstates. In the Rydberg-array setting, for example, the steady states are described as “thermal-like but not thermal”: the dynamics select eigenstates or narrow energy windows rather than Gibbs ensembles (Tian et al., 20 Apr 2026).

2. Engineered thermalization mechanisms

One major approach uses driven, dissipative ancilla pseudospins as a tunable bath. In the scheme of “Engineered thermalization and cooling of quantum many-body systems,” the many-body Hamiltonian $H_{\rm sys}$ is coupled to ancillary two-level systems through

$$H(t)=H_{\rm sys}-\sum_{m=1}^M \frac{\Omega_m(t)}{2}\tau_z^m+\sum_{m=1}^M g_m \sigma_\alpha^{k_m}\tau_x^m,$$

while each ancilla obeys

$$\frac{d\rho^m}{dt}=\gamma_+^m(t)\mathcal{D}[\tau_+^m]\rho^m(t)+\gamma_-^m(t)\mathcal{D}[\tau_-^m]\rho^m(t), \qquad \frac{\gamma_-^m(t)}{\gamma_+^m(t)}=e^{\beta\Omega_m(t)}.$$

Periodic modulation of $$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$0 sweeps the ancilla splittings across the system bandwidth, so a finite set of ancillae addresses different Bohr frequencies at different times. Under the Born-Markov hierarchy

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$1

the induced Lindbladian has Lorentzian frequency-resolved rates $$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$2, and the Gibbs state is an approximate dynamical fixed point because detailed balance is best satisfied near resonance (Metcalf et al., 2019).

A second line of work constructs continuous-time Gibbs samplers directly from filtered Heisenberg evolutions. In “Quantum Thermal State Preparation,” the jump operators take the form

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$3

and the Lindbladian

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$4

uses rates satisfying the KMS relation

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$5

Because exact energy resolution is unavailable at finite time, the analysis is organized around approximate detailed balance and finite-time secular approximation rather than the idealized infinite-time Davies limit (Chen et al., 2023).

A third mechanism is the weak-coupling collision model with ancilla resets. In this setting, a system repeatedly interacts with a fresh ancilla bath initialized in $$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$6, evolves under

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$7

with

$$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$8

and then the bath is reset. The resulting averaged channel $$\frac{d}{dt}\rho=-i[G,\rho]+\sum_j\left(K_j\rho K_j^\dagger-\frac12\{K_j^\dagger K_j,\rho\}\right)=:\mathcal L(\rho),$$9 approximates a Lindblad step at order $$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$0, and the target Lindbladian $$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$1 exactly fixes $$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$2 through KMS detailed balance (Ong et al., 4 May 2026).

3. Thermodynamic structure, dissipation, and initial-state dependence

Dissipative thermal state preparation is not only a dynamical problem but also a thermodynamic one. For diagonal initial states in the energy basis,

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$3

the dissipated heat in dissipative state preparation toward a pure eigenstate is

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$4

and the entropy change is

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$5

A central point is that if the target is an excited eigenstate, then $$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$6 can be negative: the environment can supply energy rather than extract it. This separates engineered dissipative preparation from ordinary cooling or Landauer-type reset scenarios (Liu et al., 2023).

The same work derives an initial-state-dependent quantum speed limit for Markovian dissipative preparation using the relative purity with the final pure state,

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$7

The bound is

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$8

with a looser appendix bound

$$\dot{\rho}_t=-i[H_S,\rho_t]+\sum_{\mu}\gamma_{\mu}\mathcal{D}[L_{\mu}]\rho_t, \qquad \mathcal{D}[L_\mu]\rho=L_\mu\rho L_\mu^\dagger-\frac{1}{2}\{L_\mu^\dagger L_\mu,\rho\}.$$9

Unlike a relaxation time extracted from a Liouvillian gap, this lower bound depends explicitly on the initial state and permits passive optimization over permutations of the initial populations (Liu et al., 2023).

For a fixed population multiset $H$0, minimizing the quantum-speed-limit time places the largest population on the target eigenstate,

$H$1

and minimizing heat among those arrangements orders the remaining populations in decreasing order along the increasing energy basis. If the target is the ground state, the optimal initial state is passive; if the target is an excited state, the optimal structure is passive-like except for the target slot (Liu et al., 2023). This suggests that thermal-state preparation and dissipative eigenstate preparation are both sensitive to initial-state geometry in energy space, not only to the generator itself.

The initial-state issue also has a general thermodynamic formulation. For any fixed finite-time quantum process with a protocol-specific minimally dissipative input $H$2, the extra entropy production obeys

$H$3

The mismatch cost is therefore the contraction of relative entropy between the actual input and the minimally dissipative state under the protocol. In the nonunitary setting relevant for preparation and reset, the paper further notes that mismatched expectations can lead to divergent dissipation as the actual initial state becomes orthogonal to the anticipated one (Riechers et al., 2020).

4. Protocol families and physical platforms

Programmable neutral-atom platforms support several distinct dissipative preparation paradigms. In dipolar Rydberg arrays, “source” and “sink” auxiliary atoms implement nonreciprocal, energy-selective exchange with the system through

$H$4

which in the rotating frame becomes

$H$5

Under the rotating-wave approximation, the auxiliaries produce effective frequency-selective transition rates $H$6. The dynamics can be interpreted as a directional walk in Hilbert space. For ground-state preparation in a rotated frame defined by

$H$7

the target filling is

$H$8

By reversing the roles of source and sink or selecting an energy window $H$9, the same logic extends to excited many-body states. The steady states are explicitly described as not Gibbs states of a bath at a definite temperature, but as dissipatively selected eigenstates or narrow energy-window states (Tian et al., 20 Apr 2026).

Variational protocols adapt the same philosophy to NISQ hardware. The dissipative variational quantum algorithm introduces RESET-like single-qubit channels

$\mathcal L(\sigma_\beta)=0$0

as intrinsic ansatz elements, so mixed states are generated without ancilla qubits. The target is again the Gibbs state $\mathcal L(\sigma_\beta)=0$1, but the optimization uses infidelity rather than free energy. On 1D periodic rings of size $\mathcal L(\sigma_\beta)=0$2 to $\mathcal L(\sigma_\beta)=0$3, with $\mathcal L(\sigma_\beta)=0$4, the reported noiseless performance is often above $\mathcal L(\sigma_\beta)=0$5 fidelity, while noisy simulations on the XY model remain above $\mathcal L(\sigma_\beta)=0$6 in the tested low-noise regime and outperform the unitary ansatzes of Wang et al. (2021) and Consiglio et al. (2024) (Ilin et al., 2024).

A related but dynamically distinct construction uses a filtered reservoir during adiabatic evolution,

$\mathcal L(\sigma_\beta)=0$7

Here the instantaneous ground state is dark,

$\mathcal L(\sigma_\beta)=0$8

so the bath functions as a time-dependent dissipative thermalization channel that continuously recycles nonadiabatic leakage back into the low-energy sector. In the strong engineered-dissipation regime $\mathcal L(\sigma_\beta)=0$9, the avoided-crossing analysis yields a runtime scaling improvement from $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$0 to $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$1, while finite-temperature upward transitions impose a thermal error floor

$$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$2

The protocol is thus thermal-state-like in structure, but the target is the instantaneous ground state rather than a stationary Gibbs ensemble (Zhou et al., 4 Jun 2026).

5. Rigorous convergence, rapid mixing, and algorithmic complexity

The algorithmic theory of dissipative thermal state preparation is organized around mixing time, spectral gap, and detailed-balance structure. In “Optimal quantum algorithm for Gibbs state preparation,” the Gibbs sampler Lindbladian

$$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$3

has exact fixed point $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$4, and for $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$5-local Hamiltonians on a $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$6-dimensional lattice with $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$7, the paper proves a high-temperature threshold

$$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$8

such that for every $$\rho_f=|\Phi\rangle\langle\Phi|=|E_{n^\ast}\rangle\langle E_{n^\ast}|,$$9,

$$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$0

This is a rapid-mixing theorem in the standard sense: logarithmic convergence time in system size and inverse precision. Because the Lindbladian simulation is efficient, the result implies Gibbs preparation with $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$1 Hamiltonian-simulation time and $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$2 two-qubit gates, up to polylogarithmic factors (Rouzé et al., 2024).

An earlier high-temperature theorem proves efficient thermalization for any Hamiltonian satisfying a Lieb-Robinson bound. In that setting, the transformed Lindbladian retains a constant spectral gap for sufficiently small $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$3, and the sampler converges to $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$4 in polynomial time. The same paper also gives efficient adiabatic preparation of the purification $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$5, starting from a Bell-pair product state at $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$6, and proves the computational equivalence

$$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$7

in the low-temperature regime where the inverse temperature scales polynomially with system size (Rouzé et al., 2024).

The continuous-time Gibbs-sampler framework complements these results by giving nonasymptotic control of finite-time secular approximation and approximate detailed balance. For a full-rank state $$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$8, the similarity transform

$$H_S|\Phi\rangle=E_{n^\ast}|\Phi\rangle,\qquad L_\mu|\Phi\rangle=0,\ \forall\mu.$$9

is Hermitian under exact detailed balance, and if the anti-Hermitian part is small, then the fixed point remains close to $H_{\rm sys}$0. This provides a route from physically motivated, finite-time, filter-based generators to rigorous error estimates for thermal-state preparation and for purified Gibbs-state algorithms with quantum-walk speedup (Chen et al., 2023).

Weak-coupling collision models supply a separate layer of rigor. In that analysis, the implemented channel $H_{\rm sys}$1 differs from its effective Lindbladian approximation $H_{\rm sys}$2 by controlled weak-coupling and time-window errors, and the true fixed point $H_{\rm sys}$3 satisfies a trace-distance bound to $H_{\rm sys}$4 whose key scaling is quadratic in the coupling strength,

$H_{\rm sys}$5

The additional unitary evolution generated by the system Hamiltonian is not discarded; instead, it removes the zeroth-order fixed-point degeneracy and improves perturbative control of the thermal fixed point (Ong et al., 4 May 2026).

6. Conceptual boundaries, limitations, and open issues

A persistent source of confusion is the identification of all dissipative state preparation with thermalization. The recent literature is explicit that this is not correct. Some protocols prepare genuine Gibbs states by enforcing detailed balance, while others stabilize pure eigenstates, selected fillings, or narrow microcanonical-like shells. In the latter class, the bath is engineered rather than thermal, upward and downward rates are deliberately imbalanced, and the steady state is “thermal-like but not thermal” (Tian et al., 20 Apr 2026, Liu et al., 2023).

Another limitation concerns low temperature and spectral congestion. In ancilla-sweeping thermalization schemes, the Lorentzian tails of the induced spectral densities drive off-resonant transitions that violate detailed balance more strongly when the target temperature is low or the many-body gaps are dense. The reported numerical behavior is correspondingly less accurate in those regimes, even though intermediate-temperature performance can remain strong (Metcalf et al., 2019). In dynamically cooled adiabatic protocols, the same issue appears as a finite-temperature heating channel with steady excited-state population

$H_{\rm sys}$6

which sets an irreducible thermal floor (Zhou et al., 4 Jun 2026).

Transport constraints can also dominate the asymptotics. In the number-conserving preparation of the SSH ground state on a fermionic ladder, both passive thermal coupling and engineered dissipation exhibit cooling times

$H_{\rm sys}$7

because late-time relaxation is governed by dissipatively induced diffusion of defects rather than by local elimination. The engineered protocol is faster in practice because it has a smaller diffusion constant, but it does not escape the quadratic scaling (Pokart et al., 26 Jan 2026).

Practical analog implementations must also address resonance pathology. In the weak-coupling collision model, randomized timing offsets suppress resonances between the drive and many-body Bohr frequencies; without this randomization, deterministic implementations can exhibit $H_{\rm sys}$8 fixed-point errors even as $H_{\rm sys}$9. The tradeoff is an additional observable variance that scales as $$H(t)=H_{\rm sys}-\sum_{m=1}^M \frac{\Omega_m(t)}{2}\tau_z^m+\sum_{m=1}^M g_m \sigma_\alpha^{k_m}\tau_x^m,$$0, so randomization improves correctness while increasing sampling overhead (Ong et al., 4 May 2026).

At the algorithmic level, the subject has a sharp temperature dependence. High-temperature Gibbs samplers admit rapid-mixing theorems and efficient preparation guarantees for broad classes of local and sufficiently decaying long-range Hamiltonians (Rouzé et al., 2024, Rouzé et al., 2024). By contrast, low-temperature dissipative thermalization can become computationally universal, which rules out any expectation of a generic polynomial-time classical description in that regime unless standard complexity-class collapses occur (Rouzé et al., 2024). A plausible implication is that the main frontier is no longer the existence of dissipative thermal preparation schemes, but the delineation of those Hamiltonian families, temperature windows, and hardware models for which detailed-balance engineering, rapid mixing, and experimental implementability can be made simultaneously compatible.