DO-QSD: Optimal Quantum State Diffusion

• DO-QSD is a framework for optimal quantum state diffusion that employs explicit variational principles to minimize both short-time variance growth and large-deviation costs.
• It unifies distinct optimality criteria—variance-optimal unraveling for simulation efficiency and large-deviation control for rare-event dynamics—within a single class of QSD constructions.
• DO-QSD leverages stochastic differential equations, Doob-transform methods, and advanced numerical techniques, including machine learning, to enhance control in Markovian and non-Markovian open-system simulations.

Dynamically Optimal Quantum State Diffusion (DO-QSD) denotes quantum state diffusion selected by an explicit optimality principle. The phrase is used explicitly for Lindblad unravelings that minimize the short-time growth of the variance of an observable (Cao et al., 24 Sep 2025), and related work interprets controlled quantum state diffusion and quantum Doob processes as DO-QSD when their trajectories realize prescribed long-time fluctuations with minimal large-deviation cost (Carollo et al., 2021). This suggests that DO-QSD is best understood not as a single stochastic equation, but as a class of QSD constructions in which the unraveling, drift modification, or auxiliary process is fixed by an optimization criterion.

1. Terminology and conceptual range

Quantum state diffusion is an unraveling of open-system dynamics in terms of stochastic pure-state trajectories. In the Markovian homodyne setting, the same Lindblad generator admits multiple unravelings, with photon counting yielding quantum jump trajectories and homodyne detection yielding quantum state diffusion trajectories. DO-QSD arises when this freedom is constrained by a variational principle rather than left arbitrary.

Formulation	Optimality principle	Representative realization
Large-deviation DO-QSD	Minimal large-deviation cost for a target time-averaged observable	Controlled QSD or Doob-transformed Lindblad process
Variance-optimal DO-QSD	Minimal short-time growth of the variance of an observable	Observable-dependent diffusion unraveling with optimal state-dependent phase

The two formulations are not identical. One concerns rare-event realization in the long-time limit; the other concerns Monte Carlo efficiency for simulating Lindblad equations. This suggests that “dynamically optimal” refers to the criterion being optimized—large-deviation cost, variance growth, fidelity, purity, or related dynamical functionals—rather than to a unique canonical noise process.

2. Large-deviation, control, and Doob-transform DO-QSD

For Markovian open quantum systems with homodyne detection, QSD is written as the Itô stochastic differential equation for the pure-state density matrix $\psi_t$,

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$

with

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$

The measured homodyne current is

$dQ_t^m=\mathrm{Tr}[\kappa_m(\psi_t)]\,dt+dW_m.$

The level-2.5 large-deviation formalism introduces the empirical measure

$$\mu_\tau(\psi)=\frac{1}{\tau}\int_0^\tau dt\,\delta(\psi_t-\psi),$$

and empirical noise fields

$$j_\tau^m(\psi)=\frac{1}{\tau}\int_0^\tau dt\,\delta(\psi_t-\psi)\,dW_m(t).$$

Their fluctuations satisfy a large deviation principle with rate functional

$I_{2.5}^{\mathrm{qu}}(\mu,j)= \begin{cases} \displaystyle \frac{1}{2}\int d\psi\,\sum_{m=1}^M \frac{[j^m(\psi)]^2}{\mu(\psi)}, & \text{if the continuity constraint holds,}\[0.8em] +\infty, & \text{otherwise.} \end{cases}$

The controlled dynamics associated with a rare fluctuation is

$$d\psi_t= \Big[ \mathcal{L}(\psi_t)+\sum_m \frac{\mathcal{K}^m(\psi_t)\,j^m(\psi_t)}{\mu(\psi_t)} \Big]dt +\sum_m \mathcal{K}^m(\psi_t)\,dW_m.$$

Under this dynamics, $(\mu,j)$ become typical, and the same quadratic functional is the control cost.

For observables admitting a tilted generator, the construction can be recast in terms of the tilted Lindblad operator

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$0

and the associated Doob-transformed generator

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$1

In this framework, DO-QSD is the controlled QSD process, or its Lindblad Doob transform, whose trajectories realize a specified time-averaged observable with minimal large-deviation cost (Carollo et al., 2021).

3. Variance-optimal DO-QSD for Lindblad simulation

A second, narrower meaning of DO-QSD is explicit in recent work on optimal unraveling schemes for Lindblad equations. Here the problem is not rare-event realization but stochastic simulation error. For an observable $$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$2,

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$3

and an unraveling is dynamically optimal if it minimizes the instantaneous growth of this variance.

For the case of a single Lindblad operator and one noise term, the paper gives a complete parametric description of unraveling schemes with pathwise norm-preservation and then derives the dynamically optimal diffusion. The optimal phase is chosen from the phase $$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$4 of

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$5

through

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$6

The resulting diffusion term is therefore the usual centered QSD increment multiplied by an observable-dependent state-dependent phase. The paper further shows that, compared to jump process ansatz, DO-QSD has two notable advantages: first, the variance for DO-QSD can be rigorously shown not to exceed that of any jump-process ansatz locally in time; second, it has very simple expressions. In this usage, DO-QSD is an observable-dependent, locally-in-time optimal unraveling of a Lindblad equation (Cao et al., 24 Sep 2025).

A recurrent misconception is that this optimality is global. It is not: the criterion is the short-time growth of the variance of an observable. A second misconception is that the result is observable-independent. It is not: the phase choice is built from the covariance-like quantity involving $$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$7 and $$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$8.

4. Non-Markovian, multilevel, and many-body extensions

The non-Markovian QSD literature provides the main technical machinery for extending optimality ideas beyond Markovian Lindblad simulation. In the standard bosonic-environment setting, the formal equation

$$d \psi_t = \mathcal{L}(\psi_t)\, dt + \sum_{m=1}^M \mathcal{K}^m(\psi_t)\, dW_m,$$9

is rendered time-local by introducing an $$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$0-operator,

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$1

so that

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$2

For multilevel open systems, exact time-local QSD equations have been constructed for high-spin systems, multiple-transition atomic models, and multilevel atoms driven by time-dependent classical fields; in many driven multilevel and multi-transition models, the $$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$3-operator can be noise-free (Jing et al., 2012).

The same framework supports an invariant-based control viewpoint. In non-Markovian QSD, the dynamical invariant is defined by

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$4

with

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$5

Using a bi-orthonormal eigenbasis of $$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$6, the coefficients of the stochastic state decouple and admit exact expressions. The same formalism is used to reverse-engineer Hamiltonians and couplings that drive arbitrary initial states to a target state under non-Markovian QSD dynamics (Luo et al., 2015).

For many-body open systems, exact time-local QSD equations have been constructed for $$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$7-qubit dissipative models with

$$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$8

and the exact $$\mathcal{L}(\psi) = -i[H,\psi] + \sum_{m=1}^M \left( L_m \psi L_m^\dagger - \frac{1}{2}\{\psi,L_m^\dagger L_m\} \right),$$9-operator contains noise up to order $$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$0 (Jing et al., 2010). A complementary reduction is obtained by concatenating non-Markovian QSD with Feshbach projection operator partitioning, yielding an exact one-dimensional stochastic master equation for a chosen component,

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$1

which is explicitly proposed as a general tool for controlling an arbitrary component of the system (Jing et al., 2011).

These results do not by themselves define DO-QSD, but they supply the exact time-local stochastic equations, invariant structures, and reduced control equations that any non-Markovian DO-QSD program requires.

5. Representative dynamical phenomena and model systems

A canonical Markovian example is diffusion on the Bloch sphere for the qubit Lindblad dynamics

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$2

with unravelled QSD

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$3

In Bloch coordinates this becomes isotropic diffusion on the sphere, and the time-averaged coherence

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$4

is analyzed variationally through the level-2.5 functional. This is the clearest explicit example of DO-QSD as a rare-fluctuation process made typical by optimal control (Carollo et al., 2021).

In non-Markovian two-qubit systems, diffusive quantum trajectories were used to estimate entanglement in common-bath dissipative and dephasing models. Exact time-local QSD equations reveal entanglement revival, delayed disentanglement, and the role of memory time in maintaining concurrence (Corn et al., 2011). For interacting qubits coupled to a common bosonic environment, an exact non-Markovian QSD equation without any approximations shows that entanglement generation is significantly modulated by environmental memory and that residual entanglement can survive in the steady state (Zhao et al., 2011).

For higher-dimensional systems, an exact time-local QSD equation was derived for a dissipative three-level model with

$$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$5

showing long-tailed non-Markovian relaxation and coherence behavior beyond the Markov limit (Jing et al., 2010). These examples illustrate that “optimality” in QSD is not confined to one metric: it can mean minimal large-deviation cost, best entanglement tracking, target-state steering, or control of leakage and decoherence.

A related open-system performance criterion is the quantum speed limit. A QSD-based bound written from the total-system perspective expresses the Bures angle and Fubini–Study metric through ensemble averages of QSD trajectories. For a two-level system it is shown that the infinite speedup capacity of the noiseless case is destroyed by the environment under the Born-Markovian approximation and recovered in non-Markovian dynamics as long as a bound state is formed in the energy spectrum of the total system (Wu et al., 2022). This suggests a further DO-QSD interpretation in which reservoir engineering is used to approach a QSD-based speed limit.

6. Numerical methods, scope, and current limitations

The practical utility of DO-QSD depends strongly on numerical propagation. For Markovian Lindblad dynamics, weak first- and second-order solvers for nonlinear QSD have been constructed directly from the Itô–Taylor expansion in the interaction picture. The second-order solver delivers much higher accuracy and stability with bigger time steps than the first-order scheme, with a small additional workload, although the second-order algorithm has quadratic complexity with the number of Lindblad operators as opposed to the linear complexity of the first-order algorithm (Adhikari et al., 2024).

For non-Markovian dynamics, machine-learning approaches have begun to reconstruct the stochastic time-evolution operator from ensembles of QSD trajectories. An operator-based neural-network method learns the stochastic propagator for non-Markovian QSD and then uses it to compute absorption spectra and reconstruct reduced density matrices at extended timescales (Zhang et al., 1 Sep 2025). This does not define DO-QSD directly, but it is a natural computational route for operator-level optimization of stochastic propagators and memory kernels.

A common misconception is that QSD-based optimization requires complete positivity. Standard QSD does call for the complete positivity of the open-system dynamics, but a generalized stochastic unraveling has been formulated for positive, but not completely positive evolutions, first for semigroup dynamics and then for a definite class of time-dependent generators (Caiaffa et al., 2016). This broadens the mathematical domain in which trajectory-level optimality can be posed.

The main limitation remains structural. Exact non-Markovian control constructions generally assume that the $$\mathcal{K}^m(\psi)=\kappa_m(\psi)-\psi\,\mathrm{Tr}[\kappa_m(\psi)], \qquad \kappa_m(\psi)=e^{i\alpha_m}L_m\psi+\psi e^{-i\alpha_m}L_m^\dagger.$$6-operator can be found exactly or in closed form, and this becomes difficult for large, interacting, or strongly driven systems. The currently explicit meanings of DO-QSD are also criterion-specific: large-deviation DO-QSD is a rare-event optimal-control object, whereas variance-optimal DO-QSD is an observable-dependent local Monte Carlo optimizer. This suggests that the topic is presently a family of closely related optimization principles built on QSD, rather than a single universally standardized formalism.