pith. sign in

arxiv: 2604.07856 · v1 · submitted 2026-04-09 · 🪐 quant-ph

Hardware-Aware Quantum Support Vector Machines

Adil Mubashir Chaudhry , Ali Raza Haider , Hanzla Khan , Muhammad Faryad This is my paper

Pith reviewed 2026-05-10 17:15 UTC · model grok-4.3

classification 🪐 quant-ph
keywords quantum support vector machineshardware-aware neural architecture searchgenetic algorithmsnative gatesIBM quantum processorsquantum feature mapsNISQ devicesquantum kernel methods
0
0 comments X

The pith

Hardware-aware genetic search discovers native 12-gate quantum circuits that achieve 91 percent accuracy for support vector machines on IBM processors without transpilation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that constraining neural architecture search to the exact native gate set of IBM quantum hardware lets genetic algorithms automatically design quantum feature maps for support vector machines. These maps, limited to 12 gates on 10 qubits using only ECR, SX, and RZ operations, reach 91.23 percent accuracy on the breast cancer dataset. The result matches the performance of searches free of hardware constraints yet eliminates all compilation overhead and improves 27 percentage points over hand-designed quantum feature maps. Because the circuits use 100 percent native gates, they can execute directly on current noisy devices and approach the accuracy of classical radial-basis-function support vector machines.

Core claim

Using genetic algorithms to evolve circuit architectures strictly from IBM Torino native gates (ECR, RZ, SX, X), the hardware-aware NAS identifies a 12-gate quantum feature map on 10 qubits that delivers 91.23 percent classification accuracy on the UCI Breast Cancer Wisconsin dataset. This performance equals unconstrained searches while requiring zero transpilation, exceeds hand-crafted maps at 64 percent, and nearly matches the classical RBF SVM baseline at 93 percent. Removing fixed RZ placement constraints inside the hardware-aware search adds a further 3.5 percentage points.

What carries the argument

Hardware-aware Neural Architecture Search via genetic algorithms that evolve quantum feature maps while restricting every gate to the device's native set (ECR, RZ, SX, X).

If this is right

  • 100 percent native gate usage removes all decomposition errors that arise during universal-gate compilation.
  • Zero-transpilation circuits become immediately executable on IBM processors without modification.
  • Relaxing internal placement constraints during hardware-aware search produces measurable accuracy gains of 3.5 percentage points.
  • Quantum kernel methods become deployable on current NISQ devices through automated rather than manual design.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same constrained search procedure could be applied to other quantum machine learning primitives such as quantum neural networks or variational classifiers.
  • Hardware constraints may act as a beneficial inductive bias that steers discovery toward noise-resilient circuits rather than purely expressive ones.
  • Testing the discovered circuits under realistic noise models would reveal whether native-gate compatibility also improves robustness beyond simulation accuracy.

Load-bearing premise

The genetic algorithm under the native-gate constraint consistently locates high-performing circuits and the reported accuracy gains are insensitive to random seeds, hyperparameter choices, or the particular train-test split.

What would settle it

Executing the reported 12-gate circuit on actual IBM quantum hardware and measuring whether its classification accuracy falls significantly below the simulated 91.23 percent, or whether an independent search method under identical constraints finds a circuit with materially higher accuracy.

Figures

Figures reproduced from arXiv: 2604.07856 by Adil Mubashir Chaudhry, Ali Raza Haider, Hanzla Khan, Muhammad Faryad.

Figure 1
Figure 1. Figure 1: FIGURE 1: Confusion matrices comparing classical RBF SVM (93.0% accuracy), IBM hardware-aware NAS with [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIGURE 2: Quantum circuit diagram of the IBM hardware-aware NAS-discovered feature map with fixed RZ. The [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIGURE 3: Qubit connectivity pattern discovered by [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIGURE 4: Qubit connectivity pattern for IBM [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIGURE 5: Hardware efficiency comparison across methods. (a) Native gate usage shows hardware-aware NAS [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Deploying quantum machine learning algorithms on near-term quantum hardware requires circuits that respect device-specific gate sets, connectivity constraints, and noise characteristics. We present a hardware-aware Neural Architecture Search (NAS) approach for designing quantum feature maps that are natively executable on IBM quantum processors without transpilation overhead. Using genetic algorithms to evolve circuit architectures constrained to IBM Torino native gates (ECR, RZ, SX, X), we demonstrate that automated architecture search can discover quantum Support Vector Machine (QSVM) feature maps achieving competitive performance while guaranteeing hardware compatibility. Evaluated on the UCI Breast Cancer Wisconsin dataset, our hardware-aware NAS discovers a 12-gate circuit using exclusively IBM native gates (6 ECR, 3 SX, 3 RZ) that achieves 91.23 % accuracy on 10 qubits-matching unconstrained gate search while requiring zero transpilation. This represents a 27 percentage point improvement over hand-crafted quantum feature maps (64 % accuracy) and approaches the classical RBF SVM baseline (93 %). We show that removing architectural constraints (fixed RZ placement) within hardware-aware search yields 3.5 percentage point gains, and that 100 % native gate usage eliminates decomposition errors that plague universal gate compilations. Our work demonstrates that hardware-aware NAS makes quantum kernel methods practically deployable on current noisy intermediate-scale quantum (NISQ) devices, with circuit architectures ready for immediate execution without modification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 1 minor

Summary. The paper presents a hardware-aware Neural Architecture Search (NAS) using genetic algorithms to evolve quantum feature maps for QSVMs constrained to IBM Torino native gates (ECR, RZ, SX, X). On the UCI Breast Cancer Wisconsin dataset, it reports a 12-gate circuit (6 ECR, 3 SX, 3 RZ) achieving 91.23% accuracy on 10 qubits, matching unconstrained searches, improving 27 pp over hand-crafted maps (64%), and approaching classical RBF SVM (93%). It claims benefits from removing fixed RZ constraints (3.5 pp gain) and 100% native gate usage eliminating decomposition errors, positioning the method as enabling practical NISQ deployment without transpilation.

Significance. If the empirical results prove robust, the work is significant for NISQ-era quantum ML: it shows automated search can produce immediately executable, hardware-native circuits that outperform manual designs while matching broader searches. The explicit focus on zero-transpilation compatibility and native-gate constraints is a practical strength. However, the absence of robustness metrics for the stochastic GA search limits the strength of the performance claims and their generalizability.

major comments (3)
  1. [Abstract] Abstract: The headline result (91.23% accuracy for the 12-gate native circuit) is presented without statistical error bars, number of independent GA runs, variance across random seeds, or cross-validation procedure on the UCI split. Given that evolutionary search is stochastic, a single reported circuit does not establish that hardware-aware NAS reliably matches unconstrained performance.
  2. [Abstract] Abstract: No details are supplied on GA hyperparameters (population size, generations, mutation/crossover rates, selection method) or the precise train-test split and evaluation protocol. These omissions make it impossible to assess whether the reported accuracy gains are reproducible or sensitive to implementation choices.
  3. [Abstract] Abstract: The 27 pp improvement over hand-crafted maps and the 3.5 pp gain from relaxing RZ placement are central to the hardware-aware advantage, yet the exact construction of the hand-crafted baselines and the unconstrained search is not specified, preventing verification that the comparisons are controlled for qubit count, dataset, and kernel evaluation.
minor comments (1)
  1. [Abstract] The abstract would benefit from explicitly stating the dataset size, feature dimensionality, and number of qubits used in the main experiment at the outset for immediate clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive feedback, which highlights important aspects for improving the clarity, reproducibility, and verifiability of our results. We address each major comment below and will revise the manuscript to incorporate the requested details.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The headline result (91.23% accuracy for the 12-gate native circuit) is presented without statistical error bars, number of independent GA runs, variance across random seeds, or cross-validation procedure on the UCI split. Given that evolutionary search is stochastic, a single reported circuit does not establish that hardware-aware NAS reliably matches unconstrained performance.

    Authors: We agree that statistical robustness is essential when reporting results from stochastic methods such as genetic algorithms. In the revised manuscript we will augment both the abstract and the results section with error bars derived from multiple independent GA runs, the number of runs performed, the observed variance across random seeds, and a clear description of the cross-validation procedure applied to the UCI split. This will allow readers to evaluate the consistency of the hardware-aware search relative to unconstrained performance. revision: yes

  2. Referee: [Abstract] Abstract: No details are supplied on GA hyperparameters (population size, generations, mutation/crossover rates, selection method) or the precise train-test split and evaluation protocol. These omissions make it impossible to assess whether the reported accuracy gains are reproducible or sensitive to implementation choices.

    Authors: We acknowledge that the current presentation omits these implementation details. The revised manuscript will include a dedicated subsection (or expanded Methods) that fully specifies the genetic-algorithm hyperparameters—population size, number of generations, mutation and crossover rates, and selection mechanism—together with the exact train-test split of the UCI Breast Cancer Wisconsin dataset and the complete evaluation protocol for kernel-matrix construction and QSVM training. These additions will make the experimental setup fully reproducible. revision: yes

  3. Referee: [Abstract] Abstract: The 27 pp improvement over hand-crafted maps and the 3.5 pp gain from relaxing RZ placement are central to the hardware-aware advantage, yet the exact construction of the hand-crafted baselines and the unconstrained search is not specified, preventing verification that the comparisons are controlled for qubit count, dataset, and kernel evaluation.

    Authors: We recognize the need for precise baseline definitions to substantiate the reported gains. In the updated manuscript we will provide explicit circuit descriptions of the hand-crafted feature maps that achieve the 64 % accuracy, including gate counts and placement rules, as well as the precise configuration of the unconstrained search (gate set, qubit count, and kernel-evaluation settings). All comparisons will be explicitly controlled for the same dataset split, qubit number, and evaluation protocol, thereby confirming the fairness of the 27 pp and 3.5 pp improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical NAS results are directly measured

full rationale

The paper reports performance of quantum feature maps discovered via genetic algorithm search under hardware constraints, evaluated directly on the UCI Breast Cancer Wisconsin dataset. Accuracy figures (e.g., 91.23%) are obtained from explicit model training and testing rather than any algebraic derivation, fitted parameter renamed as prediction, or self-citation chain. No equations, uniqueness theorems, or ansatzes are invoked that reduce the central claim to its own inputs. The methodology is self-contained experimental search with no load-bearing logical reductions.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no explicit free parameters, axioms, or invented entities can be extracted from the full manuscript.

pith-pipeline@v0.9.0 · 5556 in / 1217 out tokens · 56440 ms · 2026-05-10T17:15:26.433138+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

  1. [1]

    0 \\ !Spq@+ +, xxI8 jN@4Xe 1\sjdiJ yY) hT2 C ^ 1\sjdiJ yY) w5 @# & .# .# .# - ( ( &

    11em plus .33em minus .07em 4000 4000 500 `\.=1000 = biography IEEEbiography biographynophoto IEEEbiographynophoto biography [2][] 1in =0pt 1em 500 \@IEEEtrantmpdimenA=\@IEEEBIOhangdepth \@IEEEtrantmpdimenA by \@IEEEBIOskipN \@IEEEtrantmpdimenA by 1 \@IEEEtranneedspace \@IEEEtrantmpdimenA \@IEEEBIOskipN plus 1fil minus 0 \@IEEEtempbiographybox 0pt minipag...

    work page doi:10.1109/tqe.202x.doi

  2. [2]

    Havlíček, V. et al. Supervised learning with quantum-enhanced feature spaces. Nature 567, 209--212 (2019)

    work page 2019

  3. [3]

    & Killoran, N

    Schuld, M. & Killoran, N. Quantum machine learning in feature Hilbert spaces. Phys. Rev. Lett. 122, 040504 (2019)

    work page 2019

  4. [4]

    & Kandola, J

    Cristianini, N., Shawe-Taylor, J., Elisseeff, A. & Kandola, J. On kernel-target alignment. In Advances in Neural Information Processing Systems (2001)

    work page 2001

  5. [5]

    & Rostamizadeh, A

    Cortes, C., Mohri, M. & Rostamizadeh, A. Algorithms for learning kernels based on centered alignment. J. Mach. Learn. Res. 13, 795--828 (2012)

    work page 2012

  6. [6]

    Hubregtsen, T. et al. Training quantum embedding kernels on near-term quantum computers. Phys. Rev. A 106, 042431 (2022)

    work page 2022

  7. [7]

    & Schölkopf, B

    Kübler, J., Buchholz, S. & Schölkopf, B. The inductive bias of quantum kernels. In Advances in Neural Information Processing Systems (2021)

    work page 2021

  8. [8]

    & Meyer, J

    Schuld, M., Sweke, R. & Meyer, J. J. Effect of data encoding on the expressive power of variational quantum machine learning models. Phys. Rev. A 103, 032430 (2021)

    work page 2021

  9. [9]

    & Dunjko, V

    Skolik, A., Jerbi, S. & Dunjko, V. Quantum agents in the gym: a variational quantum algorithm for deep Q-learning. Quantum 6, 720 (2022)

    work page 2022

  10. [10]

    Yao, J. et al. Policy gradient based quantum approximate optimization algorithm. arXiv:2002.01068 (2021)

    work page arXiv 2002

  11. [11]

    & Tao, D

    Du, Y., Hsieh, M.-H., Liu, T. & Tao, D. Quantum circuit architecture search for variational quantum algorithms. npj Quantum Inf. 8, 62 (2022)

    work page 2022