As wireless networks transition toward 6G, the number of controllable elements in the environment is exploding. Massive MIMO arrays now feature hundreds of antennas, and Reconfigurable Intelligent Surfaces (RIS) add thousands of passive reflecting elements. While these technologies promise unprecedented spectral efficiency, they encounter a "complexity wall": traditional optimization methods, such as Semidefinite Programming (SDP) or interior-point solvers, often scale cubically ($O(N^3)$) or worse with the number of elements. In a dynamic wireless channel that changes every few milliseconds, these algorithms are simply too slow to be practical.
To bridge this gap, researchers are moving away from general-purpose numerical solvers toward structural optimization. By exploiting the specific mathematical and physical properties of wireless channels, this direction develops algorithms that achieve near-optimal performance with linear or quadratic complexity.
Bypassing Solvers with Closed-Form Solutions
The most direct way to eliminate computational lag is to replace iterative loops with direct mathematical expressions. When an optimization problem is framed correctly, the optimal "shape" of the beamforming vector can often be derived analytically using Lagrangian duality or Karush-Kuhn-Tucker (KKT) optimality conditions.
For instance, Yuanming Shi and Khaled B. Letaief demonstrated that deriving closed-form solutions for power minimization in Cloud-RANs allows for real-time adjustments that generic solvers cannot match. Similarly, in the context of massive MIMO, Guangda Zang and Ying Cui proved that hybrid beamforming can match the efficiency of fully digital systems by using semi-closed-form updates, bypassing the need for expensive semidefinite relaxation.
The intuition here is that wireless problems often have a "hidden" structure—such as a Trace Quotient Formulation—that allows the optimal solution to be expressed as a generalized eigenvector or a simple projection, rather than the result of a high-dimensional search.
From Second-Order to First-Order Methods
Traditional optimization often relies on second-order information (like the Hessian matrix) to find the fastest path to a solution. However, calculating and inverting these matrices becomes the primary bottleneck as the number of antennas grows. A major trend in this field is the shift toward first-order methods, which use only gradient information.
Yifan Ma and Xianghao Yu introduced the AO-GD framework, which combines alternating optimization with gradient descent. They showed that for large-scale RIS, simple gradient updates can achieve the same performance as complex manifold optimization but at a fraction of the cost. This is further echoed by Chong Zhang and Min Dong, whose "ultra-low-complexity" algorithms (ESCA and ASCA) leverage the optimal structure of beamforming vectors so that the complexity no longer scales with the number of antennas at all.
Element-Wise and Greedy Refinement
When dealing with thousands of passive elements on an RIS, optimizing them all simultaneously is computationally prohibitive. Instead, researchers use element-wise block coordinate descent. This approach updates one reflecting element at a time while keeping others fixed.
Yue Xiu and Zhongpei Zhang applied this to secure millimeter-wave systems, showing that sequential updates significantly reduce the search space. To scale even further, Marzieh Najafi proposed a "tiling" strategy, where elements are grouped into physical tiles and optimized as units. This greedy, tile-by-tile approach makes the configuration of massive surfaces—essential for overcoming high-frequency path loss—computationally feasible for the first time.
Exploiting Geometric and Statistical Sparsity
Finally, the research direction leverages the fact that wireless channels aren't random; they have physical and statistical boundaries.
- Statistical Optimization: Instead of re-calculating beams for every millisecond-long "instantaneous" channel change, researchers like Sadaf Syed and Wolfgang Utschick use long-term channel statistics. This reduces the frequency of optimization by orders of magnitude while maintaining near-optimal throughput.
- Geometric Search: For discrete phase shifts (where an RIS can only choose from a few set angles), the problem is usually seen as an exponential combinatorial search. However, Kaiming Shen and Shuyi Ren shattered this assumption by reframing the search as a geometric problem on a unit circle, achieving the global optimum in linear time ($O(N)$).
By moving the intelligence from generic "black box" solvers into the mathematical formulation of the wireless problem itself, this research direction enables the massive scale required for the next generation of connectivity.