---
title: Equalized Set Size in Fairness and Prediction
url: https://www.emergentmind.com/topics/equalized-set-size
type: topic
---

# Equalized Set Size in Fairness and Prediction

Searching arXiv for the cited papers to ground the article.
{"query":"arXiv:2605.14260 Equalized Set Size conformal prediction", "max_results": 5}
{"query":"arXiv:2405.11639 Fair Set Cover demographic parity selected sets equal counts", "max_results": 5}
Equalized Set Size denotes a family of cardinality-balancing constraints in which set sizes, or expected set sizes, are required to match across designated units. In the literature considered here, the most explicit formalization appears in conformal prediction, where one chooses group-specific thresholds so that the group-conditional expected prediction-set size is the same constant across groups. Closely related uses arise in fair set cover, where the balanced object is the number of selected sets from each protected group, and in fair allocation, where every agent must receive a bundle of exactly the same size. These formulations are structurally related but not identical: they balance different objects, under different feasibility constraints, and with different fairness implications [2605.14260] [2405.11639] [2507.20899].

## 1. Domain-specific meanings

In split conformal prediction, Equalized Set Size means equality of **group-conditional expected prediction-set size**. If \(\widehat C_t(X)\) is the prediction set at threshold \(t\), and \(G(X)=g\) is a prespecified group label, the relevant functional is
\[
\ell_g(t):=\mathbb E\big[|\widehat C_t(X)|\mid G=g\big].
\]
Equalized Set Size requires group-specific thresholds \(t_g\) such that
\[
\ell_g(t_g)=\lambda \qquad \forall g.
\]
The equality is therefore in expectation and conditional on group, not pointwise over individual predictions [2605.14260].

In fair set cover, the balanced object is different. There, demographic information is attached to the **available sets themselves**, not to the universe elements, and fairness constrains the composition of the selected cover. With groups \(C=\{c_1,\ldots,c_k\}\), target fractions \(f_h\), and selected cover \(X\subseteq S\), fairness requires
\[
|X\cap S_h| = f_h\,|X|.
\]
The closest literal analogue of “equalized set size” is **count-parity**, where \(f_h=\frac1k\) for all groups, so the final cover contains the same number of selected sets from each group. This is explicitly not a notion of equal coverage received by demographic groups of elements [2405.11639].

In fair division under equal-sized bundles, the central constraint is that each agent must receive exactly \(k\) items:
\[
|A_i|=k \qquad \text{for every } i\in[n].
\]
Because removing an item from another agent’s bundle would violate feasibility, the relevant corrective operation is a one-for-one **flip** between bundles rather than the removal-based comparison used in standard EF1 and EFX. This shifts the meaning of “almost envy-free” in a cardinality-constrained environment [2507.20899].

These usages suggest that Equalized Set Size is best understood as a domain-dependent balancing principle on set cardinality, selected-set representation, or expected predictive ambiguity, rather than as a single universal invariant.

## 2. Equalized expected set size in conformal prediction

The conformal-prediction formulation is population-level and is built on standard split conformal calibration. With nonconformity score \(S(X,Y)\), calibration scores \(S_i\), and threshold \(t\), the prediction set is
\[
\widehat C_t(X_{n+1}) := \{y: S(X_{n+1},y)\le t\}.
\]
Under pooled calibration one uses a single empirical threshold \(\hat q\), whereas group-conditional calibration uses thresholds \(\hat q_g\) for each group \(g\). At population level,
\[
F_S(t)=\mathbb P(S\le t), \qquad F_{S\mid g}(t)=\mathbb P(S\le t\mid G=g),
\]
with
\[
q:=\inf\{t:F_S(t)\ge 1-\alpha\}, \qquad q_g:=\inf\{t:F_{S\mid g}(t)\ge 1-\alpha\}.
\]

Equalized Set Size introduces a different policy family. Instead of forcing
\[
\mathbb P\{Y\in \widehat C_{t_g}(X)\mid G=g\}=1-\alpha,
\]
it requires a common expected size:
\[
\ell_g(t_g)=\mathbb E[|\widehat C_{t_g}(X)|\mid G=g]=\lambda \qquad \forall g.
\]
The paper later denotes the corresponding thresholds by \(\tau_g\), satisfying
\[
\ell_g(\tau_g)=\lambda.
\]

This notion is explicitly motivated as a fairness criterion on the **burden of uncertainty communication**. Larger prediction sets mean more uncertainty and more burden on the user, so Equalized Set Size asks that different groups not systematically receive larger prediction sets on average. The paper also defines
\[
\lambda_g := \ell_g(q_g),
\]
the expected set size in group \(g\) when using the group’s own coverage-calibrated threshold. These \(\lambda_g\) values are the key bridge between size parity and coverage parity: if they differ, then exact group-wise coverage already induces cross-group set-size disparity [2605.14260].

## 3. Incompatibility with Equalized Coverage

The central theoretical result in the conformal-prediction literature is that **Equalized Coverage** and **Equalized Expected Set Size** are generally incompatible. Equalized Coverage is achieved by choosing \(t_g=q_g\), so that
\[
F_{S\mid g}(q_g)=1-\alpha \qquad \forall g.
\]
Equalized Set Size instead chooses \(\tau_g\) so that
\[
\ell_g(\tau_g)=\lambda \qquad \forall g.
\]

One direction is formalized by the theorem titled “Equalized coverage induces cross-group expected set size disparity.” Under continuity of \(F_{S\mid g}\), monotonicity of \(\ell_g\), and the paper’s reference-group condition, groupwise coverage thresholds \(q_g\) necessarily create nonzero size disparity:
\[
\ell_g(q_g)-\ell_r(q_r)\ge c_g>0.
\]
The reverse direction is given by “Equalized expected set size induces cross-group coverage disparity.” With local regularity assumptions, lower-bounded score density \(f_{S\mid g}(t)\ge m_g>0\), derivative bound \(|\ell'_g(t)|\le V_g\), and
\[
\kappa_g:=\frac{m_g}{V_g},
\]
the paper proves that if \(\lambda_g<\lambda\), then
\[
F_{S\mid g}(\tau_g)-(1-\alpha)\ge \kappa_g(\lambda-\lambda_g),
\]
while if \(\lambda_{g'}>\lambda\), then
\[
(1-\alpha)-F_{S\mid g'}(\tau_{g'})\ge \kappa_{g'}(\lambda_{g'}-\lambda).
\]
Hence, whenever the common target \(\lambda\) lies strictly between two distinct \(\lambda_g\) values, coverage disparity is unavoidable [2605.14260].

The mechanism is cross-group heterogeneity in the score distributions, measured by
\[
\sigma_\Delta^2=\mathrm{Var}(q_G).
\]
If \(\sigma_\Delta=0\), the conflict disappears. If \(\sigma_\Delta>0\), calibration policy determines where the distortion appears. For pooled threshold \(q\), coverage distortion is
\[
\varepsilon_g(q):=F_{S\mid g}(q)-(1-\alpha),
\]
and the conservation law gives
\[
\sum_{g\in\mathcal G} p_g\,\varepsilon_g(q)=\delta(q),
\]
with \(\delta(q)=0\) when \(F_S\) is continuous at \(q\). The pooled-threshold uncertainty relation then yields
\[
\mathrm{Var}(\varepsilon_G(q)) \ge m_{\mathrm{eff}(q)}^2\,\mathrm{Var}(q_G).
\]

The same heterogeneity governs the cost of moving between policies. Switching from pooled calibration \(q\) to groupwise coverage \(q_g\) incurs RMS set-size distortion
\[
\sqrt{\mathbb E[(\ell_G(q_G)-\ell_G(q))^2]} \ge v_{\mathrm{eff}(q)}\,\sigma_\Delta,
\]
while moving from \(q_g\) to equalized-size thresholds \(\tau_g\) incurs RMS coverage distortion
\[
\sqrt{\mathbb E[(F_{S\mid G}(\tau_G)-F_{S\mid G}(q_G))^2]} \ge \kappa_{\mathrm{eff}(\lambda)}\,\sigma_\lambda.
\]

The empirical results mirror the theory. On Bias in Bios at \(\alpha=0.1\), pooled RMS coverage distortion is \(0.0015\), RMS set-size distortion after \(q\to q_g\) is \(0.0051\), and RMS coverage distortion under equalized expected set size is \(0.0017\). On MultiNLI, the corresponding values are \(0.0150\), \(0.0532\), and \(0.0209\). On FACET, they are \(0.0083\), \(0.1717\), and \(0.0199\). These experiments support the paper’s claim that group heterogeneity is not removed by calibration; it is displaced between the coverage and size dimensions [2605.14260].

## 4. Equalized selected-set counts in fair set cover

Fair Set Cover introduces an explicit selected-set analogue of Equalized Set Size. In the classical setting, one has a universe
\[
U=\{e_1,e_2,\ldots,e_n\},
\]
a family of sets
\[
S=\{S_1,S_2,\ldots,S_\mu\},
\]
and a cover \(X\subseteq S\) such that
\[
\bigcup_{s_i\in X}=U.
\]
The fairness extension assigns each set \(S_j\) to exactly one demographic group \(c(S_j)\in C=\{c_1,\ldots,c_k\}\), with \(S_h\subseteq S\) denoting the sets of group \(c_h\). Fairness is demographic parity over the selected sets:
\[
|X\cap S_h| = f_h\,|X| \qquad \forall h.
\]

Two special cases are singled out. In **count-parity**,
\[
f_h=\frac1k \qquad \forall h,
\]
so the selected cover contains an equal number of sets from each group. In **ratio-parity**,
\[
f_h=\frac{m_h}{\mu} \qquad \forall h,
\]
so the output preserves the original group proportions in the input family. The generalized framework allows any rational target vector \(f_1,\ldots,f_k\) summing to \(1\) [2405.11639].

Algorithmically, exact balancing is enforced through hard combinatorial constraints. In the unweighted count-parity case, each greedy round chooses exactly one set from each group, which is formalized by the max \(k\)-color cover subproblem with constraints
\[
\sum_{i:S_i\in S_h^-} x_i = 1 \qquad \forall c_h\in C.
\]
In the generalized case, if
\[
f_h=\frac{p_h}{\sum_{c_j\in C} p_j},
\]
the per-round constraints become
\[
\sum_{i:S_i\in S_h^-} x_i = p_h \qquad \forall c_h\in C.
\]
Thus exact equalized selected-set counts are not a soft penalty but a hard quota mechanism.

The paper gives three main unweighted methods. A naive post-processing method guarantees exact fairness with approximation ratio
\[
k(\ln n+1).
\]
An exact-fairness greedy algorithm achieves
\[
(\ln n+1)\text{-approximation}
\]
in
\[
O(m^k n)\text{ time}.
\]
A faster LP-based randomized algorithm achieves expected approximation
\[
\frac{e}{e-1}(\ln n+1)
\]
while still preserving exact fairness. Analogous guarantees hold in the weighted setting. The paper also proves that FSC is NP-complete and that, without additional assumptions, there is no polynomial-time approximation ratio \(f(n)\) for any computable nondecreasing \(f(n)=o(n)\), unless \(\mathsf P=\mathsf{NP}\) [2405.11639].

Empirically, the price of fairness is small. On the Resume Skills dataset, \({\sc Opt\mbox{-}SC}\) has average fairness ratio \(0.48\) and average cover size \(3.32\), while \({\sc Opt\mbox{-}FSC}\) has fairness ratio \(1.00\) and average cover size \(3.75\). The efficient approximation algorithm \({\sc EffAllPick}\) also achieves fairness ratio \(1.00\) with average size \(3.90\). In the paper’s terminology, exact parity therefore yields “zero unfairness” with only a modest increase in output size [2405.11639].

## 5. Equal-sized bundles in fair allocation

In fair allocation under equal-sized bundles, the equalization constraint is literal cardinality parity. There are \(n\) agents and \(kn\) indivisible goods, \(M=[kn]\), and an allocation
\[
\mathcal A=(A_1,\dots,A_n)
\]
must satisfy
\[
|A_i|=k \qquad \forall i\in[n].
\]
Valuations are nonnegative and additive:
\[
v_i(A)=\sum_{g\in A} v_i(g).
\]

Because bundle size must remain fixed, the relevant local repair operation is a **flip**. If agent \(i\) holds \(a\in A_i\) and agent \(j\) holds \(b\in A_j\), then the one-for-one exchange produces
\[
A_i\cup b\setminus a \qquad\text{and}\qquad A_j\cup a\setminus b.
\]
A flip is rational with respect to \(i\) if
\[
v_i(b)>v_i(a).
\]

This yields the notions of **envy-freeness up to one flip (EFF1)** and **envy-freeness up to any flip (EFFX)**. An allocation is EFF1 if, for every pair \(i,j\), either \(i\) does not envy \(j\), or there exists a rational flip after which \(i\) does not envy \(j\). It is EFFX if the same holds for **every** rational flip. The paper also studies multiplicative relaxations such as \(\gamma\)-EFFX, and proves that if an allocation is \(\gamma\)-EF, then it is also \(\gamma\)-EFFX [2507.20899].

The fixed-size constraint changes the algorithmic landscape sharply. The paper proves that any generalized Round-Robin algorithm fails to guarantee \(\gamma\)-EFFX for every \(\gamma\in(0,1]\) whenever \(k>2\), even for \(n=2\) agents with identical valuations. A natural equal-size adaptation of envy cycle elimination also fails to guarantee any positive \(\gamma\)-EFFX approximation, even for \(n=3\). At the same time, some positive exact statements survive: any allocation is either EF or \(1/k\)-EFF1; if \(k=2\), every allocation is EFF1; and when \(k=2\), EFFX allocations always exist for additive valuations.

Under additional structure, stronger guarantees are available. If agents share a common ranking over item values, the paper’s adapted envy-cycle algorithm returns a
\[
\frac12\text{-EFFX}
\]
allocation, while a lower-bound instance shows that the same algorithm cannot guarantee better than
\[
\frac23\text{-EFFX}.
\]
If agents agree on the top \(n\) items and valuations are bounded through
\[
\rho=\max_{i\in[n]}\max_{g,g'\in T^n}\frac{v_i(g)}{v_i(g')},
\]
a more elaborate envy-graph algorithm with swaps and privileged agents runs in polynomial time and returns a
\[
\min\left\{\frac13,\frac{1}{\rho+1}\right\}\text{-EF}
\]
allocation; for \(k=1\), it returns an exact EFFX allocation. With only \(\rho\)-bounded top-\(n\) valuations and no common top-\(n\) set, the guarantee becomes
\[
\frac{1}{\rho+2}\text{-EF}.
\]

The efficiency benchmark is \(k\)-bundle maximum Nash welfare. The paper proves that any \(k\)-bundle MNW-optimal allocation is
\[
\frac12\text{-EFF1},
\]
and that this bound is tight. By contrast, a social-welfare-maximizing allocation may be only
\[
\frac{1}{k-1}\text{-EFF1},
\]
and a leximin allocation may be only about
\[
\frac{2}{k}+o(1/k^2)\text{-EFF1}.
\]
The paper also gives examples showing that EFX and EFFX are largely incomparable, so equal-sized bundles do not merely add a side constraint to classical fair division; they alter the underlying fairness geometry [2507.20899].

## 6. Related combinatorial and partitioning notions

A different but related use of equalized or near-equal set size appears in additive combinatorics. For a sequence \(S\) in an abelian group, the paper on subsum representation shows that when
\[
|\Sigma_n(S)|\le |S|-n+1,
\]
one can partition the terms into nonempty sets \(A_1,\ldots,A_n\) so that
\[
\Sigma_n(S)=A_1+\cdots+A_n,
\]
and, apart from one highly structured counterexample when \(n=2\) and \(|\Sigma_n(S)|=|S|-n+1\), the partition can be chosen with sizes as equal as possible:
\[
\left\lfloor \frac{|S|}{n}\right\rfloor \le |A_i| \le \left\lceil \frac{|S|}{n}\right\rceil \qquad \forall i.
\]
Here equalization is not a fairness notion but a structural refinement that improves how sumset methods apply to subsum problems [1910.11807].

In probabilistic combinatorics, equalized set size may refer to a ground set partitioned into **equal-sized blocks**. One paper partitions \(\{1,\dots,N\}\) into blocks of common size \(S\), chooses a random \(K\)-subset, and studies
\[
\mathbb P\bigl(\forall i,\ |A\cap B_i|\ge R\bigr).
\]
Under
\[
R^2=o(S), \qquad NR=o(SK), \qquad RS\sqrt{RK}=o(N),
\]
the probability tends to
\[
e^{-c},
\]
where
\[
c=\lim_{N\to\infty} \frac{N}{S}\cdot \frac{(SK/N)^{R-1}e^{-SK/N}}{(R-1)!}.
\]
Equal block size is essential there because it makes the failure events exchangeable and yields a Poisson-type asymptotic formula [1608.07635].

Equal-cardinality partition optimization provides another nearby notion. Given an even-sized set
\[
\mathcal X=\{x_1,\ldots,x_N\},
\]
the problem is to find disjoint complementary subsets
\[
\mathcal X_1,\mathcal X_2
\]
with
\[
|\mathcal X_1|=|\mathcal X_2|=\frac N2
\]
minimizing
\[
|S_1-S_2|,
\qquad
S_1=\sum_{x\in\mathcal X_1}x,\quad S_2=\sum_{x\in\mathcal X_2}x.
\]
The paper does not solve the NP-hard global problem, but it gives a locally optimal algorithm under pairwise swaps in
\[
O(N^2)\text{ time and }O(N)\text{ space}.
\]
Here “equalized set size” means strict equality of subset cardinalities, not equality of sums or expected predictive ambiguity [2109.07882].

A useful contrast is provided by equal-sum partitioning of
\[
I_n=\{1,\dots,n\}
\]
into \(k\) subsets \(T_j\) with
\[
\sum_{x\in T_j}x=t.
\]
That construction problem does **not** impose equal subset cardinalities, and the recursive algorithm explicitly allows variable numbers of elements per subset. It is therefore about equalized subset sums, not equalized set size in the cardinality sense [1811.04014].

Source: https://www.emergentmind.com/topics/equalized-set-size