A multi-antenna base station can steer its transmission so that almost no energy reaches the users it is not serving. This is called spatial interference nulling. It works only as well as the station's knowledge of each user's channel, meaning how the radio signal is distorted on the way to that user. In practice those channels are estimated from short, overlapping pilot signals. They are uncertain, and the uncertainty is worst for distant cell-edge users. Standard precoders null the channel estimate and treat the leftover error as fixed and small, or as Gaussian with a known covariance. The error is then leaked as interference exactly onto the weakest users. So the station ends up either wasting power on users it cannot help or starving them. We propose to null the posterior: the full belief about each channel that remains after estimation. The station builds each user's expected-leakage matrix from posterior means and covariances produced by expectation propagation. Expectation propagation (EP) is a message-passing approximation that matches the mean and covariance of the true belief rather than hugging its peak. The station then balances user rates on those matrices.
Two things make this workable rather than decorative. First, EP's covariance estimates are routinely indefinite, meaning they have negative "variances" that come from subtracting one approximation from another. An indefinite covariance predicts negative interference, and a fairness loop fed that prediction diverges. Second, in a cell-free network, where many small access points jointly serve users, the covariance has a block-sparse structure: far-apart access points do not share information. That structure is what lets each access point compute its part locally. We therefore repair definiteness with a closed-form shrinkage that provably keeps the zero pattern. Standard eigenvalue clipping does not, and it would silently turn a distributed precoder into a centralized one.
Core Claim
The claim is narrow and testable. It applies when channels are estimated by data-aided joint estimation and detection under non-orthogonal pilots. In that regime the channel posterior is non-Gaussian, and the textbook linear-MMSE error covariance misstates the real uncertainty, especially for edge users. Replacing that covariance with EP second moments, made positive definite by support-preserving shrinkage, should raise the 5th-percentile user rate at equal sum-rate. It should do so while keeping signaling local to each access point. If the posterior is Gaussian, as in plain LMMSE estimation with orthogonal pilots, we expect no gain, because EP then reproduces the exact posterior. That null regime is part of the evaluation.
Research Program
The pipeline starts from the uplink. Cell-Free Massive MIMO Detection: A Distributed Expectation Propagation Approach splits EP between local linear estimation at access points and central symbol updates. It adds an iterative channel-estimation-and-detection (ICD) loop that reports up to 4.5 dB gain after four iterations under non-orthogonal pilots. Its authors state that extending this to downlink precoding is open. We reuse its ICD loop as the source of per-user channel posteriors, a mean $\hat{\mathbf h}_j$ and covariance $\mathbf C_j$. The difference is that we treat these as inputs to precoding, not merely as by-products of detection.
For user $k$ we form a weighted expected-leakage matrix and a leakage-aware precoder:
$$ \mathbf L_k(\boldsymbol\mu)=\sum_{j\neq k}\mu_j\left(\hat{\mathbf h}_j\hat{\mathbf h}_j^H+\mathbf C_j\right),\qquad \mathbf w_k\propto\left(\mathbf L_k(\boldsymbol\mu)+\mu_k\mathbf C_k+\tfrac{\sigma^2}{P}\mathbf I\right)^{-1}\hat{\mathbf h}_k . $$
Here $\mu_j\ge 0$ are dual weights (virtual uplink powers), $\sigma^2$ is noise power, and $P$ is the power budget. The weights $\mu$ are updated by the weighted SINR-balancing and uplink-downlink-duality iteration of Distributed Multicell Beamforming Design Approaching Pareto Boundary with Max-Min Fairness. That work reaches the Pareto boundary from a max-min point by exchanging only $K(B-1)$ scalars per iteration, where $K$ is the number of users and $B$ the number of base stations. It assumes bounded or perfect channel knowledge, and our change is to feed it posterior second moments. Its infeasibility problem, where hard rate floors cannot be met, appears here too. We handle it with a learned slack on the rate floor, in the spirit of Wireless Power Control via Counterfactual Optimization of Graph Neural Networks. That paper does this for scalar power control, and we carry it over to beamforming.
Definiteness enters at $\mathbf C_j$. Whenever an EP-derived $\mathbf C_j$ has minimum eigenvalue below a floor $\epsilon$, we apply the fixed-support shrinkage of Fixed support positive-definite modification of covariance matrix estimators via linear shrinkage:
$$ \Phi(\mathbf C_j)=\alpha\,\mathbf C_j+(1-\alpha)\mu_0\mathbf I . $$
The constants $\alpha\in(0,1]$ and $\mu_0>0$ are chosen in closed form so that $\lambda_{\min}\ge\epsilon$, and every zero off-diagonal block stays zero. That paper developed the shrinkage for thresholded statistical covariances and reports under one second at $p=3600$ against over two hours for competing sparse positive-definite methods. We use it inside a message-passing loop, where the preserved zeros encode which access points must communicate. We will also study the tension this creates. Shrinkage inflates uncertainty isotropically, so it spends nulling effort in directions where the posterior is already confident. We will characterize how $\alpha$ trades that waste against stability, and whether a per-block $\alpha$ helps.
Baselines
The naive reference is regularized zero-forcing on the channel estimate, meaning channel inversion with a diagonal regularizer, while ignoring estimation error. We run it under both vector normalization and matrix normalization. Large System Analysis of Power Normalization Techniques in Massive MIMO shows that vector normalization tilts toward sum-rate and matrix normalization toward fairness, so both ends of the tradeoff are represented. We found no reported result on Lacuna for the absolute sum-rate of RZF in these setups, so we will produce our own numbers on identical drops.
The strongest competitor is the same max-min/Pareto duality precoder from the Distributed Multicell Beamforming paper, driven by LMMSE channel estimates plus the analytical LMMSE error covariance. That is the principled robust design a wireless reviewer would expect. The paper reports approximately 0.5 bps/Hz worst-user-rate gain over sum-rate-optimal and Nash-bargaining beamforming, with convergence in about five iterations, in 2- and 3-BS clusters with 4 or 6 antennas and edge users at 350 m or more. We will reproduce that setup first, and we will count our reproduction as faithful only if it recovers the 0.5 bps/Hz gap. We also run WMMSE, an iterative weighted mean-squared-error method that is the standard sum-rate optimizer, as the throughput-oriented reference. Beating the strongest competitor means a higher 5th-percentile rate at no more than 2% sum-rate loss, or higher sum-rate at matched 5th percentile. It must hold on the same channel drops, the same pilot budget, the same number of ICD iterations feeding both methods, and the same outer-iteration cap and slack tuning grid.
Three ablations isolate the ingredients. The first replaces EP moments with a variational (mode-hugging) posterior covariance to test the mass-covering property. The second replaces fixed-support shrinkage with eigenvalue clipping to test the value of keeping the zero pattern. The third removes definiteness repair entirely and simply skips updates that go negative, which is standard EP practice.
Why not just…?
Why not just use the LMMSE error covariance, which is already the textbook robust precoder input? It is exact when pilots are orthogonal and estimation is linear, and in that regime we predict no gain. We report that regime explicitly. The case for EP rests on data-aided estimation under non-orthogonal pilots, where the ICD loop in the cell-free EP paper changes the channel belief using detected symbols, so the LMMSE covariance no longer describes the true error. This is a real threat, because a tuned scalar inflation of the LMMSE covariance might close the gap. We therefore include an "LMMSE × best inflation factor" baseline, tuned per SNR on validation drops. We also measure calibration directly: predicted versus realized leakage per user, binned by path loss. If inflated LMMSE matches EP on edge-user leakage calibration, the EP component is not justified and we will say so.
Why not just clip negative eigenvalues, or damp EP, as everyone does? Clipping works numerically, but it fills in the zero blocks, as the fixed-support shrinkage paper proves for spectral repair in general. It also costs a cubic-time decomposition per user per iteration. The Jaccard calibration study, Estimating Jaccard Index with Missing Observations: A Matrix Calibration Approach, reports that alternating spectral projections take over an hour at 10,000 samples. Damping and skipping reduce divergence but leave indefinite matrices in play whenever an update is accepted. The clipping and skipping ablations will show whether the zero-pattern argument matters in practice. We will count the fronthaul scalars each variant requires, alongside rate and runtime.
Why not just run WMMSE or a learned GNN power controller on the estimates? WMMSE optimizes a sum objective and, per the counterfactual-GNN paper, starves disadvantaged users in dense settings. That paper reports a 2740% 5th-percentile gain over WMMSE for scalar power control. That paper's GNN controls only scalar powers on single-antenna links, so it has no spatial nulling to exploit. Neither method uses channel uncertainty. We include WMMSE on estimates as a baseline, and we treat the slack mechanism as a component we borrow rather than a competitor.
Why not just use variational inference, which is easier to implement? Variational inference minimizes the reverse divergence and under-reports variance. For nulling, that means under-reported leakage onto edge users, which is exactly the failure we target. The VI-covariance ablation tests this directly. If it matches EP, the simpler method wins.
Proposed Experiments
We run three settings. The first is the multicell TDD setup of the Distributed Multicell Beamforming paper, used as a reproduction anchor. The second is the 3GPP urban cell-free setup of the distributed EP paper with non-orthogonal pilots, which is the main test. The third is the DeepMIMO O1 140 GHz ray-traced scenario used in Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO, with an 8×8 base-station array and a 6:2:2 split. It tests whether the gains survive realistic spatially correlated channels rather than Rayleigh fading.
Metrics are sum spectral efficiency, 5th-percentile and worst-user rate, and leakage calibration error by path-loss bin. We also track the fraction of iterations with indefinite covariances before repair, outer-loop divergence rate, per-access-point fronthaul scalars, and wall-clock time.
Two checks are decisive. The first is a pilot-overlap sweep from orthogonal to heavily reused pilots. The EP advantage over inflated LMMSE should be near zero at orthogonality and grow with contamination. A flat curve would falsify the core claim. The second is a distributed-versus-centralized comparison at matched rate. Fixed-support shrinkage should keep per-access-point signaling at the level of the distributed EP paper, which it reports as comparable to centralized linear MMSE. Clipping should force cross-access-point exchange. If both checks come out as predicted, the contribution is a concrete one: a precoder whose fairness gains come specifically from knowing where it is uncertain, delivered at the signaling cost of existing distributed designs.
References
- Cell-Free Massive MIMO Detection: A Distributed Expectation Propagation Approach
- Distributed Multicell Beamforming Design Approaching Pareto Boundary with Max-Min Fairness
- Wireless Power Control via Counterfactual Optimization of Graph Neural Networks
- Fixed support positive-definite modification of covariance matrix estimators via linear shrinkage
- Large System Analysis of Power Normalization Techniques in Massive MIMO
- Estimating Jaccard Index with Missing Observations: A Matrix Calibration Approach
- Normalization-Free Knowledge Distillation for Variable-Bit CSI Feedback in Massive MIMO