Primal-dual methods solve constrained optimization problems by updating two sets of variables in turn. The "primal" variables are the quantities being designed. The "dual" variables are prices that push the solution toward satisfying the constraints. These methods are cheap per iteration and easy to run on hardware, but they work only when two step sizes are balanced: the primal step $\tau$ and the dual step $\sigma$. In millimeter-wave and terahertz wireless links, signals travel in narrow beams that a passing pedestrian or truck can cut off entirely. When that happens, the channel matrix that couples the transmitter's design variables to the users' quality-of-service constraints changes in one step. The step sizes that were stable a moment ago can become too aggressive or too timid. A user who loses every path can also make its constraint infeasible, and its dual price then grows without limit. This is a large part of why blockage-aware designs avoid simultaneous first-order primal-dual updates and fall back on block coordinate descent or closed-form tricks.
Our claim is that the receiver already observes the blockage event and can use it to retune the method. In multi-RIS systems, where a reconfigurable intelligent surface (RIS) is a passive panel of tunable phase-shifting elements that reflects the beam around obstacles, per-surface synchronization sequences already tell the user which paths are blocked (Ryu & Yang). We propose to treat these path indicators as triggers that do three things at once. They supply a certified bound on the new operator norm with no extra matrix products. They reset the primal-to-dual step ratio. They cap the dual prices of users whose constraints have just become infeasible. The resulting single-loop method updates the transmit beamformer, the RIS phases, and the dual prices together. It should recover from a blockage in a handful of iterations with no per-scenario tuning, in regimes where closed-form alternating methods cannot handle the constraints and block coordinate descent is too slow.
Core Claim
Adaptive primal-dual theory has made the primal step track local curvature. Vladarean, Malitsky & Cevher estimate smoothness from gradient differences (APDA). On nonconvex phase retrieval it converged in every instance where Condat-Vũ diverged in 80% of cases. That work still needs two things fixed in advance: an upper bound on the operator norm $\|A\|$, and the ratio $\beta=\tau/\sigma$. Chambolle & Contreras show that the seeded ratio largely determines convergence speed and that linesearch cannot repair a bad one.
These two works motivate the proposal. Together they say the remaining sensitivity sits in exactly the two quantities that blockage perturbs.
Our contribution has two parts. First, we make the observation that in blockage-limited channels both quantities change in discrete, detectable jumps rather than continuously. Second, we build a method that exploits this. Write the effective channel at time $t$ as $A_t=\sum_p b_p(t)A_p$. Here $A_p$ is the contribution of path $p$ (direct or through a given RIS), and $b_p(t)\in\{0,1\}$ is the detected unblocked indicator. The triangle inequality then gives a bound that holds at every iteration:
$$\tau_k\,\sigma_k\Big(\sum_p b_p(t)\,\|A_p\|\Big)^2 \le 1,$$
where $\tau_k,\sigma_k$ are the primal and dual steps at iteration $k$. The per-path norms $\|A_p\|$ are computed once per coherence block. A blockage simply removes terms from the sum. This replaces APDA's fixed global bound with one that tightens when paths disappear and relaxes when they return.
Research Program
Problem. The test problem is joint active and passive beamforming in a multi-RIS link with constraints. The objective is either transmit power or weighted sum rate. The constraints are per-user rate floors (in second-order-cone form for a fixed RIS configuration), a power budget, and unit-modulus RIS phases $|e^{j\theta_n}|=1$. The unit-modulus constraint makes the joint problem nonconvex and full of local optima (RIS optimization survey).
Update rule. The primal step on the beamformer is a curvature-adaptive APDA step. The RIS update is a gradient step followed by elementwise normalization back onto the unit circle, a cheap projection. The dual step is projected onto a bounded box.
First event-triggered ingredient: the norm bound. When the indicators $b_p$ change, the step product is rescaled using the bound above.
Second ingredient: the ratio reset. The ratio $\beta$ is re-seeded from the post-event scales of the primal iterate and the dual prices. Between events it is adjusted by balancing primal and dual residuals.
Third ingredient: infeasibility handling. A user whose set of surviving paths cannot meet its rate floor under the full power budget is detected cheaply. The check is the received power when every surviving path is aligned. That user's price is capped and its constraint is relaxed into a penalty, which keeps one dead-zone user from destabilizing everyone else. This borrows the bounded, time-varying multiplier idea from LoTFair. LoTFair handles long-term constraints with dynamic multipliers but assumes convex decisions and no structured jumps. We supply neither a regret guarantee on the nonconvex manifold nor claim one.
Theory. The theoretical target is modest and specific. For a fixed RIS configuration the beamforming subproblem is convex. There we will prove convergence within each blockage epoch under the certified bound. We will also bound the total tracking cost by the number of blockage events multiplied by a per-epoch transient that depends on how loose the triangle bound is. The joint nonconvex case is studied empirically.
How this differs from the closest work. Ryu & Yang use the same path indicators. They feed them to CRPA, a closed-form Riemannian phase update that avoids step sizes altogether. That avoidance holds only for an unconstrained weighted-sum-rate objective with perfect channel knowledge. Once rate floors require dual prices, no closed form exists. Zhao et al. handle constraints in pinching-antenna systems through block coordinate descent with majorization-minimization. They acknowledge that its complexity scales superlinearly, and they assume static blockages.
Continuous unrolling (cLPD) stabilizes primal-dual networks against missing angular data in CT. That shows step-size overshoot under data truncation is real, but its fix is learned, not certified. We keep the method unlearned so that any stability we report comes from the triggers and not from training data.
Baselines
Do nothing. The naive reference is plain Condat-Vũ primal-dual splitting. Its steps are tuned once by grid search on the unblocked channel and then frozen. This is what a practitioner would deploy, and it isolates the cost of ignoring blockage. It receives the same grid budget as our initial seed.
APDA with a worst-case norm. This is APDA run with the norm bound set to the all-paths-unblocked value and a fixed ratio $\beta$. It is the strongest published adaptive primal-dual method. It differs from ours only in what it does at events. Its $\beta$ is chosen by the same grid search, and both receive the same per-iteration matrix-product budget.
Strongest existing system: CRPA inside an outer loop. This is CRPA (Ryu & Yang) wrapped in an outer dual-ascent loop on the rate-floor prices, with the inner beamformer solved by its closed-form updates. It is the strongest competitor because it is the current best on the very benchmark we use. The file reports that CRPA reaches 95% of its converged weighted sum rate in about 12 iterations at SNR = 5 dB, beating Riemannian line search and fractional-programming alternating optimization.
On the unconstrained problem we must match this to within a few iterations. Beating it means reaching feasibility within the same rate tolerance in fewer total inner iterations, and in less wall-clock time, after a blockage on the constrained problem. No reported result was found on Lacuna for CRPA with rate-floor constraints, so we generate that number ourselves.
Block coordinate descent. The fourth baseline is block coordinate descent with majorization-minimization in the style of Zhao et al.. It is compared on wall-clock time per recovery at matched final objective.
Ablations. Ablations remove each trigger in turn: the norm bound, the ratio reset, and the dual cap. A further ablation replaces the triangle bound with a few power-iteration steps at each event. This isolates whether the free bound is tight enough to matter.
Why not just…?
Why not just run a power iteration at each blockage and restart APDA? This is the most serious threat. It would fix the norm bound at the cost of a few extra matrix products. It does nothing about the ratio or the runaway dual prices, which we expect to dominate. The "power iteration plus restart" ablation settles this directly. If it matches the full method, the paper's contribution narrows to the ratio and dual-cap ingredients, and we will say so.
Why not just use residual balancing to adapt the ratio continuously? Residual-balancing heuristics are a natural reviewer suggestion. No reported result was found on Lacuna for them in this setting, so they are included in our grid as a variant of the APDA baseline. Our hypothesis is that continuous balancing reacts too slowly to a one-step jump and is corrupted by an infeasible user's diverging price. The "ratio reset off, dual cap on" ablation measures exactly this.
Why not just use block coordinate descent or fractional-programming alternating optimization, which converge reliably? Reliability is not the issue; speed is. Zhao et al. report superlinear scaling. Ryu & Yang report that fractional-programming alternating optimization needs more iterations than CRPA. We will report recovery wall-clock time as the number of RIS elements grows. If block coordinate descent stays within the coherence time at realistic sizes, the case for our method weakens to the large-surface regime.
Why not just predict beams from cameras and LiDAR? Sensor-based predictors reach 98.19% top-13 accuracy (Mollah et al.) and a distance-based accuracy score of 0.9016 (Wang et al.). They output codebook indices and do not enforce rate floors or power budgets. Both papers note degradation at night and in bad weather. These predictors are complementary rather than competing, since a predicted beam can initialize our iterate. We test that as a warm-start variant rather than as a rival.
Proposed Experiments
Main testbed. The main testbed is the 28 GHz multi-RIS simulation of Ryu & Yang: a 16-antenna base station, 4-antenna users, and several distributed surfaces. We add per-user rate floors.
Blockage timelines. Blockage timelines are not purely synthetic. They are extracted from received-power drops in DeepSense-6G vehicle-to-infrastructure scenarios 31–34, covering day and night. These scenarios have no RIS, so we use them only for event timing and duration statistics.
Sanity check. A static check on the LIBSVM logistic-regression tasks and on phase retrieval confirms that our triggers never fire spuriously. It also confirms that we do not lose APDA's reported robustness relative to Condat-Vũ.
Metrics. We measure iterations and wall-clock time to return within 5% of the post-event optimum, with all rate floors met to tolerance. We also measure cumulative constraint violation across a timeline, and the final power or sum rate.
Hyperparameter sensitivity. The key metric is the fraction of an initial $(\tau_0,\beta_0)$ grid spanning three orders of magnitude that converges within a fixed iteration budget across all events.
Decisive checks. Three outcomes decide the proposal. First, our method should keep that converging fraction near one while frozen-step Condat-Vũ and worst-case-norm APDA collapse after the first blockage. Second, the power-iteration ablation should show whether the free triangle bound is sufficient. Third, on timelines with a fully dead-zoned user, removing the dual cap should visibly degrade every other user's recovery. If that third effect does not appear, the infeasibility mechanism is not the bottleneck we claim, and the proposal should be reframed around the norm and ratio triggers alone.
References
- Blockage-Aware Multi-RIS WSR Maximization via Per-RIS Indexed Synchronization Sequences and Closed-Form Riemannian Updates
- A first-order primal-dual method with adaptivity to local smoothness
- Accelerated Bregman Primal-Dual Methods Applied to Optimal Transport and Wasserstein Barycenter Problems
- A Survey on Model-Based, Heuristic, and Machine Learning Optimization Approaches in RIS-Aided Wireless Networks
- Long-Term Fairness for Real-Time Decision Making: A Constrained Online Optimization Approach
- Robust and Secure Blockage-Aware Pinching Antenna-assisted Wireless Communication
- Continuous Learned Primal Dual
- Multi-Modality Sensing in mmWave Beamforming for Connected Vehicles Using Deep Learning
- When Vision-Language Model (VLM) Meets Beam Prediction: A Multimodal Contrastive Learning Framework