Pith. sign in

REVIEW 3 major objections 6 minor 2 cited by

Continuous-variable Quantum Diffusion Model for State Generation and Restoration

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

Pith's one-line read One continuous-variable quantum circuit, shared across diffusion steps by a time embedding, generates coherent, squeezed, Fock, and cat states above 99 percent fidelity in simulation and restores unknown coherent states from thermal loss.

desk verdict Restoration variant is the real contribution; generation claim rests on an unverified noise schedule that doesn't obviously make the terminal state thermal. read the letter →

arxiv 2506.19270 v1 pith:KLPBKMMY submitted 2025-06-24 quant-ph cs.LG

classification quant-phcs.LG PACS 03.67.-a
keywords Continuous-VariableQuantumInformationDiffusionModelsStateGenerationRestorationNeuralNetworksThermalLossChannel
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that diffusion-style training, borrowed from classical generative modeling, works for continuous-variable quantum states when the forward noise is a physically motivated thermal loss channel. The central object is a single circuit — a continuous-variable quantum neural network with a time embedding — that learns to undo one step of thermal degradation, so that applying it repeatedly walks a thermal state back to a target quantum state. In numerical simulations the framework generates coherent, squeezed, Fock, and cat states with fidelities typically above 99 percent, and a restoration variant recovers coherent states of unknown amplitude and phase from thermal loss with fidelities around 89 to 98 percent. If the framework works as claimed, state preparation and noise repair in continuous-variable systems become one adaptable procedure rather than a fixed input-output mapping.

What carries the argument

The mechanism that carries the argument is the pair of a thermal loss channel for the forward process and a learned reverse circuit. The forward map mixes the system qumode with a thermal environment of mean photon number $\bar{n}$ through a beam splitter of transmissivity $\eta_t$ per step; Theorem 1 reduces $t$ repeated steps to a single channel of cumulative transmissivity $\bar{\eta}_t = \prod_{i=1}^t \eta_i$, giving direct access to any intermediate state $\rho_t$ from $\rho_0$. The reverse map is a two-qumode circuit $U(\vartheta)$ of $L$ layers built from continuous-variable neural-network gates (displacement, rotation, squeezing, beam splitter) plus a Kerr gate for nonlinearity, applied to the tensor product of a time-embedding state and the noisy state, with the ancilla qumode traced out to produce a non-unitary one-step denoiser. The time embedding encodes timestep $t$ as the phase $\phi(t) = t\pi/T$ of a coherent state, so one parameter set serves all $T$ steps, and the loss is one minus the fidelity to the previous forward state plus a trace-normalization penalty.

What would settle it

A concrete check: the paper never reconciles its two schedule columns — the linear $\beta$ schedule ($1.0\times10^{-4}$ to $0.05$) implies a cumulative transmissivity near $0.06$ under $\eta_t = 1-\beta_t$, while the stated linear $\eta$ schedule ($0.99974$ to $0.99331$) over $T=112$ steps implies about $0.68$, an order-of-magnitude gap that leaves $\rho_T$ ill-defined. Evaluating $F(\rho_T, \rho_{\mathrm{th}}(\bar{n}))$ for each target state and running the backward algorithm from the genuinely trained terminal state $E(\rho_0, \bar{\eta}_T)$ instead of from the idealized thermal state would show whether the reported fidelities hold for the states the training actually produced.

Watch

Extended reading notes

Core claim

The paper's central claim is that the thermal-loss forward diffusion is invertible: a single parameter-shared denoising function $f_\vartheta(\rho_t, t)$, applied $T$ times to a thermal state, progressively recovers the target state $\rho_0$. The structural result that carries this is Theorem 1, which collapses $t$ successive thermal-loss steps into one effective channel with cumulative transmissivity $\bar{\eta}_t = \prod_{i=1}^t \eta_i$ acting on $\rho_0$ with a single thermal environment of the same mean photon number $\bar{n}$; this makes training cost $O(I \times B \times L)$ independent of $T$ and lets the model sample any $\rho_t$ directly from $\rho_0$. On that basis the paper reports generation fidelities typically above 99 percent for coherent, squeezed, Fock, and cat states in both pure-loss ($\bar{n}=0$) and thermal-loss ($\bar{n}=0.5$) environments, and restoration of unknown coherent states from thermal degradation with fidelities from about 89 to 98 percent.

Load-bearing premise

Two premises carry the result: that after $T = 112$ forward steps $\rho_T$ is effectively the thermal state from which generation starts, although the paper's own numbers leave the overlap $F(\rho_0, \rho_T)$ 'generally non-zero', and that a hundred or more iterations of the one-step denoiser compose without accumulating error.

Editorial extensions

If this is right

  • Coherent, squeezed, Fock, and cat states can be generated from a thermal starting state with fidelities typically above 99 percent, in both pure-loss ($\bar{n}=0$) and thermal-loss ($\bar{n}=0.5$) environments.
  • Because training cost is $O(I \times B \times L)$ and independent of the total diffusion steps $T$, longer diffusion chains do not raise per-iteration training cost, and the shared time embedding keeps the parameter count fixed as $T$ grows.
  • A single restoration-trained model can recover coherent states of unknown amplitude and phase from thermal loss channels of unknown transmissivity, reaching roughly 89 to 98 percent fidelity depending on the state's amplitude.
  • For a fixed target state, restoration fidelity is nearly independent of the corruption level ($\eta \in \{0.25, 0.5, 0.75\}$ produce near-identical fidelity trajectories), indicating the model learns a general denoising map rather than noise-level-specific ones.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because Theorem 1 makes the forward process an exact Gaussian channel, a natural test is to compare the trained denoiser with the theoretically optimal estimator for that channel; a close match would reframe the method as learned channel inversion and link it directly to quantum error mitigation.
  • The paper reports fidelity only; a sharper probe of genuine denoising would track Wigner negativity and purity of cat and Fock states through the backward chain, since true inversion of thermal loss must recreate non-classical features step by step rather than merely raising overlap.
  • The phase-rotation time embedding is generic, so a single trained model may be able to interpolate between target states (for example, $\alpha = 1$ and $\alpha = 2$ with one parameter set); the paper currently resorts to per-state hyperparameter tuning, leaving interpolation as a testable consequence of the shared-parameter design.
  • If the restoration trajectory is truly independent of the initial corruption level, the same trained model could serve as a channel-parameter estimator: the point where the backward trajectory stabilizes may reveal the unknown transmissivity and thermal photon number.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper proposes CVQD-G, a continuous-variable quantum diffusion model that uses a thermal-loss channel as the forward process and a CVQNN-based denoiser with time embedding as the backward process, together with a restoration variant CVQD-R. The forward process is reduced to a single effective thermal-loss step in Theorem 1, and the training objective combines state fidelity with a trace-normalization penalty. Numerical experiments report high-fidelity generation of coherent, squeezed, Fock, and cat states (typically >99%) and restoration of thermally corrupted coherent states with fidelities around 89%–98%. A complexity analysis claims training cost O(I×B×L), independent of the number of diffusion timesteps T.

Significance. If the central claims can be substantiated, this is a useful step toward continuous-variable quantum generative models with a physically motivated noise model. The paper has clear strengths: Theorem 1 is a clean, parameter-free derivation with a complete appendix proof; the restoration experiments in Section 6.3 are genuine out-of-sample tests on coherent states not singled out during that run's training; and the parameter-sharing time embedding is a sensible design that makes the complexity analysis meaningful. However, the generation results are evaluated in-sample, and the reported noise schedule does not establish that the generation start state lies in the training distribution. These issues currently prevent me from endorsing the paper's main generation claim, though they are fixable within the manuscript's scope.

major comments (3)
  1. [§3.2.2, §3.1, Table 4] The generation protocol in Algorithm 2 starts from the thermal state rho_th(nbar), but the forward schedule reported in Table 4 does not establish that any forward training state is close to rho_th. With the reported eta_0=0.99974 and eta_T=0.99331 over T=112, the cumulative transmissivity is eta_bar_T = product_{t=1}^{112} eta_t ≈ 0.68, so rho_T still retains roughly 68% of the original mode, and F(rho_0,rho_T) remains high (the paper itself notes in Section 6.2 that F(rho_0,rho_T) is 'generally non-zero'). Training samples are states E(rho_0, eta_bar_t) for t=1..T; if eta_bar_T≈0.68, none of these samples is near the thermal state used to initialize Algorithm 2, so the learned denoiser is asked to process an out-of-distribution input. The alternative reading of the beta entries (beta_start=1e-4, beta_end=0.05) would give eta_bar_T≈0.06 only if eta_t=1-beta_t, but no relation between beta and eta is defined anywhere in the text. The table therefore does not support the central claim that generation from rho_th is a valid instance of the trained backward process; the authors should either report a schedule that provably makes eta_bar_T small, define the beta-eta correspondence, or add generation experiments initialized from states that demonstrably occur in training.
  2. [§6.2, §3.4, Algorithm 1] The reported generation fidelities are in-sample. For each target rho_0 (coherent, squeezed, Fock, cat), CVQD-G is trained by Algorithm 1 on forward states E(rho_0, eta_bar_t) generated from that exact rho_0, and the final fidelity in Table 5 and Figures 4–8 is measured against the same rho_0. This evaluates the model's ability to invert a single known trajectory, not its ability to generate states outside the training set; the high >99% fidelities are therefore insufficient to support the general claim of 'state generation' as a generative-model capability. In contrast, the CVQD-R restoration experiments in Section 6.3 genuinely test on states not singled out during that run's training and should be presented as the primary out-of-sample evidence. I recommend reporting at least one holdout generation experiment (e.g., train on a set of states and evaluate on a disjoint set) and softening the generation claims accordingly.
  3. [§3.1, Algorithm 2, §6.2] The backward rollout is composed of T iterations of the learned single-step map f_theta, but no error propagation or fixed-point analysis is given. The single-step objective (Eq. 18) trains f_theta to map E(rho_0, eta_bar_t) to E(rho_0, eta_bar_{t-1}); at inference, however, the input to step t is never the forward state but the previous output of f_theta itself. The paper measures the composed trajectory in subplot (c) of Figures 4–8 and in Figure 10, which is useful evidence, but there is no bound or numerical study of how single-step errors accumulate over T≈112–150 iterations, or whether the iterated map has a stable fixed point near the target. This is especially important given the schedule issue above, because the composed rollout starts outside the training distribution. Without such an analysis, or an explicit statement that the claim is purely numerical for the tested cases, the general assertion that 'by sequentially applying the trained function ... the target quantum state is progressively recovered' (Section 3.1) is not justified.
minor comments (6)
  1. [§4.1, Eq. (21)] Equation (21) writes rho_out = E(rho_out, eta_ch) on the left and right; it should read rho_out = E(rho_in, eta_ch), since the right-hand side is meant to be the channel acting on the input state.
  2. [§6.2.2, Table 5] The text states that Optical GAN achieves 98.50% fidelity for the coherent state |alpha=1.0>, but Table 5 lists the corresponding value as 83.43%; these numbers should be reconciled.
  3. [§3.2.2, Table 4, Table 6] The relationship between beta_start/beta_end and eta_0/eta_T is never defined; if the beta entries are not used in the forward schedule of Eq. (13), they should be removed from the hyperparameter tables to avoid ambiguity.
  4. [§6.3, Figures 9–10] The sampling distributions for the amplitudes and phases of the training coherent states are not specified; please state the support and density (e.g., uniform in [0,1] for the X displacement and [0,2pi] for the rotation) so the out-of-sample claim is testable.
  5. [§6.2, Figures 4–8] Subplot (c) in Figures 4–8 does not label the horizontal axis; please clarify that it is the backward denoising step and specify what the varying eta values refer to for each starting state.
  6. [General] The paper does not include a data or code availability statement; providing the simulation code would greatly improve reproducibility, given that the numerical results are a central part of the contribution.

Circularity Check

1 steps flagged · score 6.0 of 10

Generation fidelities reduce to the training objective: CVQD-G is trained on a single predefined target and its reported generation performance is the same fidelity loss it was optimized to maximize.

  1. fitted input called prediction [Table 1, Section 3.4, Algorithm 1, and Section 6.2 / Table 5]
    "CVQD-G is trained by diffusing a single, predefined target state and initiates generation from a standard noise state. ... The fundamental goal of training CVQD-G is to optimize the denoising circuit parameters ϑ such that the predicted state ˜ρt−1 closely approximates the actual state ρt−1. This is achieved by maximizing the quantum fidelity between these states. ... L0 ← 1 − F(ρ0, ˜ρ0) + γP(˜ρ0)."

    The generation results in Section 6.2 (e.g., 99.95% fidelity for |α=1.0⟩) are reported as evidence that CVQD-G can generate the target state. But the target state is also the sole training state: the loss is L0 = 1 − F(ρ0, ˜ρ0) plus per-step terms 1 − F(ρ_{t−1}, ˜ρ_{t−1}) where every ρ_{t−1} is derived from the same ρ0. The final generation fidelity is therefore the very quantity the network was trained to maximize, so the reported number measures optimization success or memorization of the training target rather than an independent prediction. This is the fitted-input-called-prediction pattern: the target of the evaluation is the target of the fit.

full rationale

There is no equation-level circularity in the paper's forward-process derivation: Theorem 1 and Appendix A derive the accumulated-transmissivity formula ¯η_t = ∏ η_i by induction from the beam-splitter thermal-loss channel, and the result is not assumed as an input. The restoration experiments (Section 6.3) also carry independent content: CVQD-R is trained on a distribution of randomly sampled coherent states and then evaluated on states within that distribution, so the reported 89–98% fidelities are genuine interpolation measurements rather than re-statements of a single fitted target. The paper's self-citation [12] is used only to motivate the ancilla-based denoising design and is not load-bearing for the main derivation. The principal circular content is confined to the generation evaluation: CVQD-G is explicitly trained on a single predefined target state, and the loss directly contains 1 − F(ρ0, ˜ρ0); Section 6.2 then presents F(ρ0, generated state) for that same target as the model's generation performance. That is a fit-quality report, not a prediction. The separate concern that Table 4's β and η schedules may leave ρ_T far from the thermal state is an evidence gap or correctness issue, not a circularity, so it does not change this score.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the thermal-loss model of decoherence, the expressivity of CVQNNs with Kerr gates, and several hand-chosen hyperparameter groups (schedule endpoints, total steps, layer count, loss weights, Fock cutoff, and the unstated time-embedding amplitude). The forward-process composition (Theorem 1) is derived, not assumed. The schedule constants as printed are mutually inconsistent, which directly affects the load-bearing assumption that the terminal diffusion state is close to the thermal state. No new physical entities are introduced.

free parameters (6)
  • Noise schedule endpoints (eta_0, eta_T or beta_start, beta_end) = eta_0=0.99974, eta_T=0.99331; beta_start=1e-4, beta_end=0.05 (listed jointly in Table 4, mutually inconsistent)
    Chosen by hand (with Optuna) to control how far rho_T moves toward the thermal state. The terminal-state proximity is load-bearing for generation from the thermal state, and the two reported value sets give very different diffusion strengths.
  • Total diffusion timesteps T = 112 (generation), 150 (restoration), 117 (tuned alpha=2.5 case)
    Chosen hyperparameter; determines the number of sequential denoising applications and the cumulative effect of the schedule.
  • Time-embedding displacement alpha = not stated in the paper
    The time embedding circuit uses a displacement gate D(alpha) with fixed parameter alpha in R+ but no value is given in Tables 4, 6, or 7; the model's ability to distinguish timesteps depends on it.
  • Loss weights lambda and gamma = lambda=8.55e-5 (generation), 0.16 (restoration); gamma=100 (10.358 for tuned alpha=2.5)
    Tuned hyperparameters balancing the fidelity objective against the trace-normalization penalty.
  • CVQNN layers per step L = 30
    Circuit depth per denoising step, chosen by hand; the expressivity of the learned inverse depends on it.
  • Fock cutoff dimension = 15
    Truncation of the infinite Hilbert space. The trace penalty P is introduced specifically to compensate for truncation-induced norm loss, and large-amplitude states are limited by this truncation.
assumptions (4)
  • domain assumption Each diffusion step couples the system to an independent environment mode in the same thermal state with mean photon number n_bar, and the beam-splitter transformation (Eq. 8) describes the interaction.
    Standard quantum optics model of loss (Eqs. 9-10), used for both the forward process and the restoration training noise. The composition property of Theorem 1 relies on all environment modes sharing mean photon number n_bar.
  • domain assumption CVQNN layers combining Gaussian gates with a Kerr gate can represent the inverse of the thermal-loss channel well enough within the Fock cutoff.
    Relies on Lloyd-Braunstein universality [13, 14]. The approximation quality at L=30 layers and cutoff 15 is not proven, only demonstrated for the tested states.
  • domain assumption Fock truncation at cutoff 15, patched by the trace penalty (Eq. 20), yields faithful fidelities.
    The paper acknowledges truncation pushes states outside the computational basis (Section 3.4). The penalty keeps traces near 1, but the residual truncation error is not quantified beyond the observed fidelity losses for large alpha.
  • ad hoc to paper The linear noise schedule (Eq. 13) with the chosen endpoints drives rho_T close to the thermal state.
    The schedule form is chosen for convenience. The claim that eta_bar_T becomes small is asserted in Section 3.2.2, but the two reported constant sets in Table 4 disagree on how small.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Continuous-variable Quantum Diffusion Model for State Generation and Restoration." pith.science (2026). https://pith.science/paper/KLPBKMMY

@misc{pith2026250619270,
  author       = {Pith},
  title        = {Pith review of: Continuous-variable Quantum Diffusion Model for State Generation and Restoration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KLPBKMMY}},
  note         = {Machine review of arXiv:2506.19270}
}
read the original abstract

The generation and preservation of complex quantum states against environmental noise are paramount challenges in advancing continuous-variable (CV) quantum information processing. This paper introduces a novel framework based on continuous-variable quantum diffusion principles, synergizing them with CV quantum neural networks (CVQNNs) to address these dual challenges. For the task of state generation, our Continuous-Variable Quantum Diffusion Generative model (CVQD-G) employs a physically driven forward diffusion process using a thermal loss channel, which is then inverted by a learnable, parameter-efficient backward denoising process based on a CVQNN with time-embedding. This framework's capability is further extended for state recovery by the Continuous-Variable Quantum Diffusion Restoration model (CVQD-R), a specialized variant designed to restore quantum states, particularly coherent states with unknown parameters, from thermal degradation. Extensive numerical simulations validate these dual capabilities, demonstrating the high-fidelity generation of diverse Gaussian (coherent, squeezed) and non-Gaussian (Fock, cat) states, typically with fidelities exceeding 99%, and confirming the model's ability to robustly restore corrupted states. Furthermore, a comprehensive complexity analysis reveals favorable training and inference costs, highlighting the framework's efficiency, scalability, and its potential as a robust tool for quantum state engineering and noise mitigation in realistic CV quantum systems.

Figures

Figures reproduced from arXiv: 2506.19270 by the authors.

Figure 1
Figure 1. The Continuous-Variable (CV) Quantum Diffusion [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The framework of the proposed Continuous-Variable Quantum Diffusion Generative Model (CVQD-G). [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Schematic of the CVQD-G denoising framework. The function [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: Fidelity of generated states as a function of key state [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Generation of the squeezed vacuum state S(r = 0.5)|0⟩ in a pure loss channel (n¯ = 0). (a) Training loss. (b) Forward diffusion fidelity. (c) Backward denoising fidelity from initial states with varying noise levels (η). 6.2.2 Non-Gaussian State Generation Non-Gaussian…
Figure 7
Figure 7. Figure 7: Generation of the Fock state |1⟩ in a pure loss channel (n¯ = 0). (a) Training loss. (b) Forward diffusion fidelity. (c) Backward denoising fidelity from initial states with varying noise levels (η). Cat States (|cat(α)⟩): Cat states, superpositions of dis￾tinct cohere…
Figure 8
Figure 8. Figure 8: Generation of the even cat state |cat(1)⟩ in a pure loss channel (n¯ = 0). (a) Training loss. (b) Forward diffusion fidelity. (c) Backward denoising fidelity from initial states with varying noise levels (η). Comparative Performance with Optical GANs (¯n = 0): For a di…
Figure 9
Figure 9. Figure 9: Average restoration fidelity versus restoration [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: Restoration fidelity versus restoration timestep for [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]
Figure 11
Figure 11. Figure 11: shows the generation results for coherent states |α⟩ with amplitudes α ∈ {0.5, 1.0, 1.5, 2.0, 2.5}. For α = 2.5, the fidelity achieved with general hyperparameters is affected by the fixed Fock cutoff dimension, as discussed in Section 6.2. For results with tuned hype…
Figure 12
Figure 12. Figure 12: presents results for squeezed vacuum states S(r)|0⟩ with squeezing parameters r ∈ {0.25, 0.5, 0.75, 1.0} using general hyperparameters. A slight fidelity decrease is observed for larger r values [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Summary of training loss and fidelities. Top row: [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: shows the training loss, forward diffusion fidelity, and backward denoising fidelity for the coherent state |α = 2.5⟩ generated using these tuned hyperparameters, achieving a final fidelity exceeding 97% [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MonoPartNeRF:Human Reconstruction from Monocular Video via Part-Based Neural Radiance Fields

    cs.CV 2025-08 conditional novelty 6.0 of 10

    Randomized weak measurements provide a physically motivated forward diffusion for quantum states, with reverse recovery via learned unitary controls, Petz maps, or classical shadows.

  2. Quantum Reversibility Meets Classical Reverse Diffusion

    quant-ph 2025-10 conditional novelty 4.0 of 10

    The semiclassical limit of the Petz-reversed Lindblad equation reproduces the Bayes-rule reverse-time diffusion equation, with the reference state's Wigner function playing the role of the classical score distribution.

Reference graph

Works this paper leans on

38 extracted references · 29 canonical work pages · cited by 2 Pith papers

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P . Abbeel, “Denoising diffusion probabilistic models,”Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

  2. [2]

    Quantum channels and their entropic characteristics,

    A. S. Holevo and V . Giovannetti, “Quantum channels and their entropic characteristics,”Reports on Progress in Physics, vol. 75, p. 046001, mar 2012

  3. [3]

    Quantum information with continuous variables,

    S. L. Braunstein and P . van Loock, “Quantum information with continuous variables,”Rev. Mod. Phys., vol. 77, pp. 513–577, Jun 2005

  4. [4]

    Continuous-variable quantum neural networks,

    N. Killoran, T. R. Bromley, J. M. Arrazola, M. Schuld, N. Quesada, and S. Lloyd, “Continuous-variable quantum neural networks,” Phys. Rev. Res., vol. 1, p. 033063, Oct 2019

  5. [5]

    OpticalGAN : Generative Adversarial Networks for Continuous Variable Quantum Computation

    N. Shrivastava, N. Puri, P . Gupta, B. Krishnamurthy, and S. Verma, “Opticalgan : Generative adversarial networks for continuous variable quantum computation,”CoRR, vol. abs/1909.07806, 2019

  6. [6]

    A continuous variable born machine,

    I. ˇCepait˙e, B. Coyle, and E. Kashefi, “A continuous variable born machine,”Quantum Machine Intelligence, vol. 4, p. 6, Mar 2022

  7. [7]

    Continuous-variable quantum boltzmann machine,

    S. Bangar, L. Sunny, K. Yeter-Aydeniz, and G. Siopsis, “Continuous-variable quantum boltzmann machine,”arXiv preprint arXiv:2405.06580, 2024

  8. [8]

    On the quantum correction for thermodynamic equi- librium,

    E. Wigner, “On the quantum correction for thermodynamic equi- librium,”Physical review, vol. 40, no. 5, p. 749, 1932

Show all 38 references
  1. [9]

    Quantum in- formation with gaussian states,

    X. Wang, T. Hiroshima, A. Tomita, and M. Hayashi, “Quantum in- formation with gaussian states,”Physics Reports, vol. 448, p. 1–111, Aug. 2007

  2. [10]

    Quantum communication,

    N. Gisin and R. Thew, “Quantum communication,”Nature Photon- ics, vol. 1, pp. 165–171, Mar 2007

  3. [11]

    S. R. De Groot and P . Mazur,Non-equilibrium thermodynamics. Courier Corporation, 2013

  4. [12]

    Quantum generative diffusion model: a fully quantum-mechanical model for generating quantum state ensemble,

    C. Chen, Q. Zhao, M. Zhou, Z. He, Z. Sun, and H. Situ, “Quantum generative diffusion model: a fully quantum-mechanical model for generating quantum state ensemble,”arXiv:2401.07039, 2024

  5. [13]

    Quantum computation over contin- uous variables,

    S. Lloyd and S. L. Braunstein, “Quantum computation over contin- uous variables,”Phys. Rev. Lett., vol. 82, pp. 1784–1787, Feb 1999

  6. [14]

    Machine learning method for state preparation and gate synthesis on photonic quantum computers,

    J. M. Arrazola, T. R. Bromley, J. Izaac, C. R. Myers, K. Br ´adler, and N. Killoran, “Machine learning method for state preparation and gate synthesis on photonic quantum computers,”Quantum Science and Technology, vol. 4, p. 024004, jan 2019

  7. [15]

    General state changes in quantum theory,

    K. Kraus, “General state changes in quantum theory,”Annals of Physics, vol. 64, no. 2, pp. 311–335, 1971

  8. [16]

    Strawberry Fields: A Software Platform for Pho- tonic Quantum Computing,

    N. Killoran, J. Izaac, N. Quesada, V . Bergholm, M. Amy, and C. Weedbrook, “Strawberry Fields: A Software Platform for Pho- tonic Quantum Computing,”Quantum, vol. 3, p. 129, Mar. 2019

  9. [17]

    Applications of near-term photonic quantum computers: software and algorithms,

    T. R. Bromley, J. M. Arrazola, S. Jahangiri, J. Izaac, N. Quesada, A. D. Gran, M. Schuld, J. Swinarton, Z. Zabaneh, and N. Kil- loran, “Applications of near-term photonic quantum computers: software and algorithms,”Quantum Science and Technology, vol. 5, p. 034010, May 2020

  10. [18]

    Tensorflow: a system for large-scale machine learning,

    M. Abadi, P . Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, G. Irving, M. Isard,et al., “Tensorflow: a system for large-scale machine learning,” in12th USENIX symposium on operating systems design and implementation (OSDI 16), pp. 265–283, 2016

  11. [19]

    Adam: A method for stochastic optimization,

    D. P . Kingma, “Adam: A method for stochastic optimization,” arXiv:1412.6980, 2014

  12. [20]

    Sgdr: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “Sgdr: Stochastic gradient descent with warm restarts,”arXiv:1608.03983, 2016

  13. [22]

    Satellite-to-ground quantum key distribution,

    S.-K. Liao, W.-Q. Cai, W.-Y. Liu, L. Zhang, Y. Li, J.-G. Ren, J. Yin, Q. Shen, Y. Cao, Z.-P . Li, F.-Z. Li, X.-W. Chen, L.-H. Sun, J.-J. Jia, J.-C. Wu, X.-J. Jiang, J.-F. Wang, Y.-M. Huang, Q. Wang, Y.-L. Zhou, L. Deng, T. Xi, L. Ma, T. Hu, Q. Zhang, Y.-A. Chen, N.-L. Liu, X.-...

  14. [23]

    Quantum computers,

    T. D. Ladd, F. Jelezko, R. Laflamme, Y. Nakamura, C. Monroe, and J. L. O’Brien, “Quantum computers,”Nature, vol. 464, no. 7285, pp. 45–53, 2010

  15. [24]

    Quantum computational advantage using photons,

    H.-S. Zhong, H. Wang, Y.-H. Deng, M.-C. Chen, L.-C. Peng, Y.-H. Luo, J. Qin, D. Wu, X. Ding, Y. Hu, P . Hu, X.-Y. Yang, W.-J. Zhang, H. Li, Y. Li, X. Jiang, L. Gan, G. Yang, L. You, Z. Wang, L. Li, N.-L. Liu, C.-Y. Lu, and J.-W. Pan, “Quantum computational advantage using phot...

  16. [25]

    Quantum sensing,

    C. L. Degen, F. Reinhard, and P . Cappellaro, “Quantum sensing,” Rev. Mod. Phys., vol. 89, p. 035002, Jul 2017

  17. [26]

    M. O. Scully and M. S. Zubairy,Quantum optics. Cambridge university press, 1997

  18. [27]

    Deep unsupervised learning using nonequilibrium thermody- namics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermody- namics,” inInternational conference on machine learning, pp. 2256– 2265, PMLR, 2015

  19. [28]

    Advances in quantum metrology,

    V . Giovannetti, S. Lloyd, and L. Maccone, “Advances in quantum metrology,”Nature photonics, vol. 5, no. 4, pp. 222–229, 2011

  20. [29]

    M. A. Nielsen and I. L. Chuang,Quantum computation and quantum information. Cambridge university press, 2010

  21. [30]

    Quantum computation,

    D. P . DiVincenzo, “Quantum computation,”Science, vol. 270, no. 5234, pp. 255–261, 1995

  22. [31]

    Stimulating diffusion model for image denoising via adaptive embedding and ensembling,

    T. Li, H. Feng, L. Wang, L. Zhu, Z. Xiong, and H. Huang, “Stimulating diffusion model for image denoising via adaptive embedding and ensembling,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  23. [32]

    Restoring vision in adverse weather conditions with patch-based denoising diffusion mod- els,

    O. ¨Ozdenizci and R. Legenstein, “Restoring vision in adverse weather conditions with patch-based denoising diffusion mod- els,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 8, pp. 10346–10357, 2023

  24. [33]

    Diffusion models for imperceptible and transferable adversarial attack,

    J. Chen, H. Chen, K. Chen, Y. Zhang, Z. Zou, and Z. Shi, “Diffusion models for imperceptible and transferable adversarial attack,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  25. [34]

    An approach to cryptography based on continuous-variable quantum neural network,

    J. Shi, S. Chen, Y. Lu, Y. Feng, R. Shi, Y. Yang, and J. Li, “An approach to cryptography based on continuous-variable quantum neural network,”Scientific reports, vol. 10, no. 1, p. 2107, 2020

  26. [35]

    Generative quantum machine learning via denoising diffusion probabilistic models,

    B. Zhang, P . Xu, X. Chen, and Q. Zhuang, “Generative quantum machine learning via denoising diffusion probabilistic models,” Physical Review Letters, vol. 132, no. 10, p. 100602, 2024

  27. [36]

    Quantum diffusion models,

    A. Cacioppo, L. Colantonio, S. Bordoni, and S. Giagu, “Quantum diffusion models,”arXiv preprint arXiv:2311.15444, 2023

  28. [37]

    Optuna: A next-generation hyperparameter optimization framework,

    T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama, “Optuna: A next-generation hyperparameter optimization framework,” in Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp. 2623–2631, 2019

  29. [38]

    Diffusion models in vision: A survey,

    F.-A. Croitoru, V . Hondru, R. T. Ionescu, and M. Shah, “Diffusion models in vision: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 9, pp. 10850–10869, 2023

  30. [39]

    Motiondiffuse: Text-driven human motion generation with dif- fusion model,

    M. Zhang, Z. Cai, L. Pan, F. Hong, X. Guo, L. Yang, and Z. Liu, “Motiondiffuse: Text-driven human motion generation with dif- fusion model,”IEEE transactions on pattern analysis and machine intelligence, vol. 46, no. 6, pp. 4115–4128, 2024. 16 APPENDIXA PROOF OFTHEOREM1 We pro...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.