Pith. sign in

REVIEW 3 major objections 6 minor 53 references

Quantum-enhanced least-square support vector machine: simplified quantum algorithm and sparse solutions

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read The paper proposes quantum LS-SVM algorithms whose training cost scales as $O(\varepsilon^{-1}\log(MN))$ and whose hybrid sparse variant solves a reduced $R\times R$ system after quantum singular-value extraction.

desk verdict The paper assembles new QML variants, but its central CV matrix-inversion step hinges on a false exponential factorization, so the speed-up claim lacks support. read the letter →

arxiv 1908.01462 v1 pith:L6YA5H5B submitted 2019-08-05 quant-ph stat.ML

Jie Lin , Dan-Bo Zhang , Shuo Zhang , Xiang Wang , Tan Li , Wan-su Bao This is my paper · ORCID
classification quant-phstat.ML PACS 03.67.Ac
keywords quantumsupportvectormachineleast-squareSVMcontinuous-variablecomputingmatrixinversionsparsesolutionfeaturemapprincipalcomponentanalysisspeed-up
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 proposes two quantum algorithms for least-square support vector machines. The first uses two continuous-variable modes to perform the matrix inversion at the heart of LS-SVM training, replacing many ancillary qubits and giving an evolution cost of $O(\varepsilon^{-1}\log(MN))$ for $M$ training samples and $N$ features. The second is a hybrid quantum-classical procedure that first compresses the large dataset into a low-rank transformed dataset using quantum singular-value tools, then solves a much smaller SVM classically, with total cost $O(N+R^3+R\varepsilon^{-1}\log(MN))$ where $R$ is the number of kept singular values. If these algorithms work as described, LS-SVM training and inference would depend only logarithmically on the sample count and could exploit quantum feature maps that are hard to compute classically.

What carries the argument

The central object is the continuous-variable representation of matrix inversion as a two-mode integral, together with the assumed factorization of the phase-estimation unitary $e^{iM_s\hat{p}_1\hat{p}_2} = e^{iG_1\hat{p}_1\hat{p}_2}e^{iG_2\hat{p}_1\hat{p}_2}e^{iG_3\hat{p}_1\hat{p}_2}$, where $G_1=\mathrm{blkdiag}(K,0)$, $G_2=\mathrm{blkdiag}(\gamma^{-1}I,0)$, and $G_3=\begin{pmatrix}0&\mathbf{1}^T\\ \mathbf{1}&0\end{pmatrix}$. This factorization is what lets each block of the LS-SVM matrix be exponentiated separately, and it is the main source of the claimed logarithmic cost. The sparse hybrid algorithm is carried by the singular-value decomposition $\varphi(A)=\sum_{i=1}^{R}\sigma_i u_i v_i^T$: QPCA and QSVT extract the top singular triples, and the paper derives reduced equations for the rotated vectors $\alpha_{\Omega R}$, $\mathbf{1}_{\Omega R}$, $y_{\Omega R}$ and the diagonal reduced kernel $K_{\Omega R}=\Sigma_{R\times R}$, so the final classical step is a small linear system.

What would settle it

Evaluate Eq. (7) for a single positive eigenvalue $\lambda$: the stated integral gives $-1/\lambda$ rather than $\lambda^{-1}$, which would make the inversion routine return the wrong sign. Separately, evaluate $e^{iM_s\theta}$ against $e^{iG_1\theta}e^{iG_2\theta}e^{iG_3\theta}$ for a small noncommuting kernel matrix; the difference is nonzero, and its size is the missing approximation cost in the claimed complexity.

Watch

Extended reading notes

Core claim

The paper's central claim is that least-square SVM can be upgraded to a simpler quantum algorithm without losing the exponential speed-up of earlier proposals. The inverse matrix $M_s^{-1}$ needed for the LS-SVM solution is generated through the continuous-variable integral $M_s^{-1} = \frac{i}{\sqrt{2\pi}}\int_{-\infty}^{\infty}\int_{-\infty}^{\infty}\psi(p_1)p_2 e^{-p_2^2/2}e^{iM_s p_1 p_2}\,dp_1\,dp_2$, implemented by preparing a step-function state and a Gaussian state, applying the unitary $e^{iM_s\hat{p}_1\hat{p}_2}$, and post-selecting with homodyne detection. The paper further claims that the sample matrix can be replaced by a much smaller transformed dataset: quantum principal component analysis or quantum singular-value thresholding extracts the top $R$ singular triples of the feature matrix, and the LS-SVM parameters are then obtained by solving a reduced $R\times R$ linear system classically. The claimed costs are $O(\varepsilon^{-1}\log(MN))$ for the all-quantum version and $O(N+R^3+R\varepsilon^{-1}\log(MN))$ for the hybrid version, both giving exponential speed-up over the number of samples.

Load-bearing premise

The speed-up rests on the assumption that the matrix-inversion operation can be built by applying three simple block operations in sequence at the claimed cost; the paper does not prove this sequence is valid, and the three blocks do not commute.

Editorial extensions

If this is right

  • If HVQ-SVM works as claimed, an LS-SVM training run over $M$ samples costs $O(\varepsilon^{-1}\log(MN))$ coherent evolution time instead of any polynomial in $M$, preserving the exponential speed-up with fewer ancillary qubits.
  • If QSLS-SVM works as claimed, the classical part of the hybrid pipeline solves at most an $R\times R$ system, so for low-rank kernel matrices with $R\ll M$ the overall method is much cheaper than classical LS-SVM.
  • The error analysis treats the finite squeezing width and homodyne postselection as a form of L2 regularization, so the approximation error in the continuous-variable inversion is not simply noise but a regularizing bias.
  • The kernel-state constructions for polynomial and radial-basis kernels mean the algorithms are intended to work with quantum feature maps, not only linear data; if such feature maps are classically hard, the proposed SVMs could classify data that classical SVMs cannot efficiently represent.
  • For the soft-margin SVM dual, the paper argues that the sparse solution from QSLS-SVM is one solution of the dual problem, so the sparse extraction scheme can be carried over to the standard SVM setting.

Reading between the lines

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

  • The continuous-variable inversion gadget is a general recipe: any Hermitian matrix that can be split into efficiently exponentiating blocks could use the same two-mode integral, so the method extends beyond LS-SVM to other structured linear-algebra problems.
  • The hybrid pipeline implies that the practical quantum advantage is governed by the effective rank $R$ of the kernel matrix: when $R\ll M$, the classical piece is tiny and the quantum singular-value extraction is the part that must deliver speed-up; when $R$ is large, the $O(R^3)$ classical term dominates and the advantage disappears.
  • A direct way to test the paper's regularization claim is to compare the classifier produced by the finite-width and homodyne inversion with the classifier produced by the regularized inverse $(M_s+\chi I)^{-1}$ on the same synthetic dataset; the paper predicts they behave similarly.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript proposes two quantum algorithms for least-squares support vector machines. The first (HVQ-SVM) aims to simplify the quantum LS-SVM of Rebentrost et al. by replacing the qubit-based matrix-inversion subroutine with a continuous-variable integral representation of M_s^{-1}, claiming evolution cost O(ε^{-1} log(MN)). The second (QSLS-SVM) is a hybrid scheme that uses QPCA/QSVT to reduce the dataset to R principal components and then solves an R-dimensional classical LS-SVM, claiming complexity O(N + R^3 + R ε^{-1} log(MN)). The paper also discusses polynomial and Gaussian kernel state preparation and interprets finite-width effects as regularization.

Significance. If the central subroutine were valid, the logarithmic dependence on the number of samples would be a meaningful advance for quantum SVM, and the hybrid sparse extension is a reasonable direction. The paper correctly identifies the dense, non-sparse structure of M_s as the obstacle to direct Hamiltonian simulation and proposes a block decomposition. It also provides an appendix connecting finite-squeezing errors to L2-type regularization. However, the validity of the main algorithm rests on the exact factorization of a matrix exponential of noncommuting matrices, which is false, and the post-selection step is not justified; therefore the claimed speedups are not established.

major comments (3)
  1. [Sec. III.A, Eq. (10)] The identity e^{i M_s p1 p2} = ∏_{k=1}^3 e^{i G_k p1 p2} is false. Since M_s = G1 + G2 + G3, the equality would require the three Hermitian matrices to commute pairwise; for a generic kernel matrix K, [G1,G3] ≠ 0 and [G2,G3] ≠ 0. The text provides no Trotter-Suzuki decomposition, no ordering prescription, and no error bound. Because step 2 of the algorithm performs phase estimation with this unitary, the eigenphases read out in Eq. (11) are not those of M_s, and the post-selected state is not proportional to M_s^{-1}|ψ_y>. This invalidates the central HVQ-SVM subroutine and the claimed O(ε^{-1} log(MN)) complexity.
  2. [Sec. III.A, Eqs. (7)-(12)] The ideal CV post-selection is not justified. Direct evaluation of the scalar integral in Eq. (7) for an eigenvalue λ > 0 gives +2π/λ, so the sign of λ is preserved; the advertised output λ_k^{-1} in Eq. (12) nevertheless lacks the corresponding normalization factor. More seriously, with ψ(p1) vanishing at p1 = 0 and g(p2) containing a factor p2, the amplitude for projecting onto |0_{p1}>|0_{p2}> is zero; the statement ⟨q|0_{p1}⟩ = ⟨r|0_{p2}⟩ = 1 is incorrect. The finite-width construction introduced in Appendix B is a different procedure, and its success probability, fidelity, and effect on the complexity are not incorporated into the main algorithm.
  3. [Sec. III.B, Eqs. (14) and (35)] The reduction to the top-R components is not established at the level needed for the claimed complexity. Taking Ω = U, one obtains K_Ω = U^T K U = diag(σ_1^2, ..., σ_M^2), so the top-R submatrix is diag(σ_i^2), not Σ_{R×R} as written in Eq. (14). Moreover, the error estimate in Eq. (35) bounds the discarded contribution to g(x̂) by a factor σ_{T+1}, but it does not bound the error in α_{ΩR} and b or the number of repetitions of the quantum measurement needed to achieve a target classification accuracy; hence the cost O(N + R^3 + R ε^{-1} log(MN)) is not derived.
minor comments (6)
  1. [Title] The title contains an extraneous space in 'si mplified'; correct the typography.
  2. [Sec. II, Notations] The vector y is first defined as y = (y1, ..., yM) with y ∈ R^{1×N}, but the labels y_i are M scalars; the dimension should be M, not N.
  3. [Sec. II, Eq. (3)] The sentence 'There are f(x_i)y_i = 1 - ξ_i' is grammatically unclear; it should state that the constraint in Eq. (3) implies f(x_i)y_i = 1 - ξ_i.
  4. [Appendix A, Eq. (20)] The expression for N_{x'}^{-1} is difficult to parse and appears to have inconsistent powers; please rewrite the normalization of |φ(x')> explicitly.
  5. [Sec. III.A, before Eq. (10)] The phrase 'matrix Ms is full rank and unitary operation e^{iM_s p1 p2} can not be constructed' is confusing, since e^{iM_s t} is unitary for any Hermitian M_s; the authors likely mean that direct Hamiltonian simulation of the dense matrix M_s is unavailable.
  6. [Appendix A, after Eq. (22)] The cross-reference to 'Ref.[37]' does not match the citation numbering in the bibliography; check all cross-references.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning: the LS-SVM solve and sparse-truncation derivations are self-contained, with self-citations only as independently published subroutines.

full rationale

The central derivation chain is not circular. The target quantity is the standard LS-SVM dual solution [α;b] = M_s^{-1}[y;0] (Eq. 4), obtained from the Lagrangian of Eq. (3); the HVQ-SVM circuit is then explicitly constructed to realize that matrix inversion via the integral identity Eq. (7) and post-selection, while the QSLS-SVM part solves the reduced P-LSSVM equations Eqs. (13)-(14) after a QPCA/QSVT truncation. No parameter is fitted to a subset of data and then renamed a prediction; the only free parameters are the regularization γ and the truncation rank T, which are inputs to the stated problem, not outputs smuggled in as predictions. The self-citations that do occur (Ref. [25] by co-author D.-B. Zhang for the finite-squeezing CV inversion formalism, and Ref. [47] by the same group for a threshold-based improved QPCA) are separately published, parameter-free methods whose stated assumptions do not include the present LS-SVM result; they are used as subroutines rather than as a uniqueness argument forcing the answer, so under the hard rules they do not raise the circularity score. The reader-identified defect in Eq. (10) (noncommuting G_k make the product-of-exponentials identity false without a Trotter bound) is a mathematical correctness/implementation problem, not a circularity: the claimed algorithm would be wrong for an independent reason, but the derivation does not reduce to its own inputs. For this reason the circularity score is 0.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a sign-sensitive CV inversion identity, a noncommuting operator product, QRAM state preparation, and a low-rank assumption on the feature matrix. The first two are the most fragile: as printed, Eq (7) flips the sign of the inverse and Eq (10) ignores noncommutativity.

free parameters (2)
  • finite width L
    Width of the step-function state in Eq (8); the paper says to choose a suitable L but gives no value or criterion. It controls fidelity and contributes a regularization effect in Eqs (31)-(34).
  • rank truncation T
    Number of dominant singular components kept in QSLS-SVM, set by threshold tau = sigma_T. Error bound Eq (35) depends on T, but no selection rule or cost for finding T is given.
assumptions (5)
  • domain assumption The CV integral identity, Eq (7), holds as written: M_s^{-1} = i/sqrt(2 pi) integral integral psi(p1) p2 e^{-p2^2/2} e^{i M_s p1 p2} dp1 dp2.
    Quoted from continuous-variable matrix-inversion prior work, but a scalar check with psi(p1)=1 on (0,L) gives -1/lambda for lambda > 0, not 1/lambda. HVQ-SVM relies on this identity.
  • ad hoc to paper e^{i M_s p1 p2} can be implemented exactly as product_{k=1}^3 e^{i G_k p1 p2}.
    Used in Eq (10). G3 does not commute with G1 and G2, so exact factorization is false without ordering or Trotterization, and no error bound is supplied.
  • domain assumption QRAM can prepare the required quantum states in O(log MN) time.
    Assumed in both algorithm complexity analyses; no circuit construction is given for the specific amplitude encodings.
  • domain assumption phi(A) is low-rank and QPCA/QSVT extract its top singular components with O(epsilon^{-1} log(MN)) cost.
    The QSLS-SVM sparse solution and its complexity claim depend on low-rank structure and on the quantum matrix toolbox behaving as stated.
  • domain assumption Truncating to the top R singular components leaves the classification error E_Omega small.
    Appendix C bounds E_Omega by sigma_{T+1} times a data dependent factor, but no guarantee is given that this is small for the target dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantum-enhanced least-square support vector machine: simplified quantum algorithm and sparse solutions." pith.science (2026). https://pith.science/paper/L6YA5H5B

@misc{pith2026190801462,
  author       = {Pith},
  title        = {Pith review of: Quantum-enhanced least-square support vector machine: simplified quantum algorithm and sparse solutions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L6YA5H5B}},
  note         = {Machine review of arXiv:1908.01462}
}
read the original abstract

Quantum algorithms can enhance machine learning in different aspects. Here, we study quantum-enhanced least-square support vector machine (LS-SVM). Firstly, a novel quantum algorithm that uses continuous variable to assist matrix inversion is introduced to simplify the algorithm for quantum LS-SVM, while retaining exponential speed-up. Secondly, we propose a hybrid quantum-classical version for sparse solutions of LS-SVM. By encoding a large dataset into a quantum state, a much smaller transformed dataset can be extracted using quantum matrix toolbox, which is further processed in classical SVM. We also incorporate kernel methods into the above quantum algorithms, which uses both exponential growth Hilbert space of qubits and infinite dimensionality of continuous variable for quantum feature maps. The quantum LS-SVM exploits quantum properties to explore important themes for SVM such as sparsity and kernel methods, and stresses its quantum advantages ranging from speed-up to the potential capacity to solve classically difficult machine learning tasks.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 40 canonical work pages

  1. [1]

    J. Qiu, Q. Wu, G. Ding, Y. Xu, and S. Feng, EURASIP Journal on Advances in Signal Processing 2016, 67 (2016)

    work page 2016

  2. [2]

    Dunjko and H

    V. Dunjko and H. J. Briegel, Reports on Progress in Physics 81, 074001 (2018)

    2018

  3. [3]

    Novakovi´ c, Yugoslav Journal of Operations Research 21 (2016)

    J. Novakovi´ c, Yugoslav Journal of Operations Research 21 (2016)

    work page 2016

  4. [4]

    Hsu and C.-J

    C.-W. Hsu and C.-J. Lin, IEEE transactions on Neural Networks 13, 415 (2002)

    work page 2002

  5. [5]

    Wu and D.-X

    Q. Wu and D.-X. Zhou, Neural computation 17, 1160 (2005)

    work page 2005

  6. [6]

    J. A. Suykens and J. Vandewalle, Neural processing let- ters 9, 293 (1999)

    work page 1999

  7. [7]

    Sch¨ olkopf, S

    B. Sch¨ olkopf, S. Mika, C. J. Burges, P. Knirsch, K.-R. M¨ uller, G. R¨ atsch, and A. J. Smola, IEEE transactions on neural networks 10, 1000 (1999)

    work page 1999

  8. [8]

    V. D. S´ anchez A, Neurocomputing 55, 5 (2003)

    work page 2003

Show all 53 references
  1. [9]

    M. Gu, K. Wiesner, E. Rieper, and V. Vedral, Nature communications 3, 762 (2012). 6

    2012

  2. [10]

    S. D. Sarma, D.-L. Deng, and L.-M. Duan, Physics To- day 72, 3, 48 (2019)

    2019

  3. [11]

    Wiebe, D

    N. Wiebe, D. Braun, and S. Lloyd, Physical review let- ters 109, 050505 (2012)

    2012

  4. [12]

    Dunjko, J

    V. Dunjko, J. M. Taylor, and H. J. Briegel, Physical review letters 117, 130501 (2016)

    2016

  5. [13]

    Lloyd, S

    S. Lloyd, S. Garnerone, and P. Zanardi, Nature commu- nications 7, 10138 (2016)

    2016

  6. [14]

    Huang, Q

    H.-L. Huang, Q. Zhao, X. Ma, C. Liu, Z.-E. Su, X.-L. Wang, L. Li, N.-L. Liu, B. C. Sanders, C.-Y. Lu, et al. , Physical review letters 119, 050503 (2017)

    2017

  7. [15]

    Lloyd and C

    S. Lloyd and C. Weedbrook, Physical review letters 121, 040502 (2018)

    2018

  8. [16]

    Zhang, S.-L

    D.-B. Zhang, S.-L. Zhu, and Z. Wang, arXiv preprint arXiv:1906.03388 (2019)

    2019 arXiv

  9. [17]

    Rebentrost, M

    P. Rebentrost, M. Mohseni, and S. Lloyd, Physical re- view letters 113, 130503 (2014)

    2014

  10. [18]

    Z. Li, X. Liu, N. Xu, and J. Du, Physical review letters 114, 140504 (2015)

    2015

  11. [19]

    Schuld and N

    M. Schuld and N. Killoran, Physical review letters 122, 040504 (2019)

    2019

  12. [20]

    Havl ´ ıˇ cek, A

    V. Havl ´ ıˇ cek, A. D. C´ orcoles, K. Temme, A. W. Harrow, A. Kandala, J. M. Chow, and J. M. Gambetta, Nature 567, 209 (2019)

    2019

  13. [21]

    Benedetti, E

    M. Benedetti, E. Lloyd, and S. Sack, arXiv preprint arXiv:1906.07682 (2019)

    2019 arXiv

  14. [22]

    Rebentrost, A

    P. Rebentrost, A. Steffens, I. Marvian, and S. Lloyd, Physical review A 97, 012327 (2018)

    2018

  15. [23]

    Lloyd, in Quantum information with continuous vari- ables (Springer, 2003) pp

    S. Lloyd, in Quantum information with continuous vari- ables (Springer, 2003) pp. 37–45

    2003

  16. [24]

    H.-K. Lau, R. Pooser, G. Siopsis, and C. Weedbrook, Physical review letters 118, 080501 (2017)

    2017

  17. [25]

    Zhang, Z.-Y

    D.-B. Zhang, Z.-Y. Xue, S.-L. Zhu, and Z. Wang, Phys- ical Review A 99, 012331 (2019)

    2019

  18. [26]

    J. M. Arrazola, T. Kalajdzievski, C. Weedbrook, and S. Lloyd, arXiv preprint arXiv:1809.02622 (2018)

    2018 arXiv

  19. [27]

    Lloyd, M

    S. Lloyd, M. Mohseni, and P. Rebentrost, Nature Physics 10, 631 (2014)

    2014

  20. [28]

    B. Duan, J. Yuan, Y. Liu, and D. Li, Physical Review A 96, 032301 (2017)

    2017

  21. [29]

    B. Duan, J. Yuan, Y. Liu, and D. Li, Physical Review A 98, 012308 (2018)

    2018

  22. [30]

    A. W. Harrow, A. Hassidim, and S. Lloyd, Physical re- view letters 103, 150502 (2009)

    2009

  23. [31]

    M. J. Bremner, A. Montanaro, and D. J. Shepherd, Physical review letters 117, 080501 (2016)

    2016

  24. [32]

    Douce, D

    T. Douce, D. Markham, E. Kashefi, E. Diamanti, T. Coudreau, P. Milman, P. Van Loock, and G. Ferrini, Physical review letters 118, 070503 (2017)

    2017

  25. [33]

    Zhou, IEEE transactions on neural networks and learn - ing systems 27, 783 (2015)

    S. Zhou, IEEE transactions on neural networks and learn - ing systems 27, 783 (2015)

    2015

  26. [34]

    Sch¨ olkopf, R

    B. Sch¨ olkopf, R. Herbrich, and A. J. Smola, in In- ternational conference on computational learning theory (Springer, 2001) pp. 416–426

    2001

  27. [35]

    Giovannetti, S

    V. Giovannetti, S. Lloyd, and L. Maccone, Physical re- view letters 100, 160501 (2008)

    2008

  28. [36]

    Giovannetti, S

    V. Giovannetti, S. Lloyd, and L. Maccone, Physical Re- view A 78, 052310 (2008)

    2008

  29. [37]

    J. A. Suykens, L. Lukas, and J. Vandewalle, in 2000 IEEE International Symposium on Circuits and Systems. Emerging Technologies for the 21st Century. Proceedings (IEEE Cat No. 00CH36353) , Vol. 2 (IEEE, 2000) pp. 757–760

    2000

  30. [38]

    ling Chen, B

    H. ling Chen, B. Yang, S. jing Wang, G. Wang, H. zhong Li, W. bin Liu, et al. , Applied Mathematics and Compu- tation 239, 180 (2014)

    2014

  31. [39]

    Mall and J

    R. Mall and J. A. Suykens, IEEE transactions on neural networks and learning systems 26, 1086 (2015)

    2015

  32. [40]

    D. A. Silva, J. P. Silva, and A. R. R. Neto, Neurocom- puting 168, 908 (2015)

    2015

  33. [41]

    Zhou and M

    S. Zhou and M. Liu, in 2017 International Conference on Machine Vision and Information Technology (CMVIT) (IEEE, 2017) pp. 46–51

    2017

  34. [42]

    Chen and S

    L. Chen and S. Zhou, Neurocomputing 275, 2880 (2018)

    2018

  35. [43]

    Fowlkes, S

    C. Fowlkes, S. Belongie, F. Chung, and J. Malik, IEEE transactions on pattern analysis and machine intelligence 26, 214 (2004)

    2004

  36. [44]

    M. Li, W. Bi, J. T. Kwok, and B.-L. Lu, IEEE trans- actions on neural networks and learning systems 26, 152 (2014)

    2014

  37. [45]

    Lloyd and S

    S. Lloyd and S. L. Braunstein, in Quantum Information with Continuous Variables (Springer, 1999) pp. 9–17

    1999

  38. [46]

    B. Duan, J. Yuan, J. Xu, and D. Li, (2019)

    2019

  39. [47]

    Lin, W.-S

    J. Lin, W.-S. Bao, S. Zhang, T. Li, and X. Wang, Physics Letters A (2019)

    2019

  40. [48]

    Gily´ en, S

    A. Gily´ en, S. Lloyd, and E. Tang, arXiv preprint arXiv:1811.04909 (2018)

    2018 arXiv

  41. [49]

    Tang, arXiv preprint arXiv:1807.04271 (2018)

    E. Tang, arXiv preprint arXiv:1807.04271 (2018)

    2018 arXiv

  42. [50]

    Rahimi and B

    A. Rahimi and B. Recht, in Advances in neural informa- tion processing systems (2008) pp. 1177–1184

    2008

  43. [51]

    Talwalkar, Matrix approximation for large-scale learn- ing, Ph.D

    A. Talwalkar, Matrix approximation for large-scale learn- ing, Ph.D. thesis, Citeseer (2010)

    2010

  44. [52]

    Ring and B

    M. Ring and B. M. Eskofier, Pattern Recognition Letters 84, 107 (2016). The construction of kernel quantum state: Guass and polynomial kernel functions In this section we discuss the kernel map in Hilbert space. And in this paper we only describe the polynomial kernel function ...

    2016

  45. [53]

    Assume |φ(xp1 )⟩ = 1 √ Np1 ∑ jφ(xp1 )j|j⟩, we have ⟨φ(xp1 ),φ (xp2 )⟩ = 1√ Np1 Np2 ∑ jφ(xp1 )jφ(xp2 )j where Npi = |xpi |,i = 1 , 2

    However, normalization coefficient can bring bet- ter result. Assume |φ(xp1 )⟩ = 1 √ Np1 ∑ jφ(xp1 )j|j⟩, we have ⟨φ(xp1 ),φ (xp2 )⟩ = 1√ Np1 Np2 ∑ jφ(xp1 )jφ(xp2 )j where Npi = |xpi |,i = 1 , 2. Because the inequal- ity x+y 2 ≥ √xy,x,y ≥ 0, we have ⟨φ(xp1 ),φ (xp2 )⟩ ≤ 1 2 ( 1 N...

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.