Pith. sign in

REVIEW 3 major objections 5 minor 15 references

Entanglement for Pattern Learning in Temporal Data with Logarithmic Complexity: Benchmarking on IBM Quantum Hardware

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A fixed shallow quantum circuit with entangled qubits matches ARIMA on real weather data.

desk verdict Logarithmic complexity claim is a definitional artifact and the model is never trained, so the hardware results don't show quantum temporal learning. read the letter →

arxiv 2506.00097 v1 pith:HH2ACI24 submitted 2025-05-30 quant-ph

Mostafizur Rahaman Laskar , Richa Goel This is my paper · ORCID
classification quant-ph
keywords quantumtimeseriesforecastingentanglementparameterizedcircuitslogarithmiccomplexityIBMhardwareWeatherBenchautoregressivemodelsnoisyintermediate-scale
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

This paper argues that entanglement can serve as the memory mechanism in time-series forecasting, replacing learned recurrent or attention weights. Its Quantum Time Series (QTS) model encodes the last n normalized observations as single-qubit rotations, applies fixed forward and cross CNOT entanglement patterns, and reads the next value from the measurement statistics. On the WeatherBench Z500 geopotential-height series, adding forward plus cross entanglement lowers the simulated error from MSE 0.122392 to 0.000920, and the same circuit on IBM's Heron processor reaches 0.00043, about the same as classical ARIMA, using only 10 qubits and no trained parameters. The authors take this as evidence that quantum entanglement is a practical computational resource for temporal modelling with logarithmic data and parameter complexity.

What carries the argument

The load-bearing object is the QTS circuit: n qubits prepared in $\bigotimes_{i=1}^n R_y(\theta_{t-i})|0\rangle$ with $\theta$ encoding normalized data, followed by a fixed entangling layer $U_{\mathrm{ent}}$ of nearest-neighbour CNOTs (forward entanglement) and distance-two CNOTs (cross entanglement), then computational-basis measurement. The prediction is $\hat{x}_{t+1} = \left(\sum_k P(k)\,k\right)\frac{x_{\max}-x_{\min}}{2^n-1}+x_{\min}$. The CNOTs couple the amplitudes so that the measurement distribution $P(k)$ reflects multi-lag correlations among the input angles, which is how the circuit implements temporal memory without persistent quantum memory or recurrence.

What would settle it

Take the Z500 window, randomly permute the order of its 10 values, and run the same QTS circuit; if the forecast MSE is unchanged or depends only on the value multiset, the circuit is not exploiting temporal order. Independently, expand the Ry rotations and CNOTs into a closed-form classical function and evaluate it without fitting parameters; if that classical evaluation reproduces the reported hardware MSE, the result does not require a quantum device.

Watch

Extended reading notes

Core claim

The central claim is that a parameter-free, shallow entangling circuit can act as a functioning forecaster: encode each lagged value as an Ry rotation, entangle the qubits with CNOT gates, measure in the computational basis, and use the rescaled expectation value of the bitstring index as the prediction. The paper shows that on the real Z500 series the fully entangled variant reaches MSE 0.00043 on ibm_kingston over 15 forecast steps, compared with 0.000416 for ARIMA(2,1,0) and 0.000421 for AR(2), while TCN and Transformer exceed 0.045. It further reports that on three IBM processors forecasting 10 steps the same circuit yields MSE 0.00052–0.00104, with the best result on ibm_brisbane. The paper's conclusion is that the entangling structure, not optimization, carries the temporal pattern information, and that this structure needs only O(log N) qubits, parameters, and data points relative to series length N.

Load-bearing premise

The logarithmic-complexity claim rests on treating a single fixed 10-value window as the training data and regarding the fixed, untrained circuit as a fitted forecasting model; if that equivalence fails, the MSE comparison is between a trained classical forecaster and a fixed map.

Editorial extensions

If this is right

  • Adding forward and cross entanglement to a rotation-only QTS circuit reduces forecasting MSE from 0.122392 to 0.000920 in noisy simulation, so the entanglement pattern itself is doing the predictive work.
  • On real IBM hardware the fully entangled QTS circuit matches ARIMA and AR(2) on the Z500 series, with MSE 0.00043 over 15 steps, despite using 10 encoded values against 1024 training points for the classical models.
  • If the complexity analysis holds, forecasting models can be built with O(log N) data, parameters, and gates, which would make quantum forecasting attractive in data-limited settings.
  • The similar accuracy across three IBM devices with different gate errors and connectivities, without specialized error mitigation, suggests the approach tolerates realistic hardware noise.

Reading between the lines

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

  • The paper does not say this, but the circuit has no learning loop, so the same fixed map can be evaluated classically as a polynomial of the input window; the claimed O(log N) training cost depends on counting the n-value window as the whole training set rather than the 1024 points the classical baselines fitted.
  • A natural testable extension is to make the rotation angles variational and train them; only then would the experiment separate the expressivity of entanglement from the expressivity of a fixed nonlinear feature map.
  • Applying the same circuit to AR(1) data with varying coefficient φ would show whether the forecast tracks the dynamics or mostly reproduces the window mean; that would clarify how much temporal structure the entanglement actually captures.
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 manuscript introduces a Quantum Time Series (QTS) model for univariate forecasting and claims logarithmic complexity in training data and parameter count. The model encodes n normalized past observations into n single-qubit Ry rotations, applies fixed forward and cross CNOT entanglement layers, and defines the forecast x̂_{t+1} as the scaled expectation of the computational-basis label (Eq. 4). The authors benchmark three circuit variants against AR, ARIMA, TCN, and Transformer baselines on synthetic AR(1) data and WeatherBench Z500 geopotential-height data, reporting MSE values down to 0.00043 on ibm_kingston and 0.00052 on ibm_brisbane. The paper concludes that entanglement is a practical computational resource for temporal modelling. The central claims of the paper are, however, based on an untrained fixed circuit and a definitional complexity argument, as detailed below.

Significance. If the claims were correct, the paper would report a surprising result: a shallow, parameter-free-entanglement circuit consuming only ten data points forecasts as well as ARIMA fitted to 1024 points. The paper deserves credit for deploying the circuit on three different IBM processors, reporting transpiled gate counts, and explicitly comparing rotation-only, forward-entangled, and forward-plus-cross-entangled versions. These empirical components are useful as a characterization of a fixed quantum feature map. The significance as a learning result, however, is not established: the logarithmic scaling is a relabeling artifact, the QTS parameters are never optimized, and the baseline comparison mixes model class with training budget. The paper therefore does not support the abstract's claim that hardware benchmarking establishes entanglement as a practical resource for temporal pattern learning.

major comments (3)
  1. [Computational complexity; Supplementary Lemma 2] The claimed O(log N) data complexity is definitional rather than substantive. Eq. (1) encodes n values into n qubits, and Lemma 2 defines N=2^n, so the n-point sliding window is relabelled as log2 N; under this convention every fixed-window forecaster with window size n would be 'logarithmic'. The note to Lemma 2 confirms this by counting n+F values for F forecast steps. The actual per-prediction data use is O(n), and the total training-data use cannot be 'log(N)' without assuming that the whole series has length 2^n. The Supplementary section 'Complexity of Entanglement-based Quantum Time Series Model' repeats the claim by referring to amplitude encoding of N steps into log N qubits, but the circuit in Eq. (1) uses rotation encoding with one qubit per input value. This is load-bearing because the logarithmic complexity is the paper's headline claim.
  2. [Quantum time series model, Eqs. (1)-(4)] The QTS circuit is an untrained fixed map, not a learned model. The rotation angles are set to the last n observations, the CNOT structure is fixed, and the prediction is a deterministic expectation of measurement outcomes; no loss function, gradient, or parameter update appears anywhere in the paper. The manuscript itself states that the model 'currently lacks closed-form learning rules' (Note on Lemma 1) and that variational training is future work (Discussion). Calling the model 'trained on 10 immediate past data points' in the Fig. 2(b) caption conflates setting input angles with learning. Consequently, the hardware MSE values cannot be used as evidence that entanglement learns temporal patterns; an untrained baseline such as persistence, or a fixed classical map on the same 10 inputs, is required to interpret these numbers.
  3. [Table 1 and Fig. 2] The empirical comparison is inequitable and statistically unsupported. The classical models are fitted to 1024 training points, while QTS uses 10 values with no optimization; the reported MSE values are single numbers without error bars, shot counts, or repeated runs, so the differences between QTS variants and between hardware backends may be dominated by shot noise. To support the claim of 'learning temporal patterns using fewer data points', the authors would need to compare against (i) an untrained classical baseline, (ii) a classical model trained on the same 10-point window, or (iii) a version of QTS with genuinely optimized parameters under a matched data budget. As written, the comparison conflates model capacity with training protocol and cannot support the conclusion in the abstract.
minor comments (5)
  1. [Abstract] The phrase 'Y et' should read 'Yet'.
  2. [Experimental settings] The expressions '210 + 32 = 1056' and '210 + 16 = 1040' should be written as '2^10 + 32 = 1056' and '2^10 + 16 = 1040'; the superscript formatting appears to be lost.
  3. [Table 1 and Fig. 3] The ibm_kingston MSE is reported as 0.00043191 in Table 1 and 0.00072 in the Fig. 3 caption; if these correspond to different forecast horizons (15 vs 10 steps), state this explicitly.
  4. [Experimental settings] The number of shots per circuit and whether hardware MSEs are averaged over repeated executions are not reported; this information is needed for reproducibility.
  5. [Fig. 2(b) caption] The phrase 'trained on 10 immediate past data points' should be replaced by 'encoded 10 immediate past data points', since no training occurs in the reported experiments.

Circularity Check

2 steps flagged · score 7.0 of 10

The O(log N) training-data claim is definitional: Lemma 2 sets N=2^n and counts the n-value sliding window as training, so the headline complexity reduces to the encoding convention; no training occurs.

  1. self definitional [Supplementary Information, 'Complexity Analysis', Lemma 2 and its Note; see also 'Experimental settings'.]
    "Lemma 2. Let n be the number of qubits and N = 2n the size of the classical dataset. Then, the QTS model requires only O(logN) data points for training. ... Each quantum circuit iteration encodes n classical values into n qubits using a sliding window. Since this window has fixed size n and slides one step per forecast point, the training process for F forecast steps only requires n + F values. As n = log2 N, the number of training points scales logarithmically with the dataset size N."

    The lemma postulates N=2^n and then calls the n-point window 'training'. The conclusion O(log N) is therefore a restatement of the definition of N, not a derived property: any fixed-window forecaster that consumes n values would be O(log N) under the same convention. The actual data consumed is n+F values; 'training data requirement' is just the window size. The QTS output is the fixed map of Eqs. (1)-(4), with angles set from data and parameter-free CNOTs, so no learning loop exists; the headline logarithmic-data claim reduces by construction to the choice to encode n values into n qubits.

  2. other [Figure 2(b) caption; 'Entanglement for temporal pattern learning' section.]
    "the QTS model—comprising R(θ ) rotations with forward and cross-entanglement—was trained on 10 immediate past data points and executed 15 times on IBM’s Heron R2 quantum processor (ibm_kingston). ... While our model currently lacks closed-form learning rules, its physical transparency and architectural modularity make it a promising candidate for a near-term quantum machine learning model."

    The word 'trained' is applied to what the equations define as a closed-form encoding: each θ_{t-i} is set by the input value, the CNOT pattern is fixed, and the prediction is an expectation over measurement outcomes. There is no loss, gradient, or parameter update, and the paper itself concedes there are no closed-form learning rules. Consequently, 'trained on 10 immediate past data points' and 'capture temporal patterns using fewer data points' rename the fixed-window encoding as learning; the few-data claim is built into the construction rather than demonstrated.

full rationale

The central complexity claim is circular: Lemma 2 defines N=2^n, counts the n-value window as the training set, and then reports O(log N) data and parameter complexity. That conclusion is equivalent to the construction and would hold for any fixed n-point forecasting rule. The more serious companion issue is that no training occurs at all: Eqs. (1)-(4) give a deterministic, hand-coded map, yet the paper calls this 'trained on 10 immediate past data points' and claims learning with fewer data. The hardware MSEs are not fitted parameters and are therefore not circular in themselves, but they do not support the logarithmic-learning headline because the classical competitors were trained on 1024 points while QTS uses 10 untrained inputs. I found no load-bearing self-citation or imported uniqueness result; these are normal external references. The score reflects that the paper's most prominent theoretical claim reduces by definition, while the empirical benchmark comparison is a fairness/correctness concern rather than a formal circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central complexity result rests on the definitional identification n=log2 N; the forecasting claim rests on the assumption that a fixed expectation-value decoder generalizes, which is stated but not derived.

free parameters (1)
  • n (number of qubits / input window size) = 10
    Chosen by hand to satisfy n = log2(1024); it sets the receptive field and directly produces the O(logN) complexity claim.
assumptions (4)
  • ad hoc to paper The bitstring-label expectation value defined in Eq. 4 is a meaningful forecast of the next time step.
    No optimality or consistency argument connects this observable to the data-generating process; it is an architectural choice.
  • domain assumption CNOT entanglement between rotation-encoded qubits emulates autoregressive memory without training.
    Stated in Lemma 1 and its note; the paper provides no derivation that this improves forecast accuracy beyond being a nonlinear feature map.
  • ad hoc to paper Using n input values per prediction is equivalent to O(logN) training-data complexity when N=2^n.
    This identification in Lemma 2 conflates the input window size with the training set and is the source of the claimed logarithmic advantage.
  • domain assumption A circuit with no learnable parameters can legitimately be described as learning temporal patterns.
    The paper calls QTS a model and claims prediction from 'fewer data points', but no training, gradient updates, or hyperparameter optimization are performed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Entanglement for Pattern Learning in Temporal Data with Logarithmic Complexity: Benchmarking on IBM Quantum Hardware." pith.science (2026). https://pith.science/paper/HH2ACI24

@misc{pith2026250600097,
  author       = {Pith},
  title        = {Pith review of: Entanglement for Pattern Learning in Temporal Data with Logarithmic Complexity: Benchmarking on IBM Quantum Hardware},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HH2ACI24}},
  note         = {Machine review of arXiv:2506.00097}
}
read the original abstract

Time series forecasting is foundational in scientific and technological domains, from climate modelling to molecular dynamics. Classical approaches have significantly advanced sequential prediction, including autoregressive models and deep learning architectures such as temporal convolutional networks (TCNs) and Transformers. Yet, they remain resource-intensive and often scale poorly in data-limited or hardware-constrained settings. We propose a quantum-native time series forecasting framework that harnesses entanglement-based parameterized quantum circuits to learn temporal dependencies. Our Quantum Time Series (QTS) model encodes normalized sequential data into single-qubit rotations and embeds temporal structure through structured entanglement patterns. This design considers predictive performance with logarithmic complexity in training data and parameter count. We benchmark QTS against classical models on synthetic and real-world datasets, including geopotential height fields used in numerical weather prediction. Experiments on the noisy backend and real IBM quantum hardware demonstrate that QTS can capture temporal patterns using fewer data points. Hardware benchmarking results establish quantum entanglement as a practical computational resource for temporal modelling, with potential near-term applications in nano-scale systems, quantum sensor networks, and other forecasting scenarios.

Figures

Figures reproduced from arXiv: 2506.00097 by the authors.

Figure 1
Figure 1. Comparison of quantum circuit design before and after transpilation for hardware compatibility. The transpilation adapts the circuit to physical qubit layout and native gate set, ensuring executable fidelity on superconducting quantum processors [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Benchmarking of quantum and classical time series models on real weather data from geopotential_500hPa_1979_5.625deg.nc, demonstrating the enhanced performance of QTS circuits when entanglement is incorporated. To comprehensively assess the performance of our quantum time series (QTS) model, we benchmarked its predictive capabilities both on simulated noisy quantum backends and on real quantum hardware for the datas… view at source ↗
Figure 3
Figure 3. Hardware Benchmarking: Forecasting performance of the Quantum Time Series (QTS) model with forward and cross-entanglement executed on three IBM quantum hardware platforms: ibm_kingston, ibm_torino, and ibm_brisbane. The circuits were evaluated on real Z500 meteorological data from the WeatherBench archive over a 10-step forecasting horizon. While all hardware platforms showed consistent predictive patterns, ibm_bris… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 9 canonical work pages

  1. [1]

    & Brunet, G

    Bauer, P., Thorpe, A. & Brunet, G. The quiet revolution of numerical weather prediction. Nature 525, 47–55, DOI: 10.1038/nature14956 (2015)

    doi:10.1038/nature14956 2015

  2. [2]

    Schultz, M. G. et al. Can deep learning beat numerical weather prediction? Philos. Transactions Royal Soc. A 379, 20200097, DOI: 10.1098/rsta.2020.0097 (2021)

    arXiv 2021

  3. [3]

    & Hamad-Schifferli, K

    Gomez-Marquez, J. & Hamad-Schifferli, K. Local development of nanotechnology-based diagnostics. Nat. Nanotechnol. 16, 484–486 (2021)

    work page 2021

  4. [4]

    J., Miessen, A

    Ollitrault, P. J., Miessen, A. & Tavernelli, I. Molecular quantum dynamics: A quantum computing perspective. Accounts Chem. Res. 54, 4229–4238 (2021)

    work page 2021

  5. [5]

    & Graves, A

    Graves, A. & Graves, A. Long short-term memory. Supervised sequence labelling with recurrent neural networks 37–45 (2012)

    work page 2012

  6. [6]

    D., Vidal, R., Reiter, A

    Lea, C., Flynn, M. D., Vidal, R., Reiter, A. & Hager, G. D. Temporal convolutional networks for action segmentation and detection. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 156–165 (2017)

    work page 2017

  7. [7]

    Vaswani, A. et al. Attention is all you need. Adv. neural information processing systems 30 (2017)

    2017

  8. [8]

    & Zohren, S

    Lim, B. & Zohren, S. Time-series forecasting with deep learning: a survey. Philos. Transactions Royal Soc. A 379, 20200209 (2021)

    work page 2021

Show all 15 references
  1. [9]

    N., Carpov, D., Chapados, N

    Oreshkin, B. N., Carpov, D., Chapados, N. & Bengio, Y . N-beats: Neural basis expansion analysis for interpretable time series forecasting. arXiv preprint arXiv:1905.10437 (2019)

    2019 arXiv

  2. [10]

    Verdon, G. et al. Learning to learn with quantum neural networks via classical neural networks. arXiv preprint arXiv:1907.05415 (2019)

    2019 arXiv

  3. [11]

    & Killoran, N

    Schuld, M. & Killoran, N. Quantum machine learning in feature hilbert spaces. Phys. review letters 122, 040504 (2019)

    2019

  4. [12]

    Rasp, S. et al. Weatherbench: a benchmark data set for data-driven weather forecasting. J. Adv. Model. Earth Syst. 12, e2020MS002203 (2020)

    2020

  5. [13]

    Bai, S., Kolter, J. Z. & Koltun, V . An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arxiv. arXiv preprint arXiv:1803.01271 10 (2018)

    2018 arXiv

  6. [14]

    E., Jenkins, G

    Box, G. E., Jenkins, G. M., Reinsel, G. C. & Ljung, G. M. Time series analysis: forecasting and control (John Wiley & Sons, 2015)

    2015

  7. [15]

    Hyndman, R. J. & Athanasopoulos, G. Forecasting: principles and practice (OTexts, 2018). 8/11 Supplementary Information Open source dataset To assess the real-world applicability of our quantum time series (QTS) model, we utilize a geophysical dataset comprising geopotential h...

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.