In federated learning, phones and other devices train a shared model on data that never leaves them, then upload their model updates to a server. Over real wireless links, the devices with the weakest uplinks fail to deliver their updates before the round deadline. These devices are not a random sample. A trace study in Loss Tolerant Federated Learning estimates that standard settings would exclude about 24% of users, those with upload speeds below 2 Mbps. That paper's fix, ThrowRightAway, keeps these clients in the federation by setting their lost packets to zero and rescaling the aggregation weights. The radio itself is never asked to recover anything.
Wireless engineers have long recovered failed packets with Chase combining, a form of hybrid automatic repeat request. When a packet fails to decode, the receiver keeps the noisy received symbols rather than discarding them, asks for the same packet again, and adds the copies together. Each copy is weighted by its channel gain, which is maximal ratio combining (MRC), the linear combining rule that maximizes signal-to-noise ratio. Each retry accumulates signal energy until the packet decodes. We propose to bring this mechanism into federated learning and to make the learning objective decide which failed clients receive scarce retry slots within a round deadline. A client earns retries according to how much its missing update would bias the global model, not according to how close it is to decoding. Clients still lost after their retries fall back to loss-tolerant zero-filling.
Core Claim
Under a fixed per-round airtime budget, the server should not spend every slot on new clients or plain best-channel scheduling. A better use of some slots is to Chase-combine retransmissions from failed clients whose updates are both recoverable and statistically distinctive. We claim this recovers most of the accuracy gap between zero-filling and error-free communication, especially on classes held mainly by poorly connected clients, with no transmitter-side channel knowledge.
The new element is the allocation rule. MRC used on its own, like any radio-centric retry policy, favors whichever clients are closest to decoding. That reproduces the "strongest-gain trap" documented in Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning, where picking the strongest relay let one cluster dominate the aggregate. Our rule weights recoverability by an estimate of how far each client's update sits from the consensus.
Where This Sits
Most work at the intersection of diversity combining and federated learning is about over-the-air computation, where devices transmit analog updates simultaneously and the channel itself adds them. Several papers in this line use diversity in some form:
- Hierarchical Over-the-Air Federated Edge Learning applies multi-antenna MRC at intermediate servers but assumes an error-free backhaul to the central server.
- Relay-Assisted Cooperative Federated Learning combines direct and amplify-and-forward relayed copies. It tunes powers by alternating optimization and requires full channel knowledge.
- Over-the-Air Federated Learning in MIMO Cloud-RAN Systems replaces linear combining with learned compression and decoding across access points.
All of these combine different devices' signals in one shot and target analog aggregation. We combine repeated copies of one device's digital packet across time. Our decision variable is the retry allocation, not beamformer or power coefficients.
The work closest to ours is digital. Loss Tolerant Federated Learning motivates the proposal by showing that exclusion is biased and that inclusion via zero-filling helps, by up to 10.35% accuracy when 30% of clients were previously excluded. It also reports a roughly 1% drop in personalized accuracy, which hints at the distortion that zero-filling leaves behind. Federated Learning Over Wireless Fading Channels constrains the design. There, scheduling only the best-channel device per round struggled to exceed 80% on non-IID MNIST, while simultaneous analog aggregation reached nearly 90%, so any retry policy that collapses toward favoring strong channels should be expected to lose. Polar Coded HARQ Scheme with Chase Combining supplies the physical-layer method, which we reuse unchanged.
Research Program
Each round has a deadline of $B$ uplink slots. Clients first attempt a single transmission at a fixed code rate $R$ without transmitter channel knowledge. After $m$ Chase copies from client $k$, with per-copy SNRs $\gamma_{k,1},\dots,\gamma_{k,m}$ under independent block fading, MRC yields a combined SNR equal to the sum of the per-copy SNRs. Decoding succeeds approximately when
$$\log_2\Big(1+\sum_{j=1}^{m}\gamma_{k,j}\Big)\ \ge\ R .$$
From each failed client's first-attempt soft symbols, the server estimates $\gamma_{k,1}$. Using known fading statistics, it then computes $q_k(m)$, the probability that client $k$ decodes after $m$ total copies.
The server spends the remaining slots to minimize an expected aggregation-bias proxy:
$$\sum_k p_k\, d_k\,\big(1-q_k(m_k)\big)\quad \text{subject to}\quad \sum_k (m_k-1)\le B_{\text{left}}.$$
Here $p_k$ is client $k$'s data-size weight. The term $d_k$ estimates the distance between client $k$'s update and the aggregate, taken from its most recent successfully decoded update with the estimate decaying over rounds. The server already sees these updates, so this adds no privacy exposure. Because $q_k$ is concave in $m$ beyond the first useful copy, greedy allocation by marginal reduction per slot is a natural solver. We will check how close greedy comes to brute force on small instances. Clients still undecoded at the deadline are zero-filled and reweighted exactly as in ThrowRightAway.
The program has three parts. First, we will implement this protocol in a federated simulator with a link-level model calibrated to the FCC Mobile Broadband upload traces used by ThrowRightAway. Second, we will derive a simple convergence bound for FedAvg with partial participation in which the bias term scales with the residual missing mass $\sum_k p_k d_k (1-q_k)$, which justifies the objective above. Third, we will map when retries beat admitting fresh clients, as a function of how strongly link quality and data distribution are correlated.
Baselines
The naive reference is threshold dropping. Failed clients are excluded and the server averages over the successes. This is the default that ThrowRightAway shows to be biased.
The strongest existing method for this exact question is ThrowRightAway from Loss Tolerant Federated Learning, run with the same traces, the same deadline $B$, and the same model and tuning budget. Any slots our method spends on retries are given to ThrowRightAway as additional first-attempt clients, so total airtime is identical. Its reported numbers are the 10.35% gain over exclusion and the variance reduction from 1439 to 1031. No reported result was found on Lacuna for ThrowRightAway on CIFAR-10, so we will reproduce it ourselves. Beating it means higher final accuracy and higher worst-group accuracy at equal airtime, across at least three seeds.
A third baseline is best-channel scheduling in the style of D-DSGD from Federated Learning Over Wireless Fading Channels.
Three ablations isolate the key ingredients:
- Radio-only retry allocation. Retries go to the clients with the highest $q_k$ gain per slot, with $d_k$ set to a constant. This tests whether the learning-aware term matters.
- Selection combining of copies. The receiver keeps only the best single copy rather than MRC-summing them. This isolates the combining gain.
- Equal gain combining of copies. Copies are summed with phase alignment but no amplitude weighting.
The ablations should reproduce the known ordering of the three combining rules. More importantly, they show how much of the accuracy gain is due to MRC rather than to retrying at all.
Why not just…?
Why not just zero-fill with ThrowRightAway and skip retransmission? Zero-filling is free in airtime, and that is a genuine advantage. However, its own reported personalization penalty shows that it does not recover the information the weak clients hold. The equal-airtime comparison settles the question. If spending slots on fresh first attempts beats spending them on Chase retries, our proposal fails, and we expect this to happen when link quality is uncorrelated with data. Our experiments deliberately sweep that correlation.
Why not just send weak clients a lower-rate code, or compress their updates more? This is the most serious threat. With accurate transmitter channel knowledge, a properly chosen low-rate code beats repetition-style Chase combining. Compressing updates for slow clients, for example with side-information compression as in Federated-KLMS, shrinks the packet so that it fits. Chase combining has an advantage only when channel knowledge at the transmitter is stale or absent, because it spends extra airtime only after a failure. We will therefore add an oracle link-adaptation baseline and a per-client compression baseline, and report the channel-knowledge staleness at which each one overtakes ours. We will also test incremental-redundancy retries as a stronger physical-layer variant of our own method.
Why not just let server-side adaptive optimization absorb the noise? Adaptive Federated Learning Over the Air reaches 70% on CIFAR-10 in about 200 rounds under heavy-tailed noise, where FedAvgM-OTA stays below 40% after 600 rounds. That result concerns noisy updates that still arrive. A failed digital packet delivers nothing, and no step-size rule can restore a missing client's data. We will also run server-side Adam on top of both ThrowRightAway and our method to check whether the two gains are complementary.
Proposed Experiments
We use CIFAR-10 with ResNet-18 under Dirichlet(0.1) non-IID partitioning, following the adaptive over-the-air setup, and MNIST with the 25-device non-IID setup of Federated Learning Over Wireless Fading Channels. Each client's upload rate and loss come from the FCC 2019 traces, with Rayleigh block fading layered per attempt.
The critical manipulation is a "link–data correlation" knob. At one extreme, the clients with the lowest upload speed hold a disproportionate share of particular classes. At the other, data is assigned independently of link quality.
The metrics are:
- final and time-to-target accuracy;
- worst-class accuracy on classes concentrated among weak clients;
- effective participation mass, $\sum_k p_k$ over clients that decoded;
- airtime per round;
- variance of the aggregated update.
The decisive checks are three:
- Under strong link–data correlation, at equal airtime, our method should beat ThrowRightAway on worst-class accuracy by a margin larger than the seed spread.
- The learning-aware allocation should beat radio-only allocation. If it does not, the contribution reduces to "use HARQ," which is standard practice.
- MRC should beat selection combining of copies. If the gap is negligible, retry count rather than combining is doing the work.
We will also report where oracle link adaptation wins. This bounds the regime in which retry combining is the right tool for federated learning.
References
- Loss Tolerant Federated Learning
- Cluster-Based Cooperative Digital Over-the-Air Aggregation for Wireless Federated Edge Learning
- Hierarchical Over-the-Air Federated Edge Learning
- Relay-Assisted Cooperative Federated Learning
- Over-the-Air Federated Learning in MIMO Cloud-RAN Systems
- Federated Learning over Wireless Fading Channels
- Polar Coded HARQ Scheme with Chase Combining
- Adaptive Federated Learning Over the Air