Generating quantum ensembles via reverse-time quantum diffusions

Abstract: We establish a reverse-time denoising theory for quantum diffusions of continuously measured quantum systems. Starting from the stochastic Schrödinger equation of a forward noising dynamics, we derive the exact reverse-time dynamics for quantum trajectories, whose law coincides with the time-reversal of the original process. We prove that the denoising dynamics is a physically admissible quantum diffusion, with the same measurement-induced noise but a state-dependent feedback Hamiltonian, a direct analogue of the "score function" of generative classical diffusion models. This provides a principled framework for converting samples of a simple distribution into those of a more complex ensemble of quantum states. We show how the denoising dynamics can be directly learnt from forward trajectory data, and how to exploit purification to initialise the denoising process.

• The paper establishes an exact reverse-time construction for continuously monitored quantum systems, preserving the same measurement noise while encoding denoising in a state- and time-dependent feedback Hamiltonian analogous to a classical score function.
• The framework uses exponential mixing and purification to initialize reverse sampling from an invariant ensemble, while trajectory-level equality in law enables training the denoiser from reversed measurement records.
• A qubit demonstration shows Wasserstein convergence toward a target great-circle ensemble as the score expansion improves, but scalability, approximation error, sample complexity, and generalization remain open challenges.

The paper "Generating quantum ensembles via reverse-time quantum diffusions" (2606.03848) develops a mathematically rigorous framework for quantum generative modelling based on the theory of reverse-time diffusions. The central contribution is an exact construction of the time-reversed dynamics of a continuously measured quantum system, showing that the denoising process is itself a physically admissible quantum diffusion driven by the same measurement noise, but with a state-dependent feedback Hamiltonian that plays the role of the score function in classical score-based generative models. The framework thereby provides correctness guarantees that the authors argue are absent or incomplete in prior quantum adaptations of diffusion models, and it grounds the stochasticity entirely in the physical measurement process rather than in ad hoc classical noise sources.

Forward noising dynamics

The forward process is a stochastic Schrödinger equation (SSE) for pure states $\psi_t$ on a finite-dimensional Hilbert space, unravelling a Lindblad semigroup generated by $M$ jump operators $L_m$ coupled to Bosonic environment channels. The drift is the Lindblad generator $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$, and the diffusion coefficients $\mathcal K_m$ are the standard homodyne-type terms. The authors assume ergodicity of the master evolution together with a purification condition, which guarantees a unique invariant distribution $\mu_{\mathrm{inv}}$ over pure states and exponential mixing: $$W_1(\mu_t, \mu_{\mathrm{inv}}) \le C e^{-\lambda t}$$. This exponential convergence is what makes the forward process usable as a noising mechanism — after a sufficiently long time $T$, samples of any structured initial ensemble $\mu_0$ become indistinguishable from draws of the easy-to-sample invariant measure.

The main result: exact reverse-time dynamics

The core theorem constructs a process $(\tilde\psi_t)$ satisfying equality in law with the time-reversed forward trajectories,

$M$0

while remaining a physically admissible quantum diffusion. Three structural properties characterise the solution. First, the backward drift is again of Lindblad form. Second, the diffusion coefficients are identical to those of the forward dynamics ($M$1), so the same homodyne measurement noise drives both processes. Third, the entire reversal is encoded in a state-dependent, time-dependent feedback Hamiltonian

$M$2

with $M$3 comprising four contributions: the coherent term $M$4, the dissipative correction $M$5, the divergence of the diffusion tensor $M$6, and the score term $M$7. The supplementary material proves that this $M$8 satisfies the commutator-direction conditions $M$9 and $L_m$0, which is necessary and sufficient for it to be representable as a commutator with a Hermitian operator; hence $L_m$1 is self-adjoint and the reverse SSE preserves purity.

Two aspects of this result deserve emphasis. The guarantee holds at the level of stochastic processes (equality in law of full trajectories), not merely at the level of marginal state distributions — this is precisely what enables training the denoiser directly from reversed forward trajectory data. Additionally, the authors explicitly distinguish their construction from the continuous-time Petz map and its generalisations: the Petz map reverses only the average state $L_m$2, whereas the present reversal acts on ensembles of stochastic trajectories.

Initialisation via purification

A practical obstacle for any reverse-diffusion scheme is preparing the terminal distribution $L_m$3. The paper exploits purification under continuous monitoring to resolve this: starting from an arbitrary — possibly mixed — state and running the forward dynamics for large $L_m$4 yields a pure state $L_m$5 drawn from $L_m$6, whose value is known from the measurement record. The forward run thus serves a dual purpose, both sampling from the noise distribution and providing the tomographic knowledge of $L_m$7 required as the initial condition of the denoiser. Because the reverse control $L_m$8 depends explicitly on the instantaneous state, this continuous state estimation during the backward pass also confers intrinsic robustness against imperfections in the implemented feedback.

Learning the denoising dynamics from data

As in classical score-based diffusion, the exact score requires knowledge of $L_m$9, so in practice the feedback Hamiltonian must be learned. The equality in law permits using time-reversed forward trajectories as training samples for the backward process. Two training regimes are distinguished:

• Level 1 (full trajectories known): the backward Hamiltonian is parametrised as $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$0 over a fixed basis $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$1, and the coefficients are fitted by least-squares regression on discretised residuals of the backward SDE.
• Level 2 (measurement signal only, unknown initial states): when the initial states $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$2 are unknown, the model's observability — defined via the span of iterated adjoint superoperators applied to the constant function — guarantees identifiability of $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$3 from multi-time correlation functions of the measurement currents, after which Level 1 applies.

This observability-based identifiability result connects the learning procedure to system-identification techniques for quantum diffusions and extends the applicability of the method to settings where the target ensemble is not directly preparable or characterisable.

Qubit illustration

The framework is illustrated on a qubit with $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$4 and jump operators $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$5, targeting the uniform distribution over a great circle of the Bloch sphere. In this example the terms $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$6 and $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$7 cancel exactly — a cancellation the authors note is specific to this choice of jump operators and does not hold generally — leaving a reverse Hamiltonian of purely score-driven form,

$$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$8

where $$\mathcal L(\cdot) = -i[H,\cdot] + \mathcal D(\cdot)$$9 and $\mathcal K_m$0 is the score of the one-dimensional marginal density. Under the forward dynamics, $\mathcal K_m$1 obeys a Jacobi diffusion with Legendre polynomials as eigenfunctions (eigenvalues $\mathcal K_m$2), yielding a spectral expansion of the score truncated at order $\mathcal K_m$3. Numerically, the Wasserstein distance $\mathcal K_m$4 between the generated ensemble and the target decreases toward zero as denoising proceeds, with convergence improving as the truncation cutoff increases. The structure of $\mathcal K_m$5 mirrors classical score-based drifts: it vanishes when the state already commutes with $\mathcal K_m$6 (i.e., lies in the target ensemble) and its magnitude is modulated by the log-density gradient.

Limitations and open questions

Several caveats qualify the results. The exponential-mixing assumption underlying the noising stage restricts the admissible Lindbladians, and the purification-based initialisation relies on this condition holding. The exact reverse dynamics requires knowledge of the instantaneous density $\mathcal K_m$7 along the whole backward path; the learned approximations inherit whatever bias the parametrisation class $\mathcal K_m$8 introduces, and the paper does not quantify approximation error or sample complexity for the regression procedure. In the illustrative example, the exact cancellation between dissipative and divergence terms is special to the Pauli jump operators, so the general case retains additional drift contributions whose learnability is not demonstrated. Scalability beyond the single-qubit demonstration remains open: the authors propose combining the framework with quantum model reduction techniques to handle high-dimensional systems, but do not carry this out. Finally, questions of memorisation versus generalisation — whether trained denoisers reproduce training trajectories or genuinely sample the target ensemble, and when models trained on distinct datasets become equivalent — are raised but left unaddressed within the quantum diffusive setting.

Conclusion

The paper establishes that every quantum diffusion arising from continuous monitoring admits an exact reverse-time evolution that is itself a physically realisable feedback-controlled quantum diffusion, with the same measurement noise and a state-dependent Hamiltonian analogous to the classical score. Combined with purification-based initialisation and trajectory-level training procedures, this yields a principled and provably correct route from easily sampled invariant ensembles to structured distributions of quantum states, with the qubit example confirming quantitative convergence. The remaining challenges are those of scale and of learning-theoretic guarantees for the score approximation in high dimension.