A proposed beamforming method for wireless systems using an active intelligent reflecting surface, or IRS, was reported to deliver a higher achievable sum rate than a WMMSE comparison method while running about 260 times faster in simulations. The work appears as arXiv version 1, dated 20 August 2026, and its results come from a modeled system rather than hardware or field measurements.
The study examined whether real interference alignment could meet individual rate requirements and power limits in a multi-user channel whose direct transmitter-to-receiver links are blocked. In this setting, the active IRS shapes the modeled signal paths between users while also introducing its own noise and transmission-power constraints.
Turning interference into something receivers can separate
The proposed scheme uses real interference alignment, a signal-design technique that arranges unwanted transmissions so they can be separated from the desired signal. Its construction sets nonmatching real interference coefficients to zero, aligns the remaining interference in the imaginary part of the received signal, and recovers the desired real symbol from the received real part.
The method combines that beamforming design with an offline rate-profile learning algorithm. The researchers first use WMMSE on training channels to obtain solutions for the constrained optimization problem, then sort the resulting achievable rate profiles and use them to break the original task into multiple feasibility checks. Those checks are solved with generalized eigenvalue decomposition, a matrix calculation used here to test whether the rate and power requirements can be met.
The paper reports an overall computational complexity of O(d³M), with generalized eigenvalue decomposition dominating the feasibility-check calculation. It also uses a proposition that at most one total or per-element transmission-power constraint is active at the solution, an assumption built into the feasibility derivation.
What was tested
The simulated environment placed nodes uniformly within a disk with a 100-metre radius, kept them at least 20 metres apart and positioned the reflecting surface 15 metres above the ground at the disk’s centre. The channels used independent Rician fading with a factor of 5 and a path-loss exponent of 2.2.
The model assigned noise powers of −90 dBm at both the active reflecting surface and the receivers. Each transmitter had 20 dBm of power, while each user pair had to meet a minimum achievable rate of 1.5 bps/Hz; the per-element IRS transmission-power constraint was 15 dBm.
For the offline learning stage, the paper used 60 training channels. It compared the proposed learning-based method with interference alignment against WMMSE with interference alignment and against a system without interference alignment. The execution-time comparison used six transmitter-receiver pairs and an IRS with 16 reflecting elements.
A reported gain, with important boundaries
In the reported simulations, the proposed interference-alignment system performed better than the system without interference alignment when the IRS transmission power was higher than 9 dBm. The learning-based algorithm also achieved a higher sum rate than WMMSE while using significantly less execution time, with the paper describing the speed advantage as about 260-fold.
The paper interprets the simulations as showing that interference alignment can outperform the no-alignment system above a sufficiently high IRS power level, and that its learning-based method offers a better rate-and-runtime result than WMMSE under the tested constraints. That interpretation remains tied to the simulated setting and the particular comparisons reported.
The available report gives no exact numerical difference in sum rate, variance, confidence interval, inferential statistical test or independent simulation count. The result is therefore a directional comparison rather than a quantified performance distribution, and the size and consistency of the rate advantage cannot be assessed from the reported evidence.
Why the result is not yet a network test
This is a modeling study based on synthetic Rician channel realizations. The supplied analysis reports no real-world hardware or over-the-air validation, and it does not show that the reported rate gain or speedup will generalize beyond the simulated conditions.
The assumptions narrow the scope of the finding. The direct links between transmitters and receivers are blocked, and the IRS is assumed to know the instantaneous channel coefficients. The paper therefore does not establish how the method would perform when direct links are available or when that channel knowledge is imperfect or unavailable.
The proposed procedure is described as a suboptimal solution, and the analysis does not establish global optimality. The feasibility proposition is based on a probability argument for continuous random channels, so it does not guarantee the same behavior for every deterministic channel realization.
The design of the learning stage also leaves unanswered questions. The supplied analysis does not explain why 60 training channels were chosen or how the training channels relate to the evaluation channels. Because WMMSE is used to generate the offline training solutions and is also one of the comparison methods, the material provided does not settle questions about fairness or generalization.
The next test is outside the simulation
Open questions include whether the approach keeps its reported rate and runtime advantages under other channel models and system sizes, how it compares with stronger or multi-start WMMSE procedures, and how close its suboptimal solution comes to a global optimum. The supplied analysis also calls for hardware measurements to test whether the simulated rate gains and speedup can be reproduced.
Active operation also brings noise and power constraints into the modeled problem. The analysis leaves the energy and implementation trade-offs of those constraints unresolved, even though they are part of the system the method is designed to optimize.
Paper data and sources
Original title: Real Interference Alignment for Active IRS-Aided Systems: A Rate-Profile Learning-Based Approach
Authors: Junda Liao, Quanzhong Li, Qi Zhang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text