Federated learning trains one shared model across many devices or institutions whose data differ. A popular fairness fix, proportional fairness, maximizes the sum of the logarithms of each participant's "utility." The logarithm makes a 10% gain for a poorly served participant count for more than the same absolute gain for a well-served one. In bargaining theory, this rule is defined relative to a disagreement point: what each participant would get if cooperation failed. The leading federated version, PropFair, replaces that point with a single hand-tuned constant $M$, a nominal "worst possible loss" shared by every client. That choice is not neutral. It decides how strongly the objective bends, when gradients blow up, and which clients the objective actually protects. It has also never been measured apart from the rest of the method.
This note proposes to test the disagreement point directly. We replace PropFair's shared constant with a per-client baseline that comes from each client's own situation: the loss a client reaches by training alone on its own data, which is the literal "no deal" outcome. We then run a tuning-budget-controlled empirical study on naturally partitioned federated benchmarks. The study asks three questions: whether this removes the $M$ hyperparameter, whether it changes who benefits, and whether the guarantee bargaining theory promises actually shows up in trained models. That guarantee is individual rationality: no client ends up worse off than it would have been alone.
Core Claim
PropFair (Proportional Fairness in Federated Learning; Equality Is Not Equity) minimizes $-\sum_i p_i \log(M - f_i(\theta))$. Here $f_i(\theta)$ is client $i$'s loss under the shared model $\theta$, $p_i$ is its data share, and $M$ is a scalar the user picks. If $M$ sits close to the largest client loss, the gradient term $\nabla f_i/(M-f_i)$ explodes. If $M$ is far above it, the log flattens into a plain average and fairness disappears.
We propose the per-client version:
$$\min_\theta \; -\sum_{i=1}^m p_i \log\big(d_i - f_i(\theta)\big),$$
where $d_i$ is client $i$'s held-out loss from a model trained only on its own data.
Three predictions follow. First, there is no global scale to tune: each term is measured against that client's own baseline, so rescaling one client's loss no longer shifts the weight on everyone else. Second, the log acts as a soft barrier at $f_i = d_i$, so the trained model should rarely leave a client worse off than going alone. Third, and most interesting, this reverses whom the objective emphasizes. Data-poor clients have weak standalone baselines, so they gain a lot and are easily satisfied. Data-rich clients, whose standalone model is already good, sit near the barrier and receive the most protection. They are exactly the participants with the strongest reason to leave a federation.
Whether this is a better fairness notion than PropFair's "help the struggling" notion is itself an empirical question. We answer it by reporting both kinds of metric rather than presuming one.
Positioning Against Prior Work
The closest work fixes the disagreement point by fiat. PropFair uses the shared constant $M$ and admits it needs heuristic tuning. It also stabilizes training with a "huberized" logarithm that turns linear near zero. That truncation voids the bargaining axioms, and its effect is never quantified. Fairness-Aware Meta-Learning via Nash Bargaining applies Nash bargaining to conflicting subgroup gradients but sets the disagreement point to zero and flags that choice as unrealistic. Our proposal differs from both in two ways: the disagreement point is data-derived and different for each client, and it is treated as the object of study rather than a nuisance.
Two other methods sit on the same spectrum between utilitarian and egalitarian objectives. Fair Resource Allocation in Federated Learning introduced q-FFL, which raises each loss to a power and so reweights clients along that spectrum. It reports poor conditioning at large $q$ and client drift under local updates. Agnostic federated learning optimizes only the worst client.
Two further works motivate the evaluation design. LEAF: A Benchmark for Federated Settings argues that fairness results on synthetic partitions can hide systems costs. It supplies natural per-user partitions with communication accounting. The survey Federated Learning for Generalization, Robustness, Fairness documents that more than 100 federated methods are evaluated in stylized settings, and it never tests PropFair. We reuse LEAF's partitions and accounting, not its conclusions.
Research Program
Computing and stabilizing the baseline. Each client trains a small local model on its training split, early-stopped on a local validation split. It reports the scalar $d_i$, the same kind of loss scalar PropFair already transmits. Some clients will start with $f_i(\theta_0) > d_i$, which makes the log undefined. We handle this in two ways: a short FedAvg warm start, and a floored log below a margin $\delta$.
We instrument the floor. For every run we log how many client-rounds operate in the floored, non-axiomatic region. This turns PropFair's unmeasured huberization question into a number.
Removing the tuning advantage. Every method receives the same fixed budget of full training runs for hyperparameter search. That covers $M$ for PropFair, $q$ for q-FFL, and learning rates for all methods. We report the result at each budget level, not just the best result after unlimited search. This makes the practical cost of $M$ visible.
Separating mechanism from metric. We report five metric families side by side:
- Mean accuracy.
- Worst-10% client accuracy, the metric PropFair optimizes for.
- The individual-rationality violation rate: the fraction of clients whose shared-model test accuracy falls below their standalone accuracy.
- The magnitude of the worst such violation.
- Systems cost from LEAF's accounting: communication rounds to target accuracy and bytes transferred under partial participation.
Runs use LEAF and decentralizepy so that partial participation and local multi-step updates are realistic.
Baselines
The do-nothing references. FedAvg shows what the average-loss objective does to the tails. Local-only training defines $d_i$ and is itself a reference: a federation that loses to it for many clients has failed on its own terms. PropFair reports FedAvg reaching 25.26% worst-10% accuracy on CIFAR-100 (Equality Is Not Equity).
The strongest competitor: PropFair with a tuned global $M$, including its huberization. It is the direct object of the claim and holds the best reported numbers on the question. PropFair reports 31.84% mean and 28.85% worst-10% accuracy on CIFAR-100, and 64.75% mean and 58.14% worst-10% on CIFAR-10 (Proportional Fairness in Federated Learning). We first replicate these numbers on its Dirichlet partitions with its reported $M$, to confirm our pipeline. Then we rerun it under the shared tuning budget.
We "beat" PropFair if either of two things holds:
- We match its worst-10% accuracy within seed noise with zero $M$ search, while cutting the individual-rationality violation rate.
- We clearly exceed it when both methods get a small budget.
If a well-tuned $M$ matches us on every metric, the idea offers no accuracy gain, and we will say so.
Other fairness baselines. q-FFL is tuned over $q$ under the same budget. Its Vehicle-sensor result, a bottom-10% accuracy of 69.9% versus 43.0% for FedAvg (Fair Resource Allocation in Federated Learning), sets the bar on that dataset. Agnostic federated learning is included where full participation is feasible.
Ablation of the log. We keep the per-client baseline but drop the logarithm, optimizing $\sum_i p_i (d_i - f_i)$. Because $d_i$ is constant, this reduces exactly to FedAvg. So any gap between that variant and ours isolates the log's contribution: the curvature and the barrier.
Ablation of the baseline's source. This second ablation swaps the standalone baseline for $d_i = f_i(\theta_{\text{warm}})$, the client's loss at the warm-start model. It tests whether the standalone meaning of $d_i$ matters or merely its per-client scaling.
Why not just…?
Why not just grid-search $M$ more carefully? With enough search, a scalar $M$ may well match our worst-10% accuracy. This is the most likely null result. But a scalar cannot encode per-client baselines, so it cannot target individual rationality. The budget-controlled comparison and the violation-rate metric will settle which of the two matters in practice. If tuned $M$ also drives violations near zero, the per-client idea reduces to a tuning shortcut, and we will report it as such.
Why not just use q-FFL with a tuned $q$? q-FFL reweights by absolute loss magnitude. It has no reference point and so no notion of "worse than alone," and its own authors report conditioning and drift problems as $q$ grows (Fair Resource Allocation in Federated Learning). It is included at equal budget. If it matches us on violation rate too, our claim fails.
Why not just fine-tune the shared model locally on each client? This is a real threat. Local fine-tuning usually beats standalone training and would trivially satisfy individual rationality. Our claim is therefore restricted to settings that need one shared model: a model serving users unseen in training, devices without compute budget to fine-tune, or regulated settings with one audited model. We will still report the fine-tuned reference, so readers see how much of the gap personalization closes.
Why not impose hard "no worse than alone" constraints with Lagrange multipliers? Constrained training targets individual rationality more directly. It adds per-client multipliers with their own step size, which is another tuning burden. It also only enforces the threshold, without distributing surplus above it. We include a primal-dual variant as a secondary baseline. If it dominates, the log barrier is unnecessary.
Decisive Experiments
Replication. We reproduce PropFair on CIFAR-10 and CIFAR-100 Dirichlet partitions.
Main study. The main study runs on naturally partitioned data: LEAF's FEMNIST, Shakespeare, and Sent140, plus the Vehicle sensor data. These give realistic spreads of client size, where the reversal of emphasis toward data-rich clients should be most visible. No reported result was found on Lacuna for PropFair on Shakespeare or any LEAF split, so these runs are new measurements. Each configuration uses five seeds under partial participation with multiple local steps.
Three decisive checks. The claim holds only if all three succeed:
- Our method's worst-10% accuracy lies within PropFair's best-tuned value while spending no search budget on $M$.
- Its individual-rationality violation rate is significantly lower than that of PropFair, q-FFL, and FedAvg.
- The share of client-rounds in the floored log region falls as training proceeds. This shows the barrier is doing work rather than being silently replaced by the linear floor.
A failure of the third check would itself be a useful negative result: log-barrier fairness in federated learning would then run mostly on its heuristic smoothing, not on the bargaining rule it cites.
References
- Proportional Fairness in Federated Learning
- Equality Is Not Equity: Proportional Fairness in Federated Learning
- Fairness-Aware Meta-Learning via Nash Bargaining
- Fair Resource Allocation in Federated Learning
- LEAF: A Benchmark for Federated Settings
- Federated Learning for Generalization, Robustness, Fairness: A Survey and Benchmark