Cheap wireless devices such as sensors, tags, and micro-drones are too small to carry several antennas, so a well-known trick is to let a group of them act together as one "virtual" antenna array. Each device forwards a slightly transformed copy of the signal, and the receiver combines the copies as if they came from a single multi-antenna transmitter. This technique is called distributed space-time coding. Its theory promises that reliability, measured as how fast the error rate falls as power grows, improves in proportion to the number of cooperating devices. That theory assumes the devices sit far apart, so their radio paths fade independently. Dense swarms break this assumption. When antennas sit closer than half a wavelength, they electrically disturb one another, an effect called mutual coupling. Measurements on dense arrays show radiation efficiency collapsing and capacity saturating once spacing drops below half a wavelength. Nobody has asked what this does to distributed codes, or how a swarm should code when it cannot spread out.
We propose to treat a dense swarm the way electronic-structure methods treat a molecule. Atom positions determine a Hamiltonian matrix, the matrix describing how electron orbitals on different atoms interact. In the same way, node positions and orientations determine a coupling matrix describing how each node's antenna disturbs every other's. Machine-learned models now predict molecular Hamiltonians directly from geometry, with the correct behaviour under rotation built in, and skip expensive density functional theory (DFT) calculations. We reuse that machinery to predict the swarm's coupling matrix from geometry alone. The swarm then assigns its code streams to the cluster's few efficiently radiating collective modes rather than to its many individual nodes. This needs no extra channel training at the relays, matching the no-relay-CSI property that makes distributed coding attractive in the first place.
Core Claim
The claim has two parts. The first is a statement about physics and coding. For a cluster of $R$ cooperating single-antenna nodes packed below half-wavelength spacing, the reliability gain that standard distributed space-time coding achieves at realistic signal-to-noise ratios is not $\min\{M,N\}R$, the bound proved by Jing & Hassibi for independent relays. It is instead governed by $\min\{M,N\}R_{\text{eff}}$, where $R_{\text{eff}}$ counts the eigenmodes of the cluster's radiation-resistance matrix whose efficiency exceeds a threshold. That matrix captures how much power each joint excitation pattern actually sends to the far field. We expect $R_{\text{eff}}$ to track the cluster's physical size measured in half-wavelengths rather than its node count. The basis for this expectation is the finding in Densifying MIMO that sub-half-wavelength elements couple into evanescent, non-radiating waves. Two consequences follow: adding nodes to a fixed volume should stop helping, and codes that ignore coupling should waste power on dead modes.
The second part is a design. Write $A_r$ for the dispersion matrix, the fixed linear transformation relay $r$ applies to its received signal. In a mode-domain code, each $A_r$ becomes a geometry-dependent combination of $R_{\text{eff}}$ base dispersion matrices, $A_r = \sum_{k=1}^{R_{\text{eff}}} u_{rk} B_k$. Here $B_k$ is an ordinary full-diversity code matrix for virtual antenna $k$, and $u_{rk}$ is node $r$'s weight in the $k$-th efficient radiation mode. Every node needs only its own row $u_{r\cdot}$. That row depends only on cluster geometry, which a swarm usually knows from its own localization. It does not depend on the unknown propagation channel to the destination.
Research Program
The work builds three things.
A coupled channel model. We take a standard multipath channel and pass it through the cluster's port-coupling matrix and embedded radiation efficiencies. The coupling quantities come from full-wave electromagnetic simulation of realistic small node bodies: a dipole or patch on a small printed circuit board with a ground plane. We calibrate the efficiency-versus-spacing trend against the 74%→11% drop reported by Densifying MIMO.
An analysis of effective diversity. We derive pairwise-error-probability bounds for linear dispersion codes passed through a rank-deficient, poorly conditioned coupling transform. The goal is to state when the finite-SNR error slope follows $R_{\text{eff}}$ rather than $R$. We also re-derive the 50/50 power split of Jing & Hassibi for the coupled case, where relay power on weak modes is partly lost.
A geometry-to-coupling predictor. This is where DFT methods supply the technique. QHNet and the QH9 benchmark predict a Kohn-Sham Hamiltonian block by block, one block per pair of atoms. They use equivariant graph networks, which means rotating the molecule rotates the predicted matrix consistently. A pair of oriented antennas has the same structure: the coupling between two dipoles transforms under rotation like an angular-momentum-1 object. So the per-pair blocks of the coupling matrix can be predicted with the same tensor-product layers, applied at low angular order, where the cost problems of those layers are mild.
For large swarms we borrow the learned pair-pruning of SPHNet, which prunes up to 70% of atom-pair interactions on large molecules. Coupling decays with distance, so most node pairs in a large swarm should be prunable. The predictor replaces a full-wave simulation every time the swarm reconfigures. It has the same role a predicted Hamiltonian has in warm-starting a DFT calculation: accuracy matters only insofar as the downstream decision, here the choice of mode weights, is unchanged.
DFT is the method lens and is not claimed as a physical analogy. We do not claim that electrons and antennas obey the same equations. We claim that "geometry in, structured pair-interaction matrix out, rotation-consistent" is a shared learning problem with a mature toolkit.
Baselines
Uncoupled reference. The naive reference point is the distributed code of Jing & Hassibi: random unitary dispersion matrices per relay, 50/50 power split, no coupling awareness. It runs on the same coupled channel, the same cluster geometries, and the same total power as every other method. It is the strongest competitor on the coding side because it is the scheme the claim says breaks down. Its full-diversity guarantee holds exactly when nodes are far apart, which isolates the effect of density. Lacuna has no reported BER-versus-SNR number for this scheme under mutual coupling, so its coupled performance is something we measure, not something we take from the literature.
Energy-sorted relay selection. The decode-and-forward extension of Hohmann et al. generalizes the Alamouti code to $N$ nodes and picks relays by channel energy. That paper reports that sorted selection gives a large gain over random selection for binary phase-shift keying at 8 dB, but gives no numeric value. We give it the same perfect receiver channel knowledge it assumes, and also a version charged for the pilot overhead of probing every node.
Spacing-thinned subset. This geometric heuristic activates only nodes at least half a wavelength apart, chosen greedily. It is the simplest physically informed competitor.
Oracle mode-domain code. This is our design with the exact full-wave coupling matrix. It is the ceiling.
Ablations. Two ablations isolate the predictor. One replaces the equivariant network with a non-equivariant graph network trained with rotation augmentation on the same data and parameter budget. The other replaces it with the closed-form thin-dipole mutual-impedance formula.
What counts as beating the baselines. Beating the uncoupled reference and the thinned subset means a lower outage probability at matched total power, and a larger measured error slope, in clusters with median spacing below half a wavelength. We require at least a 3 dB gain at an outage of $10^{-3}$ as a pre-registered bar. For the predictor, success means recovering at least 90% of the oracle's gain over the uncoupled reference.
For context on the network family only: QHNet reports a Hamiltonian error of $76.31\times10^{-6}$ Hartree on QH9-stable-id (QH9). That figure is not transferable to our task, and there is no reported result on Lacuna for learned antenna-coupling prediction.
Why not just…?
Why not just keep nodes half a wavelength apart? This is the most serious threat. If spacing-thinned subsets match the mode-domain code, the coding contribution reduces to a placement rule. Our hypothesis is that thinning discards the residual efficient modes that a joint excitation can still use, and that swarms confined to small volumes, such as tags on one object or a drone cluster in a corridor, often cannot be thinned to enough nodes. The experiment plots outage against node count for fixed cluster volumes. If the thinned curve tracks the oracle within 1 dB, we report that thinning suffices and drop the coding claim.
Why not just use the closed-form dipole coupling formula, or run the electromagnetic solver? For ideal thin dipoles in free space, closed-form mutual impedance is accurate, and then no learning is needed. We test exactly this ablation. We expect it to fail on realistic node bodies with ground planes and enclosures. The full-wave solver is exact, but each new geometry costs a simulation, which defeats online reconfiguration of a moving swarm. The predictor is justified only if it beats the closed form on realistic bodies and matches the solver's downstream decisions at a small fraction of its runtime. Otherwise we drop it.
Why not just train a big unconstrained network on lots of simulated swarms? Data scaling does help. In Repurposing DFT, the HOMO-LUMO gap error of a SchNet model, a widely used molecular graph network, falls from 0.348 eV to 0.0486 eV with $10^8$ samples. The same paper reports weaker transfer off-equilibrium (0.123 eV). EG-XC reports 51% lower error than competing methods when trained on small molecules and tested on larger ones. The analogous test here is training on swarms of 4–8 nodes and testing on 16–32, where we expect the equivariant model's advantage to show. If the augmented non-equivariant model matches it, we will say so.
Why not just use a fluid antenna on each node? Fluid antenna systems get diversity from many closely spaced switchable ports on one radio chain. That paper reports that about 30 ports beat 5-branch maximum ratio combining, an ideal multi-chain receiver baseline. But it is a single-device scheme and itself faces dense-port correlation. We include a per-node fluid-antenna variant as a complementary point, not a replacement, because the swarm's aperture is larger than any one node.
Proposed Experiments
Datasets. We generate cluster geometries of 2–32 nodes at sub-6 GHz, with spacings from $\lambda/8$ to $2\lambda$, in random 3D orientations. Each geometry is simulated with full-wave electromagnetics for both an ideal dipole and a small-board node body. The channel is the Monte Carlo Rayleigh/Rician fading simulator used by the relay and fluid-antenna papers, with antenna effects grafted on through the coupling matrix. We also inject the spatial correlation statistics of the 256-point non-line-of-sight measurement grid from Densifying MIMO as a realism check.
Metrics. We measure bit error rate and outage probability against SNR, the empirical diversity slope, and ergodic rate. For the predictor we measure coupling-matrix block error, mode-subspace angle, and downstream outage loss relative to the oracle.
Check 1: does diversity follow $R_{\text{eff}}$ or $R$? For fixed cluster volume, the uncoupled reference's error slope should saturate with node count at roughly the efficient-mode count $R_{\text{eff}}$. Above half-wavelength spacing it should recover the $\min\{M,N\}R$ slope. If it does not saturate, the core claim is wrong.
Check 2: does the mode-domain code beat both uncoupled coding and thinning? At sub-half-wavelength spacing, the oracle mode-domain code should beat the uncoupled code and the thinned subset by the pre-registered margin, with gains vanishing at wide spacing.
Check 3: does the predicted coupling preserve the gain? The predicted-coupling code should retain at least 90% of the oracle gain on realistic node bodies and on swarms larger than those seen in training. It should beat both the closed-form and the non-equivariant ablations there. Failing check 3 but passing check 2 leaves a useful coding result that relies on simulation. Failing check 2 leaves a negative result: thinning plus standard coding is enough. That result is still worth publishing, because it bounds what dense swarms can gain.
References
- Cooperative diversity in wireless relay networks with multiple-antenna nodes
- Densifying MIMO: Channel Modeling, Physical Constraints, and Performance Evaluation for Holographic Communications
- Efficient and Equivariant Graph Networks for Predicting Quantum Hamiltonian
- QH9: A Quantum Hamiltonian Prediction Benchmark for QM9 Molecules
- Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity
- Reliable multimedia transmission in wireless ad-hoc networks for telehealth systems
- Repurposing Density Functional Theory to Suit Deep Learning
- Learning Equivariant Non-Local Electron Density Functionals
- Performance Analysis of Fluid Antenna System Under Spatially-Correlated Rician Fading Channels