{"id":"90a49103-0371-4f15-ae63-d45414e75271","arxiv_id":"2506.19270","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A thermal-loss diffusion process on continuous-variable quantum states, inverted by a time-embedded CV quantum neural network, generates target states and restores corrupted coherent states at high simulated fidelity.","lead":"The paper trains a continuous-variable quantum neural network to reverse a thermal-noise diffusion process on optical modes, using it to generate quantum states from thermal noise and to restore coherent states corrupted by loss. It reports generation fidelities above 99 percent and restoration fidelities of 89 to 98 percent in single-mode simulations, a step toward noise-robust CV quantum communication.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported schedule does not establish that the generation start state rho_th was ever seen in training; Table 4 implies eta_bar_T ≈ 0.68, leaving the central Algorithm 2 claim unsupported.","rationale":"The paper's core construction is coherent: Theorem 1 is a standard multistep thermal-loss composition result, the shared-parameter denoiser with time embedding is a plausible ansatz, and the reported fidelities, while not independently reproducible without code, are not implausible. I read the numerical claims as empirical demonstrations rather than proved guarantees. The single most load-bearing weakness is the mismatch between the training distribution of the forward process and the initialization of Algorithm 2. The paper's own Table 4 cannot support the claim that rho_T ≈ rho_th: the eta schedule implies eta_bar_T ≈ 0.68, while the beta schedule implies a much stronger diffusion, and no relation between beta and eta is given. This is not merely a typographical annoyance: Algorithm 2 starts from rho_th, and Algorithm 1 never trains on rho_th unless eta_bar_T ≈ 0. The claim that 'by sequentially applying f_theta T times to rho_T the target state is recovered' is therefore either an extrapolation or is based on an unreported schedule. The reader's weakest_assumption identifies exactly this issue, and the proposed test (compute eta_bar_T and compare generation from rho_th vs. rho_T) would settle it. I do not see an internal mathematical error in the forward-process derivation, and the restoration protocol for CVQD-R is less exposed to this concern because it starts from corrupted coherent states that are in the training distribution. The verdict should remain conditional pending the schedule clarification and artifacts.","tokens_in":22724,"tokens_out":6458,"duration_ms":70583,"concrete_test":"Recompute the terminal retained-signal parameter from the schedule actually used: either record eta_bar_T = product_{t=1}^{112} eta_t from Table 4 (giving ≈0.68) or recover it from Figure 4(b)'s terminal fidelity for |alpha=1>, nbar=0, where rho_T is coherent with amplitude sqrt(eta_bar_T), so F = exp(-(1-sqrt(eta_bar_T))^2). Then run Algorithm 2 twice, once initialized from rho_th as written and once from rho_T = E(rho_0, eta_bar_T). If the two final fidelities differ by more than the run-to-run spread, or if eta_bar_T is not small, the generation protocol starts out-of-distribution and the central claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central generation claim rests on Algorithm 2 starting from rho_th(nbar) and iterating the learned denoiser back to rho_0. For this to work, the training distribution must cover that starting point: the forward states seen in training, E(rho_0, eta_bar_t), must include a state near rho_th. That requires eta_bar_T = prod eta_t to be small. Table 4 does not establish this. Its eta schedule (eta_0=0.99974 to eta_T=0.99331 over T=112) gives eta_bar_T ≈ 0.68, i.e. rho_T still contains roughly 68% of the original mode. For the coherent-state example, this gives terminal fidelity F(|1>, rho_T) ≈ exp(-(1-sqrt(0.68))^2) ≈ 0.97, contradicting the paper's own description of rho_T as a 'thermal-like' state. The beta entries (beta_start=1e-4, beta_end=0.05) would give a much smaller retained fraction (~0.06 if eta_t=1-beta_t), but the paper never defines the relation between beta and eta. So the reported numbers do not pin down whether the generation start state was ever seen in training. If the schedule is the weak one implied by the eta column, the model is asked to denoise a state it was not trained on; if the schedule is the strong one, Table 4 is misreported. Either way, the paper's own data do not support the claimed generation protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":22897,"tokens_out":8448,"duration_ms":82787,"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":[{"comment":"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.","section":"§3.2.2, §3.1, Table 4"},{"comment":"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.","section":"§6.2, §3.4, Algorithm 1"},{"comment":"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.","section":"§3.1, Algorithm 2, §6.2"}],"minor_comments":[{"comment":"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.","section":"§4.1, Eq. (21)"},{"comment":"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.","section":"§6.2.2, Table 5"},{"comment":"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.","section":"§3.2.2, Table 4, Table 6"},{"comment":"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.","section":"§6.3, Figures 9–10"},{"comment":"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.","section":"§6.2, Figures 4–8"},{"comment":"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.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main issue is not the correctness of Theorem 1 but the mismatch between the reported noise schedule and the generation protocol, compounded by the in-sample nature of the generation evaluation. The restoration experiments provide independent evidence that the CVQNN denoiser learns something useful, so I would not recommend rejection. Please encourage the authors to fix the schedule definitions, add an out-of-sample generation test or substantially soften the generation claims, and address the error accumulation concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is worth a serious look but not for the reason the abstract sells. The genuinely useful piece is the restoration variant, CVQD-R; the generation story has a schedule problem that needs fixing before its headline claim stands.\n\nWhat is new: the forward process uses a thermal-loss channel, and Theorem 1 gives a closed-form multistep transition that is correct and, as the appendix shows, just beam-splitter composition with an effective thermal mode. The shared-parameter time-embedded CVQNN is a reasonable way to avoid training T separate denoisers. The restoration experiments—training on an ensemble of coherent states and then iteratively denoising unknown corrupted states—are the real contribution. The reported 89–98% restoration fidelities for unknown amplitudes and noise levels are plausible for a Fock cutoff of 15 and are measurements on states not singled out in that run's training.\n\nThe soft spot is the noise schedule. Table 4 lists eta_0=0.99974 and eta_T=0.99331 over T=112. That product is about 0.68, meaning the terminal forward state retains 68% of the original mode—not close to the thermal state the text claims. The beta schedule (1e-4 to 0.05) would imply a much smaller residual, but the paper never connects beta to eta. So the numbers don't pin down whether Algorithm 2's starting state rho_th(nbar) was ever seen in training. If the eta column is right, generation starts outside the training distribution; if the beta column is right, Table 4 is misreported. The paper's own fidelity curves in subplot (c) test robustness from forward-corrupted states, not from the thermal state, so they don't resolve this. This is fixable: report the cumulative transmissivity explicitly, verify that the generation start state is in the training distribution, or start from the actual terminal state.\n\nOther issues are minor by comparison: the Optical GAN baseline is 98.50% in the prose and 83.43% in Table 5—reconcile that. No code, data, seeds, or error bars, and the time-embedding displacement alpha is never quantified. The abstract's framing of per-target-state preparation as 'generation' is stronger than the paper's own future-work statement, which concedes dedicated hyperparameters per target state.\n\nOverall: the central math is sound, the restoration idea is useful, and the generation protocol can likely be repaired. I'd send this to a referee—not desk-reject—and ask the authors to fix the schedule issue, clean up the baseline numbers, and release artifacts. The kind of reader who benefits is someone working on CV quantum state engineering or quantum communication receivers who wants a concrete denoising recipe. I wouldn't cite it as-is for the generation claim, but I'd cite the restoration approach once the schedule is sorted.","headline":"Restoration variant is the real contribution; generation claim rests on an unverified noise schedule that doesn't obviously make the terminal state thermal.","tokens_in":23594,"tokens_out":4108,"would_cite":false,"duration_ms":40794,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.-a"],"model":"deepseek-v4-flash","headline":"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.","keywords":["Continuous-Variable Quantum Information","Quantum Diffusion Models","Quantum State Generation","Quantum State Restoration","Continuous-Variable Quantum Neural Networks","Thermal Loss Channel"],"falsifier":"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.","tokens_in":22319,"feed_emoji":"⚛️","tokens_out":27906,"duration_ms":222546,"temperature":0.7,"pith_summary":"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.","feed_headline":"One circuit generates CV quantum states at 99%+ and repairs them","feed_subtitle":"Diffusion training teaches one quantum circuit to undo thermal noise step by step.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model paradigm — forward noise process, backward denoiser, time-conditioned shared network — that motivates the entire construction.","marker":"[1]"},{"why":"Provides the continuous-variable quantum information foundations, including qumode representations and the coherent-state variance used in the restoration analysis.","marker":"[3]"},{"why":"Defines the CVQNN circuit architecture (Gaussian gates plus Kerr nonlinearity) from which the denoising circuit is built, and the Fock-backend simulation approach.","marker":"[4]"},{"why":"The Optical GAN baseline that CVQD-G is compared against for generation fidelity (Table 5) and for training and inference complexity.","marker":"[5]"},{"why":"Supplies the Gaussian-state formalism and thermal-state variance relation used to derive the output-variance equation that justifies the restoration strategy.","marker":"[9]"},{"why":"The discrete-variable quantum diffusion framework that this work extends to the continuous-variable setting; also cited for the failure mode where a denoiser simply copies its input.","marker":"[12]"},{"why":"The photonic quantum simulation platform whose Fock backend runs all numerical experiments.","marker":"[16]"},{"why":"The prior quantum denoising diffusion probabilistic model in the discrete-variable or hybrid setting that motivates a fully continuous-variable formulation.","marker":"[35]"}],"fun_headline_variants":["One diffusion circuit generates CV states at 99%+ and repairs them","CV quantum diffusion creates quantum states and reverses thermal damage","CV diffusion model: 99% state generation plus thermal-noise restoration","Thermal-loss diffusion: one circuit generates and repairs quantum states"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["One diffusion circuit generates CV states at 99%+ and repairs them","CV quantum diffusion creates quantum states and reverses thermal damage","CV diffusion model: 99% state generation plus thermal-noise restoration","Thermal-loss diffusion: one circuit generates and repairs quantum states"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001654,"raw_usage":{"total_tokens":6599,"prompt_tokens":1005,"completion_tokens":5594,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":5522}},"tokens_in":621,"tokens_out":5594,"duration_ms":39900,"temperature":1.0,"reasoning_tokens":5522,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:36:57.458852+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"Quantum information with continuous variables,","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-variable quantum information foundations, including qumode representations and the coherent-state variance used in the restoration analysis."},{"cited_title":"Continuous-variable quantum neural networks,","cited_arxiv_id":null,"evidence_quote":"Defines the CVQNN circuit architecture (Gaussian gates plus Kerr nonlinearity) from which the denoising circuit is built, and the Fock-backend simulation approach."},{"cited_title":"OpticalGAN : Generative Adversarial Networks for Continuous Variable Quantum Computation","cited_arxiv_id":"1909.07806","evidence_quote":"The Optical GAN baseline that CVQD-G is compared against for generation fidelity (Table 5) and for training and inference complexity."},{"cited_title":"Quantum in- formation with gaussian states,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-state formalism and thermal-state variance relation used to derive the output-variance equation that justifies the restoration strategy."},{"cited_title":"Strawberry Fields: A Software Platform for Pho- tonic Quantum Computing,","cited_arxiv_id":null,"evidence_quote":"The photonic quantum simulation platform whose Fock backend runs all numerical experiments."}],"review_version":2}