Discrete Modulation for Semantic and Over the Air Computation

17 concepts17 papers1 proposal

The theoretical promises of next-generation wireless networks are dazzling. In Semantic Communication, instead of transmitting every raw bit perfectly, devices extract and transmit only the underlying "meaning" of data, promising massive bandwidth savings. In Over-the-Air Computation (AirComp), instead of scheduling separate time slots for every device to upload its data, hundreds of edge devices transmit their signals simultaneously over the same frequency. The wireless channel itself acts as a natural computer, physically superimposing the waves so that the receiver instantly observes the mathematical sum of the inputs—a massive boon for federated learning.

Yet, these paradigms face a shared, fundamental roadblock: they are natively formulated as analog operations. Semantic communication naturally outputs continuous neural network feature vectors. AirComp relies on the infinite precision of physical wave superposition and precise channel inversion.

Modern commercial wireless hardware (such as 5G, Wi-Fi, and LTE) is strictly digital. It operates on discrete bits and predefined, rigid geometric signal constellations like Quadrature Amplitude Modulation (QAM). Forcing continuous analog designs onto digital chips causes severe hardware degradation (such as power amplifier distortion), demands impossible phase and amplitude precision from low-power IoT nodes, and breaks standard network protocols.

To bridge this gap, a major research direction has emerged: designing discrete modulation, quantization, and coding schemes that translate continuous semantic and aggregated signals into digital-native formats without losing their unique mathematical properties.


Digital Over-the-Air Computation: Summation in a Discrete World

To understand why digital AirComp is challenging, consider the classic analog approach. If Device A transmits a continuous value $x_1$ and Device B transmits $x_2$, the receiver gets $h_1 x_1 + h_2 x_2$, where $h_i$ represents the wireless channel fading. If the devices know their channels, they can perform "channel inversion" by scaling their transmission power by $1/h_i$, resulting in a perfect sum ($x_1 + x_2$) at the receiver.

Digital hardware cannot easily perform arbitrary power scaling or transmit infinite-precision values. Researchers have tackled this by redesigning how data is quantized and mapped to digital constellations.

1. Coarse Quantization and Phase Alignment

One way to adapt to digital hardware is to simplify the signal. Deniz Gunduz et al. developed One-Bit Over-the-Air Aggregation for Communication-Efficient Federated Edge Learning: Design and Convergence Analysis, introducing One-Bit Broadband Digital Aggregation (OBDA). By quantizing machine learning gradients down to just a single bit ($\pm 1$) and mapping them to digital modulation, the authors enabled massive device concurrency with constant latency.

To bypass the need for precise power control in low-power IoT devices, Ruichen Jiang et al. proposed a cluster-based cooperative framework in Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning. The authors demonstrated that devices can drop magnitude adjustments entirely, relying instead on simple phase correction to achieve model convergence speeds that rival noise-free environments. To mitigate the information loss inherent in 1-bit schemes, Li Qiao et al. introduced a generalized digital AirComp framework in Unsourced Massive Access-Based Digital Over-the-Air Computation for Efficient Federated Edge Learning, leveraging Unsourced Massive Access (UMA) and message-passing algorithms to decode multi-bit sums from hundreds of uncoordinated devices.

2. Constellation Engineering and Lattice Coding

When multiple devices transmit discrete digital symbols simultaneously, the receiver faces "summation ambiguity"—the sum of two standard constellation points might map to multiple conflicting combinations of original data.

The authors of An Autoencoder-Based Constellation Design for AirComp in Wireless Federated Learning (Yujia Mu et al.) solved this by using autoencoders to optimize custom digital signal constellations. These learned constellations ensure that the physical sums remain uniquely decodable at the parameter server.

Standard QAM (Ambiguous Sums)          Optimized/Lattice Constellation
      o       o                             +       +       +

      o       o                             +       +       +
(Sums overlap confusingly)             (Sums fall on distinct, valid points)

A mathematically elegant alternative is lattice coding. In geometry, a lattice is a regular repeated pattern of points where the sum of any two lattice points is guaranteed to be another point in the same lattice.

Seyed Mohammad Azimi-Abarghouyi et al. exploited this property in Compute-Update Federated Learning: A Lattice Coding Approach and Federated Learning via Lattice Joint Source-Channel Coding. By treating model quantization and wireless interference as a joint lattice coding problem, the authors enabled digital hardware to natively compute sums over the air without requiring devices to possess perfect channel state information.

3. System-Level Optimization

To make these systems practical, the underlying communication infrastructure must be co-designed with the machine learning workload. Liang Li et al. presented a high-precision digital AirComp framework in Energy and Spectrum Efficient Federated Learning via High-Precision Over-the-Air Computation, which works directly with standard QAM and optimizes how often devices compute versus transmit, yielding up to a 3x improvement in energy efficiency.

To address receiver complexity, Xin Xie et al. focused on the joint design of channel coding and digital modulation in Joint Design of Coding and Modulation for Digital Over-the-Air Computation, significantly reducing decoding complexity. Pushing the boundary of hardware integration, Zhaohui Yang et al. demonstrated in Over-the-Air Split Learning with MIMO-Based Neural Network and Constellation-Based Activation that the digital constellation diagram itself can serve as a non-linear activation function within a split neural network, blending communication and computation into a single step.


Digital Semantic Communication: Quantizing Meaning

While AirComp focuses on aggregating multiple signals, Semantic Communication focuses on compressing a single data stream down to its core intent. Traditional deep-learning-based joint source-channel coding (JSCC) outputs continuous latent vectors meant for analog transmission. To deploy these codecs on standard binary protocols, researchers must discretize semantic spaces.

Continuous Semantic Features  --->  [ Vector Quantization ]  --->  Discrete Codebook Indices  --->  Standard Digital Transceiver

1. Vector Quantization and Codebooks

The primary tool for discretizing meaning is Vector Quantization (VQ), which maps continuous multi-scale features to discrete indices in a shared, pre-trained semantic codebook.

The authors of Vector Quantized Semantic Communication System (Qifan Fu et al.) introduced VQ-DeepSC, demonstrating that mapping semantic features to discrete codebooks yields higher image reconstruction quality and significantly better robustness to channel noise than traditional digital codecs like BPG in low-signal environments.

To scale this approach for high-fidelity generative tasks, Yingbin Zhou et al. developed MOC-RVQ in MOC-RVQ: Multilevel Codebook-assisted Digital Generative Semantic Communication. By combining a multi-head octonary codebook with a Swin Transformer denoiser, the system transmits high-fidelity images over standard digital frameworks using a fraction of the bandwidth required by state-of-the-art analog alternatives.

2. Importance-Aware and Adaptive Modulation

Not all semantic information carries equal weight. In Vision Transformer-based Semantic Communications With Importance-Aware Quantization, Joohyuk Park et al. introduced an importance-aware quantization policy for Vision Transformers. The system dynamically allocates more digital bits to critical semantic tokens (such as an object in the foreground) and fewer bits to background noise, maximizing transmission efficiency.

Instead of treating quantization and digital modulation as isolated steps, Guanding Yu et al. proposed a multi-order digital joint coding-modulation (MDJCM) framework in From Analog to Digital: Multi-Order Digital Joint Coding-Modulation for Semantic Communication. By using a unique substitution training strategy, the network learns to map semantic features directly onto variable-order digital constellations based on real-time channel conditions.

3. Generalized Interfaces and Task Alignment

To formalize the bridge between these two domains, Zhicheng Bao et al. designed sDAC in sDAC - Semantic Digital Analog Converter for Semantic Communications. This architecture serves as a generalized, bi-directional hardware interface that translates continuous semantic signals into discrete digital bits and back again, ensuring backward compatibility with existing infrastructure.

Finally, Changyang She et al. demonstrated the extreme efficiency of this paradigm in Aligning Task- and Reconstruction-Oriented Communications for Edge Intelligence. By combining an "Information Reshaper" with a learnable digital modulation scheme, their framework (ATROC) aligns data reconstruction with downstream task execution (such as autonomous driving safety), achieving a 99.19% reduction in data transmission over standard digital hardware.


Why It Works: Structural Alignment

The core intuition behind this research direction is that digitization does not have to mean a loss of structure.

If you blindly apply standard scalar quantization to a neural network's gradients or semantic features, you destroy the mathematical relationships between the values—the linearity required for AirComp addition vanishes, and the topological relationships required for semantic meaning are broken.

This direction succeeds because it enforces a structural alignment between the data domain and the hardware domain:

  1. In Semantic Communication: Vector quantization and learned constellations ensure that the geometric distance between discrete digital symbols mirrors the semantic distance between the original concepts. If noise alters a symbol slightly, it still decodes to a concept with a similar meaning.
  2. In Over-the-Air Computation: Frameworks like lattice coding ensure that the physical addition performed by the wireless channel maps directly to a valid, predictable point in a discrete digital codebook.

By treating the analog-to-digital transition as a structured geometric mapping problem rather than a lossy truncation step, these approaches unlock the efficiency of edge intelligence while maintaining full compatibility with the global digital communication infrastructure.

Key Foundational Work

One-Bit Over-the-Air Aggregation for Communication-Efficient Federated Edge Learning: Design and Convergence Analysis (2020) — contributed 1 concept elements to this direction.
Energy and Spectrum Efficient Federated Learning via High-Precision Over-the-Air Computation (2022) — contributed 1 concept elements to this direction.
One-Bit Over-the-Air Aggregation for Communication-Efficient Federated Edge Learning (2020) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_analog_digital_modulation_mismatch_9e2049fe,
  title = {analog-digital modulation mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-digital-modulation-mismatch/rdr_9e2049fe7c834f9daa1bab4b64ebfa0f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_analog_digital_modulation_mismatch_9e2049fe_v1,
  title = {analog-digital modulation mismatch},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/analog-digital-modulation-mismatch/rdr_9e2049fe7c834f9daa1bab4b64ebfa0f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake