A transmitter with several antennas can steer its signal so that almost no energy reaches a chosen set of nearby receivers it is not talking to. This is called "nulling," and each receiver it protects this way uses up one antenna's worth of freedom. The textbook analysis of this trick in decentralized wireless networks (Spatial interference mitigation for multiple input multiple output ad hoc networks: MISO gains) shows that network capacity can grow linearly with the number of antennas. It gets there by quietly assuming that every transmitter already knows two things: which nearby receivers are active in the current time slot, and what the radio channel to each of them looks like. In a real network with no central controller, that knowledge costs per-slot signaling, and the pilot signals used to measure channels get mixed up whenever two nearby receivers pick the same code.
We propose to supply most of that knowledge through shared randomness, at no per-slot signaling cost. Every node holds the same secret seed and a common frame counter. From these, any node can compute, for any neighbor whose identity it knows, whether that neighbor is scheduled this slot, which pilot code it will use, and therefore which pilot collisions will happen. Each transmitter can then spend its limited nulling capacity only on receivers that are active and whose channels it can measure cleanly. The same seed can also make neighbors' activity negatively correlated, so the closest interferers are removed by scheduling and the antennas are left to handle the next ring out. The claim, in two sentences: a shared seed turns "who is on, and on which pilot" into common knowledge without messages, which lets nulling be targeted rather than blind. Beyond that, it makes scheduling and nulling complementary rather than competing uses of the network's resources.
Core Claim
In a network of multi-antenna transmitters, each with its own receiver and with no central controller, consider seed-derived per-slot activity together with seed-hopped pilot codes. We claim this combination achieves higher overhead-adjusted transmission capacity than two alternatives at the same antenna count. The first is blind partial nulling, which protects the closest receivers whether or not they are active. The second is contention-based or statically planned alternatives, once their signaling time is charged against the frame. Transmission capacity here means the density of simultaneous successful links per unit area at a fixed outage probability.
The claim should hold most clearly when three conditions meet. The pilot pool is small relative to the number of receivers within nulling range. Neighborhoods change slowly compared with slots. Nodes have more nearby interferers than antennas, which is the degrees-of-freedom depletion regime flagged in the file.
The mechanism is simple enough to state as one decision rule. At slot $t$, transmitter $j$ with $N$ antennas computes the active set $A_t$ and pilot map $\pi_t(\cdot)$ for its known neighbors from the seed. It then nulls the set
$$S_{j,t} = \arg\max_{S \subseteq C_{j,t},\ |S| \le l} \sum_{k \in S} d_{jk}^{-\alpha},$$
where $C_{j,t}$ is the set of receivers that are in $A_t$ and whose pilot $\pi_t(k)$ is not shared with any other receiver within $j$'s sensing range. The term $d_{jk}$ is the distance from $j$ to receiver $k$, $\alpha$ is the pathloss exponent, and $l \le N-1$ is the nulling budget. In words: protect the strongest victims that are actually on and actually measurable. Blind partial nulling instead picks the $l$ nearest receivers regardless of activity or contamination. Every antenna it spends on an idle or mis-measured receiver is wasted.
Where This Sits
The shared-randomness line of work supplies the method but has never touched the physical layer. Permutation Compressors for Provably Faster Distributed Nonconvex Optimization has workers use a common seed to partition gradient coordinates without handshakes, and proves up to a $\sqrt{n}$ communication speedup over independent sampling. Convergence Visualizer of Decentralized Federated Distillation with Reduced Communication Costs uses a pre-shared seed to align sample indices and reports a 72% traffic reduction on Fashion-MNIST. Both assume perfect digital links. Both run the seeded generator in lockstep, which that paper's own setting shows is fragile under node churn. We reuse the "compute it locally, identically" idea, but apply it to pilot and activity assignment.
Two game-theoretic results motivate why a common signal should matter. Signal Instructed Coordination in Cooperative Multi-agent Reinforcement Learning shows that a shared random signal lets independent agents reach any centrally correlated policy. Hardness of Independent Learning and Sparse Equilibrium Computation in Markov Games shows that without such correlation, independent learners face exponential barriers. The negatively correlated activity patterns we generate are exactly a correlated policy of this kind, realized by radios. Cooperative and Stochastic Multi-Player Multi-Armed Bandit: Optimal Regret With Neither Communication Nor Collisions is the closest analogue in spirit: a public coin lets players avoid colliding without talking. Our setting differs because collisions are spatial and partial, and because antennas can absorb some of them.
Two results constrain the design. Communication With Imperfectly Shared Randomness and Coordination Through Shared Randomness show that when the randomness is not exactly shared, communication must be spent to restore coordination. We therefore do not rely on a stateful random number generator advancing in lockstep. We use a counter-based generator, a hash of seed, node identity and frame number. A node that misses slots recovers alignment as soon as it reads the frame number again. We explicitly measure how often clock or neighbor-list disagreements break the common knowledge.
On the wireless side, the partial nulling result in Spatial interference mitigation for multiple input multiple output ad hoc networks: MISO gains is the model we build on. Its authors concede that nulling the receivers closest to the transmitter is not always the same as nulling the worst interferers. User-Centric Intercell Interference Nulling for Downlink Small Cell Networks replaces the genie with explicit nulling requests from users, and shows that base stations can run out of antennas when too many requests arrive. Random Aggregate Beamforming for Over-the-Air Federated Learning in Large-Scale Networks is the nearest work that mixes randomness and beamforming. Its randomness is sampled centrally and broadcast by one aggregator, it serves a single cell, and it does no nulling toward unselected nodes. Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air suppresses interference statistically at a large receive array, with no coordination among transmitters. Our proposal is decentralized and transmit-side, and its randomness is shared rather than broadcast.
Research Program
The system has three seed-driven pieces, all computed locally from the seed, node identities and the frame counter.
The first is activity. Each node draws a per-slot priority. A link is active if its transmitter has the highest priority among the neighbors it knows within an exclusion radius $r$. This yields a contention-free, randomly re-drawn hard-core pattern: nearby links are rarely on together. Setting $r=0$ recovers independent random access with a known coin.
The second is pilots. Each active receiver's pilot is drawn from a pool of $P$ codes by the same hash. Every transmitter therefore knows which of its neighbors share a code this slot. Collisions re-randomize every slot rather than persisting as they would under a static random code plan.
The third is nulling. Each transmitter applies the decision rule above, using channel estimates obtained by reciprocity from the uncontaminated pilots.
The one piece of state not derivable from the seed is the neighbor identity list. We assume it is maintained by slow periodic discovery and charge that cost in the overhead accounting. We then study how staleness degrades performance.
The analysis has two parts. First, using the stochastic-geometry tools of the partial nulling paper, we will derive the outage probability under seeded hard-core activity with targeted nulling. The goal is to find the capacity-maximizing split between exclusion radius $r$ and nulling budget $l$ for given $N$, $P$ and $\alpha$. This generalizes that paper's finding about how many antennas to devote to nulling. Second, we will bound the loss from imperfect common knowledge as a function of the probability that two neighbors disagree on the frame index or on each other's neighbor lists. This quantifies when the imperfect-randomness penalty from the information-theory results becomes significant.
Baselines
The naive reference point is uncoordinated random access with independent coins, where each transmitter nulls its $l$ nearest receivers blindly. It knows neither their activity nor their pilot collisions, and its channel estimates come from statically and randomly assigned pilots. This is what a network achieves without either the genie or any coordination.
The strongest existing method is partial nulling as analyzed in Spatial interference mitigation for multiple input multiple output ad hoc networks: MISO gains. We run it in two forms. The genie form is given perfect activity and channel knowledge and serves as an upper bound on what targeting can achieve. The realistic form obtains activity and channel knowledge through an explicit per-slot beacon plus orthogonal pilots, with those symbols charged to the frame. The file contains no absolute transmission-capacity figure for that paper. What it reports is linear scaling with $N$ and an optimum near $l=5$ at $N=10$ and $\alpha=4$. We will first reproduce that optimum in our simulator as a calibration check.
For the small-cell variant, the strongest baseline is the distance-triggered request scheme of User-Centric Intercell Interference Nulling for Downlink Small Cell Networks. It reports a 35–40% relative improvement in success probability over uncoordinated networks. It also reports needing one third of the uncoordinated base-station density for 90% success at a 10 dB threshold, robust with at least 10 bits of channel quantization. We match its PPP densities, thresholds and quantization. Our variant replaces its explicit requests with seed-computed request lists that are honored on a seeded rotation, so that no base station exceeds its antenna budget.
Two further baselines represent coordination by other means. Carrier-sensing access with a short contention window stands in for contention-based hard-core scheduling. A static graph-coloring schedule computed centrally from the topology stands in for planned orthogonalization, as in Decentralized SGD with Over-the-Air Computation. That paper reports coloring needing 57 slots for orthogonal point-to-point links versus 19 for over-the-air multiple access in its setting.
Fairness rules apply to every method. All methods share the same node placements, fading draws, antenna count, total frame length and transmit power. Every non-data symbol (beacons, contention mini-slots, discovery, pilots) is subtracted from the frame. The tunable parameters, namely $l$, $r$, contention window and color count, get the same grid-search budget for every method.
Two ablations isolate the key ingredient. One keeps seed-derived activity but uses static random pilots. The other uses seed-hopped pilots but independent unknown activity. A third comparison sets $r=0$ to separate the value of knowing activity from the value of correlating it.
Beating the strongest baseline means one of two outcomes. The proposal could exceed realistic partial nulling in overhead-adjusted transmission capacity at 10% outage across the depletion regime. Alternatively, it could come within a small gap of the genie bound while spending no per-slot signaling. On the small-cell setup, it means matching the reported 35–40% success-probability gain without explicit request messages.
Why not just…?
Why not just detect which receivers are active from pilot energy, with fixed pilots? This is the most serious alternative, because pilots must be sent anyway. It works when the pilot pool is large enough that nearby receivers rarely share codes. When the pool is small relative to the neighborhood, a static random plan creates persistent collisions between the same pairs of neighbors. The transmitter cannot tell a contaminated estimate from a clean one, so its nulls land in the wrong place. Seeded hopping makes collisions both transient and predictable. Our pilot-pool sweep over $P$ tests exactly this: we expect the advantage to vanish at large $P$, and we will report the crossover honestly rather than claim a uniform win.
Why not just use carrier sensing to get hard-core scheduling? Carrier sensing does produce locally exclusive activity. It costs contention time every slot, suffers hidden-terminal errors, and tells a transmitter about activity only within sensing range. The ring of receivers that nulling is supposed to protect often lies beyond that range. Carrier sensing is a real threat if its contention window can be made very short. The overhead-adjusted comparison at matched frame length, repeated across contention-window sizes and node mobility levels, settles this directly.
Why not just plan orthogonal schedules or coordinate centrally? Graph coloring and coordinated multipoint remove interference by construction. They require global topology or channel knowledge and must be recomputed whenever the topology changes. The slot-count gap reported in Decentralized SGD with Over-the-Air Computation shows how costly strict orthogonality can be. The user-centric nulling paper shows centralized requests overloading base-station antennas in dense irregular layouts. Under node churn and mobility we will measure the recomputation and backhaul cost these schemes need to stay current. We will compare it against the discovery cost our seeded scheme pays.
Proposed Experiments
The primary testbed is the ad hoc model from the partial nulling paper. Transmitters form a homogeneous Poisson point process, each with a receiver at fixed distance, under Rayleigh fading with $\alpha=4$ and $N$ ranging from 2 to 16. The main metric is overhead-adjusted transmission capacity at 10% outage, with the realized nulling efficiency as a diagnostic: the fraction of spent nulling freedom that landed on active, cleanly measured receivers.
The first decisive check is calibration. Reproducing the $l \approx 5$ optimum at $N=10$ validates the simulator before any claim is made.
The second decisive check is whether seeded targeting alone ($r=0$, hopped pilots) beats blind partial nulling at equal $N$ once receivers outnumber antennas. If it does not, the common-knowledge premise is wrong and the project reduces to a scheduling study.
The third check is whether adding seeded exclusion shifts the optimal nulling budget and raises capacity beyond targeting alone. That would support the claim that scheduling and nulling become complementary.
The fourth is the pilot-pool and carrier-sensing sweeps described above, which decide whether the simpler alternatives already capture the gain.
The fifth is the robustness test. We inject frame-index disagreements and stale neighbor lists at controlled rates and measure how quickly capacity falls toward the uncoordinated reference. We expect graceful degradation from the counter-based design, since a desynchronized node fails alone rather than drifting permanently.
The secondary testbed is the small-cell downlink model of the user-centric paper, run at its reported thresholds. It tests whether seed-computed request rotation preserves its gains while removing its request traffic.
References
- Spatial interference mitigation for multiple input multiple output ad hoc networks: MISO gains
- Permutation Compressors for Provably Faster Distributed Nonconvex Optimization
- Convergence Visualizer of Decentralized Federated Distillation with Reduced Communication Costs
- Signal Instructed Coordination in Cooperative Multi-agent Reinforcement Learning
- Hardness of Independent Learning and Sparse Equilibrium Computation in Markov Games
- Cooperative and Stochastic Multi-Player Multi-Armed Bandit: Optimal Regret With Neither Communication Nor Collisions
- Communication with Imperfectly Shared Randomness
- Coordination Through Shared Randomness
- User-Centric Intercell Interference Nulling for Downlink Small Cell Networks
- Random Aggregate Beamforming for Over-the-Air Federated Learning in Large-Scale Networks
- Quantized Analog Beamforming Enabled Multi-task Federated Learning Over-the-air
- Decentralized SGD with Over-the-Air Computation