A reconfigurable intelligent surface (RIS) is a flat panel of many small antenna elements, each ended by a tunable electronic load, that redirects radio waves toward a receiver without amplifying them. Nearly every algorithm for setting those loads treats each element as an ideal, independent knob that rotates the phase of its reflection while leaving the reflection's strength fixed. Physically Consistent Modeling of Wireless Links With Reconfigurable Intelligent Surfaces Using Multiport Network Analysis shows this picture is wrong even for one or two elements. In a passive surface, phase and amplitude are tied together: a ±45° shift costs a mandatory 3 dB in amplitude. Neighbouring elements also interact, so the best achievable gain swings from 6 dB to 0 dB with element spacing. That paper stops at two elements and has no scalable optimizer.
This note starts from an observation about the circuit equations. Once the physically consistent model is written out, the end-to-end channel is a linear fractional transformation of the diagonal matrix of load reactances. That is the same algebra as closing a static feedback gain, restricted to be diagonal, around a known multi-input multi-output plant. Tuning an RIS is therefore a structured static output feedback problem. Seen this way, the ideal-phase-shifter model becomes the decoupled limit of the feedback interconnection, and control-theory tools for that problem class transfer to it. Those tools include optimization-landscape analysis of static feedback gains and gray-box synthesis that combines physical priors with measured data. We propose to build the resulting optimizer and to measure how much the simplified model costs as surfaces become larger and more densely packed.
Core Claim
Let $X$ be the $N\times N$ diagonal matrix of element load reactances, $Z_{RR}$ the $N\times N$ mutual impedance matrix among RIS elements, $z_{RS}$ and $z_{DR}$ the $N$-vectors of impedances coupling the source to the RIS and the RIS to the destination, and $z_{DS}$ the direct source–destination term. In the multiport model the end-to-end channel is
$$h(X) = z_{DS} - z_{DR}^{\top}\,(Z_{RR} + jX)^{-1} z_{RS}.$$
This is a feedback interconnection. The loads $jX$ act as a static, diagonal, purely reactive gain closed around the "plant" $Z_{RR}$, and $h$ is the closed-loop response from source to destination. Two cases fall out of this one expression. The usual simplified model appears only when $Z_{RR}$ is diagonal (no coupling) and the reflection amplitude is forced to one; Nossek et al.'s amplitude penalty survives even after coupling is removed.
We claim three things. First, the error of the simplified model can be predicted from how far $Z_{RR}$ is from diagonal, and it grows with element density and array size. Second, gradient methods on the reactance parameterization inherit useful landscape properties from static output feedback theory, so they can be run reliably at hundreds of elements. Third, because $h(X)$ is linear in the unknowns $(z_{DS},\, z_{RS}z_{DR}^{\top})$ once $Z_{RR}$ is fixed, a geometry-derived prior on $Z_{RR}$ turns channel acquisition into a low-rank linear regression on received pilots. We can then treat the error in that prior as structured uncertainty in a robust design. This avoids estimating the cascaded channel element by element.
Relation to Prior Work
Two RIS papers supply the simplified-model baselines this note argues against. Achievable Rate maximization by Passive Intelligent Mirrors optimizes unit-modulus phases by majorization–minimization, with closed-form power allocation, and reports at least a 40% sum-rate gain over no surface. Performance Analysis of Intelligent Reflecting Surface Aided Communication Systems derives the $N^2$ power scaling and an $N+1$ diversity order. Both assume uncoupled, unit-amplitude elements. We keep their problem settings and replace their plant model.
The physics we build on comes from Nossek et al. (multiport analysis). They supply the impedance model and the two-element ground truth. We differ in scaling to many coupled elements, in the feedback reading of the model, and in identifying the channel from data.
The channel-acquisition baseline is Matrix-Calibration-Based Cascaded Channel Estimation. It estimates the factored cascaded channel by approximate message passing under angular sparsity. That paper motivates our work by showing training overhead is the bottleneck, and it constrains our design because it achieves low overhead by exploiting angular structure that our regression should also exploit. It assumes ideal elements, whereas we assume a known structure up to coupling uncertainty.
From control we reuse methods, not plants. Optimization Landscape of Policy Gradient Methods for Discrete-Time Static Output Feedback proves that static output feedback costs are smooth and coercive, and that policy gradient converges locally at a linear rate. We ask which of these properties survive when the gain is diagonal and reactive and the objective is received power instead of a quadratic cost. Combining Prior Knowledge and Data for Robust Controller Design shows that a linear-fractional prior combined with noisy data stabilizes where purely data-driven design fails. That result supplies the template for our gray-box treatment of $Z_{RR}$. Sample Complexity of LQG Control for Output Feedback Systems motivates writing performance loss as linear in model error. We expect a similar bound in coupling-matrix error.
We deliberately do not use the dynamic, trajectory-based machinery of An Efficient Data-Driven Framework for Linear Quadratic Output Feedback Control. An RIS has no internal dynamics within a coherence block, so a Willems-style substitute state would be a category error. The feedback structure here is spatial, not temporal.
Research Program
Model. We build coupled RIS plants for $N$ from 2 to 256, using analytic mutual impedances for thin dipoles and for isotropic elements at spacings from $\lambda/8$ to $\lambda/2$. We embed them in propagation drawn from Sionna's 3GPP Urban Micro channels as used in PyJama, and from the 3GPP indoor-factory parameters of the 3D non-stationary industrial model. The factory scenes add dense diffuse multipath, which makes $z_{RS}$ and $z_{DR}$ less sparse in angle. Including them tests whether the regression step degrades gracefully when the sparsity assumption weakens.
Analysis. We will bound the gain lost by optimizing under the simplified model. Expanding $(Z_{RR}+jX)^{-1}$ around its diagonal part suggests the loss scales with the norm of the off-diagonal coupling relative to the self-impedance. This is a hypothesis to check, not a result. We will also study the landscape of $|h(X)|^2$ over reactances. Questions include whether spurious local maxima appear as coupling grows, and whether a decoupled warm start puts gradient ascent in the right basin. These are the analogues of the disconnected stabilizing regions that the static output feedback literature warns about.
Gray-box identification. We transmit pilots under $M$ known reactance settings. Given a nominal $Z_{RR}$, we solve a rank-one-constrained least-squares problem for $(z_{DS}, z_{RS}z_{DR}^{\top})$. Separately, we design the reactances to maximize worst-case gain over a norm-bounded perturbation of $Z_{RR}$, following the uncertainty-multiplier approach of Berberich et al. Discrete 1- and 2-bit reactance sets are handled by projecting the continuous solution and then running rank-one coordinate refinement.
Baselines
The naive reference is a surface held at a fixed default configuration, with all reactances zero. This gives the gain that any method must exceed.
The strongest existing method in the file is ideal-phase-shifter optimization: majorization–minimization of unit-modulus phases as in Huang et al.. Each phase is then mapped to the reactance that produces it in isolation, and the result is evaluated on the true coupled plant. It is the strongest competitor because it is what deployed designs would actually run. For a fair comparison it receives the same perfect or estimated channel, the same iteration budget, and the same discrete reactance sets. The only reported number for this comparison comes from Nossek et al.: the physically consistent optimum beats this approach by "up to 6 dB depending on element spacing" in the two-element case. For more than two coupled elements, no reported result found on Lacuna. Beating this baseline means a consistent gain at spacings at or below $\lambda/4$ that grows with $N$, with no loss at $\lambda/2$.
For channel acquisition, the baseline is matrix-calibration estimation from Wang et al.. They report needing "about 30% of the training overhead of [18] and about 1% of [15]," but under ideal elements. We run it with the same pilot budget $M$ and the same angular dictionary, and compare the final received gain after optimization, not estimation error alone. Beating it means equal gain with fewer pilots on coupled plants.
Two simpler optimizers on the exact coupled model are included because a skeptic would propose them. One is coordinate ascent over reactances using Sherman–Morrison rank-one updates. The other is zeroth-order search on measured received power, given the same number of measurements as our pilots.
The key ablation runs our full pipeline with $Z_{RR}$ replaced by its diagonal. A second ablation drops the geometric prior entirely, regressing on $Z_{RR}$ as well, to isolate what the prior contributes.
Why not just…?
Why not just optimize with the ideal model and convert phases to reactances? This is the default, and at half-wavelength spacing it may be nearly lossless. Nossek et al. find exactly 0 dB normalized gain at $d=\lambda/2$, where the optimal reactances are symmetric. If the spacing sweep shows the gap vanishing at $\lambda/2$ for large $N$ too, our claim narrows to densely packed surfaces. We would report that narrowing as a finding about when the simplified model is safe, not argue around it.
Why not just run coordinate ascent on the exact coupled model? This is the real threat. Rank-one updates make each step cheap, and with perfect knowledge of $Z_{RR}$ it may match gradient methods in final gain. The feedback view must earn its place elsewhere. It should yield a warm-start rule and landscape diagnosis that coordinate ascent lacks. It should also yield robustness when $Z_{RR}$ is only approximately known: coordinate ascent has no handle on model error and will chase it. The perturbed-coupling experiment below settles this. If coordinate ascent equals the robust design under mismatch, we drop the robust-synthesis component.
Why not just search blindly on measured power, with no model? This needs no coupling model at all. With $N$ in the hundreds, however, each measurement gives one scalar, and evidence from control is that purely data-driven designs break down as noise grows: Berberich et al. report that their prior-plus-data design reached an $H_2$ norm of about 1.93 where purely data-driven methods failed to stabilize. We test the measurement-count regime directly. If blind search reaches the same gain within one coherence block's pilot budget, the identification step is unnecessary.
Why not just add a first-order coupling correction to the ideal model? A one-term expansion of the inverse captures weak coupling cheaply. It is the truncation of our own expression, so we include it as a baseline. We expect it to fail exactly where the off-diagonal norm approaches the self-impedance, which is the regime where the full model matters.
Proposed Experiments
Two-element sanity check. We first reproduce Nossek et al.'s two-element table: 6 dB at $d=0$, 4.6 dB at $\lambda/4$, and 0 dB at $\lambda/2$, with their optimal reactances. This confirms the plant and optimizer are correct before scaling.
Spacing and size sweep. We then sweep spacing and $N$ in UMi and factory scenes. We report received gain in dB relative to the naive surface, multi-user sum-rate in the Huang et al. setting, and pilots used. The first decisive check is whether the gap between ideal-model and consistent optimization grows with $N$ at sub-$\lambda/4$ spacing and tracks the predicted off-diagonal coupling norm.
Perturbed-coupling robustness. We perturb the true $Z_{RR}$ away from the geometric prior by 5% to 30% in norm, representing fabrication and mounting error. We then compare the worst-case and median gain of robust design, nominal gradient design, and exact coordinate ascent. This is the second decisive check.
Landscape study. From 1,000 random initializations we record the spread of local optima reached by gradient ascent, with and without a decoupled warm start. This tests whether the static-output-feedback convergence picture carries over to the reactance landscape.
Factory scenes. All experiments are repeated in the heavy-clutter factory scenario. This measures whether dense diffuse multipath erodes the pilot savings of the low-rank regression relative to angular-sparsity estimation.
References
- Physically Consistent Modeling of Wireless Links With Reconfigurable Intelligent Surfaces Using Multiport Network Analysis
- Achievable Rate Maximization by Passive Intelligent Mirrors
- Performance Analysis of Intelligent Reflecting Surface Aided Communication Systems
- Matrix-Calibration-Based Cascaded Channel Estimation for Reconfigurable Intelligent Surface Assisted Multiuser MIMO
- Optimization Landscape of Policy Gradient Methods for Discrete-time Static Output Feedback
- Combining Prior Knowledge and Data for Robust Controller Design
- Sample Complexity of Linear Quadratic Gaussian (LQG) Control for Output Feedback Systems
- An Efficient Data-Driven Framework for Linear Quadratic Output Feedback Control
- PyJama: Differentiable Jamming and Anti-Jamming with NVIDIA Sionna
- A 3D Non-Stationary Geometry-Based Stochastic Model for Industrial Automation Wireless Communication Systems