Joint Active and Passive Beamforming Optimization

435 concepts398 papers5 proposals

Historically, wireless engineers treated the radio propagation environment as an uncontrollable black box. If a signal encountered a concrete wall and suffered severe fading, the transmitter had to compensate by pumping more power, adding more antennas, or employing complex error-correcting codes.

The advent of Reconfigurable Intelligent Surfaces (RIS) and movable antenna arrays has fundamentally shifted this paradigm. Instead of passively accepting the channel's limitations, modern wireless systems can dynamically reshape the physical environment itself. However, unlocking this capability requires solving a formidable mathematical challenge: joint active and passive beamforming optimization.

The Core Challenge: Bilinear Coupling

To understand why this problem is difficult, we must look at how active and passive components interact. Consider a base station transmitting a signal to a user. In a traditional system, the received signal depends entirely on the active beamforming vector $w$, which controls the amplitude and phase of the transmission at the source.

When we introduce an RIS—a surface covered with hundreds of passive reflecting elements—the signal travels along two paths: a direct path and a reflected path. Each element on the RIS can independently shift the phase of the incident electromagnetic wave. We can represent these passive phase shifts as a diagonal matrix $\Phi = \text{diag}(e^{j\theta_1}, e^{j\theta_2}, \dots, e^{j\theta_N})$.

The total effective channel seen by the receiver becomes the sum of the direct and reflected paths, and the received signal power is proportional to:

$$\left| (h_{\text{direct}} + h_{\text{reflected}}^H \Phi h_{\text{incident}}) w \right|^2$$

Notice the mathematical structure here: the passive phase-shift matrix $\Phi$ and the active beamforming vector $w$ are multiplied together. This creates a bilinear coupling. If you change the active beamforming direction $w$, the optimal configuration of the passive surface $\Phi$ changes. Conversely, if you adjust the surface reflections, the ideal active beamforming weights shift. Because these variables are deeply entangled, you cannot optimize them independently without sacrificing significant performance.

The Geometry of Non-Convex Constraints

The coupling alone makes the problem difficult, but hardware limitations introduce a second layer of complexity: non-convex constraints.

Because an RIS is passive, its elements cannot amplify the incoming signal; they can only delay it. This imposes a strict unit-modulus constraint on each reflecting element, meaning the amplitude must remain exactly equal to one ($|e^{j\theta_n}| = 1$).

In geometric terms, a convex optimization problem allows variables to exist anywhere within a solid, continuous space (like a solid disk). However, the unit-modulus constraint forces the optimization variables to lie strictly on the one-dimensional boundary of a complex circle. A straight line connecting two valid points on a circle passes through the interior, which violates the constraint.

As noted by Melike Erol‐Kantarci et al. in A Survey on Model-based, Heuristic, and Machine Learning Optimization Approaches in RIS-aided Wireless Networks, combining thousands of these non-convex circular constraints with the bilinear coupling of active beamforming transforms the resource allocation problem into an NP-hard puzzle filled with countless local optima and saddle points.

Breaking the Coupling via Alternating Optimization

To solve this joint optimization problem efficiently in real-time, researchers rely on algorithmic frameworks that systematically decouple the variables. The most prominent of these is Alternating Optimization (AO), also known as Block Coordinate Descent.

Instead of trying to move both the active and passive variables simultaneously through the non-convex landscape, the AO framework plays a game of mathematical leapfrog:

  1. Fix the Passive Surface ($\Phi$): By holding the RIS phase shifts constant, the effective channel becomes a known, fixed matrix. The problem of finding the optimal active beamforming vector $w$ reduces to a standard convex quadratic program or a semidefinite program, which can be solved optimally and rapidly using interior-point methods.
  2. Fix the Active Beamformer ($w$): Once the new active beamforming vector is found, it is held constant. The algorithm then optimizes the passive phase shifts $\Phi$. While the unit-modulus constraint is still non-convex, isolating $\Phi$ allows engineers to apply specialized techniques—such as fractional programming, complex manifold optimization, or semidefinite relaxations—to find a near-optimal surface configuration.

By iterating back and forth between these two steps, the algorithm guarantees monotonic convergence to a high-quality local optimum. Hong Shen et al. in Beamforming Design with Fast Convergence for IRS-Aided Full-Duplex Communication and Xidong Mu et al. in Exploiting Intelligent Reflecting Surfaces in NOMA Networks: Joint Beamforming Optimization demonstrate that this decoupling strategy dramatically reduces computational complexity, making it feasible to update configurations within the coherence time of the wireless channel.

Another powerful alternative is the Minorization-Maximization (MM) framework. As demonstrated by Zepeng Zhang et al. in Weighted Sum-Rate Maximization for Multi-Hop RIS-Aided Multi-User Communications: A Minorization-Maximization Approach, instead of tackling the complex coupled objective function directly, one can construct a simpler "surrogate" function that globally under-approximates the true objective and touches it at the current iteration point. Maximizing this surrogate function iteratively guarantees an improvement in the true network throughput while bypassing expensive matrix inversions.

Expanding the Optimization Space: Movable Antennas

The optimization landscape becomes even more intricate when the antennas themselves are allowed to move. Traditional antenna arrays fix elements at rigid, half-wavelength intervals. In contrast, Movable Antenna (MA) or "fluid" antenna systems allow antenna elements to slide continuously within a given region to actively avoid local fading nulls.

This adds a third dimension to the optimization problem. As explored by Xiao Tang et al. in Deep Learning-Assisted Jamming Mitigation with Movable Antenna Array and Yuqi Ye et al. in Fluid Antenna-Assisted MIMO Transmission Exploiting Statistical CSI, optimizing the physical coordinates of these antennas introduces complex exponential functions that depend directly on spatial positioning.

The system must now simultaneously synchronize three highly coupled domains:

  • The spatial positions of the antenna elements.
  • The digital precoding matrices at the transmitter.
  • The passive phase shifts of the surrounding environment.

When these movable arrays are mounted on mobile platforms like Unmanned Aerial Vehicles (UAVs), the authors of UAV-Enabled Wireless Networks with Movable-Antenna Array: Flexible Beamforming and Trajectory Design demonstrate that the optimization problem expands further to include the physical flight trajectory of the aircraft, creating a deeply entangled spatio-temporal optimization problem.

In practical deployments, optimization algorithms must adapt to non-ideal hardware. For example, building an RIS that can shift phases continuously is prohibitively expensive. Instead, real-world surfaces often utilize discrete phase shifts (e.g., 1-bit or 2-bit control), which restricts the phase angles to a finite set of values.

Boya Di et al. show in Hybrid Beamforming for Reconfigurable Intelligent Surface based Multi-user Communications: Achievable Rates with Limited Discrete Phase Shifts that this discretization transforms the continuous non-convex problem into a challenging combinatorial optimization problem.

Furthermore, hardware components often exhibit physical interdependencies. For instance, in Simultaneously Transmitting and Reflecting (STAR) surfaces, an element must split an incident wave into a transmitted component and a reflected component. Jiaqi Xu et al. reveal in STAR-RISs: A Correlated T&R Phase-Shift Model and Practical Phase-Shift Configuration Strategies that the transmission and reflection coefficients are physically correlated rather than independent, adding strict structural constraints to the optimization matrix.

Multi-Functional Systems and Deep Learning

Despite these mathematical hurdles, mastering joint active and passive optimization unlocks massive performance gains across diverse applications. In Integrated Sensing and Communication (ISAC) systems, Xianxin Song et al. in Joint transmit and reflective beamforming for IRS-assisted integrated sensing and communication and Yue Xiu et al. in CRB Minimization for RIS-Aided mmWave Integrated Sensing and Communications demonstrate that joint beamforming allows a single transmitted wave to simultaneously maximize data throughput for communication users and minimize radar sensing errors (the Cramér-Rao Bound) for target tracking.

To bypass the heavy computational overhead of running iterative matrix algorithms in highly dynamic environments, researchers are increasingly turning to data-driven approaches.

Xinyu Gao et al. in Resource Allocation in IRSs Aided MISO-NOMA Networks: A Machine Learning Approach and Helin Yang et al. in Intelligent Reflecting Surface Assisted Anti-Jamming Communications Based on Reinforcement Learning demonstrate that Deep Reinforcement Learning (DRL) can implicitly learn the underlying structure of these coupled spaces. By training neural networks to map raw channel state information directly to near-optimal active and passive configurations, the system can bypass iterative optimization entirely, enabling real-time adaptation to user mobility, hardware constraints, and malicious jamming threats.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_non_convex_resource_allocation_36b6439f,
  title = {non-convex resource allocation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-convex-resource-allocation/rdr_36b6439f9abd42058045217001c92720},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_non_convex_resource_allocation_36b6439f_v1,
  title = {non-convex resource allocation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-convex-resource-allocation/rdr_36b6439f9abd42058045217001c92720/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake