Most physical signals that define our world—light, sound, and radio waves—are inherently complex-valued. They are characterized not just by their intensity (amplitude), but also by their timing relative to a reference (phase). While standard deep learning has achieved remarkable success using real-valued arithmetic, a growing body of research suggests that treating these signals as pairs of independent real numbers discards the fundamental physics of wave propagation.
Complex-valued neural networks (CVNNs) address this by performing arithmetic directly in the complex domain. By preserving the mathematical coupling between amplitude and phase, these architectures allow models to "understand" interference, diffraction, and scattering in ways that real-valued networks cannot.
The Phase Information Gap
In traditional deep learning, complex signals are typically handled by splitting them into two separate real-valued channels: the In-phase ($I$) and Quadrature ($Q$) components. However, as Tianyue Zheng et al. argue in Enhancing RF Sensing with Deep Learning: A Layered Approach, standard libraries optimized for real numbers fail to capture the underlying physics of signal propagation.
When a network treats $I$ and $Q$ as independent, it loses the "circularity" of the signal. In tasks like wireless localization or gesture recognition, the phase of a signal is often more critical than its amplitude. Guoxuan Chi et al. note in RF-Diffusion: Radio Signal Generation via Time-Frequency Diffusion that discarding these phase relationships leads to synthetic data that lacks physical realism, as phase offsets represent the precise relative timing of signal samples.
Preserving Physics through Complex Arithmetic
The core mechanism of a CVNN is the complex-valued convolution. Unlike a standard convolution that simply sums products, a complex convolution simulates complex multiplication: $(a + bi)(c + di) = (ac - bd) + (ad + bc)i$.
This operation ensures that the real and imaginary components remain mathematically coupled throughout the network. Yanxin Hu et al. demonstrate in DCCRN: Deep Complex Convolution Recurrent Network for Phase-Aware Speech Enhancement that this approach respects the physical properties of wave-based data, enabling more precise phase estimation and signal reconstruction. Similarly, in the context of optical imaging, Atreyee Saha et al. show in LWGNet - Learned Wirtinger Gradients for Fourier Ptychographic Phase Retrieval that treating amplitude and phase as a single complex entity allows the network to more accurately model the physical interactions of light waves.
Why It Works: Interference and Geometry
The advantage of CVNNs is most visible in environments where waves interfere. In multi-user wireless networks, managing constructive and destructive interference is the primary challenge for beamforming. Arindam Chowdhury et al. found in Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks that explicitly modeling the complex-valued nature of signals captures the physical properties of radio frequency propagation more effectively than models that ignore phase.
This physical alignment also leads to significant efficiency gains:
- Semantic Differentiation: In radar perception, Yizhou Wang 0005 et al. show in Vision meets mmWave Radar: 3D Object Perception Benchmark for Autonomous Driving that phase information allows systems to distinguish between objects with similar radar cross-sections, such as pedestrians and metallic poles.
- Data Efficiency: Rafael Boloix-Tortosa argues in Complex Gaussian Processes for Regression that native complex-valued frameworks can achieve optimal performance with a fraction of the training data required by real-valued methods.
- Model Size: By exploiting the geometric structure of the complex plane, Rudrasis Chakraborty et al. developed the C-SURE architecture, which is up to 100x smaller than standard ResNets while maintaining higher accuracy on signal processing tasks.
Beyond Traditional Sensing
The utility of CVNNs extends to any domain where periodic or wave-like dependencies exist. Yanfang Xue et al. demonstrate in CosNet: A Generalized Spectral Kernel Network that retaining the imaginary parts of spectral kernels allows for better modeling of time-series data. Even in environmental science, the authors of CoNOAir use complex-valued operators to capture the intricate dynamics of gas evolution in urban air quality forecasting.
By aligning the mathematical operations of the neural network with the complex-valued nature of the physical world, researchers are moving away from treating signals as mere data points and toward architectures that inherently respect the laws of wave physics.
Key Foundational Work
A FULLY COMPLEX-VALUED RADIAL BASIS FUNCTION NETWORK AND ITS LEARNING ALGORITHM (2009) — contributed 6 concept elements to this direction.