Deployed classifiers are often asked to satisfy more than one fairness rule at once. Examples include equal true-positive rates across groups, equal false-positive rates, and equal precision. Theory says some of these combinations can coexist and others cannot. Maximal Combinations of Fairness Definitions characterizes exactly which sets of seven common definitions are jointly satisfiable: 12 maximal combinations, no set of four, and demographic parity the most restrictive. That paper gives no training procedure. Practical in-processing methods such as FairGrad handle one constraint through a weighted sum of losses. Meanwhile the multi-task optimization literature has shown repeatedly that weighted sums cannot fix gradients pointing in opposite directions; reweighting only changes which objective loses (ConFIG, Gradient Surgery for Multi-Task Learning).
We propose an optimizer step built for this setting. At each step, the parameter update is projected so that it does not worsen any fairness constraint that is currently violated, while staying as close as possible to the accuracy-improving direction. The projection is applied to the update the optimizer actually takes, after Adam's momentum and rescaling, not to the raw gradient. The same projection carries a second payoff. When no such update exists, a classical theorem of the alternative (Gordan's) says some nonnegative mix of the constraint gradients cancels to zero. That gives a measurable, per-step certificate that the chosen fairness rules are in direct opposition on this model and data. Our central scientific bet is that this gradient-level certificate tracks the population-level feasibility theory of Defrance and De Bie, which would make "are my fairness requirements compatible?" something a practitioner can check during training.
Core Claim
Two claims are tested separately, so that one can fail without sinking the other.
The first is about measurement. Let $a_1,\dots,a_K$ be the gradients, with respect to the shared parameters, of smoothed versions of the $K$ requested group-disparity measures. Define the incompatibility score
$$ \rho_t = \min_{\lambda \ge 0,\ \sum_k \lambda_k = 1} \Big\| \sum_k \lambda_k \hat a_k \Big\|, $$
where $\hat a_k = a_k/\|a_k\|$ are the normalized constraint gradients at step $t$. This is the distance from the origin to their convex hull, the same minimum-norm problem the multi-gradient descent algorithm solves. A score $\rho_t > 0$ means some update reduces every disparity at once. A score of $\rho_t \approx 0$ is Gordan's certificate that none exists. We predict that, late in training, $\rho_t$ stays bounded away from zero for combinations inside one of Defrance and De Bie's maximal sets. We predict it collapses toward zero for the classic infeasible case: precision parity together with false-positive-rate parity and false-negative-rate parity, under unequal base rates.
The second is about the method. When the cone is non-empty, projecting the optimizer's actual update onto it should reach a given maximum violation at higher accuracy, and with far less oscillation, than multiplier-based training. This is the measured comparison that settles the claim.
Research Program
The building blocks are standard. Each fairness rule is turned into a differentiable disparity $c_k(\theta)$, the gap between groups in a sigmoid-smoothed group rate, with a tolerance $\epsilon$. Minibatch estimates of group rates are noisy, so each $c_k$ and $a_k$ is maintained as an exponential moving average over batches. This is the momentum-buffer idea of M-ConFIG from ConFIG, applied to constraints rather than tasks. With $K \le 4$ the stored buffers are cheap.
The step proceeds as follows. Adam first produces its proposed update $u_t$ from the task-loss gradient alone. The projection then finds the update $v$ closest to $u_t$ such that, for every active constraint (those with $c_k > \epsilon$), the descent step $-v$ reduces that constraint at a rate proportional to its excess:
$$ v_t = \arg\min_{v} \|v - u_t\|^2 \quad \text{subject to} \quad \langle v, a_k \rangle \ge \beta\,(c_k - \epsilon) \ \text{for each active } k. $$
Here $\beta > 0$ sets how aggressively violations are repaid. Because the parameter update is $\theta \leftarrow \theta - \eta v$, this inequality makes each active disparity fall to first order. The dual is a $K \times K$ problem in the Gram matrix of the $a_k$, so solving it costs almost nothing beyond the $K$ extra backward passes. Satisfied constraints impose nothing. This is the key difference from symmetric consensus methods, which force every term to decrease.
When the dual becomes unbounded, which happens exactly when $\rho_t \approx 0$, the step falls back to the minimum-norm combination and logs the certificate. Following the state-alignment idea in Gradient-Update Mismatch, Adam's first-moment buffer is then softly pulled toward $v_t$, so that stale momentum does not reintroduce the violation on the next step.
Each piece is borrowed, and the role of each source differs. The asymmetric design, where accuracy is the objective and fairness defines an allowed region, follows primary-task projection as in Directional Gradient Projection for Robust Fine-Tuning of Foundation Models. That work protects a single primary direction. We instead impose several one-sided, violation-gated inequalities. The post-optimizer placement is motivated by the Gradient-Update Mismatch paper, which found that conflict-free gradients become conflicting updates in up to 86.3% of steps under momentum optimizers on physics-informed networks. Whether the same distortion occurs for fairness constraints has not been measured. We measure it directly: the fraction of applied updates that increase an active disparity, with and without post-optimizer projection.
The work also departs from the closest "fair" gradient methods. Multi-Task Learning as a Bargaining Game and SVFair define fairness among task losses, through Nash bargaining or Shapley-weighted conflict. Both treat every objective as something to improve. Neither models group-fairness rules as constraints that only need to be met.
Baselines
The naive reference is unconstrained empirical risk minimization with Adam. It fixes the accuracy ceiling and the starting disparities.
The strongest existing competitor is FairGrad (FairGrad), which turns a fairness constraint into dynamic group reweighting, including negative weights, through Lagrange multiplier updates. It is the strongest because it already targets demographic parity, equalized odds and accuracy parity on deep networks at negligible cost; the file reports only 0.42 seconds per epoch of overhead on ResNet18. We extend it to several constraints by keeping one multiplier per constraint, which is its natural generalization. A plain multi-multiplier Lagrangian with dual ascent is run as a second scalar baseline. For FairGrad's accuracy–fairness trade-off on these datasets, no reported result found on Lacuna: the file states only "comparable to or better than FairBatch." We therefore rerun it ourselves.
The symmetric-surgery competitor is ConFIG, which treats the task loss and each disparity penalty as terms that must all decrease. PCGrad from Gradient Surgery for Multi-Task Learning is run as the pairwise version. A post-hoc per-group threshold adjustment on the ERM model gives the post-processing reference.
All methods share the architecture, data splits, number of epochs and per-constraint tolerance $\epsilon$. Each receives the same hyperparameter search budget (equal trial counts over learning rate plus its own method-specific knobs: multiplier step size, $\beta$, or penalty weights) and five seeds.
Two ablations isolate the proposal's ingredients. Projecting the raw gradient before Adam, instead of the update after it, tests the post-optimizer claim. Replacing the violation-gated one-sided cone with ConFIG's all-terms-decrease rule tests the asymmetry claim.
Beating the strongest baseline means one of two things. Either the proposal achieves higher test accuracy at matched worst-case test violation across the trade-off curve, or it achieves lower worst-case violation at matched accuracy. Either result must hold beyond seed variance on at least three of the five datasets. A lower step-to-step oscillation of disparities counts only as supporting evidence.
Why not just…?
Why not just use one Lagrange multiplier per constraint, FairGrad-style? This is the real threat. At an exact saddle point, a well-tuned Lagrangian and a projection method reach the same constrained optimum. The argument from ConFIG that weights cannot resolve opposed gradients is about the path, not necessarily the endpoint, and that argument was made for physics losses, not fairness. The equal-budget comparison settles this directly. If tuned multi-multiplier FairGrad matches our trade-off curve, we will report that the optimizer offers no accuracy advantage. In that case the contribution narrows to the incompatibility certificate and the gradient-update-mismatch measurement, which multipliers do not provide.
Why not just apply ConFIG or PCGrad with each fairness penalty as a task? Symmetric methods require every term to keep decreasing, including disparities already within tolerance. That spends accuracy on fairness slack nobody asked for, and it fails exactly when the rules are mutually opposed. PCGrad is also order-dependent (SVFair reports it far behind on NYUv2 mean rank, though that is multi-task evidence). The cone-asymmetry ablation shows whether gating by violation matters.
Why not just read feasibility off Defrance and De Bie's theorems and skip the gradient diagnostic? Their results concern exact equality over population confusion matrices for non-trivial classifiers. Practitioners impose tolerances $\epsilon$ with finite-capacity models and finite data, where a theoretically feasible combination may still be in practical conflict and vice versa. Whether $\rho_t$ tracks the theory, and where it departs from it, is precisely the first experiment.
Why not just post-process thresholds? Group thresholds can satisfy some pairs of rules. However, the confusion-matrix geometry in Maximal Combinations of Fairness Definitions constrains what thresholding a fixed score can reach, and multiple simultaneous rules often push it toward randomized predictors. We include it as a baseline rather than dismissing it, and report where it is competitive.
Why not just use plain SGD so momentum cannot distort anything? This would remove the mismatch, but at the usual cost of slower training. We include SGD with constraint projection as an extra run, so the gain from post-Adam projection is measured against this cheaper fix.
Proposed Experiments
The benchmarks are those used by FairGrad: Adult Income, Law School, UTKFace across racial groups, and Twitter Sentiment, which has 200k tweets split by African American English versus Standard American English. To these we add CelebA "smiling" with gender as the sensitive attribute.
On each dataset we run two kinds of constraint sets. The first is one triplet drawn from the maximal combinations enumerated by Defrance and De Bie. The second is the infeasible set of precision parity, false-positive-rate parity and false-negative-rate parity, on datasets where base rates differ between groups. A demographic-parity pair is added to probe the paper's claim that demographic parity is the most restrictive rule.
The first decisive check is whether late-training $\rho_t$, averaged over seeds, cleanly separates the feasible from the infeasible constraint sets. As a dose–response test, we also check whether $\rho_t$ falls as we synthetically widen the base-rate gap by resampling. If $\rho_t$ fails to separate them, the diagnostic claim is false, whatever the optimizer achieves.
The second check is the matched-violation accuracy comparison against multi-multiplier FairGrad, under the protocol described in the Baselines section.
The third check is the fraction of Adam updates that increase an active disparity, with pre-optimizer versus post-optimizer projection. For comparison, the only reported figures are from physics-informed networks: an 86.3% conflict rate, cut to 0% by post-optimizer alignment, at roughly 1.2–1.6x training time (Gradient-Update Mismatch). Finding a substantial rate in fairness training would establish that the mismatch is a general hazard for in-processing fairness methods, not only for physics losses.
References
- Maximal Combinations of Fairness Definitions
- FairGrad: Fairness Aware Gradient Descent
- ConFIG: Towards Conflict-free Training of Physics Informed Neural Networks
- Gradient Surgery for Multi-Task Learning
- Gradient-Update Mismatch: Rethinking Conflict-Free Training of Physics-Informed Neural Networks
- Directional Gradient Projection for Robust Fine-Tuning of Foundation Models
- Multi-Task Learning as a Bargaining Game
- From Gradient Volume to Shapley Fairness: Towards Fair Multi-Task Learning