Pith. sign in

REVIEW 3 major objections 4 minor 27 references

How Much Reconstruction Does Quantum Machine Learning Need? Late Fusion of Independently Trained Quantum Subcircuits

T0 review · 3 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper shows that circuit-cut quantum machine learning can replace the exponential-cost reconstruction step with late fusion—independently trained subcircuits plus a small classical head—and match full reconstruction accuracy within…

desk verdict Solid empirical case for replacing exponential reconstruction with trained late fusion in cut-circuit QML, but the 'self-characterizing' diagnostic is only validated at small scale and the paper mostly concedes this. read the letter →

arxiv 2608.05595 v1 pith:CTGS64O2 submitted 2026-08-06 quant-ph cs.DCcs.LG

Prabhjot Singh , Adel N. Toosi , Rajkumar Buyya This is my paper · ORCID
classification quant-phcs.DCcs.LG
keywords circuitcuttinglatefusionquantummachinelearningneuralnetworksquasiprobabilitydecompositionentanglementcostnoiserobustnessreconstructionoverhead
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 asks whether the exponential-cost reconstruction step in circuit cutting—a standard technique that lets a large quantum neural network run as independent subcircuits on small devices—is actually needed for machine-learning tasks. It replaces the quasiprobability sum that rebuilds cross-cut correlations with late fusion: each subcircuit is trained and measured independently, and a small classical head combines their outputs into the final prediction. Across controlled sweeps of cross-cut dependence and standard datasets, late fusion matches full reconstruction accuracy within 0.04 at every point, at a measurement cost that stays flat in the number of cuts while reconstruction's overhead grows like $O(\gamma^{2k})$. The paper also introduces a quantumness dial that interpolates from pure fusion to full reconstruction and a cut-entanglement diagnostic that indicates how much reconstruction a task needs (Spearman $\rho=0.59$ over 104 runs), and it maps the entangled-data boundary where fusion provably fails. The authors do not claim any accuracy edge over classical machine learning; the contribution is a cheaper, noise-robust readout for circuit-cutting QML and an instrument for deciding when the expensive readout is worth paying for.

What carries the argument

The load-bearing object is the operator-Schmidt decomposition of each cross-register gate, $G = \sum_\mu g_\mu (L_\mu \otimes R_\mu)$, which yields the exact reconstruction formula (Eq. 1) and the per-cut coefficient-1-norm $\gamma = \sum_\mu |g_\mu|$ whose power $\gamma^{2k}$ sets the exponential sampling overhead. The quantumness dial $Q$ carries the argument: keep the highest-magnitude terms of the reconstruction sum until a fraction $Q$ of the 1-norm mass is retained, feed the partial reconstruction together with the raw subcircuit features into a trainable classical head, so $Q=0$ is pure fusion and $Q=1$ is exact reconstruction, and the cost interpolates from $O(1)$ to the full $\gamma^{2k}$. The cut-entanglement diagnostic $S_A = S(\rho_A(x))$, the entanglement entropy across the cut of the trained coupled circuit averaged over the data, is the measurable proxy: because reconstruction overhead is lower-bounded by the entanglement cost across the cut, the resource reconstruction spends and the correlation fusion discards are the same quantity. Proposition 1 is the theoretical carrier: fusion is lossless exactly when label information is recoverable from local marginals, and provably falls short otherwise.

What would settle it

A concrete falsifier for the theoretical claim: construct a dataset whose label is fully determined by a cross-cut correlator with uninformative local marginals; if late fusion does not fall to chance while reconstruction stays accurate, Proposition 1 fails. The paper's boundary experiment already shows the predicted collapse, so the sharper open test is for the diagnostic: measure $S_A$ on trained coupled circuits across widths and test whether it rank-orders fusion gaps; the paper's cross-scale result (Spearman $-0.09$ at $n=12$) suggests it will not.

Watch

Extended reading notes

Core claim

The central claim is that for machine learning, exact reconstruction of a cut circuit is, in most tested tasks, a wasted resource. Reconstruction rewrites each cross-register gate in operator-Schmidt form and sums quasiprobability branches to reproduce the uncut circuit's expectation values, paying a sampling overhead that grows exponentially in the number of cuts. The paper shows that a fusion head $f_\theta(m_A, m_B)$ trained on the local measurement features of independently trained subcircuits matches this exact readout within 0.04 on every controlled sweep and every classical benchmark, is more robust to finite-shot and device noise because it never pays the $\gamma^2$ variance amplification of the signed sum, and fails exactly when the label information lives in a cross-cut correlator with uninformative local marginals. The locality condition (Proposition 1) states this precisely: if the Bayes-optimal predictor depends on the joint state only through the local marginals, a sufficiently expressive fusion head attains the Bayes risk; otherwise Fano's inequality lower-bounds fusion's excess risk by the residual label information in the discarded correlator. A tunable quantumness dial $Q$ retains a variable fraction of the reconstruction sum's coefficient-1-norm mass, so $Q=0$ is pure fusion and $Q=1$ is exact reconstruction, and the cut-entanglement entropy $S_A$ of the trained coupled circuit predicts where along the dial a task needs to sit.

Load-bearing premise

The load-bearing premise is that you can train the full uncut circuit that the diagnostic measures; at the scale where reconstruction becomes infeasible, that full circuit is exactly what you no longer have, and the paper's own cross-scale test found the diagnostic stops predicting the fusion gap at larger sizes.

Editorial extensions

If this is right

  • Circuit-cutting QML pipelines should treat reconstruction as a tunable budget rather than a fixed requirement: the Q-dial lets a practitioner locate the cheapest setting within 0.02 of full accuracy without paying the full $\gamma^{2k}$ sweep.
  • At scales where reconstruction is infeasible ($n=20$, $k=10$, overhead $3.5\times 10^9$), independent late fusion still trains to well above chance, so cutting-based training can outrun the exponential wall.
  • Under finite shot budgets and device noise, the cheap fusion readout is not merely cheaper but more accurate than reconstruction; the gap only closes in the many-shots limit.
  • A fusion–reconstruction tie is an audit result: it indicates that the task's label information is recoverable from local marginals, so the pipeline gains nothing from its cross-cut quantum correlation.
  • Fusion's regime of validity (local information) overlaps where tuned classical models also match the quantum model, clarifying that any future quantum advantage for fusion would have to live in the entangled-data regime where fusion fails.

Reading between the lines

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

  • An implicit consequence the authors do not push: the fusion–reconstruction agreement test can serve as a screening tool for quantum-advantage claims—if a QML model's edge over classical ML depended on cross-cut correlators, late fusion would fail exactly there, so fusion success is evidence that the modeled task carries no such label information.
  • The conjectured gap bound $R_{\text{fusion}} - R^\star \le C \sum_j (1 - 2^{-S_A^{(j)}})$ is stated as a target; a direct test would be to construct tasks with known residual label information $\epsilon = I(Y;C|M)$ and check whether the fusion excess risk tracks $h_2^{-1}(\epsilon)$ and the entanglement envelope, which the current 104-run grid only loosely supports ($R^2 = 0.10$ for a linear fit).
  • Because the diagnostic requires a trained coupled circuit, a train-free proxy—such as entanglement measured on the initial circuit or estimated from single-subcircuit statistics—would be needed to make the method self-characterizing at scale; the paper's cross-scale transfer test (Spearman $-0.09$ at $n=12$) suggests the diagnostic is at best a within-scale indicator.
  • The architecture of independently trained subcircuits with a classical fusion head is a natural deployment pattern for heterogeneous or noisy quantum hardware: registers of different sizes, topologies, or noise profiles can be trained separately and combined without cross-device quantum links.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper proposes replacing exponential-cost quasiprobability reconstruction in circuit-cutting QML with late fusion: each subcircuit is trained independently and a small classical head combines their measurement features. It introduces a quantumness dial Q interpolating between pure fusion and full reconstruction, and a cut-entanglement diagnostic S_A intended to indicate how much reconstruction a task needs. On synthetic and standard datasets, with tuned classical baselines included, late fusion matches reconstruction accuracy within small margins at far lower sampling overhead and with greater robustness to finite-shot and device noise. Controlled entangled-data experiments locate a boundary where fusion fails, and scale-up experiments show fusion continuing to train at widths where the 9^k reconstruction overhead is infeasible. The paper provides paired per-seed statistics with bootstrap CIs, a qiskit-addon-cutting verification of the exponential overhead, a limited live-hardware check, and an unusually honest limitations section.

Significance. If the empirical match holds, late fusion is a practically valuable simplification for circuit-cutting QML: it replaces the dominant exponential sampling cost with a linear-cost learned combination and shifts the question to whether task information is local. The inclusion of tuned classical baselines, the explicit non-claim of quantum advantage, and the pairing of every headline comparison with bootstrap CIs and Wilcoxon tests are strengths. The principal weakness is that the 'self-characterizing' diagnostic is validated only at fixed width n=4 and presupposes the trained coupled circuit, so the title's question is answered only in the small-scale regime where exact reconstruction is already affordable; the abstract's characterization of the method as 'self-characterizing' therefore overreaches the evidence.

major comments (3)
  1. [Abstract / Contributions; Method (Cut-entanglement diagnostic); Appendix A; Limitations (iv)] The central 'self-characterizing' claim is not supported at the scale where the decision matters. S_A is defined as the entanglement entropy of the trained coupled circuit (Method, Cut-entanglement diagnostic), and the paper itself states that it carries no signal at initialization or after five optimizer iterations. Appendix A's cross-scale transfer test finds that S_A measured at n=8 does not rank-order fusion gaps at n=12 (Spearman -0.09, not significant), and Limitations (iv) concedes that train-free proxies are open. The only positive validation is the fixed-n=4 dense grid (Spearman 0.59). Since at larger n the trained coupled circuit is exactly the object whose reconstruction cost is exponential, the abstract's 'self-characterizing alternative' overstates what is demonstrated; the paper should either provide a cheap computable proxy or explicitly restrict the diagnostic claim to a small-scale, post-hoc characterization.
  2. [Algorithm 2 (Q-dial readout)] The quantumness dial as presented is not a deployable instrument for large circuits: Algorithm 2 requires branch states and coefficients {c_mu} from the trained coupled circuit, and the fusion head is retrained per Q value. In the regime where reconstruction is infeasible (n=16 and n=20 in the scale-up study), these objects are unavailable, so the 'tunable reconstruction budget' is an analysis tool for small-scale studies rather than a practical instrument. The paper should state this restriction wherever the dial is introduced and in the Contributions bullet, or reframe the dial accordingly.
  3. [Results (Independent training); Table D (Paired statistics)] At the alpha=1 parity point, independent fusion is statistically significantly worse than reconstruction: paired gap -0.037 with 95% CI [-0.057,-0.019] and p=0.008. This is within the abstract's stated 0.04 margin, but the abstract's wording 'matches full reconstruction accuracy within 0.04 at every point of the controlled sweep' should be qualified so that readers do not infer losslessness at the boundary, particularly because this is precisely the theoretically expected failure point. The same care should be applied to the n=16 and n=20 scale-up cells, where no reconstruction reference exists and accuracy is anchored only by a generator ceiling and a classical baseline.
minor comments (4)
  1. [Method (Cut-entanglement diagnostic)] The text reports that 'partialing out alpha leaves rho=0.60', but partial Spearman correlation is not a standard statistic; the paper should state the exact procedure (e.g., partial correlation on ranks or regression residuals) used to obtain this value.
  2. [Contributions bullet] There are missing spaces in the rendered text, for example 'aregimemap' and 'quantumnessdial'; these should be corrected in the final version.
  3. [Code availability] The paper states that 'code and results available from the authors on request'; for a computational paper of this kind, an archival repository link would materially improve reproducibility.
  4. [Table 5 (Paired statistics)] The paper notes that no multiple-comparison correction is applied over 34 comparison rows; this is acceptable as exploratory, but the main-text claims should continue to emphasize the bootstrap CIs as the primary inference, as they do.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central fusion-vs-reconstruction comparison is empirical and anchored to external reconstruction baselines; prior self-citation is motivational only, and the diagnostic's scalability caveat is a limitation, not a circular step.

full rationale

The paper's claimed result—that independently trained late fusion matches full reconstruction on tasks with local information at exponentially lower cost—is established by controlled experiments against an exact statevector reconstruction reference (B1 coincides with B0; Eq. 1), across independent seeds, with paired bootstrap CIs and Wilcoxon tests. The fusion head is trained on the task loss, not derived from the reconstruction output; the Q-dial is defined as an interpolation but its knee is measured empirically; the cut-entanglement diagnostic S_A is an empirical correlate (Spearman 0.59 on the dense grid) of the fusion gap, not a quantity defined in terms of the claimed outcome. The only fitted object is an explicitly conjectural gap bound with a loose, task-dependent constant C≈10, which the paper itself labels as future theory rather than an established result. The self-citation to DistributedEstimator is used only to motivate the problem (reconstruction as dominant runtime cost) and is not load-bearing for the main claims. Appendix A's honest caveat that S_A 'presupposes a trained coupled circuit' and the failed cross-scale transfer test (Spearman −0.09) limit the deployability of the diagnostic and weaken the 'self-characterizing' wording, but they are scalability/validity limitations, not circular reductions. No step takes the claimed result as an input and renames it as an output; the derivation chain is self-contained against external reconstruction and classical baselines.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The main auxiliary constructs are the Q-dial and the S_A diagnostic, which are settings and measurements rather than postulates. The only fitted constant is the explicitly conjectural gap-bound C. All other assumptions are standard math or stated domain assumptions from the circuit-cutting literature.

free parameters (2)
  • Gap-bound constant C = about 10
    In the conjectured monotone envelope R_fusion - R_star <= C sum(1 - 2^{-S_A}), C about 10 is fitted to the 104-run dense grid and holds for 93% of runs; explicitly stated as a target for future theory, not used in the central claim.
  • Knee tolerance = 0.02
    The Q-dial knee is defined as the smallest Q within 0.02 of the Q=1 accuracy; this threshold is chosen by hand and affects reported knee values and the diagnostic correlation.
assumptions (4)
  • standard math Operator-Schmidt decomposition of each cross gate and the product-observable reconstruction formula (Eq. 1).
    The exact reconstruction baseline and the Q-dial both rest on this expansion of the uncut circuit's observables into a fixed signed sum over branch states.
  • domain assumption Reconstruction sampling overhead is exponentially lower-bounded by entanglement cost across the cut (Jing, Zhu, and Wang 2025).
    This cited bound motivates using cut entanglement S_A as a predictor of how much reconstruction a task needs; the paper does not reprove it.
  • standard math Bayes optimality, Fano's inequality, and universal approximation by one-hidden-layer MLPs.
    Appendix A's locality condition and converse lower bound rely on these to connect fusion's feature set to achievable risk.
  • domain assumption The hand-verified statevector gate cut reproduces the uncut circuit to below 1e-10, so B0 and B1 coincide.
    The gold-standard comparison treats exact reconstruction as equivalent to the ideal uncut circuit; this is asserted for the controlled simulator in Experimental Setup.

how reviews work

0 comments
Cite this review

Pith. "Pith review of How Much Reconstruction Does Quantum Machine Learning Need? Late Fusion of Independently Trained Quantum Subcircuits." pith.science (2026). https://pith.science/paper/CTGS64O2

@misc{pith2026260805595,
  author       = {Pith},
  title        = {Pith review of: How Much Reconstruction Does Quantum Machine Learning Need? Late Fusion of Independently Trained Quantum Subcircuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CTGS64O2}},
  note         = {Machine review of arXiv:2608.05595}
}
abstract

Circuit cutting lets a large quantum neural network (QNN) run as independent subcircuits on small devices, but rebuilding its outputs by reconstruction carries a classical sampling overhead exponential in the number of cuts - the dominant runtime cost in prior work. We ask whether, for machine-learning tasks, this step is necessary, and replace it with late fusion: each subcircuit is trained and measured independently, and a small classical head combines their outputs - a linear-cost, decision-level combination borrowed from multimodal learning. To characterize the trade-off we introduce a quantumness dial $Q$, a tunable reconstruction budget interpolating from pure fusion to full reconstruction, and a cut-entanglement diagnostic that indicates how much reconstruction a task needs (Spearman $\rho=0.59$ over $104$ runs). Across synthetic and standard datasets, independently trained late fusion matches full reconstruction accuracy within $0.04$ at every point of the controlled sweep and on every classical benchmark, at exponentially lower cost; it is also markedly more robust to shot and device noise. Controlled entangled-data experiments locate the boundary where fusion must fail. We do not claim advantage over classical machine learning - consistent with recent benchmarking, quantum offers no accuracy edge on these datasets. Late fusion is thus an efficient, noise-robust, self-characterizing alternative to reconstruction for circuit-cutting QML.

Figures

Figures reproduced from arXiv: 2608.05595 by the authors.

Figure 1
Figure 1. (a) Cost crossover on the controlled simulator: reconstruction overhead (red) grows exponentially with the number [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Independent training (circles) matches reconstruc [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Noise robustness. (a) Under finite shots, late fusion (blue) is more accurate than reconstruction (red), which pays a [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Q-dial accuracy/cost frontier (mean±std, 10 seeds): accuracy saturates well before full reconstruction, so most of the γ 2k overhead yields little additional accuracy; low-cost fluctuations are within seed noise. Against learned truncation. A Marshall-style read￾out (M…
Figure 6
Figure 6. Figure 6: Scaling. (a) Late fusion (blue) keeps training as cuts and width grow to [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 15 canonical work pages

  1. [1]

    Abbas, A.; Schuld, M.; and Petruccione, F. 2020. On quantum ensembles of quantum classifiers. Quantum Machine Intelligence, 2(1): 6. ArXiv:2001.10833

    work page Pith review arXiv 2020

  2. [2]

    K.; Hossain, M

    Atrey, P. K.; Hossain, M. A.; El Saddik, A.; and Kankanhalli, M. S. 2010. Multimodal fusion for multimedia analysis: a survey. Multimedia Systems, 16(6): 345--379

    work page 2010

  3. [3]

    Baltru s aitis, T.; Ahuja, C.; and Morency, L.-P. 2019. Multimodal machine learning: A survey and taxonomy. IEEE Transactions on Pattern Analysis and Machine Intelligence, 41(2): 423--443. ArXiv:1705.09406

    arXiv 2019

  4. [4]

    Bowles, J.; Ahmed, S.; and Schuld, M. 2024. Better than classical? The subtle art of benchmarking quantum machine learning models. arXiv preprint arXiv:2403.07059

    arXiv 2024

  5. [5]

    Brenner, L.; Piveteau, C.; and Sutter, D. 2025. Optimal wire cutting with classical communication. IEEE Transactions on Information Theory, 71(10): 7742--7752. ArXiv:2302.03366

    arXiv 2025

  6. [6]

    Fisher, R. A. 1936. The use of multiple measurements in taxonomic problems. Annals of Eugenics, 7(2): 179--188

    work page 1936

  7. [7]

    Huang, H.-Y.; Kueng, R.; and Preskill, J. 2020. Predicting many properties of a quantum system from very few measurements. Nature Physics, 16: 1050--1057

    work page 2020

  8. [8]

    Jing, M.; Zhu, C.; and Wang, X. 2025. Circuit knitting facing exponential sampling-overhead scaling bounded by entanglement cost. Physical Review A, 111: 012433. ArXiv:2404.03619

    arXiv 2025

Show all 27 references
  1. [9]

    Kawase, Y. 2024. Distributed quantum neural networks via partitioned features encoding. Quantum Machine Intelligence, 6: 15. ArXiv:2312.13650

    2024 arXiv

  2. [10]

    Kelly, M.; Longjohn, R.; and Nottingham, K. 2023. UCI Machine Learning Repository. https://archive.ics.uci.edu

    2023

  3. [11]

    LeCun, Y.; Bottou, L.; Bengio, Y.; and Haffner, P. 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11): 2278--2324

    1998

  4. [12]

    Li, Q.; Huang, Y.; Hou, X.; Li, Y.; Wang, X.; and Bayat, A. 2024. Ensemble-learning error mitigation for variational quantum shallow-circuit classifiers. Physical Review Research, 6: 013027. ArXiv:2301.12707

    2024

  5. [13]

    J.; Bromley, T

    Lowe, A.; Medvidovi \'c , M.; Hayes, A.; O'Riordan, L. J.; Bromley, T. R.; Arrazola, J. M.; and Killoran, N. 2023. Fast quantum circuit cutting with randomized measurements. Quantum, 7: 934. ArXiv:2207.14734

    2023 arXiv

  6. [14]

    Marchisio, A.; Sychiuco, E.; Kashif, M.; and Shafique, M. 2025. Cutting is all you need: Execution of large-scale quantum neural networks on limited-qubit devices. In IEEE International Conference on Quantum Artificial Intelligence (QAI), 330--336. ArXiv:2412.04844

    2025 arXiv

  7. [15]

    C.; Gyurik, C.; and Dunjko, V

    Marshall, S. C.; Gyurik, C.; and Dunjko, V. 2023. High-dimensional quantum machine learning with small quantum computers. Quantum, 7: 1078. ArXiv:2203.13739

    2023 arXiv

  8. [16]

    R.; Boixo, S.; Smelyanskiy, V

    McClean, J. R.; Boixo, S.; Smelyanskiy, V. N.; Babbush, R.; and Neven, H. 2018. Barren plateaus in quantum neural network training landscapes. Nature Communications, 9: 4812

    2018

  9. [17]

    McFarthing, S.; Pillay, A.; Sinayskiy, I.; and Petruccione, F. 2024. Classical ensembles of single-qubit quantum variational circuits for classification. Quantum Machine Intelligence, 6(2): 81. ArXiv:2302.02964

    2024 arXiv

  10. [18]

    Mitarai, K.; and Fujii, K. 2021. Constructing a virtual two-qubit gate by sampling single-qubit operations. New Journal of Physics, 23(2): 023021. ArXiv:1909.07534

    2021 arXiv

  11. [19]

    N.; Nguyen, P

    Nguyen, T.; Hoang, T. N.; Nguyen, P. L.; Vu, H. L.; and Thang, T. C. 2025. Expressive and scalable quantum fusion for multimodal learning. arXiv preprint arXiv:2510.06938

    2025

  12. [20]

    W.; Ozols, M.; and Wu, X

    Peng, T.; Harrow, A. W.; Ozols, M.; and Wu, X. 2020. Simulating large quantum circuits on a small quantum computer. Physical Review Letters, 125(15): 150504. ArXiv:1904.00102

    2020 arXiv

  13. [21]

    Piveteau, C.; and Sutter, D. 2024. Circuit knitting with classical communication. IEEE Transactions on Information Theory, 70(4): 2734--2745. ArXiv:2205.00016

    2024 arXiv

  14. [22]

    Schuld, M.; and Petruccione, F. 2018. Quantum ensembles of quantum classifiers. Scientific Reports, 8: 2772. ArXiv:1704.02146

    2018 arXiv

  15. [23]

    N.; and Buyya, R

    Singh, P.; Toosi, A. N.; and Buyya, R. 2026. DistributedEstimator: Distributed training of quantum neural networks via circuit cutting. arXiv preprint arXiv:2602.16233

    2026 arXiv

  16. [24]

    G.; Worring, M.; and Smeulders, A

    Snoek, C. G.; Worring, M.; and Smeulders, A. W. 2005. Early versus late fusion in semantic video analysis. In Proceedings of the 13th Annual ACM International Conference on Multimedia, 399--402

    2005

  17. [25]

    Tang, W.; Tomesh, T.; Suchara, M.; Larson, J.; and Martonosi, M. 2021. CutQC: using small quantum computers for large quantum circuit evaluations. In Proceedings of the 26th ACM International Conference on Architectural Support for Programming Languages and Operating Systems (...

    2021 arXiv

  18. [26]

    S.; Scherer, D

    Ufrecht, C.; Herzog, L. S.; Scherer, D. D.; Periyasamy, M.; Rietsch, S.; Plinge, A.; and Mutschler, C. 2024. Optimal joint cutting of two-qubit rotation gates. Physical Review A, 109: 052440. ArXiv:2312.09679

    2024 arXiv

  19. [27]

    Xiao, H.; Rasul, K.; and Vollgraf, R. 2017. Fashion-MNIST: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747

    2017 arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.