Federated learning trains a shared model across phones or sensors without collecting their data. Each round, a server sends the current model down to the devices, and the devices send their locally computed updates back up. On a wireless link, both halves of that exchange compete for the same radio spectrum. Ordinary radios are half-duplex: they cannot transmit and receive on the same frequency at the same time. The server therefore alternates between a broadcast slot and a collection slot, and each device sits idle during one of them. Wireless federated learning research has made the upload cheap. Over-the-air computation lets all devices transmit at once so that their signals add up physically into the sum the server needs. The download, meanwhile, has usually been assumed to be free, and it is not.
We propose putting a full-duplex radio only at the access point, meaning a radio that cancels its own transmitted signal well enough to receive on the same frequency while it transmits. Devices keep their ordinary half-duplex hardware. The devices are split into two spatially separated groups that take turns: while group A uploads its updates over the air, the access point simultaneously broadcasts the freshest model to group B, and then the roles swap. If this works, each channel use does double duty. The price is two kinds of interference and one learning cost. The interference is leakage of the access point's own broadcast into its receiver, and the combined upload signal of group A landing on group B's receivers. The learning cost is that the two groups train on models that are half a round apart. The research question is whether the gain in wall-clock time outweighs these costs, and in which radio regimes it does.
Core Claim
The closest wireless result is the three-node topology of Harvesting Full-Duplex Rate Gains in Cellular Networks with Half-Duplex User Terminals. That paper analyzes a full-duplex base station serving one half-duplex uplink user and one half-duplex downlink user on the same resource. It reports only a 5% rate loss relative to giving every terminal full-duplex hardware, and it introduces an overlap parameter $\alpha$ between uplink and downlink spectrum. It models ordinary point-to-point packet traffic. Our setting differs in two ways that could either break or strengthen its conclusion.
The first difference makes things harder. In over-the-air aggregation the uplink is not one user but every member of group A transmitting at full coherence. The interference reaching each downlink device is therefore the sum of all of them, which is precisely the cross-link problem that paper flags as its main open issue.
The second difference may make things easier. The access point does not need to decode each upload. It only needs a low-error estimate of the sum, and stochastic gradient descent tolerates zero-mean noise on that sum. Residual self-interference that would cost a lot of Shannon rate may therefore cost little in accuracy.
Our claim is that the break-even point for full-duplex in federated learning should be measured in time-to-accuracy, not bits per second. We further claim that, with a grouping rule that keeps the two groups physically apart, a full-duplex access point with half-duplex clients reaches target accuracy in substantially fewer channel uses than a half-duplex access point with the same spectrum. We do not predict a specific speedup. We will measure where the advantage appears and where it collapses, as a function of residual self-interference level, group separation, and $\alpha$.
Why This Is Open, and What It Builds On
The papers connecting relaying and wireless federated learning all stay half-duplex and treat the uplink alone.
Relay-Assisted Cooperative Federated Learning uses half-duplex amplify-and-forward relays across two transmission phases to help the weakest devices in over-the-air aggregation. It proves that relays reduce aggregation error only above explicit power and channel-gain thresholds, and it assumes the downlink is error-free and instantaneous.
Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning quantizes updates to one bit (signSGD, where each device sends only the sign of each gradient coordinate and the server takes a majority vote). It then aggregates in two sequential stages through cluster relays and shows that fading shrinks a 54-user network to an effective 17 voters without cooperation.
Hierarchical Over-the-Air Federated Edge Learning adds intermediary servers but assumes an ideal backhaul. Integrated Sensing, Communication, and Computation for Over-the-Air Federated Edge Learning schedules sensing, computing, and uploading as strictly sequential blocks.
These four papers motivate the proposal: they establish over-the-air aggregation as the uplink to beat, and they leave the downlink and duplexing untouched. We do not change their uplink designs. We change what the access point does while the uplink happens.
The full-duplex literature supplies the mechanism and also constrains it. In-Band Full-Duplex Wireless: Challenges and Opportunities. and Full-Duplex Wireless for 6G: Progress Brings New Opportunities and Challenges establish that the ideal doubling of spectral efficiency requires suppressing self-interference near the noise floor.
Frequency-Selective Beamforming Cancellation Design for Millimeter-Wave Full-Duplex and Hybrid Beamforming for Millimeter Wave Full-Duplex under Limited Receive Dynamic Range show that power-amplifier nonlinearity and limited converter dynamic range leave residue that cannot be cancelled. This is the negative result our design must survive. For the cancellation stage we reuse learned cancellers in the style of DSIC: Deep Learning Based Self-Interference Cancellation for In-Band Full Duplex Wireless, rather than inventing a new one.
Research Program
The system has three parts.
The two-group schedule. In each half-round, group A uploads updates computed from model $w_t$ via analog over-the-air aggregation. At the same time, the access point broadcasts $w_{t+1/2}$ to group B, which was aggregated from B's previous upload. B then starts local training while A waits for its broadcast in the next half-round. The learning algorithm is thus a two-stream staggered scheme with a fixed staleness of one half-round. We implement it and, as a separate learning-only ablation, run it with perfect channels to isolate the cost of staleness from any radio effect.
The physical model. The access point receives
$$y = \sum_{k \in A} h_k b_k s_k + r(x_{\mathrm{DL}}) + n,$$
where $h_k$ is device $k$'s uplink channel, $b_k$ its power-control coefficient, $s_k$ its update symbol, $x_{\mathrm{DL}}$ the model broadcast the access point is sending, $r(\cdot)$ the residual self-interference after cancellation, and $n$ thermal noise.
Each downlink device $j \in B$ decodes the broadcast at signal-to-interference-plus-noise ratio
$$\gamma_j = \frac{P\,|g_j|^2}{\alpha \sum_{k \in A} |c_{jk}|^2 |b_k|^2 + \sigma^2},$$
where $P$ is access-point power, $g_j$ the downlink channel, $c_{jk}$ the device-to-device channel from uploader $k$ to listener $j$, and $\alpha \in [0,1]$ the spectral overlap from the three-node paper.
The downlink rate is set by the worst listener, because the broadcast is common to all of B. The quantity $\min_j \gamma_j$ therefore drives downlink time, while $\mathbb{E}\|r\|^2$ and the weakest uplink channel drive aggregation error.
Modeling residual self-interference. We treat $r$ in two ways. The first is standard independent Gaussian noise. The second is a nonlinear, signal-dependent leakage produced by a power-amplifier model. This distinction matters for learning in a way it does not matter for communication rate. The broadcast $x_{\mathrm{DL}}$ is a model delta correlated with the updates being aggregated, so nonlinear leakage can add a biased component along the update direction rather than isotropic noise. Whether residual self-interference biases the aggregate or merely inflates its variance is the specific form of the open question raised in the file's gap analysis. It is the analytical piece we will work out.
Grouping and power control. We design a grouping rule that splits clients to maximize the minimum downlink SINR in B, subject to both groups holding comparable data. The spatial clustering of the cluster-based paper supplies the starting heuristic. We pair it with the uplink power control from the relay-assisted paper, re-derived with the self-interference term included.
Baselines
The naive reference is a half-duplex access point running synchronous federated averaging. Uplink uses analog over-the-air aggregation and downlink uses digital broadcast, in separate time slots over the full band, with every client in every round. Channel uses per round equal the uplink length plus the downlink length. We will report this reference's time-to-accuracy, but no reported result found on Lacuna gives time-to-accuracy for such a system with the downlink counted.
The strongest existing uplink design is the half-duplex relay-assisted aggregation of Relay-Assisted Cooperative Federated Learning. No reported accuracy or time-to-accuracy result found on Lacuna exists for it, so it will be reimplemented with its alternating power-and-relay optimization. For the digital variant, the relevant baseline is the cluster-based scheme of Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning. On MNIST it reports final accuracy within 0.5% of an ideal majority vote, with the non-cooperative variant about 1.2% worse. We reproduce this setup with 54 users and 6 clusters, check that our implementation matches those two gaps, and then add a counted downlink.
The frequency-division baseline uses the same half-duplex clients and splits the band into separate uplink and downlink halves. This is $\alpha = 0$ in the three-node parameterization. It already lets upload and broadcast proceed concurrently without any full-duplex hardware, so it is the most direct competitor to the core claim.
The key ablation is the same two-group staggered schedule run on a half-duplex access point by time division. It has the same staleness as our method but no simultaneous transmission. Any gain over this ablation is attributable to full-duplex operation. Any gain over the synchronous baseline that this ablation also achieves is attributable to the schedule alone.
Fairness rules apply to every method: identical total bandwidth, identical client powers and channel realizations drawn from the same seeds, the same local optimizer and learning-rate sweep, and the same number of tuning trials. We count channel uses to reach a fixed target accuracy that is set at 95% of the synchronous error-free accuracy.
Beating the baselines means reaching the target in meaningfully fewer channel uses than both the frequency-division baseline and the half-duplex staggered ablation, at final accuracy no worse than the synchronous baseline minus its seed-to-seed spread, across a stated range of residual self-interference levels.
Why not just…?
Why not just compress the updates or tune only prompts so communication stops mattering? Compression and parameter-efficient tuning shrink the upload and the broadcast by the same factor, so the fraction of each round spent broadcasting is unchanged. The full-duplex overlap still removes that fraction.
There are limits to how far compression can go. Enhancing Communication Compression via Discrepancy-aware Calibration for Federated Learning reports only 29.62% top-1 accuracy on CIFAR-10 with a small vision transformer at 1% of coordinates kept, versus 21.03% for magnitude-based selection. Extreme compression therefore costs accuracy.
The objection is still a real threat. If payloads become small enough, local computation dominates each round, and halving communication time is irrelevant. We will run our method on top of the prompt-tuning payload of Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning, which updates 2.8M parameters. We will report the ratio of compute time to communication time at which the advantage disappears.
Why not just split the band into uplink and downlink halves? This is the frequency-division baseline. The three-node paper's 5% gap versus ideal full-duplex suggests that overlapping the full band wins. However, that analysis assumes a single interfering uplink user, whereas over-the-air aggregation has all of group A transmitting coherently. If aggregate cross-link interference pushes the optimal $\alpha$ toward zero, then frequency division is the answer and our proposal fails. The sweep over $\alpha$ and group separation decides this directly.
Why not just stagger the groups on a half-duplex access point, or run asynchronous federated learning? Staggering by itself hides some client idle time but does not create extra channel uses. The half-duplex staggered ablation measures exactly how much of any gain comes from scheduling alone. If that ablation matches our method, full-duplex hardware is not justified.
Why not use full-duplex relays or passive reflecting surfaces for the uplink stragglers instead? Those address uplink stragglers, not the broadcast slot, and can be combined with our design. Reconfigurable Intelligent Surfaces vs. Relaying: Differences, Similarities, and Performance Comparison notes that a passive surface suffers multiplicative path loss and injects no power. That makes it a weak substitute for either an active relay or an active broadcast.
Proposed Experiments
Learning tasks follow the file's benchmarks, run in simulation with Rayleigh fading and stochastic placement of clients:
- MNIST, 54 clients, signSGD. This directly matches the cluster-based paper's reported gaps.
- CIFAR-10, non-IID. This follows Hierarchical Over-the-Air Federated Edge Learning, with a small CNN under analog aggregation.
- DomainNet with CLIP prompt tuning. This uses the FedAPT payload, whose reported average accuracy of 67.39% is the ceiling our channel-impaired runs should approach.
We sweep residual self-interference from the noise floor to 30 dB above it, under both the Gaussian and the nonlinear leakage models. We also sweep $\alpha$ from 0 to 1, and inter-group separation from co-located to cell-edge.
The metrics are:
- channel uses to target accuracy;
- final accuracy;
- aggregation mean-squared error;
- worst-listener downlink SINR;
- for the leakage study, the cosine similarity between the residual-induced error on the aggregate and the true update direction.
A clear positive result would require the full-duplex method to beat both the frequency-division baseline and the half-duplex staggered ablation in channel uses to target at residual self-interference within roughly 10 dB of the noise floor, with no accuracy loss. A result in which nonlinear leakage biases the aggregate along the update direction would be informative even if the speedup fails. It would show that full-duplex design for learning must target a different cancellation objective than full-duplex design for rate.
A clear negative result would be the optimal $\alpha$ collapsing to zero under aggregate cross-link interference at any realistic group separation. That outcome would settle Gap 4 of the file against concurrent broadcast and aggregation.
References
- Harvesting full-duplex rate gains in cellular networks with half-duplex user terminals
- Relay-Assisted Cooperative Federated Learning
- Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning
- Hierarchical Over-the-Air Federated Edge Learning
- Integrated Sensing, Communication, and Computation for Over-the-Air Federated Edge Learning
- In-Band Full-Duplex Wireless: Challenges and Opportunities.
- Full-Duplex Wireless for 6G: Progress Brings New Opportunities and Challenges
- Frequency-Selective Beamforming Cancellation Design for Millimeter-Wave Full-Duplex
- Hybrid Beamforming for Millimeter Wave Full-Duplex Under Limited Receive Dynamic Range
- DSIC: Deep Learning Based Self-Interference Cancellation for In-Band Full Duplex Wireless
- Enhancing Communication Compression via Discrepancy-aware Calibration for Federated Learning
- Federated Adaptive Prompt Tuning for Multi-Domain Collaborative Learning
- Reconfigurable Intelligent Surfaces vs. Relaying: Differences, Similarities, and Performance Comparison