Pith. sign in

REVIEW 4 major objections 4 minor 71 references

Quantum Learning with Tunable Loss Functions

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that tilted empirical risk minimization can be lifted to quantum process learning, with sample complexity close to ordinary quantum ERM for small tilt, plus new PAC and agnostic guarantees.

desk verdict QTERM is a sensible definition, but the main theorem rests on a false MGF identity, so the sample-complexity bounds are unproven. read the letter →

arxiv 2508.21369 v1 pith:265UTB5X submitted 2025-08-29 quant-ph

Yixian Qiu , Lirandë Pira , Patrick Rebentrost This is my paper
classification quant-ph MSC 68Q3281P6881P45
keywords quantumlearningtheorytiltedempiricalriskminimizationsamplecomplexityPACagnosticprojector-valuedfunctionsthresholdsearchcoveringnumbers
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 extends tilted empirical risk minimization (TERM), a log-sum-exp loss controlled by a tilt parameter γ, to learning from quantum data, calling the result QTERM. Its central claim is that QTERM is learnable: with enough single-copy product states, an algorithm can find a projector-valued hypothesis whose tilted risk is near-minimal, with sample complexity that for small |γ| matches quantum empirical risk minimization up to log factors. The same analysis yields a PAC uniform-convergence bound for classical TERM and an agnostic guarantee for quantum hypothesis selection when the class is not too complex. If these bounds hold, QTERM supplies a tunable implicit-regularization strategy for quantum process learning, interpolating between mean-loss and outlier-emphasizing behavior.

What carries the argument

The central object is the tilted empirical risk Rγ(h) = 1 − (1/γ) log( (1/n) Σ exp(γ Tr[ρ_i Π_i^{(c)}]) ), a log-sum-exp loss with tilt hyperparameter γ: γ→0 recovers the average loss, while large |γ| shifts weight toward high-loss samples. The argument is carried by a two-block measurement protocol: blockwise concentration of exponentiated single-copy outcomes using Hoeffding-type inequalities for multiple populations, followed by a quantum threshold-search algorithm with gentle two-outcome measurements, which checks whether a projector list's average exceeds a threshold without destroying the product state. The small-tilt condition |γ|<ε controls the log-Lipschitz factor and the variance o

What would settle it

Take any qubit state ρ and projector Π with p=Tr[ρΠ]∈(0,1), and measure once to get Y∈{0,1}. The paper's Proposition 15 predicts E[e^{γY}]=e^{γp}; direct evaluation gives E[e^{γY}]=1+(e^γ−1)p. For γ≠0 these differ, and since Theorem 7's Step 1 (Eq. (16)) and check step (Eq. (24)) invoke Proposition 15, this calculation is the concrete point that would need to be corrected for the stated sample complexity to stand.

Watch

Extended reading notes

Core claim

The paper's core claim is that tilted empirical risk minimization can be defined for projector-valued hypotheses by Rγ(h) = 1 − γ⁻¹ log( n⁻¹ Σ exp(γ Tr[ρ_i Π_i^{(c)}]) ), and that this QTERM rule is learnable. Its main theorem asserts that for |γ| in (0, ε), a quantum threshold-search algorithm outputs an index c* and a risk estimate with Pr(|μ̂_c*(γ) − max_c μ_c(γ)| ≥ ε or |μ̂_c*(γ) − μ_c*(γ)| ≥ ε) ≤ δ using n = (1/ε²) log(1/δ) log²(1/ε) × O(max((e^{|γ|}−1)²/γ² log(m/δ ...), (log m + C1)²)) samples. The paper also claims a PAC bound for classical TERM, Pr(∃h: |R(h)−Rγ(h)|≥ε) ≤ 8Γ(2n, ε/8, G_{F,L,γ}) exp(−nε²/32|e^γ−1|²), and an agnostic guarantee for quantum hypothesis selection when coveri

Load-bearing premise

The load-bearing premise in the main sample-complexity theorem is that for a single-copy two-outcome measurement Y with expectation p, E[e^{γY}] = e^{γp}; the correct identity is E[e^{γY}] = 1+(e^γ−1)p, so the concentration step in Theorem 7 rests on an equality that does not hold.

Editorial extensions

If this is right

  • For small |γ|, QTERM learns projector-valued quantum hypotheses with essentially the same sample complexity as QERM, so the extra expressive power of a tilted loss costs little in data.
  • Classical TERM gets a PAC bound: for fixed γ, the probability that some hypothesis's tilted empirical risk deviates from its population risk by ε decays exponentially in n, governed by the covering number of the exponentiated loss class.
  • Agnostic quantum hypothesis selection succeeds under a subexponential covering-number condition, returning a hypothesis within 3ε of the best achievable tilted risk even when the class contains no perfect hypothesis.
  • QTR, defined as γ⁻¹ log Tr(e^{γH}ρ), connects tilted risk to partition functions and Rényi-type divergences, giving a formal handle for Hamiltonian learning with tunable emphasis.
  • As γ→0, QTERM reduces to QERM, so the framework is a continuous generalization rather than a separate model.

Reading between the lines

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

  • A reader checking the proof's primitive will find that for a single-copy binary outcome Y with p=E[Y], E[e^{γY}] = 1+(e^γ−1)p, not e^{γp}; Proposition 15 uses the latter, and Theorem 7's concentration steps (Eq. (16), Eq. (24)) rely on it, so the stated sample complexity needs a corrected argument.
  • Because tilted losses are nonlinear functions of per-sample expectations, a practical QTERM implementation may need to split samples or use classical shadows-style estimators rather than replacing each outcome by its exponentiated expectation.
  • QTR's free-energy form suggests testing tilted-risk bounds empirically by preparing Gibbs states and comparing the γ-dependence of the generalization gap against exp(−nε²/|e^γ−1|²).
  • The theorems only cover γ within ε of 0; an open, testable question is whether the exponential-in-|γ| sample penalty is intrinsic or an artifact of the log-Lipschitz proof technique.
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

4 major / 4 minor

Summary. The paper introduces QTERM, a tilted empirical risk minimization framework for learning projector-valued functions from classical-quantum data, extending QERM [31] and classical TERM [7,8]. The main claimed contributions are: (1) a sample-complexity bound for QTERM learnability (Theorems 1 and 7), (2) a PAC generalization bound for classical TERM (Theorem 8), and (3) agnostic learning guarantees for QTERM (Theorem 10). The proofs use gentle measurements, quantum threshold search, concentration inequalities, and covering numbers. The central technical bridge is Proposition 15, which is used to concentrate the exponentiated single-copy measurement outcomes that define the tilted empirical risk.

Significance. If the results were correct, the paper would provide a meaningful extension of QERM to tilted losses, with sample complexity matching QERM up to logarithmic factors, and would contribute new PAC bounds for classical TERM. The paper also introduces a separate quantum tilted risk (QTR) definition with connections to free energy and Rényi entropies. However, the central concentration claim (Proposition 15) is false, and this invalidates the main sample-complexity theorem and the agnostic learning theorem that depends on it. The PAC bound for classical TERM also has scaling inconsistencies in its symmetrization and cover-lifting steps. The paper does not provide reproducible code or machine-checked proofs; its value rests on the analytic derivations, which are not currently sound.

major comments (4)
  1. [Proposition 15 (Appendix IX A), used in Theorem 7 Step 2 (Eq. 24)] Proposition 15 asserts that for a two-outcome measurement bit Y_s with values in {0,1} and E[Y_s]=Tr[ρ_s Π_s], the exponentiated outcome X_s = e^{cY_s} satisfies E[X_s] = e^{c E[Y_s]}. This is false. The correct moment generating function is E[e^{cY_s}] = 1 + (e^c - 1) E[Y_s], not e^{cE[Y_s]}. The equality would require Y_s to be deterministic. This error is load-bearing: Eq. (24) of Theorem 7 invokes Proposition 15 to obtain the failure exponent 2e^{-l|γ|ε²/(12(e^{|γ|}-1))}, which determines the block length l in Eq. (26) and hence the claimed sample complexity. With the correct MGF, the average of e^{γY} concentrates on 1+(e^γ-1)p, not on e^{γp}; the discrepancy is of order γ² and is not removed by the logarithmic transformation. The algorithm provides no unbiased estimator of e^{γp} from a single copy per state. Lemma 9 and Theorem 10 inherit this invalid step. Thus the main sample-co
  2. [Theorem 8 (PAC Generalization Bound of TERM), symmetrization step, Eq. (32)-(36)] The symmetrization argument contains a scaling inconsistency. The proof compares |R(h)-R̂'(h)| ≤ ε/2 and |R̃'_γ(h)-R̂'(h)| ≤ ε/4 to conclude |R̃'_γ(h)-R̃_γ(h)| ≥ ε/4. However, R̃_γ(h) is a log-average quantity, not the ordinary average R̂(h). The bound between log-average and average depends on γ, so the correct condition should involve |log Φ' - log Φ| scaled by γ, not the raw averages. The proof uses ε/4 for both quantities without accounting for the factor |e^γ-1|/γ that appears in the log-Lipschitz estimate. This makes the claimed exponent exp(-nε²/(32|e^γ-1|²)) unsupported by the preceding steps.
  3. [Theorem 8, cover-lifting step, Eq. (43)-(45)] In the lifting step, the cover is in sup-norm over the exponential class G_{F,L,γ}, so η controls |e^{γL(h,z)} - g_j(z)|. But the subsequent bound on |log Φ(h) - log Φ(g_j)| requires the ratio between the exponential values to be close to 1, i.e., η should be of order γε, not ε/8 as stated. The proof simply sets the log-difference to be ≤ η and then declares η = ε/8. This is a scaling error: e^{γL} can vary by factor e^γ; a sup-norm error η of the exponentials only gives log-ratio error η/min(1,e^{γa}), which is not ε/8 unless η is exponentially small in γ. The stated covering number Γ_1(2n, ε/8, G) is therefore not the right complexity parameter for the claimed bound.
  4. [Theorem 10 (Agnostic Learnability), Eq. (55)-(61)] Theorem 10 combines the QTERM sample-complexity bound of Theorem 7 with the uniform convergence bound of Lemma 9. Since Theorem 7's bound is invalid (see above), the perr,TERM term in Eq. (61) is unsupported. Additionally, Eq. (55) uses the condition (log Γ + C2)² ≤ C1 l ε², but Theorem 7's block-length condition (Eq. 26) involves a factor (e^{|γ|}-1)²/γ² log(m/δ), which is not reflected in Eq. (55). This mismatch further breaks the claimed error bound. The agnostic learning guarantee is therefore not established.
minor comments (4)
  1. [Throughout, notation] The symbol µ_c(γ) is used both for the tilted quantity (1/γ)log((1/n)Σ e^{γTr[...]}) and for the approximation in Eq. (13)-(14), leading to ambiguity. Also, the factor 1- in the tilted risk definition (Eq. 10) versus the log form in Theorem 7 is not consistently carried through.
  2. [Lemma 4 and Fig. 2] Lemma 4 is stated for the classical TERM loss and invoked in the quantum setting; the text should clarify the distinction. Figure 2 contains a mix of informal claims and symbols (e.g., 'Õ', QTR definition) that are not rigorously defined in the figure caption.
  3. [Section VI (QTR)] The Quantum Tilted Risk definition (Eq. 63) is interesting but not used in the main proofs. The relation to Rényi entropy is only sketched. This section could be more clearly separated as outlook or discussion.
  4. [Acknowledgments] The manuscript states that portions were drafted with ChatGPT. This is transparent but the lack of a data/code availability statement is notable for a theory paper relying on lengthy proofs.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the QTERM results extend external QERM/TERM baselines and are not reduced to their own inputs.

full rationale

The derivation chain is not circular. Definition 6 (QTERM) is an extension of the external QERM framework of Ref. [31] and classical TERM of Refs. [7,8]; Theorem 7's sample complexity is obtained by combining the external ThresholdSearch lemma (Lemma 5, from Ref. [31]) with concentration inequalities (Propositions 13, 15, 18), not by fitting a parameter to the quantity being predicted. Theorem 8 uses the standard symmetrization/covering-number argument for PAC bounds; it does not invoke Theorem 7's conclusion as a hypothesis, and the small-gamma regime is a stated assumption rather than a fitted output. Theorem 10 is built from Theorems 7 and 8, so it inherits no independent circular input. The self-citations that occur (e.g., Ref. [54] in Section VI, and other background references with overlapping authorship) are peripheral and are not load-bearing: none is used as a uniqueness theorem or as the justification for a forbidden alternative. The paper is therefore self-contained relative to external baselines. A separate mathematical concern exists: Proposition 15 asserts E[e^{cY}]=e^{cE[Y]} for binary measurement outcomes, which is false, and this makes parts of Theorem 7's concentration argument unsupported; however, that is a correctness defect in the proof, not a circular reduction of the claimed result to its own assumptions. The same applies to the algebra around Theorem 8's covering-number exponent. These issues do not affect the circularity score.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities. Its main new object is the QTERM loss function, a mathematical construct. The derivations rely on the gentle-measurement framework of Ref [31], a bounded-loss assumption, a covering-number bridge between function and operator classes, and the false exponential-expectation identity.

free parameters (1)
  • γ (tilt hyperparameter) = restricted to |γ| ∈ (0, ε)
    The theoretical guarantees hold only for small γ chosen to make proofs work; the bounds depend on γ and no selection rule is given for applications.
assumptions (5)
  • domain assumption Gentle quantum threshold search on nonidentical states (Lemma 5, Ref [31]) works with success probability 0.03 under condition (log m + C2)^2 < C1 n ε^2.
    The QTERM algorithms (Algorithms 1 and 2) are built directly on this external result; the paper does not re-prove it.
  • ad hoc to paper E[e^{γY}] = e^{γE[Y]} for a 0/1 measurement outcome Y (Propositions 15, used in Theorem 7).
    This identity is false and is introduced implicitly by the concentration step; it is load-bearing for the sample complexity result.
  • domain assumption Covering number of the exponential loss class G_{F,L,γ} is bounded by the operator-class covering number Γ1,q (Section IV Remark).
    Used to connect the classical PAC bound to quantum hypothesis classes; asserted without proof.
  • domain assumption Loss function values lie in [0,1] and γ ∈ (0, ε) for the PAC and agnostic results.
    The proofs of Theorems 8 and 10 require these restrictions to apply Hoeffding and log-Lipschitz constants.
  • domain assumption Samples are i.i.d. from a classical-quantum distribution with single-copy product states.
    This is the learning model stated in Section I B and inherited from Ref [31].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum Learning with Tunable Loss Functions." pith.science (2026). https://pith.science/paper/265UTB5X

@misc{pith2026250821369,
  author       = {Pith},
  title        = {Pith review of: Quantum Learning with Tunable Loss Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/265UTB5X}},
  note         = {Machine review of arXiv:2508.21369}
}
read the original abstract

Learning from quantum data presents new challenges to the paradigm of learning from data. This typically entails the use of quantum learning models to learn quantum processes that come with enough subtleties to modify the theoretical learning frameworks. This new intersection warrants new frameworks for complexity measures, including those on quantum sample complexity and generalization bounds. Empirical risk minimization (ERM) serves as the foundational framework for evaluating learning models in general. The diversity of learning problems leads to the development of advanced learning strategies such as tilted empirical risk minimization (TERM). Theoretical aspects of quantum learning under a quantum ERM framework are presented in [PRX Quantum 5, 020367 (2024)]. In this work, we propose a definition for TERM suitable to be employed when learning quantum processes, which gives rise to quantum TERM (QTERM). We show that QTERM can be viewed as a competitive alternative to implicit and explicit regularization strategies for quantum process learning. This work contributes to the existing literature on quantum and classical learning theory threefold. First, we prove QTERM learnability by deriving upper bounds on QTERM's sample complexity. Second, we establish new PAC generalization bounds on classical TERM. Third, we present QTERM agnostic learning guarantees for quantum hypothesis selection. These results contribute to the broader literature of complexity bounds on the feasibility of learning quantum processes, as well as methods for improving generalization in quantum learning.

Figures

Figures reproduced from arXiv: 2508.21369 by the authors.

Figure 1
Figure 1. Illustration of the learning process with QTERM. The process incorporates three main parts: Input: It begins with classical data and the corresponding quantum state representations. Algorithm: It utilizes a quantum threshold search shown in Appendix.IX B, which aims to find an optimal hypothesis class hc ∗ that minimizes the tunable loss R˜ γ(h), subject to the constraint R˜ γ(h) ≤ θc. Output: The optimal hypothesis… view at source ↗
Figure 2
Figure 2. Overview of the learning framework and main theoretical results. This diagram provides a comprehensive overview of our QTERM framework for learning quantum processes. The left side outlines the key algorithm: which takes as input a hypothesis class, projector-valued functions, product states and corresponding loss function, and a tilting parameter. Through quantum threshold search, the algorithm outputs an optimal h… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 54 canonical work pages

  1. [13]

    Generalization and Robustness of the Tilted Empirical Risk

    G. Aminian, A. R. Asadi, T. Li, A. Beirami, G. Reinert, and S. N. Cohen, Generalization error of the tilted empirical risk (2024), arXiv:2409.19431 [stat.ML]

    work page Pith review arXiv 2024

  2. [31]

    Fanizza, Y

    M. Fanizza, Y. Quek, and M. Rosati, Learning quantum processes without input control, PRX Quantum 5, 020367 (2024)

    work page 2024

  3. [1]

    Valiant, A theory of the learnable, Communications of the ACM 27, 1134 (1984)

    L. Valiant, A theory of the learnable, Communications of the ACM 27, 1134 (1984)

    work page 1984

  4. [2]

    Kearns and U

    M. Kearns and U. Vazirani, An Introduction to Computational Learning Theory (MIT Press, 1994)

    work page 1994

  5. [3]

    M. J. Kearns and R. E. Schapire, Efficient distribution-free learning of probabilistic concepts, Journal of Computer and System Sciences 48, 464 (1994)

    work page 1994

  6. [4]

    C. M. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics) (Springer-Verlag, Berlin, Heidelberg, 2006)

    work page 2006

  7. [5]

    Goodfellow, Y

    I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning (MIT Press, 2016) http://www.deeplearningbook.org

    work page 2016

  8. [6]

    LeCun, Y

    Y. LeCun, Y. Bengio, and G. Hinton, Deep learning, Nature 521, 436 (2015)

    2015

Show all 71 references
  1. [7]

    T. Li, A. Beirami, M. Sanjabi, and V. Smith, Tilted empirical risk minimization, arXiv preprint arXiv:2007.01162 (2020)

    2007 arXiv

  2. [8]

    T. Li, A. Beirami, M. Sanjabi, and V. Smith, On tilted losses in machine learning: Theory and applications, Journal of Machine Learning Research 24, 1 (2023)

    2023

  3. [9]

    Siegmund, Importance sampling in the monte carlo study of sequential tests, The Annals of Statistics , 673 (1976)

    D. Siegmund, Importance sampling in the monte carlo study of sequential tests, The Annals of Statistics , 673 (1976)

    1976

  4. [10]

    R. W. Butler, Saddlepoint approximations with applications , Vol. 22 (Cambridge University Press, 2007)

    2007

  5. [11]

    Thomas and A

    M. Thomas and A. T. Joy, Elements of information theory (Wiley-Interscience, 2006)

    2006

  6. [12]

    Dembo, Large deviations techniques and applications (Springer, 2009)

    A. Dembo, Large deviations techniques and applications (Springer, 2009)

    2009

  7. [14]

    S. L. Smith, B. Dherin, D. G. Barrett, and S. De, On the origin of implicit regularization in stochastic gradient descent, arXiv preprint arXiv:2101.12176 (2021)

    2021 arXiv

  8. [15]

    Bauer, S

    F. Bauer, S. Pereverzev, and L. Rosasco, On regularization algorithms in learning theory, Journal of Complexity 23, 52 (2007)

    2007

  9. [16]

    C. D. Mol, E. D. Vito, and L. Rosasco, Elastic-net regularization in learning theory (2008), arXiv:0807.3423 [stat.ML]

    2008 arXiv

  10. [17]

    Schuld and F

    M. Schuld and F. Petruccione, Machine Learning with Quantum Computers , Quantum Science and Technology (Springer International Publishing, 2021). 25

    2021

  11. [18]

    Biamonte, P

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, Quantum Machine Learning, Nature 549, 195 (2016)

    2016

  12. [19]

    Rebentrost, M

    P. Rebentrost, M. Mohseni, and S. Lloyd, Quantum support vector machine for big data classification, Physical review letters 113, 130503 (2014)

    2014

  13. [20]

    Schuld, A

    M. Schuld, A. Bocharov, K. M. Svore, and N. Wiebe, Circuit-centric quantum classifiers, Phys. Rev. A 101, 032308 (2020)

    2020

  14. [21]

    K. Beer, D. Bondarenko, T. Farrelly, T. J. Osborne, R. Salzmann, D. Scheiermann, and R. Wolf, Training deep quantum neural networks, Nature communications 11, 808 (2020)

    2020

  15. [22]

    M. A. Nielsen and I. L. Chuang, Quantum Computation and Quantum Information: 10th Anniversary Edition (Cambridge University Press, 2011)

    2011

  16. [23]

    Montanaro, Quantum algorithms: an overview, npj Quantum Information 2, 10.1038/npjqi.2015.23 (2016)

    A. Montanaro, Quantum algorithms: an overview, npj Quantum Information 2, 10.1038/npjqi.2015.23 (2016)

    2015 doi

  17. [24]

    Preskill, Quantum computing in the nisq era and beyond, Quantum 2, 79 (2018)

    J. Preskill, Quantum computing in the nisq era and beyond, Quantum 2, 79 (2018)

    2018

  18. [25]

    P. W. Shor, Fault-tolerant quantum computation, in Proceedings of 37th conference on foundations of computer science (IEEE, 1996) pp. 56–65

    1996

  19. [26]

    Cerezo, A

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, Variational quantum algorithms, Nat. Rev. Phys 3, 625–644 (2021)

    2021

  20. [27]

    Benedetti, E

    M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, Parameterized quantum circuits as machine learning models, Quantum Science and Technology 4, 043001 (2019)

    2019

  21. [28]

    Gily´ en, Y

    A. Gily´ en, Y. Su, G. H. Low, and N. Wiebe, Quantum singular value transformation and beyond: Exponential improvements for quantum matrix arithmetics, in Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing , STOC 2019 (Association for Computing Machiner...

    2019

  22. [29]

    Ivashkov, P.-W

    P. Ivashkov, P.-W. Huang, K. Koor, L. Pira, and P. Rebentrost, Qkan: Quantum kolmogorov-arnold networks (2024), arXiv:2410.04435 [quant-ph]

    2024 arXiv

  23. [30]

    N. Guo, Z. Yu, A. Agrawal, and P. Rebentrost, Quantum linear algebra is all you need for transformer architectures (2024), arXiv:2402.16714 [quant-ph]

    2024

  24. [32]

    V. N. Vapnik and A. Y. Chervonenkis, Teoriya raspoznavaniya obrazov [Theory of Pattern Recognition] (Nauka, Moscow, 1974)

    1974

  25. [33]

    Gyurik, D

    C. Gyurik, D. Vreumingen, van, and V. Dunjko, Structural risk minimization for quantum linear classifiers, Quantum 7, 893 (2023)

    2023

  26. [34]

    Heidari, A

    M. Heidari, A. Padakandla, and W. Szpankowski, A theoretical framework for learning from quantum data (2021), arXiv:2107.06406 [quant-ph]

    2021 arXiv

  27. [35]

    Padakandla and A

    A. Padakandla and A. Magner, Pac learning of quantum measurement classes: Sample complexity bounds and universal consistency, in International Conference on Artificial Intelligence and Statistics (PMLR, 2022) pp. 11305–11319

    2022

  28. [36]

    Heidari and W

    M. Heidari and W. Szpankowski, New bounds on quantum sample complexity of measurement classes, in 2024 IEEE International Symposium on Information Theory (ISIT) (IEEE, 2024) pp. 1515–1520

    2024

  29. [37]

    Ciliberto, A

    C. Ciliberto, A. Rocchetto, A. Rudi, and L. Wossnig, Statistical limits of supervised quantum learning, Physical Review A 102, 10.1103/physreva.102.042414 (2020)

    2020 doi

  30. [38]

    Arunachalam and R

    S. Arunachalam and R. De Wolf, Guest column: A survey of quantum learning theory, ACM Sigact News 48, 41 (2017)

    2017

  31. [39]

    Salmon, S

    W. Salmon, S. Strelchuk, and T. Gur, Provable advantage in quantum pac learning (2023), arXiv:2309.10887 [quant-ph]

    2023 arXiv

  32. [40]

    Nayak and P

    A. Nayak and P. Sinha, Proper vs improper quantum pac learning (2024), arXiv:2403.03295 [quant-ph]

    2024 arXiv

  33. [41]

    Chung and H.-H

    K.-M. Chung and H.-H. Lin, Sample Efficient Algorithms for Learning Quantum Channels in PAC Model and the Ap- proximate State Discrimination Problem, in 16th Conference on the Theory of Quantum Computation, Communication and Cryptography (TQC 2021) , Leibniz International Proc...

    2021

  34. [42]

    M. C. Caro, H.-Y. Huang, M. Cerezo, K. Sharma, A. Sornborger, L. Cincio, and P. J. Coles, Generalization in quantum machine learning from few training data, Nature Communications 13, 4919 (2022)

    2022

  35. [43]

    Abbas, D

    A. Abbas, D. Sutter, A. Figalli, and S. Woerner, Effective dimension of machine learning models (2021), arXiv:2112.04807 [cs.LG]

    2021 arXiv

  36. [44]

    Gil-Fuster, J

    E. Gil-Fuster, J. Eisert, and C. Bravo-Prieto, Understanding quantum machine learning also requires rethinking general- ization, Nature Communications 15 (2024)

    2024

  37. [45]

    N. H. Bshouty and J. C. Jackson, Learning dnf over the uniform distribution using a quantum example oracle, SIAM Journal on Computing 28, 1136 (1999), earlier version in COLT’95

    1999

  38. [46]

    Arunachalam and R

    S. Arunachalam and R. de Wolf, Optimal quantum sample complexity of learning algorithms, Journal of Machine Learning Research 19, 1 (2018)

    2018

  39. [47]

    V. N. Vapnik, Statistical Learning Theory (Wiley-Interscience, 1998)

    1998

  40. [48]

    Anthony and P

    M. Anthony and P. L. Bartlett, Neural network learning: Theoretical foundations (Cambridge University Press, 2009)

    2009

  41. [49]

    B˘ adescu and R

    C. B˘ adescu and R. O’Donnell, Improved quantum data analysis, TheoretiCSV olume 3, 10.46298/theoretics.24.7 (2024)

    2024 doi

  42. [50]

    M. M. Wolf, Mathematical foundations of supervised learning (2023)

    2023

  43. [51]

    M. C. Caro, E. Gil-Fuster, J. J. Meyer, J. Eisert, and R. Sweke, Encoding-dependent generalization bounds for parametrized quantum circuits, Quantum 5, 582 (2021)

    2021

  44. [52]

    Esscher, On the probability function in the collective theory of risk, Skandinavisk Aktuarietidskrift 15, 175 (1932)

    F. Esscher, On the probability function in the collective theory of risk, Skandinavisk Aktuarietidskrift 15, 175 (1932)

    1932

  45. [53]

    H. U. Gerber and E. S. W. Shiu, Option pricing by esscher transforms, Transactions of the Society of Actuaries 46, 99 (1994)

    1994

  46. [54]

    Y. Qiu, K. Koor, and P. Rebentrost, The quantum esscher transform, arXiv preprint arXiv:2401.07561 (2024). 26

    2024 arXiv

  47. [55]

    Bakshi, A

    A. Bakshi, A. Liu, A. Moitra, and E. Tang, Learning quantum hamiltonians at any temperature in polynomial time (2023), arXiv:2310.02243 [quant-ph]

    2023 arXiv

  48. [56]

    Anshu, S

    A. Anshu, S. Arunachalam, T. Kuwahara, and M. Soleimanifar, Sample-efficient learning of interacting quantum systems, Nature Physics 17, 931–935 (2021)

    2021

  49. [57]

    A. Gu, L. Cincio, and P. J. Coles, Practical hamiltonian learning with unitary dynamics and gibbs states, Nature Com- munications 15, 312 (2024)

    2024

  50. [58]

    W. Yu, J. Sun, Z. Han, and X. Yuan, Robust and efficient hamiltonian learning, Quantum 7, 1045 (2023)

    2023

  51. [59]

    Chen and R

    Y. Chen and R. de Wolf, Quantum algorithms and lower bounds for linear regression with norm constraints (2022), arXiv:2110.13086 [quant-ph]

    2022 arXiv

  52. [60]

    Z. Zhu, J. M. Lukens, and B. T. Kirby, On the connection between least squares, regularization, and classical shadows, Quantum 8, 1455 (2024)

    2024

  53. [61]

    Chakraborty, A

    S. Chakraborty, A. Morolia, and A. Peduri, Quantum regularized least squares, Quantum 7, 988 (2023)

    2023

  54. [62]

    Aminian, A

    G. Aminian, A. R. Asadi, T. Li, A. Beirami, G. Reinert, and S. N. Cohen, Generalization and robustness of the tilted empirical risk, in Forty-second International Conference on Machine Learning (2025)

    2025

  55. [63]

    Mitarai, M

    K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Quantum circuit learning, Phys. Rev. A 98, 032309 (2018)

    2018

  56. [64]

    Schuld, V

    M. Schuld, V. Bergholm, C. Gogolin, J. Izaac, and N. Killoran, Evaluating analytic gradients on quantum hardware, Physical Review A 99, 10.1103/physreva.99.032331 (2019)

    2019 doi

  57. [65]

    Wierichs, J

    D. Wierichs, J. Izaac, C. Wang, and C. Y.-Y. Lin, General parameter-shift rules for quantum gradients, Quantum 6, 677 (2022)

    2022

  58. [66]

    D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning representations by back-propagating errors, Nature 323, 533 (1986)

    1986

  59. [67]

    Abbas, R

    A. Abbas, R. King, H.-Y. Huang, W. J. Huggins, R. Movassagh, D. Gilboa, and J. R. McClean, On quantum backpropa- gation, information reuse, and cheating measurement collapse (2023), arXiv:2305.13362 [quant-ph]

    2023 arXiv

  60. [68]

    Hur and D

    T. Hur and D. K. Park, Understanding generalization in quantum machine learning with margins (2024), arXiv:2411.06919 [quant-ph]

    2024

  61. [69]

    Neyshabur, S

    B. Neyshabur, S. Bhojanapalli, and N. Srebro, A pac-bayesian approach to spectrally-normalized margin bounds for neural networks (2018), arXiv:1707.09564 [cs.LG]

    2018 arXiv

  62. [70]

    Hanneke and A

    S. Hanneke and A. Kontorovich, Stable sample compression schemes: New applications and an optimal svm margin bound (2020), arXiv:2011.04586 [cs.LG]

    2020 arXiv

  63. [71]

    Bousquet and A

    O. Bousquet and A. Elisseeff, Stability and generalization, J. Mach. Learn. Res. 2, 499–526 (2002)

    2002

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.