A wireless method reported in a new preprint identified the active set of reconfigurable intelligent surface (RIS) panels, meaning those classified as unblocked, with more than 95% probability and virtually no false positives in ray-tracing simulations. Under 30% per-link blockage, its weighted sum rate (WSR), a score combining users' rates, was consistently higher than that of baseline methods across the tested speeds and configurations.
A panel-by-panel blockage check
The modeled downlink is an mmWave multi-user MIMO system, meaning it serves several users at once, with M distributed passive RIS panels serving K user equipment. The panels provide RIS-assisted reflections in the simulated channel, while the detector treats each panel separately. It gives each panel an indexed synchronization sequence, then runs a correlation and energy test at each index to classify the panel as blocked or unblocked and produce the active-RIS set.
Optimization follows the blockage decision
That active set feeds the paper's optimization method, called SCRPA. It uses a sample-average approximation, representing the stochastic channel calculation with a finite set of channel samples, and block-coordinate updates with closed-form steps for both the base-station precoder, which shapes the transmitted signals, and the RIS phases, which set each panel's response. In ordinary terms, the procedure updates the transmitter and the panel settings in alternating blocks while accounting for aged or imperfect channel information.
The convergence analysis is narrower than the headline result may suggest. For continuous phases, the block-MM updates monotonically decrease the fixed-ensemble sample-average objective and converge to a block-stationary point. Finite-resolution updates converge to a discrete block-stationary point. Those are convergence statements for the specified optimization problem, not a guarantee of a global optimum.
A controlled simulation test
The authors generated numerical data by ray tracing in the Munich urban scene provided by Sionna. The dataset included mobile UEs, RIS-link blockages, aged or imperfect channel-state information (CSI), meaning the channel information available to the optimizer, and RIS-assisted reflections. The reported computational and performance results were averaged over 50 independent channel realizations per configuration, a per-configuration count rather than a pooled total.
WSR was scored on true instantaneous, geometry-consistent channels, even though SCRPA and the baseline methods optimized using imperfect and aged CSI. Unless otherwise stated, the methods used shared channel realizations. This setup separates the channel used to judge the final rate from the less-than-perfect information available during optimization.
Detection varied with the test conditions
At a false-alarm setting of alpha = 10^-3 and a sequence length of 255, the reported active-RIS identification probability exceeded 95%, with false positives virtually absent across the tested configurations. Average Jaccard-index performance, an overlap measure between the detected and actual active sets, improved with signal-to-noise ratio (SNR) and sequence length and was better when fewer RIS panels and UEs were modeled.
The detection result needs a qualification. The reported confusion-matrix definitions and denominators are not clearly explained, and one displayed confusion matrix contains a nonzero false-negative cell. The finding supports strong reported performance in this test setting, not a claim of perfect detection.
Higher rate, with an important comparison caveat
In the end-to-end test under 30% per-link blockage, SCRPA consistently produced higher WSR than the baseline methods across the considered speeds and configurations and closely approached the Strong LP Benchmark. The exact WSR values are shown in plots, with means and variability reported but no confidence intervals.
Interpretation also depends on what each method knew about blockage. The other algorithms in that comparison assumed all RIS links were available, so the numerical gap should be read as a comparison under different blockage-information assumptions, not as a universal ranking of methods.
The iteration count was not uniformly lower
At RX-SNR = 10 dB with M = 16, K = 8, 4-bit RIS phase quantization, and 5 km/h mobility, SCRPA reached 99% of converged WSR after approximately 386 iterations. AO-SPR and RMO-LS reached that target in 47 and 413 iterations, respectively. SCRPA therefore sat between those two methods in this reported setting: slower than AO-SPR, but faster than RMO-LS. These figures are approximate or averaged, and they measure iterations rather than wall-clock latency.
The hardware question is still open
Phase quantization did not materially separate the reported WSR results. SCRPA was comparable at 2-, 4-, and 8-bit RIS phase resolution, with the 4-bit versus 8-bit gap within approximately 0.3%.
For M = 32 and K = 16, the reported SCRPA cost was approximately 3.869 GFLOPs. The authors estimate that fitting the end-to-end pipeline within a 5 ms processing budget would require approximately 0.774 TFLOPs/s. Those are analytical operation and throughput estimates, not measured implementation times.
The study's practical reach remains untested beyond its simulation setting. Its evidence comes from ray-tracing channels in the Munich urban scene with simulated mobility, blockage, and RIS-assisted reflections, so the results are numerical and model-based rather than a demonstration across other scenes or configurations. The convergence result likewise stops at block-stationary solutions, and the timing case is based on throughput estimates rather than measured execution.
The manuscript is marked as submitted to IEEE Transactions on Wireless Communications and as an arXiv v1 preprint dated 26 August 2026. Acceptance or final publication is not reported.
Paper data and sources
Original title: End-to-End Mobility-Aware Multi-RIS Optimization via Blockage Detection and Closed-Form Riemannian Updates
Authors: Sehyun Ryu, Seungmin Choi, Hyun Jong Yang, John M. Cioffi
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text