High Performance Signal Design for ACO-OFDM Systems using Variational Autoencoder

Nam N. Luong, Chuyen T. Nguyen, Thanh V. Pham · 2025

Read High Performance Signal Design for ACO-OFDM Systems using Variational Autoencoder by Nam N. Luong, Chuyen T. Nguyen, Thanh V. Pham (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-07362.txt.

Introduction

Nam Luong, Chuyen Nguyen, and Thanh Pham design signals for light based data links in a 2025 paper. Their system sends data over light waves split into many channels. Good signals must keep error rates low, carry much data, and avoid sharp power peaks that distort the light source.

The design uses a learning model that shapes both the layout of signal points and how often each point is sent. A gradual training scheme tunes the two together. Simulations across a wide range of signal strengths show much lower power peaks than past designs. Error rates and data throughput also improved. The method suits visible light links in rooms where radio is weak or barred.

Catalog record

CREATORNam N. Luong, Chuyen T. Nguyen, Thanh V. Pham
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE22 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-07362

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

High Performance Signal Design for ACO-OFDM Systems using Variational Autoencoder Nam N. Luong, Chuyen T. Nguyen, Thanh V. Pham Abstract This letter proposes a design of low peak-to-average power ratio (PAPR), low symbol error rate (SER), and high data rate signal for asymmetrically clipped optical orthogonal frequency division multiplexing (ACO-OFDM) systems. The proposed design leverages a variational autoencoder (VAE) incorporating gradual loss learning to jointly optimize the geometry and probability of the constellation’s symbols. This not only enhances mutual information (MI) but also effectively reduces the PAPR while maintaining a low SER for reliable transmission. We evaluate the performance of the proposed VAE-based design by comparing the MI, SER, and PAPR against existing techniques. Simulation results demonstrate that the proposed method achieves a considerably lower PAPR while maintaining superior SER and MI performance for a wide range of SNRs. I Introduction Optical wireless communications (OWC) has emerged as one of the most rapidly developing communication technologies in recent years. Due to distinct advantages such as unlicensed spectrum and immunity to electromagnetic interference, OWC is a promising complement and/or alternative to radio frequency (RF) communications [ 1 ] . OWC systems are typically categorized into free-space optics (FSO) and visible light communications (VLC), where the former is usually used for outdoor applications, while the latter is often employed for indoor scenarios. To address challenges commonly encountered in wireless systems, such as frequency selective fading and multipath fading, orthogonal frequency division multiplexing (OFDM) has been widely adopted due to its robustness to these impairments. Conventional OFDM cannot be directly applied to OWC systems due to the real and nonnegative signal required by optical intensity modulation and direct detection (IM/DD). Therefore, several OFDM variants have been developed specifically for optical systems, notably direct current biased optical OFDM (DCO-OFDM) and asymmetrically clipped optical OFDM (ACO-OFDM) [ 2 ] . Optical OFDM utilizes Hermitian symmetry to ensure real-valued signals before the inverse fast Fourier transform (IFFT). Different methods are then applied to generate non-negative time-domain signals (e.g., DC biasing in DCO-OFDM). A major drawback of OFDM-based systems is the inherent high peak-to-average power ratio (PAPR), which arises due to the summation of multiple subcarriers with different phases, occasionally producing large peaks in the signal waveform. Due to the nonlinearity of power amplifiers at the high-power region, this high PAPR can lead to severe nonlinear signal distortions, which results in reduced system efficiency and degraded symbol error rate (SER) performance. Several techniques have been proposed for PAPR reduction, including selective mapping (SLM) [ 3 ] , partial transmit sequence (PTS) [ 4 ] , and clipping [ 5 ] . Clipping limits the signal amplitude by truncating it beyond a certain threshold. Although simple, this approach introduces signal distortions, hence deteriorating the SER. In contrast, SLM and PTS mitigate signal peaks by modifying phase factors. While SLM selects the lowest PAPR version among multiple phase-shifted signals, PTS optimizes sub-block phases. Though effective, these methods incur high computational complexity due to exhaustive searches for optimal phase factors. In recent years, deep learning (DL) has gained significant attention in tackling challenges in the physical layer design, particularly using autoencoders (AEs) to model end-to-end communication systems [ 6 , 7 ] . In this framework, the encoder represents the transmitter, while the decoder functions as the receiver. Several studies have investigated the potential of using AEs for PAPR reduction in OFDM systems [ 8 , 9 , 10 , 11 ] . In [ 8 ] , Kim et~al. proposed a PAPR reduction method using an AE called PRnet. Simulation results demonstrated that this approach achieves superior PAPR reduction compared to conventional methods while maintaining a good bit error rate (BER) performance. Using a similar approach, [ 9 ] employs an AE combined with clipping to reduce PAPR in DCO-OFDM systems. Due to the AE’s powerful signal reconstruction capability, the effect of clipping in DCO-OFDM on the BER was minimized. The authors in [ 10 ] proposed a hybrid AE, combining an AE with a phase rotator to reduce PAPR for ACO-OFDM. This design achieved better results than the AE-only approach at the expense of significantly higher computational complexity. In [ 11 ] , the authors designed an OFDM signal with a low PAPR waveform while simultaneously addressing the amplifier nonlinearity. In this study, we propose a novel approach for generating OFDM signals with low PAPR, lower SER, and high data rate for ACO-OFDM systems by leveraging a variational autoencoder (VAE) to optimize the probability and geometry of the constellation’s symbols, i.e., probabilistic and geometric constellation shaping (PCS and GCS). Unlike existing methods that consider uniform signaling, the proposed design, by properly shaping the symbol constellation, not only reduces the PAPR but also achieves lower SER and higher data rate which is calculated as the symbol-wise mutual information (MI) [ 7 ] . To achieve better learning, a two-stage training of the proposed model is performed. In the first stage, the loss function only considers MI to improve the accuracy of symbol detection. Once the first stage is completed, the second stage introduces an additional component of the weighted PAPR. This technique, known as gradual loss learning [ 11 ] , helps prevent the network from being distracted by multiple competing objectives during training. The main contributions in the paper are summarized as follows. • We propose a VAE-based signal design for ACO-OFDM systems that achieves low PAPR, low SER, and high data rate. Specifically, the proposed design employs three neural networks (NNs) for three different tasks: one for PCS optimization, one for GCS optimization and PAPR reduction (i.e., implementing the modulator), and one for signal reconstruction at the receiver (i.e., implementing the demodulator), respectively. • We present comprehensive comparisons regarding PAPR, SER, and MI between the proposed approach and some common methods such as SLM, clipping, and the PRnet [ 8 ] . Notation : In this paper, H(\cdot) , I(\cdot,\cdot) , and \mathbb{E}[\cdot] are used to denote the entropy, mutual information, and expected operation, respectively. II System model and Autoencoder II-A System model Fig. 1 : Schematic diagram of ACO-OFDM systems. As illustrated in Fig. 1 , we consider in this study an ACO-OFDM system employing M -ary quadrature amplitude modulation ( M -QAM) with 4N subcarriers where the input bitstream b is mapped to QAM symbols by a bit-to-symbol mapping block according to a parametric distribution p_{\boldsymbol{\theta}_{P}}(s) with parameters \boldsymbol{\theta}_{P} which is trained by a NN. For simplicity and clarity, in the subsequent sections, we assume the presence of this mapping block and focus on the symbols rather than the input bitstream. The symbols are then fed into a modulator, which is implemented as a NN with trainable parameters \boldsymbol{\theta}_{G} , which reconstructs the To ensure a real time-domain signal after IFFT, the symbols are then rearranged using the Hermitian symmetry operation, where one-quarter of the total subcarriers, specifically the odd-indexed subcarriers, carry the data signals \tilde{\mathbf{X}}=[\tilde{X}_{0},\tilde{X}_{1},\ldots,\tilde{X}_{N-1}] . Another quarter is assigned to the conjugates of \tilde{X} , while the remaining half, corresponding to the even-indexed subcarriers, is allocated to null data signals. Accordingly, the frequency-domain data vector is given by \mathbf{X}_{H}=[0,\tilde{X}_{0},0,\tilde{X}_{1},\ldots,0,\tilde{X}_{N-1},0,\tilde{X}_{N-1}^{*},\ldots,0,\tilde{X}^{*}_{1},0,\tilde{X}^{*}_{0}] where \tilde{X}^{*}_{n} denotes the complex conjugate of \tilde{X}_{n} . The real time-domain signal \hat{x}

Read the complete text