In the quest for 6G and beyond, wireless engineers are moving away from simply building more powerful transmitters and toward "programming" the electromagnetic environment itself. This shift is driven by two key technologies: Intelligent Reflecting Surfaces (IRS)—arrays of passive elements that steer signals—and Hybrid Beamforming, which uses a small number of expensive radio chains to control a large number of cheap antenna elements.
While these technologies promise massive gains in energy efficiency and data rates, they share a common, stubborn mathematical hurdle: the unit-modulus constraint.
The Hardware Reality: Phase Without Power
In an ideal digital system, a transmitter can change both the amplitude (strength) and the phase (timing) of a signal for every single antenna. However, doing this at the scale of hundreds of antennas is prohibitively expensive and power-hungry.
To keep costs down, hardware designers use analog phase shifters. These components can delay a signal, but they cannot easily change its power level. Mathematically, if $x$ is the weight applied to an antenna or reflecting element, it must satisfy $|x| = 1$. This means the weight can live anywhere on the unit circle in the complex plane, but never inside it.
As Wei Sun, Qingyang Song, and Lei Guo point out in Secrecy Rate Maximization for Intelligent Reflecting Surface Aided SWIPT Systems, this constraint is inherently non-convex. In optimization, "convex" problems are like bowls where any ball dropped will roll to the bottom (the global optimum). "Non-convex" problems are like rugged mountain ranges where you can easily get stuck in a local valley, far from the highest peak.
Why Standard Tools Fail
Most traditional wireless optimization tools assume a "convex" search space. If you want to minimize power, you look inside a disk of possible signal strengths. But the unit-modulus constraint forces the solution to stay exactly on the rim of that disk.
When Yang Liu, Ming Li, and Qingqing Wu attempted to minimize uplink power in Intelligent Reflecting Surface Aided MISO Uplink Communication Network, they found that standard linear tools couldn't handle this "rim-only" requirement. If you ignore the constraint and then "round" the result back to the unit circle later, you often end up with a configuration that performs poorly or fails to meet quality-of-service targets.
Strategies for Navigating the Unit Circle
Researchers have developed three primary "intuitions" for solving these problems:
1. Relax and Penalize
One common approach is to temporarily allow the weights to live inside the circle (a convex relaxation) and then gradually "punish" the algorithm if the weights don't move toward the edge. Zheng Zhang uses this in Securing NOMA Networks by Exploiting Intelligent Reflecting Surface, relaxing the feasible set into a ring and using a penalty term to drive the solution to the boundary. Similarly, Weiheng Jiang and Zhiguo Ding apply Successive Convex Approximation (SCA), which solves a sequence of simpler, convex "maps" of the rugged terrain to find a path to the peak.
2. Manifold Optimization: Walking on the Circle
Instead of treating the unit circle as a constraint on a flat plane, Manifold Optimization treats the circle as the entire world. In IRS-Assisted Millimeter Wave Communications, Yue Xiu and colleagues use Riemannian manifold tools to "walk" along the surface of these complex constraints. This ensures that every step the algorithm takes is physically realizable by the hardware, leading to faster convergence and better performance in high-frequency mmWave systems.
3. Alternating Minimization
Since the system involves both a transmitter (active) and a surface (passive), the variables are "coupled." Xianghao Yu and Khaled Ben Letaief pioneered Alternating Minimization (AltMin) for hybrid precoding in Alternating Minimization Algorithms for Hybrid Precoding in Millimeter Wave MIMO Systems. The intuition is simple: fix the surface and optimize the transmitter; then fix the transmitter and optimize the surface. By iterating back and forth, the complex joint problem becomes a series of manageable sub-problems.
The Deep Learning Shift
As networks become more complex—integrating sensing, power transfer, and communication—traditional math becomes too slow. Researchers like Foad Sohrabi and Wei Yu are now using neural networks to "learn" the optimal phase shifts.
In Deep Active Learning Approach to Adaptive Beamforming, they incorporate the unit-modulus constraint directly into the neural network's architecture. By using specialized normalization layers or Euler-mapped activation functions (as seen in Near-field Beamforming for Extremely Large-scale MIMO), the AI is forced to only "think" in terms of phase shifts, ensuring its suggestions are always hardware-compatible.
Conclusion
The research direction across these 117 papers is a move toward hardware-aware optimization. Whether through the geometric elegance of manifold optimization or the brute-force speed of deep learning, the goal is to bridge the gap between ideal mathematical models and the "phase-only" reality of the physical world. By mastering the unit-modulus constraint, researchers are turning what was once a hardware limitation into a powerful tool for shaping the future of wireless connectivity.