In wireless federated learning, many devices compute gradient updates and send them to a server over unreliable radio links. Each device's radio protocol currently decides on its own whether to resend a lost packet, and it does so with no knowledge of what the packet is for. The standard choice is HARQ (hybrid automatic repeat request, the retry-until-received scheme used in cellular links). HARQ treats a gradient update like any other packet: it must arrive with very high probability, so the device ramps up power over repeated attempts until it does. Many compressed-gradient methods, however, already carry a memory that makes lost updates recoverable. In error feedback, each device keeps the part of its update that has not yet reached the server and resends it later. A lost packet then becomes deferred information rather than lost information.
We propose to make the drop-or-retransmit decision, and the transmit power on each attempt, from a single quantity: the expected decrease of an optimization Lyapunov function per joule spent. A Lyapunov function here is a scalar "energy" that combines the loss gap with the size of every device's unsent memory, and that provably shrinks as training proceeds. The resulting rule is simple and can be computed on the device. A device whose own memory error is large pays for a reliable transmission. A device whose memory error is small transmits cheaply, accepts a likely loss, and lets its buffer absorb it. The work delivers a convergence proof for error feedback over lossy, power-controlled links, a closed-form per-device power rule derived from that proof, and an empirical test of whether optimizer-aware reliability saves energy compared with conventional reliable HARQ.
Core Claim
The key observation is that a packet erasure, when the device knows the outcome through the HARQ acknowledgement, can be modeled as a random compressor. Take EF21, an error-feedback variant in which device $i$ and the server share a running gradient estimate $g_i^t$, and the device sends a compressed correction $C(\nabla f_i(x^{t+1}) - g_i^t)$. Suppose that correction reaches the server with probability $q_i(P_i, h_i)$, which depends on transmit power $P_i$ and channel gain $h_i$. If both sides leave $g_i$ unchanged after a loss, they stay synchronized, because the acknowledgement tells the device what the server holds. The device-to-server pipeline is then a compressor whose contraction factor is roughly $q_i\alpha$, where $\alpha$ is the contraction of the compressor itself, for example Top-$K$.
Existing EF21 Lyapunov analyses take the contraction factor as their input. We can therefore use the standard function
$$\Phi_t = f(x_t) - f^\star + \frac{c}{n}\sum_{i=1}^n \|g_i^t - \nabla f_i(x_t)\|^2,$$
where $f$ is the global loss, $f^\star$ its minimum, $n$ the number of devices, and $c$ a constant set by the stepsize and contraction. This function lets us ask how much each device's power choice lowers $\Phi$. Device $i$ can compute its own contribution, because it knows both its fresh gradient and the server's copy $g_i^t$. Write $D_i = \|\nabla f_i(x_{t+1}) - g_i^t\|^2$ for this local mismatch. The per-round decision then minimizes expected drift plus an energy price $\lambda$:
$$\min_{P_i \ge 0}\; -\,c\,\alpha\, q_i(P_i,h_i)\, D_i + \lambda\, E_i(P_i),$$
where $E_i(P_i)$ is the energy of the attempt. A retransmission is worth making only if the same inequality still favors spending after the first failure. Power therefore grows with the device's memory error and shrinks in deep fades. In very poor channel states the optimal choice is to stay silent, which recovers the transmission-halting behavior that Effect of Strong Time-Varying Transmission Distance on LEO Satellite-Terrestrial Deliveries found useful for energy. Our version also says when halting is safe for learning.
How This Differs From Existing Work
Four existing lines of work come close, and each covers a different part of the problem.
The first motivates the proposal and contributes its main mechanism. Green Communication via Power-Optimized HARQ Protocols shows that starting at low power and ramping up only on failure saves energy. Power Allocation for Finite-Blocklength IR-HARQ gives a tight outage bound for short packets and casts power allocation across attempts as a geometric program. Both optimize against a fixed reliability target. We reuse the finite-blocklength outage bound as our $q_i(P,h)$. We drop the fixed target and replace it with a learning-dependent value of success.
The second line supplies the analysis. A Tight Theory of Error Feedback Algorithms in Distributed Optimization finds exact Lyapunov certificates for EF and EF21 by computer-aided proof search, and Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum builds Lyapunov functions for normalized EF21 under gradient-dependent curvature. Both assume bits arrive without error and that the compression level is fixed. We add erasures whose probability the device controls through power.
The third line is the closest wireless work, and it is myopic. Over-the-Air Federated Edge Learning with Error-Feedback One-Bit Quantization and Power Control pairs error feedback with power control, but it chooses power to minimize signal misalignment in each round, without a convergence argument linking the two. Low-Rank Gradient Compression with Error Feedback for MIMO Wireless Federated Learning similarly designs beamforming separately from any rate proof.
The fourth line is closest in spirit. Optimized Power Control Design for Over-the-Air Federated Edge Learning shows that power should depend on optimization progress, with less power early and more late. Its analysis relies on the Polyak–Łojasiewicz inequality, a condition that ties gradient size to the loss gap, and it assumes analog aggregation with no memory. Lyapunov-Based Optimization of Edge Resources for Energy-Efficient Adaptive Federated Learning uses drift-plus-penalty scheduling, a queue-stability technique from network control, but for non-convex networks it falls back on a fitted sigmoid surrogate of validation loss.
Our contribution is to put the optimizer's own Lyapunov function in the place of the network-control queue, which removes that surrogate. We also target digital links with HARQ rather than analog over-the-air aggregation. None of the papers above treats retransmission itself as a learning-aware decision.
Research Program
The theory has three parts. First, we prove that EF21 with acknowledged erasures and contractive compression converges for non-convex smooth losses, with the usual EF21 stepsize condition holding under the effective contraction $q\alpha$. We also show how heterogeneous success probabilities across devices enter the rate. Next, we prove that the per-round, state-dependent power rule keeps $\Phi$ decreasing in expectation, and we bound the extra loss gap it incurs relative to always-reliable transmission as a function of $\lambda$. This is the analogue of the familiar trade-off in drift-plus-penalty methods between cost and how far the controlled quantity is allowed to drift. Finally, for the strongly convex case, we encode the erasure compressor in the automated proof-search setup of A Tight Theory of Error Feedback Algorithms in Distributed Optimization. This checks numerically whether our hand-derived rate is tight or merely sufficient.
As a secondary extension, we carry the rule over to normalized EF21 under $(L_0,L_1)$-smoothness, where curvature grows with gradient norm. Normalization changes how $D_i$ scales, and this extension is listed in the file as an open gap.
The system work has two parts. We implement the rule using the finite-blocklength outage expression for $q_i$. To address the imperfect-CSI concern raised in the file, we also implement a variant in which $q_i$ is averaged over the posterior of a delayed channel estimate. CSI means channel state information, the device's estimate of the current link quality.
Baselines
The naive reference point is EF21 with Top-$K$ compression, transmitting every attempt at the maximum allowed power, with HARQ retrying up to a fixed cap. It shows what is spent when energy is ignored.
The strongest competitor is reliable, power-optimized HARQ underneath EF21. Per-attempt power comes from the geometric program of Power Allocation for Finite-Blocklength IR-HARQ, with outage targets swept from $10^{-1}$ to $10^{-5}$. This is what a well-engineered communications stack would do today, and it already captures most of the savings available from ramping power. We found no result on Lacuna for EF21 run over finite-blocklength IR-HARQ on any learning benchmark, so we must run this baseline ourselves. To keep the comparison fair, it gets the same channels, compressor, stepsize sweep, and number of tuning trials as our method. Our method has a single price $\lambda$, and the baseline's outage target is the matching single knob. We count our method as beating this baseline only if it reaches the same test accuracy with at least 20% less total uplink energy, measured across three seeds and both channel models, while staying within one point of the baseline's final accuracy.
The third baseline is "never retransmit": one attempt per round at a single constant power, tuned by grid search for energy-to-accuracy. It tests whether error feedback alone makes HARQ unnecessary.
The fourth baseline adapts Lyapunov-Based Optimization of Edge Resources for Energy-Efficient Adaptive Federated Learning. It uses drift-plus-penalty with a sigmoid loss surrogate, and its power is chosen by the queue rule rather than by $D_i$.
The ablation that isolates our key ingredient is the static Lyapunov-optimal power. It solves the same energy-versus-contraction trade-off once, offline, using average channel and average $D$, and then holds that policy fixed. Its gap to our method measures the value of making power depend on each device's current memory error.
For context only, published numbers come from different, analog setups. Joint Device Selection and Power Control for Wireless Federated Learning reports 71.52% CIFAR-10 accuracy against 66.47% for MSE-thresholding, and 92.75% on MNIST against 93.09% for error-free FedAvg. Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum reports up to 10% accuracy gain for normalized EF21 over EF21 on ResNet-20 at equal communication. These numbers anchor plausible accuracy ranges, but they are not directly comparable to our energy-to-accuracy metric.
Why not just…?
Why not just run reliable HARQ and let the learning layer stay unaware? This is the most direct threat. If power-optimized HARQ already operates close to the energy floor, any gain left for learning-aware dropping may be small. The argument for our method is structural. Reliable HARQ pays its steepest energy cost on retries in deep fades, which are exactly the cases where an error-feedback buffer makes a loss cheap. The outage-target sweep in the baseline settles the question empirically. If some fixed outage target matches our energy-to-accuracy on every channel model, then the honest conclusion is "relax HARQ reliability for gradients," and the state-dependent rule adds little.
Why not just drop every lost packet at a tuned constant power? Error feedback does tolerate losses. However, the EF21 analysis implies that the rate degrades with the effective contraction $q\alpha$, so a device whose memory error has grown large slows the whole system. A constant power cannot tell that device apart from one with nothing important to send. The never-retransmit baseline and the static-power ablation measure this directly. We expect the gap to appear mainly under non-i.i.d. data and heterogeneous fading, and to be small under i.i.d. data with identical channels. We will report both regimes.
Why not just reuse optimization-progress power control? Optimized Power Control Design for Over-the-Air Federated Edge Learning adapts power over training through a global schedule tied to the loss gap. Its analysis is analog and memoryless, and it uses no per-device signal that indicates whose update matters. Our rule depends on the local quantity $D_i$, which only exists once error-feedback memory is present. The drift-plus-penalty baseline, together with a global-schedule variant of our rule that ignores $D_i$, will show whether this per-device signal matters.
Proposed Experiments
The benchmarks are the ones used in the file. We train ResNet-20 on CIFAR-10, following the error-feedback setup of Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum, under both i.i.d. and Dirichlet non-i.i.d. splits across 20 devices. We also train the 6-layer CNN on MNIST and non-stationary MNIST used in the over-the-air papers. LIBSVM logistic regression serves to check theory against measurement.
There are two channel models. The first is Rayleigh block fading with heterogeneous path loss. The second is the finite-state Markov channel of the LEO study, which includes deep-fade states. Energy uses the finite-blocklength model. The main metrics are total uplink energy to reach a target accuracy (70% on CIFAR-10, 97% on MNIST), accuracy at a fixed energy budget, and average retransmissions per round.
Four checks decide the outcome. First, our method must beat optimized reliable HARQ on energy-to-accuracy by the margin stated in the Baselines section. Second, it must beat the static-power ablation under heterogeneous fading; if it does not, the per-device adaptive rule fails and only the static theory survives. Third, the per-round decrease of $\Phi$ predicted by the theory should track the measured decrease on logistic regression, where $f^\star$ is computable. On the strongly convex case, the automated proof search should confirm or tighten our rate. Fourth, under delayed CSI, the posterior-averaged rule should keep its advantage, while a variant that trusts stale CSI should lose it.
References
- Effect of Strong Time-Varying Transmission Distance on LEO Satellite-Terrestrial Deliveries
- Green Communication via Power-optimized HARQ Protocols
- Power Allocation for Finite-Blocklength IR-HARQ
- A Tight Theory of Error Feedback Algorithms in Distributed Optimization
- Error Feedback under $(L_0,L_1)$-Smoothness: Normalization and Momentum
- Over-the-Air Federated Edge Learning with Error-Feedback One-Bit Quantization and Power Control
- Low-Rank Gradient Compression with Error Feedback for MIMO Wireless Federated Learning
- Optimized Power Control Design for Over-the-Air Federated Edge Learning
- Lyapunov-Based Optimization of Edge Resources for Energy-Efficient Adaptive Federated Learning
- Joint Device Selection and Power Control for Wireless Federated Learning