A mask-based peak-to-average power ratio reduction scheme for affine frequency division multiplexing systems using Volterra Series-Driven PolyNet Authors: Xiangxin Liu, Yushuai Zhang, Jianxin Guo, Rui Zhu, Feng Wang, Liping Wang Published: 2026 ABSTRACT Affine frequency division multiplexing (AFDM) demonstrates exceptional resilience to Doppler effects in doubly dispersive channels, making it a promising waveform for 6G high-mobility communications, but its high peak-to-average power ratio (PAPR) problem severely limits system energy efficiency. When migrated to AFDM, existing mask-based deep learning PAPR reduction schemes face the challenges of contextual loss caused by local sampling and the lack of physical interpretability in black-box models. Therefore, this paper proposes a physics-aware mask-based PAPR reduction scheme tailored for AFDM. First, a full-frame input strategy is introduced to exploit the global time-domain correlation of AFDM signals for precisely recovering the impaired symbols. Second, to address the nonlinear distortion induced by masking, a PolyNet-Volterra network is proposed by integrating Volterra series theory. Departing from the traditional design of blindly stacking layers, this model explicitly constructs first-order linear and third-order power feature layers, which substantially enhances the nonlinear reconstruction accuracy while avoiding overfitting. Simulation results demonstrate that, under a threshold of 0.9, the proposed scheme achieves a significant PAPR reduction of approximately 5.9 dB. Furthermore, requiring only about 165,000 parameters, the PolyNet-Volterra model comprehensively outperforms baseline models such as DNN and ResNet in terms of both bit error rate (BER) and mean squared error (MSE). Compared with the conventional PolyNet, the proposed model reduces the parameter and computational overhead, suggesting its potential as a low-complexity receiver-side reconstruction module for resource-constrained future wireless systems. Further hardware-oriented and over-the-air validation is still needed before drawing deployment-level conclusions. Introduction With the rapid evolution of mobile communication technologies, the sixth-generation (6G) wireless network has become a global focal point for breakthrough technologies [1–3]. It aims to provide highly reliable and ubiquitous connectivity services for high-mobility scenarios, such as high-speed railways, low Earth orbit (LEO) satellite internet, and unmanned aerial vehicle (UAV) communications [4]. In such demanding scenarios, the high relative velocity between terminals and base stations induces significant Doppler shifts, causing the wireless channel to exhibit severe doubly selective fading characteristics. This leads to inter-symbol interference and frequency-selective fading, which makes channel estimation and equalization more difficult. To tackle these technical challenges, affine frequency division multiplexing (AFDM) has emerged as a novel multi-carrier modulation technique [5–7]. By incorporating the discrete affine Fourier transform (DAFT), AFDM can achieve full diversity gain in doubly dispersive channels. Furthermore, it not only delivers superior resilience to Doppler shifts compared to orthogonal frequency division multiplexing (OFDM), but also entails lower overhead for channel estimation and multi-user multiplexing than orthogonal time frequency space (OTFS) modulation [8–10]. AFDM is promising for high-mobility scenarios, but its time-domain waveform is still formed by the superposition of multiple chirp-modulated subcarriers. Therefore, the high peak-to-average power ratio (PAPR) problem still needs to be addressed [11,12]. When these signals with extremely high instantaneous peaks pass through a high power amplifier (HPA), they are driven into the saturation region, exceeding the linear amplification limit and suffering from clipping. This phenomenon inevitably causes in-band distortion, which elevates the bit error rate, and out-of-band radiation, which interferes with adjacent channels. Therefore, energy-constrained satellite, vehicular, and low-altitude communication scenarios provide a strong motivation for studying low-complexity AFDM PAPR reduction methods while maintaining communication reliability [13]. Currently, research on PAPR reduction is quite mature, encompassing traditional methods such as clipping, partial transmit sequence (PTS), and selective mapping (SLM) [14–16]. However, while clipping is simple and straightforward, it introduces irreversible nonlinear distortion, severely degrading the BER at the receiver [17,18]. Conversely, probabilistic schemes like PTS and SLM require the transmission of additional side information, and their computational complexity grows exponentially with the number of candidate sequences. In recent years, with the widespread application of deep learning in communication physical layer technologies [19], existing schemes [20–23] have successfully achieved PAPR reduction by leveraging deep neural networks (DNNs). Most notably, He et al. [24] proposed a mask-based PAPR reduction architecture for OFDM systems. This scheme masks a small fraction of high-peak symbols at the transmitter and uses a DNN at the receiver for signal reconstruction. In this way, PAPR reduction and signal recovery are handled in two separate steps, which improves the overall performance. However, several critical issues in existing research remain unresolved. First, the input to their mask reconstruction neural network comprises only the symbols at the masked positions within the received signal, along with the indices of these positions. Specifically, rather than utilizing the entire received signal, the model’s input is strictly confined to the masked symbols. Although this input strategy reduces computational complexity, it discards much of the time-domain context around the masked samples. As a result, the network has limited information for estimating the local waveform trajectory, which may lead to similar reconstruction limits across different model architectures [25]. Furthermore, these existing black-box models ignore the physical mechanisms of communication signal distortion. They typically rely on blindly stacking massive numbers of neurons and network layers to approximate the nonlinear mapping, resulting in an enormous number of parameters and sluggish training convergence. This not only exacerbates the computational burden on the receiver but also fails to meet the stringent requirements of future 6G networks for low-latency and low-power communications [26–28]. To address the aforementioned dual challenges, this paper proposes a physics-aware mask-based PAPR reduction scheme tailored for AFDM systems. Initially, the scheme employs a dynamic masking mechanism to precisely locate high-peak regions within the signal, strategically replacing these high instantaneous power peaks to significantly reduce the PAPR at the source. Subsequently, a lightweight polynomial neural network designed based on Volterra series theory (termed PolyNet-Volterra) is utilized to perform nonlinear reconstruction of the impaired symbols within these peak regions. This innovative approach substantially lowers the PAPR while maintaining a low bit error rate (BER). The main contributions of this paper are summarized as follows: We introduce a mask-based PAPR reduction scheme into the AFDM system, validating its effectiveness under chirp modulation. We also extend the masked-symbol input strategy to a full-frame input strategy. This allows the network to use frame-level signal context when reconstructing the masked samples. We introduce a mask-based PAPR reduction scheme into the AFDM system, validating its effectiveness under chirp modulation. We also extend the masked-symbol input strategy to a full-frame input strategy. This allows the network to use frame-level signal context when reconstructing the masked samples. To handle the nonlinear amplitude distortion introduced by masking, we propose PolyNet-Volterra, a neural network model inspired by the Volterra series. The model uses power-related feature layers to improve interpretability. A mask-specific loss function is also used to improve reconstruction accuracy at high-energy masked samples. To handle the nonlinear amplitude distortion introduced by masking, we propose PolyNet-Volterra, a neural network model inspired by the Volterra series. The model uses power-related feature layers to improve interpretability. A mask-specific loss function is also used to improve reconstruction accuracy at high-energy masked samples. Simulation results demonstrate that, under identical PAPR reduction levels, the proposed PolyNet-Volterra model achieves faster convergence and superior BER performance compared to conventional DNN and ResNet models, while drastically reducing the parameter count. These results indicate that PolyNet-Volterra is a promising low-complexity candidate for AFDM PAPR reduction, while its practical deployment still requires further validation under hardware and over-the-air conditions. Simulation results demonstrate that, under identical PAPR reduction levels, the proposed PolyNet-Volterra model achieves faster convergence and superior BER performance compared to conventional DNN and ResNet models, while drastically reducing the parameter count. These results indicate that PolyNet-Volterra is a promising low-complexity candidate for AFDM PAPR reduction, while its practical deployment still requires further validation under hardware and over-the-air conditions. The remainder of this paper is organized as follows. The Related Work section reviews existing approaches. The AFDM System Model section introduces the system model. The Simulation Results Analysis section presents and analyzes the experimental results. Finally, the Conclusion section summarizes the study. Related work In multi-carrier communication systems, the high peak-to-average power ratio (PAPR) has been a persistent technical challenge. For a continuous time-domain signal x(t),its PAPR is defined as the ratio of the peak power to the average power, which is mathematically expressed as: where max0≤t