TinyML-Based Adaptive Pulse Shaping for Edge Intelligence in IoT/IIoT

Afan Ali

REVIEW 3 major objections 5 minor 9 references

TinyML-Based Adaptive Pulse Shaping for Edge Intelligence in IoT/IIoT

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 The paper claims a TinyML adaptive pulse-shaping filter cuts uplink PAPR by 2 dB versus RRC in DFT-s-OFDM IoT links while trading SER against energy as SNR changes.

desk verdict Transmitter-side PAPR result is plausible and measured, but the receiver cannot know the data-dependent filter F, so the SER/end-to-end claims are not supported as written. read the letter →

arxiv 2506.05789 v1 pith:YVZSGNC6 submitted 2025-06-06 eess.SP

Afan Ali This is my paper

classification eess.SP

keywords TinyMLedgeintelligencepeak-to-averagepowerratioadaptivepulseshapingDFT-s-OFDMneuralnetworksInternetofThingsIndustrialIoT

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

The paper claims that a very small neural network can act as an adaptive pulse-shaping filter on battery-powered IoT/IIoT transmitters, replacing fixed filters such as root-raised-cosine. In a DFT-s-OFDM uplink, the network outputs coefficients of a polynomial that defines frequency-domain filter taps, and a channel-aware loss steers those taps to lower peak-to-average power ratio (PAPR) without giving up symbol error rate. Simulations put the result at about 6 dB PAPR at CCDF $10^{-3}$, a 2 dB saving over RRC and 1.5 dB over unfiltered DFT-s-OFDM, with on-device inference reported at roughly 10k FLOPs, 5 ms, and 0.8 mJ per transmission. If the claims hold, IoT sensors could save battery and adapt to noisy or fading channels entirely on the device.

What carries the argument

The load-bearing object is a polynomial tap model, $F_k = \sum_{z=0}^{Z} r_z k^z$ with $Z=5$, in which the coefficients $r_z$ are the outputs of a two-layer fully connected network with 10 hidden ReLU neurons and input $(S_{\mathrm{EXT}}, \mathrm{SNR})$. This representation turns pulse-shape design into a tiny regression problem: computing all 240 taps costs $240\times6=1440$ multiplications, and the pruned, quantized network is about 80 KB. The training objective is the joint loss $L = E + \lambda P$, with $P$ the integrated PAPR CCDF tail above $x_0 = 6$ dB and $\lambda$ taken from a lookup table indexed by SNR bins (0.1 at low SNR to 1.0 at high SNR). The paper also assumes the receiver applies matched filtering $R = \hat{Y}\odot F^*$, which is where the receiver's knowledge of $F$ enters.

What would settle it

Run the same link with a receiver that must estimate the filter taps from a short pilot block instead of being handed $F$, and compare SER and PAPR against Figs. 7b and 9. If the SER gap to RRC closes, or if freezing $F$ per SNR bin does not cost most of the 2 dB PAPR saving, the adaptive, data-dependent filter is not what is carrying the reported gain.

Watch

Extended reading notes

Core claim

The central discovery is that the filter taps in an uplink DFT-s-OFDM link do not have to come from a fixed design rule; a pruned, 8-bit quantized two-layer network with 10 hidden ReLU neurons can generate them on the fly. The network takes the spectrum-extended frequency-domain symbol vector $S_{\mathrm{EXT}}$ (length 240) plus the current SNR and outputs five coefficients $r_0,\dots,r_4$; each tap is then $F_k = \sum_{z=0}^{4} r_z k^z$. The taps are trained with the joint loss $L = E + \lambda P$, where $E$ is the simulated symbol error rate and $P = \int_{6\,\mathrm{dB}}^\infty \mathrm{CCDF}(x)\,dx$ is the tail of the PAPR distribution, with $\lambda$ chosen from a five-bin SNR lookup table. The paper reports that the adaptive filter reaches about 6 dB PAPR at CCDF $10^{-3}$, versus 8 dB for RRC and 7.5 dB for plain DFT-s-OFDM, and that the same architecture keeps working on unseen Rician fading and on 64-QAM despite training on QPSK/16-QAM.

Load-bearing premise

The result depends on the receiver being able to apply the exact filter taps the TinyML network generates, even though those taps are computed from the transmitted data and the current SNR; the paper describes no way for the receiver to obtain those taps before decoding.

Editorial extensions

If this is right

In an uplink DFT-s-OFDM link, replacing RRC filtering with the proposed TinyML filter lowers the PAPR at CCDF $10^{-3}$ from about 8 dB to about 6 dB, reducing power-amplifier back-off requirements and battery drain.
The same hardware switches its operating point according to channel SNR: with $\lambda=0.1$ at 5 dB SNR the filter protects symbol error rate, while $\lambda=1.0$ at 15 dB SNR pushes for PAPR reduction.
The reported resource footprint (about 80 KB flash, 20 KB RAM, 10k FLOPs, 5 ms inference, 0.8 mJ per transmission) fits microcontrollers such as the STM32L4, ESP32, and nRF52832, so the method is within reach of mass IoT deployments.
The filter trained on QPSK/16-QAM also performs on 64-QAM and on Rician fading channels, indicating that the learned pulse shape generalizes beyond its training modulation and channel model.
A 1000-sensor smart-factory estimate projects about 20 MWh/year saved, and a 10-sensor hardware testbed measured 1.9 dB PAPR reduction, supporting the scalability claim.

Reading between the lines

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

Editorial extension: because the taps are a five-coefficient polynomial, the same network could be asked to optimize other per-block objectives such as out-of-band emission, tone-reservation peaks, or spectral flatness by changing only the loss function.
Editorial extension: the five-bin SNR lookup table could be replaced by a continuous SNR-to-$\lambda$ mapping learned on-device, removing the quantization step between channel estimates and filter behavior.
Editorial extension: a pilot-aided or decision-directed receiver that estimates the filter taps would turn the current matched-filter assumption into a deployable end-to-end system and would test how much of the gain survives imperfect knowledge of $F$.

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.

Desk Editor's Note

read the letter

Bottom line: the transmitter-side PAPR claim is real and worth taking seriously, but the receiver-side model as written is unimplementable, and that kills the SER half of the paper as it stands. The resource arithmetic has some self-inflicted mistakes. Still, the core artifact—a pruned 8-bit network emitting polynomial filter coefficients for adaptive pulse shaping—is new in the engineering sense and has a measured on-device result.

What's actually new: unlike the cited earlier work on static pulse shaping and generic ML, this paper shows a specific deployment: a two-layer network with 10 hidden neurons, 80% sparsity, 8-bit quantization, running on STM32L4/ESP32/nRF52832, with measured 1.9 dB PAPR reduction vs. simulated 2 dB, and latency around 5 ms. That's a concrete, reproducible hardware result. The CCDF plot in Fig. 5 is internally consistent, and the paper is honest about the adversarial-robustness gap in the conclusion.

Soft spots, in order of severity. First, the receiver cannot know F as described. The filter F is computed from SEXT (the transmitted data symbols) and SNR; the receiver only has Ŷ and an SNR report. No pilot, feedback, or decision-directed mechanism is given. So matched filtering R = Ŷ ⊙ F* is not implementable, and the SER curves in Fig. 7b and Fig. 9 rest on a genie assumption. The transmitter-side PAPR claim survives this, but the end-to-end system claim does not. This is a missing piece of the system model, not a cosmetic detail.

Second, the numbers don't add up. The pruned model is 492 bytes of weights but is claimed at 80 KB flash (or 100 KB earlier). The 20 MWh/year figure doesn't follow from 1000 sensors at 1 W for 8760 hours (total 8.76 MWh, 33% saving ≈ 2.9 MWh). These inconsistencies are easy to fix but signal that the resource claims need rechecking.

Third, the PAPR gain is essentially the training objective: the loss includes the PAPR tail P, so Fig. 5 is a fitted outcome. That's not a fatal flaw per se—fitted performance is the point of learned pulse shaping—but the paper should say so and show how the hand-set λ and x0 affect the result. The λ lookup and x0=6 dB are underjustified, though they're secondary.

The paper is written for people working on PAPR reduction and low-power IoT physical layer. A reader interested in the hardware deployment and the measured PAPR reduction gets value from Section IV.B. The SER claims should be ignored until the receiver problem is solved. I'd send it to a serious referee with the expectation of heavy revision, not because the idea is bad, but because the receiver issue and arithmetic need to be resolved before the system claim is credible.

Referee Report

3 major / 5 minor

Summary. The paper proposes a TinyML-based adaptive pulse shaping filter for DFT-s-OFDM uplink in IoT/IIoT scenarios. The filter taps are computed by a small pruned and quantized neural network that takes the spectrum-extended transmitted symbols and an SNR estimate as inputs, with a loss function that trades SER against the PAPR tail. The authors report about 2 dB PAPR savings over RRC at CCDF=10^-3, similar SER performance to conventional baselines, and validate on an STM32L4 testbed with about 5 ms latency and 0.8 mJ per transmission. The main advertised contribution is a device-deployable adaptive pulse shape that optimizes both PAPR and SER in real time.

Significance. If the claims held, the paper would offer a practical contribution to low-complexity adaptive pulse shaping on microcontrollers, with a concrete energy/latency measurement on real hardware. The on-device validation is a positive feature. However, the end-to-end SER results are not supported because the receiver's matched-filtering operation requires knowledge of the data-dependent filter taps F, which cannot be obtained with the described architecture. This is a load-bearing gap that invalidates the adaptive PAPR-and-SER system as presented. The PAPR gain itself is also partly the training objective, so its status as an independent result needs clarification.

major comments (3)

[Section II and Algorithm 1] The receiver operation R = Y_hat ⊙ F* is defined in Section II, but the filter F is computed on-device from the transmitted data block SEXT and the SNR (Section III.A, Algorithm 1). Since SEXT is exactly the data to be decoded, the receiver cannot reconstruct F before detection. No pilot symbols, feedback of F, or decision-directed iterative mechanism is described. Consequently, the SER-versus-SNR results in Fig. 7b and Fig. 9 correspond to an undeclared genie-aided receiver or to a different quantity than the stated system. This is a central flaw for the advertised SER optimization.
[Section III.B and Fig. 5] The loss function L = E + λP directly minimizes P, the integral of the CCDF tail beyond x0 (Eq. 5), and Fig. 5 reports the same CCDF tail. Thus the headline 2 dB PAPR saving over RRC is largely an evaluation of the training criterion, not an independent predictive claim. To make the comparison meaningful, the authors should compare against a static pulse shape that is optimized for the same objective (e.g., an RRC with optimized roll-off, or a fixed polynomial filter with the same number of degrees of freedom) and should provide a sensitivity analysis for the hand-chosen x0 = 6 dB and the λ lookup values.
[Section III.A vs. Section III.F and Fig. 10] The computational and latency numbers are internally inconsistent: Section III.A states 10k FLOPs and 1440 multiplications, while Section III.F gives 4 ms latency and Algorithm 1 and Fig. 10a report 5 ms. The 0.8 mJ per transmission in Table I is presented without a breakdown of how it is derived or measured. These inconsistencies make it hard to evaluate the claimed energy advantage over RRC, and they should be reconciled with the exact measurement conditions of the STM32L4 testbed.

minor comments (5)

[Eq. (3)] Equation (3) defines the filter taps as Σ_{z=0}^{Z} r_z[k]^z, with Z=5 in the text, but the output layer produces coefficients for z=0,...,4; please clarify whether Z is the polynomial degree or the number of terms and align the notation throughout.
[Fig. 6] The acronyms CLF and SLM are used in the caption and text without definition; also define what the x-axis 'number of symbols' represents.
[Section IV.A, Fig. 9] For the 64-QAM generalization experiment, no baseline (RRC or DFT-s-OFDM under the same 64-QAM conditions) is shown, so it is unclear whether the proposed filter remains beneficial for the higher-order modulation; please add a comparison.
[Section IV.B] The sentence 'PAPR reduction was 1.9 dB vs. RRC's 2 dB as shown in Fig. 10b' is ambiguous; it should be rephrased to indicate the measured PAPR saving of the TinyML filter relative to RRC.
[References] Reference [7] is cited as an RRC filter, but the listed reference is a two-parametric ISI-free pulse; the citation should be corrected or the related work description should be adjusted.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the PAPR and SER results are direct evaluations of the explicitly stated training objective against external benchmarks, not fitted inputs renamed as predictions.

full rationale

The paper's central claims are empirical rather than derived from a first-principles chain. The loss function L = E + λP (Eq. 4) explicitly defines the training objective, with P being the PAPR-tail integral (Eq. 5), and the reported PAPR CCDF results (Figs. 5, 6, 8, 10) are measurements of that same objective compared externally against RRC and DFT-s-OFDM. Reporting lower PAPR after explicitly training to minimize a PAPR-tail term is the expected effect of optimization, not a fitted parameter renamed as a prediction; the comparison against fixed baselines gives independent content. The receiver-side matched filtering requires knowledge of F, which depends on transmitted data via the network; this is an implementability and correctness gap, not a circular reduction, because it does not make any output equal to an input by construction. No load-bearing self-citations appear: references [7]-[9] are external baselines or standard optimizers, and the λ lookup table is a hand-set design choice without hidden circular dependency. The paper is self-contained against external benchmarks, so no circularity is found.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The load-bearing content rests on unpaid-for premises: (1) receiver-side knowledge of the data-dependent filter F, which the paper assumes in Section II but never equips the receiver with, and (2) hand-tuned loss constants (lambda values, x0 = 6 dB) that carry the whole SER/PAPR balance with no normalization or sensitivity study. The filter is an algorithm, not a new physical entity, so the invented-entities list is empty.

free parameters (6)

lambda lookup values = 0.1, 0.3, 0.5, 0.8, 1.0 over five SNR bins
Section III.B: hand-set trade-off weights with no fitting or sensitivity analysis; the text example (15 dB gives lambda=1.0) contradicts the table (10-15 dB gives 0.5, 15-20 dB gives 0.8).
PAPR tail threshold x0 = 6 dB
Section III.B, Eq. (5): the loss integrates the CCDF tail above 6 dB, described as an average IoT PAPR threshold with no derivation.
Polynomial order Z and coefficient count = Z=5 with five output neurons (z=0..4)
Section III.A: chosen for a trade-off; the text is inconsistent about whether the sum runs to 5 or 6 terms, and there is no ablation of Z.
Hidden layer width = 10 neurons
Section III.A and Fig. 8: the ablation shows 10 and 20 neurons give the same 2 dB gain, so the choice is not principled.
Spectrum extension length NSE = 15 per side
Table III: NSE changes the DFT-s-OFDM PAPR baseline and is fixed by hand with no sweep.
Table I complexity and energy figures = 0.2 MIPS, 0.8 mJ for the proposed filter; 0.5 MIPS, 1.2 mJ for RRC
Section III.E: no measurement methodology (cycle counts, current probe, or logic analyzer) is described for these numbers.

assumptions (4)

ad hoc to paper The receiver can obtain the data-dependent filter taps F, allowing matched filtering R = Y_hat .* F*
Section II defines the receiver as applying matched filtering with F*, but Section III.A and Fig. 4 define F as the output of a network whose inputs include the transmitted symbols SEXT. No pilot, feedback, or iterative mechanism is provided, so this assumption is introduced to make the simulation's receiver work.
domain assumption The loss L = E + lambda*P with hand-set lambda and x0 encodes a valid SER/PAPR trade-off
Eqs. (4)-(5) add an MSE term E and a CCDF-tail integral P in dB units; the two have incompatible scales and the balance rests entirely on the hand-chosen lambda lookup values and the x0 = 6 dB threshold (Section III.B).
domain assumption Offline training on simulated AWGN, Rayleigh, and Rician channels transfers to deployment channels
Section III.C trains on SNR 0-20 dB simulated blocks; the deployment uses only live SNR feedback and the generalization claim rests on Figs. 7-9 without a distribution-shift analysis.
standard math Standard DFT-s-OFDM signal algebra and PAPR/CCDF statistics
Eqs. (1)-(3) and (5) use the conventional DFT-s-OFDM model, PAPR definition, and CCDF tail integral without modification.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TinyML-Based Adaptive Pulse Shaping for Edge Intelligence in IoT/IIoT." pith.science (2026). https://pith.science/paper/YVZSGNC6

@misc{pith2026250605789,
  author       = {Pith},
  title        = {Pith review of: TinyML-Based Adaptive Pulse Shaping for Edge Intelligence in IoT/IIoT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YVZSGNC6}},
  note         = {Machine review of arXiv:2506.05789}
}

read the original abstract

Edge intelligence in IoT and IIoT demands lightweight algorithms for data processing on resource-constrained devices. This paper introduces a novel adaptive pulse shape filter based on TinyML for PAPR and SER optimization on edge devices used in uplink IoT communication. Implemented on IoT nodes such as sensors, our pruned neural network provides up to 2 dB PAPR saving over root-raised-cosine (RRC) filters. Mass simulations validate its efficacy in DFT-s-OFDM systems and offer an energy-efficient and scalable solution for IoT/IIoT use cases such as smart factories and rural connectivity.

Figures

Figures reproduced from arXiv: 2506.05789 by the authors.

**Figure 2.** Proposed Method for PAPR reduction. The spectrum extension (SE) prepends and appends NSE symbols, yielding S EXT = [SNdata−NSE+1, . . . , SNdata , S1, . . . , SNdata , S1, . . . , SNSE ], where NSK = Ndata + 2NSE. The TinyML filter F = [F1, . . . , FNSK ], applied on-device, produces S ′ = S EXT ⊙ F. After IDFT (normalized by 1/ √ NFFT), the time-domain signal s ′ is transmitted. The receiver processes yˆ = s ′ + w,… view at source ↗

**Figure 5.** CCDF of PAPR for TinyML filter vs. benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗

Figures from the paper (4 more)

**Figure 6.** Figure 6: Average PAPR for CLF,SLM and proposed method. [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 8.** Figure 8: CCDF of PAPR for the TinyML-based filter (10 [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]

**Figure 11.** Figure 11: Performance comparison of the TinyML-based filter [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]

**Figure 10.** Figure 10: Latency and PAPR reduction plots for on-device [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

[1]

S. U. Amin and M. S. Hossain, ”Edge Intelligence and Internet of Things in Healthcare: A Survey,” in IEEE Access, vol. 9, pp. 45-59, 2021

work page 2021
[2]

Atzori et al., ”The Internet of Things: A survey,” Comput

L. Atzori et al., ”The Internet of Things: A survey,” Comput. Netw., vol. 54, no. 15, pp. 2787–2805, Oct. 2010

work page 2010
[3]

Beaulieu et al., ”A ’better than’ Nyquist pulse,” IEEE Commun

N. Beaulieu et al., ”A ’better than’ Nyquist pulse,” IEEE Commun. Lett., vol. 5, no. 9, pp. 367–368, Sep. 2001

work page 2001
[4]

R. H. Jhaveri, H. R. Chi and H. Wu, ”TinyML for Empowering Low- Power IoT Edge Consumer Devices,” in IEEE Transactions on Consumer Electronics, vol. 70, no. 4, pp. 7318-7321, Nov. 2024

work page 2024
[5]

Li and L

X. Li and L. J. Cimini, ”Effects of clipping and filtering on the performance of OFDM,” IEEE Commun. Lett., vol. 2, no. 5, pp. 131–133, May 1998

work page 1998
[6]

Ohkubo and T

N. Ohkubo and T. Ohtsuki, ”A peak to average power ratio reduction of multicarrier CDMA using selected mapping,” Proceedings IEEE 56th Vehicular Technology Conference, Vancouver, BC, Canada, pp. 2086- 2090 vol.4, 2002

work page 2002
[7]

Tyrowolas et al., ”A novel two-parametric ISI-free pulse,” IEEE Wireless Commun

D. Tyrowolas et al., ”A novel two-parametric ISI-free pulse,” IEEE Wireless Commun. Lett., vol. 12, no. 3, pp. 501–504, Mar. 2023

work page 2023
[8]

Ye et al., ”Channel agnostic end-to-end learning with conditional GAN,” in Proc

H. Ye et al., ”Channel agnostic end-to-end learning with conditional GAN,” in Proc. IEEE Globecom Workshops, pp. 1–5, 2018

work page 2018

Show all 9 references

[9]

Loshchilov and F

I. Loshchilov and F. Hutter, ”Decoupled weight decay regularization,” in Proc. ICLR, pp. 1–12, 2019

2019

Pith tools

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

[1] [1]

S. U. Amin and M. S. Hossain, ”Edge Intelligence and Internet of Things in Healthcare: A Survey,” in IEEE Access, vol. 9, pp. 45-59, 2021

work page 2021

[2] [2]

Atzori et al., ”The Internet of Things: A survey,” Comput

L. Atzori et al., ”The Internet of Things: A survey,” Comput. Netw., vol. 54, no. 15, pp. 2787–2805, Oct. 2010

work page 2010

[3] [3]

Beaulieu et al., ”A ’better than’ Nyquist pulse,” IEEE Commun

N. Beaulieu et al., ”A ’better than’ Nyquist pulse,” IEEE Commun. Lett., vol. 5, no. 9, pp. 367–368, Sep. 2001

work page 2001

[4] [4]

R. H. Jhaveri, H. R. Chi and H. Wu, ”TinyML for Empowering Low- Power IoT Edge Consumer Devices,” in IEEE Transactions on Consumer Electronics, vol. 70, no. 4, pp. 7318-7321, Nov. 2024

work page 2024

[5] [5]

Li and L

X. Li and L. J. Cimini, ”Effects of clipping and filtering on the performance of OFDM,” IEEE Commun. Lett., vol. 2, no. 5, pp. 131–133, May 1998

work page 1998

[6] [6]

Ohkubo and T

N. Ohkubo and T. Ohtsuki, ”A peak to average power ratio reduction of multicarrier CDMA using selected mapping,” Proceedings IEEE 56th Vehicular Technology Conference, Vancouver, BC, Canada, pp. 2086- 2090 vol.4, 2002

work page 2002

[7] [7]

Tyrowolas et al., ”A novel two-parametric ISI-free pulse,” IEEE Wireless Commun

D. Tyrowolas et al., ”A novel two-parametric ISI-free pulse,” IEEE Wireless Commun. Lett., vol. 12, no. 3, pp. 501–504, Mar. 2023

work page 2023

[8] [8]

Ye et al., ”Channel agnostic end-to-end learning with conditional GAN,” in Proc

H. Ye et al., ”Channel agnostic end-to-end learning with conditional GAN,” in Proc. IEEE Globecom Workshops, pp. 1–5, 2018

work page 2018