Channel-Constrained Quantum Diffusion

• Channel-constrained quantum state diffusion is a framework where the forward process is defined by a physical CPTP channel and the reverse denoising is learned as a recovery map.
• The method is implemented in continuous-variable and multi-qubit systems, achieving state fidelities often exceeding 99% even under thermal and depolarizing noise.
• Robust parameterizations, including CV quantum neural networks and Kraus-operator based maps, ensure physical validity, training efficiency, and strong inductive biases for state restoration and generation.

Channel-constrained quantum state diffusion denotes a class of quantum diffusion models in which the forward diffusion is fixed to a physical quantum channel and the reverse process is learned as a completely positive trace-preserving (CPTP) denoising map. In this formulation, diffusion is not an abstract corruption rule but an open-system evolution: in continuous-variable (CV) systems it may be a thermal loss channel implemented by beam-splitter coupling to a thermal bath, while in discrete-variable settings it may be a Markovian channel generated by a time-local GKSL master equation. The central objective is to preserve physical validity at every step of the trajectory, so that intermediate objects remain density operators and the learned reverse dynamics remain realizable quantum channels. This framing supports both quantum state generation and state restoration, and it has been instantiated by the CVQD-G/CVQD-R framework for CV systems and by the channel-constrained Markovian quantum diffusion model (CCMQD) for multi-qubit open-system dynamics (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

1. Definition and physical rationale

The defining constraint is that the forward process is specified by a known CPTP channel rather than by an unconstrained noise field. In the CV formulation, the forward diffusion is the evolution of the state through a thermal loss channel, and the reverse denoising map is implemented as a unitary on system plus ancilla followed by a partial trace, parameterized by a CV quantum neural network (CVQNN). In the Markovian discrete-variable formulation, both forward and backward maps are explicit quantum channels in Kraus form, and the entire trajectory is treated as a controlled Markov chain in the space of density matrices (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

This physical restriction is intended to exclude non-physical intermediate representations. In the CV setting, constraining the forward trajectory to thermal loss makes the evolution completely positive and trace-preserving by construction, while the reverse map is CPTP by Stinespring dilation. In the open-system multi-qubit setting, the same principle is enforced by explicit Kraus-operator parameterizations that satisfy trace preservation and complete positivity at every step. A common misconception is that the reverse process is an exact inverse of the forward channel. The open-system formulation states explicitly that exact inverses of CPTP maps need not be CP and are generally not TP; consequently, the learned reverse is an approximate CPTP recovery channel rather than a literal channel inverse (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

The conceptual link to classical diffusion models is direct but not identical. Like DDPMs and SDE-based models, channel-constrained quantum diffusion defines a forward noise process and learns a reverse denoiser conditioned on time. The difference is that the forward evolution is a physical quantum channel, the reverse is a CPTP map realized by ancilla-assisted dynamics, training is performed on density operators rather than score fields in phase space, and fidelity-based objectives replace classical score matching in the core implementations considered here (Huang et al., 24 Jun 2025).

2. Forward diffusion as open-system channel dynamics

In the CV thermal-loss instantiation, the forward channel is modeled in the Heisenberg picture by

$$a_{\mathrm{out}}=\sqrt{\eta}\,a_{\mathrm{in}}+\sqrt{1-\eta}\,b_{\mathrm{th}}, \qquad 0\le \eta \le 1,$$

where $b_{\mathrm{th}}$ is a thermal bath mode with mean occupancy $N_{\mathrm{th}}$. The corresponding Schrödinger-picture channel is

$$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$

For a single-mode Gaussian state with mean $\mu_0$ and covariance $V_0$, the forward evolution satisfies

$$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$

Under homogeneous thermal loss, $\eta(t)=e^{-\kappa t}$, and the paper also shows a direct multi-step-to-one-step sampling relation through $\bar{\eta}_t=\prod_{i=1}^t \eta_i$, which allows any intermediate $\rho_t$ to be sampled in one shot from $b_{\mathrm{th}}$0 (Huang et al., 24 Jun 2025).

In the CCMQD formulation, the forward process is the Markovian open-system evolution

$b_{\mathrm{th}}$1

with time-dependent rates $b_{\mathrm{th}}$2. Discretization produces stepwise CPTP channels,

$b_{\mathrm{th}}$3

This makes the forward diffusion a composition of channels, $b_{\mathrm{th}}$4, and permits generator-specific constructions such as depolarizing, dephasing, and amplitude-damping schedules by appropriate choice of $b_{\mathrm{th}}$5 and $b_{\mathrm{th}}$6 (Zhu et al., 15 Nov 2025).

The common structural point is that the corruption process is physically interpretable. In the CV case, the state is driven toward thermal statistics; in the discrete-variable case, the state is driven toward the maximally mixed state under natural decoherence. This suggests that “diffusion” is best understood not as an arbitrary interpolation between data and noise, but as a controlled decoherence trajectory inside the CPTP cone (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

3. Reverse denoising channels and training objectives

The reverse process is learned as a family of CPTP denoisers conditioned on the timestep. In CVQD, the denoising map has the form

$b_{\mathrm{th}}$7

where $b_{\mathrm{th}}$8 is a timestep-conditioned coherent ancilla generated by a time-embedding circuit $b_{\mathrm{th}}$9 with $N_{\mathrm{th}}$0. The CVQNN acts on two modes, system and ancilla, and comprises displacements, rotations, squeezers, beam splitters, and a Kerr gate

$N_{\mathrm{th}}$1

Parameter sharing across all timesteps is enabled by the time embedding, so one shared $N_{\mathrm{th}}$2 replaces $N_{\mathrm{th}}$3 distinct reverse networks (Huang et al., 24 Jun 2025).

Training in the CV formulation uses a fidelity-based loss with a normalization penalty to counteract Fock truncation,

$N_{\mathrm{th}}$4

with

$N_{\mathrm{th}}$5

The total loss averages over randomly sampled timesteps and includes a step-$N_{\mathrm{th}}$6 term that explicitly targets $N_{\mathrm{th}}$7 (Huang et al., 24 Jun 2025).

CCMQD adopts an explicitly channel-parameterized reverse family $N_{\mathrm{th}}$8 in Kraus form. To guarantee trace preservation exactly during learning, the Kraus operators are stacked into an isometry

$N_{\mathrm{th}}$9

so optimization is carried out on the complex Stiefel manifold. The model distinguishes Sequential Quantum Channel Optimization (SQCO), which trains each reverse step independently, from Holistic Quantum Trajectory Optimization (HQTO), which trains the full reverse chain end-to-end. The path-constrained loss augments endpoint fidelity with intermediate supervision,

$$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$0

thereby regularizing the learned denoising trajectory rather than only its endpoint (Zhu et al., 15 Nov 2025).

A second misconception is that physical constraints necessarily reduce expressivity below useful levels. The reported results instead associate CPTP-constrained parameterizations with stability, parameter efficiency, and physically plausible recovery trajectories. A cautious reading is that the constraints act as a strong inductive bias; this is an interpretation, but it is consistent with the observed parameter efficiency and robustness across noise structures in both the CV and discrete-variable implementations (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

4. Continuous-variable thermal-loss diffusion: CVQD-G and CVQD-R

The CVQD framework specializes channel-constrained diffusion to continuous-variable quantum optics. CVQD-G is the generative model: a target state $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$1 is diffused through the thermal channel to obtain $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$2, and the learned reverse process starts from the thermal prior and reconstructs $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$3. CVQD-R is the restoration model: it is trained on thermally degraded coherent states and learns to map $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$4 across a family of displacements, phases, and transmissivities, with the aim of restoring coherent states with unknown parameters (Huang et al., 24 Jun 2025).

The paper reports high-fidelity synthesis for both Gaussian and non-Gaussian states. For generation, coherent $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$5 states achieve approximately $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$6 fidelity under pure loss ($$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$7) and approximately $$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$8 under thermal noise ($$\rho_{\mathrm{out}}=\mathrm{Tr}_{E}\!\left[ U_{\mathrm{BS}}(\eta)\big(\rho_{\mathrm{in}}\otimes\rho_{\mathrm{th}}(N_{\mathrm{th}})\big) U_{\mathrm{BS}}^{\dagger}(\eta) \right].$$9). Squeezed states $\mu_0$0 reach approximately $\mu_0$1 for $\mu_0$2 and approximately $\mu_0$3 for $\mu_0$4. The Fock state $\mu_0$5 reaches approximately $\mu_0$6 under pure loss, and the even cat state with $\mu_0$7 reaches approximately $\mu_0$8 under pure loss; both are reported as robust under $\mu_0$9 as well. The abstract summarizes this more broadly by stating that diverse Gaussian and non-Gaussian states are generated typically with fidelities exceeding $V_0$0 (Huang et al., 24 Jun 2025).

For restoration under $V_0$1, CVQD-R attains average final fidelities of approximately $V_0$2 for $V_0$3, approximately $V_0$4 for $V_0$5, and approximately $V_0$6 for $V_0$7 at fixed noise $V_0$8. Under variable noise $V_0$9, the restoration curves are nearly identical, with a final fidelity of approximately $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$0, which the paper interprets as robustness to the initial corruption level. The forward corruption law for coherent inputs,

$$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$1

is used to motivate the model’s ability to recognize noise levels during restoration (Huang et al., 24 Jun 2025).

The computational profile is a major part of the CVQD proposal. Inference circuit depth scales as $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$2, where $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$3 is the number of CVQNN layers per reverse step. Training complexity scales as $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$4 and is notably independent of $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$5 because direct multi-step-to-one-step sampling through $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$6 avoids sequential simulation of all diffusion steps. The reported implementation uses two modes per step, system plus ancilla, and the paper argues that the framework extends naturally to pure loss, additive Gaussian noise, phase diffusion, and multi-mode entangled targets by replacing the forward channel while preserving the same channel-constrained logic (Huang et al., 24 Jun 2025).

5. Markovian multi-qubit diffusion: CCMQD

CCMQD formulates channel-constrained quantum diffusion explicitly from the perspective of open quantum systems. The forward trajectory is a sequence of CPTP-divisible Markovian channels, interpreted as natural decoherence, and the reverse trajectory is a learned sequence of CPTP recovery channels. The organizing mathematical objects are Kraus operators, Choi positivity, and Stiefel-manifold parameterizations that enforce trace preservation exactly throughout optimization. The paper emphasizes that the whole process is a controlled Markov evolution in the space of density matrices, not merely a heuristic denoising chain (Zhu et al., 15 Nov 2025).

The experimental setting spans single-qubit through seven-qubit pure states, including entangled targets. Forward channels include structured depolarizing channels with linearly increasing noise strength and unstructured Haar-random Kraus channels representing chaotic decoherence. Typical configurations use forward $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$7 and backward $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$8, with further sweeps over diffusion depth and Kraus count. HQTO plus path-constrained loss consistently outperforms SQCO, especially beyond one or two qubits, which the paper attributes to the non-local temporal structure of multi-qubit coherence (Zhu et al., 15 Nov 2025).

The quantitative results are strong throughout the reported range. HQTO plus path-constrained loss achieves fidelities exceeding $$\mu_t=\sqrt{\eta_t}\,\mu_0,\qquad V_t=\eta_t\,V_0+(1-\eta_t)\,\nu_{\mathrm{th}}\,I_2,\qquad \nu_{\mathrm{th}}=\frac{2N_{\mathrm{th}}+1}{2}.$$9 across one to three qubits under both random and depolarizing noise, and maintains fidelity above $\eta(t)=e^{-\kappa t}$0 up to seven qubits. Representative averages over five runs include $\eta(t)=e^{-\kappa t}$1 for one qubit under random noise, $\eta(t)=e^{-\kappa t}$2 for two qubits under random noise, and $\eta(t)=e^{-\kappa t}$3 for four qubits under random noise; the corresponding depolarizing results are $\eta(t)=e^{-\kappa t}$4, $\eta(t)=e^{-\kappa t}$5, and $\eta(t)=e^{-\kappa t}$6. A depth sweep with $\eta(t)=e^{-\kappa t}$7 from $\eta(t)=e^{-\kappa t}$8 to $\eta(t)=e^{-\kappa t}$9 yields fidelities from $\bar{\eta}_t=\prod_{i=1}^t \eta_i$0 to $\bar{\eta}_t=\prod_{i=1}^t \eta_i$1, and increasing the reverse Kraus count beyond approximately $\bar{\eta}_t=\prod_{i=1}^t \eta_i$2 per step produces only marginal gains (Zhu et al., 15 Nov 2025).

Comparisons to prior baselines reinforce the role of the channel-constrained design. EQGAN fidelity is reported to decline from $\bar{\eta}_t=\prod_{i=1}^t \eta_i$3 at one qubit to $\bar{\eta}_t=\prod_{i=1}^t \eta_i$4 at seven qubits, while RQGDM achieves $\bar{\eta}_t=\prod_{i=1}^t \eta_i$5 to $\bar{\eta}_t=\prod_{i=1}^t \eta_i$6 with larger variance across two to seven qubits. CCMQD, by contrast, consistently achieves approximately $\bar{\eta}_t=\prod_{i=1}^t \eta_i$7 with small variance up to seven qubits. The computational cost remains the primary classical limitation: each channel application costs $\bar{\eta}_t=\prod_{i=1}^t \eta_i$8 with naive dense algebra, total forward plus backward cost is $\bar{\eta}_t=\prod_{i=1}^t \eta_i$9, memory for one density matrix is $\rho_t$0, and $\rho_t$1 sets the practical simulation ceiling. The reported demonstrations therefore stop at $\rho_t$2 qubits (Zhu et al., 15 Nov 2025).

6. Related formulations, limitations, and research directions

Channel-constrained state diffusion is part of a broader family of quantum diffusion models that replace unconstrained reverse dynamics with physically realizable maps. QGDM drives states to the maximally mixed state through depolarizing channels and learns the reverse through variational circuits plus partial trace, with parameter sharing across timesteps and a resource-efficient variant that reduces ancilla count while preserving high fidelity on reported tasks (Chen et al., 2024). MSQuDDPM likewise replaces scrambling unitaries by depolarizing channels in the forward process and implements reverse denoising by parameterized circuits with ancilla qubits and projective measurements; it adds a cosine-exponent schedule, single-qubit random ancilla, and superfidelity-based costs for mixed-state generation (Kwun et al., 2024). Structure-preserving diffusion on density matrices takes a different route: it uses mirror diffusion with the negative von Neumann entropy mirror map so that decoding by matrix exponential and trace normalization enforces Hermiticity, positivity, and unit trace exactly (Zhu et al., 2024).

A broader interpretation of “channel-constrained diffusion” also appears outside generative modeling. In noisy discrete-time quantum walks, temporal coin noise admits a continuum Dirac–Lindblad limit in which the allowed jump operators $\rho_t$3 and $\rho_t$4 determine the effective diffusion law, with long-time coefficient $\rho_t$5 in the massless limit (Arnault et al., 2019). In graph-based multi-walker diffusion, channel constraints are literally graph edges: the one-step evolution is a CPTP channel that is a convex mixture of swap-products over graph matchings, with locality enforced by edge-selection rules (III et al., 2024). These works use the same language of channel-constrained diffusion, but the object being diffused is transport on position or graph degrees of freedom rather than a denoising trajectory toward a target quantum state.

Several limitations recur across the literature. Channel knowledge or calibration is assumed in the CVQD formulation, including the schedule $\rho_t$6 and thermal occupancy $\rho_t$7, and mismatches can affect inverse quality (Huang et al., 24 Jun 2025). CCMQD assumes Markovianity and CPTP divisibility; the paper states that in non-Markovian regimes there may not exist CPTP local inverses even approximately, and extensions would require process tensors or non-Markovian master equations (Zhu et al., 15 Nov 2025). Numerical CV training relies on a finite Fock cutoff and uses a trace penalty to mitigate normalization drift, whereas discrete-variable simulation faces exponential growth in Hilbert-space dimension (Huang et al., 24 Jun 2025, Zhu et al., 15 Nov 2025).

The main open directions follow directly from these constraints. One is extension to other channels, including additive Gaussian noise, phase diffusion, and broader multi-mode or multi-qubit noise families (Huang et al., 24 Jun 2025). Another is the incorporation of memory effects beyond Markovianity through multi-time physicality constraints rather than per-step CPTP constraints (Zhu et al., 15 Nov 2025). A plausible implication is that future progress will depend less on importing ever more classical diffusion machinery and more on designing reverse processes that are native to quantum hardware, channel structure, and open-system control.