In modern wireless communications and remote sensing, hardware efficiency dictates mathematical form. As networks transition to millimeter-wave (mmWave) frequencies and embrace dense antenna arrays, engineers face a strict physical reality: giving every single antenna element its own dedicated, power-hungry radio frequency (RF) chain is prohibitively expensive and energy-inefficient.
To circumvent this, modern architectures rely on hardware components that manipulate signals passively or semi-passively. In hybrid analog-digital beamforming and Reconfigurable Intelligent Surfaces (RIS), components like analog phase shifters can only alter the phase of an electromagnetic wave, leaving its amplitude completely untouched.
Mathematically, this hardware limitation introduces a unit-modulus constraint (also known as a constant-modulus or unimodular constraint). Every optimization variable $x_i$ representing a phase shift or a transmitted signal coefficient must lie exactly on the complex unit circle:
$$|x_i| = 1$$
While this constraint enables low-cost, power-efficient hardware, it introduces severe non-convexity into the underlying signal processing problems, rendering standard optimization toolkits useless. A major thrust of recent research focuses on developing algorithmic frameworks to navigate this constrained space.
Why the Unit Circle Breaks Convexity
To understand why this constraint is so challenging, consider the geometry of a convex set. A set is convex if, given any two valid points within it, a straight line drawn between them stays entirely inside the set.
The complex unit circle violates this definition fundamentally. If you take two valid phase configurations—say, $1$ and $-1$ on the complex plane—and average them, the result is $0$. However, $0$ does not have a modulus of $1$; it lies in the hollow center of the circle. The feasible region is not a solid disc, but a thin, rigid, one-dimensional boundary.
Non-Convex Feasible Set (Unit Circle)
* * *
* | *
* | *
*--------o--------* <-- Averaging 1 and -1 lands at 0,
* (0,0) * which is outside the set!
* *
* * *
As a result, standard optimization techniques like linear programming or unconstrained gradient descent fail. If an algorithm attempts to update a phase shift, the step will almost certainly push the variable off the unit circle. This geometric barrier turns standard engineering objectives—such as maximizing data rates in Reconfigurable Intelligent Surfaces for Energy Efficiency in Wireless Communication or minimizing radar sidelobes in Sequence Design to Minimize the Weighted Integrated and Peak Sidelobe Levels—into highly non-convex, NP-hard problems.
Algorithmic Paradigms for Phase Optimization
To optimize over this hollow landscape, researchers have developed four core algorithmic strategies. Each approaches the non-convex boundary from a different mathematical perspective.
1. Alternating Optimization: Decoupling the Variables
In many wireless systems, the overall problem involves a mix of unconstrained variables (such as digital precoding matrices at the base station) and unit-modulus variables (such as analog phase shifters). While optimizing them jointly is intractable, optimizing one while fixing the other is often straightforward.
In Alternating Minimization Algorithms for Hybrid Precoding in Millimeter Wave MIMO Systems, Xianghao Yu et al. demonstrated that by freezing the analog phase shifts, the digital precoder can be solved using standard convex quadratic programming. Conversely, when the digital precoder is frozen, the optimal phase shifts often decouple into element-by-element operations, where the algorithm simply extracts the angle of the complex numbers. By iterating back and forth, the algorithm converges to a high-quality local optimum. To scale this to massive surfaces, Yifan Ma et al. combined alternating optimization with a tailored gradient descent step size, bypassing expensive matrix inversions.
2. Majorization-Minimization: Surrounding the Obstacle
Instead of navigating the rugged, non-convex objective function directly on the unit circle, the Majorization-Minimization (MM) framework constructs a smooth, convex surrogate function (a majorant) that sits entirely above the true objective function and touches it at the current iteration point.
Objective Value
^
| ,---. <-- Convex Surrogate (Easy to optimize)
| / \
| / .---*---. <-- True Non-Convex Objective
| / / \ \
+---|---|-------|---|---> Variable Space
x_t x_t+1
Because the surrogate is convex, finding its global minimum is mathematically trivial. The algorithm moves to this minimum, constructs a new surrogate, and repeats. Because the surrogate always upper-bounds the true objective, every step is guaranteed to improve the true objective value. Junxiao Song et al. applied this to radar sequence design, proving that pairing MM with Fast Fourier Transforms (FFTs) allows systems to optimize sequences with lengths exceeding $10^4$ elements while strictly maintaining a constant signal amplitude.
3. Riemannian Manifold Optimization: Walking the Circle
Rather than treating the unit-modulus constraint as a barrier embedded within flat complex space, Manifold Optimization redefines the problem entirely. It treats the collection of unit-modulus elements as a smooth, curved geometric surface—a Riemannian manifold.
Instead of moving in a straight line and stepping off the circle, gradient steps are computed within the tangent space of the manifold. The algorithm takes a step along this flat tangent plane and then uses a mathematical operation called a retraction to map that point back down onto the nearest point on the unit circle. Tian Lin et al. utilized this approach to design hybrid beamformers under the Mean Square Error criterion, demonstrating that treating the unit circle as a native geometry yields performance that closely matches ideal, unconstrained digital systems.
4. Penalty Dual Decomposition: Relaxing and Penalizing
When phase shifts are coupled together by hardware constraints—such as in Simultaneously Transmitting and Reflecting (STAR) surfaces—the optimization space becomes even more restricted. In these scenarios, researchers decouple the constraints by introducing auxiliary variables, transforming the hard unit-modulus constraint into a flexible equality constraint.
Using frameworks like Penalty Dual Decomposition (PDD) or the Augmented Lagrangian Method (ALM), the algorithm initially allows the variables to step off the unit circle to explore the optimization landscape. However, it applies an increasingly severe mathematical penalty to the objective function as the variables deviate from a modulus of 1. Mingyi Hong 0001 et al. used this dual decomposition approach to directly optimize spectral efficiency in multi-user mmWave systems, while Zhaolin Wang et al. applied it to guarantee convergence to stationary points in surfaces with hardware-coupled phase shifts.
Real-Time Execution via Deep Unfolding
While iterative algorithms like MM, PDD, and manifold optimization provide precise solutions, they require hundreds of iterations to reach convergence. In mobile wireless networks, the physical environment changes on the order of milliseconds. By the time an iterative algorithm calculates the optimal phase shifts for an RIS, the user has already moved, rendering the solution obsolete.
To bridge this gap, researchers are replacing traditional iterative loops with deep learning architectures that can execute in a single forward pass.
A prominent approach is deep unfolding. Instead of treating a neural network as a generic black box, deep unfolding maps the individual steps of a proven iterative algorithm directly onto the layers of a neural network. Nir Shlezinger applied this to Joint Communications and Sensing (JCAS) systems, designing a network where the final layer enforces a hard normalization step:
$$x_{\text{final}} = \frac{x}{|x|}$$
This architecture guarantees that the network's output always satisfies the unit-modulus constraint. By training the internal weights of the network to mimic optimal step sizes, the unfolded model achieves near-optimal beamforming performance while reducing computational latency by over 60% compared to traditional solvers. This transition from slow offline optimization to real-time execution makes the deployment of unit-modulus hardware architectures viable in next-generation wireless networks.