acceptodds
Under review as a conference paper at ICLR 2027

Discretely Parametrized Quantum Circuits as Quantum Classifiers

Abstract

Parametrized quantum circuits (PQCs) are typically parametrized continuously, whereas low-resolution quantum control motivates restricting circuit parameters to a discrete set. This distinction can become consequential at low resolution, where projecting a continuously optimized circuit onto executable parameter values can substantially perturb the implemented circuit and degrade task performance. We therefore study *discretely parametrized quantum circuits* (DPQCs), whose trainable circuit parameters are defined and optimized directly over a discrete set, in contrast to continuously parametrized quantum circuits (CPQCs). To better preserve input structure under finite resolution, we introduce resolution-aware angular re-uploading, which progressively uploads input information across multiple resolution levels during classical data encoding into the circuit. We further develop spectrum-informed shift-compatible optimization (SISCO), which exploits the spectral structure of Pauli-generated PQCs to optimize discrete circuit parameters. Across multiple benchmarks, our DPQC framework generally outperforms alternative approaches and is particularly effective in the low-resolution regime. With a nominal **5-bit resolution** for circuit parametrization (i.e., \(2^5 = 32\) executable values), our resolution-aware DPQCs trained with SISCO achieve **93.56%**, **77.27%**, and **86.56%** mean accuracy on Imagenette, CIFAR-10, and STL-10, respectively, closely matching the corresponding CPQC references of **92.44%**, **77.43%**, and **86.62%**. These results show that jointly designing data encoding and optimization for DPQCs can retain competitive performance without relying on continuous circuit parametrization, providing an algorithmic framework for variational quantum learning under low resolutions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
quantum machine learning (QML), parametrized quantum circuits, discretely parametrized quantum circuits

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_apSAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies parametrized quantum circuits whose trainable angles are restricted to a finite set A_K of K values spaced by 2 pi over K, motivated by finite control resolution. Two components are proposed. The first is resolution-aware angular re-uploading, which writes each input as B bit-level angular contributions and uploads them coarse to fine, with trainable CZ-conjugated mixing layers between levels, recovering ordinary finite-resolution angular encoding when the mixers are set to zero. The second is SISCO, a discrete optimizer. Because each circuit output is a first-order trigonometric polynomial in one angle, the 2P parameter-shift evaluations that a gradient would need also reconstruct the full coordinate-wise loss profile over all K admissible values. SISCO takes the per-coordinate loss-optimal move as a first search direction, selects M minus one further directions by a Fourier-descriptor criterion with a projection residual, and then runs a discrete line search along each. Evaluation covers 500 random circuit-matching instances under matched evaluation budgets, a two-spiral task, and Imagenette, CIFAR-10 and STL-10 with a frozen ResNet-18 front end at 4 qubits and 6 layers across three ansatzes, plus execution on IBM Kingston.

Critical strengths
  1. The profile reconstruction is the strongest idea in the paper. Since a Pauli-generated rotation makes the output a first-order trigonometric polynomial in that angle, two shifted evaluations give the whole K-point profile for free, and the same two shifts are what a gradient would have cost anyway. Turning that into a discrete line search is a clean use of an existing budget, and the budget-matched study in Figure 4 is the right way to show it.

  2. The empirical coverage is unusual for this literature. Three ansatzes, three image datasets, four qubit counts, three resolutions, and a QPU run, with the losing cells left visible. At K = 8 on Imagenette the method falls to 83.96 against 92.44 for the continuous reference, and the paper reports it.

Critical weaknesses
  1. The premise is not tied to any device. No measurement, calibration report, or citation fixes 2 pi over 32 as a real limit on gate angle, and a single-qubit rotation on superconducting hardware is typically calibrated far below that. Koczor et al. 2024 already study PQCs under finite angle resolution, so the first-study framing in Section 1 also needs qualifying. If the constraint is hypothetical, the paper is a study of discrete optimization and not a response to a hardware barrier. And at the one resolution where the constraint bites, K = 8, the discrete model loses about 8.5 points on Imagenette, so the parity claim lives at the least constrained setting.

  2. The headline comparison is not matched on circuit evaluations, and the encoding comparison is not matched on capacity. Tables 2 and 4 use the same number of parameter updates, while Appendix J allows 2P+1+M(K-1) evaluations per SISCO update against 2P+1 for a shifted gradient, which at K = 32 and M = 3 is 93 extra evaluations per update. Separately, the resolution-aware encoder adds B-1 mixing layers with n trainable angles each plus entangling layers, so the encoding gain in Tables 1 and 2 may be a parameter-count gain.

Questions for authors
  1. Is there a device, present or near-term, for which 2 pi over 32 is the practical limit on a rotation angle? If not, what physical controller does the DPQC abstraction correspond to?

  2. At K = 8 on Imagenette the discrete model reaches 83.96 and the continuous reference 92.44. Is that not the regime the paper is motivated by, and does a parity claim at K = 32 carry the intended message?

  3. Tables 2 and 4 hold parameter updates fixed. What happens to the comparison when circuit evaluations are held fixed instead, as in Figure 4, given the 93 extra evaluations per update at K = 32?

  4. The resolution-aware encoder has more trainable angles and more entangling layers than the conventional one by construction. What happens if the conventional encoder receives the same number of extra parameters and layers, without the bit-level structure?

  5. The descriptor in Eq. (2) depends only on the step indices and the spacing, not on the circuit or the loss. So the complementarity term in Eq. (3) is the same for every circuit. Which part of the selection is actually informed by the circuit's Fourier spectrum, and does the ablation in Appendix F separate the geometric diversity from the predicted-utility term?

  6. Table 3 gives 74.50 for CIFAR-10 and 85.17 for STL-10 under the ideal simulator, while Tables 2, 4 and 5 give 77.14 to 77.27 and 86.56 to 86.81 for the same datasets. Which configuration produced the ideal column in Table 3, and which number should the hardware gap be measured against?

  7. Table 1 reports a task loss from 0.32 down to 0.14 without naming the task, the resolution, or the metric. Which experiment does that table summarize?

  8. The six hundred test images per dataset give roughly sixty per class on CIFAR-10. Is the reported hardware gap, 73.83 against 74.50, resolved at that sample size?

4 comments

Part 1/2

We thank the reviewer for the questions. We address each point below.

Q1. Physical interpretation of the finite-resolution constraint

We model a controller that can implement rotations only from a prescribed finite set of angles. This is an abstract hardware design constraint, not the measured native precision of a particular device. The setting follows prior work, including Koczor et al. (Probabilistic interpolation of quantum rotation angles, PRL, 2024), who study uniformly restricted rotation-angle sets and discuss control considerations such as DAC precision, waveform memory, control bandwidth, and power requirements.

KK is the number of executable angles, and K∈8,16,32K \in \\{8, 16, 32\\} corresponds to three-, four-, and five-bit angular control. We do not claim that 2π/322\pi/32 is a measured limit of IBM Kingston or of any current processor; it is the finest resolution in our main image-classification experiments. We will revise the introduction to separate the hardware motivation from the abstract setting studied in this work.

Q2. Performance at coarse control resolution

The two numbers refer to different datasets: 83.96% is the STL-10 accuracy at K=8K=8, whereas 92.44% is the continuous reference on Imagenette. The dataset-matched comparisons at K=8K=8 are:

DatasetCPQC reference (%)resolution-aware (RA) encoding + SISCO (%)Gap (pp)
Imagenette92.4492.390.05
CIFAR-1077.4374.123.31
STL-1086.6283.962.66

The effect of coarse control is therefore dataset-dependent: Imagenette stays close to its continuous reference at K=8K=8, whereas CIFAR-10 and STL-10 degrade more. At K=32K=32, RA + SISCO reaches 93.56%, 77.27%, and 86.56%, respectively.

We do not claim that discretization is lossless at every resolution. Our conclusion is that jointly designed encoding and optimization mitigate finite-resolution degradation, and that moderately fine discrete parameter sets retain competitive accuracy on the evaluated tasks. Note also that K=32K=32 still restricts every trainable angle to 32 values; it is a finite-resolution model, not the continuous limit.

Q3. Evaluation-matched optimization

The reviewer's count is correct: at K=32K=32 and M=3M=3, the SISCO line search adds at most M(K−1)=3×31=93M(K-1) = 3 \times 31 = 93 evaluations per update. For classification under matched circuit-evaluation budgets and per-update wall-clock costs, please refer to our response to Reviewer WnN2, Points 1 and 7. We will state the distinction between the two evaluation protocols explicitly in the revised manuscript.

Q4. Capacity-matched conventional encoding

Please refer to our response to Reviewer WnN2, Point 2.

Part 2/2

Q5. Circuit dependence of the Fourier descriptor

We thank the reviewer for raising this distinction. The descriptor v(ci,cj)v(c_i, c_j) depends on the candidate direction components, the control resolution, and the known Fourier structure of Pauli-generated rotations. It does not depend on the measured or estimated Fourier coefficients of the current circuit. It characterizes which admissible pairwise Fourier interaction patterns a candidate search line can probe.

Circuit- and loss-dependent information enters through the profile-derived utility in Eq. (3). Direction selection therefore combines two sources of information:

  1. Profile-based utility, which depends on the current circuit and objective and predicts the benefit of a candidate displacement.
  2. Fourier-structured complementarity, which measures how a candidate differs from previously selected directions in the interaction patterns it probes.

The ablation in Appendix F isolates complementarity relative to greedy utility-only selection. For more comparisons, please refer to the randomized multi-line search baseline in our response to Reviewer WnN2, Point 3, where SISCO achieves lower mean loss at every tested MM. We will describe the descriptor as a Fourier-structured diversity criterion rather than an estimate of the learned spectrum of the circuit.

Q6. Table 3 ideal-simulator results and hardware reference

Table 3 compares hardware execution with an ideal-simulator evaluation of the same trained circuits, using the same class-balanced subset of 600 test images per dataset. Tables 2, 4, and 5 report averages over independent training seeds under their respective ansatz configurations and are therefore not the reference for the hardware comparison. The corresponding comparisons are:

DatasetIdeal simulator (%)IBM Kingston (%)
Imagenette93.6786.00
CIFAR-1074.5073.83
STL-1085.1784.50

We will report this configuration and the evaluation-subset protocol in the Table 3 caption and in the description of the hardware experiment.

Q7. Experimental setting of Table 1

Table 1 reports the final test cross-entropy on the two-spiral classification task. It compares conventional angular encoding with RA under identical training protocols for Nearest STE, Stochastic STE, and SISCO. We will revise the caption to state the task, metric, control resolution, encoding depth, and ansatz.

Q8. Statistical interpretation of the hardware accuracy gap

The 600-image subset was chosen to fit the available QPU budget while keeping the classes balanced. We agree that the difference between the simulator accuracy (74.50%) and the hardware accuracy (73.83%) on CIFAR-10 does not establish equivalent performance: the subset size limits statistical precision, and finite-shot measurement adds further uncertainty. Our conclusion is that the trained DPQC retains substantial predictive performance after compilation and execution on IBM Kingston, not that hardware execution is statistically equivalent to ideal simulation. We will state this interpretation in the revised manuscript.

Dear Reviewer apSA,

Thank you again for your reviews and for highlighting aspects of our presentation and experimental evaluation that required clarification. With one day remaining in the discussion period, we would greatly appreciate your feedback on our responses. We hope our responses have clarified your questions and concerns. If any points remain unresolved, we would welcome the opportunity to address them before the discussion closes.

Thank you for your time and consideration. Best regards, The Authors.

The authors conceded two points outright, that finite resolution is an abstract constraint rather than a device limit, and that the Fourier descriptor carries no circuit-specific spectrum. I accept both.

I also withdraw my K=8 comparison. 83.96 is STL-10 and 92.44 is Imagenette, and the dataset-matched gaps are 0.05, 3.31 and 2.66 points.

The evaluation-matched comparison on image classification is not resolved. The response offers one qualitative sentence and matched-budget curves on 500 synthetic instances, but no numbers for Tables 2 and 4 under matched circuit evaluations, which is where the 93 extra evaluations per update actually land.

The capacity-matched control is not resolved either. RA gives 95.92 plus or minus 2.54 against 94.40 plus or minus 2.50 for conventional RU. That 1.52 point gap is smaller than either standard deviation, and by the authors' own test in W2, where 0.38 under standard deviations of 0.45 was read as no difference, this cell cannot carry the encoding claim.

One number from their new hardware table also goes unremarked. Imagenette drops from 93.67 on the simulator to 86.00 on Kingston, a 7.67 point gap, while the text discusses only the 0.67 point CIFAR gap. I keep my rating.

Reviewer_AYBmOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

PQCs optimize trainable parameters on continuous spaces which are limited by finite-resolution hardware. In response to this problem, the authors claim to have made three contributions:

  • DPQCs in which P trainable rotation angles are restricted to K evenly spaced values (from a finite parameter set A_k)
  • Resolution-aware angular re-uploading in which a normalized feature vector (from the input x) is B-bit quantized. Each bit then gets its' own rotation with interleaved trainable mixers inserted between consecutive bit uploads
  • SISCO in which the DPQC’s angles are optimized directly over A_k. Grid-compatible +/- pi/2 parameter shifts reconstruct each angle’s sinusoidal loss profile, and a line search runs along a few spectrum-informed, interaction-complementary directions. Simulations show SISCO consistently outperforms STE baselines at matched budgets. It nearly matches continuous solutions in two-spiral classification (0.99 vs. 1.00), avoiding the accuracy degradation of post-hoc rounding (0.94). Combined with resolution-aware re-uploading—which reduces loss across all optimizers—the 5-bit (K=32) framework matches continuous circuit performance within one standard deviation on Imagenette, CIFAR-10, and STL-10. This allowed the authors to conclude that co-designing encoding and optimization preserves near-continuous performance under finite-resolution control.
Critical strengths
  • Well-posed problem motivated by a provable worst-case projection bound and a post-hoc diagnostic isolating the cost of rounding
  • Theoretically grounded design via exact properties (such as profile reconstruction, shift closure and encoder resting) with heuristic direction selection
  • Sound progressive evaluation (with increasing complexity tasks) allowing for the separation of representation vs. learning effects
  • Appropriate use of cumulative circuit evaluations to showcase dominant costs on quantum hardware
  • Honesty of scaling by acknowledging NP-hardness and scale relative to classical algorithms
Critical weaknesses
  • Overstated Novelty due to a lack of proper literature comparison to similar hybrid quantum-classical algorithms such as Bhattacharjee et. al, 2026 (https://arxiv.org/abs/2601.04983) and discrete Clifford angle search (https://arxiv.org/abs/2202.12924)
  • Continuous reference encoding (CPQC) is never explicitly stated, so the parity claim cannot separate the effect of discretizing parameters from that of differences in encoding.
  • Mismatched image budgets, as methods get equal training steps but not equal circuit runs during training. SISCO's advantage may not account for extra computation
  • Since the only baseline is a single conventional upload, gains cannot be credited to bit decomposition over re-uploading, which results in an confounded encoder comparison
Questions for authors

Ansatz inconsistency: the default configuration is the hardware-efficient ansatz, but the headline results in Table 2, and hence the abstract, use the strongly entangling ansatz. App. K then states that the experiments are repeated with the strongly entangling and Hadamard-staggered ansatzes, yet Table 4 reports the hardware-efficient one (p. 23). Please state the ansatz for every experiment, including Table 1, Figures 6 and 9. Justify why the strongly entangling ansatz was chosen for the main results. Correct the App. K text.

2 comments

We thank the reviewer for the questions. We address each point below.

1. Novelty and relation to prior work

We thank the reviewer for pointing out these works. Bhattacharjee et al. (2026) is indeed concurrent with our work, and their method applies directly to our setting. We will therefore remove the statement that ours is the first work in this direction. Bhattacharjee et al. optimize finite-resolution circuits through temperature-controlled stochastic rounding of gradient-based parameter proposals.

We also note that Bhattacharjee et al. note that parameter-shift evaluations may fall outside the quantization grid, and they use simulator automatic differentiation in their experiments. In contrast, our Lemma 1 establishes the condition under which all shifted evaluations required for analytical gradients remain executable on the discretized grid and they not be un-executable.

We compared the optimization behavior of both methods on our large number of randomly constructed circuit instances, using all temperature settings, i.e. hyperparameters, considered in their work as we can not guarantee their optimal hyperparameter directly transfers to our settings. These resuls complete Figure 4.

Mean optimization loss:

MethodTBudget 825Budget 1,245Budget 1,665Budget 2,085
Bhattacharjee et al.0.50.079200.088550.082280.08460
Bhattacharjee et al.10.108070.105050.104720.10595
Bhattacharjee et al.50.124410.117620.124050.13256
Bhattacharjee et al.100.123510.125460.120730.13281
SISCO-0.032320.012200.008990.00671

SISCO achieves lower mean loss at every circuit evaluation budget and temperature.

CAFQA (Ravi et al., 2022) searches Clifford configurations to initialize subsequent variational optimization, exploiting efficient Clifford simulation. In our setting, RA contains non-Clifford input rotations, and restricting the trainable angles to Clifford points would change the model class. We therefore discuss CAFQA as related work and use discrete coordinate moves and randomized multi-line search as more direct baselines (please refer to our response to Reviewer WnN2, Point 3).

2. Continuous reference and the contribution of encoding

We refer to the continuously parametrized quantum circuit as CPQC. CPQC is the conventional PQC with data uploading and ansatz, followed by measurement. It uses conventional continuous-angle embedding followed by the designated variational ansatz. The comparison between CPQC and DPQC (our resolution-aware encoding + SISCO) is therefore a system-level comparison. Its purpose is to show that properly designed encoding and optimization can compensate for the performance loss caused by parameter discretization.

Table 2 provides the component-wise comparison. For example, at K=8K=8, starting from conventional encoding with Nearest STE, the gaps decompose as follows (in percentage points):

DatasetInitial gap to CPQCRecovered by SISCOFurther recovered by RARemaining gap
Imagenette8.745.553.140.05
CIFAR-107.872.901.663.31
STL-1010.164.163.342.66

3. Unequal circuit-evaluation budgets

We agree that equal update counts do not imply equal circuit-evaluation budgets for the image classification results reported. Please refer to our response to Reviewer WnN2, Point 1 (classification under matched evaluation budgets) and Point 7 (per-update overhead).

4. Conventional re-uploading and encoder capacity

Please refer to our response to Reviewer WnN2, Point 2.

5. Ansatz inconsistency

We thank the reviewer for identifying this issue. The default-configuration statement, the Table 2 caption, and Appendix K contained typos that made the configurations inconsistent. The corrected assignments are:

ExperimentAnsatz
Table 1 and Fig. 5HEA
Table 2 (CPQC reference and all DPQC configurations)HEA
Table 4SEA
Table 5HSA
Fig. 6 and Fig. 9HEA
New encoder controls and finite-shot experimentsHEA

These corrections do not affect our conclusions, though.

Dear Reviewer AYBm,

Thank you again for your reviews. We appreciate your positive assessment and thorough suggestions, which have helped us clarify the relationship to prior work and the experimental configurations. We hope our responses have clarified your questions and concerns. As the discussion period closes in one day, we would welcome your feedback on these clarifications and would be happy to respond to any further questions.

Thank you for your time and consideration. Best regards, The Authors

Reviewer_WnN2OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies quantum classifiers with discrete gate angles and proposes resolution-aware input encoding and SISCO, a spectral multi-line optimizer. Experiments show that low-resolution circuits can approach continuous-parameter accuracy on several classification benchmarks.

Critical strengths
  • Addresses a practically motivated constraint: learning with finite-resolution quantum gate parameters.

  • Combines discrete optimization and resolution-aware encoding in a coherent framework.

  • Exploits Pauli-rotation spectra to reconstruct coordinate profiles efficiently while keeping evaluations executable.

  • Evaluates optimization across 500 random instances with matched circuit-evaluation budgets.

  • Includes a controlled ablation supporting the Fourier-complementarity criterion.

Critical weaknesses
  • Table 2 matches parameter updates, although SISCO uses more circuit evaluations per update.

  • The proposed encoder adds trainable mixers and circuit depth; improvements need comparisons with equally sized conventional re-uploading circuits.

  • Comparisons mainly cover two STE methods; discrete coordinate search and randomized multi-line search would strengthen the evaluation.

  • Default experiments use four qubits, leaving scalability with qubit count and parameter count unclear.

  • Frozen ResNet-18 features may explain much of the image accuracy; matched classical classifier baselines are missing.

  • Hardware experiments test inference, not finite-resolution control constraints or SISCO training under shot noise.

  • Classical direction-selection costs and total training time are not quantified.

  • Some accuracy differences are small relative to seed variability; being within one standard deviation does not establish equivalence.

Questions for authors

See weaknesses

3 comments

We thank the reviewer for the questions. We address each point below.

Part 1:

1. SISCO uses more circuit evaluations in Table 2

Yes, this is correct. Our experiments were designed to address two complementary questions. Fig. 4 compares optimization efficiency under matched circuit-evaluation budgets on randomly constructed circuit instances, independent of specific applications. Table 2 uses image classification as a concrete application and, following standard machine-learning practice, compares converged performance at matched update counts; it therefore does not equalize quantum-evaluation cost. Each SISCO update requires at most M(K−1)M(K-1) more circuit evaluations than a parameter-shift update.

To address this concern, we extended the training of both STE baselines until their cumulative circuit evaluations matched those of SISCO.

Test accuracy (%):

Nearest STE with extended training:

KImagenetteCIFAR-10STL-10
890.41 ± 4.0771.02 ± 5.6083.22 ± 5.32
1693.15 ± 1.9476.08 ± 2.6982.87 ± 2.74
3293.57 ± 0.9676.43 ± 0.8886.24 ± 1.84

Stochastic STE with extended training:

KImagenetteCIFAR-10STL-10
887.72 ± 5.6974.29 ± 1.6478.24 ± 8.37
1692.36 ± 3.6176.27 ± 2.0481.13 ± 5.22
3291.91 ± 1.4474.52 ± 2.2885.67 ± 1.75

Compared with Table 2, the additional epochs do not substantially improve either STE baseline.

2. Encoder capacity and conventional re-uploading controls

We added two experiments in Table 1:

  • Resource-matched conventional uploading: conventional angular encoding followed by an ansatz matched to our resolution-aware re-uploading (RA) in trainable parameters.
  • Conventional re-uploading (RU): the same upload count, interleaved mixers, trainable parameters, and gate budget as RA. RU uploads the full input angle at every stage, whereas RA distributes bit-level contributions across successive stages.
EncoderTest Top-1 (%)Test CE
Single upload (resource-matched)93.68 ± 6.620.1698 ± 0.1146
Conventional re-uploading (RU)94.40 ± 2.500.1729 ± 0.1059
Resolution-aware re-uploading (RA)95.92 ± 2.540.1148 ± 0.0470

RA achieves the highest mean accuracy and approximately 34% lower cross-entropy than RU.

Spectral mechanism. We provided a theoretical analysis of our RA compared to RU since they all require repeated data uploads. For a input mjm_j, the input dependence of a measured observable is characterized by its discrete Fourier coefficients. A single conventional upload yields frequencies in 0,±1\\{0, \pm 1\\} along each input dimension. With rr uploads of the same angle, RU has support in −r,…,r\\{-r, \dots, r\\}. The number of spectral paths grows exponentially with rr, but many paths map onto the same frequency (spectral degeneracy). For example, four uploads produce 34=813^4 = 81 paths but only nine signed frequencies, −4,…,4\\{-4, \dots, 4\\}.

RA instead decomposes the input into BB binary contributions. At resolution level bb, the uploaded rotation is

$ R_X\left(\frac{2\pi \xi_{j,b}}{2^b}\right) = I + \xi_{j,b}\left[R_X\left(\frac{2\pi}{2^b}\right) - I\right], \qquad \xi_{j,b} \in \{0,1\}, $

and the corresponding bit function has discrete Fourier support

$ \Omega_b = \{0\} \cup \left\{(2r+1)\, 2^{b-1} : 0 \le r < 2^{B-b}\right\} \pmod{2^B}. $

Different bits therefore occupy distinct dyadic frequency bands before cross-level interactions are taken into account. For B=4B=4 (i.e., four uploads, as for RU), the nonnegative frequencies of real-valued responses are:

Encoder (four uploads)Frequency support
Single upload{0, 1}
RU{0, 1, 2, 3, 4}
RA{0, 1, 2, 3, 4, 5, 6, 7, 8}

In RA, each bit contributes distinct nonzero frequencies: {1, 3, 5, 7} for bit 1, {2, 6} for bit 2, {4} for bit 3, and {8} for bit 4. Together, they cover all nine nonnegative Fourier indices, whereas four conventional uploads are confined to indices 0 to 4. RA can therefore represent frequency components that an equally sized RU circuit cannot. The trainable mixers determine how these components interact.

Do trained RA models use the frequencies that RU cannot represent? We removed all Fourier components outside the RU support from the learned RA responses:

SettingTest accuracy (%)Test CE
Original RA97.25 ± 3.180.1099 ± 0.0829
Spectrally filtered RA94.50 ± 5.660.1326 ± 0.1028

This removal lowers mean accuracy by 2.75 percentage points and raises mean cross-entropy by approximately 21%. The additional spectral components not representable by RU therefore contribute to the predictive performance of trained RA models. Together with the resource-matched comparison, this indicates that the benefit of RA stems from its resolution-structured encoding rather than merely from added depth or parameters.

Part 2:

3. Discrete coordinate search and randomized multi-line search

We added both baselines.

Discrete coordinate moves (Rotosolve style). The first SISCO direction already combines the coordinate-wise loss-optimal displacements into a single joint move, i.e., a multi-parameter update. We use the large number of randomly constructed circuit instances as described in the main text, and for each of them with the same initialized parameters, we report the resulting loss:

Mean loss:

MethodM = 5M = 9M = 13
Optimal discrete coordinate-wise update0.035860.035860.03586
SISCO joint search0.031870.023210.02053

The joint search reduces the mean loss by approximately 11%, 35%, and 43% relative to the best discrete coordinate move, and the gain grows with MM. The joint move accounts for correlations among simultaneously updated parameters, which matters especially if update steps are large.

Randomized multi-line search. To isolate the effect of Fourier-guided selection, we draw the search directions uniformly at random instead of using our Fourier-guided rule. Both methods use the same number of search lines and differ only in how the lines are chosen.

Mean loss:

MethodM = 5M = 9M = 13
Random multi-line search0.015290.012180.00981
SISCO0.012200.008990.00671

Fourier-guided selection lowers the mean loss by approximately 20%, 26%, and 32% relative to random selection. The two comparisons show complementary effects: joint search reaches improvements beyond discrete coordinate search, and Fourier-guided selection identifies more effective search directions than random sampling. All values are averaged over the 500 randomly constructed circuit instances.

4. Scalability with circuit size

We indeed have the scalability study in Appendix L (Fig. 9) covers n∈4,6,8,10,12n \in \\{4, 6, 8, 10, 12\\} qubits and L∈6,8,10L \in \\{6, 8, 10\\} layers on Imagenette, CIFAR-10, and STL-10. The gap between DPQC and its continuous reference does not grow systematically with the number of qubits or layers.

5. Contribution of frozen ResNet-18 features

Our work studies parameter discretization within the established quantum transfer-learning framework of Mari et al. (2020) (https://quantum-journal.org/papers/q-2020-10-09-340/), which has been studied extensively in the QML literature. In this framework, a pretrained classical feature extractor supplies compact representations to a downstream quantum classifier, and we acknowledge the importance of these features.

Our aim is not to build circuits that perform better on these features. Our question is to what extent quantum classifiers in this framework retain their performance when parametrized discretely rather than continuously, and how encoding and optimization can relief the resulting degradation. The role of pretrained features is studied in Kim et al. (2023, https://doi.org/10.1016/j.neucom.2023.126643) and Khatun and Usman (2025, https://doi.org/10.1002/qute.202400268), which may be of interest to the reviewer.

6. Finite-shot training and hardware experiments

Our IBM Kingston experiment evaluates the execution of trained DPQCs and does not demonstrate on-device training. Full on-device SISCO training under native hardware noise exceeds our available QPU allocation and is beyond the scope of this work.

To study the effect of measurement noise during optimization, we trained SISCO and both STE baselines and will include them besides Fig.5.

Final test accuracy (%):

Shots per circuitNearest STEStochastic STESISCO
25686.50 ± 9.6384.20 ± 6.3590.40 ± 8.35
1,02492.50 ± 6.5490.70 ± 5.8493.90 ± 3.31

SISCO achieves the highest mean accuracy at both shot counts, which supports its effectiveness when training evaluations are subject to finite-shot sampling.

7. Classical direction-selection overhead and training time

We measured the per-update wall-clock time in our experimental environment:

OperationMedian wall-clock time
Parameter-shift gradient + Adam update14.48 s
Full SISCO update18.52 s
Fourier-guided direction selection (included in the full SISCO update)1.40 s

8. Statistical equivalence and seed variability

We agree that overlap within one standard deviation does not establish equivalence. We will revise such claims and report the observed gaps directly.

Dear Reviewer WnN2,

Thank you again for your thoughtful review. Your comments have been valuable in strengthening our evaluation, particularly regarding computational budgets, encoder capacity, and optimization baselines. With one day remaining in the discussion period, we hope our responses have clarified your questions and concerns. We would greatly appreciate your feedback and would be happy to address any unresolved points.

Thank you for your time and consideration.

Best regards, The Authors

Comments on the paper

2 comments

We thank all reviewers for their careful assessments and constructive suggestions. The reviewers recognized the practical relevance of finite-resolution circuit parametrization (WnN2, AYBm), the shift-compatible reconstruction of coordinate-wise circuit profiles (WnN2, AYBm, apSA), the evaluation-matched optimization benchmark (all reviewers), and the breadth of our empirical evaluation (apSA). Here, we carefully address all your concerns.

AuthorsOpenReview ↗

Rebuttal summary for Area Chair

We thank the Area Chair and reviewers for their constructive feedback.We appreciate the reviewers' recognition of the practical relevance of finite-resolution quantum circuit parametrization, the theoretically grounded optimization framework, and the breadth of our experimental evaluation. We summarize the additional evidence and clarifications below.

1. Evaluation-Matched Optimization

Concern: Unequal circuit-evaluation budgets for the image classification experiment (Rev. WnN2, AYBm, apSA).

Extending both STE baselines to match SISCO's cumulative circuit-evaluation budgets yielded no substantial improvement (Rev. WnN2, Point 1). We also compared SISCO against concurrent method: Bhattacharjee et al. (2026) under matched budgets across all reported temperature settings.

Mean loss ↓8251,2451,6652,085
Bhattacharjee et al. (T=0.5)0.079200.088550.082280.08460
Bhattacharjee et al. (T=1)0.108070.105050.104720.10595
Bhattacharjee et al. (T=5)0.124410.117620.120450.13256
Bhattacharjee et al. (T=10)0.123510.125460.127030.13281
SISCO (ours)0.032320.012200.008990.00671

SISCO achieves lower loss at every evaluated budget and temperature (Rev. AYBm, Point 1).

2. Capacity-Matched Encoding and Spectral Expressivity

Concern: Our resolution-aware (RA) encoding gains may arise from increased circuit capacity (Rev. WnN2, AYBm, apSA).

We added resource-matched single-upload and conventional re-uploading (RU) baselines.

EncoderAccuracy (%) ↑Test CE ↓
Single upload (resource-matched)93.68 ± 6.620.1698 ± 0.1146
Conventional RU94.40 ± 2.500.1729 ± 0.1059
RA (ours)95.92 ± 2.540.1148 ± 0.0470

RA reduces test CE by approximately 34% relative to RU. Our theoretical spectral analysis further supports this empirical observation. It demonstrates that conventional uploads support limited spectral expressivity, whereas our resolution-aware uploads extend the support. Removing the additional spectral components gives:

SettingAccuracy (%) ↑Test CE ↓
Original RA97.25 ± 3.180.1099 ± 0.0829
Spectrally filtered RA94.50 ± 5.660.1326 ± 0.1028

The accuracy reduction of 2.75 percentage points supports the contribution and necessasity of the additional frequencies (Rev. WnN2, Point 2).

3. Discrete Optimization and Fourier-Guided Search

Concern: The benefits of joint discrete optimization and Fourier-guided direction selection require stronger controls (Rev. WnN2, AYBm).

We added two baselines over a large number of 500 randomly constructed circuit instances.

Discrete coordinate search (mean loss ↓):

MethodM = 5M = 9M = 13
Optimal coordinate-wise update0.035860.035860.03586
SISCO joint search0.031870.023210.02053
Loss reduction11%35%43%

Randomized multi-line search (mean loss ↓):

MethodM = 5M = 9M = 13
Random multi-line0.015290.012180.00981
SISCO0.012200.008990.00671
Loss reduction20%26%32%

These comparisons separately support joint parameter updates and Fourier-guided direction selection (Rev. WnN2, Point 3).

4. Finite-Shot Training and Computational Cost

Concern: Measurement noise during training and computational overhead (Rev. WnN2, apSA).

We conducted finite-shot training experiments:

Shots/circuitNearest STE (%)Stochastic STE (%)SISCO (%) ↑
25686.50 ± 9.6384.20 ± 6.3590.40 ± 8.35
1,02492.50 ± 6.5490.70 ± 5.8493.90 ± 3.31

SISCO achieves the highest mean accuracy at both shot counts.

OperationMedian time
Parameter-shift + Adam14.48 s
Full SISCO update18.52 s
Fourier-guided selection (included above)1.40 s

We further clarified the hardware evaluation protocol and statistical limitations of the 600-image subset (Rev. WnN2, Points 6–7; Rev. apSA, Points 6–8).

5. Scope and Experimental Clarifications

We clarified that finite-resolution control is an abstract hardware design constraint, not a claimed precision limit of existing processors, and acknowledged relevant concurrent work (Rev. apSA, Point 1; Rev. AYBm, Point 1).

We also addressed continuous-reference comparisons, coarse-resolution performance, circuit scalability, frozen classical features, ansatz configurations, and statistical interpretation in the detailed responses.

We hope these additional results and clarifications provide a clearer basis for evaluating our contributions. The corresponding changes have been reflected in the work. We appreciate the Area Chair's consideration of our rebuttal and additional evidence.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.