A reconfigurable intelligent surface (RIS) is a panel of many cheap reflecting elements, each of which can shift the phase of an incoming radio wave. Steering the reflected beam means choosing one phase per element. The standard way to make that choice is semidefinite relaxation (SDR). SDR rewrites the problem in terms of a matrix of pairwise products, which makes it convex and solvable, then rounds the answer back to actual phases by random sampling. Two assumptions go into that design and are rarely questioned. The first is that every element obeys its command perfectly. The second is that the channel follows a textbook fading law such as Rayleigh or Rician. Real elements do not behave this way: they drift, lose amplitude, and get stuck, and they do so in a few recognisable health states that change over time. Real channels, as ray-tracing shows, look nothing like Rayleigh.
We propose to put the element health model inside the relaxation. The proposal rests on a simple observation. If each element moves among a few health states according to a Markov chain (a model in which the next state depends only on the current one), then the expected signal power at the moment the configuration is actually used is still a quadratic in the commanded phases. Its weighting matrix is reweighted entry by entry by the predicted health of each pair of elements. Health-awareness therefore costs nothing in tractability. What remains is the rounding step, the restriction to 1- or 2-bit phases, and the mismatch with ray-traced channels. We handle these with a discrete local search on a high-fidelity objective, seeded by the relaxation, and we test explicitly whether simulated annealing earns its place over plain greedy search.
Core Claim
Model the realised reflection of element $i$ as $\theta_i = \alpha_i v_i + \beta_i$. Here $v_i$ is the commanded unit-modulus phase, $\alpha_i$ is a random gain-and-phase-bias factor, and $\beta_i$ is a random fixed reflection that the element produces when stuck. Both are determined by the element's health state at use time. Under independence across elements, the expected received power of user $k$ is
$$ \mathbb{E}\,|h_k^H \theta|^2 = \tilde v^H \big(A_k \circ M\big)\,\tilde v + \text{const}_k . $$
The symbols are as follows. $h_k$ is the cascaded channel to user $k$. $\tilde v = [v; t]$ appends a unit-modulus homogenising variable $t$ that absorbs the stuck-element offset. $A_k$ is the usual lifted channel matrix. $\circ$ denotes the entrywise product. $M$ holds the predicted first and second moments of $(\alpha_i, \beta_i)$. $\text{const}_k$ is a user-specific diagonal term.
That diagonal term looks inert, but it is not. In max-min problems across users it shifts each user's floor differently, so it changes the optimal trade-off between users.
The moments come from propagating the current health distribution forward by the delay $\tau$ between channel estimation and use. When $\tau$ is near zero, re-estimating the cascaded channel already absorbs static degradation, and our method should reduce exactly to standard SDR. The claim is therefore falsifiable in a specific way. Health-aware lifting should beat health-oblivious design by a margin that grows with $\tau$ times the state-transition rate, and with the fraction of stuck elements. It should provide no benefit when $\tau \to 0$.
Research Program
The work has three parts.
Element health model. We build a four-state model for each element: healthy, phase-biased, amplitude-degraded, and stuck. We follow the discrete-state reliability tradition reviewed in Performability Analysis for Large-Scale Multi-State Computing Systems, which argues that treating degraded units as either fully working or dead mis-sizes capacity in both directions. For tracking hidden health states we follow the HMM-based approach of Rebello et al., which supplies the filtering method we reuse to infer current element states from pilot measurements. Our use differs from both. Those works evaluate reliability. We feed predicted state moments into an optimiser.
Health-aware relaxation. We derive the reweighted matrix $M$ and solve max-min SNR with an off-the-shelf SDP solver. This extends the continuous-health power-flow relaxation of Health-Focused Optimal Power Flow, which lifts deterministic health limits into an SDP. Our setting differs in three ways: it is wireless, the health is stochastic and Markov, and it concerns discrete reflecting hardware. That paper explicitly leaves stochastic health as open work.
Discrete refinement. We take the relaxation's leading eigenvector and quantise it to the available phase levels. We then run single-element-flip search on a high-fidelity objective. That objective is minimum SNR averaged over sampled health trajectories and ray-traced channels, and it has no closed form. The search either accepts uphill moves with probability $\exp(-\Delta E/T)$, which is simulated annealing, or accepts only improvements, which is greedy search.
The SDR optimum gives an upper bound on the simplified-model surrogate only, not on the real objective. We will report the gap between the two rather than present it as a certificate.
Baselines
The naive reference is quantised SDR with Gaussian randomization under the simplified channel and perfect hardware. This is the pipeline of Max-Min Fairness for IRS-Assisted Secure Two-Way Communications, which uses Rician RIS links, Rayleigh direct links, and randomization. That paper concedes the randomization step is sub-optimal. It reports up to a 3.6-fold minimum secrecy-rate gain over its own baselines and convergence in 4 to 10 outer iterations. Those numbers are on a secrecy objective, not ours. We found no reported result on Lacuna for discrete-phase RIS design under element degradation, so we will re-run this pipeline ourselves. The pipeline of RIS-Enhanced Cognitive ISAC has the same structure and is covered by the same baseline.
The strongest competitor is health-oblivious SDR followed by greedy element-wise search on the same high-fidelity objective, using freshly re-estimated cascaded channels. It absorbs static degradation for free and is what a careful practitioner would build.
Our ablations isolate each ingredient. Health-aware SDR with randomization isolates the search step. Health-aware SDR with greedy search versus with annealing isolates the acceptance rule. Annealing from random starts isolates the relaxation seed.
All search methods receive the same budget of objective evaluations. All tuning, including annealing schedules and randomization sample counts, uses a held-out set of user drops. Every method is scored on the same ray-traced channels and sampled health trajectories.
Beating the strongest competitor means a statistically significant gain in tenth-percentile minimum SNR across user drops at matched evaluation budget, in the large-$\tau$ regime. Matching it at $\tau = 0$ is the expected outcome, not a failure.
Why not just…?
Why not just re-estimate the cascaded channel and forget about hardware models? This is the most serious threat. Re-estimation does absorb any degradation that stays put between estimation and use, so we expect it to tie our method at small $\tau$. It cannot anticipate elements that change state during the configuration interval. Nor can it anticipate stuck elements whose reflection does not depend on the command. The $\tau$ sweep below measures exactly where, or whether, the crossover happens. If no crossover appears at realistic transition rates, the health-model contribution is refuted.
Why not just use greedy search instead of simulated annealing? With 1- or 2-bit phases, single-element greedy ascent from a good seed may land in the same basin as annealing. The file's evidence for annealing comes from other combinatorial domains, such as ACO-SA path planning and label placement, not RIS design. We therefore treat the comparison as open. For small surfaces, exhaustive enumeration will show which method reaches the true discrete optimum more often. If greedy search matches annealing, we report it and drop annealing from the method.
Why not just skip the relaxation and run a metaheuristic directly? Optimal UAV Deployment uses particle swarm only as a low-dimensional benchmark, and its search has no bound to report against. Our random-start annealing ablation measures directly how much the relaxation seed buys at a fixed budget as the surface grows to hundreds of elements.
Why not just design under Rayleigh fading? Random Beamforming in mmWave NOMA and Normalization-Free Knowledge Distillation both show that idealised or scale-normalised channel assumptions distort design quality. The second reports -12.87 dB versus -11.98 dB NMSE on DeepMIMO. Rather than assume the cost of the textbook channel, we measure it with the decomposition below.
Proposed Experiments
Channel data. We draw channels from DeepMIMO scenario O1, the ray-traced dataset used by Seo et al.. We treat a planar-array location as the surface, which yields BS-to-surface and surface-to-user links. As the simplified counterpart we use the Rician/Rayleigh model of Jayarathne et al.
Hardware settings. Surfaces range from 8 to 256 elements with 1- and 2-bit phases. For small instances (1-bit up to 12 elements, 2-bit up to 8 elements) we enumerate every configuration to obtain a true optimum. Health transition matrices are swept from slow to fast. The stuck fraction is swept from 0 to 20 percent.
First decisive check: a two-by-two decomposition. We cross the channel model used for design (simplified versus ray-traced) with the hardware model used for design (oblivious versus health-aware). Each cell is scored by its gap to the exhaustive or long-run optimum. This tells us whether the channel mismatch or the hardware mismatch dominates the loss.
Second decisive check: the $\tau$ sweep against re-estimation. This is the falsification test for the core claim.
Third decisive check: the search comparison. We compare annealing and greedy search at matched evaluation budget, against the exhaustive optimum.
Secondary analysis. We report the fraction of relaxations that come out rank-one under health reweighting. The entrywise damping of off-diagonal terms could plausibly push solutions toward higher rank, and if so, that would make the search stage matter more.
References
- Performability Analysis for Large-Scale Multi-State Computing Systems: Methodologies, Advances, and Future Directions
- An integrated approach for system functional reliability assessment using Dynamic Bayesian Network and Hidden Markov Model
- Health-Focused Optimal Power Flow
- Max-Min Fairness for IRS-Assisted Secure Two-Way Communications
- RIS-Enhanced Cognitive Integrated Sensing and Communication: Joint Beamforming and Spectrum Sensing
- A Highly Efficient ACO-SA Algorithm for Robot Path Planning
- A Hybrid Discrete Artificial Bee Colony Algorithm Based on Label Similarity for Solving Point-Feature Label Placement Problem
- Optimal UAV Deployment for Rate Maximization in IoT Networks
- Random Beamforming in Millimeter-Wave NOMA Networks
- Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO