Pith. sign in

REVIEW 3 major objections 4 minor 47 references

Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Randomized smoothing gives parameterized quantum circuit classifiers a provable noise-resilience certificate, and an evolutionary-strategy objective can train circuits to maximize it.

desk verdict The certificate is real but it certifies a majority-vote classifier, while the paper's experiments and deployment use an averaged-probability classifier; the 'provably' claim therefore doesn't apply to what was actually evaluated. read the letter →

arxiv 2505.18478 v1 pith:MXYU6AWD submitted 2025-05-24 quant-ph cs.LGphysics.comp-ph

Lucas Tecot , Di Luo , Cho-Jui Hsieh This is my paper
classification quant-phcs.LGphysics.comp-ph MSC 81P6868T05 PACS 03.67.Lx
keywords randomizedsmoothingparameterizedquantumcircuitscertifiedrobustnessparameternoiseevolutionarystrategiesphaseclassificationvariationalalgorithmsnoise-resilienttraining
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

The paper claims that randomized smoothing—sampling a classifier's parameters under Gaussian noise—transfers from classical machine learning to parameterized quantum circuit (PQC) classifiers, yielding a provable guarantee: if the smoothed circuit returns class $c_a$ at parameters $\theta$, then it still returns $c_a$ at any noisy parameters $\theta+\delta$ whose noise satisfies $\|\delta \oslash \sigma\|_2 < \frac{1}{2}(\Phi^{-1}(p_A)-\Phi^{-1}(p_B))$. The size of that certified ellipsoid depends only on the top two smoothed class probabilities, so training can enlarge it. The paper shows that the same objective used by evolutionary strategies—specifically sNES—maximizes this margin, so robustness certificates come with minimal changes to standard variational training. Demonstrated on a 12-qubit cluster-phase classification task, the smoothed classifier matches clean accuracy under small noise and beats a conventionally trained circuit under larger parameter noise. If correct, this provides near-term NISQ (noisy intermediate-scale quantum) practitioners a theory-backed way to make quantum classifiers tolerate instrumentation error.

What carries the argument

The machine doing the work is the smoothed PQC classifier $G_\sigma$ combined with the norm bound $\|\delta \oslash \sigma\|_2 < \frac{1}{2}(\Phi^{-1}(p_A)-\Phi^{-1}(p_B))$. The proof adapts the classical randomized-smoothing half-space argument to quantum classifiers: a likelihood-ratio lemma shows that probability mass under a Gaussian at $\theta$ can be transported to the Gaussian at $\theta+\delta$, and the bound follows by comparing the top class's lower probability with the runner-up's upper probability. Geometrically, the certificate is a hyper-ellipsoid in parameter space, with semi-axis lengths $\sigma_i$ times a common margin factor, so the paper's metrics—certified-area geometric mean, semi-axis average, and semi-axis standard deviation—all read off this ellipsoid. On the training side, the same $\sigma$ appears in the evolutionary-strategy search distribution, so maximizing the Gaussian-smoothed margin through sNES is a direct way to grow the ellipsoid.

What would settle it

Take a trained smoothed PQC on the 12-qubit cluster Hamiltonian, fix a test point where the smoothed classifier predicts class $c_a$ with certificate radius $R = \frac{1}{2}(\Phi^{-1}(p_A)-\Phi^{-1}(p_B))$ computed from Clopper-Pearson bounds, then evaluate the smoothed classifier at $\theta+\delta$ for a large collection of $\delta$ draws with $\|\delta \oslash \sigma\|_2 = 0.99R$. If the prediction ever changes, the central claim is false; if the certificates were computed with point estimates instead of confidence bounds, apparent violations at the nominal radius should disappear once $M$ and the confidence level are fixed.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the classical randomized-smoothing certificate is valid for parameterized quantum circuit classifiers. A smoothed PQC classifier $G_\sigma(\theta,x)$ is built by repeatedly sampling Gaussian parameter perturbations $\epsilon \sim \mathcal{N}(0,\Sigma)$ and taking the majority class; Theorem 2.1 then states that if $G_\sigma(\theta,x)=c_a$ with top probability $p_A$ and runner-up $p_B$, the smoothed prediction at $\theta+\delta$ remains $c_a$ for every $\delta$ satisfying $\|\delta \oslash \sigma\|_2 < \frac{1}{2}(\Phi^{-1}(p_A)-\Phi^{-1}(p_B))$. The paper further argues that the right-hand side of this inequality is exactly the kind of quantity an evolutionary-strategy objective optimizes, so training with sNES while choosing per-parameter variances $\sigma$ grows the certified ellipsoid. The experimental claim is that on the generalized cluster Hamiltonian, this training produces classifiers whose robustness to injected phase-shift noise is better than a conventionally trained PQC at comparable accuracy.

Load-bearing premise

The guarantee holds only when the smoothed classifier's top two class probabilities are replaced by valid statistical confidence bounds (a lower bound for the winning class, an upper bound for the runner-up), and the paper's experiments never state how many samples or what confidence level were used, so the reported certificates may rest on point estimates.

Editorial extensions

If this is right

  • A practitioner can add robustness certificates to an existing PQC classifier by re-evaluating the circuit under Gaussian parameter samples and training with an evolutionary-strategy objective; the deployment cost matches the multiple-evaluation pattern already required by probabilistic quantum measurements.
  • Certificates are anisotropic: each parameter gets its own $\sigma_i$, so the method exploits that some gate angles tolerate much larger noise than others rather than applying a uniform radius.
  • The training procedure is function-agnostic and can be layered on top of other error-mitigation techniques, and it can certify fixed parameters by optimizing only $\sigma$.
  • Experimental evidence on 12-qubit cluster-phase classification shows a smooth robustness–accuracy trade-off, with the smoothed classifier outperforming a standard classifier as injected parameter noise grows.

Reading between the lines

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

  • A rigorous deployment of the paper's certificate requires fixing the smoothing sample count $M$, a confidence level, and the confidence-bound method for $p_A$ and $p_B$, which the paper leaves to the practitioner and does not report in its experiments.
  • The optimized $\sigma$ vector can be read as a parameter-sensitivity map, suggesting applications in noise-aware ansatz selection, gate-angle freezing, or hardware-specific compilation.
  • The independent-variance Gaussian assumption is a modeling choice; testing the same argument under full-covariance smoothing or non-Gaussian noise would extend the certificate to correlated gate errors, and the cluster-phase benchmark is ready for that test.
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. The paper adapts non-uniform randomized smoothing to parameterized quantum circuit (PQC) classifiers. Theorem 2.1 certifies that a smoothed majority-vote classifier G_σ is stable to parameter perturbations δ satisfying ∥δ⊘σ∥_2 < (Φ^{-1}(pA)−Φ^{-1}(pB))/2, where pA and pB are the top-two class probabilities of the hard-label distribution under Gaussian parameter noise. The authors propose training θ and σ with an evolutionary-strategy (sNES) objective equal to this margin, regularize σ, and evaluate on 12-qubit cluster phase classification tasks. The paper reports certified ellipsoidal volumes, robustness-accuracy trade-offs, and a robustness-variance correlation.

Significance. The certificate in Theorem 2.1 is a valid extension of randomized smoothing to non-uniform Gaussian perturbations of PQC parameters, and the geometric metrics (semi-axis average, certified area geometric mean) are natural and useful. The theorem proof is mathematically sound given the imported Lemma 3 from [12]. However, the paper's central contribution—the training algorithm and the experimental validation of 'provable' robustness—is undermined by several load-bearing inconsistencies: the deployed classifier differs from the certified one, the ES training objective is degenerate as written, and the statistical confidence behind the reported certificates is never stated. These issues prevent the paper from supporting its main claims in its current form.

major comments (3)
  1. [Definition 2.3 vs. Table 1 and Section 3] The classifier certified by Theorem 2.1 is G_σ(θ,x)=argmax_z P(argmax_i C(θ+ε,x)_i = z), which is a majority vote over hard labels. The 'Deployed Model Use' box in Table 1, and the 'Smoothed Accuracy' definition in Section 3, instead compute p_i=(1/M)Σ_k C(θ+σ s_k,x)_i and output argmax_i p_i, i.e., the argmax of averaged class probabilities. These two decision rules are not equivalent: for two classes, if C_1−C_2 = +10 with probability 0.4 and −1 with probability 0.6, the majority vote returns class 2, while the averaged probabilities give p_1−p_2 = 3.4 > 0 and return class 1. Therefore the robustness guarantee in Theorem 2.1 does not apply to the model actually evaluated in Figures 1–2, and the reported certified areas are not valid certificates for that model.
  2. [Section 2.2, Eq. (13) and Table 1] The proposed ES objective is degenerate as written. O(θ+ε,x,y)=1/2(Φ^{-1}(pA)−Φ^{-1}(pB)) uses pA and pB from Theorem 2.1, which are probabilities over the Gaussian noise ε at the center θ; they do not depend on the particular sampled perturbation ε. Consequently, in Table 1 every f_k is identical for all k=1...λ. Sorting equal f_k produces an arbitrary order, and the utility-weighted sums for ∇θJ and ∇σJ become random vectors with zero expectation that do not estimate a gradient of the robust margin. The claimed 'natural connection to Evolutionary Strategies' therefore does not yield a working training algorithm for the certificate.
  3. [Section B.3 and experimental reporting] The theorem's guarantee is only valid if pA is replaced by a lower confidence bound and pB by an upper confidence bound at some level 1−δ'. The paper never states the number of smoothing samples M, the confidence level δ', or which specific estimator (e.g., Clopper–Pearson) produced the pA and pB values used in training or in the reported certified areas. Without this information, the 'provably noise-resilient training' claim is not operational, and the numerical certificates in Figure 2 are not tied to any stated probability.
minor comments (4)
  1. [Section 2.2, Eq. (13)] The left-hand side O(θ+ε,x,y) is written as a function of the sampled point θ+ε and label y, but the right-hand side depends only on θ, σ, and the class probabilities; the notation should be clarified to avoid implying that the ES sample affects the objective.
  2. [Section 4.2 and Figure 1] The text says the test dataset has 50 samples (Section 4), but the caption of Figure 1 refers to '20 test data points'; the discrepancy should be resolved.
  3. [Appendix C, Eq. (24)] The constant b in the likelihood ratio is written as −1/2∥δ⊙λ∥_1 − ∥x_i⊙λ∥_1, but x_i is not defined here; it should be the mean θ (or the notation should be made consistent).
  4. [Table 1] The sorting step 's'_k ← Sort all s_k w.r.t. f_k' is ambiguous when several f_k are equal, which is always the case under Eq. (13); a deterministic tie-breaking rule is needed for the algorithm to be well-defined.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the robustness certificate is proved from an external randomized-smoothing lemma, and the ES-training objective equals the certificate's margin only as a deliberate optimization target, not as an input to the proof.

full rationale

I walked the derivation chain. Theorem 2.1 is proved in Appendix C via Lemma C.1, which is explicitly stated to be a special case of Lemma 3 of Cohen et al. [12]; the half-space probability computations in Section C.1 are ordinary Gaussian algebra. The certificate's validity does not depend on how θ and σ are trained. The 'connection to ES' in Section 2.2 sets the objective O(θ+ε,x,y) = (Φ^{-1}(pA)-Φ^{-1}(pB))/2, which is exactly the right-hand side of the bound in Eq. (8); this is a direct optimization of the certified margin, not a case of predicting a fitted quantity, since the theorem holds for any parameters and the experiments report the bound itself. Reference [28] (Tecot and Hsieh) is cited for non-uniform randomized smoothing and Clopper-Pearson interval usage, but the load-bearing proof step cites external [12], so the self-citation is not load-bearing. Section B.3's honest statement that pA and pB must be estimated by confidence intervals, and the Table 1 deployed rule (averaged class probabilities) versus Definition 2.3's majority-vote smoothed classifier, are correctness and validity concerns rather than circularity. I therefore find no significant circularity; the low score reflects only a minor, non-load-bearing self-citation.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The certificate inherits standard randomized smoothing theory, and the training objective is the certificate margin itself; most free choices are hyperparameters and the learned sigma vector.

free parameters (3)
  • Per-parameter smoothing variance vector sigma = not reported numerically (learned)
    Central to the smoothed classifier and the certified ellipsoid; optimized by the ES training loop and regularized by a hyperparameter. The paper does not report final sigma values.
  • Variance regularization coefficient = swept, not fixed
    Controls the accuracy-robustness trade-off in Section 2.3; only the acceptable range is given in Section B.1.
  • ES hyperparameters (population lambda, learning rates eta_theta, eta_sigma, eta_r, initial sigma_0) = ranges in Section B.1
    Chosen by randomized sweep; results depend on them, though the paper claims a wide range works.
assumptions (4)
  • standard math Neyman-Pearson lemma for Gaussian likelihood ratio (Lemma 3 of Cohen et al. [12])
    Theorem 2.1 and Lemma C.1 rely on this external result; the paper does not re-derive it.
  • domain assumption The PQC classifier output C(theta,x)_i is an exact expectation value with no shot noise or measurement noise
    Definitions in Section 2.1 treat C as the ideal expectation; finite-shot noise is not modeled in the certificate.
  • domain assumption Parameter noise is additive and can be represented as a change delta in the parameter vector, with the smoothed classifier protocol used at deployment
    Definitions 2.2 and 2.3; the guarantee covers G_sigma, not the raw circuit, and assumes the smoothing protocol is actually executed.
  • standard math Concentration inequalities exist and are used to estimate pA and pB
    Section B.3 says Clopper-Pearson or similar should be used, but experiments do not report the resulting confidence level.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise." pith.science (2026). https://pith.science/paper/MXYU6AWD

@misc{pith2026250518478,
  author       = {Pith},
  title        = {Pith review of: Provably Robust Training of Quantum Circuit Classifiers Against Parameter Noise},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXYU6AWD}},
  note         = {Machine review of arXiv:2505.18478}
}
read the original abstract

Advancements in quantum computing have spurred significant interest in harnessing its potential for speedups over classical systems. However, noise remains a major obstacle to achieving reliable quantum algorithms. In this work, we present a provably noise-resilient training theory and algorithm to enhance the robustness of parameterized quantum circuit classifiers. Our method, with a natural connection to Evolutionary Strategies, guarantees resilience to parameter noise with minimal adjustments to commonly used optimization algorithms. Our approach is function-agnostic and adaptable to various quantum circuits, successfully demonstrated in quantum phase classification tasks. By developing provably guaranteed optimization theory with quantum circuits, our work opens new avenues for practical, robust applications of near-term quantum computers.

Figures

Figures reproduced from arXiv: 2505.18478 by the authors.

Figure 1
Figure 1. Accuracy over 20 test data points for a well-trained [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Phase classification for the generalized cluster Hamiltonian of 12 qubits, as outlined in Section [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 2 from Gil-Fuster et al. [ [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 19 canonical work pages

  1. [12]

    Cohen, Elan Rosenfeld, and J

    Jeremy M. Cohen, Elan Rosenfeld, and J. Zico Kolter. Certified Adversarial Ro- bustness via Randomized Smoothing, June

    work page

  2. [1]

    URL https:// github.com/qiskit-community/ qiskit-machine-learning/blob/stable/0.7/ docs/tutorials/11_quantum_convolutional_ neural_networks.ipynb

    The Quantum Convolution Neu- ral Network. URL https:// github.com/qiskit-community/ qiskit-machine-learning/blob/stable/0.7/ docs/tutorials/11_quantum_convolutional_ neural_networks.ipynb

    work page

  3. [2]

    Huggins, Ramis Movassagh, Dar Gilboa, and Jarrod R

    Amira Abbas, Robbie King, Hsin-Yuan Huang, William J. Huggins, Ramis Movassagh, Dar Gilboa, and Jarrod R. McClean. On quantum backpropagation, information reuse, and cheat- ing measurement collapse. May 2023. URL https://scirate.com/arxiv/2305.13362

    arXiv 2023

  4. [3]

    Natural evolutionary strategies for variational quantum computation

    Abhinav Anand, Matthias Degroote, and Alán Aspuru-Guzik. Natural evolutionary strategies for variational quantum computation. Machine Learning: Science and Technology, 2(4):045012, July 2021. ISSN 2632-2153. DOI: 10.1088/2632- 2153/abf3ac. URL https://dx.doi.org/10. 1088/2632-2153/abf3ac. Publisher: IOP Pub- lishing

    doi:10.1088/2632- 2021

  5. [4]

    Quantum machine learning.Nature, 549 (7671):195–202, September 2017

    Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, and Seth Lloyd. Quantum machine learning.Nature, 549 (7671):195–202, September 2017. ISSN 1476-

    work page 2017

  6. [5]

    Wang, Sepehr Ebadi, Marcin Kali- nowski, Alexander Keesling, Nishad Maskara, Hannes Pichler, Markus Greiner, Vladan Vuletić, and Mikhail D

    Dolev Bluvstein, Harry Levine, Giulia Semegh- ini, Tout T. Wang, Sepehr Ebadi, Marcin Kali- nowski, Alexander Keesling, Nishad Maskara, Hannes Pichler, Markus Greiner, Vladan Vuletić, and Mikhail D. Lukin. A quantum proces- sor based on coherent transport of entangled atom arrays. Nature, 604(7906):451–456, April

    work page

  7. [6]

    Evered, Alexan- dra A

    Dolev Bluvstein, Simon J. Evered, Alexan- dra A. Geim, Sophie H. Li, Hengyun Zhou, Tom Manovitz, Sepehr Ebadi, Madelyn Cain, MarcinKalinowski, DominikHangleiter, J.Pablo Bonilla Ataides, Nishad Maskara, Iris Cong, Xun Gao, Pedro Sales Rodriguez, Thomas Karolyshyn, Giulia Semeghini, Michael J. Gul- lans, Markus Greiner, Vladan Vuletić, and Mikhail D. Lukin...

    work page doi:10.1038/s4158 2024

  8. [7]

    Melko, and Simon Trebst

    Peter Broecker, Juan Carrasquilla, Roger G. Melko, and Simon Trebst. Machine learn- ing quantum phases of matter beyond the fermion sign problem. Scientific Re- ports, 7(1):8823, August 2017. ISSN 2045-

    work page 2017

Show all 47 references
  1. [8]

    Ben- jamin, Suguru Endo, William J

    Zhenyu Cai, Ryan Babbush, Simon C. Ben- jamin, Suguru Endo, William J. Huggins, Ying Li, Jarrod R. McClean, and Thomas E. O’Brien. Quantum Error Mitigation. Reviews of Mod- ern Physics, 95(4):045005, December 2023. ISSN 0034-6861, 1539-0756. DOI: 10.1103/RevMod- Phys.95.045005...

    2023 arXiv

  2. [9]

    Publisher: Nature Pub- lishing Group

    URL https://www.nature.com/articles/ s41586-023-06927-3. Publisher: Nature Pub- lishing Group

  3. [10]

    Juan Carrasquilla and Roger G. Melko. Ma- chine learning phases of matter.Nature Physics, 13(5):431–434, May 2017. ISSN 1745-2481. DOI: 10.1038/nphys4035. URL https://www. nature.com/articles/nphys4035. Publisher: Nature Publishing Group

    2017 doi

  4. [11]

    Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C

    M. Cerezo, Andrew Arrasmith, Ryan Babbush, Simon C. Benjamin, Suguru Endo, Keisuke Fujii, Jarrod R. McClean, Kosuke Mitarai, Xiao Yuan, Lukasz Cincio, and Patrick J. Coles. Variational Quantum Algorithms. Nature Reviews Physics, 3(9):625–644, August 2021. ISSN 2522-5820. DOI:1...

    2021 arXiv

  5. [13]

    Caro, Hsin-Yuan Huang, M

    Matthias C. Caro, Hsin-Yuan Huang, M. Cerezo, Kunal Sharma, Andrew Sornborger, Lukasz Cin- cio, and Patrick J. Coles. Generalization in quan- tum machine learning from few training data. Nature Communications, 13(1):4919, August

  6. [14]

    DOI: 10.1038/s41467- 022-32550-3

    ISSN 2041-1723. DOI: 10.1038/s41467- 022-32550-3. URL https://www.nature.com/ articles/s41467-022-32550-3. Publisher: Nature Publishing Group

    doi

  7. [15]

    A Quantum Approximate Optimiza- tion Algorithm, November 2014

    Edward Farhi, Jeffrey Goldstone, and Sam Gut- mann. A Quantum Approximate Optimiza- tion Algorithm, November 2014. URL http: //arxiv.org/abs/1411.4028. arXiv:1411.4028 [quant-ph]

    2014 arXiv

  8. [16]

    Understanding quantum ma- chine learning also requires rethinking general- ization, June 2023

    Elies Gil-Fuster, Jens Eisert, and Carlos Bravo-Prieto. Understanding quantum ma- chine learning also requires rethinking general- ization, June 2023. URL http://arxiv.org/ abs/2306.13461. arXiv:2306.13461 [cond-mat, physics:quant-ph, stat]

    2023 arXiv

  9. [17]

    The CMA Evolution Strategy: A Comparing Review

    Nikolaus Hansen. The CMA Evolution Strategy: A Comparing Review. 2006

    2006

  10. [18]

    Harrigan, Kevin J

    Matthew P. Harrigan, Kevin J. Sung, Matthew Neeley, Kevin J. Satzinger, Frank Arute, Ku- nal Arya, Juan Atalaya, Joseph C. Bardin, RamiBarends, SergioBoixo, MichaelBroughton, Bob B. Buckley, David A. Buell, Brian Bur- kett, Nicholas Bushnell, Yu Chen, Zijun Chen, Ben Chiaro, R...

    2021

  11. [19]

    Iris Cong, Soonwon Choi, and Mikhail D. Lukin. Quantum convolutional neural net- works. Nature Physics, 15(12):1273–1278, De- cember 2019. ISSN 1745-2473, 1745-2481. DOI: 10.1038/s41567-019-0648-8. URL https://www. nature.com/articles/s41567-019-0648-8

    2019 doi

  12. [20]

    Dalzell, Sam McArdle, Mario Berta, Przemyslaw Bienias, Chi-Fang Chen, An- drás Gilyén, Connor T

    Alexander M. Dalzell, Sam McArdle, Mario Berta, Przemyslaw Bienias, Chi-Fang Chen, An- drás Gilyén, Connor T. Hann, Michael J. Kas- toryano, Emil T. Khabiboulline, Aleksander Ku- bica, Grant Salton, Samson Wang, and Fernando G. S. L. Brandão. Quantum algorithms: A sur- vey of ...

    2023 arXiv

  13. [21]

    Bishop, Jerry M

    Nikolaj Moll, Panagiotis Barkoutsos, Lev S. Bishop, Jerry M. Chow, Andrew Cross, Daniel J. Egger, Stefan Filipp, Andreas Fuhrer, Jay M. Gambetta, Marc Ganzhorn, Abhinav Kandala, Antonio Mezzacapo, Peter Müller, Walter Riess, Gian Salis, John Smolin, Ivano Tavernelli, and Krist...

    2017 arXiv

  14. [22]

    Love, Alán Aspuru-Guzik, and Jeremy L

    Alberto Peruzzo, Jarrod McClean, Peter Shad- bolt, Man-Hong Yung, Xiao-Qi Zhou, Pe- ter J. Love, Alán Aspuru-Guzik, and Jeremy L. O’Brien. A variational eigenvalue solver on a photonic quantum processor.Nature Communi- cations, 5(1):4213, July 2014. ISSN 2041-1723. DOI: 10.103...

    2014 doi

  15. [23]

    Quantum Computing in the NISQ era and beyond

    John Preskill. Quantum Computing in the NISQ era and beyond. Quantum, 2:79, Au- gust 2018. DOI: 10.22331/q-2018-08-06-79. URL https://quantum-journal.org/papers/ q-2018-08-06-79/. Publisher: Verein zur Förderung des Open Access Publizierens in den Quantenwissenschaften

    2018 doi

  16. [24]

    Impact of Noise on the Resilience and the Security of Quantum Computing

    Abdullah Ash Saki, Mahabubul Alam, and Swaroop Ghosh. Impact of Noise on the Resilience and the Security of Quantum Computing. In 2021 22nd International Symposium on Quality Electronic Design (ISQED), pp. 186–191, April 2021. DOI: 10.1109/ISQED51717.2021.9424258. URL https://...

    2021

  17. [25]

    N. Klco, E. F. Dumitrescu, A. J. McCaskey, T. D. Morris, R. C. Pooser, M. Sanz, E. Solano, P. Lougovski, and M. J. Savage. Quantum- Classical Computation of Schwinger Model Dy- namics using Quantum Computers, March 2018. URL https://arxiv.org/abs/1803.03326v3

    2018 arXiv

  18. [26]

    Robustness Certificates for Sparse Adversarial Attacks by Randomized Ablation, November 2019

    Alexander Levine and Soheil Feizi. Robustness Certificates for Sparse Adversarial Attacks by Randomized Ablation, November 2019. URL https://arxiv.org/abs/1911.09272v1

    2019 arXiv

  19. [27]

    Svore, and Nathan Wiebe

    Maria Schuld, Alex Bocharov, Krysta M. Svore, and Nathan Wiebe. Circuit-centric quan- tum classifiers. Physical Review A , 101(3): 032308, March 2020. DOI: 10.1103/Phys- RevA.101.032308. URL https://link.aps. org/doi/10.1103/PhysRevA.101.032308. Pub- lisher: American Physical Society

    2020 doi

  20. [28]

    Ro- bustness Verification with Non-Uniform Ran- 8 domized Smoothing, 2021

    Lucas Matthew Tecot and Cho-Jui Hsieh. Ro- bustness Verification with Non-Uniform Ran- 8 domized Smoothing, 2021. URL https:// escholarship.org/uc/item/8ds207x6

    2021

  21. [29]

    Opti- mal quantum circuits for general two-qubit gates

    Farrokh Vatan and Colin Williams. Opti- mal quantum circuits for general two-qubit gates. Physical Review A, 69(3):032315, March

  22. [30]

    Optimal provable ro- bustness of quantum classification via quan- tum hypothesis testing

    Maurice Weber, Nana Liu, Bo Li, Ce Zhang, and Zhikuan Zhao. Optimal provable ro- bustness of quantum classification via quan- tum hypothesis testing. npj Quantum In- formation, 7(1):1–12, May 2021. ISSN 2056-6387. DOI: 10.1038/s41534-021-00410-

    2021 doi

  23. [31]

    Zico Kolter

    Hadi Salman, Mingjie Sun, Greg Yang, Ashish Kapoor, and J. Zico Kolter. Denoised Smooth- ing: A Provable Defense for Pretrained Classi- fiers, September 2020. URLhttp://arxiv.org/ abs/2003.01908. arXiv:2003.01908 [cs, stat]

    2020 arXiv

  24. [32]

    Provably Robust Deep Learn- ing via Adversarially Trained Smoothed Classi- fiers, January 2020

    Hadi Salman, Greg Yang, Jerry Li, Pengchuan Zhang, Huan Zhang, Ilya Razenshteyn, and Se- bastien Bubeck. Provably Robust Deep Learn- ing via Adversarially Trained Smoothed Classi- fiers, January 2020. URL http://arxiv.org/ abs/1906.04584. arXiv:1906.04584 [cs, stat]

    2020 arXiv

  25. [33]

    Natu- ral Evolution Strategies, June 2011

    Daan Wierstra, Tom Schaul, Tobias Glasmach- ers, Yi Sun, and Jürgen Schmidhuber. Natu- ral Evolution Strategies, June 2011. URLhttp: //arxiv.org/abs/1106.4487. arXiv:1106.4487 [cs, stat]

    2011 arXiv

  26. [34]

    Christopher J. Wood. Special Session: Noise Characterization and Error Mitigation in Near- Term Quantum Computers. In 2020 IEEE 38th International Conference on Computer Design (ICCD), pp. 13–16, October 2020. DOI: 10.1109/ICCD50377.2020.00016. URL https: //ieeexplore.ieee.org...

    2020

  27. [35]

    Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li

    Greg Yang, Tony Duan, J. Edward Hu, Hadi Salman, Ilya Razenshteyn, and Jerry Li. Ran- domized Smoothing of All Shapes and Sizes, July 2020. URL http://arxiv.org/abs/2002. 08118. arXiv:2002.08118 [cs, stat]

    2020 arXiv

  28. [36]

    Robust Quantum Gates against Correlated Noise in Integrated Quantum Chips

    Kangyuan Yi, Yong-Ju Hai, Kai Luo, Ji Chu, Libo Zhang, Yuxuan Zhou, Yao Song, Song Liu, Tongxing Yan, Xiu-Hao Deng, Yuanzhen Chen, and Dapeng Yu. Robust Quantum Gates against Correlated Noise in Integrated Quantum Chips. Physical Review Letters, 132(25):250604, June 2024. ISSN...

    2024 arXiv

  29. [38]

    Publisher: Nature Pub- lishing Group

    URL https://www.nature.com/articles/ s41534-021-00410-5. Publisher: Nature Pub- lishing Group

  30. [39]

    Kottmann, Thi Ha Kyaw, Bo Li, Alán Aspuru-Guzik, Ce Zhang, and Zhikuan Zhao

    Maurice Weber, Abhinav Anand, Alba Cervera- Lierta, Jakob S. Kottmann, Thi Ha Kyaw, Bo Li, Alán Aspuru-Guzik, Ce Zhang, and Zhikuan Zhao. Toward reliability in the NISQ era: Ro- bust interval guarantee for quantum measure- ments on approximate states. Physical Re- view Researc...

    2022 doi

  31. [40]

    West, Shu-Lok Tsang, Jia S

    Maxwell T. West, Shu-Lok Tsang, Jia S. Low, Charles D. Hill, Christopher Leckie, Lloyd C. L. Hollenberg, Sarah M. Erfani, and Muhammad Usman. Towards quantum enhanced adver- sarial robustness in machine learning. Na- ture Machine Intelligence, 5(6):581–589, May

  32. [46]

    If S ={z∈ Rd :λTz≤β} for someβ and P(h(X) = 1)≥P (X∈S), then P(h(Y ) = 1)≥P (Y ∈S)

  33. [47]

    (⊘ is hadamard division,·◦2 is element-wise square.) Proof

    If S ={z∈ Rd :λTz≥β} for someβ and P(h(X) = 1)≤P (X∈S), then P(h(Y ) = 1)≤P (Y ∈S) 11 Whereλ =δ⊘σ◦2. (⊘ is hadamard division,·◦2 is element-wise square.) Proof. This lemma is the special case of Lemma 3 in [12] whenX and Y are Gaussians with meansx and x +δ. By Lemma 3 in [12]...

  34. [2004]

    URL https://link.aps.org/doi/10.1103/ PhysRevA.69.032315

    DOI: 10.1103/PhysRevA.69.032315. URL https://link.aps.org/doi/10.1103/ PhysRevA.69.032315. Publisher: American Physical Society

    doi

  35. [2019]

    URL http://arxiv.org/abs/1902. 02918. arXiv:1902.02918 [cs, stat]

    1902 arXiv

  36. [2022]

    DOI: 10.1038/s41586- 022-04592-6

    ISSN 1476-4687. DOI: 10.1038/s41586- 022-04592-6. URL https://www.nature.com/ articles/s41586-022-04592-6. Publisher: Nature Publishing Group

    doi

  37. [2023]

    DOI:10.1038/s42256-023- 00661-1

    ISSN2522-5839. DOI:10.1038/s42256-023- 00661-1. URL http://arxiv.org/abs/2306. 12688. arXiv:2306.12688 [quant-ph]

    arXiv

  38. [2322]

    URL https://www.nature.com/articles/ s41598-017-09098-0

    DOI: 10.1038/s41598-017-09098-0. URL https://www.nature.com/articles/ s41598-017-09098-0. Publisher: Nature Publishing Group

    doi

  39. [4687]

    URLhttps:// www.nature.com/articles/nature23474

    DOI:10.1038/nature23474. URLhttps:// www.nature.com/articles/nature23474. Pub- lisher: Nature Publishing Group

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.