Balancing Variety and Sample Size: Optimal Parameter Sampling for Ariel Target Selection

Technical Summary: Balancing Variety and Sample Size in Ariel Target Selection

Problem Statement
The European Space Agency's upcoming Ariel mission, scheduled for launch in 2029, aims to characterize the atmospheres of approximately 1,000 transiting exoplanets to understand population-level trends in formation, evolution, and chemical composition. With a primary mission duration of four years and a confirmed catalog of over 6,000 transiting exoplanets, the mission faces a constrained optimization problem: selecting a subset of targets that maximizes scientific return within a finite observational time budget. The core challenge is to balance the quantity of observed planets against the diversity of their physical parameters (e.g., radius, equilibrium temperature, stellar temperature) to ensure robust constraints on global population trends, specifically linear relationships.

Methodology
The authors frame the target selection problem as a submodular maximization task with a knapsack constraint, where the "cost" is the required observation time ($t_i$) and the "budget" is the total mission time ($T$). To address this, the study evaluates three distinct objective functions ($f$) and implements various heuristic algorithms to maximize them:

1. Objective Functions:

   • Sample Size ($f = N$): Maximizes the count of observed planets, prioritizing those with the lowest observation time (Time-greedy).
   • Sample Variance ($f = V$): Maximizes the variance of selected parameters to ensure diversity, prioritizing outliers regardless of cost (Variance-greedy).
   • Sample Leverage ($f = L$): Defined as $L = N \times \text{Var}(p)$, this function balances sample size and diversity. It is motivated by the statistical principle that the uncertainty on the slope of a linear trend is inversely proportional to leverage.

2. Selection Heuristics:
   The study tests five specific strategies for maximizing the leverage objective function, alongside the greedy approaches for size and variance:

   • Leverage-greedy: An iterative algorithm that selects planets based on the marginal gain in leverage per unit of observation time ($v_i/t_i$).
   • Simulated Annealing: A probabilistic search method that explores the solution space to avoid local maxima, allowing for temporary acceptance of suboptimal moves.
   • Regular Classes: Binning parameter space into fixed intervals and selecting the cheapest target from each bin.
   • Quantile Classes: Binning parameter space such that each bin contains an equal number of planets, ensuring populated classes.
   • K-means Clustering: Using unsupervised machine learning to group planets into $K$ clusters based on parameter similarity, then cycling through clusters to select targets.

The experiments were conducted on the Ariel Mission Candidate Sample (MCS), filtered to 1,342 planets capable of achieving Tier 2 signal-to-noise ratios within 20 visits. The study performed three exercises using 1, 2, and 3 diversity axes (Planet Equilibrium Temperature; Planet Radius + Temperature; and adding Stellar Effective Temperature).

Key Results
The numerical results, summarized in Table 1 and illustrated in Figures 4–11, reveal distinct trade-offs among the strategies:

• Time-greedy selection successfully maximizes the number of observed planets (e.g., 765 in the 1D case) but yields low leverage and poor diversity, as it clusters around the most easily observable, typical planets.
• Variance-greedy selection prioritizes diversity by selecting outliers, but this comes at a severe cost to sample size (e.g., dropping to ~326 planets in the 1D case) and results in a lower leverage score (32.99) compared to the leverage-greedy approach. It oversamples rare cases and undersamples the central population, potentially biasing the understanding of typical planetary characteristics.
• Leverage-greedy selection emerges as the most effective compromise. In the 1D experiment, it selected 530 planets with a leverage score of 34.08, outperforming all other methods (including variance-greedy) in maximizing leverage while maintaining a robust sample size.
• Simulated Annealing provided the second-highest leverage scores but with significantly larger sample sizes (e.g., 698 planets in the 1D case) compared to the leverage-greedy method. The authors note that while its leverage is slightly lower, its resulting sample is more representative of the underlying population distribution.
• Class-based and Clustering methods (Regular, Quantile, K-means) produced results intermediate between the greedy extremes. They offered structured coverage of parameter space but generally achieved lower total leverage than the explicit leverage-optimizing algorithms.
• Scaling: The study confirms that leverage scales approximately with the square root of the number of diversity axes ($\sqrt{N_p}$), consistent with the mathematical definition of the objective function.

Significance and Claims
The paper claims to provide a methodological framework for optimizing exoplanet target selection under strict observational constraints. Its primary contribution is the demonstration that leverage-based selection offers a superior strategy for surveys aiming to constrain linear population-level trends, effectively balancing the need for a large sample size with the necessity of parameter diversity.

The authors assert that while time-greedy approaches fail to capture diversity and variance-greedy approaches sacrifice sample size and fail to achieve the highest leverage, the leverage-greedy algorithm (and to a slightly lesser extent, simulated annealing) provides the optimal trade-off for the Ariel mission's scientific goals. The work supports the broader community effort to ensure Ariel delivers a scientifically valuable sample of exoplanet atmospheres. The authors explicitly state that their analysis assumes perfect knowledge of planetary parameters and does not yet incorporate observational uncertainties or scheduling constraints (ephemerides), identifying these as necessary extensions for future work to bring the optimization closer to real mission planning.