When a model is trained across many phones or edge devices, a few devices may hold nearly all the examples from rare demographic intersections, such as older rural women of a particular ethnicity. The update from such a device cannot be replaced: no other device carries that data. Many wireless systems let several devices transmit at once on the same frequencies. The receiver separates their signals one at a time, decoding the strongest, subtracting it, and moving to the next. Under this scheme the device decoded first faces the most interference and is the most likely to fail, while the device decoded last sees a clean channel. Receivers usually choose this order by signal strength alone. If weak signals come from the same places where rare intersectional data lives, such as the cell edge or poorer areas with older devices, the standard order systematically loses the updates that fairness depends on most. No later reweighting at the server can recover an update that never arrived.
We propose to set the decoding order and transmit powers each training round according to how irreplaceable each device's data is for the currently worst-served intersectional subgroups. The goal is to keep rare subgroups above the sample sizes at which their error can be estimated and constrained at all. The core claim is that, when channel quality and demographic sparsity are correlated and channels change slowly, ordering the decoding by sparsity-weighted data value beats every combination of standard signal-strength ordering and server-side fairness reweighting, on both worst-subgroup error and the number of subgroups that remain auditable. When they are uncorrelated, the advantage should disappear. This prediction lets the experiments falsify the idea.
Background and the gap
Wireless research on non-orthogonal multiple access (NOMA), where users share the same time and frequency and are separated by successive interference cancellation (SIC, the decode-subtract-repeat procedure above), has shown that decoding order is a powerful control. Ding, Schober and Poor show that ordering purely by channel strength produces an outage "error floor," where more transmit power no longer improves reliability. Their hybrid SIC, which switches order when channel gains cross a threshold, removes the floor. Sun et al. show that hybrid SIC with power adaptation protects a designated primary user. Secondary users throttle their power to an exact interference threshold, which raises diversity from 1 under fixed ordering to $M$, the number of secondary users. Tang et al. show that the optimal downlink order flips with total power rather than instantaneous channel state. These papers supply our mechanism, but in all of them the protected party is defined by service contracts or channel geometry, never by the value of its data. Physical-layer fairness work such as Zhu et al. maximizes the minimum throughput across devices and explicitly leaves dynamic decoding order open. Equal throughput is not equal informational value, which is the distinction we exploit.
On the learning side, intersectional fairness research documents why rare cells matter and why they are fragile. Paes et al. prove a sharp cliff: once there are more groups than samples per group, verifying subgroup gains becomes statistically impossible. Ferrara et al. describe a matching resolution limit for audits. The practical response, in DRAF and in Němeček et al., is to constrain only "active" subgroups whose size exceeds a threshold $\gamma$, which silently drops the rarest intersections. These papers motivate our target. Every lost update from a rare-cell holder shrinks the active set, so the transmission layer directly decides which subgroups the fairness method can see. All of them assume lossless centralized data.
The closest bordering work, Ni et al., combines over-the-air federated learning with NOMA. Its decoding order is fixed by design, NOMA users first and learning signals second, and it has no notion of demographic subgroups. We differ on both counts: our order is chosen per round, and the quantity driving it is subgroup sparsity.
Core Claim
Each round, the server keeps a weight vector over intersectional subgroups that emphasizes the currently worst-served ones. This is DRAF's adversarial weight $v$, reused here as a scheduling signal rather than only a loss weight. Let $n_{k,s}$ be the (noised, coarsened) count of subgroup $s$ examples held by device $k$, and $n_s$ the total across devices. Device $k$'s irreplaceability score is
$$w_k = \sum_s v_s \frac{n_{k,s}}{n_s},$$
which is large exactly when a device holds a big share of a heavily weighted, rare subgroup. The receiver then chooses a decoding order $\pi$ and powers $p$ to maximize the expected irreplaceable mass that arrives intact,
$$\max_{\pi, p} \sum_k w_k \Pr[\text{device } k \text{ decoded} \mid \pi, p, \text{channel}] \quad \text{subject to per-device energy and rate limits.}$$
We solve this by generalizing hybrid SIC with power adaptation from one primary user to tiers. High-$w_k$ devices are treated as protected and placed late in the order. Lower-tier devices are either decoded early when their channels are strong, or have their power capped at the threshold that preserves protected-device decoding, as in Sun et al.. The $K!$ orderings are handled by restricting to tier-respecting orders and adding violated orderings lazily, borrowing the cutting-plane idea that Němeček et al. use to tame millions of subgroup constraints. The server keeps its fairness-aware aggregation (DRAF or FairGrad weights) unchanged. Our change is entirely in which updates arrive.
Research Program
The first component is a testbed in which demographics and channel quality are correlated in a realistic way. We partition folktables ACS 2018 California into devices by geographic area. Each area is placed at a distance from a simulated base station using a configurable mapping, either rurality to distance or random. A single parameter $\rho$ then controls how strongly demographic sparsity tracks path loss. Channels follow the Rayleigh fading model used by the hybrid-SIC papers, with a coherence parameter running from i.i.d. per round to quasi-static. Communities and Crime, with 18 sensitive attributes and more than 262,000 potential subgroups, provides the extreme-sparsity stress test. Adult provides a small 16-subgroup sanity case.
The second component is the scheduler itself: the score $w_k$, the tiered hybrid-SIC power rule, and the lazy order search. It will be built for single-antenna and multi-antenna receivers. The multi-antenna case matters because Zhu et al. find that orthogonal time-sharing beats NOMA when the access point has a single antenna.
The third component is analysis. We will measure how many subgroups stay above DRAF's activity threshold $\gamma$ after transmission, against round count, $\rho$, and coherence. This quantity connects the wireless choice to the Paes et al. cliff.
We will also test robustness to the privacy-preserving noise added to the counts $n_{k,s}$. Devices reporting exact per-cell counts would leak information, so we evaluate coarsened and noised counts and report how quickly the gain degrades.
Baselines
The naive reference is federated averaging over NOMA with standard strongest-first ordering and plain server averaging. This is what a deployment does by default.
The strongest existing competitor is that same channel-strength ordering, upgraded to hybrid SIC per Ding et al., combined with the best server-side intersectional method, DRAF (Kim et al.), and alternatively FairGrad (Maheshwari et al., via its PyPI package). It is the strongest because it already removes the outage floor and already applies the best fairness correction available. Its only weakness is that the fairness signal never reaches the physical layer. Neither paper reports numerical accuracy or fairness values in the file: DRAF is reported only qualitatively as outperforming GerryFair on Communities and Crime, with about 27% runtime overhead. We therefore have no reported number on Lacuna for DRAF accuracy or fairness on these benchmarks and will reproduce it ourselves. The relevant reported physical-layer figure is the diversity gain, $M$ for hybrid SIC with power adaptation versus 1 for fixed ordering (Sun et al.). Our tiered scheme must not fall below that on protected devices.
We also include max-min throughput allocation in the style of Zhu et al., and fairness-aware time-division access, where the same scores $w_k$ allocate orthogonal slots.
All methods share the same data partition, per-device energy budget, number of rounds, model architecture, and an equal hyperparameter search budget. Beating the strongest baseline means three things at matched sum rate within 5%: lower worst-active-subgroup error, lower DRAF supremum-IPM unfairness, and more subgroups above $\gamma$, all on held-out data across at least five seeds.
The ablations isolate the key ingredient. We replace $w_k$ with data volume only (no sparsity weighting), with random scores, and with scores frozen at round 0 (no dynamic $v$). We also separate order-only control from power-only control.
Why not just…?
Why not just reweight at the server and keep standard decoding? Reweighting rescales updates that arrived; it cannot create one that was lost. Under fast i.i.d. fading every device eventually gets through, and server reweighting may suffice. Our coherence sweep is designed to show exactly that null. Under quasi-static fading with $\rho>0$, the same edge devices fail round after round, and their cells drop below $\gamma$. DRAF's design then removes those cells from its constraints by construction. The experiment settles where the crossover lies.
Why not just use orthogonal access with fairness-aware slots? This is a real threat. Zhu et al. report time-division beating NOMA with a single-antenna access point, while Ding et al. argue that orthogonal access has bounded spectral efficiency. We run the orthogonal variant with the same $w_k$ scores. If it matches us in the single-antenna regime, we will say so and restrict the claim to multi-antenna receivers and high-load settings, where superposition admits more devices per round.
Why not just equalize throughput across devices? Max-min throughput spends energy lifting every weak device, including those whose data duplicates common subgroups. It ignores that device $k$'s worth depends on which cells it holds. The comparison at matched energy shows whether value-aware protection buys more fairness per joule.
Why not just enforce marginal fairness, or smooth sparse cells with a Bayesian prior? Molina and Loiseau show that marginal fairness does not bound intersectional unfairness under dependent attributes, and that Bayesian estimators are distorted by priors at small samples. Paes et al. show that no estimator escapes the sample-size cliff. Keeping real samples arriving is the only lever that moves a cell back across it.
Proposed Experiments and Decisive Checks
The main experiment sweeps correlation $\rho$ from 0 to strong and channel coherence from i.i.d. to quasi-static on folktables Income and Public Coverage, with Communities and Crime as the stress case. It reports worst-active-subgroup error, supremum-IPM, active-subgroup count at DRAF's $\gamma$, sum rate, and protected-device outage.
The decisive check is a shape prediction. The gap over hybrid SIC plus DRAF should be near zero at $\rho=0$ or under i.i.d. fading. It should grow monotonically with $\rho$ and coherence. It should be driven mostly by the sparsity term in $w_k$, not by data volume. If the data-volume ablation recovers most of the gain, the intersectional framing adds nothing and we will report it that way.
A second check uses held-out audit data to test whether subgroups kept above $\gamma$ by our scheduler actually show lower test unfairness, rather than merely appearing in the constraint set.
A third check reruns the main comparison with noised and coarsened cell counts, to show whether the method survives realistic privacy constraints on what devices disclose.
References
- Unveiling the Importance of SIC in NOMA Systems - Part 1: State of the Art and Recent Findings
- Hybrid Successive Interference Cancellation and Power Adaptation: A Win-Win Strategy for Robust Uplink NOMA Transmission
- Optimal Decoding Order and Power Allocation for Sum Throughput Maximization in Downlink NOMA Systems
- Enhancing User Fairness in Wireless Powered Communication Networks With STAR-RIS
- On the Epistemic Limits of Personalized Prediction
- Size-adaptive Hypothesis Testing for Fairness
- Doubly-Regressing Approach for Subgroup Fairness
- Intersectional Fairness via Mixed-Integer Optimization
- Integrating Over-the-Air Federated Learning and Non-Orthogonal Multiple Access: What Role Can RIS Play?
- FairGrad: Fairness Aware Gradient Descent
- Bounding and Approximating Intersectional Fairness through Marginal Fairness