Pith. sign in

REVIEW 4 major objections 5 minor 46 references

A post-Transformer forecaster built from state-space dynamics and tensor-network factorizations can match Transformer accuracy with up to 155x fewer parameters and 2.7x faster inference.

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

2026-08-03 09:47 UTC pith:6RA5ICJR

load-bearing objection Efficiency story is real; the accuracy claim against Transformers isn't supported by the paper's own baselines. the 4 major comments →

arxiv 2601.12375 v3 pith:6RA5ICJR submitted 2026-01-18 cs.NI cs.LG

LiQSS: Post-Transformer Linear Quantum-Inspired State-Space Tensor Networks for Real-Time 6G

Farhad Rezazadeh , Hatim Chergui , Amir Ashtari Gargari , Mehdi Bennis , Houbing Song , Lingjia Liu , Merouane Debbah This is my paper
classification cs.NI cs.LG
keywords 6GO-RANstate-space modelstensor networksHiPPO-LegStime-series forecastingnear-real-time controlRSRP prediction
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 tries to establish that control-grade radio forecasting in 6G O-RAN does not need self-attention. It proposes a model that replaces attention with stable HiPPO-LegS state-space kernels run as depthwise causal convolutions, and compresses the large input embedding and prediction head into tensor-train/matrix-product-state factorizations. On a bespoke dataset of 59,441 windows over 13 KPIs, the model predicts next-step reference signal received power with RMSE 0.2866, using 44,109 parameters and about 0.000456 seconds per window. The headline claim is that this matches or beats Transformer-based and structured-state-space baselines while being 10.8x-155x smaller and 1.4x-2.74x faster. If true, near-real-time RIC controllers can host accurate predictors inside tight memory and latency budgets.

Core claim

LiQSS's central claim is that a strictly causal, linear-time sequence model can deliver forecasting accuracy competitive with attention-based Transformers on O-RAN telemetry. The architecture removes self-attention entirely and instead evolves each latent channel by an independent linear state-space model whose continuous-time generator is the HiPPO-LegS operator, discretized by the bilinear (Tustin) transform and realized as a depthwise causal convolution. A mixture of such kernels with geometrically spaced time scales gives multi-timescale expressivity; squeeze-excitation gating and gated channel mixing add cross-KPI coupling; and the two densest linear maps (KPI-to-latent embedding and la

What carries the argument

Three mechanisms carry the argument. (1) TT/MPS (tensor train / matrix product state) linear operators: the large input projection and prediction head are expressed as chains of small cores whose bond dimension (TT rank) explicitly controls coupling capacity, reducing parameters from O(MN) to a sum of core sizes. (2) HiPPO-LegS state-space kernels: a fixed continuous-time generator is discretized with the Tustin transform and expanded into a causal convolution kernel, so temporal dynamics are linear-time, stable, and strictly causal. (3) Lightweight interaction operators: squeeze-excitation gating reweights channels from pooled statistics, and token-wise gated channel mixing couples channels

Load-bearing premise

The accuracy comparison assumes the compact, rebuilt versions of the Transformer and SSM baselines are fair representatives of those methods; the paper states these are not faithful re-runs of original codebases, so if those baselines were undertuned for short-horizon RSRP, the no-accuracy-loss claim over Transformers could weaken.

What would settle it

Re-run the same 59,441-window RSRP task using each baseline's original published code and hyperparameters under identical leakage-safe chronological splits; if a properly tuned Transformer matches or beats LiQSS's RMSE of 0.2866 while respecting near-real-time budgets, the accuracy claim is weakened. Independently, measure LiQSS's peak memory and latency as L grows on edge RIC hardware; super-linear growth would refute the linear-scaling claim.

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

If this is right

  • Near-real-time O-RAN controllers can host next-step KPI forecasting with a model small enough to co-locate with other xApps: 44k parameters and sub-millisecond inference per window.
  • Longer lookback windows become affordable: training time, inference time, and peak memory all scale roughly linearly with window length L, as the paper demonstrates empirically from L=8 to L=32.
  • The TT rank gives an explicit, interpretable knob for trading accuracy against footprint and latency, with the paper reporting r=16 as an accuracy-leaning alternative and r=4 as the preferred operating point.
  • Strict causality and leakage-safe normalization carry through the whole pipeline, so the reported skill scores (0.92 persistence skill, 0.99 R-squared) describe real causal predictive utility, not peeking.

Where Pith is reading between the lines

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

  • The authors test only next-step single-horizon forecasting; an unstated implication is that multi-step horizons would inherit the same linear complexity, but error-compounding behavior would need separate measurement.
  • The same TT/MPS compression of dense input and output projections could be applied to other linear-time sequence models, not just HiPPO-based ones, potentially shrinking their footprints without changing their temporal backbones.
  • Because inference was measured on a GPU, the small-memory linear scaling suggests the model could fit smaller edge processors, but actual RIC hardware timings are not reported.

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 / 5 minor

Summary. The paper proposes LiQSS, a linear-time state-space tensor-network forecaster for next-step RSRP prediction in O-RAN Near-RT RIC analytics. Self-attention is replaced by stable HiPPO-LegS state-space kernels implemented as depthwise causal convolutions; the input embedding and prediction head are parameterized as TT/MPS operators; lightweight SE gating and gated ChannelMix capture cross-KPI interactions. On a bespoke 13-KPI dataset with 59,441 windows, the authors report 44,109 parameters, RMSE 0.2866, and inference time 0.456 ms per window, claiming up to 155x parameter reduction and 2.74x faster inference than Transformer baselines 'without sacrificing forecasting accuracy.' The paper also reports linear scaling in lookback length and a sensitivity analysis of TT rank, mixture components, and state dimension.

Significance. If the headline accuracy-equivalence claim survives scrutiny, the paper is a useful contribution: it demonstrates a concrete post-Transformer architecture with strictly linear-time complexity, a very small footprint, and deployment-oriented evaluation in an O-RAN setting. The code-matched algorithms and public non-commercial code release are strengths, and the efficiency claims (parameter count, latency, linear scaling) are credible and largely independent of the baseline-fidelity issue. The main uncertainty is the accuracy comparison: the authors explicitly state that the Transformer baselines are 'not faithful re-runs of the original codebases,' and the reported errors for several published Transformers are implausibly large relative to persistence on a one-step autocorrelated KPI task. Re-running official baselines with proper hyperparameter tuning is necessary before the 'without sacrificing forecasting accuracy' claim can be accepted.

major comments (4)
  1. [Section V-B, Table IV] The accuracy half of the central claim rests on baselines the paper itself disclaims. Section V-B states: 'These implementations are not faithful re-runs of the original codebases; they are designed for consistency and comparability rather than exact replication.' In Table IV, PatchTST and iTransformer report RMSE 3.26 and 3.31 on a one-step RSRP task where persistence RMSE is roughly 1.01 (Skill(R)=0.92 for LiQSS). Published Transformers, when properly configured for short-horizon forecasting of a highly autocorrelated KPI, would not typically be three times worse than persistence. As it stands, the claim that LiQSS is competitive with or superior to Transformer baselines 'without sacrificing forecasting accuracy' is not established. Please re-run official or at least properly tuned baseline implementations and report their hyperparameters, or reframe the claim as efficiency-only.
  2. [Table IV, Section V-C] All results are reported for a single fixed seed (42), with no error bars, confidence intervals, or multi-seed statistics. The accuracy differences against the closest SSM baselines are very small: RMSE 0.2866 vs 0.2917 (WM-MS3M) and 0.2918 (MS3M), i.e. roughly 1.7-1.8%. On a dataset of this size, such differences can easily arise from optimizer stochasticity and initialization. Without at least 5-10 seeds with mean±std, or a paired bootstrap interval, the claimed accuracy improvement over prior SSM models is not statistically supported. Please add multi-seed experiments and a significance test (e.g., paired bootstrap or Diebold-Mariano).
  3. [Section VI-D, Table V] The sensitivity analysis appears to select hyperparameters using test-set RMSE: Table V reports test RMSE and R^2 for each configuration, and the text concludes that r=4, C_m=2, N_s=32 'remains the strongest accuracy-efficiency choice.' If test-set metrics informed the default configuration, the reported test RMSE 0.2866 is optimistically biased. Please clarify whether model selection and sensitivity comparisons used only validation data, and if not, report the validation-based selection procedure or use a nested split.
  4. [Section VII, Abstract] The conclusion generalizes from a single bespoke dataset, a single target KPI (RSRP), and H=1 forecasting to a broad 'post-Transformer design paradigm' for 6G O-RAN. This is a scope limitation that should be acknowledged. The architecture may be valuable even if its advantage is demonstrated only for this one-step telemetry task, but the paper should not claim a paradigm-level conclusion without additional datasets, horizons, or at least a clearly stated boundary on the claims.
minor comments (5)
  1. [Figure 3] The caption says 'last 700 test samples' but does not specify whether these are 700 consecutive time points or 700 sliding windows, nor how they relate to the chronological test split. Please clarify.
  2. [Table IV] The inference latency column reports seconds without stating the batch size used in the timing measurement. Latency per window can depend strongly on batching; please report the exact measurement configuration (batch size, warm-up, GPU clock behavior).
  3. [Section I, Abstract] The term 'control-grade prediction' is used without a formal definition. It would help to specify a quantitative threshold (e.g., RMSE below a fraction of the persistence error, or latency below a fraction of Delta_t) so the claim is falsifiable.
  4. [Section IV-F] In Eq. (12), the Tustin discretization of the HiPPO-LegS generator is stated, but it is not explained why the fixed generator leads to numerical stability for the chosen N_s=32 and Delta_t values. A brief stability comment or reference would help readers who want to reproduce the kernel construction.
  5. [References [40], [41]] Several cited works are 'accepted' or 'preprint' without complete bibliographic data. Please update these once final publication details are available.

Circularity Check

0 steps flagged

No significant circularity: LiQSS is an empirical benchmarking paper whose architecture, training, and evaluation are specified independently; no prediction reduces by construction to an input.

full rationale

I walked the derivation from Section II through VI. The claimed result — competitive accuracy with 44,109 parameters and 0.456 ms inference latency — is an empirical outcome, not a derived theorem. The TT/MPS parameterization (Section IV-D/E), HiPPO–LegS dynamics (Section IV-F), mixture kernels (Section IV-G), gating/mixing (Section IV-H), and complexity analysis (Section IV-K) are all specified in equations and code-matched algorithms; none of these definitions assumes the reported RMSE or latency. The only in-paper caveat that could undermine the accuracy claim is Section V-B's statement that compact Transformer baselines 'are not faithful re-runs of the original codebases; they are designed for consistency and comparability rather than exact replication.' That is a benchmark-fairness weakness, not a circularity: it concerns whether the comparison is generous to LiQSS, not whether LiQSS's result is its own input. The authors' self-citations to MS3M/WM-MS3M [11],[12] and to their data pipeline [40],[41] are lineage and baseline context, not load-bearing proofs; no uniqueness theorem or external prior result is invoked to rule out alternatives. Sensitivity analysis (Section VI-D) tunes hyperparameters on held-out validation, which is standard practice and does not make the test-set prediction a fit by construction. I therefore find no circular step under the requested definitions.

Axiom & Free-Parameter Ledger

9 free parameters · 6 axioms · 0 invented entities

The central empirical claims rest on hand-selected hyperparameters, several of which are justified by sensitivity analysis performed on the same test set. The main load-bearing assumption is that custom non-faithful baseline implementations fairly represent the comparison methods. No new physical entities are introduced; the 'quantum-inspired' framing is an analogy.

free parameters (9)
  • Lookback window length L = 32
    Hand-chosen window length; affects both accuracy and the linear-scaling latency measurements in Figure 4.
  • Latent width D = 64
    Hand-chosen model width; dominates parameter count and runtime alongside the TT ranks.
  • Number of SSM blocks B_l = 2
    Hand-chosen network depth; part of the default configuration in Table III.
  • SSM state dimension N_s = 32
    Selected as the default based on Table V sensitivity analysis on the same test data; N_s=16 and N_s=64 perform worse or cost more.
  • HiPPO-LegS mixture components C_m = 2
    Selected as the default from Table V; C_m=4 gives slightly lower RMSE (0.2842 vs 0.2866) but is rejected for efficiency, so the choice is somewhat arbitrary.
  • TT rank for input projection r_in = 4
    Selected as the default from Table V; r=16 gives the best RMSE (0.2836) but r=4 is chosen for latency.
  • TT rank for prediction head r_hd = 4
    Selected from the same sensitivity sweep as r_in.
  • Base time-step init dt0 and growth factor gamma = 0.1, 1.5
    Hand-set initialization rule for the learned per-component time steps; no formal justification given.
  • ChannelMix expansion factor alpha = 1
    Hand-chosen to keep D_m = D = 64; the paper does not sweep this knob.
axioms (6)
  • standard math HiPPO-LegS generators provide stable and expressive polynomial memories for continuous-time signal projection.
    Invoked in Section IV-F as the basis of the SSM kernels; treated as established theory via reference [21].
  • standard math Bilinear (Tustin) discretization of the HiPPO-LegS generator yields a stable discrete transition matrix (Eq. 12).
    Used in Algorithm 1 lines 22-25 and Section IV-F; standard numerical method, no paper-specific derivation.
  • standard math TT/MPS factorization of linear maps preserves enough expressive power while reducing parameters (Eq. 5).
    Assumed for TTIn and TTHead; standard tensor-network compression result cited from [32].
  • domain assumption The bespoke O-RAN KPI dataset from the authors' earlier studies is a faithful representation of Near-RT telemetry streams.
    Section V-A states the dataset follows the authors' earlier pipeline; no external or independent data source is used.
  • ad hoc to paper The custom compact baseline implementations fairly represent each baseline method's accuracy despite not being faithful re-runs.
    Section V-B explicitly admits the baselines are 'not faithful re-runs of the original codebases'; this assumption is load-bearing for the accuracy comparison.
  • domain assumption Latency measured with synchronized CUDA events on a single T4 GPU is representative of Near-RT RIC deployment.
    Section V-B and Table IV use GPU inference times as the deployment-relevant latency, abstracting away real xApp scheduling and data movement.

pith-pipeline@v1.3.0-alltime-deepseek · 20852 in / 12943 out tokens · 137912 ms · 2026-08-03T09:47:48.337551+00:00 · methodology

0 comments
read the original abstract

Proactive and agentic control in Sixth-Generation (6G) Open Radio Access Networks (O-RAN) requires control-grade prediction under stringent Near-Real-Time (Near-RT) latency and computational constraints. While Transformer-based models are effective for sequence modeling, their quadratic complexity limits scalability in Near-RT RAN Intelligent Controller (RIC) analytics. This paper investigates a post-Transformer design paradigm for efficient radio telemetry forecasting. We propose a quantum-inspired many-body state-space tensor network that replaces self-attention with stable structured state-space dynamics kernels, enabling linear-time sequence modeling. Tensor-network factorizations in the form of Tensor Train (TT) / Matrix Product State (MPS) representations are employed to reduce parameterization and data movement in both input projections and prediction heads, while lightweight channel gating and mixing layers capture non-stationary cross-Key Performance Indicator (KPI) dependencies. The proposed model is instantiated as an agentic perceive-predict xApp and evaluated on a bespoke O-RAN KPI time-series dataset comprising 59,441 sliding windows across 13 KPIs, using Reference Signal Received Power (RSRP) forecasting as a representative use case. Our proposed Linear Quantum-Inspired State-Space (LiQSS) model is 10.8x-15.8x smaller and approximately 1.4x faster than prior structured state-space baselines. Relative to Transformer-based models, LiQSS achieves up to a 155x reduction in parameter count and up to 2.74x faster inference, without sacrificing forecasting accuracy.

Figures

Figures reproduced from arXiv: 2601.12375 by Amir Ashtari Gargari, Farhad Rezazadeh, Hatim Chergui, Houbing Song, Lingjia Liu, Mehdi Bennis, Merouane Debbah.

Figure 1
Figure 1. Figure 1: System model and intelligence placement in the O-RAN [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: General end-to-end architecture of the proposed [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Empirical evidence of linear-time behavior in LiQSS. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ground truth versus one-step-ahead predictions for key [PITH_FULL_IMAGE:figures/full_fig_p012_3.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

46 extracted references · 7 linked inside Pith

  1. [1]

    Ultra-Low-Latency Edge Inference for Distributed Sensing,

    Z. Wanget al., “Ultra-Low-Latency Edge Inference for Distributed Sensing,”IEEE Transactions on Wireless Communications, vol. 25, pp. 1908–1922, 2025

    1908

  2. [2]

    Open RAN for 6G Networks: Architecture, Use Cases and Open Issues,

    B. Agarwalet al., “Open RAN for 6G Networks: Architecture, Use Cases and Open Issues,”IEEE Communications Surveys & Tutorials, vol. 28, pp. 2881–2924, 2025

    2025

  3. [3]

    Informer: Beyond Efficient Transformer for Long Se- quence Time-Series Forecasting,

    H. Zhouet al., “Informer: Beyond Efficient Transformer for Long Se- quence Time-Series Forecasting,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  4. [4]

    FEDformer: Frequency Enhanced Decomposed Trans- former for Long-term Series Forecasting,

    T. Zhouet al., “FEDformer: Frequency Enhanced Decomposed Trans- former for Long-term Series Forecasting,” inProceedings of the 39th International Conference on Machine Learning, vol. 162. PMLR, 2022

    2022

  5. [5]

    ETSformer: Exponential Smoothing Transformers for Time-series Forecasting,

    G. Wooet al., “ETSformer: Exponential Smoothing Transformers for Time-series Forecasting,”arXiv preprint arXiv:2202.01381, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    A Time Series is Worth 64 Words: Long-term Forecasting with Transformers,

    Y . Nieet al., “A Time Series is Worth 64 Words: Long-term Forecasting with Transformers,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  7. [7]

    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

  8. [8]

    RWKV: Reinventing RNNs for the Transformer Era,

    B. Penget al., “RWKV: Reinventing RNNs for the Transformer Era,” arXiv preprint arXiv:2305.13048, 2023

    Pith/arXiv arXiv 2023

  9. [9]

    Retentive Network: A Successor to Transformer for Large Language Models,

    Y . Sunet al., “Retentive Network: A Successor to Transformer for Large Language Models,”arXiv preprint arXiv:2307.08621, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Chronos: Learning the Language of Time Series,

    A. F. Ansariet al., “Chronos: Learning the Language of Time Series,” Transactions on Machine Learning Research, 2024

    2024

  11. [11]

    Rivaling Transformers: Multi-Scale Struc- tured State-Space Mixtures for Agentic 6G O-RAN,

    F. Rezazadehet al., “Rivaling Transformers: Multi-Scale Struc- tured State-Space Mixtures for Agentic 6G O-RAN,”arXiv preprint arXiv:2510.05255, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Agentic World Modeling for 6G: Near-Real-Time Generative State-Space Reasoning,

    F. Rezazadeh, H. Chergui, M. Debbah, H. Song, D. Niyato, and L. Liu, “Agentic World Modeling for 6G: Near-Real-Time Generative State-Space Reasoning,”arXiv preprint arXiv:2511.02748, 2025

    Pith/arXiv arXiv 2025

  13. [13]

    Simplified State Space Layers for Sequence Modeling,

    J. T. H. Smithet al., “Simplified State Space Layers for Sequence Modeling,” inICLR, 2023

    2023

  14. [14]

    HiPPO: Recurrent Memory with Optimal Polynomial Projections,

    A. Guet al., “HiPPO: Recurrent Memory with Optimal Polynomial Projections,” inNeurIPS, 2020

    2020

  15. [15]

    Matrix Product State Pre-Training for Quantum Machine Learning,

    J. Dborinet al., “Matrix Product State Pre-Training for Quantum Machine Learning,”Quantum Science and Technology, vol. 7, no. 3, 2022

    2022

  16. [16]

    Quantum computing and neuroscience for 6G/7G networks: Survey,

    S. Glisic and B. Lorenzo, “Quantum computing and neuroscience for 6G/7G networks: Survey,”Intelligent Systems with Applications, 2024

    2024

  17. [17]

    Tasaki,Physics and Mathematics of Quantum Many-Body Systems, 1st ed., ser

    H. Tasaki,Physics and Mathematics of Quantum Many-Body Systems, 1st ed., ser. Graduate Texts in Physics. Springer Cham, 2020

    2020

  18. [18]

    Modeling Sequences with Quantum States: A Look Under the Hood,

    T.-D. Bradleyet al., “Modeling Sequences with Quantum States: A Look Under the Hood,”ML: Science and Technology, vol. 1, no. 3, 2020

    2020

  19. [19]

    iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,

    Y . Liuet al., “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting,” inThe Twelfth ICLR, 2024

    2024

  20. [20]

    Rethinking Attention with Performers,

    K. Choromanskiet al., “Rethinking Attention with Performers,” inICLR, 2021

    2021

  21. [21]

    How to Train Your HiPPO: State Space Models with Generalized Orthogonal Basis Projections,

    A. Guet al., “How to Train Your HiPPO: State Space Models with Generalized Orthogonal Basis Projections,” 2022

    2022

  22. [22]

    Matrix Product Operators for Sequence-to-Sequence Learning,

    C. Guoet al., “Matrix Product Operators for Sequence-to-Sequence Learning,”Physical Review E, vol. 98, no. 4, 2018

    2018

  23. [23]

    Time-dependent Density Matrix Renormalization Group Method for Quantum Dynamics in Complex Systems,

    J. Renet al., “Time-dependent Density Matrix Renormalization Group Method for Quantum Dynamics in Complex Systems,”Wiley Interdisci- plinary Reviews: Computational Molecular Science, vol. 12, 2022

    2022

  24. [24]

    Continuous Frames in Tensor Product Hilbert Spaces, Localization Operators and Density Operators,

    P. Balazs and N. Teofanov, “Continuous Frames in Tensor Product Hilbert Spaces, Localization Operators and Density Operators,”Journal of Physics A: Mathematical and Theoretical, vol. 55, no. 14, 2022

    2022

  25. [25]

    Information Theory With Kernel Methods,

    F. Bach, “Information Theory With Kernel Methods,”IEEE Transactions on Information Theory, vol. 69, no. 2, pp. 752–775, Feb. 2023

    2023

  26. [26]

    Quantum MAC: Genuine Entanglement Access Control via Many-Body Dicke States,

    J. Illianoet al., “Quantum MAC: Genuine Entanglement Access Control via Many-Body Dicke States,”IEEE Transactions on Communications, vol. 72, no. 4, pp. 2090–2105, 2024

    2090

  27. [27]

    Highly Parallel Singular Value Decomposition for Low-Latency MIMO Processing,

    S. Chenget al., “Highly Parallel Singular Value Decomposition for Low-Latency MIMO Processing,” in2025 IEEE Workshop on Signal Processing Systems (SiPS), 2025

    2025

  28. [28]

    On Properties of Schmidt Decomposition,

    M. Kumar, “On Properties of Schmidt Decomposition,”arXiv preprint arXiv.2411.05703, 2024

    Pith/arXiv arXiv 2024

  29. [29]

    Entanglement Entropy and Quantum Field Theory,

    P. Calabrese and J. Cardy, “Entanglement Entropy and Quantum Field Theory,”Journal of Statistical Mechanics: Theory and Experiment, 2004

    2004

  30. [30]

    Area-law Entanglement from Quantum Geometry,

    N. Paul, “Area-law Entanglement from Quantum Geometry,”Physical Review B, vol. 109, 2024

    2024

  31. [31]

    Efficient Tensor Robust PCA Under Hybrid Model of Tucker and Tensor Train,

    Y . Qiuet al., “Efficient Tensor Robust PCA Under Hybrid Model of Tucker and Tensor Train,”IEEE Signal Processing Letters, vol. 29, pp. 627–631, 2022

    2022

  32. [32]

    Tensorizing Neural Networks,

    A. Novikovet al., “Tensorizing Neural Networks,” inNIPS’15: Pro- ceedings of the 29th International Conference on Neural Information Processing Systems, 2015, pp. 442–450

    2015

  33. [33]

    (2023) O-RAN Architecture Overview

    O-RAN Software Community. (2023) O-RAN Architecture Overview. [Online]. Available: https://docs.o-ran-sc.org/en/latest/architecture/ architecture.html

    2023

  34. [34]

    ETSI TS 104 040 V4.0.0: Publicly Available Specification (PAS); E2 interface: RAN Function-specific Service Models (O-RAN E2SMs),

    ETSI, “ETSI TS 104 040 V4.0.0: Publicly Available Specification (PAS); E2 interface: RAN Function-specific Service Models (O-RAN E2SMs),” European Telecommunications Standards Institute (ETSI), Tech. Rep. TS 104 040 V4.0.0, 2024. [Online]. Available: https://www.etsi.org/deliver/ etsi ts/104000 104099/104040/04.00.00 60/ts 104040v040000p.pdf

    2024

  35. [35]

    RIC-O: Efficient Placement of a Disaggregated and Distributed RAN Intelligent Controller with Dynamic Clustering of Radio Nodes,

    G. Matheus Almeida, “RIC-O: Efficient Placement of a Disaggregated and Distributed RAN Intelligent Controller with Dynamic Clustering of Radio Nodes,”IEEE Journal on Selected Areas in Communications, vol. 42, no. 2, pp. 446–459, 2023

    2023

  36. [36]

    Deep Learning for Time Series Forecasting: Tutorial and Survey,

    K. Benidiset al., “Deep Learning for Time Series Forecasting: Tutorial and Survey,”ACM Computing Surveys, vol. 55, no. 6, pp. 1–36, 2022

    2022

  37. [37]

    DeepAR: Probabilistic forecasting with autoregressive recurrent networks,

    D. Salinaset al., “DeepAR: Probabilistic forecasting with autoregressive recurrent networks,”International Journal of Forecasting, vol. 36, no. 3, pp. 1181–1191, 2020

    2020

  38. [38]

    J. P. Hespanha,Linear Systems Theory, 2nd ed. Princeton, NJ: Princeton University Press, 2018

    2018

  39. [39]

    Attention Is All You Need,

    A. Vaswaniet al., “Attention Is All You Need,” inAdvances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017, pp. 5998–6008

    2017

  40. [40]

    Learning Low-Dimensional Representation for O-RAN Testing via Transformer- ESN,

    J. Dai, R. Zhao, F. Rezazadeh, L. Zheng, H. Wang, and L. Liu, “Learning Low-Dimensional Representation for O-RAN Testing via Transformer- ESN,” 2025, accepted for publication at IEEE MASS, 2025

    2025

  41. [41]

    An Experimental Reservoir-Augmented Foundation Model: 6G O-RAN Case Study,

    F. Rezazadehet al., “An Experimental Reservoir-Augmented Foundation Model: 6G O-RAN Case Study,” 2025, accepted to publish at GACLM 2025; arXiv:2508.07778

    Pith/arXiv arXiv 2025

  42. [42]

    [Online]

    NVIDIA Corporation,CUDA Runtime API: Event Management, 2024, accessed: 2026-01-06. [Online]. Available: https://docs.nvidia.com/cuda/ cuda-runtime-api/group CUDART EVENT.html

    2024

  43. [43]

    [Online]

    PyTorch Contributors,torch.cuda.Event — PyTorch Documentation, 2024, accessed: 2026-01-06. [Online]. Available: https://pytorch.org/ docs/stable/generated/torch.cuda.Event.html

    2024

  44. [44]

    Temporal Fusion Transformers for Interpretable Multi- horizon Time Series Forecasting,

    B. Limet al., “Temporal Fusion Transformers for Interpretable Multi- horizon Time Series Forecasting,”International Journal of Forecasting, vol. 37, no. 4, pp. 1748–1764, 2021

    2021

  45. [45]

    Another Look at Forecast Accuracy Metrics for Intermittent Demand,

    R. J. Hyndman, “Another Look at Forecast Accuracy Metrics for Intermittent Demand,”Foresight: The International Journal of Applied Forecasting, no. 4, pp. 43–46, 2006

    2006

  46. [46]

    Skill Scores Based on the Mean Square Error and Their Relationships to the Correlation Coefficient,

    A. H. Murphy, “Skill Scores Based on the Mean Square Error and Their Relationships to the Correlation Coefficient,”Monthly Weather Review, vol. 116, no. 12, pp. 2417–2424, 1988

    1988