Alternating Optimization for Joint Active and Passive Beamforming

823 concepts730 papers19 proposals

In next-generation wireless networks, maximizing data rates, minimizing energy consumption, and eliminating coverage blind spots require the coordination of multiple independent subsystems. For instance, a base station must precisely shape its wireless signals using active beamforming, while simultaneously, modern hardware like Reconfigurable Intelligent Surfaces (RIS) or Intelligent Reflecting Surfaces (IRS) must alter their electromagnetic properties via passive beamforming to redirect those signals to a user.

When these systems are deployed together, engineers face a daunting mathematical challenge: the variables controlling the transmitter and the variables controlling the reflecting surface are deeply intertwined. Optimizing them simultaneously results in highly non-convex, computationally intractable problems.

To overcome this deadlock, researchers rely on Alternating Optimization (AO)—also known as Block Coordinate Descent (BCD). This design pattern breaks a single intractable joint optimization problem into a sequence of simpler, isolated subproblems that iterate until they reach a stable, high-performance equilibrium.

The Curse of Coupled Variables

To understand why alternating optimization is necessary, consider a multi-user system assisted by an IRS. The base station controls an active beamforming vector $\mathbf{w}$, which dictates the direction and power of the transmitted signal. The IRS consists of an array of passive elements, each introducing a phase shift $\theta_n$ to the incoming waves, collected into a diagonal matrix $\mathbf{\Theta}$.

The effective channel matrix $\mathbf{H}_{\text{eff}}$ experienced by a receiver is the sum of the direct path from the base station and the reflected path via the IRS:

$$\mathbf{H}_{\text{eff}} = \mathbf{H}_{\text{direct}} + \mathbf{H}_{\text{IRS-user}} \mathbf{\Theta} \mathbf{H}_{\text{base-IRS}}$$

The ultimate goal is typically to maximize a metric like the weighted sum-rate or signal-to-interference-plus-noise ratio (SINR), both of which depend on the term $|\mathbf{H}_{\text{eff}} \mathbf{w}|^2$.

Expanding this term reveals a fundamental mathematical hurdle: the expression contains cross-products of the active beamforming vector $\mathbf{w}$ and the passive phase-shift matrix $\mathbf{\Theta}$ (i.e., terms of the form $\mathbf{\Theta}\mathbf{w}$). This is a bilinear coupling. Because the variables multiply each other, the optimization landscape becomes highly non-convex, plagued by an astronomical number of local minima and saddle points. Standard convex optimization tools cannot solve for both variables at the same time.

The Intuition of Coordinate Descent

Alternating optimization resolves this coupling by exploiting a simple geometric insight: while navigating a rugged, multi-dimensional valley diagonally might be impossible, you can still find the bottom by alternating between purely North-South and East-West steps.

In the context of wireless beamforming, while the objective function is globally non-convex when optimizing $\mathbf{w}$ and $\mathbf{\Theta}$ together, it often becomes convex—or at least much easier to solve—when one of the variables is held constant.

┌────────────────────────────────────────────────────────┐
│             Joint Non-Convex Problem                   │
│               Maximize f(w, Θ)                         │
└───────────────────────────┬────────────────────────────┘
                            │
              Decompose via Alternation
                            │
                            ▼
┌────────────────────────────────────────────────────────┐
│  Subproblem 1: Fix Θ(t)                                │
│  Optimize active beamforming w(t+1)                    │
│  (Convex/Closed-form multi-user MISO problem)          │
└───────────────────────────┬────────────────────────────┘
                            │
                            ▼
┌────────────────────────────────────────────────────────┐
│  Subproblem 2: Fix w(t+1)                              │
│  Optimize passive phase shifts Θ(t+1)                  │
│  (Unimodular quadratic program via MM or SDR)          │
└───────────────────────────┴────────────────────────────┘

The algorithm proceeds in an iterative loop:

  1. Fix the Passive Phase Shifts ($\mathbf{\Theta}^{(t)}$): The IRS is treated as a static, mirror-like environment. The problem collapses into a classic active multi-user beamforming problem. Researchers can then deploy well-established convex optimization techniques, fractional programming, or closed-form updates to find the optimal $\mathbf{w}^{(t+1)}$.
  2. Fix the Active Beamforming ($\mathbf{\ w}^{(t+1)}$): The transmitter's signals are treated as fixed constants. The problem reduces to optimizing the phase shifts $\mathbf{\Theta}^{(t+1)}$ to ensure that the reflected waves arrive at the user in perfect phase alignment with the direct path. This subproblem can be solved using specialized signal processing techniques like Semidefinite Relaxation (SDR) or Majorization-Minimization (MM).

By continuously swapping which variable is frozen and which is fluid, the system incrementally climbs the performance landscape without ever needing to compute the intractable joint gradient.

Decomposing Complex Wireless Architectures

This decoupling strategy is remarkably versatile and serves as the mathematical backbone across several distinct frontiers in wireless communication.

Active vs. Passive Coordination

In distributed IRS networks designed to eliminate coverage blind spots, Yashuai Cao et al. utilize an alternating iterative framework to decouple active base station beamforming from the passive phase-shift adjustments of multiple distributed surfaces. Similarly, Xiao 潇 Liu 刘 et al. note in their survey on reconfigurable intelligent surfaces that this non-convex coordination is almost universally managed via alternating optimization or semidefinite relaxation to make 6G software-defined environments practical (Reconfigurable Intelligent Surfaces: Principles and Opportunities).

Even when feedback is constrained, the paradigm holds. Xiaochun Ge demonstrates in Beamforming Design with Partial Channel Estimation and Feedback for FDD RIS-Assisted Systems that alternating optimization can successfully refine active and passive components sequentially based on quantized, low-overhead feedback from the receiver.

Hybrid Digital and Analog Beamforming

The utility of alternating optimization extends beyond reflecting surfaces to the internal architecture of massive MIMO base stations. In hybrid beamforming, hardware constraints dictate that a large antenna array must be driven by a strictly limited number of radio frequency (RF) chains. This requires splitting the precoder into a digital stage (managing power and multi-user interference) and an analog stage (built from low-cost phase shifters).

Foad Sohrabi et al. show in Hybrid Digital and Analog Beamforming Design for Large-Scale Antenna Arrays that an iterative coordinate descent framework can decouple these stages. By tuning the analog phase shifters one by one while keeping the digital stage fixed, they prove that hybrid architectures can exactly replicate fully digital performance using a fraction of the hardware.

Movable and Fluid Antennas

A newer paradigm involves physically moving or sliding antennas within a small spatial region to optimize the local channel conditions. Here, the coupling occurs between the transmit power allocation and the physical spatial coordinates of the antennas.

To prevent an exponential explosion in processing requirements, Yuqi Ye et al. employ an alternating optimization framework in Fluid Antenna-Assisted MIMO Transmission Exploiting Statistical CSI to update the transmit signal covariance and the physical coordinates of the antennas in separate, computationally efficient sub-steps. This approach is mirrored by Wen Chen 0001 et al. in Movable Antenna Empowered Downlink NOMA Systems: Power Allocation and Antenna Position Optimization, where fixing the power allocation allows for the isolated, gradient-based refinement of 2D antenna positions.

Why Alternating Optimization Guarantees Convergence

The widespread adoption of alternating optimization is driven by a critical mathematical property: monotonic convergence.

Let $f(\mathbf{w}, \mathbf{\Theta})$ represent the objective function we wish to maximize (such as the system's sum-rate). During any given iteration $t$, the algorithm performs two distinct updates:

  1. It finds $\mathbf{w}^{(t+1)}$ by maximizing $f(\mathbf{w}, \mathbf{\Theta}^{(t)})$. Because $\mathbf{w}^{(t+1)}$ is the optimal solution for this subproblem, it is guaranteed that: $$f(\mathbf{w}^{(t+1)}, \mathbf{\Theta}^{(t)}) \geq f(\mathbf{w}^{(t)}, \mathbf{\Theta}^{(t)})$$
  2. It finds $\mathbf{\Theta}^{(t+1)}$ by maximizing $f(\mathbf{w}^{(t+1)}, \mathbf{\Theta})$. Because $\mathbf{\Theta}^{(t+1)}$ optimizes this step, it is guaranteed that: $$f(\mathbf{w}^{(t+1)}, \mathbf{\Theta}^{(t+1)}) \geq f(\mathbf{w}^{(t+1)}, \mathbf{\Theta}^{(t)})$$

Combining these two inequalities yields:

$$f(\mathbf{w}^{(t+1)}, \mathbf{\Theta}^{(t+1)}) \geq f(\mathbf{w}^{(t)}, \mathbf{\Theta}^{(t)})$$

This demonstrates that the objective value is strictly non-decreasing after every full iteration. Because physical wireless systems operate under strict resource constraints—such as finite transmit power and bounded channel capacities—the objective function $f$ is upper-bounded by a finite value. By the Monotone Convergence Theorem, any sequence that is monotonically increasing and bounded from above must converge to a stable limit point.

This reliable convergence allows researchers to confidently layer alternating optimization on top of highly complex, multi-objective wireless networks. For example, Meng Hua et al. embed dual subgradient methods and Majorization-Minimization within an alternating loop to ensure stable convergence in multi-cell coordinated networks (Intelligent Reflecting Surface-Aided Joint Processing Coordinated Multipoint Transmission).

Similarly, Chong Zhang et al. leverage this convergence pattern in Joint Downlink-Uplink Beamforming for Wireless Multi-Antenna Federated Learning, alternating between transmit beamforming, receive beamforming, and power scaling to coordinate edge-AI model aggregation without risking algorithmic instability.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_alternating_optimization_3356824c,
  title = {alternating optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alternating-optimization/rdr_3356824c3357496a99fb20e5fba8ba7c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_alternating_optimization_3356824c_v1,
  title = {alternating optimization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/alternating-optimization/rdr_3356824c3357496a99fb20e5fba8ba7c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake