{"id":"56884715-2610-4661-8c69-05f5149597d2","arxiv_id":"2508.10900","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"QVF is a quantum implicit neural representation that uses a learnable, Boltzmann-style amplitude encoding and a real-valued entangled circuit, and it outperforms prior quantum INRs while competing with classical baselines on simulator experiments.","lead":"This paper introduces QVF, a quantum-circuit-based model that encodes images and 3D shapes into the amplitudes of a quantum state and reads them back with direct measurements. It reports that QVF beats the previous quantum method QIREN and matches classical neural representation models on small test sets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unproven full-expressivity claim for the real-valued Ry+CNOT ansatz (Sec. 4.2) is the load-bearing assumption: Fig. 5 and Appendix D.3 do not establish that expressivity and gradient variance survive at larger qubit counts.","rationale":"Reader's verdict CONDITIONAL; our analysis agrees with reader's weakest assumption. The central claim is empirical and modest, but it is explicitly framed as a step toward fault-tolerant hardware, and the only theoretical guarantee offered for the architecture is the unproven 'full expressivity' claim. We located the gap in Sec. 4.2 (assertion) and Appendix D.3 (Lemma 1 not covering the non-unitary amplitude encoding). No internal contradiction in the reported numbers was found; the concern is about extrapolation beyond n=5. The proposed gradient-variance scaling test would settle whether the bounded-Haar-randomness argument lands. Because this is the same concern already reflected in the conditional verdict, we do not move the verdict; UNCHANGED is appropriate.","tokens_in":20885,"tokens_out":7759,"duration_ms":94313,"concrete_test":"Re-run the Sec. 5.1 gradient-variance protocol for n = 8, 10, 12 (scaling depth J with n, e.g., J = n, and using at least 5,000 parameter samples) and fit Var_grad as a function of n. If the variance decays like O(2^{-n}) or approaches the QIREN baseline, the 'bounded Haar randomness' claim is unsupported. As an expressivity check, train QVF with n=8, J=8 on the same CIFAR-10 image and compare PSNR to the classical baseline in Table 3; if QVF no longer outperforms MLP+Siren, then the ansatz's expressivity/trainability is not preserved at larger scales.","verdict_should_be":"UNCHANGED","load_bearing_attack":"QVF's central claim—matching classical INRs without classical post-processing—holds only if the real-valued ansatz remains both expressive and trainable as the problem grows. Section 4.2 asserts that restricting to Pauli-Y rotations and CNOT gates 'maintains full expressivity while significantly simplifying the optimisation landscape', but this is not proven. The two pieces of evidence offered are not sufficient. First, Appendix D.3's Lemma 1 is a statement about the Fourier spectrum of an energy-based encoding, and it invokes Schuld et al.'s unitary-encoding framework; QVF's encoding is amplitude preparation with non-negative real amplitudes, so that Fourier machinery does not directly apply. Second, Fig. 5 reports gradient variance for the ansatz, but only on a small-qubit simulator range and without an asymptotic fit; the real orthogonal group SO(2^n) generated by Ry and CNOT still has exponential dimension, and known Lie-algebraic barren-plateau results tie variance decay to the dynamical Lie algebra, so halving the unitary group by dropping complex phases gives no general guarantee of polynomial gradients. If the expressivity or trainability of the J=5, n=5 configuration fails to persist at n=8 or higher, the reported gains over QIREN and classical baselines may be a small-system artefact rather than a scalable QINR property. The authors themselves only claim simulator-scale results and future fault-tolerant hardware, so this is the exact point where the central claim is least secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Quantum Visual Fields (QVF), a quantum implicit neural representation for 2D images and 3D signed distance fields. QVF encodes query coordinates into quantum statevector amplitudes through a learned energy-based Boltzmann distribution, processes them with a fully entangled parameterized circuit built from Pauli-Y rotations and CNOT gates, and extracts the represented field directly from local Pauli-Z measurements without a classical decoder. The method is evaluated on a noiseless PennyLane simulator with n=5 qubits and J=5 layers, compared against the QINR method QIREN on equal qubit budgets and against classical INR baselines (MLP+ReLU, DeepSDF, SIREN) on image and shape collections. The authors report that QVF outperforms QIREN and is competitive with or better than the classical baselines in MSE, PSNR, and 3D MAE, and they demonstrate image inpainting, shape completion, and latent-space interpolation as applications.","tokens_in":21159,"tokens_out":8479,"duration_ms":101812,"significance":"If the reported results hold, QVF is a meaningful advance for QINRs: it is the first QINR in this line to support collections of images and 3D SDFs, it avoids the classical decoder used by QIREN, and it includes an ablation of the ansatz depth, width, and latent dimension. The empirical comparisons are run with multiple repetitions and reported with means and standard deviations, and the equal-qubit-budget comparison with QIREN is a fair and useful baseline. The significance is conditional, however, on two issues: the theoretical claim that the real-valued ansatz retains full expressivity and good gradient behavior at larger scales is not established, and key experimental hyperparameters (notably beta and the measurement-shot count) are not reported. These do not invalidate the n=5 simulator findings, but they do limit the paper's broader claim that QVF is a scalable QINR architecture.","major_comments":[{"comment":"The claim in Section 4.2 that restricting the ansatz to real-valued Pauli-Y rotations and CNOT gates 'maintains full expressivity while significantly simplifying the optimisation landscape' is load-bearing for the scalability of QVF, but the evidence provided is insufficient. Appendix D.3's Lemma 1 invokes the Fourier framework of Schuld et al. for circuits of the form U(x)=W2 g(x) W1, whereas QVF's input state is prepared by amplitude encoding with non-negative real amplitudes derived from a Boltzmann distribution; the lemma does not directly transfer to this setting. Figure 5 reports gradient variance over a small range of qubit counts and without an asymptotic fit, so it does not establish that the real-valued restriction avoids barren plateaus at larger n. Please either provide a proof of the expressivity and variance-scaling claims, or explicitly restrict the claims to the tested n=5 regime.","section":"Section 4.2; Appendix D.3"},{"comment":"Two central experimental hyperparameters are never reported. The inverse temperature beta governing the Boltzmann encoding in Eq. (2) is listed as an input in Algorithm 1, but no value is given in the main text or appendix. Similarly, the number of measurement shots used for the quantitative results in Tables 2--4 is not stated; shot counts (100 to 10^4) appear only in Appendix E for the qualitative figure. If the simulator uses analytic expectation values rather than finite-shot sampling, that should be stated explicitly; otherwise the finite-shot statistics should be reported. Without these values the central empirical comparisons are not reproducible, and the effect of measurement noise on the reported differences cannot be assessed.","section":"Section 5.2 and Algorithm 1"},{"comment":"The paper's headline distinction over QIREN is that QVF 'does not rely on classical post-processing'. This is accurate for the readout stage, since circuit outputs are obtained by direct measurements. However, the application protocol in Eq. (11) performs a classical MAP optimization of a latent code z after training, and the energy encoder is a classical MLP before the circuit. The scope of the no-post-processing claim should be stated precisely, because as written it overstates the purity of the quantum pipeline and invites a misleading comparison with QIREN's classical decoder.","section":"Section 4.3, Eq. (11)"}],"minor_comments":[{"comment":"Equation (3) is typeset incorrectly: the integral sign and the differential dTheta are misplaced, and the intended expression Z = integral exp(-beta E(Theta,z)) dTheta should be written clearly.","section":"Eq. (3)"},{"comment":"The summation notation in Eq. (10), written as sum_{i,j=1}^{W,M}, is unconventional; please use explicit double sums or a standard multi-index notation.","section":"Eq. (10)"},{"comment":"Table 3 labels one baseline as 'DeepSDF [41]', while the text describes it as an MLP with ReLU activations; please clarify the exact architecture, latent dimension, and parameter count used for this baseline so that the comparison is unambiguous.","section":"Table 3"},{"comment":"The statement that QVF 'trains faster than all compared methods' is not supported by wall-clock or per-epoch runtime measurements; Figure 6 shows learning curves, which measure convergence behaviour, not training speed.","section":"Section 6"}],"recommendation":"major_revision","confidential_remarks":"This is a reasonable methods paper for quantum implicit neural representations, and simulator-only evaluation is standard in this subfield. The main risk is overclaiming scalability: the expressivity and trainability of the real-valued ansatz are asserted more strongly than the evidence supports. The missing beta and shot-count reporting are easily fixable, but they are currently a reproducibility barrier. I do not see a novelty or attribution problem; the comparisons to QIREN and 3D-QAE are appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a real step forward in QINRs: the learnable energy-based amplitude encoding is a genuine break from hand-crafted encodings, the real-valued ansatz with direct measurements removes classical post-processing, and they are the first to push this to 3D fields and collections. The experiments are fair—same qubit budget as QIREN, honest classical baselines—and the ablations on depth, width, and latent size are useful. Appendix analyses of shot noise and circuit noise are a credit; most QINR papers skip that.\n\nThe main soft spot is exactly what the stress-test flags: the claim that the Ry+CNOT ansatz 'maintains full expressivity' while taming barren plateaus is asserted but not proven. Appendix D.3's Lemma 1 is a Fourier argument for unitary encodings; the amplitude encoding here is a non-negative real state prep, so that machinery does not obviously apply. Fig. 5's gradient variance only spans a small qubit range, so the advantage at scale is unknown. This is not fatal for a simulator-era paper, but the sentence in Sec. 4.2 should be downgraded from 'maintains full expressivity' to 'preserves expressivity on the tested configurations' unless they can prove otherwise.\n\nMinor but annoying: beta in the Boltzmann encoding is never reported (Algorithm 1 lists it, but no experiment says its value), and the main text never specifies how many measurement shots the headline numbers use—App. E only shows a single image at various shot counts. Code is promised but not yet available.\n\nThe 3D experiments are small (three shapes), but for a first QINR in this domain, acceptable. The paper is honest about simulator limits and does not oversell against classical models.\n\nRecommendation: this deserves a serious referee. It's a novel architecture with credible but not yet generalizable evidence. I would accept conditionally, requiring the authors to clarify the expressivity claim (or show larger qubit counts), report beta and shots, and release code. If the expressivity claim fails at n=8, the whole approach needs revision—but that is exactly what peer review is for.\n\nI'd bring it to a reading group and would cite it in QINR-related work.","headline":"A novel QINR architecture worth refereeing; the real-valued ansatz expressivity claim is unproven, but the work is honest and the flaws are fixable.","tokens_in":21713,"tokens_out":3273,"would_cite":true,"duration_ms":34624,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Quantum Visual Field claims that a 5-qubit circuit with a learned Boltzmann amplitude encoder can represent 2D images and 3D signed distance fields more accurately than the prior quantum method QIREN and competitively with classical INR…","keywords":["quantum implicit neural representation","neural amplitude encoding","Gibbs–Boltzmann encoding","real-valued ansatz","parameterised quantum circuit","implicit neural representation","signed distance field","3D shape completion"],"falsifier":"Run the QVF ansatz with 20 or more qubits, keeping depth per qubit and the same Boltzmann encoder, and plot gradient variance over random parameters: if the variance decays exponentially in qubit count as in the strongly entangled ansatz, the barren-plateau-avoidance claim fails. Then, on the same 50-image benchmark, fix the circuit and replace the learned energy encoding with a random amplitude encoding of equal dimension: if PSNR stays at the reported level, the learned energy manifold is not the source of the gains. Finally, on near-term hardware with finite shots, measure PSNR versus shot count; the claim that QVF is competitive with classical INRs should be re-evaluated at the shot budget a real device can afford.","tokens_in":20665,"feed_emoji":"⚛️","tokens_out":9585,"duration_ms":92404,"temperature":0.7,"pith_summary":"The paper introduces Quantum Visual Field (QVF), an implicit neural representation built on a parameterised quantum circuit. QVF encodes a query coordinate (plus an optional latent code) into the amplitudes of an n-qubit state through a learned energy spectrum: a small MLP predicts an energy E, the Gibbs–Boltzmann factor turns it into probabilities, and those probabilities set the squared amplitudes with zero phase. A fully entangled circuit restricted to real-valued unitaries (Pauli-Y rotations plus CNOT entanglers) then evolves the state, and Pauli-Z measurements on the first m qubits produce the image pixel or signed-distance output directly, with no classical decoder. The central claim is that this design beats the previous QINR method QIREN on 2D image reconstruction and is competitive with (and in the reported tables, better than) classical INR baselines such as Siren and DeepSDF-style MLPs for both 2D images and 3D signed distance fields, on a noiseless simulator. If the claim holds, it would show that a quantum circuit can carry the whole implicit representation rather than acting as a Fourier feature extractor for classical layers, and that learned amplitude encoding is a viable way to align classical data with Hilbert-space processing.","feed_headline":"A 5-qubit circuit rivals classical networks on images and 3D shapes","feed_subtitle":"Learned energy-based encoding plus real-valued circuit beats the prior quantum method, matches classical baselines.","key_machinery":"The load-bearing mechanism is the neural amplitude encoder, not the circuit alone. A minimal MLP maps positionally-encoded coordinates and a latent code to an energy spectrum; a Gibbs–Boltzmann factor with inverse temperature β converts that spectrum into probabilities; and those probabilities set the squared amplitudes of the computational-basis input state, with the phase fixed to zero. This is what injects data-dependent non-linearity and a learned frequency structure into the circuit, and it replaces the hand-crafted amplitude embeddings used by earlier quantum auto-encoders. Around it sits the fully entangled real-valued ansatz S(θ): alternating Pauli-Y single-qubit rotations and CNOT entanglers, initialised either as identity blocks or with Gaussian noise scaled by depth, trained by the parameter-shift rule. Local Pauli-Z measurements on the first m qubits extract the output. The paper's trainability argument is the claim that restricting unitaries to the real subspace and using local observables keeps gradient variance from decaying exponentially, which is exactly the role the ansatz plays in the architecture.","core_discovery":"On its own terms, QVF's discovery is that a QINR can produce visual fields directly from projective measurement once two design choices are made together. First, classical data is embedded by learning an energy manifold: the conditional energy spectrum E(Θ,z) inferred by a tiny MLP is converted by the Gibbs distribution P = exp(−βE)/Z into probabilities that become the squared moduli of the input statevector amplitudes (phases set to zero). Second, the ansatz is restricted to the real Hilbert subspace, built from alternating Pauli-Y rotations and CNOT entanglers, so that complex phases never enter the evolution; the paper argues this 'bounded Haar randomness' keeps the optimisation landscape free of barren plateaus while preserving everything measurable with Pauli-Z observables. Measured expectation values are mapped injectively to signal dimensions. The reported consequence is that with n=5 qubits, depth J=5 and a 128-unit MLP, QVF's best variant reaches 0.54×10^−3 MSE and 32.67 dB PSNR on 50 CIFAR-10 images versus 0.78×10^−3 and 31.03 dB for QIREN, and on three ShapeNet SDFs it reaches 0.27×10^−3 MAE versus 0.48×10^−3 for MLP+Siren and 1.43×10^−3 for the MLP+ReLU baseline. Since the ansatz adds only about 170 parameters, the paper attributes the gains to the encoding plus ansatz inductive bias, not to raw parameter count.","pith_inferences":["If the real-valued ansatz claim survives larger qubit counts, it points to architecture-level barren-plateau avoidance that could be combined with smarter initialisation and error-mitigation, extending beyond QINRs to variational quantum algorithms generally.","A natural ablation the paper does not run: replace the learned Boltzmann encoding with a fixed random amplitude encoding of the same dimension; if performance drops sharply, the learned energy manifold—not the circuit—is doing most of the work.","The shot-noise curves imply the simulator numbers are effectively the infinite-shot limit; a meaningful comparison on real hardware would fix a wall-clock shot budget and measure PSNR per unit time.","Because the energy encoder outputs a frequency spectrum in the Fourier sense of variational quantum models, the QVF construction could be lifted to scene reconstruction or neural rendering tasks where spectral bias matters."],"forward_implications":["QVF supports representation of image collections and 3D shape collections through a single latent-conditioned circuit, a capability previous QINR methods did not offer.","With identical quantum resources (5 qubits, same simulator), QVF's best variant reports roughly 30% lower MSE and 1.6 dB higher PSNR than QIREN on 50 images.","On 3D signed distance fields, QVF's MAE of 0.27×10^−3 is below the classical Siren baseline (0.48×10^−3) and far below the ReLU/DeepSDF baseline (1.43×10^−3), so the ansatz improves rather than degrades representation quality.","Finite-shot measurements degrade image quality predictably, and gate-noise simulations show PSNR stays around 25–28 dB at σ=0.1, so near-term hardware use becomes a shot-fidelity trade-off rather than a conceptual obstacle.","Parameter counts scale as O(p^2) for the classical encoder and O(nJ) for the ansatz, meaning the quantum part is a low-cost inductive-bias addition to an otherwise classical representation."],"supporting_citations":[{"why":"Defines QIREN, the prior QINR method with classical sandwich layers that QVF is compared against and improves upon.","marker":"[65]"},{"why":"Introduces 3D-QAE with hand-crafted amplitude embeddings; it is the baseline encoding idea QVF replaces with a learned energy-based encoding.","marker":"[43]"},{"why":"Supplies the Siren periodic activation baseline and the Siren activation used in QVF's classical encoder.","marker":"[53]"},{"why":"Supplies the DeepSDF-style MLP+ReLU baseline and the Gaussian latent-code prior QVF borrows for shape collections.","marker":"[41]"},{"why":"Provides the noiseless gate-based simulator on which all QVF experiments and gradient-variance measurements are run.","marker":"[8]"},{"why":"Establishes the truncated Fourier expansion of variational quantum models, used to argue the learned encoding controls the accessible frequency spectrum.","marker":"[50]"},{"why":"Formalises barren plateaus and the exponential gradient-variance decay that the real-valued ansatz is designed to avoid.","marker":"[34]"},{"why":"Shows how cost-function and observable locality affect gradient magnitudes, supporting the use of local Pauli measurements.","marker":"[12]"},{"why":"Introduces the parameter-shift rule used to optimise the quantum circuit parameters during training.","marker":"[38]"}],"fun_headline_variants":["5-qubit circuit beats prior quantum, rivals classical on images and shapes","Quantum visual fields: tiny real-valued circuit matches classical nets","Learnable energy manifold powers 5-qubit visual field learning","Direct measurement, real-valued ansatz: QVF beats prior quantum method","5-qubit quantum visual fields: no post-processing, matches classical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central bet is that a circuit built only from real-valued rotations and CNOT entanglers loses no measurement-relevant expressivity while keeping gradients from vanishing; the paper supports this with gradient-variance measurements on a simulator, not with a proof, so it is the premise that would most quickly sink the approach at scale.","fun_headline_variants_meta":{"raw":{"variants":["5-qubit circuit beats prior quantum, rivals classical on images and shapes","Quantum visual fields: tiny real-valued circuit matches classical nets","Learnable energy manifold powers 5-qubit visual field learning","Direct measurement, real-valued ansatz: QVF beats prior quantum method","5-qubit quantum visual fields: no post-processing, matches classical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000747,"raw_usage":{"total_tokens":3423,"prompt_tokens":1133,"completion_tokens":2290,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":749,"completion_tokens_details":{"reasoning_tokens":2200}},"tokens_in":749,"tokens_out":2290,"duration_ms":15311,"temperature":1.0,"reasoning_tokens":2200,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:29:23.854828+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the QVF ansatz with 20 or more qubits, keeping depth per qubit and the same Boltzmann encoder, and plot gradient variance over random parameters: if the variance decays exponentially in qubit count as in the strongly entangled ansatz, the barren-plateau-avoidance claim fails. Then, on the same 50-image benchmark, fix the circuit and replace the learned energy encoding with a random amplitude encoding of equal dimension: if PSNR stays at the reported level, the learned energy manifold is not the source of the gains. Finally, on near-term hardware with finite shots, measure PSNR versus shot count; the claim that QVF is competitive with classical INRs should be re-evaluated at the shot budget a real device can afford.","supporting_citations":[{"cited_title":"Quan- tum implicit neural representations","cited_arxiv_id":null,"evidence_quote":"Defines QIREN, the prior QINR method with classical sandwich layers that QVF is compared against and improves upon."},{"cited_title":"3D-QAE: Fully quantum auto-encoding of 3d point clouds","cited_arxiv_id":null,"evidence_quote":"Introduces 3D-QAE with hand-crafted amplitude embeddings; it is the baseline encoding idea QVF replaces with a learned energy-based encoding."},{"cited_title":"Implicit neural represen- tations with periodic activation functions","cited_arxiv_id":null,"evidence_quote":"Supplies the Siren periodic activation baseline and the Siren activation used in QVF's classical encoder."},{"cited_title":"Deepsdf: Learning continuous signed distance functions for shape representation","cited_arxiv_id":null,"evidence_quote":"Supplies the DeepSDF-style MLP+ReLU baseline and the Gaussian latent-code prior QVF borrows for shape collections."},{"cited_title":"Ef- fect of data encoding on the expressive power of variational quantum-machine-learning models","cited_arxiv_id":null,"evidence_quote":"Establishes the truncated Fourier expansion of variational quantum models, used to argue the learned encoding controls the accessible frequency spectrum."},{"cited_title":"Barren plateaus in quan- tum neural network training landscapes","cited_arxiv_id":null,"evidence_quote":"Formalises barren plateaus and the exponential gradient-variance decay that the real-valued ansatz is designed to avoid."},{"cited_title":"Cost function dependent barren plateaus in shallow parametrized quantum circuits","cited_arxiv_id":null,"evidence_quote":"Shows how cost-function and observable locality affect gradient magnitudes, supporting the use of local Pauli measurements."},{"cited_title":"Quantum circuit learning","cited_arxiv_id":null,"evidence_quote":"Introduces the parameter-shift rule used to optimise the quantum circuit parameters during training."}],"review_version":1}