Deconvolving the Passive Surface: Channel Estimation for Reflecting Surfaces as a Masked-Heterogeneity Problem

At millimeter-wave frequencies a radio link fails whenever something stands between transmitter and receiver. A reconfigurable intelligent surface can rescue the link. This is a wall-mounted panel of hundreds of cheap elements, each adding a programmable phase shift to the wave it reflects. Tuned correctly, the reflections add up in phase at the user and form a "virtual line of sight" around the obstacle. The catch is that the elements are passive, so they cannot measure anything themselves. The receiver only ever sees one number per pilot slot: the sum of all elements' contributions under whatever phases were set. Recovering the per-element channels from these sums takes pilot overhead that grows with element count times user count. The research file lists this as the first open problem of the area.

Genomics has a long-studied version of the same structure. Bulk RNA sequencing returns one aggregate profile per tissue sample, and averaging over cells hides rare subpopulations such as drug-resistant stem-like clones. The answer there was not to estimate every cell's profile. It was to use a reference atlas as a prior and to infer which components are present, sampling explicitly so the result stays consistent with the observed aggregate. This proposal carries that lesson to the reflecting surface. The receiver's pilot sums are a "bulk measurement" of a channel drawn from a small number of geometric regimes. The pilot budget should go toward identifying the regime, not toward estimating $N$ complex coefficients. Estimators that average over regimes fail in a specific, predictable way, the same way bulk averaging hides minority cell states.

Core Claim

Write the per-pilot observation as $y_t = v_t^{H} g + n_t$. Here $g \in \mathbb{C}^N$ is the cascaded base-station-to-surface-to-user channel seen through each of the $N$ elements, $v_t$ is the unit-modulus phase configuration applied in slot $t$, and $n_t$ is noise. Given an estimate of $g$, the surface sets each phase to cancel the phase of $g_n$, and the received power is the beamforming gain $|v^{H}g|^2$.

In a blocked indoor or street scene, $g$ is not one smooth distribution. It is a mixture of a few geometric regimes. The surface may see the user directly. A moving blocker may cut the surface-to-user path. The user may sit in one of several reflection lobes. A plug-in estimator that uses the posterior mean of $g$ averages these regimes. Because the regimes have different phase patterns, the averaged phases cancel, and the resulting configuration loses most of the coherent gain.

The more careful conventional choice is to maximize expected power, $\mathbb{E}|v^Hg|^2$, whose solution is the top eigenvector of $\mathbb{E}[gg^H]$. This does not cancel the regimes, but it does lock onto the dominant one and systematically abandons users in rarer regimes. That is the radio analogue of a bulk profile that is accurate for the majority cell type and blind to the minority clone.

The claim is that a receiver which keeps an explicit multimodal posterior over $g$ will recover near-full surface gain with a number of pilots that scales with the number of plausible regimes rather than with $N$. The posterior is built from a site-specific "reference library" of channels, in the role a single-cell atlas plays in deconvolution. The receiver chooses each next configuration to separate the remaining regimes, and it commits only once the posterior concentrates. The gain over averaging estimators should be concentrated in the lower tail of users. Headline averages such as the ergodic coverage and sum-rate figures of Stochastic Geometry Analysis for Distributed RISs-Assisted mmWave Communications would therefore mask it, so evaluation must be per-user and per-regime.

Where the Pieces Come From

The surface-side problem and its existing solutions come from the virtual-line-of-sight literature. Joint active and passive precoding, as in Intelligent Reflecting Surface-Assisted Millimeter Wave Communications: Joint Active and Passive Precoding Design, alternates between optimizing the base-station precoder and the surface phases. It assumes the cascaded channel is known, which is exactly what this proposal removes. Two learning methods skip channel estimation by regressing the answer directly. Phase Configuration Learning in Wireless Networks with Multiple Reconfigurable Intelligent Surfaces maps user coordinates to phases with a multilayer perceptron. Reconfigurable Intelligent Surface-Assisted Aerial-Terrestrial Communications via Multi-Task Learning uses a shared network with a classification head and a regression head. Both are trained as predictors of a single best output. When the same input is consistent with several regimes, a squared-error regressor learns their conditional mean, which is the failure mode this proposal is built around. They therefore motivate the work: their success in unambiguous settings, and their architecture, set the bar to beat in ambiguous ones.

The closest existing work is the active-sensing line: Active Sensing for Localization with Reconfigurable Intelligent Surface, Codebook Learning for Active Sensing with Reconfigurable Intelligent Surface, and Learning Beamforming Codebooks for Active Sensing With Reconfigurable Intelligent Surface. These train a recurrent network that observes past pilot sums, picks the next surface configuration, and outputs a final beamformer, all trained end to end. They already adapt measurements sequentially, so the proposal does not claim adaptivity as new. What differs is that the proposal keeps an explicit, inspectable posterior built from a reference library. That posterior gives three things an end-to-end policy does not give by construction: a principled stopping rule (stop piloting when one regime dominates), calibrated outage risk, and reuse under a new objective or a new blockage mix without retraining the policy. Whether these matter in practice is the main empirical question, and the experiments are designed so this baseline can win.

The deconvolution side supplies the diagnosis and part of the method. Two papers supply the diagnosis that averaging hides minority states: Genome 2011 and the resistance study Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties. In the latter, a rare stem-like population is indistinguishable from a uniform shift in bulk.

Count Bridges enable Modeling and Deconvolving Transcriptomic Data supplies two techniques we reuse. The first is expectation-maximization with the per-component profiles treated as latent variables. The second is projection-guided sampling, which forces every generated sample to reproduce the observed aggregate exactly. In our setting this means every posterior sample of $g$ must satisfy the pilot equations up to noise. Count Bridges' Poisson birth-death machinery is specific to integer counts and does not transfer, since channels are complex-valued.

Cross-modal transfer learning for mapping bulk transcriptomes at cellular level (POPPY) contributes the idea of querying a reference atlas by attention to attribute an aggregate to components. It also contributes its key known weakness: components absent from the reference cannot be recovered. That weakness constrains our design. We must test regimes missing from the library and include an explicit "none of the above" component whose detection triggers fallback to exhaustive beam training.

Research Program

Reference library. We first build, for each simulated site, a library of cascaded channels sampled under the Line Boolean obstacle model of the stochastic-geometry paper. Obstacles are random line segments, and surfaces and users are placed as Poisson points. We run it at 28 GHz rather than the 3 GHz used there, which closes one of the file's listed gaps. The geometric regime of each sample is labeled from the geometry: whether each link is blocked, and which reflection lobe the user occupies.

Prior. On the library we fit two priors of increasing capacity. The first is a mixture of complex Gaussians, one per regime, each with a low-rank covariance in the angular domain. The second is a conditional generative model that produces per-element channels, sampled with the projection-guided step from Count Bridges so that every sample is consistent with the pilots observed so far.

Pilot policy. Configurations come from a quantized codebook, at 1-bit and 2-bit resolution to respect the hardware limit the file highlights. Each next configuration greedily maximizes the expected reduction in posterior regime entropy. The receiver stops when the posterior probability of the leading regime exceeds a threshold, or when an estimate of outage under the best configuration falls below target.

Final configuration. The commit step maximizes the posterior expected rate, not power. This keeps it from sacrificing the minority regime for the sake of average power.

Analysis. We will derive a simple bound on the coherent-gain loss of the posterior-mean plug-in as a function of the angular separation between regimes. The bound makes the averaging failure quantitative and predicts when the proposal should and should not help. When regimes overlap in angle, averaging is harmless, and we expect no advantage.

Baselines

The naive reference point is exhaustive least-squares estimation. The receiver uses $N$ orthogonal pilot configurations, estimates every element's channel, and applies the cancelling phases. This gives the full-overhead upper bound on gain and the cost we are trying to cut. A second naive point is a fixed random or specular configuration with no pilots, which marks how much the surface helps at all.

The geometry baseline is hierarchical beam training over a quantized angular codebook, which needs on the order of $\log N$ pilots in a single-path channel. It is the obvious practical competitor and is run with an identical per-slot pilot budget.

The regression baselines are the position-to-phase multilayer perceptron of Phase Configuration Learning in Wireless Networks with Multiple Reconfigurable Intelligent Surfaces and a shared-trunk multi-task network patterned on Reconfigurable Intelligent Surface-Assisted Aerial-Terrestrial Communications via Multi-Task Learning. Both are given the same library as training data and the same pilot budget.

On the multi-task network's own simulator, the authors report throughput rising from 0.06 Mbps to 28 Mbps under blocked line of sight, and computation time of about 0.017 ms versus 3.8 ms for spatial branch-and-bound with 8 users. Those numbers are measured against a no-surface or exact-optimizer reference, not against active sensing. They cannot be compared directly to our metric, so we re-run the architecture on our scenes.

The strongest competitor is the end-to-end learned active-sensing policy of Learning Beamforming Codebooks for Active Sensing With Reconfigurable Intelligent Surface. It already chooses pilots adaptively from past observations and directly optimizes the final gain. We found no reported result on Lacuna for this method's beamforming gain at a given pilot count, so we re-implement it on our scenes with the same library, pilot budgets, phase quantization, and hyperparameter search budget as our method. The search budget is counted in GPU-hours and in validation configurations tried.

Beating it means two things at matched pilot count. First, a statistically significant gain in the 5th-percentile user rate on in-distribution scenes. Second, a larger gain when the regime proportions at test time differ from training. At equal pilots, it should not lose more than a small margin in mean rate.

Two ablations isolate the key ingredient. The first replaces the multimodal posterior with a single Gaussian of the same total covariance (posterior-mean plug-in and top-eigenvector variants). This removes multimodality while keeping everything else. The second replaces the generative prior with the Gaussian mixture. If the mixture matches the generator, the contribution is regime identification, not generative modeling, and we will say so.

Why not just…?

Why not just do hierarchical beam search? In a channel with one dominant reflected path, beam search is close to optimal, and we expect it to match our method there. It degrades when several surface lobes or partially blocked paths produce comparable energy, because each binary split compares summed energies that the regimes can confuse. The per-regime breakdown in our experiments will show where this crossover happens. If beam search matches us across all regimes at equal pilots, the averaging argument has no practical bite at these array sizes, and the proposal's main claim fails.

Why not just train the end-to-end active-sensing network and let it learn the posterior implicitly? This is the real threat. A recurrent policy trained on enough data can in principle discriminate regimes without any explicit posterior. Our bet is narrower. When the mix of blockage regimes shifts at test time, or when the operator changes the objective from mean rate to outage, the explicit posterior adapts by reweighting the prior or changing the commit rule, while the trained policy must be retrained. The distribution-shift and objective-swap experiments test exactly this. If the trained policy stays robust under them, the defensible contribution shrinks to calibrated stopping and interpretability, and we will report it that way.

Why not just regress phases from user position, with no pilots at all? Position is cheap and often good enough, and the position-based multilayer perceptron needs zero pilots. But position does not reveal a moving blocker or a body near the handset. In those cases one position maps to several regimes, and a squared-error regressor outputs their average. Our scenes contain dynamic blockers, so we can measure how often this happens and how much gain it costs. We will also test a hybrid that uses position to set the prior and then spends a few pilots, which may well be the best practical system.

Why not just use a Gaussian-mixture prior with hypothesis testing? It may be enough, which is why it is an ablation rather than a strawman. If it matches the generative prior, we adopt it as the recommended method, because it is cheaper and analyzable.

Proposed Experiments

All experiments use the 28 GHz Line Boolean simulator described above. Surfaces have 64, 256, and 1024 elements, phases are quantized to 1 or 2 bits, and blockers are static or dynamic. Reported metrics are beamforming gain relative to perfect-channel optimum as a function of pilot count, mean and 5th-percentile user rate, outage probability, and the calibration of predicted outage against observed outage.

The first decisive check tests the averaging pathology itself. On scenes with two well-separated regimes, the posterior-mean plug-in should lose gain roughly as the derived bound predicts, while the multimodal posterior should not. On scenes with overlapping regimes, the gap should vanish. If the predicted pattern does not appear, the core mechanism is wrong.

The second check is scaling. If regime identification is what matters, the pilots needed to reach 90% of optimal gain should grow slowly with $N$ at fixed regime count. Least squares, by contrast, needs a number of pilots that grows linearly with $N$.

The third check is distribution shift against the end-to-end active-sensing policy. We train every method on one blockage density and test on denser and sparser ones, mirroring the dense-versus-low-blockage contrast in the stochastic-geometry analysis, where reported sum-rate gains were 1.91x and 1.15x respectively.

The fourth check targets the out-of-reference regime, the POPPY-style failure. We hold one regime out of the library entirely and measure two things: whether the "none of the above" component detects it, and how much gain is lost before fallback to beam training.

A positive result on the first and third checks, together with honest reporting of where beam search and the trained policy tie, would establish that the bulk-deconvolution view gives a reflecting-surface receiver something it does not already have.