Pith. sign in

REVIEW 3 major objections 5 minor 49 references

A physics-informed classical neural-network latent, copied into a quantum IQP circuit's latent block, improves the circuit's generative accuracy on Burgers' equation solutions compared with random initialization.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 01:27 UTC pith:UUSTKS4J

load-bearing objection A clean empirical study of PINN-initialized IQP latent adaptation, but the headline comparison is missing a norm-matched control and multi-seed statistics, so the 'physics-informed' part of the claim is not yet established. the 3 major comments →

arxiv 2607.28866 v1 pith:UUSTKS4J submitted 2026-07-30 quant-ph

Generative IQP Circuit Learning with Physics-Informed Latent Initialization

Chen-Yu Liu , Leonardo Placidi , Marco Ballarin , Enrico Rinaldi This is my paper
classification quant-ph
keywords IQP circuitsgenerative modelinglatent adaptationphysics-informed neural networksBurgers' equationwarm startmaximum mean discrepancyPlatonic representation hypothesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that a quantum generative model based on an instantaneous quantum polynomial-time (IQP) circuit can be trained better if its low-dimensional latent variable is initialized with a latent vector extracted from a classical physics-informed neural network (PINN) surrogate, rather than randomly. On Burgers' equation solutions, the PINN-initialized circuit consistently reaches lower mean-squared reconstruction error at unseen viscosity values, across three different initial-condition families and across surrogate resolutions down to 12x12. The paper explicitly does not claim quantum advantage; the IQP training remains classically simulable. The point of the result is that classical physics-informed representations can supply a structured warm start for quantum generative latent adaptation, which could matter for future, harder settings where a classical surrogate is all one has.

Core claim

The central claim is that the transfer defined by z_PINN(ν0) → θ_lat^(ν0) — copying the reference-viscosity latent of a latent-conditioned PINN into the first 50 parameters of an 18-qubit IQP Born machine — improves the subsequent two-stage latent-adaptation protocol. In stage one, the transferred latent is held fixed while a shared circuit core is trained on the reference viscosity via an MMD loss over bitstrings. In stage two, the core is frozen and instance latents are adapted sequentially for other viscosities. The paper's numerical experiments show lower MSE at every tested unseen viscosity for the dual_exp, exp, and sin initial conditions, with all three PINN surrogate resolutions, tha

What carries the argument

Latent-conditioned PINN surrogate: an MLP G_Θ(x,t,z) trained to output Burgers' solutions with a combined data-loss and PDE-residual loss, one latent z^(ν) per viscosity. IQP Born machine: U(θ)=H⊗n exp(i Σ_{s⊆[n]} θ_s Z_s) H⊗n, whose bitstring output is trained with an MMD objective that for IQP circuits reduces to a sum over Pauli expectation values. The latent-adaptation partition θ=(θ_core, θ_lat) with d_lat=50, and the elementwise copy in Eq. (44), are the mechanism that injects physics information.

Load-bearing premise

The load-bearing premise is that the 50 coordinates of the PINN's latent vector line up, element by element, with the first 50 IQP circuit parameters (single-qubit terms then low-order multi-qubit terms), so that an elementwise copy is a meaningful transfer; the paper gives no argument that this ordering correspondence holds.

What would settle it

Train the same pipeline with the PINN latent coordinates randomly permuted before transfer. If the MSE improvement over random initialization survives arbitrary permutations, the specific coordinate structure is irrelevant and the benefit is a warm-start/magnitude artifact; if it does not survive, the paper's chosen identity ordering is doing unexplained work.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If correct, physics-informed latent initialization is a reliable improvement over random initialization for IQP generative modeling on parameterized PDE solution families.
  • The improvement does not depend on high surrogate resolution; even a 12×12 grid transfers useful structure.
  • The classical and quantum latent spaces retain compatible inter-instance ordering, suggesting a principled way to use classical surrogates as warm starts for quantum models.
  • The same two-stage protocol can be applied to other parameterized PDEs where a latent-conditioned classical surrogate is available.
  • No quantum-advantage claim follows; this is a training-efficiency gain within a classically simulable setting.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Inference: A decisive untested control is to randomly permute the coordinates of the PINN latent before the Eq. (44) copy. Because the paper's similarity analyses use cosine similarity, they are permutation-invariant and cannot detect whether elementwise alignment matters. If the benefit survives permutation, the effect is a magnitude/warm-start artifact; if it does not, the paper's chosen orderin
  • Inference: The elementwise copy assumes the PINN MLP bottleneck units and the IQP Pauli-string parameter ordering share semantic alignment; nothing in the construction fixes this. A direct test is to compare elementwise alignment statistics across random permutations of the latent coordinates.
  • Inference: The method's generality could be probed by moving to a PDE whose solution manifold is less smooth in ν or to higher latent dimension; the paper's warm-start interpretation would predict the benefit shrinks as the manifold becomes less structured.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The manuscript proposes a physics-informed initialization for the latent variable used in IQP latent adaptation. A latent-conditioned PINN is trained on lower-resolution Burgers' equation solutions; the PINN's latent vector for the reference viscosity is copied (Eq. 44) into the first 50 IQP circuit parameters, which are then held fixed while the shared IQP core is learned. Subsequent latent adaptation for other viscosities proceeds with the core frozen. The paper's central claim is that this PINN-initialized latent consistently improves reconstruction MSE at unseen viscosities relative to random latent initialization, across three initial-condition families and three surrogate resolutions, and that the PINN and IQP latent spaces preserve compatible inter-instance structure. The paper explicitly disclaims any quantum-advantage claim and presents the method as a practical initialization strategy.

Significance. If validated, the result gives a concrete, controlled demonstration that a classical physics-informed surrogate representation can act as an effective warm start for a quantum generative model, extending the warm-starting literature from QAOA to IQP latent adaptation. The main strengths are the clean comparison setup (identical architecture, objective, optimizer, and optimization budget, differing only in latent initialization), the explicit reproducibility appendix with hyperparameters and seeds, the honest limitations section, and the inclusion of classical baselines (Appendix D). The central claim is falsifiable and moderately scoped, which makes the missing controls and statistical details particularly consequential.

major comments (3)
  1. [Sec. VI.B / Fig. 5 / Table V] The headline comparison is confounded by initial latent scale. Table V states that the random initial parameter scale is 10^-4, while Sec. VI.D notes that the transferred PINN latent has 'comparatively large overall magnitude relative to the subsequent latent updates.' Without a norm-matched random-initialization control, the gain in Fig. 5 could be a warm-start magnitude effect rather than a consequence of the physics-informed structure. Please add a control in which a random latent is rescaled to the same L2 norm as z_{PINN}(nu0) (and, ideally, a zero-mean random vector of the same norm) and rerun the same protocol. This is a concrete, fixable omission and is load-bearing for the phrase 'physics-informed initialization.'
  2. [Sec. V, Eq. (44) / Sec. VI.A] The transfer is an elementwise copy of the PINN latent vector into the first 50 IQP parameters ordered by Pauli strings (single-qubit terms first, then low-order multi-qubit terms). The PINN's MLP bottleneck coordinates have no intrinsic ordering, and no argument is given that they correspond to the IQP Pauli-string ordering. The pairwise-similarity analysis (Eq. 55) uses cosine similarity and Spearman rank correlation, both invariant to coordinate permutations, so it cannot validate the identity map. Please add a permutation ablation: transfer a randomly permuted version of z_{PINN}(nu0) and report the resulting MSE curves. If the permuted transfer reproduces the benefit, the specific structure is irrelevant; if not, the identity map needs a mechanical or empirical justification.
  3. [Fig. 5, Tables II and VI] All reported MSE curves and Spearman correlations come from single fixed-seed runs, with no error bars, multiple seeds, or significance tests. The central claim is that the PINN-initialized variant 'consistently outperforms' the random baseline across all conditions. Given the stochastic MMD estimator and the dependence of the latent adaptation on random initialization, single runs cannot support that consistency claim. Please repeat the main comparison (random vs. PINN-init, at least for one surrogate resolution) with 5–10 independent seeds and report mean±std (or a paired comparison across the unseen viscosity set). This is directly relevant to the paper's main quantitative conclusion.
minor comments (5)
  1. [Sec. VII.B, Eq. (58)] The 'effective pre-optimization' assumption is essentially a restatement of the desired conclusion. The text labels it as an interpretation, which is appropriate, but the paper would be strengthened by testing a concrete prediction of that model, e.g., that optimization from the PINN-init latent reaches a given loss in fewer explicit optimization steps than from random init.
  2. [Sec. VI.A / Table V] Please specify the exact distribution used for random latent initialization (e.g., uniform in [-1e-4, 1e-4] or Gaussian with std 1e-4), not only the scale, for reproducibility.
  3. [Appendix B, Eq. (B1)] The notation 'R(Y k)' should read 'R(Y_k)' for consistency with the preceding sentence.
  4. [Fig. 5] The non-monotonic resolution dependence (12×12 best for the exp case) is interesting but underexplained; adding a brief comment on whether this reflects optimization noise or a surrogate-quality effect would improve the discussion.
  5. [Table X] The table says 'Fixed random seeds used for each reported training run' but does not list those seeds for the main IQP runs; please list them or state why they are omitted.

Circularity Check

1 steps flagged

Empirical PINN-vs-random comparison is self-contained; only the optional warm-start 'explanation' (Eq. 58) assumes its own conclusion, and it is not load-bearing.

specific steps
  1. self definitional [Sec. VII.B, Eqs. (58)-(64)]
    "The central modeling assumption is that the transferred PINN initialization can be interpreted as an approximate m-step pre-optimized iterate of the latent adaptation dynamics: θ(ν)lat,0 ≈ ĝmν (θ(ν)rand,0), m≥1 ... Combining this with the pre-optimization assumption (58), we obtain ... Thus, under the effective pre-optimization interpretation, the transferred initialization is closer to the target minimizer than a random initialization..."

    The 'interpretation' postulates that the PINN initialization already equals m iterations of the IQP latent optimizer. From that postulate the paper derives that the initialization is closer to the minimizer, has smaller initial loss, and needs fewer explicit steps. The benefit is therefore a logical consequence of the assumption itself, not an independent explanation. However, the paper explicitly labels it as 'a local theoretical interpretation rather than a formal proof,' and the main central claim rests on the direct empirical MSE comparison in Fig. 5, so this circularity is not load-bearing.

full rationale

The central claim of the paper is an empirical comparison: physics-informed latent initialization from a PINN surrogate yields lower reconstruction MSE than random initialization in IQP latent adaptation. That claim is supported by direct experiments (Fig. 5) over multiple initial conditions, unseen viscosities, and surrogate resolutions, with optimization settings stated as identical between conditions. The comparison does not reduce to an input by construction, and it is externally falsifiable via the reported MSE curves. The transferred-latent step Eq. (44) is a warm-start heuristic; while the coordinate-ordering correspondence is not justified, the paper itself lists 'parameter ordering' as a possible dependency in Sec. VII.D, and Sec. VI.D acknowledges that the transferred latent's large magnitude compresses cosine similarities. These are robustness or missing-justification concerns, not circular derivations. The pairwise similarity analysis is presented as supporting interpretation, not as the central evidence; it is scale-invariant and thus cannot by itself rule out a warm-start magnitude effect, but that is a correctness/confound issue, not circularity. The only genuine self-definitional passage is the 'effective pre-optimization' interpretation in Sec. VII.B, where Eq. (58) assumes the transfer equals m optimizer steps and then derives faster convergence. The paper flags this as an interpretation rather than a proof, and the main claim does not rest on it. Self-citations to the latent-adaptation framework [22] and PRH [37] are contextual; the present method is implemented and benchmarked directly, so no load-bearing self-citation chain forces the result. Overall, the central empirical derivation is self-contained; the minor circular interpretation justifies a low score of 2.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

The ledger shows the paper's contribution is an empirical transfer recipe on top of a large existing machinery. The transferred latent is itself a fitted object (a PINN optimization output), and the paper's interpretation rests on two explicit ad hoc assumptions (PRH alignment, effective pre-optimization) plus one unflagged assumption (coordinate correspondence). No new physical entities are invented, and the main free parameters are hand-chosen hyperparameters rather than physics constants.

free parameters (6)
  • Transferred PINN latent z(nu0) in R^50 = not reported (optimized vector)
    The initialization itself is the output of PINN Stage-I/Stage-II optimization (Secs. III.D-F); its values are fitted to the Burgers data family, not derived. The paper does not report the latent values or test sensitivity to them.
  • Latent dimension d_lat = 50
    Set by hand following Ref. [22] (Sec. VI.A); no ablation over d_lat is performed, though the transfer's informativeness plausibly depends on it.
  • MMD kernel bandwidth scale = median heuristic / 5
    A hand-chosen rescaling of the Gaussian kernel (Table V) that directly shapes the training landscape Eq. (8).
  • MMD stochastic estimator sizes = 1500 random operators; 2000 samples
    Hand-chosen Monte Carlo sizes (Table V) that set gradient noise; no sensitivity analysis is reported.
  • Reference viscosity and sweep = nu0 = 0.06; anchors {0.11, ..., 0.51}
    Choice of anchor instance and anchor spacing is arbitrary; the method's sensitivity to which instance anchors the shared core is not tested.
  • PINN loss weights lambda_data, lambda_phys = not reported
    The surrogate latent used for transfer is the solution of Eq. (34), whose balance between data and physics terms is set by these weights; Appendix E omits their values.
axioms (6)
  • standard math The IQP MMD loss collapses to a sum over Pauli-Z moment differences (Eq. 9).
    Taken from Refs. [19,22]; it is what makes classical training of the 'quantum' model possible.
  • domain assumption The fixed-core IQP family is locally expressive enough to track the target distribution family, and the target family is smooth in viscosity (Appendix A, Eqs. A1-A9).
    Necessary for latent adaptation to work at all; shared with the random-init baseline, so it does not distinguish the proposed method.
  • ad hoc to paper PRH relational alignment between PINN and IQP latent spaces (Eq. 42).
    The paper's stated motivation for the transfer (Sec. IV); it is a hypothesis, tested only via cosine-similarity/Spearman analyses that also hold for random init.
  • ad hoc to paper Effective pre-optimization: PINN init is approximately m repeated IQP latent-gradient steps (Eq. 58), plus local contractivity and PL inequality (Eqs. 59-62).
    Sec. VII.B explicitly labels this the 'central modeling assumption'; the convergence bounds that follow are consequences of the assumption, not independent support.
  • ad hoc to paper Coordinate-wise semantic correspondence between PINN latent units and the first 50 IQP Pauli-string parameters (identity map in Eq. 44).
    Elementwise copy is performed in Sec. V with no argument that the PINN bottleneck ordering matches the single-/low-order Pauli-term ordering; untested by permutation controls.
  • domain assumption Six-bit uniform quantization of (x, t, u) is an adequate sample representation for MMD training.
    The float-to-bit map (Sec. II.A) fixes the modeled distribution; quantization error and its effect on the reconstructed-field MSE are not analyzed.

pith-pipeline@v1.3.0-alltime-deepseek · 26368 in / 20022 out tokens · 209417 ms · 2026-08-03T01:27:49.224944+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Generative IQP Circuit Learning with Physics-Informed Latent Initialization." pith.science (2026). https://pith.science/paper/UUSTKS4J

@misc{pith2026260728866,
  author       = {Pith},
  title        = {Pith review of: Generative IQP Circuit Learning with Physics-Informed Latent Initialization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UUSTKS4J}},
  note         = {Machine review of arXiv:2607.28866}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Quantum generative learning based on instantaneous quantum polynomial-time (IQP) circuits can benefit from efficient classical training strategies. A recent latent adaptation framework for IQP-based generative modeling transfers shared circuit parameters across instances of the same task with different hyperparameters while adapting only a low-dimensional latent variable for each new instance. However, existing approaches initialize this latent variable randomly, which can limit optimization efficiency and performance. In this work, we introduce a physics-informed latent initialization scheme for IQP generative learning to improve upon existing random initialization schemes. Motivated by the platonic representation hypothesis, we use latent representations extracted from a classical physics-informed neural network (PINN) surrogate to initialize the latent variables of the quantum model for the solution of the Burgers' equation. The initialized IQP model is then adapted on a higher-resolution solution domain. We find that this structured initialization consistently outperforms random latent initialization, yielding improved adaptation behavior and stronger generative accuracy across multiple viscosity settings. These results show that classical surrogate representations can provide useful inductive bias for quantum generative models and offer a practical route to improved initialization in IQP-based learning.

Figures

Figures reproduced from arXiv: 2607.28866 by Chen-Yu Liu, Enrico Rinaldi, Leonardo Placidi, Marco Ballarin.

Figure 1
Figure 1. Figure 1: FIG. 1. Schematic overview of the proposed physics-informed latent initialization scheme for IQP generative learning. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2. PRH-motivated view of latent alignment between the PINN surrogate and the IQP generative model. A parameterized [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3. Training loss during the reference-core learning step [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4. Qualitative comparison between physics-informed latent initialization and random latent initialization at the unseen [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5. Mean-squared error (MSE) between the reconstructed IQP predictions and the numerical solutions across unseen [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6. Pairwise similarity preservation between the adapted [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: FIG. 8. Viscosity-distance control. Raw Spearman correlations between pairwise PINN and IQP latent similarities are [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: FIG. 9. Trivial viscosity-only baseline. Pairwise similarities induced by a one-dimensional viscosity embedding are compared [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: FIG. 10. Randomized adaptation-order ablation. The original monotonic-order PINN-IQP latent correlation is compared [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: FIG. 11. Classical cDCGAN baselines for Burgers’ equation solution generation. We compare the IQP model initial [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 5 linked inside Pith

  1. [1]

    Schuld and N

    M. Schuld and N. Killoran, Quantum machine learning in feature hilbert spaces, Physical Review Letters122, 040504 (2019)

    2019

  2. [2]

    Benedetti, E

    M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, Pa- rameterized quantum circuits as machine learning mod- els, Quantum Science and Technology4, 043001 (2019)

    2019

  3. [3]

    A. Mari, T. R. Bromley, J. Izaac, M. Schuld, and N. Killo- ran, Transfer learning in hybrid classical-quantum neural 16 networks, Quantum4, 340 (2020)

    2020

  4. [4]

    Liu, E.-J

    C.-Y. Liu, E.-J. Kuo, C.-H. Abraham Lin, J. Gem- sun Young, Y.-J. Chang, M.-H. Hsieh, and H.-S. Goan, Quantum-train: Rethinking hybrid quantum-classical machine learning in the model compression perspective, Quantum Machine Intelligence7, 80 (2025)

    2025

  5. [5]

    Huang, M

    H.-Y. Huang, M. Broughton, M. Mohseni, R. Babbush, S. Boixo, H. Neven, and J. R. McClean, Power of data in quantum machine learning, Nature Communications12, 2631 (2021)

    2021

  6. [6]

    Belis, J

    V. Belis, J. Bowles, R. Gupta, E. Peters, and M. Schuld, Spectral methods: Crucial for machine learning, natu- ral for quantum computers? (2026), arXiv:2603.24654 [quant-ph]

    Pith/arXiv arXiv 2026

  7. [7]

    Liu, K.-C

    C.-Y. Liu, K.-C. Chen, K. Murota, S. Y.-C. Chen, and E. Rinaldi, Quantum relational knowledge distillation, in Proceedings of UniReps: the Third Edition of the Work- shop on Unifying Representations in Neural Models, Pro- ceedings of Machine Learning Research, Vol. 322 (PMLR,

  8. [8]

    D. Dong, C. Chen, H. Li, and T.-J. Tarn, Quantum rein- forcement learning, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics)38, 1207 (2008)

    2008

  9. [9]

    C.-Y. Liu, L. Placidi, K.-C. Chen, S. Y.-C. Chen, and G. Matos, You only measure once: On designing single-shot quantum machine learning models (2025), arXiv:2509.20090 [cs.LG]

    arXiv 2025

  10. [10]

    S. Y.-C. Chen, C.-H. H. Yang, J. Qi, P.-Y. Chen, X. Ma, and H.-S. Goan, Variational quantum circuits for deep reinforcement learning, IEEE Access8, 141007 (2020)

    2020

  11. [11]

    Liu, E.-J

    C.-Y. Liu, E.-J. Kuo, C.-H. A. Lin, S. Chen, J. G. Young, Y.-J. Chang, and M.-H. Hsieh, Training classical neural networks by quantum machine learning, in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), Vol. 02 (2024) pp. 34–38

    2024

  12. [12]

    C.-H. A. Lin, C.-Y. Liu, and K.-C. Chen, Quantum-train long short-term memory: Application on flood predic- tion problem, in2024 IEEE International Conference on Quantum Computing and Engineering (QCE), Vol. 02 (2024) pp. 268–273

    2024

  13. [13]

    H. Wang, Y. Ding, J. Gu, Y. Lin, D. Z. Pan, F. T. Chong, and S. Han, Quantumnas: Noise-adaptive search for robust quantum circuits, in2022 IEEE International Symposium on High-Performance Computer Architecture (HPCA)(2022) pp. 692–708

    2022

  14. [14]

    Du, M.-H

    Y. Du, M.-H. Hsieh, T. Liu, and D. Tao, A grover-search based quantum learning scheme for classification, New Journal of Physics23, 023020 (2021)

    2021

  15. [15]

    Liu, C.-H

    C.-Y. Liu, C.-H. H. Yang, H.-S. Goan, and M.-H. Hsieh, A quantum circuit-based compression perspective for parameter-efficient learning, inThe Thirteenth Interna- tional Conference on Learning Representations(2025)

    2025

  16. [16]

    Zhang, P

    B. Zhang, P. Xu, X. Chen, and Q. Zhuang, Generative quantum machine learning via denoising diffusion prob- abilistic models, Physical Review Letters132, 100602 (2024)

    2024

  17. [17]

    Lloyd and C

    S. Lloyd and C. Weedbrook, Quantum generative ad- versarial learning, Physical Review Letters121, 040502 (2018)

    2018

  18. [18]

    Huang, M

    H.-Y. Huang, M. Broughton, N. Eassa, H. Neven, R. Babbush, and J. R. McClean, Generative quantum advantage for classical and quantum problems (2025), arXiv:2509.09033 [quant-ph]

    Pith/arXiv arXiv 2025

  19. [19]

    Recio-Armengol, S

    E. Recio-Armengol, S. Ahmed, and J. Bowles, Train on classical, deploy on quantum: Scaling generative quantum machine learning to a thousand qubits (2025), arXiv:2503.02934 [quant-ph]

    arXiv 2025

  20. [20]

    Placidi, E

    L. Placidi, E. Rinaldi, K. Fujii, and C.-Y. Liu, The im- pact of qubit connectivity on quantum advantage in noisy iqp circuits (2026), arXiv:2604.12635 [quant-ph]

    Pith/arXiv arXiv 2026

  21. [21]

    Ball´ o-Gimbernat, M

    O. Ball´ o-Gimbernat, M. Arroyo-S´ anchez, P. Garc ´ ıa- Molina, A. Garriga, and F. Vilari˜ no, Shallow instanta- neous quantum polynomial-time circuits for generative modeling on noisy intermediate-scale quantum hardware, Physical Review A113, 042617 (2026)

    2026

  22. [22]

    C.-Y. Liu, L. Placidi, E. Brunner, and E. Rinaldi, Toward generative quantum utility via correlation-complexity map (2026), arXiv:2603.06440 [quant-ph]

    Pith/arXiv arXiv 2026

  23. [23]

    M. J. Bremner, R. Jozsa, and D. J. Shepherd, Classical simulation of commuting quantum computations implies collapse of the polynomial hierarchy, Proceedings of the Royal Society A: Mathematical, Physical and Engineer- ing Sciences467, 459 (2011)

    2011

  24. [24]

    Leontica and D

    S. Leontica and D. Amaro, Exploring the neighborhood of 1-layer qaoa with instantaneous quantum polynomial circuits, Physical Review Research6, 013071 (2024)

    2024

  25. [25]

    Coyle, D

    B. Coyle, D. Mills, V. Danos, and E. Kashefi, The born supremacy: quantum advantage and training of an ising born machine, npj Quantum Information6, 60 (2020)

    2020

  26. [26]

    K. Shen, S. Pielawa, V. Dunjko, and H. Wang, Character- izing trainability of instantaneous quantum polynomial circuit born machines (2026), arXiv:2602.11042 [quant- ph]

    arXiv 2026

  27. [27]

    A. R. Khojasteh, S. Laizet, D. Heitz, and Y. Yang, La- grangian and eulerian dataset of the wake downstream of a smooth cylinder at a reynolds number equal to 3900, Data in Brief40, 107725 (2022)

    2022

  28. [28]

    J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Bab- bush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nature Communications9, 4812 (2018)

    2018

  29. [29]

    Grant, L

    E. Grant, L. Wossnig, M. Ostaszewski, and M. Benedetti, An initialization strategy for addressing barren plateaus in parametrized quantum circuits, Quantum3, 214 (2019)

    2019

  30. [30]

    R. Puig, M. Drudis, S. Thanasilp, and Z. Holmes, Vari- ational Quantum Simulation: A Case Study for Under- standing Warm Starts, PRX Quantum6, 010317 (2025)

    2025

  31. [31]

    ˇZunkoviˇ c, M

    B. ˇZunkoviˇ c, M. Ballarin, L. Wright, and M. Lubasch, Scalable, self-verifying variational quantum eigensolver using adiabatic warm starts (2026), arXiv:2602.17612 [quant-ph]

    arXiv 2026

  32. [32]

    D. J. Egger, J. Mareˇ cek, and S. Woerner, Warm-starting quantum optimization, Quantum5, 479 (2021)

    2021

  33. [33]

    Raissi, P

    M. Raissi, P. Perdikaris, and G. E. Karniadakis, Physics- informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations, Journal of Computational Physics378, 686 (2019)

    2019

  34. [34]

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, Physics-informed machine learn- ing, Nature Reviews Physics3, 422 (2021)

    2021

  35. [35]

    W. Cho, M. Jo, H. Lim, K. Lee, D. Lee, S. Hong, and N. Park, Parameterized physics-informed neural net- works for parameterized pdes, inProceedings of the 41st International Conference on Machine Learning, ICML’24 (JMLR.org, 2024)

    2024

  36. [36]

    H. Zhou, S. Cheng, and R. Arcucci, Multi-fidelity 17 physics constrained neural networks for dynamical sys- tems, Computer Methods in Applied Mechanics and En- gineering420, 116758 (2024)

    2024

  37. [37]

    M. Huh, B. Cheung, T. Wang, and P. Isola, Position: The platonic representation hypothesis, inForty-first In- ternational Conference on Machine Learning(2024)

    2024

  38. [38]

    Spearman, The proof and measurement of association between two things, The American Journal of Psychology 100, 441 (1987)

    C. Spearman, The proof and measurement of association between two things, The American Journal of Psychology 100, 441 (1987)

    1987

  39. [39]

    R. N. Forthofer and R. G. Lehnen, Rank correlation methods, inPublic Program Analysis: A New Categorical Data Approach(Springer, 1981) pp. 146–163

    1981

  40. [40]

    Cervera-Lierta, J

    A. Cervera-Lierta, J. S. Kottmann, and A. Aspuru- Guzik, Meta-variational quantum eigensolver: Learning energy profiles of parameterized hamiltonians for quan- tum simulation, PRX Quantum2, 020329 (2021)

    2021

  41. [41]

    Shaydulin, P

    R. Shaydulin, P. C. Lotshaw, J. Larson, J. Ostrowski, and T. S. Humble, Parameter transfer for quantum approx- imate optimization of weighted maxcut, ACM Transac- tions on Quantum Computing4, 1 (2023)

    2023

  42. [42]

    Galda, E

    A. Galda, E. Gupta, J. Falla, X. Liu, D. Lykov, Y. Alex- eev, and I. Safro, Similarity-based parameter transfer- ability in the quantum approximate optimization algo- rithm, Frontiers in Quantum Science and Technology2, 1200975 (2023)

    2023

  43. [43]

    Radford, L

    A. Radford, L. Metz, and S. Chintala, Unsupervised rep- resentation learning with deep convolutional generative adversarial networks (2016), arXiv:1511.06434 [cs.LG]. 18 Appendix A: Geometric motivation for latent adaptation In this appendix, we provide a simple geometric motivation for the latent-adaptation assumption used in the main text. The purpose i...

    Pith/arXiv arXiv 2016

  44. [44]

    Viscosity-Distance Control First, we test whether the PINN-IQP latent alignment remains after explicitly controlling for viscosity distance. For every pair of viscosities, we compute Dij =|ν i −ν j|,(C1) 20 dual_exp32x32 PINN init dual_exp16x16 PINN init dual_exp12x12 PINN init dual_expbaseline random init exp 32x32 PINN init exp 16x16 PINN init exp 12x12...

  45. [45]

    T rivial Viscosity-Only Baseline Second, we compare the learned latent similarities against a trivial one-dimensional viscosity embedding. In this baseline, each viscosity is represented only by itself, zvisc(νi) = [νi],(C3) 21 dual_exp32x32 PINN init dual_exp16x16 PINN init dual_exp12x12 PINN init dual_expbaseline random init exp 32x32 PINN init exp 16x1...

  46. [46]

    In the original experiment, viscosities are adapted in increasing order, ν0, ν1,

    Randomized Adaptation Order Finally, we test whether the latent relation is induced by the monotonic warm-start procedure used during sequential IQP latent adaptation. In the original experiment, viscosities are adapted in increasing order, ν0, ν1, . . . , νK.(C6) This could impose a smooth trajectory in IQP latent space even if the learned relation were ...

  47. [47]

    Conditional DCGAN baseline We use a conditional deep convolutional generative adversarial network (cDCGAN) [43] as the classical generative baseline. In contrast to the IQP model, which represents samples in the binarized (x, t, u) domain and reconstructs the solution field through sampling and inverse decoding, the cDCGAN treats each Burgers’ equation so...

  48. [48]

    Comparison with IQP-PINN latent adaptation The results are shown in Fig. 11. The comparison shows that the relative performance of the IQP-PINN model and the cDCGAN baselines depends strongly on the initial-condition family. For thedual expinitial condition, the IQP model with 12×12 PINN initialization achieves the lowest MSE across the tested viscosity r...

  49. [49]

    Scope of the classical comparison We emphasize that these experiments do not rule out the existence of stronger classical baselines. Modern classical generative modeling includes many architectures beyond DCGANs, including U-Nets, diffusion models, transformer- based generators, neural operators, and autoregressive models. A comprehensive search over clas...

    2000