Pith. sign in

REVIEW 4 major objections 4 minor 41 references

A compact variational quantum circuit with a linear readout can match transformer-based forecasters on short-horizon multivariate time-series tasks while using far fewer parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 22:14 UTC pith:LF2FVDXQ

load-bearing objection A worthwhile comparison with an unusually honest limitations section, but the experiments don't implement the joint-encoding architecture described in the methods, so the headline claims about cross-channel entanglement and the necessity of trainable quantum features are unsupported as written. the 4 major comments →

arxiv 2607.16358 v1 pith:LF2FVDXQ submitted 2026-07-17 cs.IR cs.AIcs.ET

A Quantum-Classical Hybrid Framework for Multivariate Time-Series Forecasting Complexity-Fidelity Trade-offs and Limitations

Sanjay Chakraborty , Fredrik Heintz This is my paper
classification cs.IR cs.AIcs.ET
keywords variational quantum circuitquantum reservoir computingmultivariate time-series forecastingmulti-horizon predictionparameter-shift rulequantum-classical hybridangle encodingMIMO readout
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a shallow variational quantum circuit — six qubits per channel, depth two, with DCT-compressed angle encoding — paired with a linear readout can match or beat strong transformer forecasters on short-horizon multivariate forecasts, and that this accuracy comes with substantially lower parameter counts. It contrasts this VQF-F model against QRC-F, a fixed random quantum reservoir, and reports that trainable quantum feature extraction is essential: the fixed reservoir is systematically inferior in accuracy, though it is more robust to circuit noise and needs no gradients. The paper also frames a complexity–fidelity trade-off: short horizons favor the cheap variational circuit, while long horizons degrade toward near-mean predictions because of the DCT bottleneck and rank-limited linear readout. The stated goal is a practical quantum-native forecasting framework deployable on near-term NISQ hardware.

Core claim

The central claim is that a trainable variational quantum circuit, using uniform quantization and RY-angle encoding followed by cross-channel controlled-Z entanglement and a shared MIMO linear readout, provides competitive or superior multivariate forecasting accuracy across seven benchmark datasets and four horizons while using far fewer parameters than attention-based baselines. The paper further claims that trainable variational feature extraction is necessary — fixed random quantum reservoirs (QRC-F) are systematically insufficient for modelling real-world multivariate dynamics — and that QRC-F's advantage lies instead in circuit fidelity and training stability under hardware noise. The

What carries the argument

The variational quantum circuit (VQF-F) — composed of RY and RZ single-qubit rotations and CNOT entangling layers, trained via the parameter-shift rule — is the central trainable object. It extracts Pauli expectation values (single-qubit and pairwise correlators) that feed a classical linear MIMO head that predicts all horizons at once. The cross-channel controlled-Z entanglement layer is the designed mechanism for inter-variable dependency capture, while DCT-based compression (retaining 6 of 96 coefficients) and uniform b-bit quantization form the encoding bridge from real-valued series to rotation angles.

Load-bearing premise

The paper's conclusions rest on the assumption that the reported experiments implemented the jointly entangled M·L-qubit circuit described in the methods; the reported resource counts instead match a per-channel model with no cross-channel entanglement, so the central dependency-capture mechanism may never have been evaluated.

What would settle it

Inspect the actual circuit construction used in the experiments, or reproduce the benchmarks with and without the cross-channel CZ gates: if the logged gate counts and memory figures contain no cross-channel operations, or if accuracy is identical after removing them, the claim that entanglement captures inter-variable dependencies is not supported. A second check is to compare VQF-F against a purely classical linear forecaster using the same DCT-compressed features; matching accuracy would indicate the quantum circuit is not the source of the reported gains.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, short-horizon multivariate forecasting can be done with roughly 52,000 trainable parameters and linear complexity instead of quadratic self-attention, making deployment practical on resource-constrained and near-term quantum hardware.
  • The reported failure of QRC-F implies that fixed random quantum mappings are not a viable forecasting backend; any practical quantum forecaster needs trainable parameters or a different mechanism.
  • The long-horizon degradation to near-mean predictions is attributed to the DCT bottleneck and rank-limited linear readout, not overfitting, so improvements should target those components through deeper ansätze, nonlinear readouts, or relaxed DCT compression.
  • Both variants are memory-efficient compared with transformer baselines, with the largest dataset reported at about 2.1 GB, and training stability is claimed to be high even in the variational case.
  • The exponential density-matrix simulation cost O(2^{2n}) limits the joint-encoding version to small qubit counts, motivating channel-independent or patch-based encoding variants for larger systems.

Where Pith is reading between the lines

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

  • Editorial inference: The resource tables and memory figures describe a per-channel model with 6 qubits per channel and no explicit cross-channel gates, whereas the method section defines a joint M·L-qubit state with CZ entanglement across channels; the reported experiments therefore appear not to test the cross-channel dependency claim as designed.
  • Editorial inference: The performance gap between VQF-F and QRC-F may be driven more by the trainable linear readout and DCT feature vector than by the quantum circuit itself; a classical linear model using the same DCT-compressed features would isolate the quantum contribution.
  • Editorial inference: A direct ablation that disables the CZ entanglement in VQF-F and re-runs the benchmarks would settle whether the entanglement layer contributes anything beyond the classical feature extraction; if accuracy is unchanged, the stated mechanism is not load-bearing.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces QTSF, a hybrid quantum-classical forecasting framework with two feature-extraction backends: QRC-F (fixed random unitary reservoir) and VQF-F (trainable variational circuit). The formal method in §IV applies b-bit quantization and angle encoding to multivariate series, constructs a joint tensor-product state over M channels and L timesteps (Eq. 24), adds cross-channel CZ entanglement (Eq. 25), and measures local and pairwise observables; a shared linear MIMO head predicts all horizons and an arccos-based inverse map reconstructs the signal. Experiments on seven benchmarks report VQF-F as consistently superior/competitive against classical (Tables III–IV) and quantum (Table VI) baselines, with QRC-F as a poorer parameter-free baseline. Resource, noise, and complexity analyses are given in Tables VII–XV. The conclusion asserts that trainable variational quantum feature extraction is essential and that cross-channel entanglement captures multivariate dependencies.

Significance. If correct, the paper would provide a compact NISQ-era alternative to transformer forecasters and a systematic comparison of fixed vs trainable quantum feature extraction. The idea of benchmarking a quantum reservoir against a variational circuit under a shared readout, and of analyzing complexity–fidelity tradeoffs, is genuinely relevant. The paper also contains extensive resource and complexity tables and an unusually candid limitations section. However, the reported experiments do not implement the architecture defined in §IV, the headline entanglement claim is untested, and several broad conclusions outrun the data. No code or supplement is provided, so the empirical claims cannot be independently checked. The potential significance does not compensate for these load-bearing inconsistencies.

major comments (4)
  1. [§IV-B/C vs §V-B/C, Tables II/XIII] Eq. (24) defines n=M·L qubits, Eq. (25) adds cross-channel CZ entanglement, and Eqs. (29)/(31) set feature dimension n+C(n,2). For ETTh1 (M=7, L=96) this gives n=672 (or n=42 after DCT compression); 2^42×16 bytes ≈64 TiB, impossible with the reported 95.6 GB GPU. Table II instead gives 6 qubits per channel, 168 quantum params (=7×6×2×2), and 52,416 classical params at H=96, which imply d_f=11 per channel — seven independent 6-qubit circuits with classical concatenation, not the joint entangled state. Table XIII memory (708–2004 MiB) confirms. Hence cross-channel entanglement was never implemented/evaluated; the abstract's entanglement claim is unsupported, and §VI.6 admits no verification mechanism.
  2. [§IV-A vs §V-B/C, Figs. 4–5] The formal pipeline is normalization, min-max scaling, b-bit quantization, and angle encoding of raw look-back values. The experiments add DCT compression (retaining 6 of 96 coefficients) before encoding (Table II, §V-C4, Fig. 4). This DCT step is absent from §IV, changes the quantity actually encoded, and is later cited as the cause of long-horizon collapse (§VII). Either fold DCT into the formal method with a correctness analysis, or report results without this extra transformation.
  3. [Table IV, §V-C, §VII] The conclusion that 'fixed random quantum mappings are systematically insufficient' is drawn from QRC-F results reported for only ETTh1, ETTh2, and Exchange; the remaining four rows are '-'. Table VI contains no QRC-F entries. A claim of systematic insufficiency across seven benchmarks needs results on all of them (or explicit justification for exclusion). The current evidence supports only a dataset-specific comparison.
  4. [Abstract/Introduction vs Tables X–XI] The abstract credits QRC-F with 'enhanced robustness and circuit fidelity under hardware noise,' but Table X reports overall fidelity QRC-F=0.856 and VQF-F=0.928, and Table XI lists VQF-F as the fidelity winner. These values are also computed from assumed depolarizing rates (p≈1e-3, p_2q≈1e-2), not measured on hardware. The claims about QRC-F should be aligned with the evidence (e.g., training stability/gradient-free operation) and the fidelity numbers presented as model-based estimates.
minor comments (4)
  1. [Table IV, §VII] The variant is called QRC-F in the abstract and most of the paper, but Table IV and §VII use 'QCR-F'. Please correct the typo.
  2. [Eqs. (35)–(37)] The arccos inverse mapping is applied to the output of a linear layer, which is not constrained to [-1,1]; clipping is mentioned only in Eq. (35) and not analyzed. Please justify this decoding choice or replace it with a defined readout.
  3. [Table II / Eq. (31)] The reported 52,416 classical parameters at H=96 imply d_f=11 per channel, but Eq. (31) gives d_f = n + C(n,2) = 6+15 = 21 for n=6. The parameter-count arithmetic is inconsistent with the formal feature dimension; please define d_f explicitly.
  4. [General reproducibility] No code, supplement, or random seeds are provided, although the paper claims details are in a supplementary file and that baselines were 'reproduced' using their settings. Tables III, IV, and VI report averages over four horizons without standard deviations; given small margins (e.g., ETTh2: VQF-F 0.397 vs iTransformer 0.393), 'consistently superior' is not supported by the reported data.

Circularity Check

1 steps flagged

Fidelity/robustness conclusion is computed from the assumed depolarizing noise model, not measured; the central accuracy results are genuine external benchmarks and are not circular.

specific steps
  1. self definitional [Table X (Noise and Fidelity Analysis), §V-C.3; abstract robustness claim]
    "Single-qubit depolarizing: E(ρ) = (1−p)ρ + p/3 (XρX + YρY + ZρZ), p∼10^-3; Two-qubit depolarizing: p2q ∼10p1q ∼10^-2; Overall fidelity QRC-F Product model ≈0.856; Overall fidelity VQF-F Product model ≈0.928"

    The reported 'overall fidelity' values are obtained by plugging the paper's own assumed per-gate depolarizing rates and each architecture's gate count into a product-fidelity formula. The Table XI winner (VQF-F, 0.928 vs 0.856) is therefore the arithmetic consequence of VQF-F having fewer two-qubit gates, and the conclusion that QRC-F provides 'enhanced robustness and circuit fidelity under quantum noise' (abstract) merely restates these assumed error rates and gate counts. No hardware measurement, calibrated noise model, or error-mitigation experiment is reported, so the robustness result is equivalent to its input assumptions by construction.

full rationale

The paper's forecasting comparison is a genuine experimental evaluation against external classical baselines (Autoformer, Informer, iTransformer, PatchTST, etc.) and published quantum-classical models, so the headline accuracy claims are not derived from the framework's own equations and do not reduce to their inputs. The only load-bearing 'prediction' that is definitionally forced is the noise/fidelity analysis: the per-gate depolarizing rates in Table X are assumed inputs, and the overall fidelity values are computed from them, making the subsequent robustness/fidelity comparison a restatement of the assumed noise model and gate-count arithmetic rather than an empirical result. I do not count the QRC-F vs VQF-F accuracy gap as circular: it is an empirical comparison of two different feature extractors, and a fixed random feature map could in principle be sufficient, so the conclusion that trainable features are needed is not tautological. The manuscript's own limitation §VI.6 ('the framework currently lacks a mechanism to verify whether quantum entanglement effectively captures inter-channel relationships') and the resource mismatch between the joint n=M·L encoding (Eq. 24-25) and the reported 6-qubit-per-channel configuration (Table II, Table XIII) concern validity and reproducibility of the cross-channel entanglement claim, not circularity of a derivation. Score 3 reflects one secondary circular/self-definitional pillar while the central benchmark results remain independent.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The framework rests on hand-chosen architectural constants (b = 8, n = 6, D = 2, DCT retention = 6) that are neither derived nor ablated, on a domain assumption that quantization-plus-angle-encoding preserves forecasting information, on assumed (not measured) noise rates, and on the experiment-vs-method mismatch described in the soundness rationale. No new physical entities are introduced.

free parameters (7)
  • quantization bits b = 8
    Hand-chosen (Table II); sets encoding precision and the claimed reconstruction bound Eq. (49).
  • qubits per channel n = 6
    Hand-chosen (Table II); under joint encoding the total would be M·n, which is inconsistent with the feasibility of the reported experiments.
  • circuit depth D = 2
    Hand-chosen (Table II); the barren-plateau discussion concerns deeper circuits, but all results use D = 2.
  • DCT coefficients retained = 6
    Hand-chosen (Table II); discards high-frequency content, and the paper attributes long-horizon degradation to this choice (Figs. 4–5 analysis).
  • reservoir couplings J_jk = random, distribution unspecified
    Eq. (27): the reservoir unitary depends on random couplings; no distribution or seed is given, which matters for QRC-F reproducibility.
  • noise rates p, p_2q = 1e-3, 1e-2
    Table X: assumed, not measured, depolarizing rates that feed the fidelity claims; the abstract's QRC-F fidelity claim contradicts the resulting numbers.
  • quantum/classical learning rates and early stopping = 5e-3 / 1e-3; patience 3/5
    Table II: hand-chosen; with max 10 epochs and early stopping, optimization quality depends on these settings.
axioms (6)
  • domain assumption A shallow 6-qubit parameterized circuit followed by expectation-value features and a linear map can represent useful forecasting structure
    Premise of the whole framework, §IV-C/D.
  • domain assumption Uniform b-bit quantization plus RY angle encoding preserves the information needed for multi-horizon forecasting
    §IV-A/B; no information-theoretic justification is given.
  • ad hoc to paper The joint tensor-product state over M channels and L timesteps (Eq. 24) is the encoding used in the experiments
    This is the method as written but appears inconsistent with the reported resource counts (Table II, XIII), suggesting the experiments used channel-independent encoding instead.
  • domain assumption Depolarizing noise with p ≈ 1e-3, p_2q ≈ 1e-2 approximates NISQ devices; fidelities computed from this model support robustness claims
    Table X; no hardware measurement is reported.
  • standard math Parameter-shift rule exactness and barren-plateau scaling Var[∂L/∂θ] ∝ 2^(-n)
    §IV-C Step 4.6 and §VI.8; standard results, cited to ref [2].
  • ad hoc to paper Arccos inverse mapping (Eq. 35) can decode linear readout outputs into forecast angles
    The readout outputs are unconstrained linear values clipped to [-1,1]; treating them as encoding angles of the forecast is a design choice, and the claim of 'faithful' reconstruction is stronger than the pipeline supports.

pith-pipeline@v1.3.0-alltime-deepseek · 18641 in / 26751 out tokens · 238890 ms · 2026-08-01T22:14:48.139694+00:00 · methodology

0 comments
read the original abstract

This paper presents a unified quantum-classical hybrid framework for multi-horizon time-series forecasting, introducing two model variants Quantum Reservoir Forecaster (QRC-F) and Variational Quantum Forecaster (VQF-F). The proposed framework investigates the complexity-fidelity trade-off of quantum forecasting under near-term NISQ hardware constraints. Continuous time-series signals are transformed into binary representations through uniform quantization and encoded into quantum states using angle encoding with parameterized RY rotation gates. Cross-channel entanglement layers capture dependencies among multiple variables. QRC-F utilizes a fixed random unitary quantum reservoir for stable, gradient-free temporal feature extraction, whereas VQF-F employs a trainable variational quantum circuit optimized through the parameter-shift rule to learn temporal and inter-variable patterns from Pauli expectation values. Both models replace computationally expensive quadratic self-attention with efficient linear transformations, reducing parameter complexity. A shared MIMO-based multi-horizon prediction head simultaneously generates forecasts across multiple horizons, avoiding error accumulation in recursive forecasting. Experimental evaluations on benchmark datasets, including ETTh1, ETTh2, ETTm1, ETTm2, Weather, electricity, and exchange-rate, demonstrate that VQF-F achieves superior training stability and parameter efficiency, while QRC-F provides enhanced robustness and circuit fidelity under quantum noise. The results establish a practical quantum-native forecasting framework with strong potential for deployment on near-term NISQ devices.

Figures

Figures reproduced from arXiv: 2607.16358 by Fredrik Heintz, Sanjay Chakraborty.

Figure 1
Figure 1. Figure 1: Variational Quantum Circuit IV. METHODOLOGY [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of the proposed QTSF framework [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: depicts a quantum circuit diagram of the proposed framework suitable for multivariate time series forecasting. Six qubits and three channels are considered, where the total number of qubits is defined as n = M·L = 6, corresponding to M = 3 channels and L = 2 timesteps per channel. The qubits are grouped channel-wise as (q0, q1) for channel 1, (q2, q3) for channel 2, and (q4, q5) for channel 3, thereby pres… view at source ↗
Figure 4
Figure 4. Figure 4: VQF-F (QTSF) forecasting results on the ETTh2 dataset across four prediction horizons (H ∈ {96, 192, 336, 720}) with look-back length L = 96, n = 6 qubits per channel, and circuit depth D = 2 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: VQF-F (QTSF) forecasting results on the Exchange [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

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

41 extracted references · 11 linked inside Pith

  1. [1]

    Variational quantum classifier for binary classification: Real vs synthetic dataset,

    D. Maheshwari, D. Sierra-Sosa, and B. Garcia-Zapirain, “Variational quantum classifier for binary classification: Real vs synthetic dataset,”IEEE access, vol. 10, pp. 3705–3715, 2021

    2021

  2. [2]

    Investigating and miti- gating barren plateaus in variational quantum circuits: a survey: J. cunningham, j. zhuang,

    J. Cunningham and J. Zhuang, “Investigating and miti- gating barren plateaus in variational quantum circuits: a survey: J. cunningham, j. zhuang,”Quantum Information Processing, vol. 24, no. 2, p. 48, 2025

    2025

  3. [3]

    Reformer: The efficient transformer,

    N. Kitaev, Ł. Kaiser, and A. Levskaya, “Reformer: The efficient transformer,”arXiv preprint arXiv:2001.04451, 2020

    Pith/arXiv arXiv 2001

  4. [4]

    Non-stationary transformers: Exploring the stationarity in time series forecasting,

    Y . Liu, H. Wu, J. Wang, and M. Long, “Non-stationary transformers: Exploring the stationarity in time series forecasting,”Advances in Neural Information Processing Systems, vol. 35, pp. 9881–9893, 2022

    2022

  5. [5]

    Are transform- ers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are transform- ers effective for time series forecasting?” inProceedings of the AAAI conference on artificial intelligence, vol. 37, no. 9, 2023, pp. 11 121–11 128

    2023

  6. [6]

    A transformer-based framework for multi- variate time series representation learning,

    G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eickhoff, “A transformer-based framework for multi- variate time series representation learning,” inProceed- ings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp. 2114–2124

    2021

  7. [7]

    Multivariate time series forecasting by graph attention networks with theoretical guarantees,

    Z. Zhang, W. Li, and H. Liu, “Multivariate time series forecasting by graph attention networks with theoretical guarantees,” inInternational Conference on Artificial Intelligence and Statistics. PMLR, 2024, pp. 2845– 2853

    2024

  8. [8]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” inProceed- ings of the AAAI conference on artificial intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  9. [9]

    Autoformer: Searching transformers for visual recognition,

    M. Chen, H. Peng, J. Fu, and H. Ling, “Autoformer: Searching transformers for visual recognition,” inPro- ceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 270–12 280

    2021

  10. [10]

    itransformer: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “itransformer: Inverted transformers are effective for time series forecasting,”arXiv preprint arXiv:2310.06625, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,

    T. Zhou, Z. Ma, Q. Wen, X. Wang, L. Sun, and R. Jin, “Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting,” inInterna- tional conference on machine learning. PMLR, 2022, pp. 27 268–27 286

    2022

  12. [12]

    A time series is worth 64 words: Long-term forecasting with transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, 14 “A time series is worth 64 words: Long-term forecasting with transformers,”arXiv preprint arXiv:2211.14730, 2022

    Pith/arXiv arXiv 2022

  13. [13]

    Ets- former: Exponential smoothing transformers for time- series forecasting,

    G. Woo, C. Liu, D. Sahoo, A. Kumar, and S. Hoi, “Ets- former: Exponential smoothing transformers for time- series forecasting,”arXiv preprint arXiv:2202.01381, 2022

    Pith/arXiv arXiv 2022

  14. [14]

    Pyraformer: Low- complexity pyramidal attention for long-range time series modeling and forecasting,

    S. Liu, H. Yu, C. Liao, J. Li, W. Lin, A. X. Liu, and S. Dustdar, “Pyraformer: Low- complexity pyramidal attention for long-range time series modeling and forecasting,” in# PLACEHOLDER PARENT METADATA VALUE#, 2022

    2022

  15. [15]

    Ed- former: Embedded decomposition transformer for in- terpretable multivariate time series predictions,

    S. Chakraborty, I. Delibasoglu, and F. Heintz, “Ed- former: Embedded decomposition transformer for in- terpretable multivariate time series predictions,”arXiv preprint arXiv:2412.12227, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting,

    Y . Zhang and J. Yan, “Crossformer: Transformer utilizing cross-dimension dependency for multivariate time series forecasting,” inThe eleventh international conference on learning representations, 2023

    2023

  17. [17]

    Ts-fastformer: Fast transformer for time-series forecasting,

    S. Lee, J. Hong, L. Liu, and W. Choi, “Ts-fastformer: Fast transformer for time-series forecasting,”ACM Trans- actions on Intelligent Systems and Technology, vol. 15, no. 2, pp. 1–20, 2024

    2024

  18. [18]

    Flowformer: Linearizing transformers with conservation flows,

    H. Wu, J. Wu, J. Xu, J. Wang, and M. Long, “Flowformer: Linearizing transformers with conservation flows,” inProceedings of the 39th International Confer- ence on Machine Learning. PMLR, 2022, pp. 24 226– 24 242

    2022

  19. [19]

    Quantum machine learning,

    J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, 2017

    2017

  20. [20]

    Variational quantum algorithms,

    M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincioet al., “Variational quantum algorithms,”Na- ture Reviews Physics, vol. 3, no. 9, pp. 625–644, 2021

    2021

  21. [21]

    Quantum long short-term memory,

    S. Y .-C. Chen, S. Yoo, and Y .-L. L. Fang, “Quantum long short-term memory,” inICASSP 2022 – IEEE In- ternational Conference on Acoustics, Speech and Signal Processing. IEEE, 2022, pp. 8622–8626

    2022

  22. [22]

    Rapid training of quantum recurrent neural networks,

    M. Siemaszko, A. Buraczewski, B. Le Saux, and M. Stobi ´nska, “Rapid training of quantum recurrent neural networks,”Quantum Machine Intelligence, vol. 5, pp. 1–16, 2023

    2023

  23. [23]

    QSegRNN: Quantum segment recur- rent neural network for time series forecasting,

    J.-Y . Kanget al., “QSegRNN: Quantum segment recur- rent neural network for time series forecasting,”EPJ Quantum Technology, 2025

    2025

  24. [24]

    Multivariate time series forecasting with gate-based quantum reservoir computing on NISQ hard- ware,

    W. Hamhoum, S. Cherkaoui, J.-F. Laprade, O. Ahmed, and S. Wang, “Multivariate time series forecasting with gate-based quantum reservoir computing on NISQ hard- ware,”arXiv preprint arXiv:2510.13634, 2025

    Pith/arXiv arXiv 2025

  25. [25]

    Quantum adaptive self- attention for quantum transformer models,

    C.-S. Chen and E.-J. Kuo, “Quantum adaptive self- attention for quantum transformer models,”arXiv preprint arXiv:2504.05336, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    Quantum neural network architectures for multivariate time-series forecasting,

    e. a. Wang, “Quantum neural network architectures for multivariate time-series forecasting,”arXiv preprint arXiv:2510.21168, 2025

    arXiv 2025

  27. [27]

    Integrating quantum- classical attention in patch transformers for enhanced time series forecasting,

    S. Chakraborty and F. Heintz, “Integrating quantum- classical attention in patch transformers for enhanced time series forecasting,”arXiv preprint arXiv:2504.00068, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    A quantum-based attention mechanism in scene text detection,

    H. Wu, J. Zhou, Q. Zhang, Y . Lei, K. Yu, W. An, and J. Zhang, “A quantum-based attention mechanism in scene text detection,” inChinese Conference on Pattern Recognition and Computer Vision (PRCV). Springer, 2023, pp. 3–14

    2023

  29. [29]

    Quantum kernel-based long short-term memory for climate time-series forecasting,

    Y .-C. Hsu, N.-Y . Chen, T.-Y . Li, K.-C. Chenet al., “Quantum kernel-based long short-term memory for climate time-series forecasting,”arXiv preprint arXiv:2412.08851, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    A fusion of supervised contrastive learning and variational quantum classifiers,

    A. K. K. Don, I. Khalil, and M. Atiquzzaman, “A fusion of supervised contrastive learning and variational quantum classifiers,”IEEE Transactions on Consumer Electronics, 2024

    2024

  31. [31]

    Diffusion-enhanced optimization of variational quantum eigensolver for general hamiltonians,

    S. Zhang, Z. Qin, Y . Zhang, Y . Zhou, R. Li, C. Du, and Z. Xiao, “Diffusion-enhanced optimization of variational quantum eigensolver for general hamiltonians,”arXiv preprint arXiv:2501.05666, 2025

    arXiv 2025

  32. [32]

    A multi-classification classifier based on variational quantum computation,

    J. Zhou, D. Li, Y . Tan, X. Yang, Y . Zheng, and X. Liu, “A multi-classification classifier based on variational quantum computation,”Quantum Information Process- ing, vol. 22, no. 11, p. 412, 2023

    2023

  33. [33]

    Sa-timeseries: Self-attention time series models,

    UV A-MLSys, “Sa-timeseries: Self-attention time series models,” 2024, last accessed: 10 March 2025. [Online]. Available: https://github.com/UV A-MLSys/ SA-Timeseries

    2024

  34. [34]

    Quantum- optimized selective state space model for efficient time series prediction,

    S ¸.-A. Jura, M. Udrescu, and A. Top ˆırceanu, “Quantum- optimized selective state space model for efficient time series prediction,” in2025 IEEE International Confer- ence on Big Data (BigData). IEEE, 2025, pp. 995– 1004

    2025

  35. [35]

    Qnn- former: quantum-enhanced neural networks with multi- attention mechanisms for long-term time series forecast- ing: Y . tang et al

    Y . Tang, Z. Cai, Y . Zhang, Z. Gao, and J. Yu, “Qnn- former: quantum-enhanced neural networks with multi- attention mechanisms for long-term time series forecast- ing: Y . tang et al.”Quantum Information Processing, vol. 25, no. 2, p. 66, 2026

    2026

  36. [36]

    Quantum long short- term memory (qlstm) vs. classical lstm in time series forecasting: a comparative study in solar power forecast- ing,

    S. Z. Khan, N. Muzammil, S. Ghafoor, H. Khan, S. M. H. Zaidi, A. J. Aljohani, and I. Aziz, “Quantum long short- term memory (qlstm) vs. classical lstm in time series forecasting: a comparative study in solar power forecast- ing,”Frontiers in Physics, vol. 12, p. 1439180, 2024

    2024

  37. [37]

    Bench- marking quantum and classical sequential models for urban telecommunication forecasting,

    C.-S. Chen, S. Y .-C. Chen, and Y .-C. Tsai, “Bench- marking quantum and classical sequential models for urban telecommunication forecasting,”arXiv preprint arXiv:2508.04488, 2025

    arXiv 2025

  38. [38]

    Exploring hy- brid quantum-classical methods for practical time-series forecasting,

    M. Dimitrijevs, M. K ¯alis, and I. Repko, “Exploring hy- brid quantum-classical methods for practical time-series forecasting,”arXiv preprint arXiv:2412.05615, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Qultsf: Long-term time series forecast- ing with quantum machine learning,

    H. H. S. Chittoor, P. R. Griffin, A. Neufeld, J. Thompson, and M. Gu, “Qultsf: Long-term time series forecast- ing with quantum machine learning,”arXiv preprint arXiv:2412.13769, 2024

    Pith/arXiv arXiv 2024

  40. [40]

    Quantum- enhanced dual-layer graph attention network for time- series forecasting,

    Y . Tang, Z. Cai, Y . Zhang, Z. Gao, and J. Yu, “Quantum- enhanced dual-layer graph attention network for time- series forecasting,”Scientific Reports, vol. 15, no. 1, p. 15 39969, 2025

    2025

  41. [41]

    Quantum temporal fusion trans- former,

    K. Barik and G. Paul, “Quantum temporal fusion trans- former,”arXiv preprint arXiv:2508.04048, 2025

    arXiv 2025