Selection Possibility Frontier (SPF)

• Selection Possibility Frontier (SPF) is a framework that defines Pareto-efficient cohorts by jointly optimizing performance and diversity under fixed-size constraints.
• It quantifies trade-offs by specifying the maximum attainable diversity for a given performance level and vice versa, using submodular functions and greedy approximation.
• The approach aids both retrospective evaluation and prospective decision support by identifying efficient choices and revealing potential improvements in cohort selection.

Searching arXiv for the specified paper to ground the article in the cited source. The Selection Possibility Frontier (SPF) is a representation of Pareto-efficient cohort selection when an organization must jointly consider average talent or performance and cohort diversity under a fixed cohort size. In the formulation introduced in "A Possibility Frontier Approach to Diverse Talent Selection" (Natarajan et al., 7 Oct 2025), the SPF is the set of all feasible cohorts for which no alternative cohort can improve both performance and diversity simultaneously. Operationally, it specifies, for any target level of performance, the maximum diversity achievable, and conversely, for any target level of diversity, the maximum performance achievable. The construct is intended to quantify the trade-off between talent and diversity and to reveal whether an observed cohort is Pareto-inferior relative to other feasible selections (Natarajan et al., 7 Oct 2025).

1. Conceptual definition and decision-theoretic role

In diverse cohort selection, the decision-maker faces two competing desiderata: average performance of the cohort and diversity across one or more attributes such as gender, nationality, or socioeconomic status (Natarajan et al., 7 Oct 2025). The SPF formalizes this tension by restricting attention to cohorts of fixed size $k$ and identifying those that are Pareto-efficient in the two-dimensional space of performance and diversity.

A cohort lies on the SPF if there is no other feasible cohort that improves on both dimensions simultaneously. This makes the SPF a decision-theoretic boundary rather than a single optimal policy. It does not prescribe one universally best cohort; instead, it characterizes the efficient set of attainable trade-offs. In this sense, the SPF functions both as a diagnostic tool for retrospective evaluation and as a prescriptive tool for prospective decision support (Natarajan et al., 7 Oct 2025).

The paper’s framing is notable because it treats diversity and talent as jointly optimizable quantities rather than as sequential filters or informal balancing criteria. This suggests a shift from ad hoc selection to explicit frontier analysis. A plausible implication is that organizations can distinguish genuine trade-offs from avoidable inefficiencies once the feasible boundary is approximated.

2. Mathematical formulation

The applicant universe is denoted $U = \{a_1,\ldots,a_n\}$, and the fixed cohort size is $k$. The feasible cohort space is

$C = \{c \subseteq U : |c| = k\}.$

Each applicant $a \in U$ receives an individual performance score $p(a) \in [0,1]$, such as a psychometric test score. Cohort performance is measured by

$P(c) = \sum_{a \in c} p(a).$

Because all compared cohorts have the same size $k$, maximizing the sum is equivalent to maximizing the mean $P(c)/k$ (Natarajan et al., 7 Oct 2025).

The performance function $P(\cdot)$ is specified as non-negative, finite, monotone, and submodular. Diversity is defined over a set of demographic attributes $U = \{a_1,\ldots,a_n\}$0 with target proportions $U = \{a_1,\ldots,a_n\}$1. For each attribute $U = \{a_1,\ldots,a_n\}$2 and cohort $U = \{a_1,\ldots,a_n\}$3, the count and proportion are

$U = \{a_1,\ldots,a_n\}$4

A proportional-target component is then defined as

$U = \{a_1,\ldots,a_n\}$5

which rewards meeting or exceeding the target count and otherwise increases linearly in the count. The overall diversity score is

$U = \{a_1,\ldots,a_n\}$6

with weights $U = \{a_1,\ldots,a_n\}$7 (Natarajan et al., 7 Oct 2025).

The same source also describes an alternative entropy-based measure,

$U = \{a_1,\ldots,a_n\}$8

plus a constant so that $U = \{a_1,\ldots,a_n\}$9 is non-negative. Under mild conditions, $k$0 is also monotone and submodular (Natarajan et al., 7 Oct 2025).

The frontier itself is defined in terms of the attainable pairs $k$1:

$k$2

Equivalently, a point lies on the frontier if no other cohort strictly dominates it in both dimensions. This definition aligns the SPF with standard Pareto-frontier concepts, but specialized to fixed-size cohort selection under submodular objectives (Natarajan et al., 7 Oct 2025).

3. Optimization characterization and approximation algorithm

The paper characterizes the SPF through two families of knapsack-style problems. For any performance threshold $k$3, one may solve

$k$4

and, symmetrically, for any diversity threshold $k$5,

$k$6

Sweeping the thresholds across their feasible ranges traces out the frontier (Natarajan et al., 7 Oct 2025).

The same paper gives a scalarization approach:

$k$7

Varying $k$8 produces points that lie approximately along the convex hull of the true SPF. This scalarization is the basis of the proposed approximation method (Natarajan et al., 7 Oct 2025).

Exact computation is described as computationally difficult. The paper states that even computing the single most diverse cohort, $k$9 subject to $C = \{c \subseteq U : |c| = k\}.$0, is NP-hard when there are multiple, non-mutually-exclusive attributes, and that finding any Pareto-optimal extreme point other than the pure-performance maximum is likewise intractable for large $C = \{c \subseteq U : |c| = k\}.$1 (Natarajan et al., 7 Oct 2025).

To address this, the paper applies the classical greedy guarantee for non-negative, monotone, submodular objectives. For

$C = \{c \subseteq U : |c| = k\}.$2

the greedy algorithm that iteratively adds the element of largest marginal gain achieves

$C = \{c \subseteq U : |c| = k\}.$3

Because each $C = \{c \subseteq U : |c| = k\}.$4 is also monotone submodular, each scalarized problem can be approximated to within a factor $C = \{c \subseteq U : |c| = k\}.$5 (Natarajan et al., 7 Oct 2025).

The algorithmic workflow is specified as follows: initialize an empty list of frontier points; for $C = \{c \subseteq U : |c| = k\}.$6 to $C = \{c \subseteq U : |c| = k\}.$7, set $C = \{c \subseteq U : |c| = k\}.$8; define $C = \{c \subseteq U : |c| = k\}.$9; compute $a \in U$0; add $a \in U$1 to the frontier list; then return the upper-convex envelope of the collected points (Natarajan et al., 7 Oct 2025). The subroutine $a \in U$2 initializes $a \in U$3, repeats $a \in U$4 times, selects the applicant maximizing the marginal increment $a \in U$5, updates $a \in U$6, and returns $a \in U$7.

4. Computational properties

For an applicant pool of size $a \in U$8 and $a \in U$9 discretization points for $p(a) \in [0,1]$0, the total runtime is

$p(a) \in [0,1]$1

The paper states that, in practice, $p(a) \in [0,1]$2, $p(a) \in [0,1]$3, and $p(a) \in [0,1]$4–$p(a) \in [0,1]$5, yielding runtime in seconds or minutes (Natarajan et al., 7 Oct 2025).

Each point $p(a) \in [0,1]$6 satisfies

$p(a) \in [0,1]$7

and the resulting set of $p(a) \in [0,1]$8 pairs is therefore a $p(a) \in [0,1]$9-approximate convex hull of the true SPF (Natarajan et al., 7 Oct 2025). The same source notes that, in many real datasets, the greedy approximation is much closer to optimal than the worst-case bound.

These properties situate SPF computation within the broader class of submodular optimization problems with cardinality constraints. The significance of the result is not exact frontier recovery in the worst case, but the construction of a tractable, principled upper boundary that can be used for empirical diagnosis and committee guidance. A plausible implication is that organizations can operationalize multi-criteria selection without requiring exact combinatorial optimization.

5. Empirical case study: talent investment program, 2021–2023

The main empirical demonstration concerns a talent investment program evaluated across the 2021, 2022, and 2023 cycles (Natarajan et al., 7 Oct 2025). The applicant pools are reported as approximately $P(c) = \sum_{a \in c} p(a).$0–$P(c) = \sum_{a \in c} p(a).$1 per year, with finalist cohort size fixed at $P(c) = \sum_{a \in c} p(a).$2. Individual performance scores are standardized psychometric assessments in $P(c) = \sum_{a \in c} p(a).$3.

The diversity attributes used in the case study are gender identity, nationality, and socioeconomic status. The targets are specified as $P(c) = \sum_{a \in c} p(a).$4 non-male for gender identity, uniform distribution over the top $P(c) = \sum_{a \in c} p(a).$5 countries for nationality, and $P(c) = \sum_{a \in c} p(a).$6 low-SES for socioeconomic status. The diversity function is

$P(c) = \sum_{a \in c} p(a).$7

where each component is of the form $P(c) = \sum_{a \in c} p(a).$8 (Natarajan et al., 7 Oct 2025).

Implementation used discretization $P(c) = \sum_{a \in c} p(a).$9, corresponding to $k$0. The system was implemented in Python using NumPy for vectorized marginal-gain computations. Post-processing computed the convex hull of the raw $k$1 points in order to enforce monotonicity of the frontier (Natarajan et al., 7 Oct 2025).

The numerical results are reported explicitly:

Year	Reported SPF comparison
2021	Actual cohort at $k$2; vertical gap implies $k$3 potential diversity gain at same performance; horizontal gap implies $k$4 potential performance gain at same diversity
2022	Frontier lies strictly above the actual cohort; max diversity improvement $k$5; max performance improvement $k$6
2023	After committee access to the SPF, chosen cohort lies essentially on the boundary; Pareto-improvement opportunities effectively zero $k$7

The 2023 cycle is especially important in the paper’s interpretation. After the selection committee was given access to the SPF in mid-2023, the committee reportedly swapped out a small number of borderline high-score, low-diversity candidates for moderate-score, high-diversity candidates, thereby moving the selected cohort to the frontier (Natarajan et al., 7 Oct 2025). The empirical claim is therefore not only retrospective—that earlier cohorts were Pareto-inferior with respect to the measured dimensions—but also prospective, in that presentation of the frontier altered decision-making.

6. Interpretation, scope, and limitations

The SPF is explicitly limited to precisely defined, quantitative dimensions. The paper states that qualitative factors such as interpersonal dynamics and holistic review judgments remain outside the model (Natarajan et al., 7 Oct 2025). Accordingly, SPF-based evaluation should not be interpreted as an exhaustive account of all selection-relevant considerations. It is a frontier over the variables encoded in $k$8 and $k$9, not over all values a committee may care about.

The approximation guarantee also has clear limits. The worst-case bound of $P(c)/k$0 can be loose, and actual gaps may vary by dataset (Natarajan et al., 7 Oct 2025). In addition, the discretization of $P(c)/k$1 controls the resolution of the frontier, so finer resolution increases runtime linearly. This establishes a practical trade-off between computational expense and frontier granularity.

A further technical limitation concerns the diversity functional form. The paper notes that entropic, cross-entropy diversity functions can fail submodularity if improperly combined, and that the implementation therefore relies on piecewise linear $P(c)/k$2 surrogates that preserve submodularity (Natarajan et al., 7 Oct 2025). This is significant because the greedy guarantee depends on monotonicity and submodularity of the scalarized objective. A plausible implication is that modeling choices for diversity are not merely normative; they are also algorithmically consequential.

The paper also identifies several extensions: alternative optimization heuristics such as stochastic local search, simulated annealing, and MIP relaxations; multi-objective frontiers that incorporate geographic spread, disciplinary balance, or budget constraints; and robust-optimization variants that account for uncertainty in performance scores or demographic self-reports (Natarajan et al., 7 Oct 2025). These extensions indicate that the SPF is a framework rather than a fixed metric.

7. Relation to Pareto efficiency in cohort selection

Within the paper’s framework, the SPF serves as the efficient boundary for fixed-size selection under jointly valued performance and diversity objectives. Its central analytical contribution is to make Pareto inefficiency observable in settings where organizations previously lacked a tractable way to measure it (Natarajan et al., 7 Oct 2025). In the case study, the 2021 and 2022 cohorts were identified as Pareto-inferior on the measured dimensions, whereas the 2023 cohort was selected on the frontier after the committee had access to the approximation.

This distinguishes the SPF from rank-order admission, quota systems, or unconstrained diversity maximization. Rank-order selection optimizes performance alone; pure diversity maximization ignores performance; quota-like approaches may target attribute counts without mapping the full feasible trade-off surface. By contrast, the SPF enumerates efficient trade-offs at a given cohort size and thereby supports decisions that are Pareto-efficient with respect to the specified metrics (Natarajan et al., 7 Oct 2025).

The broader significance of the SPF is therefore methodological. It reframes diverse talent selection as a frontier-estimation problem over a combinatorial cohort space, using monotone submodular objectives and greedy approximation to recover an actionable boundary. This suggests that the principal value of the SPF lies less in prescribing a single correct cohort than in clarifying which choices are efficient, which are dominated, and how much improvement is feasible under the organization’s own formal criteria.