Introduction
Statistical classification models have become increasingly prevalent in high-stakes decision-making applications, including hiring screening [19 ] , criminal risk assessment [1 ] , and healthcare resource allocation [18 ] .
While these models often achieve strong predictive performance, a growing body of empirical evidence [5 ] has revealed that they may systematically disadvantage certain demographic groups defined by sensitive attributes such as gender, race, or ethnicity. Such disparities can lead to unequal treatment across groups and may perpetuate or even amplify existing social inequalities when algorithmic decisions influence access to employment opportunities, judicial outcomes, or healthcare services. Consequently, developing classification models that balance predictive accuracy and fairness has become an important objective in the modern statistical learning community.
To address fairness concerns, prior work has proposed a range of group fairness notions for classification that formalize equity requirements across groups. Prominent examples include Demographic Parity (DP) [7 ] , which requires equal selection rates across groups and Equal Opportunity (EO) [10 ] , which mandates equal true positive rates across groups. These definitions provide measures of algorithmic fairness and are widely used to regulate classifier behavior and mitigate discriminatory outcomes.
Based on these fairness notions, researchers have developed a variety of fairness-aware classification approaches by incorporating fairness violations directly into the learning objective. A common approach [17 , 3 ] is to formulate fairness-aware classification as a fairness-regularized risk minimization problem (See Eq. (6) ), balancing predictive accuracy against fairness violations through a trade-off parameter. Because both the classification error and fairness violations are typically defined through indicator functions over discrete events (See Eq. (3) ), the corresponding optimization problem is generally non-convex and discontinuous.
From an optimization perspective, existing approaches to solve the fairness-regularized risk minimization problem fall into two methodological streams. The first stream attempts to directly optimize the original fairness-aware problem based on the 0 0 0 –1 1 1 loss and exact fairness violations, resulting in a discrete optimization problem. To handle the discrete optimization, mixed-integer programming approaches [23 , 29 ] are adopted for fairness-aware classification. Although the first methodological stream targets the original optimization problem, it typically suffers from substantial computational complexity and limited scalability on large-scale datasets [6 ] .
The second stream takes a different approach by replacing the discontinuous components with convex and smooth surrogate functions. In this paradigm, the 0 0 0 –1 1 1 loss is substituted with surrogate losses such as the logistic, hinge, or exponential loss, while fairness violations are also approximated using surrogate functions [3 ] . To fix notation, let L L L and L ˉ \bar{L} L ˉ denote the surrogate losses for the accuracy and fairness terms, with risks R L ( f ) \mathcal{R}_{L}(f) R L ( f ) and R L ˉ ( f ) \mathcal{R}_{\bar{L}}(f) R L ˉ ( f ) . Surrogate-based methods then minimize
R f a i r L , L ˉ ( f ) ≔ R L ( f ) − λ ⋅ R L ˉ ( f ) , \mathcal{R}^{L,\bar{L}}_{\mathrm{fair}}(f) \coloneqq\mathcal{R}_{L}(f) - \lambda\cdot\mathcal{R}_{\bar{L}}(f), R fair L , L ˉ ( f ) : = R L ( f ) − λ ⋅ R L ˉ ( f ) ,
where λ ≥ 0 \lambda\ge0 λ ≥ 0 is the trade-off parameter. We formalize these definitions in § 2.4 and denote R L ( f ) \mathcal{R}_{L}(f) R L ( f ) , R L ˉ ( f ) \mathcal{R}_{\bar{L}}(f) R L ˉ ( f ) and R f a i r L , L ˉ ( f ) \mathcal{R}^{L,\bar{L}}_{\mathrm{fair}}(f) R fair L , L ˉ ( f ) as L L L -risk, L ˉ \bar{L} L ˉ -risk and ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk with respect to f f f respectively. This optimization problem can be efficiently optimized using standard gradient-based algorithms. As a result, surrogate-based methods are computationally scalable and have become the dominant approach for fairness-aware classification [30 , 11 ] .
The widespread adoption of surrogate-based methods naturally raises a fundamental question: when does minimizing a surrogate objective lead to a statistically valid solution for the original learning problem? In standard classification, this question is addressed through two foundational theoretical concepts. The first is Fisher consistency , which examines whether minimizing the surrogate L L L -risk recovers the Bayes-optimal classifier for the underlying 0–1 loss [16 ] (Yang, Huang, Huang & Chang 2021; Yang, Guo & Chang 2021). The second is classification calibration , which establishes quantitative relationships between the excess L L L -risk and the excess risk through calibrated inequalities [2 ] . However, much less is known about whether these guarantees continue to hold in fairness-aware classification, where recent finite-sample and distribution-free results focus on the fairness-accuracy trade-off [12 ] (Li et al. 2023) rather than on the consistency and calibration of surrogate fairness-aware risks. Unlike standard classification, fairness-aware objectives simultaneously balance predictive accuracy and fairness violation through a fairness-regularized risk formulation. This raises two fundamental questions:
Q1. (Fairness-Aware Fisher Consistency): Under what conditions does the population minimizer of a ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk induce the Bayes-optimal classifier for the corresponding fairness-aware risk?
Q2. (Fairness-Aware Calibration): Can we establish a calibration inequality that relates the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk to the excess fairness-aware risk?
This paper addresses these questions by developing a statistical learning theory for surrogate-based fairness-aware classification. To answer the first question, we introduce the notion of Fairness-Aware Fisher Consistency (FAFC). We begin with a pointwise analysis of the fairness-aware conditional risk and identify a fairness-aware discriminant that determines the Bayes-optimal classifier for the original fairness-aware risk. We then define FAFC for the fairness-aware surrogate loss pointwise. At every pair of conditional probabilities at which the fairness-aware discriminant does not vanish, the smallest conditional risk achievable by a wrong-sign prediction is strictly larger than the optimal conditional risk. We then establish necessary and sufficient conditions under which L f a i r L_{\mathrm{fair}} L fair satisfies FAFC. To answer the second question, we construct a fairness-aware calibration function by minimizing the pointwise wrong-sign excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk over level sets of the fairness-aware discriminant. By taking its convex envelope, we obtain a calibrated inequality that controls the excess fairness-aware risk in terms of the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk. Consequently, under the proposed FAFC conditions, convergence of the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk to zero implies convergence of the excess fairness-aware risk to zero, thereby providing a formal justification for surrogate-based fairness-aware classification.
Our main contributions are summarized as follows:
A pointwise framework and FAFC characterization. We derive the fairness-aware conditional risk for DP and EO in both the attribute-blind and attribute-aware settings, and identify a fairness-aware discriminant that determines the Bayes-optimal classifier. Based on it, we introduce FAFC for surrogate loss pairs ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) , and give necessary and sufficient conditions for FAFC in terms of the derivatives of the partial losses at the origin, which can be checked directly from the loss pair.
A calibration inequality for fairness-aware risks. We construct a fairness-aware calibration function by minimizing the pointwise wrong-sign excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk over level sets of the discriminant, and show that its convex envelope yields an inequality bounding the excess fairness-aware risk by the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk. We further show that this envelope is invertible if and only if the pair is FAFC, so that FAFC is exactly the condition under which surrogate risk convergence implies convergence of the fairness-aware risk.
Closed-form guarantees for widely used surrogates. We verify FAFC and derive explicit calibration functions and excess-risk bounds for fairness-aware logistic, hinge, and exponential losses, covering DP and EO in both feature settings, and we propose a general surrogate loss-based algorithm that can be used with any FAFC-
compliant pair. Simulations and a credit-default application illustrate the accuracy–fairness trade-off.
Preliminaries
We begin by introducing the notation and basic theory of cost-sensitive classification, since the fairness-aware objectives studied in this paper can be naturally represented and analyzed within the cost-sensitive learning framework [17 ] .
Cost-Sensitive Classification and Surrogate Loss
Let ( X , Y ) (X,Y) ( X , Y ) be a random pair drawn from a distribution D D D on X × { − 1 , 1 } \mathcal{X} \times\{-1,1\} X × { − 1 , 1 } , where X \mathcal{X} X is the feature space and Y ∈ { − 1 , 1 } Y \in\{-1,1\} Y ∈ { − 1 , 1 } is the ground-truth label. Let F \mathcal{F} F denote the set of all measurable functions f : X → R f:\mathcal{X} \to\mathbb{R} f : X → R ; each f f f induces a classifier through the decision rule Y ^ = sign ( f ( X ) ) \widehat{Y} = \operatorname{sign}(f(X)) Y = sign ( f ( X )) , with the convention sign ( 0 ) = − 1 \operatorname{sign}(0) = -1 sign ( 0 ) = − 1 .
Definition 1 (Cost-Sensitive Risk and Bayes-Optimal Classifier). For a classifier f ∈ F f \in\mathcal{F} f ∈ F and a distribution D D D , the cost-sensitive risk with cost parameter α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) is defined as
R α ( f ; D ) ≔ ( 1 − α ) ⋅ P ( Y ^ = − 1 , Y = 1 ) + α ⋅ P ( Y ^ = 1 , Y = − 1 ) , (1) \mathcal{R}_{\alpha}(f;D) \coloneqq(1-\alpha) \cdot\mathbb{P}(\widehat{Y}=-1,Y=1) + \alpha\cdot\mathbb{P}(\widehat{Y}=1,Y=-1), \tag*{(1)} R α ( f ; D ) : = ( 1 − α ) ⋅ P ( Y = − 1 , Y = 1 ) + α ⋅ P ( Y = 1 , Y = − 1 ) , (1) where Y ^ = sign ( f ( X ) ) \widehat{Y} = \operatorname{sign}(f(X)) Y = sign ( f ( X )) is the prediction label induced by f f f . Then Bayes-optimal classifier is:
γ ( x ) ∈ arg min f ∈ F R α ( f ; D ) , (2) \gamma(x) \in\arg\min_{f \in\mathcal{F}} \mathcal{R}_{\alpha}(f;D), \tag*{(2)} γ ( x ) ∈ arg f ∈ F min R α ( f ; D ) , (2) where γ ( x ) = η ( x ) − α \gamma(x) = \eta(x) - \alpha γ ( x ) = η ( x ) − α and η ( x ) ≔ P ( Y = 1 ∣ X = x ) \eta(x) \coloneqq\mathbb{P}(Y=1 \mid X=x) η ( x ) : = P ( Y = 1 ∣ X = x ) is the posterior probability of the positive class. The cost parameter α \alpha α controls the relative penalty on false negatives versus false positives; when α = 0.5 \alpha= 0.5 α = 0.5 , R α ( f ; D ) \mathcal{R}_{\alpha}(f;D) R α ( f ; D ) reduces to the misclassification error rate P ( Y ^ ≠ Y ) \mathbb{P}(\widehat{Y} \ne Y) P ( Y = Y ) .
Equivalently, the risk can be written as the expectation of a pointwise cost-sensitive 0–1 loss [21 ] . Let us define
U α ( y , t ) ≔ ( 1 − α ) ⋅ 1 { y = 1 } 1 { t ≤ 0 } + α ⋅ 1 { y = − 1 } 1 { t > 0 } , (3) U_{\alpha}(y,t) \coloneqq(1-\alpha)\cdot1\{y=1\}1\{t\le0\}+\alpha\cdot1\{y=-1\}1\{t>0\}, \tag*{(3)} U α ( y , t ) : = ( 1 − α ) ⋅ 1 { y = 1 } 1 { t ≤ 0 } + α ⋅ 1 { y = − 1 } 1 { t > 0 } , (3)
then R α ( f ; D ) = E ( X , Y ) ∼ D [ U α ( Y , f ( X ) ) ] \mathcal{R}_{\alpha}(f;D)=\mathbb{E}_{(X,Y)\sim D}[U_{\alpha}(Y,f(X))] R α ( f ; D ) = E ( X , Y ) ∼ D [ U α ( Y , f ( X ))] . Here 1 { ⋅ } 1\{\cdot\} 1 { ⋅ } denotes the indicator function.
The 0–1 loss U α U_{\alpha} U α is discontinuous and non-convex, so direct minimization of R α ( ⋅ ; D ) \mathcal{R}_{\alpha}(\cdot;D) R α ( ⋅ ; D ) from finite data is computationally intractable. The standard remedy in statistical learning is to replace U α U_{\alpha} U α by a smooth, convex surrogate loss L : { − 1 , 1 } × R → [ 0 , ∞ ) L:\{-1,1\}\times\mathbb{R}\to[0,\infty) L : { − 1 , 1 } × R → [ 0 , ∞ ) . Any surrogate loss admits the partial-loss decomposition
L ( y , t ) ≔ 1 { y = 1 } L 1 ( t ) + 1 { y = − 1 } L − 1 ( t ) , (4) L(y,t)\coloneqq1\{y=1\}L_1(t)+1\{y=-1\}L_{-1}(t), \tag*{(4)} L ( y , t ) : = 1 { y = 1 } L 1 ( t ) + 1 { y = − 1 } L − 1 ( t ) , (4)
where L 1 ( t ) ≔ L ( 1 , t ) L_1(t)\coloneqq L(1,t) L 1 ( t ) : = L ( 1 , t ) and L − 1 ( t ) ≔ L ( − 1 , t ) L_{-1}(t)\coloneqq L(-1,t) L − 1 ( t ) : = L ( − 1 , t ) denote the partial losses for the positive and negative labels, respectively. Replacing U α U_{\alpha} U α by L L L in Eq. (1) yields the L L L -risk
R L ( f ; D ) ≔ E ( X , Y ) ∼ D [ L ( Y , f ( X ) ) ] . \mathcal{R}_L(f;D)\coloneqq\mathbb{E}_{(X,Y)\sim D}[L(Y,f(X))]. R L ( f ; D ) : = E ( X , Y ) ∼ D [ L ( Y , f ( X ))] .
We denote the minimal risk by R α ∗ ≔ inf f ∈ F R α ( f ; D ) \mathcal{R}_{\alpha}^{*}\coloneqq\inf_{f\in\mathcal{F}}\mathcal{R}_{\alpha}(f;D) R α ∗ : = inf f ∈ F R α ( f ; D ) and the minimal L L L -risk by R L ∗ ≔ inf f ∈ F R L ( f ; D ) \mathcal{R}_{L}^{*}\coloneqq\inf_{f\in\mathcal{F}}\mathcal{R}_{L}(f;D) R L ∗ : = inf f ∈ F R L ( f ; D ) .
To validate that minimizing the L L L -risk yields a statistically valid solution for the original optimization problem, we require two theoretical guarantees: (1) a qualitative asymptotic guarantee that the surrogate recovers the Bayes-optimal classifier, and (2) a quantitative bound that links the excess L L L -risk to the excess risk. We formalize these two properties in the remainder of this subsection.
Fisher Consistency
Fisher consistency is the fundamental asymptotic guarantee for surrogate optimization: the population minimizer of the L L L -risk must recover the Bayes-optimal classifier of the original risk [8 , 16 ] .
Definition 2 (Fisher Consistency for Cost-Sensitive Classification). A surrogate loss L L L is α \alpha α -Fisher consistent for cost-sensitive classification with cost α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) if, for every distribution D D D and every minimizer f ∗ ∈ arg min f ∈ F R L ( f ; D ) f^* \in\operatorname*{arg\,min}_{f\in\mathcal{F}} \mathcal{R}_L(f;D) f ∗ ∈ arg min f ∈ F R L ( f ; D ) , its sign matches that of the cost-sensitive Bayes-optimal classifier:
sign ( f ∗ ( x ) ) = sign ( γ ( x ) ) for P X -a.e. x ∈ X with γ ( x ) ≠ 0 , \operatorname{sign}(f^*(x)) = \operatorname{sign}(\gamma(x)) \quad\text{for } P_X\text{-a.e. } x \in\mathcal{X} \text{ with } \gamma(x) \ne0, sign ( f ∗ ( x )) = sign ( γ ( x )) for P X -a.e. x ∈ X with γ ( x ) = 0 , where γ ( x ) = η ( x ) − α \gamma(x) = \eta(x) - \alpha γ ( x ) = η ( x ) − α is the cost-sensitive discriminant in Eq. (2) .
Satisfying Fisher consistency guarantees that, given the population distribution, minimizing the surrogate loss yields the same classification decisions as the Bayes-optimal classifier almost surely. However, this guarantee is stated at the level of population minimizers and is not directly checkable from the loss. The following result of Scott [21 ] shows that, for convex partial losses differentiable at 0 0 0 , Fisher consistency at cost α \alpha α reduces to a local condition on the derivatives of L 1 L_1 L 1 and L − 1 L_{-1} L − 1 at the origin, which can be read off from the loss alone.
Lemma 3. Let L L L be a surrogate loss with convex partial losses L 1 , L − 1 L_1, L_{-1} L 1 , L − 1 differentiable at 0 0 0 , and let α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) . Then L L L is Fisher consistent at cost α \alpha α if and only if
L 1 ′ ( 0 ) < 0 , L − 1 ′ ( 0 ) > 0 , α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0. L_1'(0) < 0,\qquad L_{-1}'(0) > 0,\qquad\alpha L_1'(0) + (1-\alpha)L_{-1}'(0) = 0. L 1 ′ ( 0 ) < 0 , L − 1 ′ ( 0 ) > 0 , α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0.
Classification Calibration
Beyond the asymptotic guarantee of Fisher consistency, classification calibration provides a quantitative bound relating the excess L L L -risk to the excess risk through a calibration function [2 , 21 ] .
Lemma 4 (Classification Calibration for Cost-Sensitive Classification). Let L L L be a non-negative surrogate loss and α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) .
There exists a convex, non-decreasing calibration function ψ L , α : [ 0 , ∞ ) → [ 0 , ∞ ) \psi_{L,\alpha} : [0,\infty) \to[0,\infty) ψ L , α : [ 0 , ∞ ) → [ 0 , ∞ ) with ψ L , α ( 0 ) = 0 \psi_{L,\alpha}(0) = 0 ψ L , α ( 0 ) = 0 such that, for every measurable classifier f ∈ F f \in\mathcal{F} f ∈ F and every distribution D D D ,
ψ L , α ( R α ( f ; D ) − R α ∗ ) ≤ R L ( f ; D ) − R L ∗ . \psi_{L,\alpha}\left(\mathcal{R}_{\alpha}(f;D)-\mathcal{R}_{\alpha}^{*}\right) \le\mathcal{R}_{L}(f;D)-\mathcal{R}_{L}^{*}. ψ L , α ( R α ( f ; D ) − R α ∗ ) ≤ R L ( f ; D ) − R L ∗ .
The calibration function ψ L , α \psi_{L,\alpha} ψ L , α is invertible if and only if L L L is Fisher consistent in the sense of Definition 2 .
Fairness-Aware Classification
Fairness-aware classification extends the cost-sensitive setting of § 2.1 by adding a binary sensitive attribute A ∈ { − 1 , 1 } A \in\{-1,1\} A ∈ { − 1 , 1 } (e.g., gender or race) alongside the target label Y Y Y . We consider both the attribute-aware setting, in which A A A is available at training and prediction time, and the attribute-blind setting, in which A A A is available at training time only [17 ] . In either setting, the goal is to learn a classifier f ∈ F f \in\mathcal{F} f ∈ F that balances accuracy with respect to Y Y Y against group fairness with respect to A A A .
Group fairness is formalized as a statistical parity condition on the predicted label Y ^ = sign ( f ( X ) ) \widehat{Y} = \operatorname{sign}(f(X)) Y = sign ( f ( X )) across the two groups. We focus on two of the most widely adopted criteria, Demographic Parity (DP) and Equal Opportunity (EO) .
Definition 5 (DP [7 ] ). A classifier satisfies DP if its prediction Y ^ \widehat{Y} Y is statistically independent of the sensitive attribute A A A :
P ( Y ^ = 1 ∣ A = 1 ) = P ( Y ^ = 1 ∣ A = − 1 ) . \mathbb{P}(\widehat{Y}=1 \mid A=1)=\mathbb{P}(\widehat{Y}=1 \mid A=-1). P ( Y = 1 ∣ A = 1 ) = P ( Y = 1 ∣ A = − 1 ) . This criterion demands identical positive prediction rates across the two groups, regardless of their ground-truth label distributions.
Definition 6 (EO [10 ] ). A classifier satisfies EO if it achieves equal true
positive rates across all groups:
P ( Y ^ = 1 ∣ Y = 1 , A = 1 ) = P ( Y ^ = 1 ∣ Y = 1 , A = − 1 ) . \mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = 1) = \mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = -1). P ( Y = 1 ∣ Y = 1 , A = 1 ) = P ( Y = 1 ∣ Y = 1 , A = − 1 ) .
EO relaxes the independence requirement of DP to conditional independence between Y ^ \widehat{Y} Y and A A A given Y = 1 Y = 1 Y = 1 .
Exact parity in Definitions 5 and 6 is generally incompatible with accuracy and cannot be guaranteed on finite data [22 ] . In practice, fairness requirements are therefore relaxed to approximate fairness and quantified by discrimination metrics [10 , 14 ] . We use the two standard metrics, Mean Difference (MD) and Mean Ratio (MR) .
Definition 7 (MD [4 ] ). For a classifier f f f ,
MD D P ( f ) = P ( Y ^ = 1 ∣ A = − 1 ) − P ( Y ^ = 1 ∣ A = 1 ) , \operatorname{MD}_{\mathrm{DP}}(f) = \mathbb{P}(\widehat{Y} = 1 \mid A = -1) - \mathbb{P}(\widehat{Y} = 1 \mid A = 1), MD DP ( f ) = P ( Y = 1 ∣ A = − 1 ) − P ( Y = 1 ∣ A = 1 ) , MD E O ( f ) = P ( Y ^ = 1 ∣ Y = 1 , A = − 1 ) − P ( Y ^ = 1 ∣ Y = 1 , A = 1 ) . \operatorname{MD}_{\mathrm{EO}}(f) = \mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = -1) - \mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = 1). MD EO ( f ) = P ( Y = 1 ∣ Y = 1 , A = − 1 ) − P ( Y = 1 ∣ Y = 1 , A = 1 ) .
Definition 8 (MR [9 ] ). For a classifier f f f ,
MR D P ( f ) = P ( Y ^ = 1 ∣ A = − 1 ) P ( Y ^ = 1 ∣ A = 1 ) , \operatorname{MR}_{\mathrm{DP}}(f) = \frac{\mathbb{P}(\widehat{Y} = 1 \mid A = -1)}{\mathbb{P}(\widehat{Y} = 1 \mid A = 1)}, MR DP ( f ) = P ( Y = 1 ∣ A = 1 ) P ( Y = 1 ∣ A = − 1 ) , MR E O ( f ) = P ( Y ^ = 1 ∣ Y = 1 , A = − 1 ) P ( Y ^ = 1 ∣ Y = 1 , A = 1 ) . \operatorname{MR}_{\mathrm{EO}}(f) = \frac{\mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = -1)}{\mathbb{P}(\widehat{Y} = 1 \mid Y = 1, A = 1)}. MR EO ( f ) = P ( Y = 1 ∣ Y = 1 , A = 1 ) P ( Y = 1 ∣ Y = 1 , A = − 1 ) .
The key observation of Menon & Williamson [17 ] is that approximate fairness measures can be written as a regularization term. Let D ˉ D P \bar{D}_{\mathrm{DP}} D ˉ DP be the joint law of ( X , A ) (X, A) ( X , A ) and D ˉ E O \bar{D}_{\mathrm{EO}} D ˉ EO the conditional law of ( X , A ) (X, A) ( X , A ) given Y = 1 Y = 1 Y = 1 . For α ˉ ∈ ( 0 , 1 ) \bar{\alpha} \in(0, 1) α ˉ ∈ ( 0 , 1 ) , define
R α ˉ ( f ; D ˉ ) ≔ ( 1 − α ˉ ) P ( Y ^ = − 1 , A = 1 ) + α ˉ P ( Y ^ = 1 , A = − 1 ) , (5) \mathcal{R}_{\bar{\alpha}}(f; \bar{D}) \coloneqq(1 - \bar{\alpha}) \mathbb{P}(\widehat{Y} = -1, A = 1) + \bar{\alpha} \mathbb{P}(\widehat{Y} = 1, A = -1), \tag*{(5)} R α ˉ ( f ; D ˉ ) : = ( 1 − α ˉ ) P ( Y = − 1 , A = 1 ) + α ˉ P ( Y = 1 , A = − 1 ) , (5)
where D ˉ \bar{D} D ˉ is either D ˉ D P \bar{D}_{\mathrm{DP}} D ˉ DP or D ˉ E O \bar{D}_{\mathrm{EO}} D ˉ EO .
The choice of α ˉ \bar{\alpha} α ˉ in Eq. (5) is dictated by whether MD or MR is used; Appendix § 8.1 records the four standard cases. Motivated by this equivalence, we treat λ \lambda λ as a trade-off parameter λ ≥ 0 \lambda\ge0 λ ≥ 0 chosen by the decision maker and study the fairness-regularized risk
R f a i r α , α ˉ ( f ; D , D ˉ ) ≔ R α ( f ; D ) − λ ⋅ R α ˉ ( f ; D ˉ ) = E ( X , Y ) ∼ D [ U α ( Y , f ( X ) ) ] − λ E ( X , A ) ∼ D ˉ [ U α ˉ ( A , f ( X ) ) ] , (6) \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f;D,\bar{D}) \coloneqq\mathcal{R}_{\alpha}(f;D)-\lambda\cdot\mathcal{R}_{\bar{\alpha}}(f;\bar{D})
= \mathbb{E}_{(X,Y)\sim D}[U_{\alpha}(Y,f(X))]-\lambda\mathbb{E}_{(X,A)\sim\bar{D}}[U_{\bar{\alpha}}(A,f(X))], \tag*{(6)} R fair α , α ˉ ( f ; D , D ˉ ) : = R α ( f ; D ) − λ ⋅ R α ˉ ( f ; D ˉ ) = E ( X , Y ) ∼ D [ U α ( Y , f ( X ))] − λ E ( X , A ) ∼ D ˉ [ U α ˉ ( A , f ( X ))] , (6)
where R α ˉ \mathcal{R}_{\bar{\alpha}} R α ˉ is the auxiliary risk in Eq. (5) and U α ˉ U_{\bar{\alpha}} U α ˉ is defined as in Eq. (3) with α \alpha α replaced by α ˉ \bar{\alpha} α ˉ .
The regularized form in Eq. (6) and the Bayes-optimal classifier below are due to Menon & Williamson [17 ] , Zeng et al. [31 ] , and Yang et al. [28 ] , who derived them through the equivalence recorded in Lemma 21 (Appendix § 8.1 ). We adopt this formulation as our starting point.
To state the classifier, recall that η ( x ) = P ( Y = 1 ∣ X = x ) \eta(x)=\mathbb{P}(Y=1\mid X=x) η ( x ) = P ( Y = 1 ∣ X = x ) ; let η ˉ D P ( x ) ≔ P ( A = 1 ∣ X = x ) \bar{\eta}_{\mathrm{DP}}(x)\coloneqq\mathbb{P}(A=1\mid X=x) η ˉ DP ( x ) : = P ( A = 1 ∣ X = x ) , η ˉ E O ( x ) ≔ P ( A = 1 ∣ X = x , Y = 1 ) \bar{\eta}_{\mathrm{EO}}(x)\coloneqq\mathbb{P}(A=1\mid X=x,Y=1) η ˉ EO ( x ) : = P ( A = 1 ∣ X = x , Y = 1 ) , π A ≔ P ( A = 1 ) \pi_A\coloneqq\mathbb{P}(A=1) π A : = P ( A = 1 ) , π A ∣ Y = 1 ≔ P ( A = 1 ∣ Y = 1 ) \pi_{A\mid Y=1}\coloneqq\mathbb{P}(A=1\mid Y=1) π A ∣ Y = 1 : = P ( A = 1 ∣ Y = 1 ) , and π ≔ P ( Y = 1 ) \pi\coloneqq\mathbb{P}(Y=1) π : = P ( Y = 1 ) . In the attribute-aware setting, let η ( x , a ) ≔ P ( Y = 1 ∣ X = x , A = a ) \eta(x,a)\coloneqq\mathbb{P}(Y=1\mid X=x,A=a) η ( x , a ) : = P ( Y = 1 ∣ X = x , A = a ) .
Theorem 9 (Bayes-Optimal Fairness-Aware Classifiers). For both D P DP D P and E O EO E O , the Bayes-optimal classifier is Y ^ = sign ( f ∗ ( x ) ) \hat{Y}=\operatorname{sign}(f^*(x)) Y ^ = sign ( f ∗ ( x )) in the attribute-blind setting and Y ^ = sign ( f ∗ ( x , a ) ) \hat{Y}=\operatorname{sign}(f^*(x,a)) Y ^ = sign ( f ∗ ( x , a )) in the attribute-aware setting, where a minimizer f ∗ ∈ arg min f ∈ F R f a i r α , α ˉ ( f ; D , D ˉ ) f^*\in\arg\min_{f\in\mathcal{F}}\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f;D,\bar{D}) f ∗ ∈ arg min f ∈ F R fair α , α ˉ ( f ; D , D ˉ ) is given by the following forms:
D P ( attribute-blind ) : f D P ∗ ( x ) = η ( x ) − α − λ ( η ˉ D P ( x ) − α ˉ ) , D P ( attribute-aware ) : f D P ∗ ( x , a ) = η ( x , a ) − α − λ ( a + 1 2 − α ˉ ) , E O ( attribute-blind ) : f E O ∗ ( x ) = η ( x ) − α − λ π − 1 η ( x ) ( η ˉ E O ( x ) − α ˉ ) , E O ( attribute-aware ) : f E O ∗ ( x , a ) = η ( x , a ) − α − λ π − 1 η ( x , a ) ( a + 1 2 − α ˉ ) . \begin{aligned}
DP\ (\text{attribute-blind}):\quad&f^*_{DP}(x)=\eta(x)-\alpha-\lambda\left(\bar{\eta}_{DP}(x)-\bar{\alpha}\right),\\
DP\ (\text{attribute-aware}):\quad&f^*_{DP}(x,a)=\eta(x,a)-\alpha-\lambda\left(\frac{a+1}{2}-\bar{\alpha}\right),\\
EO\ (\text{attribute-blind}):\quad&f^*_{EO}(x)=\eta(x)-\alpha-\lambda\pi^{-1}\eta(x)\left(\bar{\eta}_{EO}(x)-\bar{\alpha}\right),\\
EO\ (\text{attribute-aware}):\quad&f^*_{EO}(x,a)=\eta(x,a)-\alpha-\lambda\pi^{-1}\eta(x,a)\left(\frac{a+1}{2}-\bar{\alpha}\right).
\end{aligned} D P ( attribute-blind ) : D P ( attribute-aware ) : E O ( attribute-blind ) : E O ( attribute-aware ) : f D P ∗ ( x ) = η ( x ) − α − λ ( η ˉ D P ( x ) − α ˉ ) , f D P ∗ ( x , a ) = η ( x , a ) − α − λ ( 2 a + 1 − α ˉ ) , f E O ∗ ( x ) = η ( x ) − α − λ π − 1 η ( x ) ( η ˉ E O ( x ) − α ˉ ) , f E O ∗ ( x , a ) = η ( x , a ) − α − λ π − 1 η ( x , a ) ( 2 a + 1 − α ˉ ) .
Theorem 9 shows that the fairness-aware Bayes classifier retains the cost-sensitive discriminant η − α \eta-\alpha η − α of Eq. (2) and subtracts an instance-dependent adjustment. For DP, the adjustment is λ ( η ˉ D P − α ˉ ) \lambda(\bar{\eta}_{\mathrm{DP}}-\bar{\alpha}) λ ( η ˉ DP − α ˉ ) , the deviation of the sensitive-attribute posterior from its base rate; in the attribute-aware case the posterior is replaced by ( a + 1 ) / 2 (a+1)/2 ( a + 1 ) /2 . For EO, the corresponding deviation is weighted by π − 1 η \pi^{-1}\eta π − 1 η , with η \eta η evaluated at x x x or ( x , a ) (x,a) ( x , a ) according to the setting, reflecting that the fairness term is evaluated only on the positive class. Setting λ = 0 \lambda=0 λ = 0 recovers the cost-sensitive classifier in Eq. (2) .
Surrogate Losses for Fairness-Aware Classification
Although Eq. (6) admits the closed-form Bayes-optimal classifier of Theorem 9 , the fairness-regularized risk is built on the discontinuous, non-convex 0–1 loss and is therefore intractable to minimize directly from finite data. Following the standard paradigm of § 2.1 , practical fairness-aware methods replace the two cost-sensitive 0–1 losses by surrogate losses. Let L : { − 1 , 1 } × R → [ 0 , ∞ ) L:\{-1,1\}\times\mathbb{R}\to[0,\infty) L : { − 1 , 1 } × R → [ 0 , ∞ ) be a surrogate loss for the accuracy term (with partial losses L 1 , L − 1 L_1,L_{-1} L 1 , L − 1 as in Eq. (4) ), and let L ˉ : { − 1 , 1 } × R → R \bar{L}:\{-1,1\}\times\mathbb{R}\to\mathbb{R} L ˉ : { − 1 , 1 } × R → R be a surrogate loss for the fairness term (with partial losses L ˉ 1 , L ˉ − 1 \bar{L}_1,\bar{L}_{-1} L ˉ 1 , L ˉ − 1 ). We call ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) a fairness-aware surrogate loss pair , and define the composite fairness-aware surrogate loss
L f a i r ( y , a , t ) : = L ( y , t ) − λ L ˉ ( a , t ) , L_{\mathrm{fair}}(y,a,t):=L(y,t)-\lambda\bar{L}(a,t), L fair ( y , a , t ) := L ( y , t ) − λ L ˉ ( a , t ) ,
and the corresponding ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk
R f a i r L , L ˉ ( f ; D , D ˉ ) : = R L ( f ; D ) − λ ⋅ R L ˉ ( f ; D ˉ ) = E [ L f a i r ( Y , A , f ( X ) ) ] . = E ( X , Y ) ∼ D [ L ( Y , f ( X ) ) ] − λ E ( X , A ) ∼ D ˉ [ L ˉ ( A , f ( X ) ) ] . \begin{aligned}
\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f;D,\bar{D}):=\mathcal{R}_{L}(f;D)-\lambda\cdot\mathcal{R}_{\bar{L}}(f;\bar{D})=\mathbb{E}[L_{\mathrm{fair}}(Y,A,f(X))]. \\
&=\mathbb{E}_{(X,Y)\sim D}[L(Y,f(X))]-\lambda\mathbb{E}_{(X,A)\sim\bar{D}}[\bar{L}(A,f(X))].
\end{aligned} R fair L , L ˉ ( f ; D , D ˉ ) := R L ( f ; D ) − λ ⋅ R L ˉ ( f ; D ˉ ) = E [ L fair ( Y , A , f ( X ))] . = E ( X , Y ) ∼ D [ L ( Y , f ( X ))] − λ E ( X , A ) ∼ D ˉ [ L ˉ ( A , f ( X ))] .
We denote the minimal values of the two risks by R f a i r α , α ˉ , ∗ : = inf f ∈ F R f a i r α , α ˉ ( f ; D , D ˉ ) \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}:=\inf_{f\in\mathcal{F}}\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f;D,\bar{D}) R fair α , α ˉ , ∗ := inf f ∈ F R fair α , α ˉ ( f ; D , D ˉ ) and
R f a i r L , L ˉ , ∗ : = inf f ∈ F R f a i r L , L ˉ ( f ; D , D ˉ ) \mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}:=\inf_{f\in\mathcal{F}}\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f;D,\bar{D}) R fair L , L ˉ , ∗ := inf f ∈ F R fair L , L ˉ ( f ; D , D ˉ ) .
When the surrogate losses L L L and L ˉ \bar{L} L ˉ are convex and smooth, the ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk can be minimized efficiently using standard gradient-based methods, which is why surrogate-based methods are the dominant approach in fairness-aware classification practice. However, the classical guarantees of § 2.1 do not transfer directly: the fairness-aware objective is a difference of two cost-sensitive risks, so neither the Bayes-optimal classifier nor the calibration function of a single surrogate loss suffices to control the fairness-aware excess risk. This raises the central question of this paper: can we establish Fisher consistency and classification calibration in the fairness-aware classification in Eq. (6) ? We answer this question in the following sections.
Fisher Consistency for Fairness-Aware Classification
This section adapts the concept of Fisher consistency to surrogate losses in fairness-aware classification defined in § 2.4 . Specifically, we begin with a pointwise analysis of the fairness-aware conditional risk, then define FAFC, and finally establish a necessary and sufficient condition for this property.
Fairness-Aware Conditional Risk
To analyze the properties of the ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk, we derive fairness-aware conditional risk results for the attribute-blind and attribute-aware settings individually, and unify both cases into a generalized framework. For clarity, we develop the attribute-blind setting in detail first and then show that the attribute-aware setting arises as a special case.
The Attribute-Blind Setting
In the attribute-blind setting, the classifier f f f makes decision solely based on the features X X X . For a fixed instance x ∈ X x \in\mathcal{X} x ∈ X , we will analyze the two fairness notions (i.e., DP and EO) discussed in § 2.3 .
For DP, the fairness penalty is evaluated over the marginal distribution of X X X and A A A . The fairness-aware conditional risk is:
E ( X , Y ) ∼ D ∣ X = x [ U α ( Y , f ( X ) ) ] − λ E ( X , A ) ∼ D ‾ D P ∣ X = x [ U ‾ α ‾ ( A , f ( X ) ) ] = [ ( 1 − α ) η ( x ) − λ ( 1 − α ‾ ) η ‾ D P ( x ) ] 1 { f ( x ) ≤ 0 } + [ α ( 1 − η ( x ) ) − λ α ‾ ( 1 − η ‾ D P ( x ) ) ] 1 { f ( x ) > 0 } . \begin{aligned}
\mathbb{E}_{(X,Y)\sim D\mid X=x}\left[U_{\alpha}(Y,f(X))\right]-\lambda\mathbb{E}_{(X,A)\sim\overline{D}_{\mathrm{DP}}\mid X=x}\left[\overline{U}_{\overline{\alpha}}(A,f(X))\right] \\
&=\left[(1-\alpha)\eta(x)-\lambda(1-\overline{\alpha})\overline{\eta}_{\mathrm{DP}}(x)\right]\mathbf{1}\{f(x)\leq0\}
+\left[\alpha(1-\eta(x))-\lambda\overline{\alpha}(1-\overline{\eta}_{\mathrm{DP}}(x))\right]\mathbf{1}\{f(x)>0\}.
\end{aligned} E ( X , Y ) ∼ D ∣ X = x [ U α ( Y , f ( X )) ] − λ E ( X , A ) ∼ D DP ∣ X = x [ U α ( A , f ( X )) ] = [ ( 1 − α ) η ( x ) − λ ( 1 − α ) η DP ( x ) ] 1 { f ( x ) ≤ 0 } + [ α ( 1 − η ( x )) − λ α ( 1 − η DP ( x )) ] 1 { f ( x ) > 0 } .
For EO, the fairness penalty is conditioned on the positive class Y = 1 Y=1 Y = 1 . The fairness-aware conditional risk is:
E ( X , Y ) ∼ D ∣ X = x [ U α ( Y , f ( X ) ) ] − λ E ( X , A ) ∼ D ‾ E O ∣ X = x [ U ‾ α ‾ ( A , f ( X ) ) ] = [ ( 1 − α ) η ( x ) − λ π − 1 η ( x ) ( 1 − α ‾ ) η ‾ E O ( x ) ] 1 { f ( x ) ≤ 0 } + [ α ( 1 − η ( x ) ) − λ π − 1 η ( x ) α ‾ ( 1 − η ‾ E O ( x ) ) ] 1 { f ( x ) > 0 } . \begin{aligned}
\mathbb{E}_{(X,Y)\sim D\mid X=x}\left[U_{\alpha}(Y,f(X))\right]-\lambda\mathbb{E}_{(X,A)\sim\overline{D}_{\mathrm{EO}}\mid X=x}\left[\overline{U}_{\overline{\alpha}}(A,f(X))\right] \\
&=\left[(1-\alpha)\eta(x)-\lambda\pi^{-1}\eta(x)(1-\overline{\alpha})\overline{\eta}_{\mathrm{EO}}(x)\right]\mathbf{1}\{f(x)\leq0\}
+\left[\alpha(1-\eta(x))-\lambda\pi^{-1}\eta(x)\overline{\alpha}(1-\overline{\eta}_{\mathrm{EO}}(x))\right]\mathbf{1}\{f(x)>0\}.
\end{aligned} E ( X , Y ) ∼ D ∣ X = x [ U α ( Y , f ( X )) ] − λ E ( X , A ) ∼ D EO ∣ X = x [ U α ( A , f ( X )) ] = [ ( 1 − α ) η ( x ) − λ π − 1 η ( x ) ( 1 − α ) η EO ( x ) ] 1 { f ( x ) ≤ 0 } + [ α ( 1 − η ( x )) − λ π − 1 η ( x ) α ( 1 − η EO ( x )) ] 1 { f ( x ) > 0 } .
We unify the above cases by defining a general fairness-aware conditional risk. Define that
C f a i r α , α ‾ ( η , η ‾ , t ) = A α , α ‾ ( η , η ‾ ) 1 { t ≤ 0 } + B α , α ‾ ( η , η ‾ ) 1 { t > 0 } , (7) C_{\mathrm{fair}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta},t)=A^{\alpha,\overline{\alpha}}(\eta,\overline{\eta})\mathbf{1}\{t\leq0\}+B^{\alpha,\overline{\alpha}}(\eta,\overline{\eta})\mathbf{1}\{t>0\}, \tag*{(7)} C fair α , α ( η , η , t ) = A α , α ( η , η ) 1 { t ≤ 0 } + B α , α ( η , η ) 1 { t > 0 } , (7)
where η ‾ \overline{\eta} η denotes η ‾ D P \overline{\eta}_{\mathrm{DP}} η DP or η ‾ E O \overline{\eta}_{\mathrm{EO}} η EO depending on the context, and A α , α ‾ ( η , η ‾ ) A^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}) A α , α ( η , η ) and B α , α ‾ ( η , η ‾ ) B^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}) B α , α ( η , η ) are cost functions derived from the specific fairness notion. Since A α , α ‾ ( η , η ‾ ) A^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}) A α , α ( η , η ) and B α , α ‾ ( η , η ‾ ) B^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}) B α , α ( η , η ) depend on η \eta η and η ‾ \overline{\eta} η , which is associated with every datapoint, we call them example-dependent fair cost functions .
For DP, the example-dependent cost functions are
A D P α , α ‾ ( η , η ‾ D P ) = η ( 1 − α ) − λ η ‾ D P ( 1 − α ‾ ) , A_{\mathrm{DP}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}_{\mathrm{DP}})=\eta(1-\alpha)-\lambda\overline{\eta}_{\mathrm{DP}}(1-\overline{\alpha}), A DP α , α ( η , η DP ) = η ( 1 − α ) − λ η DP ( 1 − α ) ,
B D P α , α ‾ ( η , η ‾ D P ) = ( 1 − η ) α − λ ( 1 − η ‾ D P ) α ‾ B_{\mathrm{DP}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}_{\mathrm{DP}})=(1-\eta)\alpha-\lambda(1-\overline{\eta}_{\mathrm{DP}})\overline{\alpha} B DP α , α ( η , η DP ) = ( 1 − η ) α − λ ( 1 − η DP ) α
and the conditional risk is defined as
C D P α , α ‾ ( η , η ‾ D P , t ) : = A D P α , α ‾ ( η , η ‾ D P ) 1 { t ≤ 0 } + B D P α , α ‾ ( η , η ‾ D P ) 1 { t > 0 } . C_{\mathrm{DP}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}_{\mathrm{DP}},t):=A_{\mathrm{DP}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}_{\mathrm{DP}})\mathbf{1}\{t\leq0\}+B_{\mathrm{DP}}^{\alpha,\overline{\alpha}}(\eta,\overline{\eta}_{\mathrm{DP}})\mathbf{1}\{t>0\}. C DP α , α ( η , η DP , t ) := A DP α , α ( η , η DP ) 1 { t ≤ 0 } + B DP α , α ( η , η DP ) 1 { t > 0 } .
For EO, the example-dependent cost functions are
A E O α , α ˉ ( η , η ˉ E O ) = η ( 1 − α ) − λ π − 1 η η ˉ E O ( 1 − α ˉ ) , A_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}_{\mathrm{EO}})=\eta(1-\alpha)-\lambda\pi^{-1}\eta\bar{\eta}_{\mathrm{EO}}(1-\bar{\alpha}), A EO α , α ˉ ( η , η ˉ EO ) = η ( 1 − α ) − λ π − 1 η η ˉ EO ( 1 − α ˉ ) ,
B E O α , α ˉ ( η , η ˉ E O ) = ( 1 − η ) α − λ π − 1 η ( 1 − η ˉ E O ) α ˉ B_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}_{\mathrm{EO}})=(1-\eta)\alpha-\lambda\pi^{-1}\eta(1-\bar{\eta}_{\mathrm{EO}})\bar{\alpha} B EO α , α ˉ ( η , η ˉ EO ) = ( 1 − η ) α − λ π − 1 η ( 1 − η ˉ EO ) α ˉ
and the conditional risk is defined as
C E O α , α ˉ ( η , η ˉ E O , t ) ≔ A E O α , α ˉ ( η , η ˉ E O ) 1 { t ≤ 0 } + B E O α , α ˉ ( η , η ˉ E O ) 1 { t > 0 } . C_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}_{\mathrm{EO}},t)\coloneqq A_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}_{\mathrm{EO}})1\{t\le0\}+B_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}_{\mathrm{EO}})1\{t>0\}. C EO α , α ˉ ( η , η ˉ EO , t ) : = A EO α , α ˉ ( η , η ˉ EO ) 1 { t ≤ 0 } + B EO α , α ˉ ( η , η ˉ EO ) 1 { t > 0 } .
Both the risk of DP and EO satisfies
R f a i r α , α ˉ ( f ) = E X [ C f a i r α , α ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) ] . \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f)=\mathbb{E}_{X}\left[C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X),f(X))\right]. R fair α , α ˉ ( f ) = E X [ C fair α , α ˉ ( η ( X ) , η ˉ ( X ) , f ( X )) ] .
We define C f a i r α , α ˉ , ∗ ( η , η ˉ ) ≔ inf t ∈ R C f a i r α , α ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta})\coloneqq\inf_{t\in\mathbb{R}}C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t) C fair α , α ˉ , ∗ ( η , η ˉ ) : = inf t ∈ R C fair α , α ˉ ( η , η ˉ , t ) to denote the optimal conditional risk. Consequently, R f a i r α , α ˉ , ∗ = E X [ C f a i r α , α ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}=\mathbb{E}_{X}\left[C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta(X),\bar{\eta}(X))\right] R fair α , α ˉ , ∗ = E X [ C fair α , α ˉ , ∗ ( η ( X ) , η ˉ ( X )) ] .
The same pointwise analysis applies to any surrogate loss pair. Recall that
R L ( f ; D ) = E ( X , Y ) ∼ D [ L ( Y , f ( X ) ) ] , R L ˉ ( f ; D ˉ ) = E ( X , A ) ∼ D ˉ [ L ˉ ( A , f ( X ) ) ] , \mathcal{R}_{L}(f;D)=\mathbb{E}_{(X,Y)\sim D}[L(Y,f(X))],\qquad\mathcal{R}_{\bar{L}}(f;\bar{D})=\mathbb{E}_{(X,A)\sim\bar{D}}[\bar{L}(A,f(X))], R L ( f ; D ) = E ( X , Y ) ∼ D [ L ( Y , f ( X ))] , R L ˉ ( f ; D ˉ ) = E ( X , A ) ∼ D ˉ [ L ˉ ( A , f ( X ))] ,
where L ( y , t ) = 1 { y = 1 } L 1 ( t ) + 1 { y = − 1 } L − 1 ( t ) L(y,t)=1\{y=1\}L_{1}(t)+1\{y=-1\}L_{-1}(t) L ( y , t ) = 1 { y = 1 } L 1 ( t ) + 1 { y = − 1 } L − 1 ( t ) and L ˉ ( a , t ) = 1 { a = 1 } L ˉ 1 ( t ) + 1 { a = − 1 } L ˉ − 1 ( t ) \bar{L}(a,t)=1\{a=1\}\bar{L}_{1}(t)+1\{a=-1\}\bar{L}_{-1}(t) L ˉ ( a , t ) = 1 { a = 1 } L ˉ 1 ( t ) + 1 { a = − 1 } L ˉ − 1 ( t ) by the partial-loss decomposition of Eq. (4) .
Conditioning on X = x X=x X = x , the accuracy term contributes η ( x ) L 1 ( f ( x ) ) + ( 1 − η ( x ) ) L − 1 ( f ( x ) ) \eta(x)L_{1}(f(x))+(1-\eta(x))L_{-1}(f(x)) η ( x ) L 1 ( f ( x )) + ( 1 − η ( x )) L − 1 ( f ( x )) .
For DP, the fairness term contributes η ˉ D P ( x ) L ˉ 1 ( f ( x ) ) + ( 1 − η ˉ D P ( x ) ) L ˉ − 1 ( f ( x ) ) \bar{\eta}_{\mathrm{DP}}(x)\bar{L}_{1}(f(x))+(1-\bar{\eta}_{\mathrm{DP}}(x))\bar{L}_{-1}(f(x)) η ˉ DP ( x ) L ˉ 1 ( f ( x )) + ( 1 − η ˉ DP ( x )) L ˉ − 1 ( f ( x )) , and hence we define
C D P L , L ˉ ( η , η ˉ D P , t ) ≔ [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) ] − λ [ η ˉ D P L ˉ 1 ( t ) + ( 1 − η ˉ D P ) L ˉ − 1 ( t ) ] . (8) C_{\mathrm{DP}}^{L,\bar{L}}(\eta,\bar{\eta}_{\mathrm{DP}},t)\coloneqq[\eta L_{1}(t)+(1-\eta)L_{-1}(t)]-\lambda[\bar{\eta}_{\mathrm{DP}}\bar{L}_{1}(t)+(1-\bar{\eta}_{\mathrm{DP}})\bar{L}_{-1}(t)]. \tag*{(8)} C DP L , L ˉ ( η , η ˉ DP , t ) : = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t )] − λ [ η ˉ DP L ˉ 1 ( t ) + ( 1 − η ˉ DP ) L ˉ − 1 ( t )] . (8)
For EO, the fairness term contributes π − 1 η ( x ) [ η ˉ E O ( x ) L ˉ 1 ( f ( x ) ) + ( 1 − η ˉ E O ( x ) ) L ˉ − 1 ( f ( x ) ) ] \pi^{-1}\eta(x)[\bar{\eta}_{\mathrm{EO}}(x)\bar{L}_{1}(f(x))+(1-\bar{\eta}_{\mathrm{EO}}(x))\bar{L}_{-1}(f(x))] π − 1 η ( x ) [ η ˉ EO ( x ) L ˉ 1 ( f ( x )) + ( 1 − η ˉ EO ( x )) L ˉ − 1 ( f ( x ))] , and hence we define
C E O L , L ˉ ( η , η ˉ E O , t ) ≔ [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) ] − λ π − 1 η [ η ˉ E O L ˉ 1 ( t ) + ( 1 − η ˉ E O ) L ˉ − 1 ( t ) ] . (9) C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta}_{\mathrm{EO}},t)\coloneqq[\eta L_{1}(t)+(1-\eta)L_{-1}(t)]-\lambda\pi^{-1}\eta[\bar{\eta}_{\mathrm{EO}}\bar{L}_{1}(t)+(1-\bar{\eta}_{\mathrm{EO}})\bar{L}_{-1}(t)]. \tag*{(9)} C EO L , L ˉ ( η , η ˉ EO , t ) : = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t )] − λ π − 1 η [ η ˉ EO L ˉ 1 ( t ) + ( 1 − η ˉ EO ) L ˉ − 1 ( t )] . (9)
Recall that η ˉ \bar{\eta} η ˉ denote η ˉ D P \bar{\eta}_{\mathrm{DP}} η ˉ DP or η ˉ E O \bar{\eta}_{\mathrm{EO}} η ˉ EO according to the fairness notion. By the tower property,
R f a i r L , L ˉ ( f ) = E X [ C f a i r L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) ] . \mathcal{R}^{L,\bar{L}}_{\mathrm{fair}}(f)=\mathbb{E}_{X}\left[C^{L,\bar{L}}_{\mathrm{fair}}\left(\eta(X),\bar{\eta}(X),f(X)\right)\right]. R fair L , L ˉ ( f ) = E X [ C fair L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) ] .
Similarly, we define C f a i r L , L ˉ , ∗ ( η , η ˉ ) ≔ inf t ∈ R C f a i r L , L ˉ ( η , η ˉ , t ) C^{L,\bar{L},*}_{\mathrm{fair}}(\eta,\bar{\eta}) \coloneqq\inf_{t\in\mathbb{R}} C^{L,\bar{L}}_{\mathrm{fair}}(\eta,\bar{\eta},t) C fair L , L ˉ , ∗ ( η , η ˉ ) : = inf t ∈ R C fair L , L ˉ ( η , η ˉ , t ) to denote the optimal conditional ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk. Since F \mathcal{F} F denote the set of all measurable functions f : X → R f:\mathcal{X}\to\mathbb{R} f : X → R , consequently,
R f a i r L , L ˉ , ∗ = E X [ C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] . \mathcal{R}^{L,\bar{L},*}_{\mathrm{fair}}=\mathbb{E}_{X}\left[C^{L,\bar{L},*}_{\mathrm{fair}}\left(\eta(X),\bar{\eta}(X)\right)\right]. R fair L , L ˉ , ∗ = E X [ C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] .
The Attribute-Aware Setting
In the attribute-aware setting, the classifier f f f utilizes both the feature X X X and the sensitive attribute A A A . For a fixed instance ( X = x , A = a ) (X=x,A=a) ( X = x , A = a ) , the posterior probability of the positive class becomes η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) \eta(x,a)=\mathbb{P}(Y=1\mid X=x,A=a) η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) . Since the true sensitive attribute is given, the inferred probabilities collapse to indicator functions: η ˉ D P ( x , a ) = P ( A = 1 ∣ X = x , A = a ) = 1 { a = 1 } \bar{\eta}_{\mathrm{DP}}(x,a)=\mathbb{P}(A=1\mid X=x,A=a)=1\{a=1\} η ˉ DP ( x , a ) = P ( A = 1 ∣ X = x , A = a ) = 1 { a = 1 } and η ˉ E O ( x , a ) = 1 { a = 1 } \bar{\eta}_{\mathrm{EO}}(x,a)=1\{a=1\} η ˉ EO ( x , a ) = 1 { a = 1 } .
Consequently, the conditional risk no longer depends on the posterior probability η ˉ ∈ [ 0 , 1 ] \bar{\eta}\in[0,1] η ˉ ∈ [ 0 , 1 ] , but directly on the binary sensitive attribute a ∈ { − 1 , 1 } a\in\{-1,1\} a ∈ { − 1 , 1 } . We define the attribute-aware conditional risk as
C f a i r α , α ˉ ( η , a , t ) = A α , α ˉ ( η , a ) 1 { t ≤ 0 } + B α , α ˉ ( η , a ) 1 { t > 0 } . (10) C^{\alpha,\bar{\alpha}}_{\mathrm{fair}}(\eta,a,t)=A^{\alpha,\bar{\alpha}}(\eta,a)1\{t\leq0\}+B^{\alpha,\bar{\alpha}}(\eta,a)1\{t>0\}. \tag*{(10)} C fair α , α ˉ ( η , a , t ) = A α , α ˉ ( η , a ) 1 { t ≤ 0 } + B α , α ˉ ( η , a ) 1 { t > 0 } . (10)
For DP, the cost functions A A A and B B B are simplified to
A D P α , α ˉ ( η , a ) = η ( 1 − α ) − λ ( 1 − α ˉ ) 1 { a = 1 } , (11) A^{\alpha,\bar{\alpha}}_{\mathrm{DP}}(\eta,a)=\eta(1-\alpha)-\lambda(1-\bar{\alpha})1\{a=1\}, \tag*{(11)} A DP α , α ˉ ( η , a ) = η ( 1 − α ) − λ ( 1 − α ˉ ) 1 { a = 1 } , (11)
B D P α , α ˉ ( η , a ) = ( 1 − η ) α − λ α ˉ ( 1 − 1 { a = 1 } ) . (12) B^{\alpha,\bar{\alpha}}_{\mathrm{DP}}(\eta,a)=(1-\eta)\alpha-\lambda\bar{\alpha}(1-1\{a=1\}). \tag*{(12)} B DP α , α ˉ ( η , a ) = ( 1 − η ) α − λ α ˉ ( 1 − 1 { a = 1 }) . (12)
Similarly, for EO, the cost functions are
A E O α , α ˉ ( η , a ) = η ( 1 − α ) − λ π − 1 η ( 1 − α ˉ ) 1 { a = 1 } , A^{\alpha,\bar{\alpha}}_{\mathrm{EO}}(\eta,a)=\eta(1-\alpha)-\lambda\pi^{-1}\eta(1-\bar{\alpha})1\{a=1\}, A EO α , α ˉ ( η , a ) = η ( 1 − α ) − λ π − 1 η ( 1 − α ˉ ) 1 { a = 1 } ,
B E O α , α ˉ ( η , a ) = ( 1 − η ) α − λ π − 1 η α ˉ ( 1 − 1 { a = 1 } ) . B^{\alpha,\bar{\alpha}}_{\mathrm{EO}}(\eta,a)=(1-\eta)\alpha-\lambda\pi^{-1}\eta\bar{\alpha}(1-1\{a=1\}). B EO α , α ˉ ( η , a ) = ( 1 − η ) α − λ π − 1 η α ˉ ( 1 − 1 { a = 1 }) .
Given an arbitrary surrogate loss pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) , the attribute-aware conditional ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk is given by
C D P L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) ] − λ [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 } ) L ˉ − 1 ( t ) ] . C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t) = [\eta L_1(t) + (1-\eta)L_{-1}(t)] - \lambda[1\{a=1\}\bar{L}_1(t) + (1-1\{a=1\})\bar{L}_{-1}(t)]. C DP L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t )] − λ [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 }) L ˉ − 1 ( t )] .
C E O L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) ] − λ π − 1 η [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 } ) L ˉ − 1 ( t ) ] . C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t) = [\eta L_1(t) + (1-\eta)L_{-1}(t)] - \lambda\pi^{-1}\eta[1\{a=1\}\bar{L}_1(t) + (1-1\{a=1\})\bar{L}_{-1}(t)]. C EO L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t )] − λ π − 1 η [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 }) L ˉ − 1 ( t )] .
Remark 10. Compared with (7) and (10) , the attribute-aware conditional risk is mathematically a special case of the attribute-blind conditional risk, where the continuous variable η ˉ ∈ [ 0 , 1 ] \bar{\eta} \in[0,1] η ˉ ∈ [ 0 , 1 ] is restricted to the discrete set { 0 , 1 } \{0,1\} { 0 , 1 } .
Fairness-Aware Fisher Consistency
We now use the pointwise conditional risk to define FAFC. For any surrogate loss L f a i r ( y , a , t ) = L ( y , t ) − λ L ˉ ( a , t ) L_{\mathrm{fair}}(y,a,t) = L(y,t) - \lambda\bar{L}(a,t) L fair ( y , a , t ) = L ( y , t ) − λ L ˉ ( a , t ) , define the fairness-aware discriminant as
D α , α ˉ ( η , η ˉ ) : = A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) , (13) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) := A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) - B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}), \tag*{(13)} D α , α ˉ ( η , η ˉ ) := A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) , (13)
and the excess conditional risk as
H L , L ˉ α , α ˉ ( η , η ˉ ) : = C f a i r L , L ˉ , − ( η , η ˉ ) − C f a i r L , L ˉ , ∗ ( η , η ˉ ) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) := C_{\mathrm{fair}}^{L,\bar{L},-}(\eta,\bar{\eta}) - C_{\mathrm{fair}}^{L,\bar{L},*}(\eta,\bar{\eta}) H L , L ˉ α , α ˉ ( η , η ˉ ) := C fair L , L ˉ , − ( η , η ˉ ) − C fair L , L ˉ , ∗ ( η , η ˉ )
for η , η ˉ ∈ [ 0 , 1 ] \eta,\bar{\eta} \in[0,1] η , η ˉ ∈ [ 0 , 1 ] , where C f a i r L , L ˉ , − ( η , η ˉ ) : = inf { t : t ⋅ D α , α ˉ ( η , η ˉ ) ≤ 0 } C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L},-}(\eta,\bar{\eta}) := \inf_{\{t:t\cdot D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\leq0\}} C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ , − ( η , η ˉ ) := inf { t : t ⋅ D α , α ˉ ( η , η ˉ ) ≤ 0 } C fair L , L ˉ ( η , η ˉ , t ) .
Before defining FAFC, we fix the domain on which the pointwise condition is imposed and clarify the role of the trade-off parameter.
Remark 11 (Well-posedness and the admissible range of λ \lambda λ ). For a fixed loss pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) and λ ≥ 0 \lambda\geq0 λ ≥ 0 , let K \mathcal{K} K be a domain on which the conditional risk is well posed:
C f a i r L , L ˉ , ∗ ( η , η ˉ ) > − ∞ for every ( η , η ˉ ) ∈ K , C_{\mathrm{fair}}^{L,\bar{L},*}(\eta,\bar{\eta}) > -\infty\quad\text{for every }(\eta,\bar{\eta}) \in\mathcal{K}, C fair L , L ˉ , ∗ ( η , η ˉ ) > − ∞ for every ( η , η ˉ ) ∈ K ,
The parameter λ \lambda λ is admissible for a distribution ( D , D ˉ ) (D,\bar{D}) ( D , D ˉ ) if
( η ( X ) , η ˉ ( X ) ) ∈ { ( η , η ˉ ) : C f a i r L , L ˉ , ∗ ( η , η ˉ ) > − ∞ } (\eta(X),\bar{\eta}(X)) \in\{(\eta,\bar{\eta}): C_{\mathrm{fair}}^{L,\bar{L},*}(\eta,\bar{\eta}) > -\infty\} ( η ( X ) , η ˉ ( X )) ∈ {( η , η ˉ ) : C fair L , L ˉ , ∗ ( η , η ˉ ) > − ∞ }
almost surely.
For a fixed distribution this is a condition on λ \lambda λ , not on the classifier. When L ˉ ≥ 0 \bar{L} \ge0 L ˉ ≥ 0 , the maximal well-posed domain shrinks as λ \lambda λ increases, so the admissible λ \lambda λ form an interval starting at 0. Its right endpoint is
λ max : = sup { λ ≥ 0 : ( η ( X ) , η ˉ ( X ) ) ∈ K almost surely } . \lambda_{\max} := \sup\{\lambda\ge0 : (\eta(X), \bar{\eta}(X)) \in\mathcal{K}\text{ almost surely}\}. λ m a x := sup { λ ≥ 0 : ( η ( X ) , η ˉ ( X )) ∈ K almost surely } .
For the linear L ˉ \bar{L} L ˉ in the examples, the same conclusion follows from the explicit linear inequalities, and this endpoint is computed there. If the condition fails, then C f a i r L , L ˉ , ∗ = − ∞ C_{\mathrm{fair}}^{L,\bar{L},*} = -\infty C fair L , L ˉ , ∗ = − ∞ on a set of positive probability, so R f a i r L , L ˉ , ∗ = − ∞ \mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*} = -\infty R fair L , L ˉ , ∗ = − ∞ and the excess risk is undefined. All subsequent results are stated for admissible λ \lambda λ .
On K \mathcal{K} K , H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ge0 H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ 0 , and it is finite by the definition of K \mathcal{K} K .
Definition 12 (FAFC). For a fixed λ \lambda λ , let K \mathcal{K} K be a well-posed domain as in Remark 11. L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K with respect to α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) if
H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 for all ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) ≠ 0. H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 \quad\text{for all }(\eta,\bar{\eta}) \in\mathcal{K}\text{ with }D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0. H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 for all ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) = 0.
Intuitively, L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K if, at every point ( η , η ˉ ) ∈ K (\eta,\bar{\eta}) \in\mathcal{K} ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0 D α , α ˉ ( η , η ˉ ) = 0 , the value of t = f ( x ) t = f(x) t = f ( x ) minimizing the conditional ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk has the same sign as D α , α ˉ ( η , η ˉ ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) D α , α ˉ ( η , η ˉ ) . Corollary 13 shows that this pointwise requirement is inherited by every minimizer of the population risk whenever the support of ( η ( X ) , η ˉ ( X ) ) (\eta(X),\bar{\eta}(X)) ( η ( X ) , η ˉ ( X )) lies in K \mathcal{K} K .
Corollary 13. Recall the definition of fairness-aware discriminant D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ in (13) . If L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K , then for any minimizer f ∗ ∈ arg min f ∈ F R f a i r L , L ˉ ( f ; D , D ˉ ) f^* \in\arg\min_{f\in\mathcal{F}} \mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f;D,\bar{D}) f ∗ ∈ arg min f ∈ F R fair L , L ˉ ( f ; D , D ˉ ) , its sign matches the fairness-aware discriminant:
sign ( f ∗ ( x ) ) = sign ( D α , α ˉ ( η ( x ) , η ˉ ( x ) ) ) for P X -a.e. x with D α , α ˉ ( η ( x ) , η ˉ ( x ) ) ≠ 0. \operatorname{sign}(f^*(x)) = \operatorname{sign}\left(D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x))\right) \quad\text{for }P_X\text{-a.e. }x\text{ with }D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x)) \ne0. sign ( f ∗ ( x )) = sign ( D α , α ˉ ( η ( x ) , η ˉ ( x )) ) for P X -a.e. x with D α , α ˉ ( η ( x ) , η ˉ ( x )) = 0.
Note that Corollary 13 is one-directional for a fixed distribution: FAFC on K \mathcal{K} K implies the population-level statement, but the population-level statement only constrains ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) almost surely on the support. Example 32 (See Appendix § 8.7 ) illustrates this fixed-distribution gap.
The following theorem provides a necessary and sufficient condition for FAFC on K \mathcal{K} K . The condition is stated purely in terms of the derivatives of the four partial losses L 1 , L − 1 , L ‾ 1 , L ‾ − 1 L_1,L_{-1},\overline{L}_1,\overline{L}_{-1} L 1 , L − 1 , L 1 , L − 1 at 0.
Theorem 14. Let α , α ‾ ∈ ( 0 , 1 ) \alpha,\overline{\alpha}\in(0,1) α , α ∈ ( 0 , 1 ) , λ > 0 \lambda>0 λ > 0 , let K \mathcal{K} K be a well-posed domain as in Remark 11, and let L f a i r L_{\mathrm{fair}} L fair be a fairness-aware surrogate loss. Assume the partial losses L 1 , L − 1 , L ‾ 1 , L ‾ − 1 L_1,L_{-1},\overline{L}_1,\overline{L}_{-1} L 1 , L − 1 , L 1 , L − 1 are convex and differentiable at 0. Set r : = λ / π r:=\lambda/\pi r := λ / π , where π = P ( Y = 1 ) \pi=\mathbb{P}(Y=1) π = P ( Y = 1 ) . Furthermore, suppose that, for every ( η , η ‾ ) ∈ K (\eta,\overline{\eta})\in\mathcal{K} ( η , η ) ∈ K , the map t ↦ C f a i r L , L ‾ ( η , η ‾ , t ) t\mapsto C_{\mathrm{fair}}^{L,\overline{L}}(\eta,\overline{\eta},t) t ↦ C fair L , L ( η , η , t ) is strictly convex on R \mathbb{R} R , and that K \mathcal{K} K contains a neighborhood of ( α , α ‾ ) (\alpha,\overline{\alpha}) ( α , α ) . Then, in the attribute-blind setting, L f a i r L_{\mathrm{fair}} L fair is ( α , α ‾ ) (\alpha,\overline{\alpha}) ( α , α ) -FAFC on K \mathcal{K} K if and only if the following conditions hold for the corresponding criterion:
( D P ) : ( D 1 ) L 1 ′ ( 0 ) < L − 1 ′ ( 0 ) , ( D 2 ) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ‾ 1 ′ ( 0 ) − L ‾ − 1 ′ ( 0 ) , ( D 3 ) α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = λ [ α ‾ L ‾ 1 ′ ( 0 ) + ( 1 − α ‾ ) L ‾ − 1 ′ ( 0 ) ] ; \begin{aligned}
(DP):\quad&(D1)\quad L_1'(0)<L_{-1}'(0),\\
&(D2)\quad L_1'(0)-L_{-1}'(0)=\overline{L}_1'(0)-\overline{L}_{-1}'(0),\\
&(D3)\quad\alpha L_1'(0)+(1-\alpha)L_{-1}'(0)
=\lambda\left[\overline{\alpha}\,\overline{L}_1'(0)+(1-\overline{\alpha})\overline{L}_{-1}'(0)\right];
\end{aligned} ( D P ) : ( D 1 ) L 1 ′ ( 0 ) < L − 1 ′ ( 0 ) , ( D 2 ) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) , ( D 3 ) α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = λ [ α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) ] ; ( E O ) : ( E 1 ) L ‾ 1 ′ ( 0 ) − L ‾ − 1 ′ ( 0 ) < 0 , ( E 2 ) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ‾ 1 ′ ( 0 ) − L ‾ − 1 ′ ( 0 ) + r [ α ‾ L ‾ 1 ′ ( 0 ) + ( 1 − α ‾ ) L ‾ − 1 ′ ( 0 ) ] , ( E 3 ) α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = r α [ α ‾ L ‾ 1 ′ ( 0 ) + ( 1 − α ‾ ) L ‾ − 1 ′ ( 0 ) ] . \begin{aligned}
(EO):\quad&(E1)\quad\overline{L}_1'(0)-\overline{L}_{-1}'(0)<0,\\
&(E2)\quad L_1'(0)-L_{-1}'(0)=\overline{L}_1'(0)-\overline{L}_{-1}'(0)
+r\left[\overline{\alpha}\,\overline{L}_1'(0)+(1-\overline{\alpha})\overline{L}_{-1}'(0)\right],\\
&(E3)\quad\alpha L_1'(0)+(1-\alpha)L_{-1}'(0)
=r\alpha\left[\overline{\alpha}\,\overline{L}_1'(0)+(1-\overline{\alpha})\overline{L}_{-1}'(0)\right].
\end{aligned} ( E O ) : ( E 1 ) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) < 0 , ( E 2 ) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) + r [ α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) ] , ( E 3 ) α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = r α [ α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) ] .
Remark 15 (Attribute-aware setting). Theorem 14 is stated for the attribute-blind setting. Its sufficiency direction applies verbatim to the attribute-aware domain. For necessity, one additional nondegeneracy condition is required: for each a ∈ { − 1 , 1 } a\in\{-1,1\} a ∈ { − 1 , 1 } , the section K a = { η : ( η , a ) ∈ K } \mathcal{K}_a=\{\eta:(\eta,a)\in\mathcal{K}\} K a = { η : ( η , a ) ∈ K } is an interval, and there exists an interior point η a \eta_a η a of the interval K a \mathcal{K}_a K a such that D α , α ‾ ( η a , a ) = 0 D^{\alpha,\overline{\alpha}}(\eta_a,a)=0 D α , α ( η a , a ) = 0 . This condition ensures that the decision boundary crosses each attribute edge, so that FAFC controls both sides of the boundary. Since D ( ⋅ , a ) D(\cdot,a) D ( ⋅ , a ) is affine, it fails only when the boundary misses an entire edge or meets it at an endpoint; it therefore restricts neither the loss class nor the data distribution and is a mild nondegeneracy condition. It holds, for instance, in the weighted-logistic example below. With this condition, the proof is the edgewise analogue of the attribute-blind proof.
Theorem 14 provides the Fisher consistency property for the fairness-aware setting: whenever the fairness-regularized objective is well-posed, standard surrogate losses retain their consistency guarantees provided the loss pair satisfies the corresponding derivative conditions.
Remark 16 (Connection to Classical Fisher Consistency). Each condition in Theorem 14 traces back to the characterization of classical Fisher consistency (Theorem 3). We discuss the DP case; the EO case reduces to it when α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) = 0 \bar{\alpha}\bar{L}'_1(0) + (1-\bar{\alpha})\bar{L}'_{-1}(0) = 0 α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) = 0 , which holds in the examples below. In the degenerate case λ = 0 \lambda= 0 λ = 0 (equivalently, as λ → 0 \lambda\to0 λ → 0 in the conditions above), (D3) is the classical decision-boundary condition α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0 \alpha L'_1(0) + (1-\alpha)L'_{-1}(0) = 0 α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0 . (D1) is the classical sign requirement L 1 ′ ( 0 ) < 0 < L − 1 ′ ( 0 ) L'_1(0) < 0 < L'_{-1}(0) L 1 ′ ( 0 ) < 0 < L − 1 ′ ( 0 ) , written in the equivalent form L 1 ′ ( 0 ) < L − 1 ′ ( 0 ) L'_1(0) < L'_{-1}(0) L 1 ′ ( 0 ) < L − 1 ′ ( 0 ) , which suffices once (D3) holds. (D2) has no classical counterpart: it is the fairness-specific requirement that the two surrogate losses share the same derivative gap at 0.
Calibration Inequality for Fairness-Aware Classification
Having characterized when the composite fairness-aware loss L f a i r L_{\mathrm{fair}} L fair is FAFC, we now answer the next central question: establishing a quantitative bound between the excess fairness-aware risk and the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk under the attribute-blind setting. Then, the attribute-aware setting will be discussed in Remark 18 .
Throughout this section, K ⊆ [ 0 , 1 ] 2 \mathcal{K} \subseteq[0,1]^2 K ⊆ [ 0 , 1 ] 2 is a compact, connected subset of a well-posed domain as in Remark 11 that contains ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) , and every distribution considered satisfies ( η ( X ) , η ˉ ( X ) ) ∈ K (\eta(X),\bar{\eta}(X)) \in\mathcal{K} ( η ( X ) , η ˉ ( X )) ∈ K almost surely. Define the maximum discriminant value as B max : = max ( η , η ˉ ) ∈ K ∣ D α , α ˉ ( η , η ˉ ) ∣ B_{\max} := \max_{(\eta,\bar{\eta})\in\mathcal{K}} |D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})| B m a x := max ( η , η ˉ ) ∈ K ∣ D α , α ˉ ( η , η ˉ ) ∣ . For any ϵ ∈ [ 0 , B max ] \epsilon\in[0,B_{\max}] ϵ ∈ [ 0 , B m a x ] , the calibration function ν L , L ˉ α , α ˉ : [ 0 , B max ] → [ 0 , ∞ ] \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} : [0,B_{\max}] \to[0,\infty] ν L , L ˉ α , α ˉ : [ 0 , B m a x ] → [ 0 , ∞ ] is defined as
ν L , L ˉ α , α ˉ ( ϵ ) : = inf ( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ H L , L ˉ α , α ˉ ( η , η ˉ ) . \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) := \inf_{(\eta,\bar{\eta})\in\mathcal{K}:|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})|=\epsilon} H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). ν L , L ˉ α , α ˉ ( ϵ ) := ( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ inf H L , L ˉ α , α ˉ ( η , η ˉ ) .
If the level set { ( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ } \{(\eta,\bar{\eta}) \in\mathcal{K} : |D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})| = \epsilon\} {( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ } is empty for some ϵ ∈ ( 0 , B max ] \epsilon\in(0,B_{\max}] ϵ ∈ ( 0 , B m a x ] , we define ν L , L ˉ α , α ˉ ( ϵ ) = ∞ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) = \infty ν L , L ˉ α , α ˉ ( ϵ ) = ∞ .
Since the convexity of ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ cannot be guaranteed in general, we require its convex envelope to utilize Jensen’s inequality for the excess risk bound. Concretely, we define
ψ L , L ˉ α , α ˉ : = ( ν L , L ˉ α , α ˉ ) ∗ ∗ , for all ϵ ∈ [ 0 , B max ] , \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} := \left(\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}\right)^{**}, \quad\text{for all } \epsilon\in[0,B_{\max}], ψ L , L ˉ α , α ˉ := ( ν L , L ˉ α , α ˉ ) ∗∗ , for all ϵ ∈ [ 0 , B m a x ] ,
where g ∗ ∗ g^{**} g ∗∗ denotes the Fenchel-Legendre biconjugate of g g g . Equivalently, ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is the largest convex, lower semi-continuous function majorized by ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ , and it can be characterized via the relation Epi g ∗ ∗ = co Epi g ‾ \operatorname{Epi} g^{**} = \overline{\operatorname{co}\operatorname{Epi} g} Epi g ∗∗ = co Epi g , where Epi g = { ( r , s ) : g ( r ) ≤ s } \operatorname{Epi} g = \{(r,s) : g(r) \le s\} Epi g = {( r , s ) : g ( r ) ≤ s } denotes the epigraph of g g g ; “co” stands for the convex hull of a set, and the overbar signifies the closure of the set.
With ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ and its convex envelope ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ defined, we now state the theorem relating the fairness-aware excess risk to the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk.
Theorem 17. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) , let L f a i r = L − λ L ˉ L_{\mathrm{fair}} = L - \lambda\bar{L} L fair = L − λ L ˉ be a fairness-aware loss , and let K \mathcal{K} K be as above.
For any f ∈ F f \in\mathcal{F} f ∈ F and distribution D , D ˉ D,\bar{D} D , D ˉ with ( η ( X ) , η ˉ ( X ) ) ∈ K (\eta(X),\bar{\eta}(X)) \in\mathcal{K} ( η ( X ) , η ˉ ( X )) ∈ K almost surely, we have
ψ L , L ˉ α , α ˉ ( R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ) ≤ R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}\left(\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f) - \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}\right) \le\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f) - \mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}. ψ L , L ˉ α , α ˉ ( R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ) ≤ R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ . Furthermore, if the map t ↦ C f a i r L , L ˉ ( η , η ˉ , t ) t \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) t ↦ C fair L , L ˉ ( η , η ˉ , t ) is convex on R \mathbb{R} R , then ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is invertible if and only if L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K .
Application to Fairness-Aware Classification
This section applies the framework of § 3 and § 4 to concrete classifiers. To illustrate the FAFC conditions in a concrete model, we first revisit the fairness-aware logistic regression proposed by Zafar et al. [30 ] and verify FAFC on the relevant domain for its composite loss L f a i r L_{\mathrm{fair}} L fair , including the well-posedness range of λ \lambda λ and a closed-form calibration function. We then present a general surrogate loss-based fairness-aware classification algorithm that can be used with any surrogate loss pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) and trade-off λ \lambda λ for which L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on a domain containing the support.
Fairness-Aware Logistic Regression
We begin with fairness-aware logistic regression under DP in the attribute-blind setting—the configuration studied by Zafar et al. [30 ] . Let the accuracy surrogate be L ( y , t ) = L(y,t) = L ( y , t ) = log ( 1 + e − y t ) \log(1+e^{-yt}) log ( 1 + e − y t ) and the fairness surrogate be L ˉ ( a , t ) = ( π A − a + 1 2 ) t \bar{L}(a,t)=\left(\pi_A-\frac{a+1}{2}\right)t L ˉ ( a , t ) = ( π A − 2 a + 1 ) t , where π A = P ( A = 1 ) \pi_A=\mathbb{P}(A=1) π A = P ( A = 1 ) .
The fairness-aware surrogate loss is
L f a i r ( y , a , t ) = L ( y , t ) − λ L ˉ ( a , t ) = log ( 1 + e − y t ) − λ ( π A − a + 1 2 ) t . (14) L_{\mathrm{fair}}(y,a,t)=L(y,t)-\lambda\bar{L}(a,t)=\log(1+e^{-yt})-\lambda\left(\pi_A-\frac{a+1}{2}\right)t. \tag*{(14)} L fair ( y , a , t ) = L ( y , t ) − λ L ˉ ( a , t ) = log ( 1 + e − y t ) − λ ( π A − 2 a + 1 ) t . (14)
This pair fits the framework of § 2.4 . The partial losses are L 1 ( t ) = log ( 1 + e − t ) L_1(t)=\log(1+e^{-t}) L 1 ( t ) = log ( 1 + e − t ) , L − 1 ( t ) = log ( 1 + e t ) L_{-1}(t)=\log(1+e^t) L − 1 ( t ) = log ( 1 + e t ) , L ˉ 1 ( t ) = ( π A − 1 ) t \bar{L}_1(t)=(\pi_A-1)t L ˉ 1 ( t ) = ( π A − 1 ) t , and L ˉ − 1 ( t ) = π A t \bar{L}_{-1}(t)=\pi_A t L ˉ − 1 ( t ) = π A t . The following proposition proves that the logistic composite loss L f a i r L_{\mathrm{fair}} L fair is FAFC on the domain K \mathcal{K} K defined below.
Proposition 19. Let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be the logistic surrogate loss pair defined in Eq. (14) , and let
K : = { ( η , η ˉ ) ∈ [ 0 , 1 ] 2 : 0 ≤ η − λ ( η ˉ − π A ) ≤ 1 } . \mathcal{K}:=\left\{(\eta,\bar{\eta})\in[0,1]^2: 0\leq\eta-\lambda(\bar{\eta}-\pi_A)\leq1\right\}. K := { ( η , η ˉ ) ∈ [ 0 , 1 ] 2 : 0 ≤ η − λ ( η ˉ − π A ) ≤ 1 } . If 0 < λ < λ max 0<\lambda<\lambda_{\max} 0 < λ < λ m a x , with λ max \lambda_{\max} λ m a x given in (15) , then ( η ( X ) , η ˉ ( X ) ) ∈ K (\eta(X),\bar{\eta}(X))\in\mathcal{K} ( η ( X ) , η ˉ ( X )) ∈ K almost surely, and L f a i r L_{\mathrm{fair}} L fair is ( 0.5 , π A ) (0.5,\pi_A) ( 0.5 , π A ) -FAFC on K \mathcal{K} K for the DP criterion in the attribute-blind setting.
Proof. We verify (D1)–(D3) of Theorem 14 . (D1) holds since L 1 ′ ( 0 ) = − 1 2 < L − 1 ′ ( 0 ) = 1 2 L'_1(0)=-\frac{1}{2}<L'_{-1}(0)=\frac{1}{2} L 1 ′ ( 0 ) = − 2 1 < L − 1 ′ ( 0 ) = 2 1 .
(D2) holds since L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 1 = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) L'_1(0)-L'_{-1}(0)=-1=\bar{L}'_1(0)-\bar{L}'_{-1}(0) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 1 = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) .
For (D3), with α = 0.5 \alpha=0.5 α = 0.5 ,
( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = 1 2 ⋅ 1 2 + 1 2 ⋅ ( − 1 2 ) = 0 , (1-\alpha)L'_{-1}(0)+\alpha L'_1(0)=\frac{1}{2}\cdot\frac{1}{2}+\frac{1}{2}\cdot\left(-\frac{1}{2}\right)=0, ( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = 2 1 ⋅ 2 1 + 2 1 ⋅ ( − 2 1 ) = 0 ,
and with α ˉ = π A \bar{\alpha}=\pi_A α ˉ = π A ,
( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 ) = ( 1 − π A ) π A + π A ( π A − 1 ) = 0. (1-\bar{\alpha})\bar{L}'_{-1}(0)+\bar{\alpha}\bar{L}'_1(0)=(1-\pi_A)\pi_A+\pi_A(\pi_A-1)=0. ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 ) = ( 1 − π A ) π A + π A ( π A − 1 ) = 0.
Hence (D3)
( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = λ [ ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 ) ] , (1-\alpha)L'_{-1}(0)+\alpha L'_1(0)=\lambda\left[(1-\bar{\alpha})\bar{L}'_{-1}(0)+\bar{\alpha}\bar{L}'_1(0)\right], ( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = λ [ ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 ) ] ,
holds. On K \mathcal{K} K the conditional risk is strictly convex and bounded below: strict convexity is shown in the next paragraph for all ( η , η ˉ ) ∈ [ 0 , 1 ] 2 (\eta,\bar{\eta})\in[0,1]^2 ( η , η ˉ ) ∈ [ 0 , 1 ] 2 , and boundedness below follows from 0 ≤ η − λ ( η ˉ − π A ) ≤ 1 0\leq\eta-\lambda(\bar{\eta}-\pi_A)\leq1 0 ≤ η − λ ( η ˉ − π A ) ≤ 1 and the tail expansion below. Hence the sufficiency direction of Theorem 14 gives that L f a i r L_{\mathrm{fair}} L fair is ( 0.5 , π A ) (0.5,\pi_A) ( 0.5 , π A ) -FAFC on K \mathcal{K} K . The inclusion of the support in K \mathcal{K} K is verified in the next paragraph.
It remains to ensure that the fairness-aware risk is well-posed (Remark 11). Substituting the surrogate loss pair into the conditional ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk gives
C f a i r L , L ˉ ( η , η ˉ , t ) = η log ( 1 + e − t ) + ( 1 − η ) log ( 1 + e t ) − λ [ η ˉ ( π A − 1 ) t + ( 1 − η ˉ ) π A t ] = η log ( 1 + e − t ) + ( 1 − η ) log ( 1 + e t ) + λ ( η ˉ − π A ) t . \begin{aligned}
C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) = \eta\log(1+e^{-t}) + (1-\eta)\log(1+e^t) - \lambda[\bar{\eta}(\pi_A-1)t + (1-\bar{\eta})\pi_A t] \\
&= \eta\log(1+e^{-t}) + (1-\eta)\log(1+e^t) + \lambda(\bar{\eta}-\pi_A)t.
\end{aligned} C fair L , L ˉ ( η , η ˉ , t ) = η log ( 1 + e − t ) + ( 1 − η ) log ( 1 + e t ) − λ [ η ˉ ( π A − 1 ) t + ( 1 − η ˉ ) π A t ] = η log ( 1 + e − t ) + ( 1 − η ) log ( 1 + e t ) + λ ( η ˉ − π A ) t .
The conditional risk is strictly convex in t t t , since the second derivatives of log ( 1 + e − t ) \log(1+e^{-t}) log ( 1 + e − t ) and log ( 1 + e t ) \log(1+e^t) log ( 1 + e t ) both equal e t / ( 1 + e t ) 2 e^t/(1+e^t)^2 e t / ( 1 + e t ) 2 ,
∂ 2 ∂ t 2 C f a i r L , L ˉ ( η , η ˉ , t ) = e t ( 1 + e t ) 2 > 0 \frac{\partial^2}{\partial t^2}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) = \frac{e^t}{(1+e^t)^2} > 0 ∂ t 2 ∂ 2 C fair L , L ˉ ( η , η ˉ , t ) = ( 1 + e t ) 2 e t > 0
for all ( η , η ˉ ) ∈ [ 0 , 1 ] 2 (\eta,\bar{\eta}) \in[0,1]^2 ( η , η ˉ ) ∈ [ 0 , 1 ] 2 and t ∈ R t \in\mathbb{R} t ∈ R .
With S : = η − λ ( η ˉ − π A ) S := \eta-\lambda(\bar{\eta}-\pi_A) S := η − λ ( η ˉ − π A ) , the conditional risk can be written as
C f a i r L , L ˉ ( η , η ˉ , t ) = log ( 1 + e t ) − S t . C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) = \log(1+e^t)-St. C fair L , L ˉ ( η , η ˉ , t ) = log ( 1 + e t ) − S t .
Since log ( 1 + e t ) = t + o ( t ) \log(1+e^t)=t+o(t) log ( 1 + e t ) = t + o ( t ) as t → + ∞ t\to+\infty t → + ∞ and log ( 1 + e t ) = o ( t ) \log(1+e^t)=o(t) log ( 1 + e t ) = o ( t ) as t → − ∞ t\to-\infty t → − ∞ , the tails satisfy
C f a i r L , L ˉ ( η , η ˉ , t ) = { ( 1 − S ) t + o ( t ) , t → + ∞ , − S t + o ( t ) , t → − ∞ , C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) =
\begin{cases}
(1-S)t+o(t), & t\to+\infty,\\
-St+o(t), & t\to-\infty,
\end{cases} C fair L , L ˉ ( η , η ˉ , t ) = { ( 1 − S ) t + o ( t ) , − S t + o ( t ) , t → + ∞ , t → − ∞ ,
so the conditional risk is bounded below if and only if 0 ≤ S ≤ 1 0\le S\le1 0 ≤ S ≤ 1 . Because σ ( R ) = ( 0 , 1 ) \sigma(\mathbb{R})=(0,1) σ ( R ) = ( 0 , 1 ) and the conditional risk is strictly convex, the pointwise minimizer is uniquely characterized by σ ( t ∗ ) = S \sigma(t^*)=S σ ( t ∗ ) = S precisely when 0 < S < 1 0<S<1 0 < S < 1 . Consequently, we require 0 < S ( x ) < 1 0<S(x)<1 0 < S ( x ) < 1 for P X P_X P X -a.e. x x x . For λ > 0 \lambda>0 λ > 0 , this is ensured by λ < λ max \lambda<\lambda_{\max} λ < λ m a x , where
λ max : = min { inf x ∈ supp ( X ) : η ˉ ( x ) > π A η ( x ) η ˉ ( x ) − π A , inf x ∈ supp ( X ) : η ˉ ( x ) < π A 1 − η ( x ) π A − η ˉ ( x ) } . (15) \lambda_{\max}:=\min\left\{\inf_{\substack{x\in\operatorname{supp}(X):\\\bar{\eta}(x)>\pi_A}}\frac{\eta(x)}{\bar{\eta}(x)-\pi_A},\ \inf_{\substack{x\in\operatorname{supp}(X):\\\bar{\eta}(x)<\pi_A}}\frac{1-\eta(x)}{\pi_A-\bar{\eta}(x)}\right\}. \tag*{(15)} λ m a x := min ⎩ ⎨ ⎧ x ∈ supp ( X ) : η ˉ ( x ) > π A inf η ˉ ( x ) − π A η ( x ) , x ∈ supp ( X ) : η ˉ ( x ) < π A inf π A − η ˉ ( x ) 1 − η ( x ) ⎭ ⎬ ⎫ . (15)
and the infimum over an empty set is understood as + ∞ +\infty + ∞ . Under λ ∈ ( 0 , λ max ) \lambda\in(0,\lambda_{\max}) λ ∈ ( 0 , λ m a x ) , the conditional risk is strictly convex with a unique pointwise minimizer characterized by σ ( t ∗ ) = S ∈ ( 0 , 1 ) \sigma(t^*) = S \in(0,1) σ ( t ∗ ) = S ∈ ( 0 , 1 ) , hence sign ( t ∗ ) = sign ( D α , α ˉ ) \operatorname{sign}(t^*) = \operatorname{sign}(D^{\alpha,\bar{\alpha}}) sign ( t ∗ ) = sign ( D α , α ˉ ) and the support of ( η ( X ) , η ˉ ( X ) ) (\eta(X),\bar{\eta}(X)) ( η ( X ) , η ˉ ( X )) lies in the interior of K \mathcal{K} K . Both bounds in (15) are support-dependent: the first degenerates as η ( x ) → 0 \eta(x) \to0 η ( x ) → 0 on { η ˉ ( x ) > π A } \{\bar{\eta}(x) > \pi_A\} { η ˉ ( x ) > π A } and the second degenerates as η ( x ) → 1 \eta(x) \to1 η ( x ) → 1 on { η ˉ ( x ) < π A } \{\bar{\eta}(x) < \pi_A\} { η ˉ ( x ) < π A } , so no bound on λ \lambda λ that is independent of the support can be given.
Having verified FAFC on K \mathcal{K} K for L f a i r L_{\mathrm{fair}} L fair , we now derive the calibration function and excess-risk bound for this example.
Proposition 20. Suppose 0 < λ < λ max 0 < \lambda< \lambda_{\max} 0 < λ < λ m a x , and let K \mathcal{K} K be the domain of Proposition 19. Then, on K \mathcal{K} K , for every ϵ ∈ [ 0 , 0.5 ) \epsilon\in[0,0.5) ϵ ∈ [ 0 , 0.5 ) the calibration function of the pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) is
ν L , L ˉ α , α ˉ ( ϵ ) = ( 0.5 + ϵ ) ln ( 1 + 2 ϵ ) + ( 0.5 − ϵ ) ln ( 1 − 2 ϵ ) . \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) = (0.5+\epsilon)\ln(1+2\epsilon) + (0.5-\epsilon)\ln(1-2\epsilon). ν L , L ˉ α , α ˉ ( ϵ ) = ( 0.5 + ϵ ) ln ( 1 + 2 ϵ ) + ( 0.5 − ϵ ) ln ( 1 − 2 ϵ ) . Its convex envelope coincides with it, ψ L , L ˉ α , α ˉ = ν L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} = \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ = ν L , L ˉ α , α ˉ , the quadratic lower bound ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 2 ϵ 2 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) \ge2\epsilon^2 ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 2 ϵ 2 holds, and consequently, for every f ∈ F f \in\mathcal{F} f ∈ F ,
R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ≤ min { 1 2 , 1 2 ( R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ ) } . (16) \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f) - \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*} \le\min\left\{\frac{1}{2},\sqrt{\frac{1}{2}\left(\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f) - \mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}\right)}\right\}. \tag*{(16)} R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ≤ min { 2 1 , 2 1 ( R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ ) } . (16) Proof. Since C f a i r L , L ˉ ( η , η ˉ , t ) = log ( 1 + e t ) − S t C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) = \log(1+e^t) - St C fair L , L ˉ ( η , η ˉ , t ) = log ( 1 + e t ) − S t and S = η − λ ( η ˉ − π A ) S = \eta- \lambda(\bar{\eta}-\pi_A) S = η − λ ( η ˉ − π A ) , then we have the derivative of the conditional risk is
∂ ∂ t C f a i r L , L ˉ ( η , η ˉ , t ) = σ ( t ) − S , \frac{\partial}{\partial t}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) = \sigma(t) - S, ∂ t ∂ C fair L , L ˉ ( η , η ˉ , t ) = σ ( t ) − S , where σ ( t ) = 1 / ( 1 + e − t ) \sigma(t) = 1/(1+e^{-t}) σ ( t ) = 1/ ( 1 + e − t ) . Setting the derivative to zero yields σ ( t ∗ ) = S \sigma(t^*) = S σ ( t ∗ ) = S , and D α , α ˉ ( η , η ˉ ) = η − 0.5 − λ ( η ˉ − π A ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = \eta- 0.5 - \lambda(\bar{\eta}-\pi_A) D α , α ˉ ( η , η ˉ ) = η − 0.5 − λ ( η ˉ − π A ) . Hence D α , α ˉ = S − 0.5 D^{\alpha,\bar{\alpha}} = S - 0.5 D α , α ˉ = S − 0.5 . So, we can evaluate the conditional risk at the unique minimizer gives
C f a i r L , L ˉ , ∗ ( η , η ˉ ) = − S log S − ( 1 − S ) log ( 1 − S ) , C_{\mathrm{fair}}^{L,\bar{L},*}(\eta,\bar{\eta}) = -S\log S - (1-S)\log(1-S), C fair L , L ˉ , ∗ ( η , η ˉ ) = − S log S − ( 1 − S ) log ( 1 − S ) ,
and the wrong-sign infimum is attained at t = 0 t=0 t = 0 :
C f a i r L , L ˉ , − ( η , η ˉ ) = C f a i r L , L ˉ ( η , η ˉ , 0 ) = log 2. C_{\mathrm{fair}}^{L,\bar{L},-}(\eta,\bar{\eta})=C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},0)=\log2. C fair L , L ˉ , − ( η , η ˉ ) = C fair L , L ˉ ( η , η ˉ , 0 ) = log 2.
Hence the wrong-sign excess conditional ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk is
H L , L ˉ α , α ˉ ( η , η ˉ ) = S log S + ( 1 − S ) log ( 1 − S ) + log 2. H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=S\log S+(1-S)\log(1-S)+\log2. H L , L ˉ α , α ˉ ( η , η ˉ ) = S log S + ( 1 − S ) log ( 1 − S ) + log 2.
Note that H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ and D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ depend on η \eta η and η ˉ \bar{\eta} η ˉ solely through S S S .
To find ν L , L ˉ α , α ˉ ( ϵ ) \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) ν L , L ˉ α , α ˉ ( ϵ ) , we minimize H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ over the set { ( η , η ˉ ) : ∣ S − 0.5 ∣ = ϵ } \{(\eta,\bar{\eta}): |S-0.5|=\epsilon\} {( η , η ˉ ) : ∣ S − 0.5∣ = ϵ } . This constraint implies S = 0.5 ± ϵ S=0.5\pm\epsilon S = 0.5 ± ϵ . Since H H H depends on S S S only through the symmetric quantity S log S + ( 1 − S ) log ( 1 − S ) + log 2 S\log S+(1-S)\log(1-S)+\log2 S log S + ( 1 − S ) log ( 1 − S ) + log 2 , we have H L , L ˉ α , α ˉ ( 0.5 + ϵ ) = H L , L ˉ α , α ˉ ( 0.5 − ϵ ) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0.5+\epsilon)=H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0.5-\epsilon) H L , L ˉ α , α ˉ ( 0.5 + ϵ ) = H L , L ˉ α , α ˉ ( 0.5 − ϵ ) . Thus, the minimization yields the exact calibration function:
ν L , L ˉ α , α ˉ ( ϵ ) = ( 0.5 + ϵ ) ln ( 1 + 2 ϵ ) + ( 0.5 − ϵ ) ln ( 1 − 2 ϵ ) . \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)=(0.5+\epsilon)\ln(1+2\epsilon)+(0.5-\epsilon)\ln(1-2\epsilon). ν L , L ˉ α , α ˉ ( ϵ ) = ( 0.5 + ϵ ) ln ( 1 + 2 ϵ ) + ( 0.5 − ϵ ) ln ( 1 − 2 ϵ ) .
ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ is convex on [ 0 , 0.5 ) [0,0.5) [ 0 , 0.5 ) , since its second derivative equals ν ′ ′ = 4 / ( 1 − 4 ϵ 2 ) > 0 \nu''=4/(1-4\epsilon^2)>0 ν ′′ = 4/ ( 1 − 4 ϵ 2 ) > 0 . So
ψ L , L ˉ α , α ˉ ( ϵ ) = ν L , L ˉ α , α ˉ ( ϵ ) . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)=\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon). ψ L , L ˉ α , α ˉ ( ϵ ) = ν L , L ˉ α , α ˉ ( ϵ ) .
Finally, noting that H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ is the KL divergence between the Bernoulli distributions with parameters S S S and 0.5 0.5 0.5 , and applying Pinsker’s inequality D K L ( p ∥ q ) ≥ 2 ( p − q ) 2 D_{\mathrm{KL}}(p\|q)\ge2(p-q)^2 D KL ( p ∥ q ) ≥ 2 ( p − q ) 2 , we obtain the quadratic lower bound
ψ L , L ˉ α , α ˉ ( ϵ ) = ν L , L ˉ α , α ˉ ( ϵ ) ≥ 2 ( ( 0.5 + ϵ ) − 0.5 ) 2 = 2 ϵ 2 . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)=\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)\ge2((0.5+\epsilon)-0.5)^2=2\epsilon^2. ψ L , L ˉ α , α ˉ ( ϵ ) = ν L , L ˉ α , α ˉ ( ϵ ) ≥ 2 (( 0.5 + ϵ ) − 0.5 ) 2 = 2 ϵ 2 .
Combining this lower bound with the calibration inequality of Theorem 17 (which applies because the support lies in K \mathcal{K} K ) yields 2 ( R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ) 2 ≤ R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ 2(\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f)-\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*})^2\le\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*} 2 ( R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ) 2 ≤ R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ , i.e.
R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ≤ 1 2 ( R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ ) . \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f)-\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}\le\sqrt{\frac{1}{2}\left(\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}\right)}. R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ≤ 2 1 ( R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ ) .
By Lemma 22 , since ∣ D α , α ˉ ∣ = ∣ S − 1 2 ∣ < 1 2 |D^{\alpha,\bar{\alpha}}|=|S-\frac{1}{2}|<\frac{1}{2} ∣ D α , α ˉ ∣ = ∣ S − 2 1 ∣ < 2 1 , the trivial bound R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ≤ 1 2 \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f)-\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}\le\frac{1}{2} R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ≤ 2 1 holds as well, and taking the minimum of the two gives (16) .
Up to now, we have focused on the DP criterion and the attribute-blind setting of fairness-aware logistic regression; the EO criterion, the attribute-aware setting, and further examples of FAFC composite losses are deferred (e.g., hinge and exponential loss) to Appendix § 8.6 .
A General Surrogate Loss-Based Fairness-Aware Classification Algorithm
The fairness-aware risk in Eq. (6) is built from discontinuous 0–1 losses and is not amenable to gradient-based learning from finite samples. The surrogate formulation in § 2.4 instead minimizes the population ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk R f a i r L , L ˉ ( f ; D , D ˉ ) \mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f;D,\bar{D}) R fair L , L ˉ ( f ; D , D ˉ ) . When L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K , every population minimizer aligns with the fairness-aware discriminant (Corollary 13 ); when in addition the calibration inequality of Theorem 17 applies, controlling the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk controls the excess fairness-aware risk. Algorithm 1 implements the same principle empirically: it replaces the population objective by its sample average, parameterizes the score as f θ f_\theta f θ , and minimizes the resulting smooth (or piecewise smooth) empirical ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk, optionally with a penalty Ω ( θ ) \Omega(\theta) Ω ( θ ) for stability. The DP and EO criteria enter only through the construction of the fairness term in Step 2, mirroring the population definitions in § 2.3 ; any pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) and trade-off λ \lambda λ for which L f a i r L_{\mathrm{fair}} L fair is FAFC can be plugged in at Step 1. Step 3 then calls a standard unconstrained optimizer (e.g., L-BFGS when the objective is differentiable).
§ 6 turns to numerical experiments that show the empirical performance of Algorithm 1.
Algorithm 1 General Surrogate Loss-Based Fairness-Aware Classification Algorithm
Input:
Cost parameters α , α ˉ ∈ ( 0 , 1 ) \alpha, \bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) ;
Dataset D = { ( x i , a i , y i ) } i = 1 n D = \{(x_i, a_i, y_i)\}_{i=1}^{n} D = {( x i , a i , y i ) } i = 1 n , where x i x_i x i denotes the feature, a i a_i a i the protected attribute, and y i y_i y i the label;
Trade-off parameter λ ≥ 0 \lambda\ge0 λ ≥ 0 ;
Fairness criterion (DP or EO).
Step 1 (Select the fairness-aware surrogate loss pair): Select a pair of fairness-aware surrogate loss functions ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) such that, for the given trade-off λ \lambda λ , L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on a domain K \mathcal{K} K containing the support of ( η ( X ) , η ˉ ( X ) ) (\eta(X),\bar{\eta}(X)) ( η ( X ) , η ˉ ( X )) .
Step 2 (Construct the empirical ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk objective):
For the DP fairness criterion, construct the empirical risk as:
R ^ D P L , L ˉ ( f θ ) = 1 n ∑ i = 1 n L ( y i , f θ ( x i ) ) − λ n ∑ i = 1 n L ˉ ( a i , f θ ( x i ) ) ; \widehat{\mathcal{R}}_{\mathrm{DP}}^{L,\bar{L}}(f_\theta)=\frac{1}{n}\sum_{i=1}^{n}L(y_i,f_\theta(x_i))-\frac{\lambda}{n}\sum_{i=1}^{n}\bar{L}(a_i,f_\theta(x_i)); R DP L , L ˉ ( f θ ) = n 1 i = 1 ∑ n L ( y i , f θ ( x i )) − n λ i = 1 ∑ n L ˉ ( a i , f θ ( x i )) ;
For the EO fairness criterion, first estimate the positive class prior π ^ = 1 n ∑ i = 1 n 1 { y i = 1 } \hat{\pi}=\frac{1}{n}\sum_{i=1}^{n}\mathbf{1}\{y_i=1\} π ^ = n 1 ∑ i = 1 n 1 { y i = 1 } , then construct the empirical risk as:
R ^ E O L , L ˉ ( f θ ) = 1 n ∑ i = 1 n L ( y i , f θ ( x i ) ) − λ π ^ n ∑ i : y i = 1 L ˉ ( a i , f θ ( x i ) ) . \widehat{\mathcal{R}}_{\mathrm{EO}}^{L,\bar{L}}(f_\theta)=\frac{1}{n}\sum_{i=1}^{n}L(y_i,f_\theta(x_i))-\frac{\lambda}{\hat{\pi}n}\sum_{i:y_i=1}\bar{L}(a_i,f_\theta(x_i)). R EO L , L ˉ ( f θ ) = n 1 i = 1 ∑ n L ( y i , f θ ( x i )) − π ^ n λ i : y i = 1 ∑ L ˉ ( a i , f θ ( x i )) .
Step 3 (Solve the unconstrained optimization problem): Solve the following problem using an optimization algorithm (e.g., L-BFGS):
θ ∗ = arg min θ R ^ f a i r L , L ˉ ( f θ ) + Ω ( θ ) , \theta^*=\arg\min_{\theta}\widehat{\mathcal{R}}_{\mathrm{fair}}^{L,\bar{L}}(f_\theta)+\Omega(\theta), θ ∗ = arg θ min R fair L , L ˉ ( f θ ) + Ω ( θ ) ,
where Ω ( θ ) \Omega(\theta) Ω ( θ ) is an optional regularization term (e.g., L 2 L_2 L 2 regularization).
Output: Decision function f ^ ∗ ( x ) = f θ ∗ ( x ) \hat{f}^*(x)=f_{\theta^*}(x) f ^ ∗ ( x ) = f θ ∗ ( x ) and classification rule Y ^ = sign ( f ^ ∗ ( x ) ) \hat{Y}=\operatorname{sign}(\hat{f}^*(x)) Y ^ = sign ( f ^ ∗ ( x )) .
Numerical Experiments
This section aims to validate the theoretical properties and the practical performance of Algorithm 1, through simulation experiments and a real-data study on credit card default prediction.
Simulation
Two distinct data-generating processes are constructed to correspond to the DP and EO fairness criteria, respectively. The first setup adopts the fairness-aware logistic loss under the attribute-blind setting for DP (§ 5.1 ), while the second setup employs the fairness-aware exponential loss under the attribute-aware setting for EO (Appendix § 8.6.3 ).
Fairness-Aware Logistic Loss
We generate binary class labels Y ∈ { − 1 , 1 } Y \in\{-1,1\} Y ∈ { − 1 , 1 } uniformly at random and assign a two-dimensional feature vector per label sampled from two Gaussian distributions:
X ∣ Y = 1 ∼ N ( ( 0 3 ) , ( 4 0 0 6 ) ) , X \mid Y = 1 \sim N\left(\begin{pmatrix}0\\3\end{pmatrix},\begin{pmatrix}4&0\\0&6\end{pmatrix}\right), X ∣ Y = 1 ∼ N ( ( 0 3 ) , ( 4 0 0 6 ) ) ,
X ∣ Y = − 1 ∼ N ( ( 0 − 3 ) , ( 4 0 0 6 ) ) . X \mid Y = -1 \sim N\left(\begin{pmatrix}0\\-3\end{pmatrix},\begin{pmatrix}4&0\\0&6\end{pmatrix}\right). X ∣ Y = − 1 ∼ N ( ( 0 − 3 ) , ( 4 0 0 6 ) ) .
Then, we restrict the population to the box B = [ − 6 , 6 ] 2 \mathcal{B} = [-6,6]^2 B = [ − 6 , 6 ] 2 by conditioning the Gaussians on X ∈ B X \in\mathcal{B} X ∈ B , using the conditional distribution over B \mathcal{B} B rather than truncated samples. This bounded support is required because the fairness surrogate loss L ˉ \bar{L} L ˉ is linear in the score, and unbounded support would make the conditional risk unbounded below, collapsing λ max \lambda_{\max} λ m a x in ((15) ) to zero. Over B \mathcal{B} B , the well-posedness threshold ((15) ) satisfies λ max = 1.1808 \lambda_{\max} = 1.1808 λ m a x = 1.1808 . The sensitive attribute is drawn from a Bernoulli distribution,
A ∣ X = x ∼ Binomial ( 1 , σ ( w ⊤ x + b ) ) , A \mid X = x \sim\operatorname{Binomial}(1,\sigma(w^\top x+b)), A ∣ X = x ∼ Binomial ( 1 , σ ( w ⊤ x + b )) ,
where σ ( u ) = 1 / ( 1 + e − u ) \sigma(u)=1/(1+e^{-u}) σ ( u ) = 1/ ( 1 + e − u ) denotes the logistic function and w = ( − 0.25 , 0.6 ) w=(-0.25,0.6) w = ( − 0.25 , 0.6 ) , b = 0.1 b=0.1 b = 0.1 . The resulting marginal probability is π A = P ( A = 1 ) = E [ σ ( w ⊤ X + b ) ] = 0.5139 \pi_A=\mathbb{P}(A=1)=\mathbb{E}[\sigma(w^\top X+b)]=0.5139 π A = P ( A = 1 ) = E [ σ ( w ⊤ X + b )] = 0.5139 . We set α = 0.5 \alpha=0.5 α = 0.5 , α ˉ = π A \bar{\alpha}=\pi_A α ˉ = π A and λ = 0.6 \lambda=0.6 λ = 0.6 .
Let ( L , L ‾ ) (L,\overline{L}) ( L , L ) be the logistic surrogate loss pair defined in Eq. (14) . We minimize the empirical ( L , L ‾ ) (L,\overline{L}) ( L , L ) -risk of Algorithm 1 for the DP criterion over the linear class
F = { f θ ( x ) = θ 0 + θ 1 η ( x ) + θ 2 η ‾ D P ( x ) : θ 0 , θ 1 , θ 2 ∈ R } . \mathcal{F} = \{f_{\theta}(x) = \theta_0 + \theta_1\eta(x) + \theta_2\overline{\eta}_{\mathrm{DP}}(x) : \theta_0,\theta_1,\theta_2 \in\mathbb{R}\}. F = { f θ ( x ) = θ 0 + θ 1 η ( x ) + θ 2 η DP ( x ) : θ 0 , θ 1 , θ 2 ∈ R } .
For f θ ∈ F f_{\theta} \in\mathcal{F} f θ ∈ F , the corresponding empirical ( L , L ‾ ) (L,\overline{L}) ( L , L ) -risk is
R ^ f a i r L , L ‾ ( f θ ) = 1 2 n ∑ i = 1 n log ( 1 + e − y i f θ ( x i ) ) + λ 2 n ∑ i = 1 n ( a i + 1 2 − π A ) f θ ( x i ) + ρ ∥ θ ∥ 2 2 . \widehat{\mathcal{R}}_{\mathrm{fair}}^{L,\overline{L}}(f_{\theta}) = \frac{1}{2n}\sum_{i=1}^{n}\log(1+e^{-y_i f_{\theta}(x_i)})+\frac{\lambda}{2n}\sum_{i=1}^{n}\left(\frac{a_i+1}{2}-\pi_A\right)f_{\theta}(x_i)+\rho\lVert\theta\rVert_2^2. R fair L , L ( f θ ) = 2 n 1 i = 1 ∑ n log ( 1 + e − y i f θ ( x i ) ) + 2 n λ i = 1 ∑ n ( 2 a i + 1 − π A ) f θ ( x i ) + ρ ∥ θ ∥ 2 2 .
We solve it with L-BFGS using a ridge penalty ρ = 10 − 4 \rho= 10^{-4} ρ = 1 0 − 4 .
Figure 1 reports one-at-a-time hyperparameter sweeps for the DP setting. We fix two parameters to the FAFC reference configuration ( α , α ‾ , λ ) = ( 1 / 2 , π A , 0.6 ) (\alpha,\overline{\alpha},\lambda)=(1/2,\pi_A,0.6) ( α , α , λ ) = ( 1/2 , π A , 0.6 ) and vary the third: accuracy cost parameter α \alpha α (left), fairness cost parameter α ‾ \overline{\alpha} α (middle), and trade-off parameter λ \lambda λ (right). Fairness-aware excess risk is evaluated at sample size n = 20000 n=20000 n = 20000 . The vertical axis uses a log scale, with error bars representing ± 2 \pm2 ± 2 standard errors over 20 independent replications. Dashed lines mark the FAFC reference values, and the dotted line in the right panel marks the well-posedness threshold λ max = 1.1808 \lambda_{\max}=1.1808 λ m a x = 1.1808 .
Figure 1. Sensitivity sweep over hyperparameters under the DP target.
Each curve attains its minimum precisely at the FAFC reference point, and excess risk grows monotonically as parameters deviate away from this optimum. This observation illustrates that satisfying all three FAFC conditions is necessary for low excess risk. For the λ \lambda λ sweep, excess risk stays small below λ max \lambda_{\max} λ m a x but rises steeply once the regularization strength exceeds this well-posedness bound.
Fairness-Aware Exponential Loss
We simulate synthetic data under the following data-generating process. Binary class labels Y ∈ { − 1 , 1 } Y \in\{-1,1\} Y ∈ { − 1 , 1 } are sampled uniformly at random. The binary sensitive attribute A ∈ { − 1 , 1 } A \in\{-1,1\} A ∈ { − 1 , 1 } is also drawn uniformly. Conditional on the sensitive-attribute value A = a A = a A = a , the feature vector X X X follows a zero-mean Gaussian distribution with group-dependent variance:
X ∣ A = a ∼ N ( 0 , σ a 2 ) . X \mid A = a \sim N(0,\sigma_a^2). X ∣ A = a ∼ N ( 0 , σ a 2 ) .
We set group-specific variances σ 1 = 2 \sigma_1 = 2 σ 1 = 2 and σ − 1 = 1 \sigma_{-1} = 1 σ − 1 = 1 . The difference in variances across groups creates inherent disparities in true-positive rates, making this setup suitable for studying equal-opportunity fairness.
The positive-class posterior is a logistic function of the feature,
P ( Y = 1 ∣ X = x , A = a ) = σ ( x ) , \mathbb{P}(Y = 1 \mid X = x, A = a) = \sigma(x), P ( Y = 1 ∣ X = x , A = a ) = σ ( x ) ,
and Y ∣ X , A ∼ Binomial ( 1 , σ ( X ) ) Y \mid X,A \sim\operatorname{Binomial}(1,\sigma(X)) Y ∣ X , A ∼ Binomial ( 1 , σ ( X )) . The design has unbounded support because the exponential accuracy loss has exponential tails on both sides (Appendix § 8.6.3 ), so the conditional risk is bounded below for every λ \lambda λ and no well-posedness threshold restricts the trade-off.
The classifier observes ( X , A ) (X,A) ( X , A ) , so the sensitive-attribute posterior collapses to η ˉ = 1 { a = 1 } \bar{\eta} = 1\{a = 1\} η ˉ = 1 { a = 1 } . We set α = 0.5 \alpha= 0.5 α = 0.5 , α ˉ = 0.5 \bar{\alpha} = 0.5 α ˉ = 0.5 and λ = 0.3 \lambda= 0.3 λ = 0.3 .
Let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be the exponential surrogate loss pair defined in Eq. (33) . We minimize the empirical ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk of Algorithm 1 for the EO criterion over the group-wise linear class
F = { f θ ( x , a ) = θ 0 , a + θ 1 , a η ( x , a ) : θ 0 , a , θ 1 , a ∈ R , a ∈ { − 1 , 1 } } . \mathcal{F} = \{f_\theta(x,a) = \theta_{0,a} + \theta_{1,a}\eta(x,a) : \theta_{0,a},\theta_{1,a} \in\mathbb{R},\ a \in\{-1,1\}\}. F = { f θ ( x , a ) = θ 0 , a + θ 1 , a η ( x , a ) : θ 0 , a , θ 1 , a ∈ R , a ∈ { − 1 , 1 }} .
For f θ ∈ F f_\theta\in\mathcal{F} f θ ∈ F , the corresponding empirical ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk is
R ^ f a i r L , L ˉ ( f θ ) = 1 2 n ∑ i = 1 n e − y i f θ ( x i , a i ) + λ 2 π ^ n ∑ i : y i = 1 a i f θ ( x i , a i ) + ρ ∥ θ ∥ 2 \widehat{\mathcal{R}}_{\mathrm{fair}}^{L,\bar{L}}(f_\theta) = \frac{1}{2n}\sum_{i=1}^{n} e^{-y_i f_\theta(x_i,a_i)} + \frac{\lambda}{2\hat{\pi}n}\sum_{i:y_i=1} a_i f_\theta(x_i,a_i) + \rho\lVert\theta\rVert^2 R fair L , L ˉ ( f θ ) = 2 n 1 i = 1 ∑ n e − y i f θ ( x i , a i ) + 2 π ^ n λ i : y i = 1 ∑ a i f θ ( x i , a i ) + ρ ∥ θ ∥ 2
where π ^ = 1 n ∑ i = 1 n 1 { y i = 1 } \hat{\pi} = \frac{1}{n}\sum_{i=1}^{n} \mathbb{1}\{y_i = 1\} π ^ = n 1 ∑ i = 1 n 1 { y i = 1 } . We solve it with L-BFGS using a ridge penalty ρ = 10 − 4 \rho= 10^{-4} ρ = 1 0 − 4 .
Figure 2 reports one-at-a-time hyperparameter sweeps for the EO setting, in the same format as Figure 1 . Two parameters are fixed at the FAFC reference configuration ( α , α ˉ , λ ) = ( 0.5 , 0.5 , 0.3 ) (\alpha,\bar{\alpha},\lambda) = (0.5, 0.5, 0.3) ( α , α ˉ , λ ) = ( 0.5 , 0.5 , 0.3 ) and the third is varied: accuracy cost α \alpha α (left), fairness cost α ˉ \bar{\alpha} α ˉ (middle), and trade-off λ \lambda λ (right), with the fairness-aware excess risk evaluated at n = 20000 n = 20000 n = 20000 (log vertical axis, ± 2 \pm2 ± 2 standard errors over 20 replications). Each curve attains its minimum at the FAFC reference value and increases monotonically away from it.
Figure 2. Sensitivity sweep over hyperparameters under the EO target.
A Real Case Study: Personal Credit Default Prediction
Identifying borrowers at risk of credit default is important for retail lenders to manage portfolio risk. Misclassifying a high-risk applicant who will default incurs substantially larger financial losses than rejecting a creditworthy low-risk borrower, which naturally introduces asymmetric misclassification costs in credit scoring [24 , 8 ] . Beyond cost imbalance, standard risk models trained solely for prediction accuracy often encode demographic disparities and yield unequal outcomes across protected groups. This makes consumer credit default prediction a setting where cost-sensitive classification and fairness regularization are both practically indispensable. Our proposed fairness-aware surrogate loss framework is particularly valuable here, as it unifies asymmetric error weighting and fairness regularization under a theoretically grounded classification procedure.
The Give Me Some Credit (GMC) dataset1 encompasses 150,000 anonymized consumer credit records; the separate official competition test set has no available ground-truth labels and is not used here. It includes features such as revolving utilization of unsecured credit lines, borrower age, counts of past-due episodes across 30–59, 60–89 and 90+ day delinquency tiers, debt-to-income ratio, monthly income, number of open credit accounts, number of real estate loans, and number of dependents. The objective is to predict whether a borrower will experience serious delinquency, defined as 90 or more days past due, within a two-year window.
After preprocessing, we retain all 150,000 records. The binary label Y = 1 Y = 1 Y = 1 indicates serious delinquency (90 or more days past due within two years) and Y = − 1 Y = -1 Y = − 1 otherwise, with P ( Y = 1 ) = 0.06684 \mathbb{P}(Y = 1) = 0.06684 P ( Y = 1 ) = 0.06684 empirically estimated. We select borrower age as the sensitive attribute A A A . Age is a formally protected ground for non-discrimination under the EU Charter of Fundamental Rights.2 We encode age as a binary variable A A A : A = − 1 A = -1 A = − 1 for applicants younger than 25 years and A = 1 A = 1 A = 1 otherwise, giving P ( A = 1 ) = 0.9862 \mathbb{P}(A = 1) = 0.9862 P ( A = 1 ) = 0.9862 . The 25-year threshold is chosen for two reasons. First, it follows the standard binarization proposed by Kamiran & Calders [13 ] and widely adopted in the fair credit scoring literature. Second, this cutoff aligns with the United Nations official statistical definition of youth (aged 15–24), which corresponds to the demographic group that typically has limited credit history and faces systematically differential treatment in lending decisions.
We adopt the EO criterion under the attribute-blind setting: age is only used to compute the fairness penalty during training and is excluded from the feature matrix at prediction time. We evaluate with five folds stratified jointly on the four ( Y , A ) (Y, A) ( Y , A ) cells, training on 120,000 records and testing on the remaining 30,000 in each fold.
We first train a cost-sensitive SVM. In credit-scoring applications, failing to detect a delinquent borrower generates substantially greater financial losses than denying credit to a low-risk applicant. We set α = 0.2 \alpha= 0.2 α = 0.2 to reflect this asymmetry: false negatives receive weight 1 − α = 0.8 1-\alpha= 0.8 1 − α = 0.8 and false positives weight α = 0.2 \alpha= 0.2 α = 0.2 , corresponding to a 4 : 1 4:1 4 : 1 cost ratio between false negatives and false positives. The trained classifier achieves an overall false negative rate P ( Y ^ = − 1 ∣ Y = 1 ) = 0.6358 \mathbb{P}(\widehat{Y}=-1 \mid Y=1)=0.6358 P ( Y = − 1 ∣ Y = 1 ) = 0.6358 . For comparison, the standard SVM ( α = 1 / 2 ) (\alpha= 1/2) ( α = 1/2 ) yields an even worse false negative rate of 1.0. Under this setup, the asymmetric cost weighting is a necessary prerequisite to obtain a non-degenerate classifier. Hence all subsequent experiments use α = 0.2 \alpha= 0.2 α = 0.2 .
Cost-sensitive classification alone, however, fails to align classification performance across demographic groups. Without fairness regularization, ∣ M D E O ∣ = 0.1273 |\mathrm{MDEO}| = 0.1273 ∣ MDEO ∣ = 0.1273 . This observation motivates incorporating our fairness regularizer.
Figure 3a plots the empirical cost-sensitive risk R ^ 0.2 \widehat{R}_{0.2} R 0.2 (blue, left axis) and equal-opportunity gap M D E O \mathrm{MDEO} MDEO (red, right axis) as functions of λ \lambda λ , with α = 0.2 \alpha= 0.2 α = 0.2 and α ˉ = 0.9784 \bar{\alpha} = 0.9784 α ˉ = 0.9784 held fixed. Here α ^ = P ^ ( A = 1 ∣ Y = 1 ) = 0.9784 \widehat{\alpha} = \widehat{\mathbb{P}}(A=1 \mid Y=1)=0.9784 α = P ( A = 1 ∣ Y = 1 ) = 0.9784 is estimated on the training folds of each split. All plotted values are aggregated from five-fold out-of-fold predictions, with error bars representing ± 2 \pm2 ± 2 standard errors estimated across the cross-validation folds.
The absolute equal-opportunity gap ∣ M D E O ∣ |\mathrm{MDEO}| ∣ MDEO ∣ exhibits a U-shaped trajectory as λ \lambda λ increases: it starts at 0.127 for the pure cost-sensitive SVM ( λ = 0 ) (\lambda= 0) ( λ = 0 ) , decreases to a minimum close to zero near λ = 1 \lambda= 1 λ = 1 , and then rises to 0.489 at λ = 2 \lambda= 2 λ = 2 . The cost-sensitive risk remains nearly constant when λ ∈ [ 0 , 1.1 ] \lambda\in[0,1.1] λ ∈ [ 0 , 1.1 ] . When λ > 1.1 \lambda> 1.1 λ > 1.1 , however, both curves rise together, and should not be read as a continuation of the trade-off. The fairness term is linear in the score, so its pull grows in proportion to λ \lambda λ , whereas the accuracy surrogate term is a hinge loss: its subgradient is bounded and vanishes once the margin is met, so it can resist the penalty only up to a fixed level. Once the penalty exceeds that local restoring force, the minimizer is no longer pinned by the accuracy term and the fitted score drifts in the direction the penalty favors, which here is the direction that raises the scores of the group coded A = − 1 A=-1 A = − 1 . Both the empirical risk and the disparity then increase with λ \lambda λ , and the pair leaves the range covered by (30) .
At λ = 0.9 \lambda=0.9 λ = 0.9 , the equal-opportunity gap is nearly eliminated: ∣ M D E O ∣ = 0.0015 |\mathrm{MD}_{\mathrm{EO}}|=0.0015 ∣ MD EO ∣ = 0.0015 , compared with 0.1273 for the cost-sensitive baseline. This is the optimal tradeoff parameter λ \lambda λ , where the cost-sensitive empirical risk R ^ 0.2 \widehat{R}_{0.2} R 0.2 remains essentially unchanged, moving only from 0.0401 to 0.0399. The vertical dashed line in Figure 3 a marks this point.
Figure 3. Cost-sensitive risk and equal-opportunity disparity across the λ \lambda λ sweep for the GMC credit dataset.
To verify that the above findings are not specific to the hinge-loss SVM, we repeat the analysis using a fairness-aware logistic regression classifier under the same equal-opportunity criterion, with α = 0.2 \alpha=0.2 α = 0.2 and α ˉ = 0.9784 \bar{\alpha}=0.9784 α ˉ = 0.9784 held fixed as in the SVM setup. Figure 3 b presents the resulting trade-off curves. The blue trace shows the empirical cost-sensitive risk R ^ 0.2 \widehat{R}_{0.2} R 0.2 on the left axis, and the red trace shows the equal-opportunity disparity M D E O \mathrm{MD}_{\mathrm{EO}} MD EO on the right axis. All points are computed from five-fold out-of-fold predictions, consistent with SVM.
The curve follows the same qualitative pattern observed for the SVM. For λ ∈ [ 0 , 0.4 ] \lambda\in[0,0.4] λ ∈ [ 0 , 0.4 ] , the empirical cost-sensitive risk remains nearly flat, while the EO gap narrows monotonically toward zero. When λ > 0.4 \lambda> 0.4 λ > 0.4 , however, the fairness penalty begins to dominate the objective, and the cost-sensitive risk rises noticeably with further increases in λ \lambda λ . The disparity crosses zero around λ = 0.4 \lambda= 0.4 λ = 0.4 . At this point we obtain ∣ M D E O ∣ = 0.0051 |\mathrm{MD}_{\mathrm{EO}}| = 0.0051 ∣ MD EO ∣ = 0.0051 , with the cost-sensitive risk still close to its baseline value. This confirms that equal-opportunity fairness can be achieved at negligible cost to the primary risk objective.
Conclusion
In this paper, we develop a statistical learning theory framework for surrogate-based fairness-aware classification. Within this framework, we introduce FAFC and the associated calibration inequality, covering both the DP and EO criteria under both the attribute-blind and attribute-aware settings.
First, we establish a necessary and sufficient condition for FAFC, stated purely in terms of the partial losses at the origin. Building on FAFC, we derive a calibration inequality showing that the excess fairness-aware risk is controlled by the excess ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk through an explicit calibration function. We then instantiate the general framework with concrete surrogate loss pairs and provide a general algorithm for fairness-aware classification.
Turning to empirical validation, our simulations corroborate the theoretical predictions: the excess fairness-aware risk is minimized at the FAFC reference configuration, grows monotonically as the parameters deviate from it, and rises sharply once λ \lambda λ exceeds the well-posedness threshold. A real-data study on the Give Me Some Credit credit default dataset further bears out the practical relevance of these results: the equal-opportunity disparity can be driven to nearly zero at negligible cost to the cost-sensitive risk.
Finally, we note a key limitation of the present theory: it presumes the fairness-regularized risk to be well posed. Beyond this threshold the conditional risk need not be bounded below, in which case no minimizer exists and neither the FAFC notion nor the calibration inequality applies.
Appendix
Choice of alpha: Regularization Form of MD and MR
The following lemma records the regularization form of the four standard fairness constraints and, in particular, the value of α ˉ \bar{\alpha} α ˉ induced by each constraint. It is due to Menon & Williamson [17 ] .
Lemma 21. (Regularization form of MD and MR [17 ] ). For any distribution of ( X , A , Y ) (X,A,Y) ( X , A , Y ) and any α , τ ∈ ( 0 , 1 ) \alpha,\tau\in(0,1) α , τ ∈ ( 0 , 1 ) , each of the four cases below specifies a constraint and a pair ( α ˉ , D ˉ ) (\bar{\alpha},\bar{D}) ( α ˉ , D ˉ ) such that the problem min f ∈ F R α ( f ; D ) \min_{f\in\mathcal{F}} \mathcal{R}_{\alpha}(f;D) min f ∈ F R α ( f ; D ) subject to that constraint is equivalent, for a suitable λ > 0 \lambda> 0 λ > 0 , to the penalized problem
min f ∈ F R α ( f ; D ) − λ R α ˉ ( f ; D ˉ ) . \min_{f\in\mathcal{F}}\mathcal{R}_{\alpha}(f;D)-\lambda\mathcal{R}_{\bar{\alpha}}(f;\bar{D}). f ∈ F min R α ( f ; D ) − λ R α ˉ ( f ; D ˉ ) . (i) ( D P , M D ) (DP, MD) ( D P , M D ) : MD D P ( f ) ≥ τ \operatorname{MD}_{DP}(f) \ge\tau MD D P ( f ) ≥ τ , with α ˉ = π A \bar{\alpha}=\pi_A α ˉ = π A and D ˉ = D ˉ D P \bar{D}=\bar{D}_{DP} D ˉ = D ˉ D P ;
(ii) ( D P , M R ) (DP, MR) ( D P , M R ) : MR D P ( f ) ≥ τ \operatorname{MR}_{DP}(f) \ge\tau MR D P ( f ) ≥ τ , with α ˉ = π A / [ π A + ( 1 − π A ) τ ] \bar{\alpha}=\pi_A/[\pi_A+(1-\pi_A)\tau] α ˉ = π A / [ π A + ( 1 − π A ) τ ] and D ˉ = D ˉ D P \bar{D}=\bar{D}_{DP} D ˉ = D ˉ D P ;
(iii) ( E O , M D ) (EO, MD) ( E O , M D ) : MD E O ( f ) ≥ τ \operatorname{MD}_{EO}(f) \ge\tau MD E O ( f ) ≥ τ , with α ˉ = π A ∣ Y = 1 \bar{\alpha}=\pi_{A\mid Y=1} α ˉ = π A ∣ Y = 1 and D ˉ = D ˉ E O \bar{D}=\bar{D}_{EO} D ˉ = D ˉ E O ;
(iv) ( E O , M R ) (EO, MR) ( E O , M R ) : MR E O ( f ) ≥ τ \operatorname{MR}_{EO}(f) \ge\tau MR E O ( f ) ≥ τ , with α ˉ = π A ∣ Y = 1 / [ π A ∣ Y = 1 + ( 1 − π A ∣ Y = 1 ) τ ] \bar{\alpha}=\pi_{A\mid Y=1}/[\pi_{A\mid Y=1}+(1-\pi_{A\mid Y=1})\tau] α ˉ = π A ∣ Y = 1 / [ π A ∣ Y = 1 + ( 1 − π A ∣ Y = 1 ) τ ] and D ˉ = D ˉ E O \bar{D}=\bar{D}_{EO} D ˉ = D ˉ E O .
Proof of Corollary 13
Proof. Let f ∗ f^* f ∗ be a minimizer. Since C f a i r L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x ) ) ≥ C f a i r L , L ˉ , ∗ ( η ( x ) , η ˉ ( x ) ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta(x),\bar{\eta}(x),f^*(x)) \ge C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(x),\bar{\eta}(x)) C fair L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x )) ≥ C fair L , L ˉ , ∗ ( η ( x ) , η ˉ ( x )) for every x x x , we have
E [ C f a i r L , L ˉ ( η ( X ) , η ˉ ( X ) , f ∗ ( X ) ) ] = R f a i r L , L ˉ ( f ∗ ; D , D ˉ ) = R f a i r L , L ˉ , ∗ = E [ C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] , \mathbb{E}\left[C_{\mathrm{fair}}^{L,\bar{L}}(\eta(X),\bar{\eta}(X),f^*(X))\right]=\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f^*;D,\bar{D})=\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}=\mathbb{E}\left[C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(X),\bar{\eta}(X))\right], E [ C fair L , L ˉ ( η ( X ) , η ˉ ( X ) , f ∗ ( X )) ] = R fair L , L ˉ ( f ∗ ; D , D ˉ ) = R fair L , L ˉ , ∗ = E [ C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X )) ] ,
so C f a i r L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x ) ) = C f a i r L , L ˉ , ∗ ( η ( x ) , η ˉ ( x ) ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta(x),\bar{\eta}(x),f^*(x))=C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(x),\bar{\eta}(x)) C fair L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x )) = C fair L , L ˉ , ∗ ( η ( x ) , η ˉ ( x )) for P X P_X P X -a.e. x x x .
Let x x x be such that ( η ( x ) , η ˉ ( x ) ) ∈ K (\eta(x),\bar{\eta}(x))\in\mathcal{K} ( η ( x ) , η ˉ ( x )) ∈ K , D α , α ˉ ( η ( x ) , η ˉ ( x ) ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x))\ne0 D α , α ˉ ( η ( x ) , η ˉ ( x )) = 0 , and the last equality holds. Suppose that f ∗ ( x ) f^*(x) f ∗ ( x ) has the wrong sign, i.e., f ∗ ( x ) D α , α ˉ ( η ( x ) , η ˉ ( x ) ) ≤ 0 f^*(x)D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x)) \le0 f ∗ ( x ) D α , α ˉ ( η ( x ) , η ˉ ( x )) ≤ 0 . Then
C f a i r L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x ) ) ≥ C f a i r L , L ˉ , − ( η ( x ) , η ˉ ( x ) ) = C f a i r L , L ˉ , ∗ ( η ( x ) , η ˉ ( x ) ) + H L , L ˉ α , α ˉ ( η ( x ) , η ˉ ( x ) ) > C f a i r L , L ˉ , ∗ ( η ( x ) , η ˉ ( x ) ) , C_{\mathrm{fair}}^{L,\bar{L}}(\eta(x),\bar{\eta}(x),f^*(x)) \ge C_{\mathrm{fair}}^{L,\bar{L},-}(\eta(x),\bar{\eta}(x))
= C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(x),\bar{\eta}(x)) + H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x))
> C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(x),\bar{\eta}(x)), C fair L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x )) ≥ C fair L , L ˉ , − ( η ( x ) , η ˉ ( x )) = C fair L , L ˉ , ∗ ( η ( x ) , η ˉ ( x )) + H L , L ˉ α , α ˉ ( η ( x ) , η ˉ ( x )) > C fair L , L ˉ , ∗ ( η ( x ) , η ˉ ( x )) ,
where the strict inequality uses FAFC on K \mathcal{K} K . This contradicts C f a i r L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x ) ) = C f a i r L , L ˉ , ∗ ( η ( x ) , η ˉ ( x ) ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta(x),\bar{\eta}(x),f^*(x)) = C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(x),\bar{\eta}(x)) C fair L , L ˉ ( η ( x ) , η ˉ ( x ) , f ∗ ( x )) = C fair L , L ˉ , ∗ ( η ( x ) , η ˉ ( x )) . Hence f ∗ ( x ) D α , α ˉ ( η ( x ) , η ˉ ( x ) ) > 0 f^*(x)D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x)) > 0 f ∗ ( x ) D α , α ˉ ( η ( x ) , η ˉ ( x )) > 0 for P X P_X P X -a.e. x x x with D α , α ˉ ( η ( x ) , η ˉ ( x ) ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta(x),\bar{\eta}(x)) \ne0 D α , α ˉ ( η ( x ) , η ˉ ( x )) = 0 . □
Proof of Theorem 14
Proof. Attribute-blind setting. For this setting, write
Φ ( η , η ˉ ) = ∂ ∂ t C f a i r L , L ˉ ( η , η ˉ , t ) ∣ t = 0 . \Phi(\eta,\bar{\eta}) = \left.\frac{\partial}{\partial t}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t)\right|_{t=0}. Φ ( η , η ˉ ) = ∂ t ∂ C fair L , L ˉ ( η , η ˉ , t ) t = 0 .
The quantities Φ D P \Phi_{\mathrm{DP}} Φ DP and Φ E O \Phi_{\mathrm{EO}} Φ EO below are the DP and EO versions of this derivative.
We use the following fact twice. Suppose that Φ \Phi Φ is continuous on K \mathcal{K} K , that D α , α ˉ ( α , α ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha}) = 0 D α , α ˉ ( α , α ˉ ) = 0 , and that ∂ η D α , α ˉ ( α , α ˉ ) ≠ 0 \partial_{\eta}D^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha}) \ne0 ∂ η D α , α ˉ ( α , α ˉ ) = 0 . Then Φ = 0 \Phi= 0 Φ = 0 on the part of the decision boundary { D α , α ˉ = 0 } \{D^{\alpha,\bar{\alpha}} = 0\} { D α , α ˉ = 0 } near ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) . Indeed, near ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) the boundary separates points with D α , α ˉ > 0 D^{\alpha,\bar{\alpha}} > 0 D α , α ˉ > 0 from points with D α , α ˉ < 0 D^{\alpha,\bar{\alpha}} < 0 D α , α ˉ < 0 ; at such points FAFC gives Φ < 0 \Phi< 0 Φ < 0 and Φ > 0 \Phi> 0 Φ > 0 , respectively, and continuity gives Φ = 0 \Phi= 0 Φ = 0 on the boundary. In particular, Φ ( α , α ˉ ) = 0 \Phi(\alpha,\bar{\alpha}) = 0 Φ ( α , α ˉ ) = 0 .
DP. Let
Φ D P ( η , η ˉ ) = ∂ ∂ t C D P L , L ˉ ( η , η ˉ , t ) ∣ t = 0 = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 ) ] − λ [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 ) ] = [ L − 1 ′ ( 0 ) − λ L ˉ − 1 ′ ( 0 ) ] + [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ] η + [ λ L ˉ 1 ′ ( 0 ) − λ L ˉ − 1 ′ ( 0 ) ] η ˉ . \begin{aligned}
\Phi_{\mathrm{DP}}(\eta,\bar{\eta}) = \left.\frac{\partial}{\partial t}C_{\mathrm{DP}}^{L,\bar{L}}(\eta,\bar{\eta},t)\right|_{t=0} \\
&= [\eta L'_1(0) + (1-\eta)L'_{-1}(0)] - \lambda[\bar{\eta}\bar{L}'_1(0) + (1-\bar{\eta})\bar{L}'_{-1}(0)] \\
&= [L'_{-1}(0) - \lambda\bar{L}'_{-1}(0)] + [L'_1(0) - L'_{-1}(0)]\eta+ [\lambda\bar{L}'_1(0) - \lambda\bar{L}'_{-1}(0)]\bar{\eta}.
\end{aligned} Φ DP ( η , η ˉ ) = ∂ t ∂ C DP L , L ˉ ( η , η ˉ , t ) t = 0 = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 )] − λ [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 )] = [ L − 1 ′ ( 0 ) − λ L ˉ − 1 ′ ( 0 )] + [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 )] η + [ λ L ˉ 1 ′ ( 0 ) − λ L ˉ − 1 ′ ( 0 )] η ˉ .
Since Φ D P ( η , η ˉ ) \Phi_{\mathrm{DP}}(\eta,\bar{\eta}) Φ DP ( η , η ˉ ) is affine to ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) , it is continuous on K \mathcal{K} K .
Since t ↦ C D P L , L ˉ ( η , η ˉ , t ) t \mapsto C_{\mathrm{DP}}^{L,\bar{L}}(\eta,\bar{\eta},t) t ↦ C DP L , L ˉ ( η , η ˉ , t ) is strictly convex on K \mathcal{K} K (with boundedness below guaranteed by the definition of K \mathcal{K} K ), Lemma 23 shows that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K if and only if
Φ D P ( η , η ˉ ) { < 0 if D D P α , α ˉ ( η , η ˉ ) > 0 , > 0 if D D P α , α ˉ ( η , η ˉ ) < 0 , = 0 if D D P α , α ˉ ( η , η ˉ ) = 0. \Phi_{\mathrm{DP}}(\eta,\bar{\eta}) \begin{cases} < 0 & \text{if } D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0, \\ > 0 & \text{if } D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) < 0, \\ = 0 & \text{if } D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0. \end{cases} Φ DP ( η , η ˉ ) ⎩ ⎨ ⎧ < 0 > 0 = 0 if D DP α , α ˉ ( η , η ˉ ) > 0 , if D DP α , α ˉ ( η , η ˉ ) < 0 , if D DP α , α ˉ ( η , η ˉ ) = 0.
Necessity ( ⟹ ) (\Longrightarrow) ( ⟹ ) . Assume L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for this fixed λ \lambda λ .
By the continuity fact (applied to D D P α , α ˉ D_{\mathrm{DP}}^{\alpha,\bar{\alpha}} D DP α , α ˉ , whose η \eta η -derivative is 1 1 1 ), Φ D P ( α , α ˉ ) = 0 \Phi_{\mathrm{DP}}(\alpha,\bar{\alpha}) = 0 Φ DP ( α , α ˉ ) = 0 .
Since
Φ D P ( α , α ˉ ) = [ α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) ] − λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] = 0 , \Phi_{\mathrm{DP}}(\alpha,\bar{\alpha}) = [\alpha L'_1(0) + (1-\alpha)L'_{-1}(0)] - \lambda[\bar{\alpha}\bar{L}'_1(0) + (1-\bar{\alpha})\bar{L}'_{-1}(0)] = 0, Φ DP ( α , α ˉ ) = [ α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 )] − λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 )] = 0 ,
we obtain
α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] . \alpha L'_1(0) + (1-\alpha)L'_{-1}(0) = \lambda[\bar{\alpha}\bar{L}'_1(0) + (1-\bar{\alpha})\bar{L}'_{-1}(0)]. α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 )] .
This is exactly condition (D3).
By using condition (D3), Φ D P \Phi_{\mathrm{DP}} Φ DP simplifies to
Φ D P ( η , η ˉ ) = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 ) ] − λ [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 ) ] − ( α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) − λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] ) = ( η − α ) ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) − λ ( η ˉ − α ˉ ) ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ) . \begin{aligned}
\Phi_{\mathrm{DP}}(\eta,\bar{\eta}) ={}& [\eta L'_1(0) + (1-\eta)L'_{-1}(0)] - \lambda[\bar{\eta}\bar{L}'_1(0) + (1-\bar{\eta})\bar{L}'_{-1}(0)] \\
&- \bigl(\alpha L'_1(0) + (1-\alpha)L'_{-1}(0) - \lambda[\bar{\alpha}\bar{L}'_1(0) + (1-\bar{\alpha})\bar{L}'_{-1}(0)]\bigr) \\
={}& (\eta-\alpha)(L'_1(0)-L'_{-1}(0)) - \lambda(\bar{\eta}-\bar{\alpha})(\bar{L}'_1(0)-\bar{L}'_{-1}(0)).
\end{aligned} Φ DP ( η , η ˉ ) = = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 )] − λ [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 )] − ( α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) − λ [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 )] ) ( η − α ) ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) − λ ( η ˉ − α ˉ ) ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 )) .
Consider any point on the decision boundary where D D P α , α ˉ ( η , η ˉ ) = 0 D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 D DP α , α ˉ ( η , η ˉ ) = 0 , i.e., η − α = λ ( η ˉ − α ˉ ) \eta-\alpha= \lambda(\bar{\eta}-\bar{\alpha}) η − α = λ ( η ˉ − α ˉ ) .
By the continuity fact, Φ D P = 0 \Phi_{\mathrm{DP}} = 0 Φ DP = 0 there as well, hence
λ ( η ˉ − α ˉ ) [ ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) − ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ) ] = 0. \lambda(\bar{\eta}-\bar{\alpha}) [(L'_1(0)-L'_{-1}(0)) - (\bar{L}'_1(0)-\bar{L}'_{-1}(0))] = 0. λ ( η ˉ − α ˉ ) [( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) − ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ))] = 0.
The decision boundary contains two distinct points with η ˉ ≠ α ˉ \bar{\eta} \ne\bar{\alpha} η ˉ = α ˉ (the line η = α + λ ( η ˉ − α ˉ ) \eta= \alpha+ \lambda(\bar{\eta}-\bar{\alpha}) η = α + λ ( η ˉ − α ˉ ) lies in K \mathcal{K} K for all η ˉ \bar{\eta} η ˉ in a neighborhood of α ˉ \bar{\alpha} α ˉ ), and the fixed value of λ \lambda λ is nonzero, so L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) L'_1(0)-L'_{-1}(0) = \bar{L}'_1(0)-\bar{L}'_{-1}(0) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) must hold. This is exactly condition (D2).
By condition (D2), we continue simplifying Φ D P \Phi_{\mathrm{DP}} Φ DP :
Φ D P ( η , η ˉ ) = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) ⋅ [ ( η − α ) − λ ( η ˉ − α ˉ ) ] = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) ⋅ D D P α , α ˉ ( η , η ˉ ) . \Phi_{\mathrm{DP}}(\eta,\bar{\eta}) = (L'_1(0) - L'_{-1}(0)) \cdot[(\eta-\alpha)-\lambda(\bar{\eta}-\bar{\alpha})] = (L'_1(0)-L'_{-1}(0))\cdot D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). Φ DP ( η , η ˉ ) = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) ⋅ [( η − α ) − λ ( η ˉ − α ˉ )] = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) ⋅ D DP α , α ˉ ( η , η ˉ ) .
Since FAFC on K \mathcal{K} K requires Φ D P \Phi_{\mathrm{DP}} Φ DP and D D P α , α ˉ D_{\mathrm{DP}}^{\alpha,\bar{\alpha}} D DP α , α ˉ to have strictly opposite signs, L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) L'_1(0)-L'_{-1}(0) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) must be strictly negative. Thus, L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) < 0 L'_1(0)-L'_{-1}(0)<0 L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) < 0 . This is exactly condition (D1).
Sufficiency (⟸ \Longleftarrow ⟸ ). Assume conditions (D1)–(D3) hold.
For the fixed λ \lambda λ , using condition (D3), we have
Φ D P ( η , η ˉ ) = ( η − α ) ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) − λ ( η ˉ − α ˉ ) ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ) . \Phi_{\mathrm{DP}}(\eta,\bar{\eta})=(\eta-\alpha)(L'_1(0)-L'_{-1}(0))-\lambda(\bar{\eta}-\bar{\alpha})(\bar{L}'_1(0)-\bar{L}'_{-1}(0)). Φ DP ( η , η ˉ ) = ( η − α ) ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) − λ ( η ˉ − α ˉ ) ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 )) .
Using condition (D2), we have
Φ D P ( η , η ˉ ) = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ) [ ( η − α ) − λ ( η ˉ − α ˉ ) ] \Phi_{\mathrm{DP}}(\eta,\bar{\eta})=(L'_1(0)-L'_{-1}(0))[(\eta-\alpha)-\lambda(\bar{\eta}-\bar{\alpha})] Φ DP ( η , η ˉ ) = ( L 1 ′ ( 0 ) − L − 1 ′ ( 0 )) [( η − α ) − λ ( η ˉ − α ˉ )]
Since condition (D1) ensures L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) < 0 L'_1(0)-L'_{-1}(0)<0 L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) < 0 , Φ D P ( η , η ˉ ) \Phi_{\mathrm{DP}}(\eta,\bar{\eta}) Φ DP ( η , η ˉ ) has precisely the opposite sign of D D P α , α ˉ ( η , η ˉ ) D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) D DP α , α ˉ ( η , η ˉ ) . Thus, L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K .
EO. Let r : = λ / π r := \lambda/\pi r := λ / π . Then
Φ E O ( η , η ˉ ) = ∂ ∂ t C E O L , L ˉ ( η , η ˉ , t ) ∣ t = 0 = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 ) ] − r η [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 ) ] , \Phi_{\mathrm{EO}}(\eta,\bar{\eta})=\left.\frac{\partial}{\partial t}C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},t)\right|_{t=0}
=[\eta L'_1(0)+(1-\eta)L'_{-1}(0)]-r\eta[\bar{\eta}\bar{L}'_1(0)+(1-\bar{\eta})\bar{L}'_{-1}(0)], Φ EO ( η , η ˉ ) = ∂ t ∂ C EO L , L ˉ ( η , η ˉ , t ) t = 0 = [ η L 1 ′ ( 0 ) + ( 1 − η ) L − 1 ′ ( 0 )] − r η [ η ˉ L ˉ 1 ′ ( 0 ) + ( 1 − η ˉ ) L ˉ − 1 ′ ( 0 )] ,
and D E O α , α ˉ ( η , η ˉ ) = ( η − α ) − r η ( η ˉ − α ˉ ) D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=(\eta-\alpha)-r\eta(\bar{\eta}-\bar{\alpha}) D EO α , α ˉ ( η , η ˉ ) = ( η − α ) − r η ( η ˉ − α ˉ ) .
Since t ↦ C E O L , L ˉ ( η , η ˉ , t ) t\mapsto C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},t) t ↦ C EO L , L ˉ ( η , η ˉ , t ) is strictly convex on K \mathcal{K} K (with boundedness below guaranteed by the definition of K \mathcal{K} K ), Lemma 23 shows that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K if and only if
Φ E O ( η , η ˉ ) { < 0 if D E O α , α ˉ ( η , η ˉ ) > 0 , > 0 if D E O α , α ˉ ( η , η ˉ ) < 0 , = 0 if D E O α , α ˉ ( η , η ˉ ) = 0. \Phi_{\mathrm{EO}}(\eta,\bar{\eta})\begin{cases}
<0 & \text{if }D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})>0,\\
>0 & \text{if }D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})<0,\\
=0 & \text{if }D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0.
\end{cases} Φ EO ( η , η ˉ ) ⎩ ⎨ ⎧ < 0 > 0 = 0 if D EO α , α ˉ ( η , η ˉ ) > 0 , if D EO α , α ˉ ( η , η ˉ ) < 0 , if D EO α , α ˉ ( η , η ˉ ) = 0.
Necessity (⇒ \Rightarrow ⇒ ). Assume L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for this fixed λ \lambda λ .
By the continuity fact (applied to D E O α , α ˉ D_{\mathrm{EO}}^{\alpha,\bar{\alpha}} D EO α , α ˉ , whose η \eta η -derivative at ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) is 1), Φ E O ( α , α ˉ ) = 0 \Phi_{\mathrm{EO}}(\alpha,\bar{\alpha})=0 Φ EO ( α , α ˉ ) = 0 . That is,
α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = r α [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] . \alpha L'_1(0)+(1-\alpha)L'_{-1}(0)=r\alpha\left[\bar{\alpha}\bar{L}'_1(0)+(1-\bar{\alpha})\bar{L}'_{-1}(0)\right]. α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = r α [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] .
which is (E3).
For general ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) , subtracting Φ E O ( α , α ˉ ) \Phi_{\mathrm{EO}}(\alpha,\bar{\alpha}) Φ EO ( α , α ˉ ) gives
Φ E O ( η , η ˉ ) − Φ E O ( α , α ˉ ) = ( η − α ) [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ] − r ( η − α ) [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] − r η ( η ˉ − α ˉ ) [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] . \begin{aligned}
\Phi_{\mathrm{EO}}(\eta,\bar{\eta})-\Phi_{\mathrm{EO}}(\alpha,\bar{\alpha})
={}&(\eta-\alpha)\left[L'_1(0)-L'_{-1}(0)\right]-r(\eta-\alpha)\left[\bar{\alpha}\bar{L}'_1(0)+(1-\bar{\alpha})\bar{L}'_{-1}(0)\right]\\
&-r\eta(\bar{\eta}-\bar{\alpha})\left[\bar{L}'_1(0)-\bar{L}'_{-1}(0)\right].
\end{aligned} Φ EO ( η , η ˉ ) − Φ EO ( α , α ˉ ) = ( η − α ) [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) ] − r ( η − α ) [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] − r η ( η ˉ − α ˉ ) [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] .
On the decision boundary D E O α , α ˉ ( η , η ˉ ) = 0 D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0 D EO α , α ˉ ( η , η ˉ ) = 0 , i.e., η − α = r η ( η ˉ − α ˉ ) \eta-\alpha=r\eta(\bar{\eta}-\bar{\alpha}) η − α = r η ( η ˉ − α ˉ ) , the continuity fact gives Φ E O = 0 \Phi_{\mathrm{EO}}=0 Φ EO = 0 . Substituting this into the last display yields
r η ( η ˉ − α ˉ ) [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) − r ( α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ) − ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ) ] = 0. r\eta(\bar{\eta}-\bar{\alpha})\left[L'_1(0)-L'_{-1}(0)-r\left(\bar{\alpha}\bar{L}'_1(0)+(1-\bar{\alpha})\bar{L}'_{-1}(0)\right)-\left(\bar{L}'_1(0)-\bar{L}'_{-1}(0)\right)\right]=0. r η ( η ˉ − α ˉ ) [ L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) − r ( α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ) − ( L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ) ] = 0.
The boundary contains points with η ˉ ≠ α ˉ \bar{\eta}\ne\bar{\alpha} η ˉ = α ˉ arbitrarily close to ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) (the curve η = α / [ 1 − r ( η ˉ − α ˉ ) ] \eta=\alpha/[1-r(\bar{\eta}-\bar{\alpha})] η = α / [ 1 − r ( η ˉ − α ˉ )] lies in K \mathcal{K} K for all η ˉ \bar{\eta} η ˉ in a neighborhood of α ˉ \bar{\alpha} α ˉ ), and r η > 0 r\eta>0 r η > 0 there, so
L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) + r [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] . L'_1(0)-L'_{-1}(0)=\bar{L}'_1(0)-\bar{L}'_{-1}(0)+r\left[\bar{\alpha}\bar{L}'_1(0)+(1-\bar{\alpha})\bar{L}'_{-1}(0)\right]. L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) + r [ α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) ] .
This is (E2).
Using (E2) and Φ E O ( α , α ˉ ) = 0 \Phi_{\mathrm{EO}}(\alpha,\bar{\alpha})=0 Φ EO ( α , α ˉ ) = 0 , we obtain
Φ E O ( η , η ˉ ) = [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] [ ( η − α ) − r η ( η ˉ − α ˉ ) ] = [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] D E O α , α ˉ ( η , η ˉ ) . \Phi_{\mathrm{EO}}(\eta,\bar{\eta})=\left[\bar{L}'_1(0)-\bar{L}'_{-1}(0)\right]\left[(\eta-\alpha)-r\eta(\bar{\eta}-\bar{\alpha})\right]=\left[\bar{L}'_1(0)-\bar{L}'_{-1}(0)\right]D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). Φ EO ( η , η ˉ ) = [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] [ ( η − α ) − r η ( η ˉ − α ˉ ) ] = [ L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) ] D EO α , α ˉ ( η , η ˉ ) .
Since FAFC on K \mathcal{K} K requires Φ E O \Phi_{\mathrm{EO}} Φ EO and D E O α , α ˉ D_{\mathrm{EO}}^{\alpha,\bar{\alpha}} D EO α , α ˉ to have strictly opposite signs,
L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) < 0. \bar{L}'_1(0)-\bar{L}'_{-1}(0)<0. L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) < 0.
This is (E1).
Sufficiency (⇐). Assume (E1)–(E3) hold. By (E2) and (E3), the last display holds for every ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) . Since (E1) gives L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) < 0 \bar{L}'_{1}(0)-\bar{L}'_{-1}(0)<0 L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) < 0 , Φ E O \Phi_{\mathrm{EO}} Φ EO has precisely the opposite sign of D E O α , α ˉ D^{\alpha,\bar{\alpha}}_{\mathrm{EO}} D EO α , α ˉ , and Lemma 23 yields that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K .
Attribute-aware setting. The sufficiency direction is the same as in the attribute-blind case: it is pointwise and needs no additional assumption. Only the necessity direction requires the condition in Remark 15. That condition is used below to ensure that each boundary point η a \eta_a η a has points of K a \mathcal{K}_a K a on both sides, so that FAFC and continuity give Φ ( η a , a ) = 0 \Phi(\eta_a,a)=0 Φ ( η a , a ) = 0 . Assume this condition. Write
Φ ( η , a ) = ∂ L f a i r L , L ˉ ( η , a , t ) ∂ t ∣ t = 0 . \Phi(\eta,a)=\left.\frac{\partial L_{\mathrm{fair}}^{L,\bar{L}}(\eta,a,t)}{\partial t}\right|_{t=0}. Φ ( η , a ) = ∂ t ∂ L fair L , L ˉ ( η , a , t ) t = 0 .
On the edge a = 1 a=1 a = 1 , let η + \eta_+ η + be the point where D ( η , 1 ) = 0 D(\eta,1)=0 D ( η , 1 ) = 0 ; on a = − 1 a=-1 a = − 1 , let η − \eta_- η − be the corresponding point. Since D ( ⋅ , a ) D(\cdot,a) D ( ⋅ , a ) changes sign at these points, FAFC and continuity give
Φ ( η + , 1 ) = 0 , Φ ( η − , − 1 ) = 0. \Phi(\eta_+,1)=0,\qquad\Phi(\eta_-,-1)=0. Φ ( η + , 1 ) = 0 , Φ ( η − , − 1 ) = 0.
For DP, D ( η , 1 ) = 0 D(\eta,1)=0 D ( η , 1 ) = 0 gives η + = α + λ ( 1 − α ˉ ) \eta_+=\alpha+\lambda(1-\bar{\alpha}) η + = α + λ ( 1 − α ˉ ) , and D ( η , − 1 ) = 0 D(\eta,-1)=0 D ( η , − 1 ) = 0 gives η − = α − λ α ˉ \eta_-=\alpha-\lambda\bar{\alpha} η − = α − λ α ˉ . Hence
0 = η + L 1 ′ + ( 1 − η + ) L − 1 ′ − λ L ˉ 1 ′ , 0=\eta_+L'_1+(1-\eta_+)L'_{-1}-\lambda\bar{L}'_1, 0 = η + L 1 ′ + ( 1 − η + ) L − 1 ′ − λ L ˉ 1 ′ ,
0 = η − L 1 ′ + ( 1 − η − ) L − 1 ′ − λ L ˉ − 1 ′ . 0=\eta_-L'_1+(1-\eta_-)L'_{-1}-\lambda\bar{L}'_{-1}. 0 = η − L 1 ′ + ( 1 − η − ) L − 1 ′ − λ L ˉ − 1 ′ .
Subtracting yields L 1 ′ − L − 1 ′ = L ˉ 1 ′ − L ˉ − 1 ′ L'_1-L'_{-1}=\bar{L}'_1-\bar{L}'_{-1} L 1 ′ − L − 1 ′ = L ˉ 1 ′ − L ˉ − 1 ′ , which is (D2). Substituting back yields (D3).
After (D2) and (D3), Φ \Phi Φ is a constant multiple of D D D ; FAFC forces the constant to be negative, which is (D1).
For EO, D ( η , a ) = η − α − r η ( 1 { a = 1 } − α ˉ ) D(\eta,a)=\eta-\alpha-r\eta(1\{a=1\}-\bar{\alpha}) D ( η , a ) = η − α − r η ( 1 { a = 1 } − α ˉ ) , so
η + − α = r η + ( 1 − α ˉ ) , η − − α = − r η − α ˉ . \eta_+-\alpha=r\eta_+(1-\bar{\alpha}),\qquad\eta_--\alpha=-r\eta_-\bar{\alpha}. η + − α = r η + ( 1 − α ˉ ) , η − − α = − r η − α ˉ .
Equivalently,
0 = η + L 1 ′ + ( 1 − η + ) L − 1 ′ − r η + L ˉ 1 ′ , 0=\eta_+L'_1+(1-\eta_+)L'_{-1}-r\eta_+\bar{L}'_1, 0 = η + L 1 ′ + ( 1 − η + ) L − 1 ′ − r η + L ˉ 1 ′ ,
0 = η − L 1 ′ + ( 1 − η − ) L − 1 ′ − r η − L ‾ − 1 ′ . 0 = \eta_{-}L'_{1} + (1-\eta_{-})L'_{-1} - r\eta_{-}\overline{L}'_{-1}. 0 = η − L 1 ′ + ( 1 − η − ) L − 1 ′ − r η − L − 1 ′ .
Combining these two equations with the two relations for η ± \eta_{\pm} η ± first gives (E2); substituting (E2) back gives (E3). The sign of the resulting constant multiple of D D D gives (E1). This proves the edgewise analogue in the attribute-aware setting. □
Proof of Theorem 17
Proof. By Lemma 22 , the excess fairness-aware risk is given by
R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ = E X [ 1 { sign ( f ( X ) ) ≠ sign ( D α , α ˉ ( X ) ) } ⋅ ∣ D α , α ˉ ( X ) ∣ ] ≤ E X [ 1 { f ( X ) ⋅ D α , α ˉ ( X ) ≤ 0 } ⋅ ∣ D α , α ˉ ( X ) ∣ ] . \begin{aligned}
\mathcal{R}^{\alpha,\bar{\alpha}}_{\mathrm{fair}}(f)-\mathcal{R}^{\alpha,\bar{\alpha},*}_{\mathrm{fair}} = \mathbb{E}_{X}\left[1\left\{\operatorname{sign}(f(X)) \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(X))\right\}\cdot\left|D^{\alpha,\bar{\alpha}}(X)\right|\right] \\
&\le\mathbb{E}_{X}\left[1\left\{f(X)\cdot D^{\alpha,\bar{\alpha}}(X)\le0\right\}\cdot\left|D^{\alpha,\bar{\alpha}}(X)\right|\right].
\end{aligned} R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ = E X [ 1 { sign ( f ( X )) = sign ( D α , α ˉ ( X )) } ⋅ D α , α ˉ ( X ) ] ≤ E X [ 1 { f ( X ) ⋅ D α , α ˉ ( X ) ≤ 0 } ⋅ D α , α ˉ ( X ) ] .
The event { sign ( f ( X ) ) ≠ sign ( D α , α ˉ ( X ) ) } \{\operatorname{sign}(f(X)) \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(X))\} { sign ( f ( X )) = sign ( D α , α ˉ ( X ))} is contained in { f ( X ) D α , α ˉ ( X ) ≤ 0 } \{f(X)D^{\alpha,\bar{\alpha}}(X)\le0\} { f ( X ) D α , α ˉ ( X ) ≤ 0 } .
Since ψ L , L ˉ α , α ˉ \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}} ψ L , L ˉ α , α ˉ is convex, ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 , and ψ L , L ˉ α , α ˉ ≥ 0 \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\ge0 ψ L , L ˉ α , α ˉ ≥ 0 (Lemma 24 ), it is nondecreasing on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) . We apply Jensen’s inequality.
ψ L , L ˉ α , α ˉ ( R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ ) ≤ ψ L , L ˉ α , α ˉ ( E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ∣ D α , α ˉ ( X ) ∣ ] ) ≤ E X [ ψ L , L ˉ α , α ˉ ( 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ∣ D α , α ˉ ( X ) ∣ ) ] . \begin{aligned}
\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(\mathcal{R}^{\alpha,\bar{\alpha}}_{\mathrm{fair}}(f)-\mathcal{R}^{\alpha,\bar{\alpha},*}_{\mathrm{fair}}\right)
\le\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(\mathbb{E}_{X}\left[1\left\{f(X)D^{\alpha,\bar{\alpha}}(X)\le0\right\}\left|D^{\alpha,\bar{\alpha}}(X)\right|\right]\right) \\
&\le\mathbb{E}_{X}\left[\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(1\left\{f(X)D^{\alpha,\bar{\alpha}}(X)\le0\right\}\left|D^{\alpha,\bar{\alpha}}(X)\right|\right)\right].
\end{aligned} ψ L , L ˉ α , α ˉ ( R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ ) ≤ ψ L , L ˉ α , α ˉ ( E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } D α , α ˉ ( X ) ] ) ≤ E X [ ψ L , L ˉ α , α ˉ ( 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } D α , α ˉ ( X ) ) ] .
Since ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 , we have ψ L , L ˉ α , α ˉ ( I E ⋅ z ) = I E ⋅ ψ L , L ˉ α , α ˉ ( z ) \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}(I_E\cdot z)=I_E\cdot\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}(z) ψ L , L ˉ α , α ˉ ( I E ⋅ z ) = I E ⋅ ψ L , L ˉ α , α ˉ ( z ) for any event E E E . Furthermore, by the definition of the biconjugate, ψ L , L ˉ α , α ˉ ≤ ν L , L ˉ α , α ˉ \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\le\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} ψ L , L ˉ α , α ˉ ≤ ν L , L ˉ α , α ˉ . Thus,
E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ψ L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( X ) ∣ ) ] ≤ E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ν L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( X ) ∣ ) ] . \mathbb{E}_{X}\left[1\left\{f(X)D^{\alpha,\bar{\alpha}}(X)\le0\right\}\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(\left|D^{\alpha,\bar{\alpha}}(X)\right|\right)\right]
\le\mathbb{E}_{X}\left[1\left\{f(X)D^{\alpha,\bar{\alpha}}(X)\le0\right\}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(\left|D^{\alpha,\bar{\alpha}}(X)\right|\right)\right]. E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ψ L , L ˉ α , α ˉ ( D α , α ˉ ( X ) ) ] ≤ E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ν L , L ˉ α , α ˉ ( D α , α ˉ ( X ) ) ] .
Recall ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ K { H L , L ˉ α , α ˉ ( η , η ˉ ) : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ } \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon)=\inf_{(\eta,\bar{\eta})\in\mathcal{K}}\left\{H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta,\bar{\eta}):\left|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\right|=\epsilon\right\} ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ K { H L , L ˉ α , α ˉ ( η , η ˉ ) : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ } . It follows that ν L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( X ) ∣ ) ≤ H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X ) ) \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}\left(\left|D^{\alpha,\bar{\alpha}}(X)\right|\right)\le H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta(X),\bar{\eta}(X)) ν L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( X ) ∣ ) ≤ H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X )) . Substituting this back, we have
R H S ≤ E X [ 1 { f ( X ) D ( X ) ≤ 0 } H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X ) ) ] = E X [ 1 { f ( X ) D ( X ) ≤ 0 } ( C f a i r L , L ˉ , − ( η ( X ) , η ˉ ( X ) ) − C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ) ] . \begin{aligned}
\mathrm{RHS}\le\mathbb{E}_{X}\left[1\left\{f(X)D(X)\le0\right\}H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta(X),\bar{\eta}(X))\right] \\
&=\mathbb{E}_{X}\left[1\left\{f(X)D(X)\le0\right\}\left(C^{L,\bar{L},-}_{\mathrm{fair}}(\eta(X),\bar{\eta}(X))-C^{L,\bar{L},*}_{\mathrm{fair}}(\eta(X),\bar{\eta}(X))\right)\right].
\end{aligned} RHS ≤ E X [ 1 { f ( X ) D ( X ) ≤ 0 } H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X )) ] = E X [ 1 { f ( X ) D ( X ) ≤ 0 } ( C fair L , L ˉ , − ( η ( X ) , η ˉ ( X )) − C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X )) ) ] .
By the definition of C f a i r L , L ˉ , − ( η , η ˉ ) = inf t : D ( η , η ˉ ) ≤ 0 C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L},-}(\eta,\bar{\eta})=\inf_{t:D(\eta,\bar{\eta})\leq0}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ , − ( η , η ˉ ) = inf t : D ( η , η ˉ ) ≤ 0 C fair L , L ˉ ( η , η ˉ , t ) , if f ( X ) f(X) f ( X ) satisfies the condition f ( X ) D ( X ) ≤ 0 f(X)D(X)\leq0 f ( X ) D ( X ) ≤ 0 , then C f a i r L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) ≥ C f a i r L , L ˉ , − ( η ( X ) , η ˉ ( X ) ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta(X),\bar{\eta}(X),f(X))\geq C_{\mathrm{fair}}^{L,\bar{L},-}(\eta(X),\bar{\eta}(X)) C fair L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X )) ≥ C fair L , L ˉ , − ( η ( X ) , η ˉ ( X )) . Therefore,
1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ( C f a i r L , L ˉ , − ( η ( X ) , η ˉ ( X ) ) − C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ) ≤ C f a i r L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) − C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) . \mathbf{1}\{f(X)D^{\alpha,\bar{\alpha}}(X)\leq0\}\left(C_{\mathrm{fair}}^{L,\bar{L},-}(\eta(X),\bar{\eta}(X))-C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(X),\bar{\eta}(X))\right)
\leq C_{\mathrm{fair}}^{L,\bar{L}}(\eta(X),\bar{\eta}(X),f(X))-C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(X),\bar{\eta}(X)). 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } ( C fair L , L ˉ , − ( η ( X ) , η ˉ ( X )) − C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X )) ) ≤ C fair L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X )) − C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X )) .
Note that if the indicator is 0, the LHS is 0, and the RHS is non-negative by the definition of the optimal risk, so the inequality holds generally. Taking the expectation over X X X :
E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X ) ) ] ≤ E X [ C f a i r L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) − C f a i r L , L ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] = R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ . \mathbb{E}_{X}\left[\mathbf{1}\{f(X)D^{\alpha,\bar{\alpha}}(X)\leq0\}H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))\right]
\leq\mathbb{E}_{X}\left[C_{\mathrm{fair}}^{L,\bar{L}}(\eta(X),\bar{\eta}(X),f(X))-C_{\mathrm{fair}}^{L,\bar{L},*}(\eta(X),\bar{\eta}(X))\right]
=\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}. E X [ 1 { f ( X ) D α , α ˉ ( X ) ≤ 0 } H L , L ˉ α , α ˉ ( η ( X ) , η ˉ ( X )) ] ≤ E X [ C fair L , L ˉ ( η ( X ) , η ˉ ( X ) , f ( X )) − C fair L , L ˉ , ∗ ( η ( X ) , η ˉ ( X )) ] = R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ .
This completes the proof of the first part.
Now consider the second part. Suppose first that ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is invertible on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] . By Part 1 of Lemma 24 , ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 . Since an invertible function has at most one zero and ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)\geq0 ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 0 on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] , it follows that
ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B max ] . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)>0\qquad\text{for every }\epsilon\in(0,B_{\max}]. ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B m a x ] .
Now let ( η , η ˉ ) ∈ K (\eta,\bar{\eta})\in\mathcal{K} ( η , η ˉ ) ∈ K satisfy D α , α ˉ ( η , η ˉ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\ne0 D α , α ˉ ( η , η ˉ ) = 0 . Then ∣ D α , α ˉ ( η , η ˉ ) ∣ ∈ ( 0 , B max ] \lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert\in(0,B_{\max}] ∣ D α , α ˉ ( η , η ˉ )∣ ∈ ( 0 , B m a x ] . By the definition of ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ ,
H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ ν L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( η , η ˉ ) ∣ ) ≥ ψ L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( η , η ˉ ) ∣ ) > 0. H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\geq\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}\left(\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert\right)\geq\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}\left(\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert\right)>0. H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ ν L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( η , η ˉ )∣ ) ≥ ψ L , L ˉ α , α ˉ ( ∣ D α , α ˉ ( η , η ˉ )∣ ) > 0.
Hence L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K .
Conversely, assume that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K . Then ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 by Part 1 of Lemma 24 , ψ L , L ˉ α , α ˉ ( ϵ ) > 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)>0 ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B max ] \epsilon\in(0,B_{\max}] ϵ ∈ ( 0 , B m a x ] by Part 2, and ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is continuous on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] by Part 3. Let 0 ≤ a < b ≤ B max 0\leq a<b\leq B_{\max} 0 ≤ a < b ≤ B m a x . Convexity together with ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 gives
ψ L , L ˉ α , α ˉ ( a ) ≤ a b ψ L , L ˉ α , α ˉ ( b ) + ( 1 − a b ) ψ L , L ˉ α , α ˉ ( 0 ) = a b ψ L , L ˉ α , α ˉ ( b ) . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(a)\leq\frac{a}{b}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(b)+\left(1-\frac{a}{b}\right)\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=\frac{a}{b}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(b). ψ L , L ˉ α , α ˉ ( a ) ≤ b a ψ L , L ˉ α , α ˉ ( b ) + ( 1 − b a ) ψ L , L ˉ α , α ˉ ( 0 ) = b a ψ L , L ˉ α , α ˉ ( b ) .
If a = 0 a=0 a = 0 , then ψ L , L ˉ α , α ˉ ( 0 ) = 0 < ψ L , L ˉ α , α ˉ ( b ) \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0<\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(b) ψ L , L ˉ α , α ˉ ( 0 ) = 0 < ψ L , L ˉ α , α ˉ ( b ) ; if a > 0 a>0 a > 0 , then ψ L , L ˉ α , α ˉ ( a ) > 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(a)>0 ψ L , L ˉ α , α ˉ ( a ) > 0 , so ψ L , L ˉ α , α ˉ ( b ) ≥ b a ψ L , L ˉ α , α ˉ ( a ) > ψ L , L ˉ α , α ˉ ( a ) \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(b)\ge\frac{b}{a}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(a)>\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(a) ψ L , L ˉ α , α ˉ ( b ) ≥ a b ψ L , L ˉ α , α ˉ ( a ) > ψ L , L ˉ α , α ˉ ( a ) . Hence ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is strictly increasing on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] . Since it is continuous as well, it is invertible on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] .
Useful Lemmas
Unified Excess-Risk Lemma
The following lemma characterizes the optimal decision boundary under the fairness-aware risk. It establishes that the excess risk is determined by the alignment between the prediction’s sign and fairness-aware discriminant D α , α ˉ ( η , η ˉ ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) D α , α ˉ ( η , η ˉ ) .
Lemma 22. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha}\in(0,1) α , α ˉ ∈ ( 0 , 1 ) . For both the attribute-blind and attribute-aware settings, the conditional excess risk satisfies
C f a i r α , α ˉ ( η , η ˉ , t ) − C f a i r α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) ≠ sign ( D α , α ˉ ( η , η ˉ ) ) } ⋅ ∣ D α , α ˉ ( η , η ˉ ) ∣ . C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t)-C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta})=\mathbf{1}\{\operatorname{sign}(t)\ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}))\}\cdot\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert. C fair α , α ˉ ( η , η ˉ , t ) − C fair α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) = sign ( D α , α ˉ ( η , η ˉ ))} ⋅ ∣ D α , α ˉ ( η , η ˉ )∣ . Consequently, for any f ∈ F f\in\mathcal{F} f ∈ F ,
R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ = E X [ 1 { sign ( f ( X ) ) ≠ sign ( D α , α ˉ ( η ( X ) , η ˉ ( X ) ) ) } ⋅ ∣ D α , α ˉ ( η ( X ) , η ˉ ( X ) ) ∣ ] . \mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(f)-\mathcal{R}_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}=\mathbb{E}_{X}\left[\mathbf{1}\{\operatorname{sign}(f(X))\ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X)))\}\cdot\lvert D^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))\rvert\right]. R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ = E X [ 1 { sign ( f ( X )) = sign ( D α , α ˉ ( η ( X ) , η ˉ ( X )))} ⋅ ∣ D α , α ˉ ( η ( X ) , η ˉ ( X ))∣ ] .
Proof. Let x ∈ X x\in\mathcal{X} x ∈ X be fixed. Recall that the fairness-aware conditional risk is given by
C f a i r α , α ˉ ( η , η ˉ , t ) = A α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } + B α , α ˉ ( η , η ˉ ) 1 { t > 0 } . C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t)=A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\mathbf{1}\{t\le0\}+B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\mathbf{1}\{t>0\}. C fair α , α ˉ ( η , η ˉ , t ) = A α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } + B α , α ˉ ( η , η ˉ ) 1 { t > 0 } .
Using the identity 1 { t > 0 } = 1 − 1 { t ≤ 0 } \mathbf{1}\{t>0\}=1-\mathbf{1}\{t\le0\} 1 { t > 0 } = 1 − 1 { t ≤ 0 } , we can rewrite the conditional risk as
C f a i r α , α ˉ ( η , η ˉ , t ) = A α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } + B α , α ˉ ( η , η ˉ ) ( 1 − 1 { t ≤ 0 } ) = B α , α ˉ ( η , η ˉ ) + ( A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) ) 1 { t ≤ 0 } . \begin{aligned}
C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t)
&=A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\mathbf{1}\{t\le0\}+B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\left(1-\mathbf{1}\{t\le0\}\right)\\
&=B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})+\left(A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})-B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\right)\mathbf{1}\{t\le0\}.
\end{aligned} C fair α , α ˉ ( η , η ˉ , t ) = A α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } + B α , α ˉ ( η , η ˉ ) ( 1 − 1 { t ≤ 0 } ) = B α , α ˉ ( η , η ˉ ) + ( A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) ) 1 { t ≤ 0 } .
Recall the fairness-aware discriminant D α , α ˉ ( η , η ˉ ) = A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})-B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) D α , α ˉ ( η , η ˉ ) = A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) . Since B α , α ˉ B^{\alpha,\bar{\alpha}} B α , α ˉ is constant with respect to t t t , minimizing the risk is equivalent to minimizing the term D α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\mathbf{1}\{t\le0\} D α , α ˉ ( η , η ˉ ) 1 { t ≤ 0 } . We consider three cases based on the sign of D α , α ˉ ( η , η ˉ ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) D α , α ˉ ( η , η ˉ ) . Case D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 D α , α ˉ ( η , η ˉ ) > 0 . To minimize the risk, we must set 1 { t ≤ 0 } = 0 \mathbf{1}\{t \le0\} = 0 1 { t ≤ 0 } = 0 , i.e., t > 0 t > 0 t > 0 . The optimal fairness-aware conditional risk is
C f a i r α , α ˉ , ∗ ( η , η ˉ ) = B α , α ˉ ( η , η ˉ ) . C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta}) = B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). C fair α , α ˉ , ∗ ( η , η ˉ ) = B α , α ˉ ( η , η ˉ ) .
Since D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 D α , α ˉ ( η , η ˉ ) > 0 , we have D α , α ˉ ( η , η ˉ ) = ∣ D α , α ˉ ( η , η ˉ ) ∣ D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = \lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \rvert D α , α ˉ ( η , η ˉ ) = ∣ D α , α ˉ ( η , η ˉ )∣ and sign ( D α , α ˉ ( η , η ˉ ) ) = 1 \operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})) = 1 sign ( D α , α ˉ ( η , η ˉ )) = 1 . Furthermore, 1 { t ≤ 0 } \mathbf{1}\{t \le0\} 1 { t ≤ 0 } implies sign ( t ) = − 1 ≠ sign ( D α , α ˉ ( η , η ˉ ) ) \operatorname{sign}(t) = -1 \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})) sign ( t ) = − 1 = sign ( D α , α ˉ ( η , η ˉ )) . Thus, we have
C f a i r α , α ˉ ( η , η ˉ , t ) − C f a i r α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) ≠ sign ( D α , α ˉ ( η , η ˉ ) ) } ∣ D α , α ˉ ( η , η ˉ ) ∣ . C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t) - C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta}) = \mathbf{1}\{\operatorname{sign}(t) \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}))\}\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert. C fair α , α ˉ ( η , η ˉ , t ) − C fair α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) = sign ( D α , α ˉ ( η , η ˉ ))} ∣ D α , α ˉ ( η , η ˉ )∣ .
Case D α , α ˉ ( η , η ˉ ) < 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) < 0 D α , α ˉ ( η , η ˉ ) < 0 . To minimize the risk, we must set 1 { t ≤ 0 } = 1 \mathbf{1}\{t \le0\} = 1 1 { t ≤ 0 } = 1 , i.e., t ≤ 0 t \le0 t ≤ 0 . The optimal fairness-aware conditional risk is
C f a i r α , α ˉ , ∗ ( η , η ˉ ) = A α , α ˉ ( η , η ˉ ) . C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta}) = A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). C fair α , α ˉ , ∗ ( η , η ˉ ) = A α , α ˉ ( η , η ˉ ) .
Therefore, the excess fairness-aware conditional risk for any t t t is
C f a i r α , α ˉ ( η , η ˉ , t ) − C f a i r α , α ˉ , ∗ ( η , η ˉ ) = − D α , α ˉ ( η , η ˉ ) 1 { t > 0 } . C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t) - C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta}) = -D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\mathbf{1}\{t > 0\}. C fair α , α ˉ ( η , η ˉ , t ) − C fair α , α ˉ , ∗ ( η , η ˉ ) = − D α , α ˉ ( η , η ˉ ) 1 { t > 0 } .
Since D α , α ˉ ( η , η ˉ ) < 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) < 0 D α , α ˉ ( η , η ˉ ) < 0 , we have D α , α ˉ ( η , η ˉ ) = − ∣ D α , α ˉ ( η , η ˉ ) ∣ D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = -\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert D α , α ˉ ( η , η ˉ ) = − ∣ D α , α ˉ ( η , η ˉ )∣ and sign ( D α , α ˉ ( η , η ˉ ) ) = − 1 \operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})) = -1 sign ( D α , α ˉ ( η , η ˉ )) = − 1 . Furthermore, 1 { t > 0 } \mathbf{1}\{t > 0\} 1 { t > 0 } implies sign ( t ) = 1 ≠ sign ( D α , α ˉ ( η , η ˉ ) ) = − 1 \operatorname{sign}(t) = 1 \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})) = -1 sign ( t ) = 1 = sign ( D α , α ˉ ( η , η ˉ )) = − 1 . Thus, we also have
C f a i r α , α ˉ ( η , η ˉ , t ) − C f a i r α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) ≠ sign ( D α , α ˉ ( η , η ˉ ) ) } ∣ D α , α ˉ ( η , η ˉ ) ∣ . C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t) - C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta}) = \mathbf{1}\{\operatorname{sign}(t) \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}))\}\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert. C fair α , α ˉ ( η , η ˉ , t ) − C fair α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) = sign ( D α , α ˉ ( η , η ˉ ))} ∣ D α , α ˉ ( η , η ˉ )∣ .
Case D α , α ˉ ( η , η ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 D α , α ˉ ( η , η ˉ ) = 0 . Since the risk is constant with respect to t t t , the excess risk is always 0.
Combining these cases, the conditional excess risk is
C f a i r α , α ˉ ( η , η ˉ , t ) − C f a i r α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) ≠ sign ( D α , α ˉ ( η , η ˉ ) ) } ∣ D α , α ˉ ( η , η ˉ ) ∣ = 1 { sign ( t ) ≠ sign ( A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) ) } ∣ A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ) ∣ . \begin{aligned}
C_{\mathrm{fair}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta},t) - C_{\mathrm{fair}}^{\alpha,\bar{\alpha},*}(\eta,\bar{\eta})
&= \mathbf{1}\{\operatorname{sign}(t) \ne\operatorname{sign}(D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}))\}\lvert D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert\\
&= \mathbf{1}\{\operatorname{sign}(t) \ne\operatorname{sign}(A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) - B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}))\}\lvert A^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) - B^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\rvert.
\end{aligned} C fair α , α ˉ ( η , η ˉ , t ) − C fair α , α ˉ , ∗ ( η , η ˉ ) = 1 { sign ( t ) = sign ( D α , α ˉ ( η , η ˉ ))} ∣ D α , α ˉ ( η , η ˉ )∣ = 1 { sign ( t ) = sign ( A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ ))} ∣ A α , α ˉ ( η , η ˉ ) − B α , α ˉ ( η , η ˉ )∣ .
Consequently,
R f a i r α , α ˉ ( f ) − R f a i r α , α ˉ , ∗ = E X [ C f a i r α , α ˉ ( η ( X ) , η ˉ ( X ) , f ( X ) ) − C f a i r α , α ˉ , ∗ ( η ( X ) , η ˉ ( X ) ) ] = E X [ 1 { sign ( f ( X ) ) ≠ sign ( A α , α ˉ ( η ( X ) , η ˉ ( X ) ) − B α , α ˉ ( η ( X ) , η ˉ ( X ) ) ) } ⋅ ∣ A α , α ˉ ( η ( X ) , η ˉ ( X ) ) − B α , α ˉ ( η ( X ) , η ˉ ( X ) ) ∣ ] . \mathcal{R}^{\alpha,\bar{\alpha}}_{\mathrm{fair}}(f)-\mathcal{R}^{\alpha,\bar{\alpha},*}_{\mathrm{fair}}
= \mathbb{E}_{X}\left[C^{\alpha,\bar{\alpha}}_{\mathrm{fair}}(\eta(X),\bar{\eta}(X),f(X))-C^{\alpha,\bar{\alpha},*}_{\mathrm{fair}}(\eta(X),\bar{\eta}(X))\right]
= \mathbb{E}_{X}\left[1\left\{\operatorname{sign}(f(X))\ne\operatorname{sign}\left(A^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))-B^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))\right)\right\}\cdot\left|A^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))-B^{\alpha,\bar{\alpha}}(\eta(X),\bar{\eta}(X))\right|\right]. R fair α , α ˉ ( f ) − R fair α , α ˉ , ∗ = E X [ C fair α , α ˉ ( η ( X ) , η ˉ ( X ) , f ( X )) − C fair α , α ˉ , ∗ ( η ( X ) , η ˉ ( X )) ] = E X [ 1 { sign ( f ( X )) = sign ( A α , α ˉ ( η ( X ) , η ˉ ( X )) − B α , α ˉ ( η ( X ) , η ˉ ( X )) ) } ⋅ A α , α ˉ ( η ( X ) , η ˉ ( X )) − B α , α ˉ ( η ( X ) , η ˉ ( X )) ] .
Equivalent Conditions of FAFC
Lemma 23. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha}\in(0,1) α , α ˉ ∈ ( 0 , 1 ) , let K \mathcal{K} K be a well-posed domain as in Remark 11, and let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be a surrogate loss pair such that, for every ( η , η ˉ ) ∈ K (\eta,\bar{\eta})\in\mathcal{K} ( η , η ˉ ) ∈ K , the map t ↦ C f a i r L , L ˉ ( η , η ˉ , t ) t\mapsto C^{L,\bar{L}}_{\mathrm{fair}}(\eta,\bar{\eta},t) t ↦ C fair L , L ˉ ( η , η ˉ , t ) is convex on R \mathbb{R} R and differentiable at 0 0 0 . Let
Φ L , L ˉ ( η , η ˉ ) : = ∂ t C f a i r L , L ˉ ( η , η ˉ , t ) ∣ t = 0 . \Phi_{L,\bar{L}}(\eta,\bar{\eta}) := \left.\partial_t C^{L,\bar{L}}_{\mathrm{fair}}(\eta,\bar{\eta},t)\right|_{t=0}. Φ L , L ˉ ( η , η ˉ ) := ∂ t C fair L , L ˉ ( η , η ˉ , t ) t = 0 . Then the following three statements are equivalent:
L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K ;
H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta,\bar{\eta})>0 H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 for every ( η , η ˉ ) ∈ K (\eta,\bar{\eta})\in\mathcal{K} ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\ne0 D α , α ˉ ( η , η ˉ ) = 0 (Definition 12 );
Φ L , L ˉ ( η , η ˉ ) ⋅ D α , α ˉ ( η , η ˉ ) < 0 \Phi_{L,\bar{L}}(\eta,\bar{\eta})\cdot D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})<0 Φ L , L ˉ ( η , η ˉ ) ⋅ D α , α ˉ ( η , η ˉ ) < 0 for every ( η , η ˉ ) ∈ K (\eta,\bar{\eta})\in\mathcal{K} ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\ne0 D α , α ˉ ( η , η ˉ ) = 0 .
Proof. The equivalence of (i) and (ii) is Definition 12 . Since K \mathcal{K} K is a well-posed domain, the conditional risk is bounded below on K \mathcal{K} K .
For (ii)⇔ \Leftrightarrow ⇔ (iii), fix ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) , write C ( t ) : = C f a i r L , L ˉ ( η , η ˉ , t ) C(t):=C^{L,\bar{L}}_{\mathrm{fair}}(\eta,\bar{\eta},t) C ( t ) := C fair L , L ˉ ( η , η ˉ , t ) , and let Φ : = Φ L , L ˉ ( η , η ˉ ) = C ′ ( 0 ) \Phi:=\Phi_{L,\bar{L}}(\eta,\bar{\eta})=C'(0) Φ := Φ L , L ˉ ( η , η ˉ ) = C ′ ( 0 ) .
Convexity and differentiability at 0 0 0 give
C ( t ) ≥ C ( 0 ) + Φ t for every t ∈ R . (17) C(t)\ge C(0)+\Phi t \qquad\text{for every }t\in\mathbb{R}. \tag*{(17)} C ( t ) ≥ C ( 0 ) + Φ t for every t ∈ R . (17)
Assume first that D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 D α , α ˉ ( η , η ˉ ) > 0 , so the wrong-sign set is t ≤ 0 t \le0 t ≤ 0 .
If Φ < 0 \Phi< 0 Φ < 0 , then (17) gives C ( t ) ≥ C ( 0 ) C(t) \ge C(0) C ( t ) ≥ C ( 0 ) for every t ≤ 0 t \le0 t ≤ 0 , while C ( s ) < C ( 0 ) C(s) < C(0) C ( s ) < C ( 0 ) for all sufficiently small s > 0 s > 0 s > 0 . Hence
inf t ≤ 0 C ( t ) ≥ C ( 0 ) > inf t ∈ R C ( t ) , \inf_{t\le0} C(t) \ge C(0) > \inf_{t\in\mathbb{R}} C(t), t ≤ 0 inf C ( t ) ≥ C ( 0 ) > t ∈ R inf C ( t ) ,
so H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 .
If Φ ≥ 0 \Phi\ge0 Φ ≥ 0 , then C ( t ) ≥ C ( 0 ) C(t) \ge C(0) C ( t ) ≥ C ( 0 ) for every t > 0 t > 0 t > 0 . Since 0 0 0 belongs to the wrong-sign set,
inf t ≤ 0 C ( t ) ≤ C ( 0 ) ≤ inf t > 0 C ( t ) , \inf_{t\le0} C(t) \le C(0) \le\inf_{t>0} C(t), t ≤ 0 inf C ( t ) ≤ C ( 0 ) ≤ t > 0 inf C ( t ) ,
so the wrong-sign infimum equals the global infimum and H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 .
Thus, for D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 D α , α ˉ ( η , η ˉ ) > 0 , H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) > 0 H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 if and only if Φ < 0 \Phi< 0 Φ < 0 , i.e., Φ D α , α ˉ ( η , η ˉ ) < 0 \Phi D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) < 0 Φ D α , α ˉ ( η , η ˉ ) < 0 . The case D α , α ˉ ( η , η ˉ ) < 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) < 0 D α , α ˉ ( η , η ˉ ) < 0 is symmetric, with the roles of t ≤ 0 t \le0 t ≤ 0 and t ≥ 0 t \ge0 t ≥ 0 interchanged. This proves (ii)⇔ \Leftrightarrow ⇔ (iii). □
Proof of Lemma 24
Lemma 24. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) and let K \mathcal{K} K be as in § 4 . Assume the map t ↦ C f a i r L , L ˉ ( η , η ˉ , t ) t \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) t ↦ C fair L , L ˉ ( η , η ˉ , t ) is convex on R \mathbb{R} R for every ( η , η ˉ ) ∈ K (\eta,\bar{\eta}) \in\mathcal{K} ( η , η ˉ ) ∈ K . Then,
H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 whenever D α , α ˉ ( η , η ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 D α , α ˉ ( η , η ˉ ) = 0 . In particular,
H L , L ˉ α , α ˉ ( α , α ˉ ) = ν L , L ˉ α , α ˉ ( 0 ) = ψ L , L ˉ α , α ˉ ( 0 ) = 0. H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha}) = \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0) = \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0) = 0. H L , L ˉ α , α ˉ ( α , α ˉ ) = ν L , L ˉ α , α ˉ ( 0 ) = ψ L , L ˉ α , α ˉ ( 0 ) = 0. If L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K , then
ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B max ] . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) > 0 \text{ for every } \epsilon\in(0,B_{\max}]. ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B m a x ] . If L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K , then ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is continuous on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] .
Proof. Part 1. Let ( η , η ˉ ) ∈ K (\eta,\bar{\eta}) \in\mathcal{K} ( η , η ˉ ) ∈ K satisfy D α , α ˉ ( η , η ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 D α , α ˉ ( η , η ˉ ) = 0 . Then t D α , α ˉ ( η , η ˉ ) = 0 ≤ 0 tD^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = 0 \le0 t D α , α ˉ ( η , η ˉ ) = 0 ≤ 0 for every t ∈ R t \in\mathbb{R} t ∈ R , so the feasible set in the definition of C f a i r L , L ˉ ( η , η ˉ ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta}) C fair L , L ˉ ( η , η ˉ ) is all of R \mathbb{R} R and
C f a i r L , L ˉ ( η , η ˉ ) = inf t ∈ R C f a i r L , L ˉ ( η , η ˉ , t ) = C f a i r L , L ˉ , ∗ ( η , η ˉ ) . C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta})=\inf_{t\in\mathbb{R}} C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t)=C_{\mathrm{fair}}^{L,\bar{L},*}(\eta,\bar{\eta}). C fair L , L ˉ ( η , η ˉ ) = t ∈ R inf C fair L , L ˉ ( η , η ˉ , t ) = C fair L , L ˉ , ∗ ( η , η ˉ ) .
Hence H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0 H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 . In particular, since A D P α , α ˉ ( α , α ˉ ) = α ( 1 − α ) − λ α ˉ ( 1 − α ˉ ) = B D P α , α ˉ ( α , α ˉ ) A_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha})=\alpha(1-\alpha)-\lambda\bar{\alpha}(1-\bar{\alpha})=B_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha}) A DP α , α ˉ ( α , α ˉ ) = α ( 1 − α ) − λ α ˉ ( 1 − α ˉ ) = B DP α , α ˉ ( α , α ˉ )
and A E O α , α ˉ ( α , α ˉ ) = α ( 1 − α ) − λ π − 1 α α ˉ ( 1 − α ˉ ) = B E O α , α ˉ ( α , α ˉ ) A_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha})=\alpha(1-\alpha)-\lambda\pi^{-1}\alpha\bar{\alpha}(1-\bar{\alpha})=B_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha}) A EO α , α ˉ ( α , α ˉ ) = α ( 1 − α ) − λ π − 1 α α ˉ ( 1 − α ˉ ) = B EO α , α ˉ ( α , α ˉ ) , D α , α ˉ ( α , α ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha})=0 D α , α ˉ ( α , α ˉ ) = 0 and therefore
H L , L ˉ α , α ˉ ( α , α ˉ ) = 0. H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha})=0. H L , L ˉ α , α ˉ ( α , α ˉ ) = 0.
Hence the level set L 0 = { ( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = 0 } = { ( η , η ˉ ) ∈ K : D α , α ˉ ( η , η ˉ ) = 0 } \mathcal{L}_0=\{(\eta,\bar{\eta})\in\mathcal{K}:|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})|=0\}=\{(\eta,\bar{\eta})\in\mathcal{K}:D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0\} L 0 = {( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = 0 } = {( η , η ˉ ) ∈ K : D α , α ˉ ( η , η ˉ ) = 0 } is nonempty, since it contains ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) and H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0 H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 on all of L 0 \mathcal{L}_0 L 0 . Thus,
ν L , L ˉ α , α ˉ ( 0 ) = inf ( η , η ˉ ) ∈ L 0 H L , L ˉ α , α ˉ ( η , η ˉ ) = 0. \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=\inf_{(\eta,\bar{\eta})\in\mathcal{L}_0}H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0. ν L , L ˉ α , α ˉ ( 0 ) = ( η , η ˉ ) ∈ L 0 inf H L , L ˉ α , α ˉ ( η , η ˉ ) = 0.
By the definition of the biconjugate function, ψ L , L ˉ α , α ˉ ( ϵ ) ≤ ν L , L ˉ α , α ˉ ( ϵ ) \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)\leq\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) ψ L , L ˉ α , α ˉ ( ϵ ) ≤ ν L , L ˉ α , α ˉ ( ϵ ) and ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is the largest convex, lower semi-continuous function majorized by ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ . Hence ψ L , L ˉ α , α ˉ ( 0 ) ≤ ν L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)\leq\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) ≤ ν L , L ˉ α , α ˉ ( 0 ) = 0 .
Conversely, since H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\geq0 H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ 0 , ν L , L ˉ α , α ˉ ( ϵ ) ≥ 0 \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)\geq0 ν L , L ˉ α , α ˉ ( ϵ ) ≥ 0 on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] . The constant function 0 0 0 is therefore convex, lower semi-continuous and majorized by ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ . As ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is the largest such function, ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)\geq0 ψ L , L ˉ α , α ˉ ( ϵ ) ≥ 0 if ϵ ∈ [ 0 , B max ] \epsilon\in[0,B_{\max}] ϵ ∈ [ 0 , B m a x ] . Hence ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 .
Part 2. By the explicit formulas for the example-dependent fair cost functions, D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ is a polynomial in ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) ; for instance, D D P α , α ˉ ( η , η ˉ ) = ( η − α ) − λ ( η ˉ − α ˉ ) D_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=(\eta-\alpha)-\lambda(\bar{\eta}-\bar{\alpha}) D DP α , α ˉ ( η , η ˉ ) = ( η − α ) − λ ( η ˉ − α ˉ ) and D E O α , α ˉ ( η , η ˉ ) = ( η − α ) − λ π − 1 η ( η ˉ − α ˉ ) D_{\mathrm{EO}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=(\eta-\alpha)-\lambda\pi^{-1}\eta(\bar{\eta}-\bar{\alpha}) D EO α , α ˉ ( η , η ˉ ) = ( η − α ) − λ π − 1 η ( η ˉ − α ˉ ) . Hence D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ is continuous on K \mathcal{K} K . Since K \mathcal{K} K is compact and connected and D α , α ˉ ( α , α ˉ ) = 0 D^{\alpha,\bar{\alpha}}(\alpha,\bar{\alpha})=0 D α , α ˉ ( α , α ˉ ) = 0 , the range of ∣ D α , α ˉ ∣ |D^{\alpha,\bar{\alpha}}| ∣ D α , α ˉ ∣ is [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] , so the level set
L ϵ = { ( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ } \mathcal{L}_{\epsilon}=\{(\eta,\bar{\eta})\in\mathcal{K}:|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})|=\epsilon\} L ϵ = {( η , η ˉ ) ∈ K : ∣ D α , α ˉ ( η , η ˉ ) ∣ = ϵ }
is nonempty for every ϵ ∈ [ 0 , B max ] \epsilon\in[0,B_{\max}] ϵ ∈ [ 0 , B m a x ] .
Fix ( η , η ˉ ) ∈ K (\eta,\bar{\eta})\in\mathcal{K} ( η , η ˉ ) ∈ K with D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})>0 D α , α ˉ ( η , η ˉ ) > 0 (the case D α , α ˉ ( η , η ˉ ) < 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})<0 D α , α ˉ ( η , η ˉ ) < 0 is symmetric). Since L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K , we have H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})>0 H L , L ˉ α , α ˉ ( η , η ˉ ) > 0 , that is,
inf t ≤ 0 C f a i r L , L ˉ ( η , η ˉ , t ) > inf t ∈ R C f a i r L , L ˉ ( η , η ˉ , t ) , \inf_{t\leq0}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t)>\inf_{t\in\mathbb{R}}C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t), t ≤ 0 inf C fair L , L ˉ ( η , η ˉ , t ) > t ∈ R inf C fair L , L ˉ ( η , η ˉ , t ) ,
so there exists s ∈ R + s \in\mathbb{R}^{+} s ∈ R + with C f a i r L , L ˉ ( η , η ˉ , s ) < inf t ≤ 0 C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},s) < \inf_{t \le0} C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ ( η , η ˉ , s ) < inf t ≤ 0 C fair L , L ˉ ( η , η ˉ , t ) . For every t ≤ 0 t \le0 t ≤ 0 the point 0 0 0 is a convex combination of t t t and s s s , so convexity of u ↦ C f a i r L , L ˉ ( η , η ˉ , u ) u \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},u) u ↦ C fair L , L ˉ ( η , η ˉ , u ) together with C f a i r L , L ˉ ( η , η ˉ , s ) < C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},s) < C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ ( η , η ˉ , s ) < C fair L , L ˉ ( η , η ˉ , t ) gives C f a i r L , L ˉ ( η , η ˉ , 0 ) < C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},0) < C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ ( η , η ˉ , 0 ) < C fair L , L ˉ ( η , η ˉ , t ) . As t = 0 t=0 t = 0 itself is feasible, we know
C f a i r L , L ˉ , − α ˉ ( η , η ˉ ) = C f a i r L , L ˉ ( η , η ˉ , 0 ) , C_{\mathrm{fair}}^{L,\bar{L},-\bar{\alpha}}(\eta,\bar{\eta}) = C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},0), C fair L , L ˉ , − α ˉ ( η , η ˉ ) = C fair L , L ˉ ( η , η ˉ , 0 ) ,
and therefore
H L , L ˉ α , α ˉ ( η , η ˉ ) = C f a i r L , L ˉ ( η , η ˉ , 0 ) − C f a i r L , L ˉ , ∗ ( η , η ˉ ) (18) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) = C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},0) - C_{\mathrm{fair}}^{L,\bar{L},*\!}(\eta,\bar{\eta}) \tag*{(18)} H L , L ˉ α , α ˉ ( η , η ˉ ) = C fair L , L ˉ ( η , η ˉ , 0 ) − C fair L , L ˉ , ∗ ( η , η ˉ ) (18)
on { ( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) ≠ 0 } \{(\eta,\bar{\eta}):D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0\} {( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) = 0 } .
For every fixed t ∈ R t \in\mathbb{R} t ∈ R the map ( η , η ˉ ) ↦ C f a i r L , L ˉ ( η , η ˉ , t ) (\eta,\bar{\eta}) \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) ( η , η ˉ ) ↦ C fair L , L ˉ ( η , η ˉ , t ) is continuous on K \mathcal{K} K since (8) and (9) are polynomials of degree at most two in ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) whose coefficients L 1 ( t ) , L − 1 ( t ) , L ˉ 1 ( t ) , L ˉ − 1 ( t ) L_{1}(t),L_{-1}(t),\bar{L}_{1}(t),\bar{L}_{-1}(t) L 1 ( t ) , L − 1 ( t ) , L ˉ 1 ( t ) , L ˉ − 1 ( t ) are real numbers.
Now we know ( η , η ˉ ) ↦ C f a i r L , L ˉ ( η , η ˉ , 0 ) (\eta,\bar{\eta}) \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},0) ( η , η ˉ ) ↦ C fair L , L ˉ ( η , η ˉ , 0 ) is continuous. And since C f a i r L , L ˉ , ∗ ( η , η ˉ ) = inf t ∈ R C f a i r L , L ˉ ( η , η ˉ , t ) C_{\mathrm{fair}}^{L,\bar{L},*\!}(\eta,\bar{\eta}) = \inf_{t \in\mathbb{R}} C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) C fair L , L ˉ , ∗ ( η , η ˉ ) = inf t ∈ R C fair L , L ˉ ( η , η ˉ , t ) is an infimum of functions that are continuous in ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) , hence is upper semi-continuous. By (18) , H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ is a continuous function minus an upper semi-continuous function, hence lower semi-continuous on { ( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) ≠ 0 } \{(\eta,\bar{\eta}):D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0\} {( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) = 0 } .
Fix ϵ ∈ ( 0 , B max ] \epsilon\in(0,B_{\max}] ϵ ∈ ( 0 , B m a x ] . By the definition of L ϵ \mathcal{L}_{\epsilon} L ϵ , we have L ϵ ⊆ { ( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) ≠ 0 } \mathcal{L}_{\epsilon} \subseteq\{(\eta,\bar{\eta}):D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0\} L ϵ ⊆ {( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) = 0 } . Since H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ is lower semi-continuous on { ( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) ≠ 0 } \{(\eta,\bar{\eta}):D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) \ne0\} {( η , η ˉ ) : D α , α ˉ ( η , η ˉ ) = 0 } , H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ is also lower semi-continuous on L ϵ \mathcal{L}_{\epsilon} L ϵ .
Recall that ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ L ϵ H L , L ˉ α , α ˉ ( η , η ˉ ) \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) = \inf_{(\eta,\bar{\eta})\in\mathcal{L}_{\epsilon}} H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}) ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ L ϵ H L , L ˉ α , α ˉ ( η , η ˉ ) , which is finite because H L , L ˉ α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ is nonnegative on L ϵ \mathcal{L}_{\epsilon} L ϵ . Choose ( η k , η ˉ k ) ∈ L ϵ (\eta_k,\bar{\eta}_k) \in\mathcal{L}_{\epsilon} ( η k , η ˉ k ) ∈ L ϵ with H L , L ˉ α , α ˉ ( η k , η ˉ k ) → ν L , L ˉ α , α ˉ ( ϵ ) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta_k,\bar{\eta}_k) \to\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon) H L , L ˉ α , α ˉ ( η k , η ˉ k ) → ν L , L ˉ α , α ˉ ( ϵ ) as k → ∞ k \to\infty k → ∞ . By compactness of L ϵ \mathcal{L}_{\epsilon} L ϵ a subsequence converges to ( η ∗ , η ˉ ∗ ) ∈ L ϵ (\eta^{*},\bar{\eta}^{*}) \in\mathcal{L}_{\epsilon} ( η ∗ , η ˉ ∗ ) ∈ L ϵ , and lower semi-continuity gives
H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) ≤ lim inf k → ∞ H L , L ˉ α , α ˉ ( η k , η ˉ k ) = ν L , L ˉ α , α ˉ ( ϵ ) . (19) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta^{*},\bar{\eta}^{*}) \le\liminf_{k \to\infty} H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta_k,\bar{\eta}_k) = \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon). \tag*{(19)} H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) ≤ k → ∞ lim inf H L , L ˉ α , α ˉ ( η k , η ˉ k ) = ν L , L ˉ α , α ˉ ( ϵ ) . (19)
Since ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ L ϵ H L , L ˉ α , α ˉ ( η , η ˉ ) \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon)=\inf_{(\eta,\bar{\eta})\in\mathcal{L}_{\epsilon}}H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta,\bar{\eta}) ν L , L ˉ α , α ˉ ( ϵ ) = inf ( η , η ˉ ) ∈ L ϵ H L , L ˉ α , α ˉ ( η , η ˉ ) , equality holds, i.e.
ν L , L ˉ α , α ˉ ( ϵ ) = H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) . (20) \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon)=H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta^*,\bar{\eta}^*). \tag*{(20)} ν L , L ˉ α , α ˉ ( ϵ ) = H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) . (20)
As ( η ∗ , η ˉ ∗ ) ∈ L ϵ (\eta^*,\bar{\eta}^*)\in\mathcal{L}_{\epsilon} ( η ∗ , η ˉ ∗ ) ∈ L ϵ , D α , α ˉ ( η ∗ , η ˉ ∗ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta^*,\bar{\eta}^*)\ne0 D α , α ˉ ( η ∗ , η ˉ ∗ ) = 0 , so FAFC on K \mathcal{K} K yields H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) > 0 H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta^*,\bar{\eta}^*)>0 H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) > 0 . Therefore
ν L , L ˉ α , α ˉ ( ϵ ) = H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) > 0 for all ϵ ∈ ( 0 , B max ] . (21) \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon)=H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta^*,\bar{\eta}^*)>0\text{ for all }\epsilon\in(0,B_{\max}]. \tag*{(21)} ν L , L ˉ α , α ˉ ( ϵ ) = H L , L ˉ α , α ˉ ( η ∗ , η ˉ ∗ ) > 0 for all ϵ ∈ ( 0 , B m a x ] . (21)
Next, we prove ν L , L ˉ α , α ˉ \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} ν L , L ˉ α , α ˉ is lower semi-continuous. Let ϵ n → ϵ \epsilon_n\to\epsilon ϵ n → ϵ in [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] and set a : = lim inf n → ∞ ν L , L ˉ α , α ˉ ( ϵ n ) a:=\liminf_{n\to\infty}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_n) a := lim inf n → ∞ ν L , L ˉ α , α ˉ ( ϵ n ) . Choose a subsequence n j n_j n j with ν L , L ˉ α , α ˉ ( ϵ n j ) → a \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_{n_j})\to a ν L , L ˉ α , α ˉ ( ϵ n j ) → a . If ϵ > 0 \epsilon>0 ϵ > 0 , then ϵ n j > 0 \epsilon_{n_j}>0 ϵ n j > 0 for all sufficiently large j j j ; after discarding finitely many terms, choose minimizers ( η n j , η ˉ n j ) ∈ L ϵ n j (\eta_{n_j},\bar{\eta}_{n_j})\in\mathcal{L}_{\epsilon_{n_j}} ( η n j , η ˉ n j ) ∈ L ϵ n j with ν L , L ˉ α , α ˉ ( ϵ n j ) = H L , L ˉ α , α ˉ ( η n j , η ˉ n j ) \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_{n_j})=H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta_{n_j},\bar{\eta}_{n_j}) ν L , L ˉ α , α ˉ ( ϵ n j ) = H L , L ˉ α , α ˉ ( η n j , η ˉ n j ) . Such minimizers exist by (19) and (20) .
Since K \mathcal{K} K is compact, pass to a further subsequence n j i n_{j_i} n j i such that ( η n j i , η ˉ n j i ) → ( η , η ˉ ) ∈ K (\eta_{n_{j_i}},\bar{\eta}_{n_{j_i}})\to(\eta,\bar{\eta})\in\mathcal{K} ( η n j i , η ˉ n j i ) → ( η , η ˉ ) ∈ K as i → ∞ i\to\infty i → ∞ . By the continuity of D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ ,
∣ D α , α ˉ ( η , η ˉ ) ∣ = lim i → ∞ ∣ D α , α ˉ ( η n j i , η ˉ n j i ) ∣ = lim i → ∞ ϵ n j i = ϵ . \left|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\right|=\lim_{i\to\infty}\left|D^{\alpha,\bar{\alpha}}(\eta_{n_{j_i}},\bar{\eta}_{n_{j_i}})\right|=\lim_{i\to\infty}\epsilon_{n_{j_i}}=\epsilon. D α , α ˉ ( η , η ˉ ) = i → ∞ lim D α , α ˉ ( η n j i , η ˉ n j i ) = i → ∞ lim ϵ n j i = ϵ .
Hence ( η , η ˉ ) ∈ L ϵ (\eta,\bar{\eta})\in\mathcal{L}_{\epsilon} ( η , η ˉ ) ∈ L ϵ . Since ϵ > 0 \epsilon>0 ϵ > 0 , D α , α ˉ ( η , η ˉ ) ≠ 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\ne0 D α , α ˉ ( η , η ˉ ) = 0 , and lower semi-continuity of H L , L ˉ α , α ˉ H^{\alpha,\bar{\alpha}}_{L,\bar{L}} H L , L ˉ α , α ˉ at ( η , η ˉ ) (\eta,\bar{\eta}) ( η , η ˉ ) yields
a = lim i → ∞ ν L , L ˉ α , α ˉ ( ϵ n j i ) = lim i → ∞ H L , L ˉ α , α ˉ ( η n j i , η ˉ n j i ) ≥ H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ ν L , L ˉ α , α ˉ ( ϵ ) . a=\lim_{i\to\infty}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_{n_{j_i}})=\lim_{i\to\infty}H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta_{n_{j_i}},\bar{\eta}_{n_{j_i}})\ge H^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\eta,\bar{\eta})\ge\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon). a = i → ∞ lim ν L , L ˉ α , α ˉ ( ϵ n j i ) = i → ∞ lim H L , L ˉ α , α ˉ ( η n j i , η ˉ n j i ) ≥ H L , L ˉ α , α ˉ ( η , η ˉ ) ≥ ν L , L ˉ α , α ˉ ( ϵ ) .
If ϵ = 0 \epsilon=0 ϵ = 0 , then ν L , L ˉ α , α ˉ ( 0 ) = 0 ≤ a \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(0)=0\le a ν L , L ˉ α , α ˉ ( 0 ) = 0 ≤ a because ν L , L ˉ α , α ˉ ≥ 0 \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}\ge0 ν L , L ˉ α , α ˉ ≥ 0 . Hence lim inf n → ∞ ν L , L ˉ α , α ˉ ( ϵ n ) ≥ ν L , L ˉ α , α ˉ ( ϵ ) \liminf_{n\to\infty}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_n)\ge\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon) lim inf n → ∞ ν L , L ˉ α , α ˉ ( ϵ n ) ≥ ν L , L ˉ α , α ˉ ( ϵ ) in all cases, so ν L , L ˉ α , α ˉ \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} ν L , L ˉ α , α ˉ is lower semi-continuous on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] .
Since ψ L , L ˉ α , α ˉ ≥ 0 \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}\ge0 ψ L , L ˉ α , α ˉ ≥ 0 on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] , so it suffices to exclude ψ L , L ˉ α , α ˉ ( ϵ 0 ) = 0 \psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon_0)=0 ψ L , L ˉ α , α ˉ ( ϵ 0 ) = 0 for some ϵ 0 ∈ ( 0 , B max ] \epsilon_0\in(0,B_{\max}] ϵ 0 ∈ ( 0 , B m a x ] .
Suppose such an ϵ 0 \epsilon_0 ϵ 0 exists. By the definition of the biconjugate function, Epi ψ L , L ˉ α , α ˉ = co Epi ν L , L ˉ α , α ˉ \operatorname{Epi}\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}}=\operatorname{co}\operatorname{Epi}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} Epi ψ L , L ˉ α , α ˉ = co Epi ν L , L ˉ α , α ˉ and ( ϵ 0 , 0 ) ∈ Epi ψ L , L ˉ α , α ˉ (\epsilon_0,0)\in\operatorname{Epi}\psi^{\alpha,\bar{\alpha}}_{L,\bar{L}} ( ϵ 0 , 0 ) ∈ Epi ψ L , L ˉ α , α ˉ . Hence there exist w k ∈ co Epi ν L , L ˉ α , α ˉ w_k\in\operatorname{co}\operatorname{Epi}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} w k ∈ co Epi ν L , L ˉ α , α ˉ with w k → ( ϵ 0 , 0 ) w_k\to(\epsilon_0,0) w k → ( ϵ 0 , 0 ) .
By Carathéodory’s theorem in R 2 \mathbb{R}^2 R 2 , for every k k k there are weights θ 1 ( k ) , θ 2 ( k ) , θ 3 ( k ) ≥ 0 \theta^{(k)}_1,\theta^{(k)}_2,\theta^{(k)}_3\ge0 θ 1 ( k ) , θ 2 ( k ) , θ 3 ( k ) ≥ 0 with ∑ i = 1 3 θ i ( k ) = 1 \sum_{i=1}^{3}\theta^{(k)}_i=1 ∑ i = 1 3 θ i ( k ) = 1 and points ( r i ( k ) , s i ( k ) ) ∈ Epi ν L , L ˉ α , α ˉ (r^{(k)}_i,s^{(k)}_i)\in\operatorname{Epi}\nu^{\alpha,\bar{\alpha}}_{L,\bar{L}} ( r i ( k ) , s i ( k ) ) ∈ Epi ν L , L ˉ α , α ˉ such that w k = ∑ i = 1 3 θ i ( k ) ( r i ( k ) , s i ( k ) ) w_k=\sum_{i=1}^{3}\theta^{(k)}_i(r^{(k)}_i,s^{(k)}_i) w k = ∑ i = 1 3 θ i ( k ) ( r i ( k ) , s i ( k ) ) . Since ν L , L ˉ α , α ˉ ( ϵ ) = ∞ \nu^{\alpha,\bar{\alpha}}_{L,\bar{L}}(\epsilon)=\infty ν L , L ˉ α , α ˉ ( ϵ ) = ∞ outside [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] , every r i ( k ) r^{(k)}_i r i ( k ) lies in [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] . Since θ i ( k ) \theta_i^{(k)} θ i ( k ) is in a standard simplex and r i ( k ) ∈ [ 0 , B max ] r_i^{(k)} \in[0,B_{\max}] r i ( k ) ∈ [ 0 , B m a x ] , by the compactness of the simplex and of [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] , there exist subsequences θ i ( k m ) \theta_i^{(k_m)} θ i ( k m ) and r i ( k m ) r_i^{(k_m)} r i ( k m ) such that θ i ( k m ) → θ i \theta_i^{(k_m)} \to\theta_i θ i ( k m ) → θ i and r i ( k m ) → r i r_i^{(k_m)} \to r_i r i ( k m ) → r i for i = 1 , 2 , 3 i=1,2,3 i = 1 , 2 , 3 as m → ∞ m \to\infty m → ∞ .
As the second coordinate ( w k ) 2 (w_k)_2 ( w k ) 2 of w k w_k w k tends to 0 as k → ∞ k \to\infty k → ∞ and s i ( k ) ≥ ν L , L ˉ α , α ˉ ( r i ( k ) ) ≥ 0 s_i^{(k)} \ge\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(r_i^{(k)}) \ge0 s i ( k ) ≥ ν L , L ˉ α , α ˉ ( r i ( k ) ) ≥ 0 , we have
0 ≤ θ i ( k ) s i ( k ) ≤ ( w k ) 2 → 0 , k → ∞ . 0 \le\theta_i^{(k)}s_i^{(k)} \le(w_k)_2 \to0,\quad k \to\infty. 0 ≤ θ i ( k ) s i ( k ) ≤ ( w k ) 2 → 0 , k → ∞.
For each i i i , consider the limit weight θ i \theta_i θ i . If θ i > 0 \theta_i > 0 θ i > 0 , then θ i ( k m ) → θ i > 0 \theta_i^{(k_m)} \to\theta_i > 0 θ i ( k m ) → θ i > 0 and θ i ( k m ) s i ( k m ) → 0 \theta_i^{(k_m)}s_i^{(k_m)} \to0 θ i ( k m ) s i ( k m ) → 0 , hence s i ( k m ) → 0 s_i^{(k_m)} \to0 s i ( k m ) → 0 . Lower semi-continuity of ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ then gives
0 ≤ ν L , L ˉ α , α ˉ ( r i ) ≤ lim inf m → ∞ ν L , L ˉ α , α ˉ ( r i ( k m ) ) ≤ lim inf m → ∞ s i ( k m ) = 0 , 0 \le\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(r_i) \le\liminf_{m \to\infty}\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(r_i^{(k_m)}) \le\liminf_{m \to\infty}s_i^{(k_m)} = 0, 0 ≤ ν L , L ˉ α , α ˉ ( r i ) ≤ m → ∞ lim inf ν L , L ˉ α , α ˉ ( r i ( k m ) ) ≤ m → ∞ lim inf s i ( k m ) = 0 ,
so ν L , L ˉ α , α ˉ ( r i ) = 0 \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}}(r_i)=0 ν L , L ˉ α , α ˉ ( r i ) = 0 , and (21) gives r i = 0 r_i=0 r i = 0 . If θ i = 0 \theta_i=0 θ i = 0 , then θ i ( k m ) r i ( k m ) ≤ θ i ( k m ) B max → 0 \theta_i^{(k_m)}r_i^{(k_m)} \le\theta_i^{(k_m)}B_{\max} \to0 θ i ( k m ) r i ( k m ) ≤ θ i ( k m ) B m a x → 0 , so this index contributes 0 to the limit of the first coordinate. Therefore,
ϵ 0 = lim m → ∞ ∑ i = 1 3 θ i ( k m ) r i ( k m ) = 0 , \epsilon_0=\lim_{m\to\infty}\sum_{i=1}^{3}\theta_i^{(k_m)}r_i^{(k_m)}=0, ϵ 0 = m → ∞ lim i = 1 ∑ 3 θ i ( k m ) r i ( k m ) = 0 ,
contradicting ϵ 0 > 0 \epsilon_0>0 ϵ 0 > 0 . Hence ψ L , L ˉ α , α ˉ ( ϵ ) > 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon)>0 ψ L , L ˉ α , α ˉ ( ϵ ) > 0 for every ϵ ∈ ( 0 , B max ] \epsilon\in(0,B_{\max}] ϵ ∈ ( 0 , B m a x ] .
Part 3. Since 0 ≤ ψ L , L ˉ α , α ˉ ≤ ν L , L ˉ α , α ˉ 0\le\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}\le\nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} 0 ≤ ψ L , L ˉ α , α ˉ ≤ ν L , L ˉ α , α ˉ and ν L , L ˉ α , α ˉ \nu_{L,\bar{L}}^{\alpha,\bar{\alpha}} ν L , L ˉ α , α ˉ is finite on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] (proved in Part 2), ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is real-valued on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] .
Let ϵ ∈ ( 0 , B max ) \epsilon\in(0,B_{\max}) ϵ ∈ ( 0 , B m a x ) . Since ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is convex and finite on a neighborhood of ϵ \epsilon ϵ , it is continuous at ϵ \epsilon ϵ ([20 Theorem 10.1] ).
Let ϵ n → 0 + \epsilon_n\to0^+ ϵ n → 0 + . By convexity of ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ and ψ L , L ˉ α , α ˉ ( 0 ) = 0 \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=0 ψ L , L ˉ α , α ˉ ( 0 ) = 0 ,
0 ≤ ψ L , L ˉ α , α ˉ ( ϵ n ) ≤ ϵ n B max ψ L , L ˉ α , α ˉ ( B max ) + ( 1 − ϵ n B max ) ψ L , L ˉ α , α ˉ ( 0 ) = ϵ n B max ψ L , L ˉ α , α ˉ ( B max ) ⟶ 0 , 0\le\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon_n)\le\frac{\epsilon_n}{B_{\max}}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(B_{\max})+\left(1-\frac{\epsilon_n}{B_{\max}}\right)\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)=\frac{\epsilon_n}{B_{\max}}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(B_{\max})\longrightarrow0, 0 ≤ ψ L , L ˉ α , α ˉ ( ϵ n ) ≤ B m a x ϵ n ψ L , L ˉ α , α ˉ ( B m a x ) + ( 1 − B m a x ϵ n ) ψ L , L ˉ α , α ˉ ( 0 ) = B m a x ϵ n ψ L , L ˉ α , α ˉ ( B m a x ) ⟶ 0 ,
where finiteness of ψ L , L ˉ α , α ˉ ( B max ) \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(B_{\max}) ψ L , L ˉ α , α ˉ ( B m a x ) is used in the last step. Hence ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is continuous at 0.
Let ϵ n → B max − \epsilon_n\to B_{\max}^{-} ϵ n → B m a x − . Convexity gives
ψ L , L ˉ α , α ˉ ( ϵ n ) ≤ B max − ϵ n B max ψ L , L ˉ α , α ˉ ( 0 ) + ϵ n B max ψ L , L ˉ α , α ˉ ( B max ) ⟶ ψ L , L ˉ α , α ˉ ( B max ) . \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\epsilon_n)\le\frac{B_{\max}-\epsilon_n}{B_{\max}}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(0)+\frac{\epsilon_n}{B_{\max}}\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(B_{\max})\longrightarrow\psi_{L,\bar{L}}^{\alpha,\bar{\alpha}}(B_{\max}). ψ L , L ˉ α , α ˉ ( ϵ n ) ≤ B m a x B m a x − ϵ n ψ L , L ˉ α , α ˉ ( 0 ) + B m a x ϵ n ψ L , L ˉ α , α ˉ ( B m a x ) ⟶ ψ L , L ˉ α , α ˉ ( B m a x ) .
Since ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is lower semi-continuous (proved in Part 2), the lower semi-continuity gives the reverse inequality in the limit. Hence ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is continuous at B max B_{\max} B m a x as well.
Therefore ψ L , L ˉ α , α ˉ \psi_{L,\bar{L}}^{\alpha,\bar{\alpha}} ψ L , L ˉ α , α ˉ is continuous on [ 0 , B max ] [0,B_{\max}] [ 0 , B m a x ] .
Other Surrogate Loss Functions with FAFC
The examples in this subsection, together with the logistic example in § 5.1 , cover all four combinations of the fairness criterion (DP or EO) and the feature setting (attribute-blind or attribute-aware): § 5.1 treats DP in the attribute-blind setting, § 8.6.1 treats DP in the attribute-aware setting, § 8.6.2 treats EO in the attribute-blind setting, and § 8.6.3 treats EO in the attribute-aware setting.
Throughout, we write Φ ( η , η ˉ ) = ∂ t C f a i r L , L ˉ ( η , η ˉ , t ) ∣ t = 0 \Phi(\eta,\bar{\eta})=\left.\partial_t C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t)\right|_{t=0} Φ ( η , η ˉ ) = ∂ t C fair L , L ˉ ( η , η ˉ , t ) t = 0 for the derivative of the conditional risk at the origin. Unlike the logistic surrogate loss pair in § 5.1 , which comes from a specific published method and therefore carries its own class and group costs, the three examples below are constructed for illustration and are not tied to a particular published method. We therefore treat cost parameters α , α ˉ \alpha,\bar{\alpha} α , α ˉ as free parameters in ( 0 , 1 ) (0,1) ( 0 , 1 ) and state the FAFC and well-posedness results for the whole weighted family. The general calibration inequality, however, does not lead to a simple bound when α \alpha α and α ˉ \bar{\alpha} α ˉ are kept general. To keep the calibration bounds explicit, we therefore compute them in the symmetric case α = α ˉ = 1 2 \alpha=\bar{\alpha}=\frac{1}{2} α = α ˉ = 2 1 .
Fairness-Aware Weighted Logistic Loss
In this example, we consider the DP criterion in the attribute-aware setting. Recall from § 3.1.2 that in this setting the classifier has access to ( X , A ) (X,A) ( X , A ) ; for a fixed instance ( x , a ) (x,a) ( x , a ) , the positive-class posterior is η = η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) \eta=\eta(x,a)=\mathbb{P}(Y=1\mid X=x,A=a) η = η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) , and the sensitive-attribute posterior collapses to η ˉ = 1 { a = 1 } \bar{\eta} = \mathbf{1}\{a = 1\} η ˉ = 1 { a = 1 } .
For the accuracy cost parameter α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) and the fairness cost parameter α ˉ ∈ ( 0 , 1 ) \bar{\alpha} \in(0,1) α ˉ ∈ ( 0 , 1 ) , we define the partial loss functions of the accuracy surrogate loss L L L and the fairness surrogate loss L ˉ \bar{L} L ˉ as weighted logistic losses:
L 1 ( t ) = ( 1 − α ) log ( 1 + e − t ) , L − 1 ( t ) = α log ( 1 + e t ) , L_{1}(t) = (1-\alpha)\log(1+e^{-t}), \qquad L_{-1}(t) = \alpha\log(1+e^{t}), L 1 ( t ) = ( 1 − α ) log ( 1 + e − t ) , L − 1 ( t ) = α log ( 1 + e t ) ,
L ˉ 1 ( t ) = ( 1 − α ˉ ) log ( 1 + e − t ) , L ˉ − 1 ( t ) = α ˉ log ( 1 + e t ) . \bar{L}_{1}(t) = (1-\bar{\alpha})\log(1+e^{-t}), \qquad\bar{L}_{-1}(t) = \bar{\alpha}\log(1+e^{t}). L ˉ 1 ( t ) = ( 1 − α ˉ ) log ( 1 + e − t ) , L ˉ − 1 ( t ) = α ˉ log ( 1 + e t ) .
This gives the compact form of the fairness-aware surrogate loss:
L f a i r ( y , a , t ) = [ 1 + y ( 1 − 2 α ) 2 ] log ( 1 + e − y t ) − λ [ 1 + a ( 1 − 2 α ˉ ) 2 ] log ( 1 + e − a t ) . L_{\mathrm{fair}}(y,a,t) = \left[\frac{1+y(1-2\alpha)}{2}\right]\log(1+e^{-yt})-\lambda\left[\frac{1+a(1-2\bar{\alpha})}{2}\right]\log(1+e^{-at}). L fair ( y , a , t ) = [ 2 1 + y ( 1 − 2 α ) ] log ( 1 + e − y t ) − λ [ 2 1 + a ( 1 − 2 α ˉ ) ] log ( 1 + e − a t ) .
The following proposition proves that the composite loss L f a i r L_{\mathrm{fair}} L fair is FAFC.
Proposition 25. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) , let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be the weighted logistic surrogate pair defined above, let λ max w l o g \lambda_{\max}^{\mathrm{wlog}} λ m a x wlog be given in (22) , and let K \mathcal{K} K be the set of ( η , a ) ∈ [ 0 , 1 ] × { − 1 , 1 } (\eta,a) \in[0,1] \times\{-1,1\} ( η , a ) ∈ [ 0 , 1 ] × { − 1 , 1 } on which t ↦ C D P L , L ˉ ( η , a , t ) t \mapsto C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t) t ↦ C DP L , L ˉ ( η , a , t ) is bounded below. If 0 < λ < λ max w l o g 0 < \lambda< \lambda_{\max}^{\mathrm{wlog}} 0 < λ < λ m a x wlog , then L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the DP criterion in the attribute-aware setting.
Proof. Substituting the partial losses into the attribute-aware conditional risk gives
C D P L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) ] − λ [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 } ) L ˉ − 1 ( t ) ] . C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t) = [\eta L_{1}(t)+(1-\eta)L_{-1}(t)]-\lambda[\mathbf{1}\{a=1\}\bar{L}_{1}(t)+(1-\mathbf{1}\{a=1\})\bar{L}_{-1}(t)]. C DP L , L ˉ ( η , a , t ) = [ η L 1 ( t ) + ( 1 − η ) L − 1 ( t )] − λ [ 1 { a = 1 } L ˉ 1 ( t ) + ( 1 − 1 { a = 1 }) L ˉ − 1 ( t )] .
By the sufficiency direction of Theorem 14 it suffices to check (D1)–(D3) and strict convexity. The partial losses have derivatives at the origin
L 1 ′ ( 0 ) = − 1 2 ( 1 − α ) , L − 1 ′ ( 0 ) = 1 2 α , L ˉ 1 ′ ( 0 ) = − 1 2 ( 1 − α ˉ ) , L ˉ − 1 ′ ( 0 ) = 1 2 α ˉ . L_{1}'(0)=-\frac{1}{2}(1-\alpha), \qquad L_{-1}'(0)=\frac{1}{2}\alpha, \qquad\bar{L}_{1}'(0)=-\frac{1}{2}(1-\bar{\alpha}), \qquad\bar{L}_{-1}'(0)=\frac{1}{2}\bar{\alpha}. L 1 ′ ( 0 ) = − 2 1 ( 1 − α ) , L − 1 ′ ( 0 ) = 2 1 α , L ˉ 1 ′ ( 0 ) = − 2 1 ( 1 − α ˉ ) , L ˉ − 1 ′ ( 0 ) = 2 1 α ˉ .
(D1) holds because L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 1 2 < 0 L_{1}'(0)-L_{-1}'(0)=-\frac{1}{2}<0 L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 2 1 < 0 . (D2) holds because both derivative gaps equal − 1 2 -\frac{1}{2} − 2 1 . (D3) holds because
( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = 1 2 α ( 1 − α ) − 1 2 α ( 1 − α ) = 0 , (1-\alpha)L_{-1}'(0)+\alpha L_{1}'(0)=\frac{1}{2}\alpha(1-\alpha)-\frac{1}{2}\alpha(1-\alpha)=0, ( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = 2 1 α ( 1 − α ) − 2 1 α ( 1 − α ) = 0 ,
and identically for L ˉ \bar{L} L ˉ with α ˉ \bar{\alpha} α ˉ in place of α \alpha α . Hence
( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = λ [ ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 ) ] . (1-\alpha)L'_{-1}(0)+\alpha L'_1(0)=\lambda[(1-\bar{\alpha})\bar{L}'_{-1}(0)+\bar{\alpha}\bar{L}'_1(0)]. ( 1 − α ) L − 1 ′ ( 0 ) + α L 1 ′ ( 0 ) = λ [( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) + α ˉ L ˉ 1 ′ ( 0 )] .
For the strict convexity, write η ˉ = 1 { a = 1 } \bar{\eta}=1\{a=1\} η ˉ = 1 { a = 1 } and collect the two logarithmic terms of the conditional risk:
C D P L , L ˉ ( η , a , t ) = c 1 log ( 1 + e − t ) + c 2 log ( 1 + e t ) , C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t)=c_1\log(1+e^{-t})+c_2\log(1+e^t), C DP L , L ˉ ( η , a , t ) = c 1 log ( 1 + e − t ) + c 2 log ( 1 + e t ) ,
where c 1 : = η ( 1 − α ) − λ η ˉ ( 1 − α ˉ ) c_1:=\eta(1-\alpha)-\lambda\bar{\eta}(1-\bar{\alpha}) c 1 := η ( 1 − α ) − λ η ˉ ( 1 − α ˉ ) , c 2 : = ( 1 − η ) α − λ ( 1 − η ˉ ) α ˉ c_2:=(1-\eta)\alpha-\lambda(1-\bar{\eta})\bar{\alpha} c 2 := ( 1 − η ) α − λ ( 1 − η ˉ ) α ˉ .
Its second derivative is ( c 1 + c 2 ) e t / ( 1 + e t ) 2 (c_1+c_2)e^t/(1+e^t)^2 ( c 1 + c 2 ) e t / ( 1 + e t ) 2 , which is positive whenever c 1 + c 2 > 0 c_1+c_2>0 c 1 + c 2 > 0 . On K \mathcal{K} K , c 1 , c 2 ≥ 0 c_1,c_2\ge0 c 1 , c 2 ≥ 0 by definition. If c 1 + c 2 = 0 c_1+c_2=0 c 1 + c 2 = 0 at some point of K \mathcal{K} K , then c 1 = c 2 = 0 c_1=c_2=0 c 1 = c 2 = 0 , which would force one of the two ratios in (22) to equal λ \lambda λ , contradicting 0 < λ < λ max w l o g 0<\lambda<\lambda_{\max}^{\mathrm{wlog}} 0 < λ < λ m a x wlog . Hence c 1 + c 2 > 0 c_1+c_2>0 c 1 + c 2 > 0 on K \mathcal{K} K , and the conditional risk is strictly convex there. □ \square □
The derivative conditions and strict convexity on K \mathcal{K} K are therefore verified; the well-posedness is checked in the next paragraph, and the sufficiency direction of Theorem 14 yields that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the DP criterion in the attribute-aware setting.
It remains to verify well-posedness. Since log ( 1 + e − t ) = o ( t ) \log(1+e^{-t})=o(t) log ( 1 + e − t ) = o ( t ) and log ( 1 + e t ) = t + o ( t ) \log(1+e^t)=t+o(t) log ( 1 + e t ) = t + o ( t ) as t → + ∞ t\to+\infty t → + ∞ , while log ( 1 + e − t ) = − t + o ( t ) \log(1+e^{-t})=-t+o(t) log ( 1 + e − t ) = − t + o ( t ) and log ( 1 + e t ) = o ( t ) \log(1+e^t)=o(t) log ( 1 + e t ) = o ( t ) as t → − ∞ t\to-\infty t → − ∞ , its tails are
C D P L , L ˉ ( η , a , t ) = { c 2 t + o ( t ) , t → + ∞ , − c 1 t + o ( t ) , t → − ∞ . C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t)=
\begin{cases}
c_2t+o(t), & t\to+\infty,\\
-c_1t+o(t), & t\to-\infty.
\end{cases} C DP L , L ˉ ( η , a , t ) = { c 2 t + o ( t ) , − c 1 t + o ( t ) , t → + ∞ , t → − ∞.
so the risk is bounded below exactly when c 1 ≥ 0 c_1\ge0 c 1 ≥ 0 and c 2 ≥ 0 c_2\ge0 c 2 ≥ 0 . Over the support, the two strict inequalities read 0 < λ < λ max w l o g 0<\lambda<\lambda_{\max}^{\mathrm{wlog}} 0 < λ < λ m a x wlog , where
λ max w l o g : = min { inf x ∈ supp ( X ) : a = 1 η ( x , 1 ) ( 1 − α ) 1 − α ˉ , inf x ∈ supp ( X ) : a = − 1 ( 1 − η ( x , − 1 ) ) α α ˉ } . (22) \lambda_{\max}^{\mathrm{wlog}}:=\min\left\{
\inf_{\substack{x\in\operatorname{supp}(X):\\a=1}}\frac{\eta(x,1)(1-\alpha)}{1-\bar{\alpha}},
\inf_{\substack{x\in\operatorname{supp}(X):\\a=-1}}\frac{(1-\eta(x,-1))\alpha}{\bar{\alpha}}
\right\}. \tag*{(22)} λ m a x wlog := min ⎩ ⎨ ⎧ x ∈ supp ( X ) : a = 1 inf 1 − α ˉ η ( x , 1 ) ( 1 − α ) , x ∈ supp ( X ) : a = − 1 inf α ˉ ( 1 − η ( x , − 1 )) α ⎭ ⎬ ⎫ . (22)
By definition of K \mathcal{K} K , c 1 , c 2 ≥ 0 c_1,c_2\ge0 c 1 , c 2 ≥ 0 there, so the risk is bounded below on K \mathcal{K} K . Under (22) , the sufficiency direction of Theorem 14 applies and L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the DP criterion in the attribute-aware setting.
Having verified FAFC, we now derive the calibration function and the excess-risk bound.
We consider the symmetric member of the weighted logistic surrogate loss family, in which
L 1 ( t ) = 1 2 log ( 1 + e − t ) , L − 1 ( t ) = 1 2 log ( 1 + e t ) , L_1(t) = \frac{1}{2}\log(1+e^{-t}), \qquad L_{-1}(t) = \frac{1}{2}\log(1+e^t), L 1 ( t ) = 2 1 log ( 1 + e − t ) , L − 1 ( t ) = 2 1 log ( 1 + e t ) ,
L ˉ 1 ( t ) = 1 2 log ( 1 + e − t ) , L ˉ − 1 ( t ) = 1 2 log ( 1 + e t ) . \bar{L}_1(t) = \frac{1}{2}\log(1+e^{-t}), \qquad\bar{L}_{-1}(t) = \frac{1}{2}\log(1+e^t). L ˉ 1 ( t ) = 2 1 log ( 1 + e − t ) , L ˉ − 1 ( t ) = 2 1 log ( 1 + e t ) .
Up to the common positive factor 1 2 \frac{1}{2} 2 1 , this is the standard unweighted logistic loss for both the accuracy and fairness terms.
Proposition 26. For the symmetric member of the weighted logistic family ( α = α ˉ = 1 2 ) (\alpha= \bar{\alpha} = \frac{1}{2}) ( α = α ˉ = 2 1 ) , suppose that 0 < λ < λ max w l o g 0 < \lambda< \lambda_{\max}^{\mathrm{wlog}} 0 < λ < λ m a x wlog , with λ max w l o g \lambda_{\max}^{\mathrm{wlog}} λ m a x wlog given in (22) , and let K \mathcal{K} K be the well-posed region of Proposition 25 . Then, on K \mathcal{K} K , the calibration function of the pair ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) for the DP criterion in the attribute-aware setting satisfies
ψ L , L ˉ 1 2 , 1 2 ( ϵ ) ≥ ϵ 2 1 − λ (23) \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon) \ge\frac{\epsilon^2}{1-\lambda} \tag*{(23)} ψ L , L ˉ 2 1 , 2 1 ( ϵ ) ≥ 1 − λ ϵ 2 (23) and, consequently, for every f ∈ F f \in\mathcal{F} f ∈ F ,
R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ min { 1 − λ 2 , ( 1 − λ ) ( R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ ) } . (24) \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f) - \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*} \le\min\left\{\frac{1-\lambda}{2},\sqrt{(1-\lambda)\left(\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}\right)}\right\}. \tag*{(24)} R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ min { 2 1 − λ , ( 1 − λ ) ( R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ ) } . (24)
Proof. Write w 1 : = η − λ 1 { a = 1 } w_1 := \eta-\lambda1\{a=1\} w 1 := η − λ 1 { a = 1 } , w 2 : = ( 1 − η ) − λ 1 { a = − 1 } w_2 := (1-\eta)-\lambda1\{a=-1\} w 2 := ( 1 − η ) − λ 1 { a = − 1 } . The conditional risk is
C D P L , L ˉ ( η , a , t ) = 1 2 w 1 log ( 1 + e − t ) + 1 2 w 2 log ( 1 + e t ) , C_{\mathrm{DP}}^{L,\bar{L}}(\eta,a,t) = \frac{1}{2}w_1\log(1+e^{-t})+\frac{1}{2}w_2\log(1+e^t), C DP L , L ˉ ( η , a , t ) = 2 1 w 1 log ( 1 + e − t ) + 2 1 w 2 log ( 1 + e t ) ,
Its derivative with respect to t t t is 1 2 [ ( w 1 + w 2 ) σ ( t ) − w 1 ] \frac{1}{2}[(w_1+w_2)\sigma(t)-w_1] 2 1 [( w 1 + w 2 ) σ ( t ) − w 1 ] , where σ ( t ) = 1 / ( 1 + e − t ) \sigma(t)=1/(1+e^{-t}) σ ( t ) = 1/ ( 1 + e − t ) . Under (22) , w 1 , w 2 > 0 w_1,w_2>0 w 1 , w 2 > 0 on the support, so the conditional risk is strictly convex and its unique minimizer is characterized by σ ( t ∗ ) = w 1 / ( w 1 + w 2 ) \sigma(t^*)=w_1/(w_1+w_2) σ ( t ∗ ) = w 1 / ( w 1 + w 2 ) . Let
S ~ : = σ ( t ~ ∗ ) = η − λ 1 { a = 1 } 1 − λ . \widetilde{S}:=\sigma(\widetilde{t}^*)=\frac{\eta-\lambda1\{a=1\}}{1-\lambda}. S := σ ( t ∗ ) = 1 − λ η − λ 1 { a = 1 } .
Since w 1 + w 2 = 1 − λ w_1+w_2=1-\lambda w 1 + w 2 = 1 − λ ,
C D P L , L ˉ , ∗ ( η , a ) = − 1 2 [ w 1 log S ~ + w 2 log ( 1 − S ~ ) ] , C_{\mathrm{DP}}^{L,\bar{L},*}(\eta,a)=-\frac{1}{2}\left[w_1\log\widetilde{S}+w_2\log(1-\widetilde{S})\right], C DP L , L ˉ , ∗ ( η , a ) = − 2 1 [ w 1 log S + w 2 log ( 1 − S ) ] ,
To compute C D P L ˉ , L ˉ , − ( η , a ) C_{\mathrm{DP}}^{\bar{L},\bar{L},-}(\eta,a) C DP L ˉ , L ˉ , − ( η , a ) , we first compute the discriminant D 1 2 , 1 2 ( η , a ) D^{\frac{1}{2},\frac{1}{2}}(\eta,a) D 2 1 , 2 1 ( η , a ) . By (11) ,
D 1 2 , 1 2 ( η , a ) = A D P 1 2 , 1 2 ( η , a ) − B D P 1 2 , 1 2 ( η , a ) = η − 1 2 − λ 1 { a = 1 } + λ 2 . D^{\frac{1}{2},\frac{1}{2}}(\eta,a)=A_{\mathrm{DP}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)-B_{\mathrm{DP}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\eta-\frac{1}{2}-\lambda1\{a=1\}+\frac{\lambda}{2}. D 2 1 , 2 1 ( η , a ) = A DP 2 1 , 2 1 ( η , a ) − B DP 2 1 , 2 1 ( η , a ) = η − 2 1 − λ 1 { a = 1 } + 2 λ .
Since S ~ = η − λ 1 { a = 1 } 1 − λ \widetilde{S}=\frac{\eta-\lambda1\{a=1\}}{1-\lambda} S = 1 − λ η − λ 1 { a = 1 } , we have
D 1 2 , 1 2 ( η , a ) = ( 1 − λ ) ( S ~ − 1 2 ) , D^{\frac{1}{2},\frac{1}{2}}(\eta,a)=(1-\lambda)\left(\widetilde{S}-\frac{1}{2}\right), D 2 1 , 2 1 ( η , a ) = ( 1 − λ ) ( S − 2 1 ) ,
which means the unique minimizer lies on the correct side of the origin. Hence
C D P L ˉ , L ˉ , − ( η , a ) = C D P L ˉ , L ˉ ( η , a , 0 ) = 1 − λ 2 log 2. C_{\mathrm{DP}}^{\bar{L},\bar{L},-}(\eta,a)=C_{\mathrm{DP}}^{\bar{L},\bar{L}}(\eta,a,0)=\frac{1-\lambda}{2}\log2. C DP L ˉ , L ˉ , − ( η , a ) = C DP L ˉ , L ˉ ( η , a , 0 ) = 2 1 − λ log 2.
Therefore, the excess conditional risk is
H L , L ˉ 1 2 , 1 2 ( η , a ) = 1 − λ 2 [ log 2 + S ~ log S ~ + ( 1 − S ~ ) log ( 1 − S ~ ) ] = 1 − λ 2 D K L ( S ~ ∥ 1 2 ) . H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\frac{1-\lambda}{2}\left[\log2+\widetilde{S}\log\widetilde{S}+(1-\widetilde{S})\log(1-\widetilde{S})\right]=\frac{1-\lambda}{2}D_{\mathrm{KL}}\left(\widetilde{S}\mathbin{\|}\frac{1}{2}\right). H L , L ˉ 2 1 , 2 1 ( η , a ) = 2 1 − λ [ log 2 + S log S + ( 1 − S ) log ( 1 − S ) ] = 2 1 − λ D KL ( S ∥ 2 1 ) .
Since D 1 2 , 1 2 ( η , a ) = ( 1 − λ ) ( S ~ − 1 2 ) D^{\frac{1}{2},\frac{1}{2}}(\eta,a)=(1-\lambda)(\widetilde{S}-\frac{1}{2}) D 2 1 , 2 1 ( η , a ) = ( 1 − λ ) ( S − 2 1 ) , on a discriminant level set ∣ D 1 2 , 1 2 ( η , a ) ∣ = ϵ \left|D^{\frac{1}{2},\frac{1}{2}}(\eta,a)\right|=\epsilon D 2 1 , 2 1 ( η , a ) = ϵ we have S ~ = 1 2 ± ϵ / ( 1 − λ ) \widetilde{S}=\frac{1}{2}\pm\epsilon/(1-\lambda) S = 2 1 ± ϵ / ( 1 − λ ) . The two signs give the same value because S ~ ↦ D K L ( S ~ ∥ 1 2 ) \widetilde{S}\mapsto D_{\mathrm{KL}}(\widetilde{S}\mathbin{\|}\frac{1}{2}) S ↦ D KL ( S ∥ 2 1 ) is symmetric about 1 2 \frac{1}{2} 2 1 , so the level-set minimum is
ν L , L ˉ 1 2 , 1 2 ( ϵ ) = 1 − λ 2 D K L ( 1 2 + ϵ 1 − λ ∥ 1 2 ) . \nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\frac{1-\lambda}{2}D_{\mathrm{KL}}\left(\frac{1}{2}+\frac{\epsilon}{1-\lambda}\mathbin{\|}\frac{1}{2}\right). ν L , L ˉ 2 1 , 2 1 ( ϵ ) = 2 1 − λ D KL ( 2 1 + 1 − λ ϵ ∥ 2 1 ) .
Set g ( u ) : = D K L ( 1 2 + u ∥ 1 2 ) g(u):=D_{\mathrm{KL}}\left(\frac{1}{2}+u\mathbin{\|}\frac{1}{2}\right) g ( u ) := D KL ( 2 1 + u ∥ 2 1 ) . g g g is convex and increasing, so ν L , L ˉ 1 2 , 1 2 ( ϵ ) = 1 − λ 2 g ( ϵ / ( 1 − λ ) ) \nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\frac{1-\lambda}{2}g(\epsilon/(1-\lambda)) ν L , L ˉ 2 1 , 2 1 ( ϵ ) = 2 1 − λ g ( ϵ / ( 1 − λ )) is convex and increasing as well. Hence ψ L , L ˉ 1 2 , 1 2 = ν L , L ˉ 1 2 , 1 2 \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}=\nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}} ψ L , L ˉ 2 1 , 2 1 = ν L , L ˉ 2 1 , 2 1 . Finally, by Pinsker’s inequality
D K L ( p ∥ q ) ≥ 2 ( p − q ) 2 , D_{\mathrm{KL}}(p\mathbin{\|}q)\geq2(p-q)^2, D KL ( p ∥ q ) ≥ 2 ( p − q ) 2 ,
ψ L , L ˉ 1 2 , 1 2 ( ϵ ) = 1 − λ 2 g ( ϵ 1 − λ ) ≥ ϵ 2 1 − λ , \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\frac{1-\lambda}{2}g\left(\frac{\epsilon}{1-\lambda}\right)\geq\frac{\epsilon^2}{1-\lambda}, ψ L , L ˉ 2 1 , 2 1 ( ϵ ) = 2 1 − λ g ( 1 − λ ϵ ) ≥ 1 − λ ϵ 2 ,
which is (23) .
By Theorem 17 and Remark 18 , (23) gives R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ ( 1 − λ ) ( R f a i r L ˉ , L ˉ ( f ) − R f a i r L ˉ , L ˉ , ∗ ) \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f)-\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*}\leq\sqrt{(1-\lambda)(\mathcal{R}_{\mathrm{fair}}^{\bar{L},\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{\bar{L},\bar{L},*})} R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ ( 1 − λ ) ( R fair L ˉ , L ˉ ( f ) − R fair L ˉ , L ˉ , ∗ ) .
In addition, since S ~ ∈ ( 0 , 1 ) \widetilde{S}\in(0,1) S ∈ ( 0 , 1 ) , we have ∣ D 1 2 , 1 2 ( η , a ) ∣ < ( 1 − λ ) / 2 \left|D^{\frac{1}{2},\frac{1}{2}}(\eta,a)\right|<(1-\lambda)/2 D 2 1 , 2 1 ( η , a ) < ( 1 − λ ) /2 . By Lemma 22 , the trivial bound ϵ ≤ 1 − λ 2 \epsilon\leq\frac{1-\lambda}{2} ϵ ≤ 2 1 − λ holds.
Taking the smaller of the two bounds gives (24) .
Fairness-Aware Hinge Loss
The Support Vector Machine (SVM) is one of the most classic classifiers in statistical learning, with the hinge loss as its standard loss function. In this example we consider the EO criterion in the attribute-blind setting, where η ˉ = η ˉ E O ( x ) \bar{\eta}=\bar{\eta}_{\mathrm{EO}}(x) η ˉ = η ˉ EO ( x ) is the sensitive-attribute posterior among the positive class.
Let the accuracy cost parameter be α ∈ ( 0 , 1 ) \alpha\in(0,1) α ∈ ( 0 , 1 ) , the fairness cost parameter be α ˉ ∈ ( 0 , 1 ) \bar{\alpha} \in(0,1) α ˉ ∈ ( 0 , 1 ) , and define the partial losses
L 1 ( t ) = ( 1 − α ) ( 1 − t ) + , L − 1 ( t ) = α ( 1 + t ) + , L ˉ 1 ( t ) = − ( 1 − α ˉ ) t , L ˉ − 1 ( t ) = α ˉ t , (25) L_1(t)=(1-\alpha)(1-t)_+,\quad L_{-1}(t)=\alpha(1+t)_+,\quad\bar{L}_1(t)=-(1-\bar{\alpha})t,\quad\bar{L}_{-1}(t)=\bar{\alpha}t, \tag*{(25)} L 1 ( t ) = ( 1 − α ) ( 1 − t ) + , L − 1 ( t ) = α ( 1 + t ) + , L ˉ 1 ( t ) = − ( 1 − α ˉ ) t , L ˉ − 1 ( t ) = α ˉ t , (25)
where ( z ) + = max { z , 0 } (z)_+=\max\{z,0\} ( z ) + = max { z , 0 } : the accuracy surrogate is the cost-sensitive hinge of the classical theory, and the fairness surrogate is linear.
The following proposition proves that the composite loss L f a i r L_{\mathrm{fair}} L fair is FAFC.
Proposition 27. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha} \in(0,1) α , α ˉ ∈ ( 0 , 1 ) , let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be the surrogate loss pair in (25) , let λ \lambda λ be given in (30) , and let K \mathcal{K} K be the set of ( η , η ˉ ) ∈ [ 0 , 1 ] 2 (\eta,\bar{\eta}) \in[0,1]^2 ( η , η ˉ ) ∈ [ 0 , 1 ] 2 on which t ↦ C f a i r L , L ˉ ( η , η ˉ , t ) t \mapsto C_{\mathrm{fair}}^{L,\bar{L}}(\eta,\bar{\eta},t) t ↦ C fair L , L ˉ ( η , η ˉ , t ) is bounded below. If 0 < λ < λ max s v m 0 < \lambda< \lambda_{\max}^{\mathrm{svm}} 0 < λ < λ m a x svm , then L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the EO criterion in the attribute-blind setting.
Proof. The hinge is piecewise linear and hence not strictly convex, so the strict-convexity premise of Theorem 14 is not met and L f a i r L_{\mathrm{fair}} L fair cannot be certified through that theorem. We therefore verify FAFC directly through Definition 12 , which requires no convexity.
The attribute-blind EO conditional risk is
C E O L , L ˉ ( η , η ˉ , t ) = η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) − λ π − 1 η [ η ˉ L ˉ 1 ( t ) + ( 1 − η ˉ ) L ˉ − 1 ( t ) ] . C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},t)=\eta L_1(t)+(1-\eta)L_{-1}(t)-\lambda\pi^{-1}\eta\left[\bar{\eta}\bar{L}_1(t)+(1-\bar{\eta})\bar{L}_{-1}(t)\right]. C EO L , L ˉ ( η , η ˉ , t ) = η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) − λ π − 1 η [ η ˉ L ˉ 1 ( t ) + ( 1 − η ˉ ) L ˉ − 1 ( t ) ] .
With
θ : = λ π − 1 η ( η ˉ − α ˉ ) , \theta:= \lambda\pi^{-1}\eta(\bar{\eta}-\bar{\alpha}), θ := λ π − 1 η ( η ˉ − α ˉ ) ,
the conditional risk becomes
C E O L , L ˉ ( η , η ˉ , t ) = η ( 1 − α ) ( 1 − t ) + + ( 1 − η ) α ( 1 + t ) + + θ t . (26) C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},t)=\eta(1-\alpha)(1-t)_{+}+(1-\eta)\alpha(1+t)_{+}+\theta t. \tag*{(26)} C EO L , L ˉ ( η , η ˉ , t ) = η ( 1 − α ) ( 1 − t ) + + ( 1 − η ) α ( 1 + t ) + + θ t . (26)
The function t ↦ C E O L , L ˉ ( η , η ˉ , t ) t\mapsto C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},t) t ↦ C EO L , L ˉ ( η , η ˉ , t ) of (26) is affine on each of the three intervals cut out by t = ± 1 t=\pm1 t = ± 1 . Expanding ( 1 − t ) + (1-t)_{+} ( 1 − t ) + and ( 1 + t ) + (1+t)_{+} ( 1 + t ) + on these intervals,
C E O L , L ˉ ( η , η ˉ , t ) = { η ( 1 − α ) + σ − t , t ≤ − 1 , η ( 1 − α ) + ( 1 − η ) α + σ 0 t , − 1 ≤ t ≤ 1 , ( 1 − η ) α + σ + t , t ≥ 1 , C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},t)=
\begin{cases}
\eta(1-\alpha)+\sigma_{-}t, & t\le-1,\\
\eta(1-\alpha)+(1-\eta)\alpha+\sigma_{0}t, & -1\le t\le1,\\
(1-\eta)\alpha+\sigma_{+}t, & t\ge1,
\end{cases} C EO L , L ˉ ( η , η ˉ , t ) = ⎩ ⎨ ⎧ η ( 1 − α ) + σ − t , η ( 1 − α ) + ( 1 − η ) α + σ 0 t , ( 1 − η ) α + σ + t , t ≤ − 1 , − 1 ≤ t ≤ 1 , t ≥ 1 ,
where
σ − = − η ( 1 − α ) + θ , σ 0 = ( α − η ) + θ = − D α , α ˉ ( η , η ˉ ) , σ + = ( 1 − η ) α + θ . (27) \sigma_{-}=-\eta(1-\alpha)+\theta,\qquad\sigma_{0}=(\alpha-\eta)+\theta=-D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}),\qquad\sigma_{+}=(1-\eta)\alpha+\theta. \tag*{(27)} σ − = − η ( 1 − α ) + θ , σ 0 = ( α − η ) + θ = − D α , α ˉ ( η , η ˉ ) , σ + = ( 1 − η ) α + θ . (27)
Since
σ 0 − σ − = α ( 1 − η ) ≥ 0 , σ + − σ 0 = η ( 1 − α ) ≥ 0 , \sigma_{0}-\sigma_{-}=\alpha(1-\eta)\ge0,\qquad\sigma_{+}-\sigma_{0}=\eta(1-\alpha)\ge0, σ 0 − σ − = α ( 1 − η ) ≥ 0 , σ + − σ 0 = η ( 1 − α ) ≥ 0 ,
we have σ − ≤ σ 0 ≤ σ + \sigma_{-}\le\sigma_{0}\le\sigma_{+} σ − ≤ σ 0 ≤ σ + , i.e. C E O L , L ˉ ( η , η ˉ , ⋅ ) C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},\cdot) C EO L , L ˉ ( η , η ˉ , ⋅ ) is convex in t t t . In particular, when − 1 ≤ t ≤ 1 -1\le t\le1 − 1 ≤ t ≤ 1 ,
C E O L , L ˉ ( η , η ˉ , t ) = C E O L , L ˉ ( η , η ˉ , 0 ) − t D α , α ˉ ( η , η ˉ ) . (28) C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},t)=C^{L,\bar{L}}_{\mathrm{EO}}(\eta,\bar{\eta},0)-tD^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). \tag*{(28)} C EO L , L ˉ ( η , η ˉ , t ) = C EO L , L ˉ ( η , η ˉ , 0 ) − t D α , α ˉ ( η , η ˉ ) . (28)
In particular, the middle branch carries the whole dependence on the discriminant: its slope is − D α , α ˉ -D^{\alpha,\bar{\alpha}} − D α , α ˉ , so the risk is strictly monotone at the origin whenever the discriminant is nonzero, and it is flat precisely on the discriminant boundary.
Assume 0 < λ < λ max s v m 0<\lambda<\lambda_{\max}^{\mathrm{svm}} 0 < λ < λ m a x svm , where λ max s v m \lambda_{\max}^{\mathrm{svm}} λ m a x svm is defined in (30) . Then the slopes satisfy σ − < 0 < σ + \sigma_{-}<0<\sigma_{+} σ − < 0 < σ + and the risk is bounded below.
Suppose first that D α , α ˉ ( η , η ˉ ) > 0 D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})>0 D α , α ˉ ( η , η ˉ ) > 0 . Then σ − ≤ σ 0 = − D α , α ˉ ( η , η ˉ ) < 0 ≤ σ + \sigma_{-}\le\sigma_{0}=-D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})<0\le\sigma_{+} σ − ≤ σ 0 = − D α , α ˉ ( η , η ˉ ) < 0 ≤ σ + , so the conditional risk is strictly decreasing on ( − ∞ , 1 ) (-\infty,1) ( − ∞ , 1 ) and non-decreasing on ( 1 , ∞ ) (1,\infty) ( 1 , ∞ ) . Hence its minimizers are t = 1 t=1 t = 1 , or the whole interval [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) in the boundary case σ + = 0 \sigma_{+}=0 σ + = 0 .
On the wrong-sign set { t : t D α , α ˉ ≤ 0 } = ( − ∞ , 0 ] \{t : tD^{\alpha,\bar{\alpha}} \le0\}=(-\infty,0] { t : t D α , α ˉ ≤ 0 } = ( − ∞ , 0 ] the risk is strictly decreasing, so the wrong-sign infimum is attained at t = 0 t=0 t = 0 . Since C E O L , L ˉ ( η , η ˉ , 1 ) = 2 ( 1 − η ) α + θ C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},1)=2(1-\eta)\alpha+\theta C EO L , L ˉ ( η , η ˉ , 1 ) = 2 ( 1 − η ) α + θ and C E O L , L ˉ ( η , η ˉ , 0 ) = η ( 1 − α ) + ( 1 − η ) α C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},0)=\eta(1-\alpha)+(1-\eta)\alpha C EO L , L ˉ ( η , η ˉ , 0 ) = η ( 1 − α ) + ( 1 − η ) α , we get
H L , L ˉ α , α ˉ ( η , η ˉ ) = C E O L , L ˉ ( η , η ˉ , 0 ) − C E O L , L ˉ ( η , η ˉ , 1 ) = ( η − α ) − θ = D α , α ˉ ( η , η ˉ ) . H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},0)-C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},1)=(\eta-\alpha)-\theta=D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta}). H L , L ˉ α , α ˉ ( η , η ˉ ) = C EO L , L ˉ ( η , η ˉ , 0 ) − C EO L , L ˉ ( η , η ˉ , 1 ) = ( η − α ) − θ = D α , α ˉ ( η , η ˉ ) .
The case D α , α ˉ < 0 D^{\alpha,\bar{\alpha}}<0 D α , α ˉ < 0 is symmetric: the risk is strictly increasing at the origin, its minimizers are t = − 1 t=-1 t = − 1 , or ( − ∞ , − 1 ] (-\infty,-1] ( − ∞ , − 1 ] when σ − = 0 \sigma_{-}=0 σ − = 0 , the wrong-sign infimum is attained at t = 0 t=0 t = 0 , and C E O L , L ˉ ( η , η ˉ , − 1 ) = 2 η ( 1 − α ) − θ C_{\mathrm{EO}}^{L,\bar{L}}(\eta,\bar{\eta},-1)=2\eta(1-\alpha)-\theta C EO L , L ˉ ( η , η ˉ , − 1 ) = 2 η ( 1 − α ) − θ gives H L , L ˉ α , α ˉ = − D α , α ˉ H_{L,\bar{L}}^{\alpha,\bar{\alpha}}=-D^{\alpha,\bar{\alpha}} H L , L ˉ α , α ˉ = − D α , α ˉ .
Finally, if D α , α ˉ = 0 D^{\alpha,\bar{\alpha}}=0 D α , α ˉ = 0 then the middle branch (28) is flat on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] , so t = 0 t=0 t = 0 is a minimizer and H L , L ˉ α , α ˉ = 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}=0 H L , L ˉ α , α ˉ = 0 . Combining the three cases yields
H L , L ˉ α , α ˉ ( η , η ˉ ) = ∣ D α , α ˉ ( η , η ˉ ) ∣ . (29) H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=\left|D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})\right|. \tag*{(29)} H L , L ˉ α , α ˉ ( η , η ˉ ) = D α , α ˉ ( η , η ˉ ) . (29)
Consequently H L , L ˉ α , α ˉ > 0 H_{L,\bar{L}}^{\alpha,\bar{\alpha}}>0 H L , L ˉ α , α ˉ > 0 whenever D α , α ˉ ≠ 0 D^{\alpha,\bar{\alpha}}\ne0 D α , α ˉ = 0 , and Definition 12 yields FAFC on K \mathcal{K} K for L f a i r L_{\mathrm{fair}} L fair . □ \square □
It remains to ensure that the conditional risk is bounded below (Remark 11) and that its pointwise minimizer exists. By (27) the risk is convex with σ − ≤ σ 0 ≤ σ + \sigma_{-}\le\sigma_{0}\le\sigma_{+} σ − ≤ σ 0 ≤ σ + , so it is bounded below if and only if σ − ≤ 0 ≤ σ + \sigma_{-}\le0\le\sigma_{+} σ − ≤ 0 ≤ σ + , that is, if and only if
− ( 1 − η ) α ≤ λ π − 1 η ( η ˉ − α ˉ ) ≤ η ( 1 − α ) -(1-\eta)\alpha\le\lambda\pi^{-1}\eta(\bar{\eta}-\bar{\alpha})\le\eta(1-\alpha) − ( 1 − η ) α ≤ λ π − 1 η ( η ˉ − α ˉ ) ≤ η ( 1 − α )
holds for P X P_X P X -a.e. x x x ; the two inequalities are strict exactly when the risk is coercive, in which case the pointwise minimizer exists and is unique whenever D α , α ˉ ≠ 0 D^{\alpha,\bar{\alpha}}\ne0 D α , α ˉ = 0 (as noted above, on the boundary D α , α ˉ = 0 D^{\alpha,\bar{\alpha}}=0 D α , α ˉ = 0 the risk is flat on [ − 1 , 1 ] [-1,1] [ − 1 , 1 ] ). Written pointwise over the support, the strict inequalities are 0 < λ < λ max s v m 0<\lambda<\lambda_{\max}^{\mathrm{svm}} 0 < λ < λ m a x svm , where
λ max s v m ≔ π min { inf x ∈ supp ( X ) : η ˉ ( x ) > α ˉ 1 − α η ˉ ( x ) − α ˉ , inf x ∈ supp ( X ) : η ˉ ( x ) < α ˉ ( 1 − η ( x ) ) α η ( x ) ( α ˉ − η ˉ ( x ) ) } . (30) \lambda_{\max}^{\mathrm{svm}}\coloneqq\pi\min\left\{\inf_{\substack{x\in\operatorname{supp}(X):\\\bar{\eta}(x)>\bar{\alpha}}}\frac{1-\alpha}{\bar{\eta}(x)-\bar{\alpha}},\ \inf_{\substack{x\in\operatorname{supp}(X):\\\bar{\eta}(x)<\bar{\alpha}}}\frac{(1-\eta(x))\alpha}{\eta(x)(\bar{\alpha}-\bar{\eta}(x))}\right\}. \tag*{(30)} λ m a x svm : = π min ⎩ ⎨ ⎧ x ∈ supp ( X ) : η ˉ ( x ) > α ˉ inf η ˉ ( x ) − α ˉ 1 − α , x ∈ supp ( X ) : η ˉ ( x ) < α ˉ inf η ( x ) ( α ˉ − η ˉ ( x )) ( 1 − η ( x )) α ⎭ ⎬ ⎫ . (30)
with the infimum over an empty set understood as + ∞ +\infty + ∞ . The threshold is support-dependent: both infima are evaluated over the support of X X X , so no support-free constant can replace (30) . At the boundary λ = λ max s v m \lambda= \lambda_{\max}^{\mathrm{svm}} λ = λ m a x svm the risk is still bounded below, but one of the slopes σ ± \sigma_{\pm} σ ± vanishes and the set of minimizers becomes the half-line [ 1 , ∞ ) [1,\infty) [ 1 , ∞ ) or ( − ∞ , − 1 ] (-\infty,-1] ( − ∞ , − 1 ] , whose sign still agrees with D α , α ˉ D^{\alpha,\bar{\alpha}} D α , α ˉ . Under 0 < λ < λ max s v m 0 < \lambda< \lambda_{\max}^{\mathrm{svm}} 0 < λ < λ m a x svm all premises of Definition 12 hold and L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the EO criterion in the attribute-blind setting.
Having verified FAFC, we now derive the calibration function and the excess-risk bound.
We consider the symmetric member of the hinge loss family, in which
L 1 ( t ) = 1 2 ( 1 − t ) + , L − 1 ( t ) = 1 2 ( 1 + t ) + , L ˉ 1 ( t ) = − 1 2 t , L ˉ − 1 ( t ) = 1 2 t . L_1(t)=\frac{1}{2}(1-t)_+,\qquad L_{-1}(t)=\frac{1}{2}(1+t)_+,\qquad\bar{L}_1(t)=-\frac{1}{2}t,\qquad\bar{L}_{-1}(t)=\frac{1}{2}t. L 1 ( t ) = 2 1 ( 1 − t ) + , L − 1 ( t ) = 2 1 ( 1 + t ) + , L ˉ 1 ( t ) = − 2 1 t , L ˉ − 1 ( t ) = 2 1 t .
We record the resulting calibration function and excess-risk bound.
Proposition 28. For the symmetric member of the hinge family ( α = α ˉ = 1 2 ) (\alpha=\bar{\alpha}=\frac{1}{2}) ( α = α ˉ = 2 1 ) , suppose 0 < λ < λ max s v m 0<\lambda<\lambda_{\max}^{\mathrm{svm}} 0 < λ < λ m a x svm , with λ max s v m \lambda_{\max}^{\mathrm{svm}} λ m a x svm given in (30) , and let K \mathcal{K} K be the well-posed region of Proposition 27 . Then, on K \mathcal{K} K , the calibration function of the pair satisfies ψ L , L ˉ 1 2 , 1 2 ( ϵ ) = ϵ \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\epsilon ψ L , L ˉ 2 1 , 2 1 ( ϵ ) = ϵ for 0 ≤ ϵ ≤ 1 2 0\leq\epsilon\leq\frac{1}{2} 0 ≤ ϵ ≤ 2 1 , so it is strictly increasing and invertible there; consequently, for every f ∈ F f\in\mathcal{F} f ∈ F ,
R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ min { 1 2 , R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ } . (31) \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f)-\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*}\leq\min\left\{\frac{1}{2},\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}\right\}. \tag*{(31)} R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ min { 2 1 , R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ } . (31)
Proof. By (29) with α = α ˉ = 1 2 \alpha=\bar{\alpha}=\frac{1}{2} α = α ˉ = 2 1 ,
H L , L ˉ 1 2 , 1 2 ( η , η ˉ ) = ∣ D 1 2 , 1 2 ( η , η ˉ ) ∣ . H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\eta,\bar{\eta})=\left|D^{\frac{1}{2},\frac{1}{2}}(\eta,\bar{\eta})\right|. H L , L ˉ 2 1 , 2 1 ( η , η ˉ ) = D 2 1 , 2 1 ( η , η ˉ ) .
Since D 1 2 , 1 2 ( η , η ˉ ) = η − 1 2 − θ D^{\frac{1}{2},\frac{1}{2}}(\eta,\bar{\eta})=\eta-\frac{1}{2}-\theta D 2 1 , 2 1 ( η , η ˉ ) = η − 2 1 − θ and the well-posedness gives
− 1 − η 2 ≤ θ ≤ η 2 , -\frac{1-\eta}{2}\leq\theta\leq\frac{\eta}{2}, − 2 1 − η ≤ θ ≤ 2 η ,
we have − 1 2 ≤ D 1 2 , 1 2 ( η , η ˉ ) ≤ 1 2 -\frac{1}{2}\leq D^{\frac{1}{2},\frac{1}{2}}(\eta,\bar{\eta})\leq\frac{1}{2} − 2 1 ≤ D 2 1 , 2 1 ( η , η ˉ ) ≤ 2 1 . Hence every level set ∣ D 1 2 , 1 2 ( η , η ˉ ) ∣ = ϵ \left|D^{\frac{1}{2},\frac{1}{2}}(\eta,\bar{\eta})\right|=\epsilon D 2 1 , 2 1 ( η , η ˉ ) = ϵ with 0 ≤ ϵ ≤ 1 2 0\leq\epsilon\leq\frac{1}{2} 0 ≤ ϵ ≤ 2 1 contains a well-posed point, and therefore
ν L , L ˉ 1 2 , 1 2 ( ϵ ) = ψ L , L ˉ 1 2 , 1 2 ( ϵ ) = ϵ , 0 ≤ ϵ ≤ 1 2 . (32) \nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\epsilon,\qquad0\leq\epsilon\leq\frac{1}{2}. \tag*{(32)} ν L , L ˉ 2 1 , 2 1 ( ϵ ) = ψ L , L ˉ 2 1 , 2 1 ( ϵ ) = ϵ , 0 ≤ ϵ ≤ 2 1 . (32)
The calibration inequality of Theorem 17 therefore applies with ψ L , L ˉ 1 2 , 1 2 ( ϵ ) = ϵ \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\epsilon)=\epsilon ψ L , L ˉ 2 1 , 2 1 ( ϵ ) = ϵ and yields
R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ . \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f)-\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*}\leq\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}. R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ .
Finally, by Lemma 22, the same wrong-sign-average representation gives the trivial bound
R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ ∣ D 2 1 2 , 1 2 ∣ ≤ 1 2 . \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f)-\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*}\leq\left|D_{2}^{\frac{1}{2},\frac{1}{2}}\right|\leq\frac{1}{2}. R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ D 2 2 1 , 2 1 ≤ 2 1 .
Taking the smaller of the two bounds gives (31) .
Fairness-Aware Exponential Loss
The exponential loss is another classical surrogate in binary classification (e.g., in boosting).
In this example we consider the EO criterion in the attribute-aware setting, where the accuracy surrogate is an asymmetric exponential loss and the fairness surrogate is linear.
As in § 3.1.2 , the classifier has access to ( X , A ) (X,A) ( X , A ) : for a fixed instance ( x , a ) (x,a) ( x , a ) the positive-class posterior is η = η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) \eta=\eta(x,a)=\mathbb{P}(Y=1\mid X=x,A=a) η = η ( x , a ) = P ( Y = 1 ∣ X = x , A = a ) , the sensitive-attribute posterior collapses to η ˉ = 1 { a = 1 } \bar{\eta}=\mathbb{1}\{a=1\} η ˉ = 1 { a = 1 } , and the fairness cost parameter and positive-class probability are α ˉ = P ( A = 1 ∣ Y = 1 ) \bar{\alpha}=\mathbb{P}(A=1\mid Y=1) α ˉ = P ( A = 1 ∣ Y = 1 ) and π = P ( Y = 1 ) \pi=\mathbb{P}(Y=1) π = P ( Y = 1 ) .
Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha}\in(0,1) α , α ˉ ∈ ( 0 , 1 ) and define the partial losses
L 1 ( t ) = ( 1 − α ) e − t , L − 1 ( t ) = α e t , L ˉ 1 ( t ) = − ( 1 − α ˉ ) t , L ˉ − 1 ( t ) = α ˉ t . (33) L_1(t)=(1-\alpha)e^{-t},\qquad L_{-1}(t)=\alpha e^t,\qquad\bar{L}_1(t)=-(1-\bar{\alpha})t,\qquad\bar{L}_{-1}(t)=\bar{\alpha}t. \tag*{(33)} L 1 ( t ) = ( 1 − α ) e − t , L − 1 ( t ) = α e t , L ˉ 1 ( t ) = − ( 1 − α ˉ ) t , L ˉ − 1 ( t ) = α ˉ t . (33)
The following proposition proves that the composite loss L f a i r L_{\mathrm{fair}} L fair is FAFC.
Proposition 29. Let α , α ˉ ∈ ( 0 , 1 ) \alpha,\bar{\alpha}\in(0,1) α , α ˉ ∈ ( 0 , 1 ) , let ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) be the surrogate loss pair in (33) , and let K = ( 0 , 1 ) × { − 1 , 1 } \mathcal{K}=(0,1)\times\{-1,1\} K = ( 0 , 1 ) × { − 1 , 1 } . Assume ( η ( X , A ) , A ) ∈ K (\eta(X,A),A)\in\mathcal{K} ( η ( X , A ) , A ) ∈ K almost surely. For every λ ∈ R \lambda\in\mathbb{R} λ ∈ R , the map t ↦ C E O L , L ˉ ( η , a , t ) t\mapsto C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t) t ↦ C EO L , L ˉ ( η , a , t ) is strictly convex on K \mathcal{K} K , and L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K for the EO criterion in the attribute-aware setting.
Proof. Substituting the partial losses into the attribute-aware EO conditional risk gives
C E O L , L ˉ ( η , a , t ) = η ( 1 − α ) e − t + ( 1 − η ) α e t + λ π − 1 η ( 1 { a = 1 } − α ˉ ) t . (34) C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t)=\eta(1-\alpha)e^{-t}+(1-\eta)\alpha e^{t}+\lambda\pi^{-1}\eta\left(\mathbf{1}\{a=1\}-\bar{\alpha}\right)t. \tag*{(34)} C EO L , L ˉ ( η , a , t ) = η ( 1 − α ) e − t + ( 1 − η ) α e t + λ π − 1 η ( 1 { a = 1 } − α ˉ ) t . (34)
Differentiating the partial losses at t = 0 t=0 t = 0 gives
L 1 ′ ( 0 ) = − ( 1 − α ) , L − 1 ′ ( 0 ) = α , L ˉ 1 ′ ( 0 ) = − ( 1 − α ˉ ) , L ˉ − 1 ′ ( 0 ) = α ˉ . L'_1(0)=-(1-\alpha),\qquad L'_{-1}(0)=\alpha,\qquad\bar{L}'_1(0)=-(1-\bar{\alpha}),\qquad\bar{L}'_{-1}(0)=\bar{\alpha}. L 1 ′ ( 0 ) = − ( 1 − α ) , L − 1 ′ ( 0 ) = α , L ˉ 1 ′ ( 0 ) = − ( 1 − α ˉ ) , L ˉ − 1 ′ ( 0 ) = α ˉ .
Here
α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) = 0 , L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) = − 1 < 0 , \bar{\alpha}\bar{L}'_1(0)+(1-\bar{\alpha})\bar{L}'_{-1}(0)=0,\qquad\bar{L}'_1(0)-\bar{L}'_{-1}(0)=-1<0, α ˉ L ˉ 1 ′ ( 0 ) + ( 1 − α ˉ ) L ˉ − 1 ′ ( 0 ) = 0 , L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) = − 1 < 0 ,
and
L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 1 = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) . L'_1(0)-L'_{-1}(0)=-1=\bar{L}'_1(0)-\bar{L}'_{-1}(0). L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = − 1 = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) .
Hence (E1)–(E3) hold: the bracket in (E2)–(E3) vanishes, so (E2) reduces to L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) L'_1(0)-L'_{-1}(0)=\bar{L}'_1(0)-\bar{L}'_{-1}(0) L 1 ′ ( 0 ) − L − 1 ′ ( 0 ) = L ˉ 1 ′ ( 0 ) − L ˉ − 1 ′ ( 0 ) , and (E3) to α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0 \alpha L'_1(0)+(1-\alpha)L'_{-1}(0)=0 α L 1 ′ ( 0 ) + ( 1 − α ) L − 1 ′ ( 0 ) = 0 .
The conditional risk (34) is strictly convex in t t t for every ( η , a ) ∈ ( 0 , 1 ) × { − 1 , 1 } (\eta,a)\in(0,1)\times\{-1,1\} ( η , a ) ∈ ( 0 , 1 ) × { − 1 , 1 } , because
∂ 2 ∂ t 2 C E O L , L ˉ ( η , a , t ) = η ( 1 − α ) e − t + ( 1 − η ) α e t > 0 ; (35) \frac{\partial^2}{\partial t^2}C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t)=\eta(1-\alpha)e^{-t}+(1-\eta)\alpha e^t>0; \tag*{(35)} ∂ t 2 ∂ 2 C EO L , L ˉ ( η , a , t ) = η ( 1 − α ) e − t + ( 1 − η ) α e t > 0 ; (35)
in particular λ \lambda λ does not enter the second derivative, so convexity holds for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R .
Since K = ( 0 , 1 ) × { − 1 , 1 } \mathcal{K}=(0,1)\times\{-1,1\} K = ( 0 , 1 ) × { − 1 , 1 } is a well-posed domain, the sufficiency direction of Theorem 14 gives that L f a i r L_{\mathrm{fair}} L fair is ( α , α ˉ ) (\alpha,\bar{\alpha}) ( α , α ˉ ) -FAFC on K \mathcal{K} K .
It remains to verify well-posedness. Write A = η ( 1 − α ) > 0 A=\eta(1-\alpha)>0 A = η ( 1 − α ) > 0 , B = ( 1 − η ) α > 0 B=(1-\eta)\alpha>0 B = ( 1 − η ) α > 0 and δ a = r η ( 1 { a = 1 } − α ˉ ) \delta_a=r\eta(\mathbf{1}\{a=1\}-\bar{\alpha}) δ a = r η ( 1 { a = 1 } − α ˉ ) with r = λ / π r=\lambda/\pi r = λ / π , so that C E O L , L ˉ ( η , a , t ) = A e − t + B e t + δ a t C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t)=Ae^{-t}+Be^t+\delta_at C EO L , L ˉ ( η , a , t ) = A e − t + B e t + δ a t . Because A A A and B B B are positive, the two exponential terms dominate the linear drift: A e − t → + ∞ Ae^{-t}\to+\infty A e − t → + ∞ as t → − ∞ t\to-\infty t → − ∞ and B e t → + ∞ Be^t\to+\infty B e t → + ∞ as t → + ∞ t\to+\infty t → + ∞ , for every value of δ a \delta_a δ a . Hence the risk is bounded below for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R , and by (35) it is strictly convex, so it possesses a unique minimizer.
We now compute the calibration function and the resulting excess-risk bound. We take the symmetric member of the exponential/linear family, in which
L 1 ( t ) = 1 2 e − t , L − 1 ( t ) = 1 2 e t , L ˉ 1 ( t ) = − 1 2 t , L ˉ − 1 ( t ) = 1 2 t . L_1(t)=\frac{1}{2}e^{-t},\qquad L_{-1}(t)=\frac{1}{2}e^t,\qquad\bar{L}_1(t)=-\frac{1}{2}t,\qquad\bar{L}_{-1}(t)=\frac{1}{2}t. L 1 ( t ) = 2 1 e − t , L − 1 ( t ) = 2 1 e t , L ˉ 1 ( t ) = − 2 1 t , L ˉ − 1 ( t ) = 2 1 t .
Up to the common positive factor 1 2 \frac{1}{2} 2 1 , this is the standard unweighted exponential loss paired with the symmetric linear fairness penalty.
Well-posedness holds for every λ ∈ R \lambda\in\mathbb{R} λ ∈ R , so restricting λ \lambda λ is not needed for FAFC or for the existence of a minimizer. The restriction below comes from the calibration bound: writing r = λ / π r = \lambda/\pi r = λ / π , the level-set minimization that produces the constant 0.95 0.95 0.95 is carried out over 0 ≤ r ≤ 1 0 \le r \le1 0 ≤ r ≤ 1 . For r > 1 r > 1 r > 1 the same uniform lower bound does not hold.
Proposition 30. Let α = α ˉ = 1 2 \alpha= \bar{\alpha} = \frac{1}{2} α = α ˉ = 2 1 , let K \mathcal{K} K be the domain of Proposition 29 , and suppose that the trade-off parameter lies in the range 0 ≤ λ ≤ π 0 \le\lambda\le\pi 0 ≤ λ ≤ π . Set B max : = sup ( η , a ) ∈ K ∣ D 1 2 , 1 2 ( η , a ) ∣ B_{\max} := \sup_{(\eta,a)\in\mathcal{K}} \left|D^{\frac{1}{2},\frac{1}{2}}(\eta,a)\right| B m a x := sup ( η , a ) ∈ K D 2 1 , 2 1 ( η , a ) , and define ν L , L ˉ 1 2 , 1 2 , a w \nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}} ν L , L ˉ 2 1 , 2 1 , aw as in § 4 but with an infimum in place of the minimum. Then, on K \mathcal{K} K ,
ψ L , L ˉ 1 2 , 1 2 , a w ( ϵ ) ≥ 0.95 ϵ 2 (36) \psi_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}}(\epsilon) \ge0.95\epsilon^2 \tag*{(36)} ψ L , L ˉ 2 1 , 2 1 , aw ( ϵ ) ≥ 0.95 ϵ 2 (36) and, consequently, for every f ∈ F f \in\mathcal{F} f ∈ F ,
R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ ≤ min { 1 2 + λ 2 π , 20 19 ( R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ ) } . (37) \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f) - \mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*} \le\min\left\{\frac{1}{2} + \frac{\lambda}{2\pi},\sqrt{\frac{20}{19}\left(\mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f) - \mathcal{R}_{\mathrm{fair}}^{L,\bar{L},*}\right)}\right\}. \tag*{(37)} R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ ≤ min { 2 1 + 2 π λ , 19 20 ( R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ ) } . (37)
Proof. By Proposition 29 and the well-posedness established above, the conditional risk C E O L , L ˉ ( η , a , t ) = A e − t + B e t + δ a t C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t) = A e^{-t} + B e^t + \delta_a t C EO L , L ˉ ( η , a , t ) = A e − t + B e t + δ a t , with A = 1 2 η A = \frac{1}{2}\eta A = 2 1 η , B = 1 2 ( 1 − η ) B = \frac{1}{2}(1-\eta) B = 2 1 ( 1 − η ) , δ a = r η ( 1 { a = 1 } − 1 2 ) \delta_a = r\eta(\mathbb{1}\{a=1\} - \frac{1}{2}) δ a = r η ( 1 { a = 1 } − 2 1 ) , r = λ / π r = \lambda/\pi r = λ / π . Since A , B > 0 A,B > 0 A , B > 0 , the minimizer in t t t below is attained at an interior point of K \mathcal{K} K .
Moreover, B max B_{\max} B m a x is a supremum and ν L , L ˉ 1 2 , 1 2 , a w \nu_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}} ν L , L ˉ 2 1 , 2 1 , aw is an infimum; the level-set computation below shows that the relevant infima are attained at the displayed interior points, so the formulas from the compact case apply.
Since C E O L , L ˉ C_{\mathrm{EO}}^{L,\bar{L}} C EO L , L ˉ is strictly convex, its unique minimizer solves ∂ t C E O L , L ˉ ( η , a , t ∗ ) = − A e − t ∗ + B e t ∗ + δ a = 0 \partial_t C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t^*) = -Ae^{-t^*} + Be^{t^*} + \delta_a = 0 ∂ t C EO L , L ˉ ( η , a , t ∗ ) = − A e − t ∗ + B e t ∗ + δ a = 0 , i.e., B w 2 + δ a w − A = 0 Bw^2 + \delta_a w - A = 0 B w 2 + δ a w − A = 0 with w = e t ∗ w=e^{t^*} w = e t ∗ :
w = − δ a + δ a 2 + 4 A B 2 B , t ∗ = log w . w = \frac{-\delta_a + \sqrt{\delta_a^2 + 4AB}}{2B}, \qquad t^* = \log w. w = 2 B − δ a + δ a 2 + 4 A B , t ∗ = log w .
To locate the wrong-sign half-line we first compute the discriminant. With ( α , α ˉ ) = ( 1 2 , 1 2 ) (\alpha,\bar{\alpha}) = \left(\frac{1}{2},\frac{1}{2}\right) ( α , α ˉ ) = ( 2 1 , 2 1 ) the attribute-aware EO cost functions of § 3.1.2 become
A E O 1 2 , 1 2 ( η , a ) = η 2 − r η 2 1 { a = 1 } , B E O 1 2 , 1 2 ( η , a ) = 1 − η 2 − r η 2 1 { a = − 1 } , A_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\frac{\eta}{2}-\frac{r\eta}{2}\mathbf{1}\{a=1\},\qquad B_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\frac{1-\eta}{2}-\frac{r\eta}{2}\mathbf{1}\{a=-1\}, A EO 2 1 , 2 1 ( η , a ) = 2 η − 2 r η 1 { a = 1 } , B EO 2 1 , 2 1 ( η , a ) = 2 1 − η − 2 r η 1 { a = − 1 } ,
Using 1 { a = − 1 } = 1 − 1 { a = 1 } \mathbf{1}\{a=-1\}=1-\mathbf{1}\{a=1\} 1 { a = − 1 } = 1 − 1 { a = 1 } and the definition of δ a \delta_a δ a ,
D E O 1 2 , 1 2 ( η , a ) = η − 1 2 − δ a . D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\eta-\frac{1}{2}-\delta_a. D EO 2 1 , 2 1 ( η , a ) = η − 2 1 − δ a .
Writing Φ ( η , a ) = ∂ t C E O L , L ˉ ( η , a , t ) ∣ t = 0 \Phi(\eta,a)=\partial_t C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,t)|_{t=0} Φ ( η , a ) = ∂ t C EO L , L ˉ ( η , a , t ) ∣ t = 0 , we have
Φ ( η , a ) = − A + B + δ a = 1 2 − η + δ a . \Phi(\eta,a)=-A+B+\delta_a=\frac{1}{2}-\eta+\delta_a. Φ ( η , a ) = − A + B + δ a = 2 1 − η + δ a .
Hence Φ ( η , a ) = − D E O 1 2 , 1 2 ( η , a ) \Phi(\eta,a)=-D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a) Φ ( η , a ) = − D EO 2 1 , 2 1 ( η , a ) and the risk is therefore monotone away from t = 0 t=0 t = 0 on the wrong-sign half-line, so
C E O L , L ˉ , − ( η , a ) = C E O L , L ˉ ( η , a , 0 ) = A + B . C_{\mathrm{EO}}^{L,\bar{L},-}(\eta,a)=C_{\mathrm{EO}}^{L,\bar{L}}(\eta,a,0)=A+B. C EO L , L ˉ , − ( η , a ) = C EO L , L ˉ ( η , a , 0 ) = A + B .
Subtracting the value at t ∗ t^* t ∗ gives
H L , L ˉ 1 2 , 1 2 ( η , a ) = C E O L , L ˉ , − ( η , a ) − C E O L , L ˉ , ∗ ( η , a ) = A + B − A w − B w − δ a log w . H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=C_{\mathrm{EO}}^{L,\bar{L},-}(\eta,a)-C_{\mathrm{EO}}^{L,\bar{L},*}(\eta,a)=A+B-\frac{A}{w}-Bw-\delta_a\log w. H L , L ˉ 2 1 , 2 1 ( η , a ) = C EO L , L ˉ , − ( η , a ) − C EO L , L ˉ , ∗ ( η , a ) = A + B − w A − B w − δ a log w .
Since D E O 1 2 , 1 2 ( η , a ) = η ( 1 − r 2 1 { a = 1 } + r 2 1 { a = − 1 } ) − 1 2 D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)=\eta\left(1-\frac{r}{2}\mathbf{1}\{a=1\}+\frac{r}{2}\mathbf{1}\{a=-1\}\right)-\frac{1}{2} D EO 2 1 , 2 1 ( η , a ) = η ( 1 − 2 r 1 { a = 1 } + 2 r 1 { a = − 1 } ) − 2 1 , for each a a a , D E O 1 2 , 1 2 ( η , a ) D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a) D EO 2 1 , 2 1 ( η , a ) is affine in η \eta η , so the equation ∣ D E O 1 2 , 1 2 ( η , a ) ∣ = ϵ \left|D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)\right|=\epsilon D EO 2 1 , 2 1 ( η , a ) = ϵ contains at most four solutions in all:
η = 1 2 ∓ ϵ 1 − r 2 ( a = 1 ) , η = 1 2 ∓ ϵ 1 + r 2 ( a = − 1 ) , \eta=\frac{\frac{1}{2}\mp\epsilon}{1-\frac{r}{2}}\quad(a=1),\qquad\eta=\frac{\frac{1}{2}\mp\epsilon}{1+\frac{r}{2}}\quad(a=-1), η = 1 − 2 r 2 1 ∓ ϵ ( a = 1 ) , η = 1 + 2 r 2 1 ∓ ϵ ( a = − 1 ) ,
Because the calibration inequality is inverted through a quadratic lower bound ν ≥ c ϵ 2 \nu\geq c\epsilon^2 ν ≥ c ϵ 2 , the relevant object is the ratio H L , L ˉ 1 2 , 1 2 / ϵ 2 H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}/\epsilon^2 H L , L ˉ 2 1 , 2 1 / ϵ 2 . On a level set ϵ = ∣ D E O 1 2 , 1 2 ∣ \epsilon=\left|D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}\right| ϵ = D EO 2 1 , 2 1 , this ratio equals H L , L ˉ 1 2 , 1 2 / ∣ D E O 1 2 , 1 2 ∣ 2 H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}/\left|D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}\right|^2 H L , L ˉ 2 1 , 2 1 / D EO 2 1 , 2 1 2 and is free of ϵ \epsilon ϵ :
H L , L ˉ 1 2 , 1 2 ϵ 2 = A + B − A w − B w − δ a log w ( η − 1 2 − δ a ) 2 . (38) \frac{H_{L,\bar{L}}^{\frac{1}{2},\frac{1}{2}}}{\epsilon^2}=\frac{A+B-\frac{A}{w}-Bw-\delta_a\log w}{\left(\eta-\frac{1}{2}-\delta_a\right)^2}. \tag*{(38)} ϵ 2 H L , L ˉ 2 1 , 2 1 = ( η − 2 1 − δ a ) 2 A + B − w A − B w − δ a log w . (38)
Minimizing (38) over 0 ≤ r ≤ 1 0 \le r \le1 0 ≤ r ≤ 1 , a ∈ { − 1 , 1 } a \in\{-1,1\} a ∈ { − 1 , 1 } and 0 < η < 1 0 < \eta< 1 0 < η < 1 gives its minimum at r = 1 r=1 r = 1 on the branch a = 1 a=1 a = 1 with D E O 1 2 , 1 2 < 0 D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}<0 D EO 2 1 , 2 1 < 0 . There δ 1 = A = η / 2 \delta_1=A=\eta/2 δ 1 = A = η /2 and ϵ = B = ( 1 − η ) / 2 \epsilon=B=(1-\eta)/2 ϵ = B = ( 1 − η ) /2 , so the first-order condition B w 2 + A w − A = 0 Bw^2+Aw-A=0 B w 2 + A w − A = 0 is equivalent to η = w 2 / ( w 2 − w + 1 ) \eta=w^2/(w^2-w+1) η = w 2 / ( w 2 − w + 1 ) , and (38) reduces to the one-variable function
R ( w ) = 2 ( w 2 − w + 1 ) [ ( 1 − 2 w ) ( 1 − w ) − w 2 log w ] ( 1 − w ) 2 , 0 < w < 1 , R(w)=\frac{2(w^2-w+1)\left[(1-2w)(1-w)-w^2\log w\right]}{(1-w)^2},\qquad0<w<1, R ( w ) = ( 1 − w ) 2 2 ( w 2 − w + 1 ) [ ( 1 − 2 w ) ( 1 − w ) − w 2 log w ] , 0 < w < 1 ,
whose unique stationary point is w ∗ = 0.7667913 … w^*=0.7667913\ldots w ∗ = 0.7667913 … , with
R ( w ∗ ) = 0.9570795 … . R(w^*)=0.9570795\ldots. R ( w ∗ ) = 0.9570795 … .
Since ν L , L ˉ 1 2 , 1 2 , a w ≥ 0.95 ϵ 2 \nu_{L,\bar L}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}}\ge0.95\epsilon^2 ν L , L ˉ 2 1 , 2 1 , aw ≥ 0.95 ϵ 2 and ϵ ↦ 0.95 ϵ 2 \epsilon\mapsto0.95\epsilon^2 ϵ ↦ 0.95 ϵ 2 is convex on [ 0 , ∞ ) [0,\infty) [ 0 , ∞ ) , the convex envelope satisfies
ψ L , L ˉ 1 2 , 1 2 , a w = ( ν L , L ˉ 1 2 , 1 2 , a w ) ∗ ∗ ≥ 0.95 ϵ 2 \psi_{L,\bar L}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}}=\left(\nu_{L,\bar L}^{\frac{1}{2},\frac{1}{2},\mathrm{aw}}\right)^{**}\ge0.95\epsilon^2 ψ L , L ˉ 2 1 , 2 1 , aw = ( ν L , L ˉ 2 1 , 2 1 , aw ) ∗∗ ≥ 0.95 ϵ 2
for every ϵ ≥ 0 \epsilon\ge0 ϵ ≥ 0 .
It remains to bound ϵ = R f a i r 1 2 , 1 2 ( f ) − R f a i r 1 2 , 1 2 , ∗ \epsilon=\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2}}(f)-\mathcal{R}_{\mathrm{fair}}^{\frac{1}{2},\frac{1}{2},*} ϵ = R fair 2 1 , 2 1 ( f ) − R fair 2 1 , 2 1 , ∗ , and this is done from two independent directions.
The calibration inequality of Theorem 17 in its attribute-aware form (Remark 18 ) together with (36) gives
ϵ ≤ 20 19 ( R f a i r L , L ˉ ( f ) − R f a i r L , L ˉ , ∗ ) . \epsilon\le\sqrt{\frac{20}{19}\left(\mathcal{R}_{\mathrm{fair}}^{L,\bar L}(f)-\mathcal{R}_{\mathrm{fair}}^{L,\bar L,*}\right)}. ϵ ≤ 19 20 ( R fair L , L ˉ ( f ) − R fair L , L ˉ , ∗ ) .
By Lemma 22 , the excess fairness-aware risk is a wrong-sign average of the discriminant, so it cannot exceed the largest discriminant value,
ϵ ≤ sup ( η , a ) ∣ D E O 1 2 , 1 2 ( η , a ) ∣ = 1 2 + r 2 , \epsilon\le\sup_{(\eta,a)}\left|D_{\mathrm{EO}}^{\frac{1}{2},\frac{1}{2}}(\eta,a)\right|=\frac{1}{2}+\frac{r}{2}, ϵ ≤ ( η , a ) sup D EO 2 1 , 2 1 ( η , a ) = 2 1 + 2 r ,
Taking their minimum gives (37) .
Population-level form versus pointwise form
We first record the pointwise characterization of cost-sensitive Fisher consistency used in § 2.1 . For η ∈ [ 0 , 1 ] \eta\in[0,1] η ∈ [ 0 , 1 ] , let γ : = η − α \gamma:=\eta-\alpha γ := η − α and
C η ( t ) : = η L 1 ( t ) + ( 1 − η ) L − 1 ( t ) C_\eta(t):=\eta L_1(t)+(1-\eta)L_{-1}(t) C η ( t ) := η L 1 ( t ) + ( 1 − η ) L − 1 ( t )
denote the conditional L L L -risk, and define
H L α ( η ) : = inf { t : t γ ≤ 0 } C η ( t ) − inf t ∈ R C η ( t ) ≥ 0. H_L^\alpha(\eta) := \inf_{\{t:t\gamma\le0\}} C_\eta(t) - \inf_{t\in\mathbb{R}} C_\eta(t) \ge0. H L α ( η ) := { t : t γ ≤ 0 } inf C η ( t ) − t ∈ R inf C η ( t ) ≥ 0.
Proposition 31. Assume that, for every η ∈ [ 0 , 1 ] \eta\in[0,1] η ∈ [ 0 , 1 ] , the conditional risk C η C_\eta C η is convex, bounded below, and coercive. Then L L L is α \alpha α -Fisher consistent in the sense of Definition 2 if and only if
H L α ( η ) > 0 for every η ∈ [ 0 , 1 ] with γ ≠ 0. H_L^\alpha(\eta) > 0 \quad\text{for every } \eta\in[0,1] \text{ with } \gamma\ne0. H L α ( η ) > 0 for every η ∈ [ 0 , 1 ] with γ = 0. Proof. For any distribution D D D , the L L L -risk decomposes as
R L ( f ; D ) = E X [ C η ( X ) ( f ( X ) ) ] . \mathcal{R}_L(f;D) = \mathbb{E}_X\left[C_{\eta(X)}(f(X))\right]. R L ( f ; D ) = E X [ C η ( X ) ( f ( X )) ] . Suppose first that H L α ( η ) > 0 H_L^\alpha(\eta) > 0 H L α ( η ) > 0 for every η \eta η with γ ≠ 0 \gamma\ne0 γ = 0 . If a minimizer f ∗ f^* f ∗ were wrong-signed on a set of positive probability, then changing f ∗ ( x ) f^*(x) f ∗ ( x ) on that set to a point on the correct side would reduce the conditional risk by at least H L α ( η ( x ) ) > 0 H_L^\alpha(\eta(x)) > 0 H L α ( η ( x )) > 0 , contradicting minimality. Hence every minimizer is correctly signed P X P_X P X -a.e.
Conversely, suppose that H L α ( η 0 ) = 0 H_L^\alpha(\eta_0) = 0 H L α ( η 0 ) = 0 for some η 0 \eta_0 η 0 with γ 0 ≠ 0 \gamma_0 \ne0 γ 0 = 0 . Since C η 0 C_{\eta_0} C η 0 is convex, bounded below, and coercive, it attains its minimum, and the wrong-sign half-line contains a minimizer. Let D D D be supported on a single point x 0 x_0 x 0 with P ( Y = 1 ∣ X = x 0 ) = η 0 \mathbb{P}(Y=1 \mid X=x_0) = \eta_0 P ( Y = 1 ∣ X = x 0 ) = η 0 . Then
R L ( f ; D ) = C η 0 ( f ( x 0 ) ) , \mathcal{R}_L(f;D) = C_{\eta_0}(f(x_0)), R L ( f ; D ) = C η 0 ( f ( x 0 )) , so this distribution admits a minimizer whose sign disagrees with γ 0 \gamma_0 γ 0 . Thus L L L is not α \alpha α -Fisher consistent.
Proposition 31 shows that, for cost-sensitive classification, the pointwise and population forms are equivalent when the population statement is quantified over all distributions. The following example shows that, for a fixed fairness-aware distribution, the population statement of Corollary 13 can hold while FAFC fails off the support; Remark 33 explains the role of the distribution-dependent α ˉ , π \bar{\alpha}, \pi α ˉ , π .
Example 32. We consider the DP criterion and attribute-blind setting. Let ( D , D ˉ ) (D,\bar{D}) ( D , D ˉ ) be supported on two equiprobable points x 1 , x 2 x_1,x_2 x 1 , x 2 , with ( η ( x 1 ) , η ˉ ( x 1 ) ) = ( 0.9 , 0.8 ) (\eta(x_1),\bar{\eta}(x_1))=(0.9,0.8) ( η ( x 1 ) , η ˉ ( x 1 )) = ( 0.9 , 0.8 ) and ( η ( x 2 ) , η ˉ ( x 2 ) ) = ( 0.4 , 0.2 ) (\eta(x_2),\bar{\eta}(x_2))=(0.4,0.2) ( η ( x 2 ) , η ˉ ( x 2 )) = ( 0.4 , 0.2 ) . Take α = α ˉ = 1 2 \alpha=\bar{\alpha}=\frac{1}{2} α = α ˉ = 2 1 , λ = 1 \lambda=1 λ = 1 , and
L 1 ( t ) = t 2 , L − 1 ( t ) = ( t − 1 ) 2 , L ˉ 1 ( t ) = t , L ˉ − 1 ( t ) = − t . L_1(t)=t^2,\qquad L_{-1}(t)=(t-1)^2,\qquad\bar{L}_1(t)=t,\qquad\bar{L}_{-1}(t)=-t. L 1 ( t ) = t 2 , L − 1 ( t ) = ( t − 1 ) 2 , L ˉ 1 ( t ) = t , L ˉ − 1 ( t ) = − t . The conditional risk
C D P L , L ˉ ( η , η ˉ , t ) = t 2 − [ 2 ( 1 − η ) + ( 2 η ˉ − 1 ) ] t + ( 1 − η ) C_{\mathrm{DP}}^{L,\bar{L}}(\eta,\bar{\eta},t)=t^2-\left[2(1-\eta)+(2\bar{\eta}-1)\right]t+(1-\eta) C DP L , L ˉ ( η , η ˉ , t ) = t 2 − [ 2 ( 1 − η ) + ( 2 η ˉ − 1 ) ] t + ( 1 − η ) is strictly convex and bounded below, with unique minimizer
t ∗ ( η , η ˉ ) = 1 2 − η + η ˉ . t^*(\eta,\bar{\eta})=\frac{1}{2}-\eta+\bar{\eta}. t ∗ ( η , η ˉ ) = 2 1 − η + η ˉ . Since D α , α ˉ ( η , η ˉ ) = A D P α , α ˉ ( η , η ˉ ) − B D P α , α ˉ ( η , η ˉ ) = ( η − α ) − λ ( η ˉ − α ˉ ) D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=A_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})-B_{\mathrm{DP}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=(\eta-\alpha)-\lambda(\bar{\eta}-\bar{\alpha}) D α , α ˉ ( η , η ˉ ) = A DP α , α ˉ ( η , η ˉ ) − B DP α , α ˉ ( η , η ˉ ) = ( η − α ) − λ ( η ˉ − α ˉ ) , for α = α ˉ = 1 2 \alpha=\bar{\alpha}=\frac{1}{2} α = α ˉ = 2 1 and λ = 1 \lambda=1 λ = 1 , the discriminant is
D α , α ˉ ( η , η ˉ ) = η − η ˉ . D^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=\eta-\bar{\eta}. D α , α ˉ ( η , η ˉ ) = η − η ˉ . On { η < η ˉ } \{\eta<\bar{\eta}\} { η < η ˉ } the discriminant is negative, so the wrong-sign half-line is t ≥ 0 t\geq0 t ≥ 0 . Since
t ∗ = 1 2 − ( η − η ˉ ) > 1 2 , H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 t^*=\frac{1}{2}-(\eta-\bar{\eta})>\frac{1}{2},\qquad H_{L,\bar{L}}^{\alpha,\bar{\alpha}}(\eta,\bar{\eta})=0 t ∗ = 2 1 − ( η − η ˉ ) > 2 1 , H L , L ˉ α , α ˉ ( η , η ˉ ) = 0 on { η < η ˉ } \{\eta<\bar{\eta}\} { η < η ˉ } , which implies that L f a i r L_{\mathrm{fair}} L fair is not FAFC on [ 0 , 1 ] 2 [0,1]^2 [ 0 , 1 ] 2 .
Since both support points lie on 0 < η − η ˉ < 1 2 0<\eta-\bar{\eta}<\frac{1}{2} 0 < η − η ˉ < 2 1 , t ∗ t^* t ∗ equals 0.4 and 0.3, and the conditional minimizer lies on the correct side. The ( L , L ˉ ) (L,\bar{L}) ( L , L ˉ ) -risk of ( D , D ˉ ) (D,\bar{D}) ( D , D ˉ ) is
R f a i r L , L ˉ ( f ) = p 1 C D P L , L ˉ ( η 1 , η ˉ 1 , f ( x 1 ) ) + p 2 C D P L , L ˉ ( η 2 , η ˉ 2 , f ( x 2 ) ) , p i = P ( X = x i ) > 0. \mathcal{R}_{\mathrm{fair}}^{L,\bar{L}}(f)=p_1C_{\mathrm{DP}}^{L,\bar{L}}(\eta_1,\bar{\eta}_1,f(x_1))+p_2C_{\mathrm{DP}}^{L,\bar{L}}(\eta_2,\bar{\eta}_2,f(x_2)),\qquad p_i=\mathbb{P}(X=x_i)>0. R fair L , L ˉ ( f ) = p 1 C DP L , L ˉ ( η 1 , η ˉ 1 , f ( x 1 )) + p 2 C DP L , L ˉ ( η 2 , η ˉ 2 , f ( x 2 )) , p i = P ( X = x i ) > 0. A minimizer must minimize both C D P L , L ˉ ( η 1 , η ˉ 1 , f ( x 1 ) ) C_{\mathrm{DP}}^{L,\bar{L}}(\eta_1,\bar{\eta}_1,f(x_1)) C DP L , L ˉ ( η 1 , η ˉ 1 , f ( x 1 )) and C D P L , L ˉ ( η 2 , η ˉ 2 , f ( x 2 ) ) C_{\mathrm{DP}}^{L,\bar{L}}(\eta_2,\bar{\eta}_2,f(x_2)) C DP L , L ˉ ( η 2 , η ˉ 2 , f ( x 2 )) , so
f ∗ ( x i ) = t ∗ ( η i , η ˉ i ) > 0 f^*(x_i)=t^*(\eta_i,\bar{\eta}_i)>0 f ∗ ( x i ) = t ∗ ( η i , η ˉ i ) > 0 for i = 1 , 2 i=1,2 i = 1 , 2 . These values have the same sign as D α , α ˉ ( η 1 , η ˉ 1 ) = 0.1 D^{\alpha,\bar{\alpha}}(\eta_1,\bar{\eta}_1)=0.1 D α , α ˉ ( η 1 , η ˉ 1 ) = 0.1 , D α , α ˉ ( η 2 , η ˉ 2 ) = 0.2 D^{\alpha,\bar{\alpha}}(\eta_2,\bar{\eta}_2)=0.2 D α , α ˉ ( η 2 , η ˉ 2 ) = 0.2 , and the support contains no other points, so every minimizer matches the sign of the discriminant on the whole support. This is exactly the conclusion of Corollary 13 for this composite loss.