REVIEW 3 major objections 6 minor 15 references
On the Accuracy and Precision of Moving Averages to Estimate Wi-Fi Link Quality
T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash
Pith's one-line read This paper derives closed-form mean squared error formulas for simple and exponential moving averages that estimate Wi-Fi frame delivery ratio, exact in steady channels and good when the filter cutoff is a decade below the disturbance.
desk verdict A correct, modest analytical result: closed-form MSE for SMA/EMA estimation of Wi-Fi FDR, well validated on synthetic data but with an unaddressed independence assumption and thin experimental support. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing identities are the variance decompositions of the estimation errors. For the SMA, the error $d=z-u$ splits into a future half-window and a past half-window, yielding $\mathrm{Var}(d)=\mathrm{Var}(x)/(2m)$. For the EMA, the error $e=z-y$ splits into $v/2$ and $y-u/2$, and because the two sums run over disjoint outcome indices they are uncorrelated, which gives $\mathrm{Var}(e)=\mathrm{Var}(x)\left[\frac{\alpha}{2-\alpha}+\frac{1}{m}\left((1-\alpha)^m-\frac12\right)\right]$ in terms of the smoothing parameter $\alpha$ and window half-width $m$. The paper uses these identities to predict the sample MSE from the per-attempt variance $\epsilon(1-\epsilon)$, and to state the cutoff-frequency rule for non-stationary validity.
What would settle it
Generate a stationary binary process with fixed mean failure probability but controlled positive autocorrelation between consecutive outcomes, run the SMA and EMA filters, and compare the measured MSE against Eq. (24) and Eq. (30); systematic overshoot of the formulas would show the uncorrelated-outcome assumption is the weak point.
Extended reading notes
Core claim
The paper's central claim is that the mean squared error of SMA and EMA estimates of the centered frame delivery ratio equals the probabilistic variance of the estimation error under stationary conditions, with explicit closed forms given by Eq. (24) for the SMA and Eq. (30) for the EMA. The target quantity is the frame delivery ratio over a window of $2m$ transmission attempts centered on the current time, which is non-causal and so only usable offline; the SMA estimates it from the most recent $m$ attempts, while the EMA weights recent attempts exponentially. Because both filters are unbiased in stationary conditions, the MSE reduces to the variance, which is expressed in terms of the per-attempt variance $\mathrm{Var}(x)=\epsilon(1-\epsilon)$ and the filter parameters $m$ and $\alpha$. Numerical runs on synthetic stationary processes confirm the formulas, and tests on synthetic and experimental non-stationary logs show that the approximation stays acceptable as long as the filter cutoff frequency is roughly one decade below the disturbance frequency.
Load-bearing premise
The variance formulas assume that successive transmission outcomes are uncorrelated, so if real failures cluster in bursts, the true estimation error will be larger than the formula predicts even in stationary conditions.
Editorial extensions
If this is right
- In stationary channels, SMA error variance is inversely proportional to window size $m$, and EMA error variance grows roughly linearly with $\alpha$, giving a direct rule for sizing the filters.
- For slowly varying channels, keeping the filter cutoff about one decade below the disturbance frequency keeps the closed-form formulas within acceptable accuracy, so the formulas double as a tuning rule.
- Because MSE equals variance for the unbiased estimators, the sample variance of observed estimation errors is a legitimate proxy for MSE in stationary conditions, and the formulas give the value before any data is collected.
- Any machine-learning link-quality estimator can be held to this baseline: to justify its added complexity, its MSE should beat the SMA and EMA values at the same responsiveness.
Reading between the lines
- If real transmission outcomes are positively correlated, for instance under bursty interference, the independence assumption will make the closed-form variance understate the true error; an extension with an autocorrelation correction would cover such channels.
- The same variance-splitting technique applies to other linear filters, such as multi-pole IIR designs, so closed-form MSE baselines for the more sophisticated estimators mentioned in the conclusions seem within reach.
- Because the centered frame delivery ratio is the target used to train neural-network predictors, these formulas also provide a lower bound on what such predictors can achieve in stationary conditions.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper analyzes the accuracy and precision of simple moving average (SMA) and exponential moving average (EMA) estimators of Wi-Fi link quality, quantified as the frame delivery ratio (FDR) over an interval centered on the current time. Under stationary conditions with uncorrelated transmission outcomes, it derives closed-form expressions for the mean squared error (MSE) of the SMA and EMA estimation errors: Eq. (24) for the SMA and Eq. (30) for the EMA. The derivations are validated against synthetic i.i.d. Bernoulli processes with more than ten million samples, where the sample MSE closely matches the formulas. The paper then considers non-stationary conditions, both synthetic (sinusoidal failure probability) and experimental (one week of logs from four Wi-Fi channels), and claims that the stationary formulas provide an acceptable approximation to the MSE provided the filter cutoff is about a decade below the disturbance frequency, with larger deviations for slow filters. The paper positions the results as a parameter-free baseline against which machine-learning-based link quality estimators can be compared.
Significance. If the central claims hold, the paper provides simple, closed-form expressions for the MSE of SMA and EMA link-quality estimators that require no fitted parameters and can serve as a baseline for runtime monitoring and for evaluating more complex ML-based approaches. The algebraic derivation is self-contained and the synthetic validation is convincing: the formulas match simulation results across a wide range of filter settings and failure probabilities. The main significance is in the utility of the formulas as a design tool, rather than in methodological novelty. However, the result is conditional on the assumption that transmission outcomes are uncorrelated, which is asserted but not verified on the experimental traces, and the non-stationary/experimental validation relies on subjective color-coded criteria without confidence intervals. These limitations currently bound the paper's generality.
major comments (3)
- [Section IV-B, Table III] The closed-form variance formulas are derived under the explicit assumption that the random variables xi are uncorrelated, but this assumption is never validated for the experimental data. The text asserts that Ts = 0.5 s is 'spaced adequately', yet no autocorrelation diagnostics are reported. Real Wi-Fi transmission failures are frequently bursty (e.g., due to interference episodes or slow fading), and for a correlated binary process the variance of an m-sample mean contains additional terms of the form (p(1-p)/m)[1+2Σ(1-k/m)ρ_k]; similar corrections apply to the block-difference terms in Eqs. (24) and (30). Without quantifying the autocorrelation of the experimental logs, the claim that the formulas describe the MSE of real Wi-Fi FDR estimation is not established. I recommend adding (a) sample autocorrelation plots or lag-1/lag-k correlation estimates for the four experimental logs, and (b) a sensitivity analysis using a correlated binary process with realistic ρ_k to show how much the true MSE can deviate from Eqs. (24)/(30).
- [Section IV-A, Table II] The experimental evaluation uses a single week-long log per channel and reports no confidence intervals or quantitative criterion for when the approximation is 'acceptable'. Table III shows large discrepancies for slow filters; for example, on channel 13 with m = 10000, Var(e) = 0.000012 while the MSE is 0.000320, and on channel 1 with m = 1000, Var(d) = 0.000033 while the MSE is 0.000127. Even for faster filters where the match is better, the sampling variability of the MSE estimate from a single trace is not quantified. The paper should define an explicit acceptable-error threshold (e.g., relative error between MSE and Var), and report bootstrap confidence intervals for the experimental MSE values so that the reader can assess whether the observed agreement is statistically meaningful.
- [Section II-C, Eq. (13)] The conclusion that the stationary formulas are 'acceptable' only when the filter cutoff is about a decade below the disturbance frequency is based on a small, hand-picked set of parameter combinations (f ∈ {0.0001, 0.001} Hz, m ∈ {10, 100, 1000, 10000}) and on a color code that is never defined in the table or text. The text states that in some rows 'the MSE becomes poorer but still acceptable' without giving a quantitative rule. To support the 'one decade' claim, the paper should sweep the ratio of filter cutoff to disturbance frequency more finely and plot the relative discrepancy between the simulated MSE and the stationary variance as a function of that ratio, or alternatively derive an analytic bound for the non-stationary case. As written, the claim is an observation about a few rows rather than a demonstrated property.
minor comments (6)
- [Section II-C, Eq. (13)] The color code (green, yellow, red) mentioned in the text of Section IV-A is not reproduced or defined in the table itself; please add a legend or explicit thresholds for what counts as acceptable.
- [Section II] The term 'uncorrected variance s²_ε[i]' is used but never defined; please define it as the sample variance of the ε_j values in the interval without Bessel's correction.
- [Section IV-B] The notation xi is used both for the observed binary sequence and for the random variable model. Using uppercase X_i for the random variable would avoid ambiguity, especially in the probabilistic derivations of Section II-C.
- [References] The experimental setup does not state the actual probing rate Ts used for the four channel logs, even though the adequacy of Ts = 0.5 s is a key assumption in Section II-C; please report the probing period for the experimental traces.
- [Eq. (13)] Several references used for motivation and prior art are self-citations (refs. 10, 11, 12, 15); please consider citing independent evaluations of moving-average or regression-based link-quality estimation to strengthen the literature context.
- [Section IV-A] In Eq. (13), the expansion from ϵ_j(1-ϵ_j) to the terms involving ¯ϵ[i], ¯ϵ²[i], and s²_ϵ[i] has a skipped step; adding one intermediate line would improve readability.
Circularity Check
No significant circularity: the MSE formulas are derived under explicit assumptions and validated against external synthetic and experimental data.
full rationale
The paper's central claim is a conditional mathematical derivation: under the stated stationary, uncorrelated-outcome model, Eqs. (24) and (30) give the variances of the SMA and EMA FDR estimation errors, and Tables I-III check these closed forms against synthetic and real traces. No parameter is fitted to make the formulas agree; the inputs are only m, alpha, and the failure probability epsilon, and the comparison is external to the derivation. The independence assumption in Section II-C ('Probabilistic variance can be easily computed under the assumption that random variables xi are uncorrelated') is asserted with a spacing justification (Ts = 0.5 s) but is not quantitatively verified on the experimental logs; however, that is a limitation of the model's applicability to correlated or bursty channels, not a circularity, because the derivation does not presuppose the result it claims. The self-citations (refs. 10, 11, 12, 15) are used for motivation and prior art ('As pointed out in [10]', 'As show in [15]') and are not load-bearing for Eqs. (24) and (30). There is no step in which a fitted parameter is renamed a prediction, no uniqueness theorem imported from the authors' prior work, and no definition that encodes the target result. The derivation chain is self-contained: Eq. (24) follows from splitting the centered FDR into disjoint past and future sums, and Eq. (30) follows from the analogous splitting of the EMA error; both rely on the explicitly stated uncorrelated-samples assumption rather than on any calibrated quantity. Therefore the paper receives a circularity score of 0.
Assumptions & free parameters
free parameters (2)
- SMA window size m =
10, 100, 1000, 10000
- EMA smoothing factor alpha =
alpha = 2/m (0.2, 0.02, 0.002, 0.0002)
assumptions (4)
- domain assumption Transmission outcomes are uncorrelated
- domain assumption The random process is stationary in Section III
- standard math Standard probability and variance algebra
- standard math EMA steady-state limit is used after initial transient
Cite this review
Pith. "Pith review of On the Accuracy and Precision of Moving Averages to Estimate Wi-Fi Link Quality." pith.science (2026). https://pith.science/paper/UHIIGZU7
@misc{pith2026241112265,
author = {Pith},
title = {Pith review of: On the Accuracy and Precision of Moving Averages to Estimate Wi-Fi Link Quality},
year = {2026},
howpublished = {\url{https://pith.science/paper/UHIIGZU7}},
note = {Machine review of arXiv:2411.12265}
}
read the original abstract
The radio spectrum is characterized by a noticeable variability, which impairs performance and determinism of every wireless communication technology. To counteract this aspect, mechanisms like Minstrel are customarily employed in real Wi-Fi devices, and the adoption of machine learning for optimization is envisaged in next-generation Wi-Fi 8. All these approaches require communication quality to be monitored at runtime. In this paper, the effectiveness of simple techniques based on moving averages to estimate wireless link quality is analyzed, to assess their advantages and weaknesses. Results can be used, e.g., as a baseline when studying how artificial intelligence can be employed to mitigate unpredictability of wireless networks by providing reliable estimates about current spectrum conditions.
Figures
Reference graph
Works this paper leans on
-
[1]
What Will Wi-Fi 8 Be? A Primer on IEEE 802.11bn Ultra High Reliability
L. G. Giordano, G. Geraci, M. Carrascosa, and B. Bellalta, “What Will Wi-Fi 8 Be? A Primer on IEEE 802.11bn Ultra High Reliability,” arXiv, vol. 2303.10442, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2303.10442
work page Pith review arXiv doi:10.48550/arxiv.2303.10442 2023
-
[2]
H. M. F. Noman, E. Hanafi, K. A. Noordin, K. Dimyati, M. N. Hindia, A. Abdrabou, and F. Qamar, “Machine Learning Empowered Emerging Wireless Networks in 6G: Recent Advancements, Challenges and Future Trends,” IEEE Access , vol. 11, pp. 83 017–83 051, 2023
work page 2023
-
[3]
Machine Learning for Large-Scale Optimization in 6G Wireless Networks,
Y . Shi, L. Lian, Y . Shi, Z. Wang, Y . Zhou, L. Fu, L. Bai, J. Zhang, and W. Zhang, “Machine Learning for Large-Scale Optimization in 6G Wireless Networks,” IEEE Communications Surveys & Tutorials , vol. 25, no. 4, pp. 2088–2132, 2023
2023
-
[4]
W. Tashan, I. Shayea, S. Aldirmaz-C ¸ olak, O. A. Aziz, A. Alhammadi, and Y . I. Daradkeh, “Advanced Mobility Robustness Optimization Mod- els in Future Mobile Networks Based on Machine Learning Solutions,” IEEE Access , vol. 10, pp. 111 134–111 152, 2022
work page 2022
-
[5]
Knowledge-Driven Machine Learning and Applications in Wireless Communications,
D. Li, Y . Xu, M. Zhao, J. Zhu, and S. Zhang, “Knowledge-Driven Machine Learning and Applications in Wireless Communications,”IEEE Transactions on Cognitive Communications and Networking , vol. 8, no. 2, pp. 454–467, 2022
work page 2022
-
[6]
Dynamic Multichannel Access Based on Deep Reinforcement Learning in Dis- tributed Wireless Networks,
Q. Cui, Z. Zhang, Y . Shi, W. Ni, M. Zeng, and M. Zhou, “Dynamic Multichannel Access Based on Deep Reinforcement Learning in Dis- tributed Wireless Networks,” IEEE Systems Journal , vol. 16, no. 4, pp. 5831–5834, 2022
work page 2022
-
[7]
Wi-Fi Meets ML: A Survey on Improving IEEE 802.11 Performance With Machine Learning,
S. Szott, K. Kosek-Szott, P. Gawłowicz, J. T. G ´omez, B. Bellalta, A. Zubow, and F. Dressler, “Wi-Fi Meets ML: A Survey on Improving IEEE 802.11 Performance With Machine Learning,” IEEE Communica- tions Surveys and Tutorials , vol. 24, no. 3, pp. 1843–1893, 2022
work page 2022
-
[8]
DeepWiFi: Cognitive WiFi with Deep Learning,
K. Davaslioglu, S. Soltani, T. Erpek, and Y . E. Sagduyu, “DeepWiFi: Cognitive WiFi with Deep Learning,” IEEE Transactions on Mobile Computing, vol. 20, no. 2, pp. 429–444, 2021
work page 2021
Show all 15 references
-
[9]
Feature-Based Deep Neural Networks for Short-Term Prediction of WiFi Channel Occupancy Rate,
H. I. J. L. A. Al-Tahmeesschi, K. Umebayashi and M. L ´opez-Ben´ıtez, “Feature-Based Deep Neural Networks for Short-Term Prediction of WiFi Channel Occupancy Rate,”IEEE Access, vol. 9, pp. 85 645–85 660, 2021
2021
-
[10]
Predicting Wire- less Channel Quality by Means of Moving Averages and Regression Models,
G. Formis, S. Scanzio, G. Cena, and A. Valenzano, “Predicting Wire- less Channel Quality by Means of Moving Averages and Regression Models,” in 2023 IEEE 19th International Conference on Factory Communication Systems (WFCS) , Apr. 2023, pp. 1–8
2023
-
[11]
Linear Combination of Exponential Moving Averages for Wire- less Channel Prediction,
——, “Linear Combination of Exponential Moving Averages for Wire- less Channel Prediction,” in 2023 IEEE 21st International Conference on Industrial Informatics (INDIN) , 2023, pp. 1–6
2023
-
[12]
Predicting Wi-Fi link quality through artificial neural networks,
S. Scanzio, F. Xia, G. Cena, and A. Valenzano, “Predicting Wi-Fi link quality through artificial neural networks,” Internet Technology Letters, vol. 5, no. 2, p. e326, 2022. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/itl2.326
2022 doi
-
[13]
1–4379, 2021
“IEEE Standard for Information Technology–Telecommunications and Information Exchange between Systems - Local and Metropolitan Area Networks–Specific Requirements - Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications,” IEEE Std 802.11-2020...
2020
-
[14]
SDMAC: A Software-Defined MAC for Wi-Fi to Ease Implementation of Soft Real-Time Applica- tions,
G. Cena, S. Scanzio, and A. Valenzano, “SDMAC: A Software-Defined MAC for Wi-Fi to Ease Implementation of Soft Real-Time Applica- tions,” IEEE Transactions on Industrial Informatics , vol. 15, no. 6, pp. 3143–3154, 2019
2019
-
[15]
On the Use of Artificial Neural Networks to Predict the Quality of Wi-Fi Links,
A. S. Colletto, S. Scanzio, G. Formis, and G. Cena, “On the Use of Artificial Neural Networks to Predict the Quality of Wi-Fi Links,” IEEE Access, vol. 11, pp. 120 082–120 094, 2023
2023
Reviewed August 12, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.