In crowded wireless networks, a transmitter's signal is usually treated as circularly symmetric noise at every receiver it does not serve. Improper signaling breaks that symmetry on purpose. A transmitter concentrates its power along one direction in the complex plane, which leaves the perpendicular direction quiet, and a nearby victim receiver can then listen along that quiet direction. Small studies show large gains in interference-limited settings. On a 3-cell, 18-user network, energy efficiency rises from about 8.2 to 9.5 bits/Joule (Improper Gaussian Signaling for Broadcast Interference Networks). In rank-deficient downlinks, which have more users than base-station antennas, throughput rises by 15–25% (Regularized Zero-Forcing Aided Hybrid Beamforming for Millimeter-Wave Multiuser MIMO Systems). Every one of these results, however, comes from a centralized solver that collects all channels in a handful of cells. The broadcast-interference paper itself names scaling as unresolved.
We propose a distributed optimizer for improper signaling across many cells, borrowing the decomposition machinery used for large power grids. The central observation is that improper interference is not described by a single number. A victim's rate depends on the interference power and also on its pseudo-covariance, the part that records how elongated and how rotated the interference is. Distributed schemes for ordinary (proper) signaling let neighbors exchange a scalar interference price. Our claim is that improper signaling needs an augmented price: a real power price plus a complex "pseudo-price" per coupled link. With it, base stations that only talk to their interference neighbors should nearly match centralized improper optimization, at a per-iteration cost that grows linearly with network size. A useful side effect is that the pseudo-prices tell each cell whether being improper is worth it at all. That turns the open proper-versus-improper switching question raised by A Journey From Improper Gaussian Signaling to Asymmetric Signaling into something the optimizer answers locally.
Core Claim
Consider user $k$ treating interference as noise. Its achievable rate under improper Gaussian signals is
$$R_k = \tfrac12 \log_2 \frac{\det \underline{\mathbf R}_{y_k}}{\det \underline{\mathbf R}_{\iota_k}},$$
where $\underline{\mathbf R}_{y_k}$ is the augmented covariance of user $k$'s received signal and $\underline{\mathbf R}_{\iota_k}$ is the augmented covariance of its interference plus noise. "Augmented" means the covariance of the signal stacked with its complex conjugate, so the matrix carries both the covariance and the pseudo-covariance. Each interfering cell $j$ adds a term $\underline{\mathbf Q}_{jk}$ to $\underline{\mathbf R}_{\iota_k}$, and this term depends only on cell $j$'s widely linear beamformer. A widely linear beamformer is one that processes a signal and its conjugate as separate inputs. Because significant interference only comes from nearby cells, the coupling terms $\underline{\mathbf Q}_{jk}$ form a sparse interference graph.
We treat each $\underline{\mathbf Q}_{jk}$ as a shared variable. Cell $j$ holds a local copy of what it emits toward user $k$, and cell $i$, which serves $k$, holds a local copy of what it will tolerate. Consensus ADMM (the alternating direction method of multipliers, which splits a coupled problem into local solves plus agreement updates) forces the two copies to agree. The dual variable attached to each $\underline{\mathbf Q}_{jk}$ is the augmented price. Its off-diagonal block, the pseudo-price, is what tells cell $j$ which direction in the complex plane its leakage should occupy. If a link's pseudo-price stays at zero, the ordinary scalar price already captures the tradeoff for that link. If every link touching a cell has zero pseudo-price, that cell's best response is proper signaling.
Relation to Prior Work
The improper-signaling literature supplies the objective and the local solvers. The centralized path-following successive convex approximation (SCA, which repeatedly maximizes a concave lower bound of the rate) in Improper Gaussian Signaling for Broadcast Interference Networks handles max-min rate and energy efficiency over augmented covariances. It converges in 17–32 iterations but needs global channel knowledge. We reuse its concave quadratic rate bounds inside each cell's local subproblem and remove the global collection step.
The closed-form alternating descent in Maximizing the Geometric Mean of User-Rates to Improve Rate-Fairness: Proper vs. Improper Gaussian Signaling and the polar-form Frank-Wolfe updates in Regularized Zero-Forcing Aided Hybrid Beamforming for Millimeter-Wave Multiuser MIMO Systems show that improper designs can avoid heavy semidefinite solvers. Both are single-cell with 4–10 users. They motivate cheap local updates but offer no way to coordinate cells.
The two-user analyses comparing improper signaling with time-sharing, Two-User SIMO Interference Channel With Treating Interference as Noise: Improper Signaling Versus Time-Sharing and Improper Signaling versus Time-Sharing in the Two-User Gaussian Interference Channel with TIN, constrain the design. Impropriety is not always the best way to enlarge the rate region, so our evaluation must include orthogonal scheduling.
The large-scale optimization literature supplies the decomposition. ADMM for distribution feeders (Scalable Optimization Methods for Distribution Networks With High PV Integration) and dual decomposition into a coordinator plus local agents (Information-constrained Optimal Control of Distributed Systems with Power Constraints) are the templates we follow. The second paper also warns that dual variables multiply with coupling constraints and slow convergence across wide networks. Our pseudo-prices roughly double that count, so convergence speed is a first-order risk rather than a detail.
Research Program
We will build three things. The first is the algorithm. It is a convexified consensus ADMM: in each round, every cell solves the SCA-linearized version of its local weighted rate or max-min problem over its widely linear beamformers and its copies of incoming and outgoing $\underline{\mathbf Q}_{jk}$. Neighbors then exchange augmented prices and residuals. The rate objective is nonconvex, so we do not claim global convergence. We will prove the weaker property that fixed points are stationary points of the centralized SCA problem, and we will measure convergence empirically.
The second is a message-reduction variant. For single-antenna receivers, $\underline{\mathbf Q}_{jk}$ has only three real parameters (a power and a complex pseudo-power). For multi-antenna receivers we will compress it to its dominant augmented directions, and we will measure how much rate the compression costs.
The third is the analysis. We will characterize when the optimal pseudo-price at a link is zero, which is the local criterion for staying proper. We will then test whether that criterion tracks simple local statistics such as the users-to-antennas ratio in a cell's interference neighborhood.
Baselines
The naive reference point is proper signaling with independent per-cell beamforming and no coordination. It shows how much any coordination buys.
The standard proper competitor is distributed proper-signaling weighted minimum mean square error beamforming (WMMSE) with scalar interference pricing, run over the same interference graph and the same message budget. Beating it shows that impropriety survives decentralization.
The strongest existing improper method is the centralized path-following SCA of Improper Gaussian Signaling for Broadcast Interference Networks. On its 3-cell model with 4 transmit antennas it reports about 9.5 bits/Joule for improper versus about 8.2 bits/Joule for proper signaling. We will first reproduce that setup: 500 m hexagonal cells, 6 users per cell, 4–6 transmit and 2 receive antennas, a 2 GHz carrier, Rayleigh fading, and path-loss exponent 3.1. Then we run it on every larger layout until its runtime or memory becomes prohibitive. It serves as an upper reference, not as something to beat on objective value. Our success criterion against it is recovering at least 90% of its improper-over-proper gain, in max-min rate and in energy efficiency, while per-cell compute and message size stay flat as the network grows.
The key ablation keeps local widely linear beamforming but lets cells exchange only the scalar power price, with no pseudo-price. If this ablation matches the full method, our central claim is false.
A cluster baseline runs the centralized SCA inside fixed 3-cell clusters and treats inter-cluster interference as proper noise. Every method gets the same channel draws, the same initialization protocol, and the same wall-clock budget on the same hardware.
Why not just…?
Why not just rewrite the improper problem in real coordinates and run ordinary WMMSE? Stacking real and imaginary parts turns widely linear design into a real-valued problem of twice the size, and a real-composite WMMSE could plausibly handle it centrally. That is a genuine threat to any claim that improper signaling needs new local solvers, so we do not make that claim. We include real-composite WMMSE both as a centralized comparison and as an alternative local solver inside our ADMM. The contribution is the decomposition and the augmented price, not the per-cell solver. If real-composite WMMSE plugged into our decomposition works best, we will report that.
Why not just use proper signaling with better scheduling or time-sharing? Time-sharing can enlarge the two-user rate region (Improper Signaling versus Time-Sharing in the Two-User Gaussian Interference Channel with TIN), but it needs synchronized slot assignment among cells. That is exactly the coordination a distributed scheme is trying to avoid. We will include a proportional-fair time-division baseline with the same message budget. If it matches our method in lightly loaded layouts, we expect that to show up as zero pseudo-prices there, which would be a result about where impropriety matters rather than a failure.
Why not just run centralized improper optimization inside small clusters? Clustering is the most practical threat. The 3-cell results suggest improper gains are local, and if they are, fixed clusters may capture most of them. Our experiments decide this directly. We will measure how the full method's advantage over the cluster baseline changes with cell density and with the users-to-antennas ratio. If the advantage vanishes, the honest conclusion is that clustered centralized improper signaling is sufficient. Our zero-pseudo-price criterion would then still be useful for deciding where to draw cluster boundaries.
Why not just transmit maximally improper signals everywhere? Fixed maximal impropriety needs no optimization and performs well in some cognitive-radio settings (Maximally Improper Signaling in Underlay MIMO Cognitive Radio Networks). But improper signaling gives up transmit entropy and is worse than proper signaling when noise dominates (A Journey From Improper Gaussian Signaling to Asymmetric Signaling). We will run "all maximally improper" and "all proper" as the two fixed-mode extremes. We expect each to lose in the regimes the other wins.
Proposed Experiments
The testbed extends the 3-cell model above to hexagonal layouts of 7, 19, 37, 61 and 91 cells. We vary the users-to-antennas ratio from well below 1 to above 1, so the network passes into the rank-deficient regime where improper signaling is known to help. We also vary cell radius to move between noise-limited and interference-limited operation.
The metrics are max-min rate, 5th-percentile user rate, sum rate, and energy efficiency in bits/Joule. We also measure the gap to centralized SCA wherever that solver can still run, ADMM rounds to a fixed residual, messages per cell per round, per-cell wall-clock time, and the fraction of cells whose solution is improper.
Three outcomes are decisive. First, on the 3-cell reproduction, our method must approach the reported 9.5 versus 8.2 bits/Joule gap. If it cannot, the decomposition loses too much. Second, at 37 cells and beyond, the full method must beat both the scalar-price ablation and the 3-cell cluster baseline on 5th-percentile rate with a margin exceeding seed-to-seed variation. This is what establishes that the pseudo-price carries information that neither local impropriety nor clustering captures. Third, the set of cells that end up proper should agree with the zero-pseudo-price criterion, and it should track the local users-to-antennas ratio. That would give the first scalable, locally computed answer to when a node should turn improper.
The main risk is slow ADMM convergence as network diameter grows, the failure mode noted by Information-constrained Optimal Control of Distributed Systems with Power Constraints. We will therefore report performance at fixed round budgets of 10, 30 and 100, not only at convergence.
References
- Improper Gaussian Signaling for Broadcast Interference Networks
- Regularized Zero-Forcing Aided Hybrid Beamforming for Millimeter-Wave Multiuser MIMO Systems
- A Journey From Improper Gaussian Signaling to Asymmetric Signaling
- Maximizing the Geometric Mean of User-Rates to Improve Rate-Fairness: Proper vs. Improper Gaussian Signaling
- Two-User SIMO Interference Channel With Treating Interference as Noise: Improper Signaling Versus Time-Sharing
- Improper Signaling versus Time-Sharing in the Two-User Gaussian Interference Channel with TIN
- Scalable Optimization Methods for Distribution Networks With High PV Integration
- Information-Constrained Optimal Control of Distributed Systems with Power Constraints
- Maximally Improper Signaling in Underlay MIMO Cognitive Radio Networks