Pith. sign in

REVIEW 3 major objections 5 minor 27 references

Supervised learning with a quantum classifier using a multi-level system

T0 review · 3 major / 5 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 A quNit quantum classifier trains each class in a single pass

desk verdict A neat single-quNit classifier with an honest accuracy discussion, but the advertised 'single shot' speedup breaks once you ask how the class density matrix is prepared. read the letter →

arxiv 1908.08385 v1 pith:SINUJXPB submitted 2019-08-22 quant-ph

Soumik Adhikary , Siddharth Dangwal , Debanjan Bhowmik This is my paper · ORCID
classification quant-ph
keywords quantummachinelearningneuralnetworkvariationalalgorithmquNitsingleshottrainingsuperviseddensitymatrixencodingSU(N)unitary
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 a supervised quantum classifier that encodes input vectors into a single $N$-level quantum system, a quNit, rather than into entangled multi-qubit registers. Its central claim is that "single shot training", in which all training samples of one class are combined into a density matrix $\rho_k$ and sent through the variational circuit at once, significantly speeds up training and uses far fewer parameters than a classical neural network. The paper demonstrates the classifier on four benchmark datasets, reporting accuracy comparable to a classical network with no hidden layer and higher accuracy on the three-class Iris dataset. If the speedup holds, classification tasks on near-term hardware could run with a quantum register whose size depends only on the number of classes, not on the input dimension.

What carries the argument

The load-bearing object is the class density matrix $\rho_k$: the uniform mixture of all encoded states in class $k$, which enters the circuit as a single quantum input, enabling "single shot training" (one forward pass per class per epoch). Around it stand the encoding map $|\psi(x)\rangle = e^{iS_3(w\cdot x)}H^{(N)}|0\rangle$, which folds the input features into phases via the diagonal matrix $S_3$; the generalized Hadamard $H^{(N)}$, which creates the superposition over all $N$ levels; the full $SU(N)$ unitary with $N^2-1$ real angles; and the weighted class-error $E$, which the classical optimizer minimizes. Together they make the quantum register size depend on the number of classes rather than on the input dimensionality.

What would settle it

Implement the proposed circuit on the CANCER dataset and measure the wall-clock time per epoch as the training set grows: if the time scales with the total number of samples rather than with the number of classes, the single-shot-training speedup is not realized. A direct count of the state-preparation steps in Eq. (5) gives the same result analytically: forming $\rho_k$ costs $O(m_k)$ encodings.

Watch

Extended reading notes

Core claim

The central discovery is that a quNit, with Hilbert-space dimension equal to the number of output classes, can serve as the feature space for supervised classification. Each input vector $x$ is mapped to a state $|\psi(x)\rangle = e^{iS_3(\sum_j w_j x_j)} H^{(N)}|0\rangle$; a parameterized $SU(N)$ unitary (via the generalized Euler-angle decomposition) transforms the state, and the label is chosen by the most probable outcome of a measurement of the diagonal operator $S_3$. For training, the samples of class $k$ are pooled into the density matrix $\rho_k = \frac{1}{m_k}\sum_i |\psi(x_i)_k\rangle\langle\psi(x_i)_k|$, and the error function $E=\frac{1}{M}\sum_k m_k(1-p_k)$ is minimized by gradient descent over the encoding weights $\{w_j\}$ and the $N^2-1$ circuit angles $\{\alpha_i\}$. This arrangement yields a training pass per class per epoch instead of per sample, and a parameter count of $d+N^2-1$, which the paper contrasts with the $(N+1)d$ weights of a no-hidden-layer classical FCNN.

Load-bearing premise

The speedup claim assumes that the class-averaged density matrix $\rho_k$ can be supplied to the circuit as a ready-made quantum state in constant time; the paper does not show how to prepare $\rho_k$ from the $m_k$ individual samples without iterating over all of them.

Editorial extensions

If this is right

  • Per-epoch training cost scales with the number of classes $N$, not the number of samples $M$: one circuit pass per class instead of one per sample.
  • The quantum register is a single quNit of dimension $N$; adding more input features only changes the encoding weights, not the Hilbert-space size.
  • The number of trainable parameters is $d+N^2-1$, which for the tested datasets is smaller than the weight count of a classical one-hidden-layer network.
  • On Fisher's Iris dataset, the quNit classifier reaches roughly 82 percent train and 83 percent test accuracy, above the 66 percent ceiling of a no-hidden-layer classical FCNN on that data.
  • The classifier uses superposition and interference only; entangled resources are not required for this form of quantum machine learning.

Reading between the lines

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

  • The paper leaves open how $\rho_k$ is physically prepared; if building $\rho_k$ requires applying the encoding map to each of the $m_k$ samples separately, the per-epoch cost is still linear in $M$, which would erase the claimed "single shot" speedup.
  • Because the $SU(N)$ circuit explores the full unitary group on the class-index space, the model's capacity is set by the number of classes, suggesting that linearly inseparable data would need either a nonlinear encoding or a nonlinear measurement to exceed the classical hidden-layer baseline the paper reports.
  • A concrete extension would be to replace the fixed $S_3$ measurement with a learned observable or a POVM; this could push the quNit classifier past the roughly 41 percent test accuracy it gets on the SONAR dataset without enlarging the register.
  • The same single-quNit pipeline could apply to regression or multi-label classification by reading out multiple expectation values from the final state instead of a single projective outcome.
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 / 5 minor

Summary. The paper proposes a variational quantum classifier in which input feature vectors are encoded into a single N-level quantum system (qudit) rather than into an entangled multi-qubit register. The circuit consists of a parameter-dependent encoding layer, a general SU(N) rotation with N^2-1 trainable angles, and a projective measurement used for the decision rule. The main claimed novelty is "single shot training": all training samples belonging to the same class are combined into the class-level density matrix ρ_k, and the circuit is trained on these density matrices rather than on individual samples. Numerical experiments on four benchmark datasets (CANCER, SONAR, IRIS, IRIS(2)) report train and test accuracies, which are compared with a classical fully connected neural network with zero or one hidden layer. The authors also claim that the number of training parameters is significantly smaller than in classical classifiers and that single-shot training significantly speeds up the training procedure.

Significance. If the single-shot training procedure could be realized with a preparation cost independent of class size, the work would be a useful qudit-based alternative to multi-qubit variational classifiers, with a compact parameter count and no reliance on entanglement. The paper is explicit about the limitation on linearly inseparable classes and makes a clear architectural statement. However, the central speedup claim is unsupported as written, and the numerical comparisons are made against deliberately weak classical baselines; the contribution is therefore more limited than the abstract suggests.

major comments (3)
  1. [§2.4, Eq. (5)] The "single shot training" claim rests on the assumption that the class-mixture density matrix ρ_k can be formed and passed through the circuit in one go, but the paper never specifies how ρ_k is prepared from the classical training set. In the absence of a quantum random-access memory or similar oracle, the natural construction ρ_k = (1/m_k) Σ_i |ψ(x_i)^k⟩⟨ψ(x_i)^k| requires Ω(m_k) state-preparation calls per epoch. Because the encoding weights w_j are trainable and are updated every epoch, ρ_k changes each epoch and must be rebuilt. Moreover, the gradient of E with respect to w_j involves derivatives of the individual constituent states, again requiring per-sample access. Thus the training cost is not independent of class size, and the claimed speedup over classical per-sample training is not established.
  2. [Abstract and §2.4] The statement that single-shot training "significantly speeds up the training procedure" is never supported by a runtime model, a query-complexity analysis, or an empirical timing comparison. To make the speedup claim meaningful, the authors need to state the assumed cost of preparing ρ_k (e.g., by specifying a quantum data-access model) or to compare wall-clock training times against classical classifiers under the same conditions. As written, the claim is not falsifiable.
  3. [§3, Tables 3 and 4] The claimed "advantage over classical machine learning classifiers" is based on a zero-hidden-layer FCNN with (N+1)d parameters, which the authors themselves show is outperformed by a one-hidden-layer FCNN on every dataset. The parameter-count comparison is therefore against a deliberately weak baseline; a fair comparison for a classifier that cannot separate linearly inseparable classes would use classical models of comparable expressivity, and the advantage claim should be correspondingly qualified.
minor comments (5)
  1. [§2.1 vs §2.3] The symbol S3 is used for two different diagonal matrices: in Eq. (1) it is diag(-(N-1)/2,...,(N-1)/2), while in the measurement discussion it is Diag(1,2,...,N). Since the decision rule in Eq. (4) depends on S3, the notation should be disambiguated or the encoding operator renamed.
  2. [§3, Tables 2-4] The quantum classifier results are averaged over 100 runs, while the classical FCNN results are averaged over only 5 runs; the disparity in the number of runs is not justified and makes the variance estimates (e.g., the zero reported for CANCER) difficult to interpret.
  3. [§2.4] The gradient descent procedure is not specified. The paper does not describe how the gradients of E with respect to the SU(N) angles {α_i} and the encoding weights {w_j} are computed (e.g., by parameter-shift rules or numerical differentiation), which is necessary for the hybrid algorithm to be reproducible.
  4. [§3, Table 1] The CANCER dataset is listed with d=27, but the standard Wisconsin Breast Cancer (Diagnostic) dataset has 30 features; the source of the 27-dimensional version is not cited, and the authors should clarify which pre-processing, if any, was applied.
  5. [Throughout] The spelling "quNit" is nonstandard; "qudit" is the common term. This is purely cosmetic but may confuse readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the classifier's error function, training procedure, and test accuracies are defined from the circuit without presupposing the results.

full rationale

The paper's derivation chain is self-contained against external benchmarks and does not reduce to its own inputs. The encoding |ψ(x)⟩ = Z(w_1,...,w_d) H^(N)|0⟩ (Eq. 1), the general SU(N) unitary (Eq. 3), and the measurement rule f*(x) = l_b iff p_b is maximal (Eq. 4) are all stated independently of the test outcomes. The class-mixture density matrix ρ_k = (1/m_k) Σ |ψ(x)^k_i⟩⟨ψ(x)^k_i| (Eq. 5) is defined from the training samples, and the training error E = (1/M) Σ m_k E_k is minimized by gradient descent on the parameters {w_i} and {α_i}; the reported train and test accuracies are genuine holdout evaluations after parameter freezing. No fitted parameter is renamed as a prediction, no load-bearing claim is justified only by a self-citation, and no uniqueness theorem is imported from the authors' prior work. The main weakness identified by the skeptic — that preparing ρ_k from the m_k individual encoded states costs O(m_k) operations per epoch, so the 'single shot' speedup is not demonstrated — is an unsupported efficiency claim about state preparation, not a circular step: the mathematical definitions and the accuracy results do not presuppose the speedup conclusion. Accordingly, the paper earns a 0 on the circularity scale.

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

The central claim relies on: (1) d trainable encoding weights and N^2-1 circuit angles, all learned from data; (2) the standard SU(N) parameterization; (3) the unproven utility of the class-averaged loss as a surrogate for accuracy; and (4) an unspecified gradient computation method. The most consequential is the unstated cost model for preparing the class-mixture density matrix, which the paper assumes to be cheap but does not specify.

free parameters (3)
  • Encoding weights {w_j} = not reported (optimized via gradient descent)
    These d parameters are trained on the training data; final values are not given, so the model cannot be reproduced from the paper alone.
  • SU(N) rotation angles {α_i} = not reported (optimized via gradient descent)
    The N^2-1 angles parameterize the circuit and are learned; without their final or initial values, the exact decision boundary is unspecified.
  • Learning rate (quantum) = 0.002
    Set by hand, fixed during training; affects convergence but not the final classifier if optimization succeeds.
assumptions (4)
  • domain assumption The encoding map x -> |ψ(x)⟩ = e^{i S3 Σ w_j x_j} H^(N)|0⟩ is a valid feature map that can be implemented on quantum hardware.
    Used in Sec. 2.1 without demonstration that the phase gate with arbitrary real coefficients is efficiently implementable on a qudit.
  • standard math The generalized Euler angle parameterization of SU(N) in Eq. (3) covers all unitaries and is differentiable for gradient descent.
    The paper cites [26] (Tilma & Sudarshan) for this parameterization; it is a known mathematical result.
  • domain assumption Minimizing the class-averaged error E = (1/M) Σ m_k (1 - p_k) will drive per-sample classification accuracy upward.
    Assumed without proof; the loss is a surrogate for accuracy but the paper does not show they are monotonically related. This is the basis of the training procedure in Sec. 2.4.
  • domain assumption Gradients of the quantum expectation values with respect to the gate parameters can be computed on a classical computer (or via parameter-shift rule).
    The paper simply says 'classical gradient descent technique [8]' without specifying a quantum gradient method, which is a nontrivial requirement.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Supervised learning with a quantum classifier using a multi-level system." pith.science (2026). https://pith.science/paper/SINUJXPB

@misc{pith2026190808385,
  author       = {Pith},
  title        = {Pith review of: Supervised learning with a quantum classifier using a multi-level system},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SINUJXPB}},
  note         = {Machine review of arXiv:1908.08385}
}
abstract

We propose a quantum classifier, which can classify data under the supervised learning scheme using a quantum feature space. The input feature vectors are encoded in a single qu$N$it (a $N$ level quantum system), as opposed to more commonly used entangled multi-qubit systems. For training we use the much used quantum variational algorithm -- a hybrid quantum-classical algorithm -- in which the forward part of the computation is performed on a quantum hardware whereas the feedback part is carried out on a classical computer. We introduce "single shot training" in our scheme, with all input samples belonging to the same class being used to train the classifier simultaneously. This significantly speeds up the training procedure and provides an advantage over classical machine learning classifiers. We demonstrate successful classification of popular benchmark datasets with our quantum classifier and compare its performance with respect to some classical machine learning classifiers. We also show that the number of training parameters in our classifier is significantly less than the classical classifiers.

Figures

Figures reproduced from arXiv: 1908.08385 by the authors.

Figure 1
Figure 1. A schematic representation of our quantum classifier. The gates [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Training results for CANCER dataset. (a) Variation of error with number of [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Training results for SONAR dataset. (a) Variation of error with number of [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Training results for IRIS dataset. (a) Variation of error with number of [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Training results for IRIS(2) dataset. (a) Variation of error with number of [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 19 canonical work pages

  1. [1]

    Bennett C H, Brassard G, Cr´ epeau C, Jozsa R, Peres A and Wootters W K 1993 Phys. Rev. Lett. 70(13) 1895–1899

    work page 1993

  2. [2]

    IEEE Int

    Bennett C and Brassard G 1984 Proc. IEEE Int. Conf. on Comp. Sys. Signal Process (ICCSSP) 175

    work page 1984

  3. [3]

    2017 Nature 549 70

    Ren J G, Xu P, Yong H L, Zhang L, Liao S K, Yin J, Liu W Y, Cai W Q, Yang M, Li L et al. 2017 Nature 549 70

    work page 2017

  4. [4]

    Nielsen M A and Chuang I 2002 Quantum computation and quantum information

    work page 2002

  5. [5]

    Grover L K 1996 ( Preprint quant-ph/9605043)

    arXiv 1996

  6. [6]

    Ekert A K 1991 Phys. Rev. Lett. 67(6) 661–663 URL https://link.aps.org/doi/10.1103/ PhysRevLett.67.661

    work page 1991

  7. [7]

    Biamonte J, Wittek P, Pancotti N, Rebentrost P, Wiebe N and Lloyd S 2017 Nature 549 195

    2017

  8. [8]

    LeCun Y, YBengio and GHinton 2015 Nature 521 436–444 URL https://rdcu.be/bwlmL

    work page 2015

Show all 27 references
  1. [9]

    Bishop C M 2006 Pattern recognition and machine learning (springer)

    2006

  2. [10]

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

    2019

  3. [11]

    Schuld M and Killoran N 2019 Phys. Rev. Lett. 122(4) 040504 URL https://link.aps.org/doi/ 10.1103/PhysRevLett.122.040504

    2019 doi

  4. [12]

    Wan K H, Dahlsten O, Kristj´ ansson H, Gardner R and Kim M 2017 npj Quantum Information 3 36

    2017

  5. [13]

    Farhi E and Neven H 2018 arXiv preprint arXiv:1802.06002

    2018 arXiv

  6. [14]

    Rebentrost P, Bromley T R, Weedbrook C and Lloyd S 2018 Phys. Rev. A 98(4) 042308 URL https://link.aps.org/doi/10.1103/PhysRevA.98.042308

    2018 doi

  7. [15]

    McClean J R, Romero J, Babbush R and Aspuru-Guzik A 2016 New Journal of Physics 18 023023 URL https://doi.org/10.1088%2F1367-2630%2F18%2F2%2F023023

    2016

  8. [16]

    Mitarai K, Negoro M, Kitagawa M and Fujii K 2018 Phys. Rev. A 98(3) 032309 URL https: //link.aps.org/doi/10.1103/PhysRevA.98.032309

    2018 doi

  9. [17]

    Russell S J and Norvig P 2016 Artificial intelligence: a modern approach (Malaysia; Pearson Education Limited,)

    2016

  10. [18]

    Hinton G E, Sejnowski T J and Poggio T A 1999 Unsupervised learning: foundations of neural computation (MIT press)

    1999

  11. [19]

    Sutton R S and Barto A G 1998 Cambridge, MA

    1998

  12. [20]

    Schuld M, Bocharov A, Svore K and Wiebe N 2018 arXiv preprint arXiv:1804.00633

    2018 arXiv

  13. [21]

    Preskill J 2018 Quantum 2 79 ISSN 2521-327X URL https://doi.org/10.22331/ q-2018-08-06-79

    2018

  14. [22]

    Malik M, Mirhosseini M, Lavery M P, Leach J, Padgett M J and Boyd R W 2014 Nature communications 5 3115

    2014

  15. [23]

    Department of Information and Computer Science, University of California Irvine 1998 UCI repository of machine learning databases https://archive.ics.uci.edu/ml/datasets/Iris Supervised learning with a quantum classifier using a multi-level system 12

    1998

  16. [24]

    Department of Information and Computer Science, University of California Irvine 1998 UCI repository of machine learning databases https://archive.ics.uci.edu/ml/datasets/ connectionist+bench+(sonar,+mines+vs.+rocks)

    1998

  17. [25]

    Department of Information and Computer Science, University of California Irvine 1998 UCI repository of machine learning databases http://archive.ics.uci.edu/ml/datasets/ Breast+Cancer+Wisconsin+%28Diagnostic%29

    1998

  18. [26]

    Tilma T and Sudarshan E C G 2002 Journal of Physics A: Mathematical and General 35 10467– 10501 URL https://doi.org/10.1088%2F0305-4470%2F35%2F48%2F316

    2002

  19. [27]

    Hornik K, Stinchcombe M and White H 1989 Neural Networks 2 359 – 366 ISSN 0893-6080 URL http://www.sciencedirect.com/science/article/pii/0893608089900208

    1989

Pith tools

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