A reconfigurable intelligent surface (RIS) is a wall-sized panel of cheap printed antenna patches. Each patch can shift the phase of a radio wave bouncing off it, so the panel steers signals toward users. Because the patches are passive, with no radio receiver behind them, the surface cannot "hear" anything. The base station only sees the combined base-station-to-surface-to-user channel, and learning that channel element by element costs pilot time that grows with the number of patches. Two papers break this bottleneck by adding active hardware. One wires a subset of elements to full receivers (Deep Denoising Neural Network Assisted Compressive Channel Estimation for mmWave Intelligent Reflecting Surfaces). The other puts a cheap power detector on every element and needs only 3 pilot slots per user, against roughly 200 for linear minimum-mean-square-error estimation at $N=200$ (Sensing RISs: Enabling Dimension-Independent CSI Acquisition for Beamforming). Both papers name the same unsolved problem: active parts bring wiring, power and heat. They cannot go on every element, so which $k$ of the $N$ elements should carry them?
This note argues that the question is really about robustness, not about solving a combinatorial search quickly. The sensor layout is fixed when the panel is built. It must keep working as users move and the channel's spatial statistics change. This is the same situation as a long-only portfolio: a portfolio with $k$ holdings and non-negative weights is chosen on an estimated covariance and then must survive out of sample. That literature has a documented failure mode. With non-negative weights, the optimizer has many near-tied solutions, and the chosen weights jump under tiny changes to the estimated covariance (Benzaquen et al. call this "disorder chaos"). It also has a solver that handles the combination of cardinality, non-negativity and covariance uncertainty (Cardinality-constrained Distributionally Robust Portfolio Optimization). We propose to carry that formulation over to sensor placement. We then test whether a layout chosen to minimize worst-case estimation error over plausible future channel statistics beats layouts that are optimal for one nominal estimate, or merely evenly spread.
Core Claim
Choosing $k$ sensor elements and splitting a pilot-power budget among them is a cardinality-constrained, long-only design problem with the same structure as robust portfolio selection. We claim three things. First, the layout that is optimal for a nominal channel covariance is unstable: small perturbations of the estimate change which elements are chosen. Second, that layout degrades sharply when the user population drifts. Third, a distributionally robust version, solved with a bilevel cutting-plane method adapted from the portfolio literature, gives lower worst-case estimation error across held-out channel regimes at the same hardware budget.
To make this precise, let the cascaded channel across the $N$ elements be $h \sim \mathcal{CN}(0, R)$, where $R$ is its $N\times N$ spatial covariance. Let $s\in\{0,1\}^N$ mark which elements carry a sensor, with $\|s\|_0\le k$. Let $p\ge 0$ be the pilot power given to each sensed element, with $\sum_i p_i \le P$ and $p_i=0$ wherever $s_i=0$. With noise variance $\sigma^2$, the posterior error of the whole channel is $\mathrm{tr}\big((R^{-1} + \mathrm{diag}(p)/\sigma^2)^{-1}\big)$. The robust design solves
$$ \min_{s,\,p}\;\max_{R\in\mathcal{U}(\hat R,\rho)}\;\mathrm{tr}\Big(\big(R^{-1} + \mathrm{diag}(p)/\sigma^2\big)^{-1}\Big), $$
where $\mathcal{U}(\hat R,\rho)$ is a ball of radius $\rho$ around the estimated covariance $\hat R$, in Frobenius or $L_1$ norm, matching the ambiguity sets of the portfolio paper. Two facts make this tractable.
The objective is concave in $R$. The map $R\mapsto(R^{-1}+D)^{-1}$ is a parallel sum, and parallel sums are operator concave. The inner worst case is therefore a concave maximization. The objective is also convex in the non-negative weights $p$. For fixed $s$ the problem is a convex–concave saddle problem, and only the binary selection is hard. That is the structure the CPA+ cutting-plane method handles in the portfolio setting.
Research Program
The first component is the solver. Cardinality-constrained Distributionally Robust Portfolio Optimization formulates long-only, $k$-sparse robust portfolio choice as a mixed-integer semidefinite program. It solves this with an outer master problem over the binary selection and inner semidefinite subproblems that generate cuts. The paper's key speedup comes from a matrix completion theorem: the inner subproblem only needs the $k\times k$ sub-covariance of the selected assets. We reuse the method, and we flag one place where it may not transfer. Portfolio variance depends only on the selected assets. Our error, by contrast, includes the unsensed elements, which are predicted from the sensed ones through cross-covariances. So the $k\times k$ decoupling does not hold in general. We will work out what does survive. One candidate is a block structure in which only the $k\times N$ cross-covariance and the selected block enter, which happens when the ambiguity set perturbs $\hat R$ in a low-rank or banded way. If no decoupling survives, we run the cutting-plane method on the full $N\times N$ inner problem. This is feasible for $N\le 256$ and is still a meaningful scale for sensor layouts.
The second component measures instability, which is the empirical heart of the claim. We bootstrap the covariance estimate $\hat R$ from finite channel snapshots and re-solve the nominal and robust problems. We then measure how far apart the resulting layouts are. Plain Hamming distance between selection masks overstates instability, because moving a sensor one patch over on a strongly correlated surface hardly matters. We therefore borrow the shifted Hamming distance from read mapping (Shifted Hamming distance: a fast and accurate SIMD-friendly filter to accelerate alignment verification in read mapping). That method aligns two bit strings while tolerating shifts of up to $e$ positions, using cheap bitwise AND operations. Here we apply it in two dimensions to the sensor masks. The genomics direction plays a small, honest role: it supplies a cheap, shift-tolerant distance between layouts. It lets us state "disorder chaos" for sensor placement as a measured quantity rather than a metaphor. The same bitwise distance also lets us scan thousands of candidate layouts from the cutting-plane runs and discard near-duplicates cheaply.
The third component is the channel estimator that consumes the sensor readings. We hold it fixed so that only the layout changes. The main estimator is linear MMSE prediction of unsensed elements from sensed ones. As a second estimator we use orthogonal matching pursuit followed by the complex-valued denoising network of the hybrid-IRS paper. This checks whether a learned reconstructor washes out layout differences.
Baselines
The naive reference is a uniform grid: $k$ sensors spread evenly across the panel, with equal power. It needs no statistics at all. The strongest competitor in practice is likely this grid or random placement, especially if spatial correlation is close to isotropic. Random placement, averaged over 50 draws, is the second naive reference.
The strongest simple optimizing method is greedy selection. It adds, one at a time, the element whose sensor most reduces nominal posterior error on $\hat R$, then sets power with the same convex solver. This is the standard approach in sensor-selection practice, and it is what a reviewer would reach for first. We found no reported result on Lacuna for greedy or convex-relaxation sensor placement on a reflecting surface, so its performance will be measured here, not quoted.
The strongest existing hybrid estimation pipeline is the one in Deep Denoising Neural Network Assisted Compressive Channel Estimation for mmWave Intelligent Reflecting Surfaces. It reports about a 4 dB improvement in normalized mean square error (NMSE) over matching pursuit alone at 10 dB SNR, but on its own unoptimized active-element layout. We will rerun that pipeline on its native layout and on ours.
Two references bracket the hardware budget. The upper one is a sensor on every element, as in Sensing RISs. The lower one is a fully passive surface estimated by tensor decomposition with alternating least squares. That method is shown to reach the Cramér–Rao bound, with nearly 10× lower NMSE than Khatri–Rao least squares at 20 dB SNR, in Channel Estimation for RIS-Empowered Multi-User MISO Wireless Communications. It is run at an equal pilot-time budget.
The key ablation solves the same mixed-integer problem with the same solver on nominal $\hat R$, with $\rho=0$. This isolates robustness from exact combinatorial optimization. A second ablation removes power allocation by forcing equal $p$. This isolates the long-only weighting from the selection itself.
All methods see the same training snapshots for $\hat R$, the same $k$, the same total pilot power, and the same downstream estimator. Hyperparameters ($\rho$, greedy stopping, network training) are tuned on a validation regime disjoint from test. Each method gets the same wall-clock budget, one hour per instance, matching the timeout used in the portfolio paper. Beating the baselines means lower worst-case and 90th-percentile NMSE across held-out regimes than both greedy and nominal-exact selection. The bar is at least 2 dB on the worst case, with average-case NMSE no more than 0.5 dB worse. A gain on the average alone does not count.
Why not just…?
Why not just use greedy selection on the estimated covariance? Greedy is near-optimal for the nominal problem when the error is close to submodular, so we do not expect to beat it on the training covariance. The proposal's bet is that nominal optimality is the wrong target. Benzaquen et al.'s analysis of long-only portfolios suggests that boundary-dominated, cardinality-limited designs have many near-tied optima and flip under estimation noise. Greedy inherits that fragility because it commits to the first element chosen on noisy $\hat R$. The instability measurement and the held-out-regime test show directly whether greedy layouts swing as much as nominal-exact ones. If greedy is both stable and within 1 dB of the robust method on the worst case, greedy wins, and we will say so.
Why not just spread sensors uniformly? This is the real threat. If the cascaded channel's spatial correlation is nearly translation-invariant, an even grid is close to robust by symmetry, and no optimization is needed. Robust selection should only help when correlation is anisotropic or non-stationary across the panel. Examples are near-field users, partial blockage, or users clustered in one region, which is also the regime the near-field survey (Channel Estimation for 6G Near-Field Wireless Communications: A Comprehensive Survey) highlights for extra-large surfaces. We sweep a controlled anisotropy parameter. The predicted outcome is a crossover: the grid matches robust selection at isotropy and falls behind as anisotropy grows. If no crossover appears, the method is not worth deploying, and that negative result is itself the answer to the open question both hybrid-sensing papers raise.
Why not just put a power detector on every element? The Sensing RIS paper shows this is pilot-efficient. But it notes that per-element detectors bring manufacturing and power-routing overhead, and it lists sparse sensing as open. The active-RIS literature documents wiring and thermal limits for dense active cells (Active RIS vs. Passive RIS: Which Will Prevail in 6G?; Zhang et al. (2023)). We report NMSE as a function of $k/N$ against the full-sensing reference. If 10% sensing comes within 1 dB of full sensing, then the extra hardware has a quantified cost and little benefit.
Why not stay fully passive and use tensor decomposition or blind beamforming? Tensor decomposition reaches the Cramér–Rao bound, but its pilot cost grows with $N$ (Opportunistic Beamforming Using an Intelligent Reflecting Surface Without Instantaneous CSI). Blind conditional-sample-mean beamforming (Blind Passive Beamforming for MIMO System) avoids estimation entirely but loses accuracy as $N$ grows. The equal-pilot-time comparison against alternating least squares, run at $N\in\{64,128,256\}$, settles where sparse sensing starts to pay.
Proposed Experiments
Channels come from the synthetic geometric and Rician models used across this literature, at 28 GHz and 3.5 GHz with SNR from −10 to 25 dB. Panels have $N\in\{64,128,256\}$ and sensor budgets $k/N\in\{5,10,20\}\%$. To create drift, we define user regimes as distinct regions and angular clusters, add a near-field setting, and add partial-blockage settings. The layout is chosen on regimes A–C and evaluated on unseen regimes D–F. As a correctness check on our solver port, we first reproduce the reported CPA+ result on sbm100 with $k=5$: global optimality in 2.7 s, where SCIP-SDP times out after 3,600 s (Cardinality-constrained Distributionally Robust Portfolio Optimization).
Metrics are NMSE (mean, 90th percentile and worst case across held-out regimes), the beamforming SNR obtained from the estimate, solver time, and layout instability. Instability is the shifted Hamming distance between masks across 100 bootstrap resamples of $\hat R$, at shift tolerance $e\in\{0,1,2\}$.
Three results would decide the question. The first is whether nominal-exact and greedy layouts show high bootstrap instability that robust layouts suppress. If none of the methods is unstable, the portfolio analogy fails at its root. The second is whether that stability turns into the stated worst-case NMSE margin on held-out regimes, and whether the margin grows with anisotropy, as the robustness argument predicts. The third is whether the advantage survives a learned downstream reconstructor, or whether the denoising network makes the layout irrelevant. A positive answer on all three gives the first principled answer to the sparse-sensing problem posed in Sensing RISs. A negative answer on the second, specifically a uniform grid that holds up, is equally publishable and more useful to hardware designers than another estimator.
References
- Deep Denoising Neural Network Assisted Compressive Channel Estimation for mmWave Intelligent Reflecting Surfaces
- Sensing RISs: Enabling Dimension-Independent CSI Acquisition for Beamforming
- Cardinality-constrained distributionally robust portfolio optimization
- Shifted Hamming distance: a fast and accurate SIMD-friendly filter to accelerate alignment verification in read mapping
- Channel Estimation for RIS-Empowered Multi-User MISO Wireless Communications
- Channel Estimation for 6G Near-Field Wireless Communications: A Comprehensive Survey
- Active RIS vs. Passive RIS: Which Will Prevail in 6G?
- Active RISs: Signal Modeling, Asymptotic Analysis, and Beamforming Design
- Opportunistic Beamforming Using an Intelligent Reflecting Surface Without Instantaneous CSI
- Blind Passive Beamforming for MIMO System