Backing Off Power Amplifiers Under Maximum-Entropy Channel Uncertainty

A multi-antenna base station drives each antenna with its own power amplifier, and each amplifier has a hard output ceiling. A recent result shows that when the transmitter knows the channel exactly, every efficient beamformer must run every amplifier at its ceiling (Per-antenna power constraints: constructing Pareto-optimal precoders with cubic complexity under non-negligible noise conditions). Real transmitters never know the channel exactly. The error in their channel estimate also differs from antenna to antenna, and how large it is often goes unmeasured. This happens because pilot quality depends on distance and blockage, and because the channel drifts at a rate set by user motion. We propose to model that error with the maximum entropy principle, which picks the least-committed probability distribution consistent with what is actually known, here an upper bound and possibly a mean on each antenna's error energy. We then design beamformers that respect each amplifier's limit and protect users' worst-case rates under this model.

The claim is that the full-power theorem breaks in a specific and predictable way. When an antenna's error energy is uncertain rather than merely nonzero, the maximum-entropy error distribution is heavier-tailed than a Gaussian with the same variance. Under that distribution, the rare-event interference a user suffers grows with the power placed on the poorly known antennas. The rate-protecting beamformer should therefore deliberately run those amplifiers below their ceiling, and run well-known antennas at full power. The method decides which amplifiers to back off and by how much. It also predicts when the effect vanishes: if every antenna's error energy is known exactly, the maximum-entropy model collapses to an ordinary Gaussian error, and the design reduces to known robust beamforming.

Background and the gap

Three lines of work meet here without touching. The first is per-antenna precoding, meaning the choice of how much signal each antenna sends to each user. Petrov et al. build Pareto-optimal precoders with cubic complexity and prove the full-power property (their Proposition 1), but assume perfect channel knowledge and name estimation error and channel aging as open problems (Per-antenna power constraints). Related methods solve the same constraint family with different tools. One handles rate guarantees through a weighted minimum-mean-square-error reformulation split by an alternating direction method (MIMO Precoding Design with QoS and Per-Antenna Power Constraints). Another uses a tight semidefinite relaxation with fast first-order dual solvers (Joint Beamforming and Compression Design for Per-Antenna Power Constrained Cooperative Cellular Networks). A third uses uplink-downlink duality for distributed access points (UL-DL duality for cell-free massive MIMO with per-AP power and information constraints). These papers supply our inner solvers and our test scenarios. None of them asks how channel uncertainty interacts with the per-antenna structure.

The second line is robust transceiver design. Kang and Lim handle estimation error and spatial correlation under general power constraints, including per-antenna ones, by alternating Wiener receive filtering with a second-order cone program for the precoder (Robust Transceiver Design for Correlated MIMO Interference Channels in the Presence of CSI Errors under General Power Constraints). They assume a Gaussian or norm-bounded error with known size. This is our main competitor. Our proposal differs only where the error size itself is uncertain.

The third line is maximum-entropy channel modeling. Guillaud, Debbah and Moustakas show that when total channel energy is unknown but bounded, the maximum-entropy channel is no longer Gaussian, and its 1% outage capacity in a 4×4 link at 10 dB is about 3.9 nats rather than the 4.5 nats of the Gaussian model (Maximum Entropy MIMO Wireless Channel Models). This motivates our work, since it shows the tails matter. It also limits our scope in two ways. That paper constrains only total energy, and it never designs a transmitter. We move the energy constraint from the whole array to each antenna, and apply it to the estimation error rather than to the channel.

One tempting version of this intersection is to define the robust uncertainty set by maximum entropy subject to the error's mean and covariance. That version adds nothing. The maximum-entropy distribution under first and second moments is exactly the Gaussian already used by Kang and Lim. The idea only does work when the constraints are inequalities on unknown per-antenna energies, so that the solution is a mixture rather than a single Gaussian.

Core Claim

Write the true channel as the estimate plus an error whose entry on transmit antenna $i$ has random energy $\varepsilon_i$. What the transmitter knows is $0 \le \varepsilon_i \le \bar\varepsilon_i$, and possibly the mean $\mathbb{E}[\varepsilon_i] = m_i$. Applying maximum entropy at both levels gives the following model. Conditioned on $\varepsilon_i$, the error is a circular Gaussian, because that is the maximum-entropy choice for a fixed energy. Each $\varepsilon_i$ is independent, and is uniform on $[0,\bar\varepsilon_i]$ if only the bound is known, or truncated exponential if the mean is also known; this follows the bounded-energy construction of Guillaud et al. applied per antenna.

Let $w_{ij}$ be the precoder weight from antenna $i$ to user $j$, and let $p_i=\sum_j |w_{ij}|^2 \le P_i$ be the power antenna $i$ actually emits. Given the per-antenna error energies, the extra interference leaking to user $k$ has conditional mean

$$ \mathbb{E}\left[I_k \mid \varepsilon\right] = \sum_{i} \varepsilon_i \sum_{j\ne k} |w_{ij}|^2 , $$

so its tail is governed by a weighted sum of independent bounded random scales, with antenna powers as the weights. A uniform power back-off cannot shape this tail efficiently, because it lowers useful signal on well-known antennas along with harmful leakage on poorly known ones. Our design problem maximizes a weighted sum of users' $q$-outage rates, meaning the rate each user achieves with probability at least $1-q$ under the maximum-entropy model, subject to $p_i \le P_i$.

We expect a threshold structure. Antennas with small or well-known $\bar\varepsilon_i$ stay saturated, in line with Petrov's Proposition 1. Antennas with large, uncertain $\bar\varepsilon_i$ fall below $P_i$.

Research Program

The theory part extends Proposition 1. We will characterize when $p_i = P_i$ remains necessary at the outage-optimal point, as a function of $\bar\varepsilon_i$, the outage level $q$, and the noise level. We will prove that the property survives when all $\bar\varepsilon_i$ are equal and known, which is the Gaussian collapse case. We will also give a sufficient condition under which back-off on antenna $i$ strictly improves the outage objective.

The algorithm part uses two nested loops. The inner loop fixes per-antenna power targets $p_i$ and computes the precoder with Petrov's cubic-complexity parametric method, with the conditional mean interference added to the noise term. The outer loop updates the $p_i$ by projected gradient on the outage objective. The outage objective is cheap to evaluate because only the $\varepsilon_i$ need integrating; conditioned on them, interference is a Gaussian quadratic form. In cell-free networks we tie $\varepsilon_i$ across the antennas of one access point, which leaves only 16 scale variables in the target scenario. Gradients come from sampling those scales, with a Chernoff-type bound as a cheaper deterministic surrogate that we will also test.

The modeling part estimates $\bar\varepsilon_i$ and $m_i$ from information a base station actually has: per-antenna pilot signal-to-noise ratio, plus a Doppler range implied by the scenario's user speed distribution. This gives the aging component, which is the part of the error the transmitter cannot measure.

Baselines

The naive reference is the perfect-CSI-assuming parametric precoder of Petrov et al. applied directly to the channel estimate, with all amplifiers saturated. Under perfect CSI on QuaDRiGa Berlin urban macro non-line-of-sight, that paper reports a mean throughput of "7.26" against "5.56" for zero-forcing in an 8×3 system. We will reproduce both numbers first as a sanity check. We have no reported result on Lacuna for this precoder under imperfect CSI, so that is a number we will generate ourselves.

The strongest existing competitor is Gaussian robust design under per-antenna constraints, following Kang and Lim. It is strongest because it already treats the same constraints and estimation error, and differs from us only in how the error is modeled. We will run it two ways. In the first it uses the same alternating Wiener/cone-program solver as the original paper. In the second it is re-implemented with our inner solver and per-antenna Gaussian error matched to the mean energies $m_i$, so that solver quality is not confounded with the error model. No reported numerical result on Lacuna was found for this method on our benchmarks.

We also include the first-order method of Joint Beamforming and Compression Design as an exact solver on the small instances, to verify that our inner loop is not leaving performance behind.

The key ablation keeps our full method but forces $p_i = P_i$. This isolates whether gains come from selective back-off or merely from the robust interference term. A second ablation keeps back-off but replaces the maximum-entropy mixture with the matched Gaussian. This isolates the maximum-entropy ingredient.

All methods get the same channel estimates, the same per-antenna information ($\bar\varepsilon_i$, $m_i$), the same iteration budget, and the same tuning of regularization or noise inflation. Each method gets the grid search over inflation scale that we use for our own sampling budget.

Beating the Gaussian baseline means a higher 5%-outage per-user rate at equal or lower radiated power, with confidence intervals over channel drops excluding zero. It must hold in the heterogeneous-uncertainty regimes and must not lose more than noise-level amounts in the homogeneous regime.

Why not just…?

Why not just solve with sum power and clip or scale per antenna? Rescaling each row after a sum-power design breaks the interference cancellation the design relied on. MIMO Precoding Design with QoS and Per-Antenna Power Constraints reports that such heuristics lose sum rate and violate rate guarantees relative to joint optimization. We include clipping only as a reference curve. It also backs off amplifiers for the wrong reason: because the design overshot their limit, not because their channel is poorly known.

Why not just use the Gaussian error model with inflated noise? This is the real threat. When per-antenna error energies are well known, our model provably reduces to it. Its weakness should appear only when error energy is uncertain and differs across antennas, because the Gaussian has lighter tails than the mixture at the same variance. That is the mechanism behind the 4.5 versus 3.9 nat gap in Maximum Entropy MIMO Wireless Channel Models. Our experiments sweep uncertainty heterogeneity from zero upward. If the Gaussian baseline with tuned inflation matches us at 5% outage across that sweep, the maximum-entropy ingredient is not worth its cost, and we will report it that way.

Why not just back off all amplifiers uniformly by a tuned margin? A uniform margin is one scalar and is easy to deploy, so we tune it on the same validation drops. It cannot distinguish a nearby, well-estimated access point from a distant, aging one. Our prediction is that it trades too much signal on the good antennas. The per-antenna saturation histogram, compared against the uniform rule at matched outage, will show whether the selectivity matters.

Why not just use a worst-case ellipsoidal error bound? A norm ball sized to the largest $\bar\varepsilon_i$ protects against an error realization that the stated knowledge makes very unlikely. It therefore tends to over-shrink all beams. We include it, sized per antenna, and expect it to win only at very small $q$.

Proposed Experiments

The first testbed is QuaDRiGa Berlin urban macro non-line-of-sight, used in Petrov et al., at 8×3 and 192×24. We add pilot-based estimation and user-speed-dependent aging, so that $\bar\varepsilon_i$ varies with geometry.

The second is the 3GPP urban microcell cell-free layout of Miretti et al.: 16 access points with 4 antennas each, 16 users, and service by the 4 strongest access points. That paper reports that a minimum rate of "2.5 b/s/Hz" is supported for all users with per-access-point limits. We will report the minimum rate our method sustains at 5% outage once estimation error is introduced. Path loss and penetration parameters follow 5G 3GPP-like Channel Models, which makes per-access-point uncertainty naturally heterogeneous.

The metrics are 5%- and 1%-outage per-user rate, mean throughput, minimum user rate, total radiated power, fraction of amplifiers at saturation, and runtime. The 192×24 case uses Petrov's reported 15-minute desktop solve as the budget reference.

Three checks would decide the question. The first is the homogeneity sweep. As the spread of $\bar\varepsilon_i$ goes to zero, our method should coincide with the Gaussian baseline, and saturation should return to 100%. Failure here would mean an implementation bug or a wrong theory. The second is the heterogeneous regime, meaning cell-free with mobile users. Here our method should beat the matched-Gaussian ablation at 5% outage while using less radiated power. Backed-off amplifiers should be concentrated on the access points with the largest uncertainty bounds. The third is the forced-saturation ablation, which should lose most of the gain. If it does not, the benefit comes from the interference model rather than from breaking the full-power property, and the central claim is wrong.