Sizing Reduced-RF Antenna Arrays for the Worst Users, Not the Average One

Large antenna arrays at millimeter-wave and sub-terahertz frequencies are made affordable by giving them far fewer radio-frequency chains than antennas. Each chain is the expensive, power-hungry path of converters, mixers and amplifiers behind an antenna. The workaround is "hybrid" beamforming: a cheap analog network of phase shifters does most of the spatial steering, and a small digital stage handles the rest. The field's main design rule says how many chains are "enough": with at least twice as many chains as data streams, a hybrid array can exactly reproduce any fully digital beamformer (Hybrid Digital and Analog Beamforming Design for Large-Scale Antenna Arrays). That result, and the algorithms built on it, assume perfect channel knowledge. They report average spectral efficiency, measured on synthetic clustered channel models. Real deployments have none of these properties. With fewer chains the base station also sees less of the channel when it estimates it. The users who matter for reliability are the unlucky few in blocked or scattering-poor locations, and those locations are exactly what synthetic models describe worst.

We propose to re-answer the hardware sizing question, "how many RF chains and phase shifters does this array need?", using a tail criterion. The criterion is Conditional Value-at-Risk (CVaR), the average outcome over the worst α-fraction of cases. We will measure it on ray-traced channels rather than stochastic models, with channel estimates that degrade realistically as chains are removed. The core claim has two parts. First, the chain count needed to meet a tail-rate target is larger than the count needed to meet a mean-rate target, and the synthetic models understate this gap. Second, a precoder optimized directly for tail rate can buy back part of that hardware, so part of the reliability cost of fewer chains can be paid in algorithm rather than silicon. If either part fails, the negative result is itself useful: it would mean the existing average-case sizing rules transfer safely.

Core Claim and What Is New

Existing reduced-RF work optimizes the average and assumes perfect channels. This covers the fully connected designs of Sohrabi et al., the manifold alternating-minimization algorithms of Yu et al., the fixed-phase-shifter switch networks of Yu et al., and the sub-array energy-efficient designs of He et al.. Two papers come closest to our question.

Lavi & Shlezinger make hybrid precoding robust. They maximize the rate under the worst channel error inside a fixed-radius ball, and they speed up that max-min solver by unrolling it into a learned network. They validate on Rayleigh and QuaDRiGa channels with 1,000 realizations. Their design treats all errors inside the ball as equally threatening and ignores errors outside it. It does not report tail statistics or architecture sizing.

Zhao et al. show a "digital liability": with coarse channel feedback, adding a digital zero-forcing stage to analog beamforming lowers average sum rate, because miscomputed interference nulls create interference. That finding comes from analytical approximations on synthetic channels and concerns only the mean.

Our proposal differs from both in three ways, all of which are needed to answer the sizing question. The outcome measured is a tail statistic of the worst-user rate. The channel estimation error is not an assumed ball or Gaussian; it comes from actually estimating the channel through $N_{RF}$ observations. The channels come from ray tracing via the DeepMIMO dataset, which Seo et al. used at 140 GHz, but only for fully digital feedback compression with average metrics.

Research Program

The program has three components. The first is a measurement pipeline. The second is a tail-optimized precoder, needed so that the hardware comparison is not biased by an average-case algorithm. The third is a transfer study that asks whether synthetic channels would have given the same engineering answer.

Measurement pipeline. We generate multi-user downlink channels from DeepMIMO scenario O1 at 140 GHz. We use the 8×8 base-station array of Seo et al. for continuity, and a larger 12×12 (144-element) array to match the size used by Yu et al. and Yu et al.. Groups of four to eight users are drawn from the user grid, with train, validation and test locations split spatially so that test locations are unseen.

For each architecture and each $N_{RF}$, the base station estimates the channel from pilots observed through its own $N_{RF}$-dimensional analog front end. We use a standard on-grid sparse estimator of the kind used for partially connected arrays (Low-Complexity On-Grid Channel Estimation for Partially-Connected Hybrid XL-MIMO). This couples the two effects of removing a chain: less steering flexibility and a worse channel estimate. The residual error distribution is recorded per location rather than assumed.

The architectures compared are fully connected phase shifters, sub-connected arrays, and fixed-phase-shifter switch networks. For each we sweep $N_{RF}$ and, for the switch network, the number of fixed shifters. We also sweep phase-shifter resolution, because rounding to 1–2 bits is known to cost about 5 dB on average (Sohrabi et al.), and its tail cost is unknown.

Tail-optimized hybrid precoder. Given a channel estimate $\hat H$, we draw $M$ plausible true channels $\hat H + \Delta_m$ from the empirical error distribution learned on training locations. We then solve the Rockafellar–Uryasev form of lower-tail CVaR for the worst-user rate:

$$\max_{F_{RF},\,F_{BB},\,\gamma}\; \gamma - \frac{1}{\alpha M}\sum_{m=1}^{M}\big(\gamma - R_{\min}(F_{RF}F_{BB};\,\hat H+\Delta_m)\big)^{+}$$

Here $R_{\min}$ is the smallest per-user rate under that channel draw, and $\gamma$ is an auxiliary threshold that ends at the α-quantile of that rate. $F_{RF}$ is constrained to unit-modulus entries, or to switch selections, and $F_{BB}$ to the power budget.

We solve this by alternating two steps. The analog step is a Riemannian gradient step on the complex circle, reusing the manifold machinery of MO-AltMin (Yu et al.). The digital and threshold updates are subgradient steps that act only on the currently worst draws. That hard-sample structure is exactly why CVaR equals averaging over the hardest examples (Rawat et al.).

The file raises two cautions that constrain the design. When α is small, plain CVaR gradients have high variance and can stall training; Curi et al. document truncated CVaR failing to pass 70% CIFAR-10 accuracy. Their adaptive sampler is therefore the fallback if variance blocks the α = 0.01 runs. Their no-regret guarantee holds only for convex problems, however, so on the unit-modulus manifold we claim nothing beyond empirical convergence and will report it.

By the duality of Chow et al., this objective equals maximizing expected rate under the worst reweighting of the error distribution, with density ratio at most $1/\alpha$. That gives α a physical reading: the precoder is hedged against error distributions up to $1/\alpha$ times more concentrated on bad draws than the one seen in training.

Transfer study. For each DeepMIMO configuration, we fit a Saleh–Valenzuela clustered model to the same cluster counts and angle spreads. We then rerun the entire sizing pipeline on it. This is the empirical-validation question stated concretely: would a designer who only had the synthetic model pick the same $N_{RF}$, the same architecture, and the same algorithm?

Baselines

The naive reference point is the standard pipeline. MO-AltMin or PE-AltMin factorizes a fully digital zero-forcing precoder computed on the estimated channel, treating the estimate as the truth (Yu et al.). No reported result was found on Lacuna for its tail rate on DeepMIMO. For the average, the file reports only that MO-AltMin is "nearly identical" to fully digital on Saleh–Valenzuela channels. Sohrabi et al. report hybrid with $N_{RF}=N_s$ "within ~1 dB of the fully digital limit" (Hybrid Digital and Analog Beamforming Design for Large-Scale Antenna Arrays).

The strongest existing robust method is the max-min design of Lavi & Shlezinger. Their reported figures are a 20× iteration reduction (5 versus 100 iterations) with sum rate "similar to or even better than conventional methods" on QuaDRiGa and Rayleigh channels. No reported tail or DeepMIMO result was found on Lacuna for it. To make the comparison fair, we run their mirror-prox solver to convergence rather than the unrolled version, since speed is not our question. It gets the same estimated channels and the same number of error draws for calibration. Its ball radius $\epsilon$ is tuned on validation locations to maximize the same CVaR metric, with the same number of tuning trials we give α.

The ablation that isolates our key ingredient is sample-average robust precoding. It uses the identical solver, error draws and initialization, but maximizes the mean worst-user rate over the draws instead of its lower α-tail. If CVaR does not beat it on test-location CVaR, the tail objective is not what matters.

Two further references bound the comparison. Pure analog beamforming follows Zhao et al., who report that with $B_2=2$ digital feedback bits pure analog beats hybrid zero-forcing across the SNR range. A fully digital precoder with the same estimation budget is the hardware ceiling. For switch networks, FPS-AltMin is the average-case reference (Yu et al.), reported near-optimal with 10 fixed shifters single-user and 30 multi-user, against over 2,300 adaptive shifters for double-phase-shifter designs.

Beating the strongest competitor means two things, both judged on held-out locations with confidence intervals from resampling locations. First, a higher 5% and 1% CVaR of worst-user rate at matched $N_{RF}$. Second, reaching a fixed tail target with strictly fewer RF chains. Either must hold with a mean-rate loss no larger than the tail gain.

Why not just…?

Why not just maximize average rate and let retransmissions absorb the tail? In latency-bound traffic, retransmissions cost time the budget does not have. The file also documents risk-neutral optimizers leaving a heavy failure tail that CVaR-constrained optimizers remove (Chow et al.), and ERM trading hard examples for mean loss (Curi et al.). Those findings come from finance, recommendation and vision, not from beamforming. The sample-average ablation above is the direct test here. It is a real threat: if errors are light-tailed on DeepMIMO, averaging over error draws may already protect the worst users. We will report that outcome as a finding that simplifies practice.

Why not just use the max-min error ball, which already works for hybrid precoding? A ball either misses the heavy tail of estimation errors, if small, or makes every user pay for rare errors, if large. It also has no way to know which locations produce large errors. Lavi & Shlezinger themselves note that their error patterns were random and that $\epsilon$ must be fixed in advance (Learn to Rapidly and Robustly Optimize Hybrid Precoding). Tuning $\epsilon$ on our own tail metric gives it every chance. If the tuned ball matches CVaR, the extra formalism is not justified.

Why not just drop the digital stage when channel estimates are poor? Zhao et al.'s result makes this a strong heuristic. We include a per-instance switch between analog-only and hybrid, triggered by the estimator's residual, as an additional baseline. The argument against relying on it is that analog-only cannot multiplex multiple streams (Sohrabi et al.). The CVaR precoder can also shrink the digital stage partially rather than all-or-nothing. Whether that partial hedging beats the hard switch in the tail is an open empirical question the experiments settle.

Why not just add RF chains until the tail is fine? That is exactly the cost the field is trying to avoid, and our sizing curves quantify it. The question is how many chains the tail demands, and how many a tail-aware algorithm saves. We express the result as chain and phase-shifter counts rather than watts, because the file gives no power model we could cite.

Proposed Experiments and Decisive Checks

The primary outputs are sizing curves: the minimal $N_{RF}$, and the minimal fixed-shifter count for switch networks, needed to reach a target 5% CVaR of worst-user rate. They are plotted against SNR and phase resolution, for each architecture and precoder, on DeepMIMO and on the fitted synthetic model. Mean sum rate, outage probability at a target rate, and the mean-versus-tail frontier traced by sweeping α are reported alongside.

Three checks decide the claim. The first compares the mean-target and tail-target chain counts on DeepMIMO. If they coincide, the first part of the claim fails. The second checks whether CVaR precoding lowers the tail-target chain count relative to the sample-average and tuned-ball baselines. If not, the second part fails. The third is the rank agreement between synthetic-model and DeepMIMO rankings of architectures and algorithms, computed separately for the mean and for the tail. We expect agreement on the mean and disagreement in the tail. Agreement in both would say synthetic models are adequate for this decision.

As a secondary check, we test whether the digital liability of Zhao et al. appears in the tail at estimation qualities where the mean shows no liability. We test this by conditioning on the worst 5% of locations. Wideband operation over DeepMIMO's 128 subcarriers, where the frequency-flat analog stage causes beams to squint, is a natural extension. It is deferred until the narrowband results are in.