Separating Oblivious and Adaptive Models of Variable Selection

A new theoretical study establishes a sharp separation between two models of variable selection in sparse recovery, showing that the sample complexity required for optimal error guarantees can differ dramatically depending on whether the signal is fixed or can change adaptively. The work, posted to arXiv by researcher Yusong Zhu, examines sparse recovery with ℓ∞ error guarantees, a framework motivated by variable selection tasks where the goal is to estimate the support of a k-sparse signal in ℝ^d [1][2]. Sparse recovery itself is a foundational problem in learning theory and high-dimensional statistics, fields that underpin many modern machine learning algorithms [2][3]. Machine learning broadly relies on statistical algorithms that learn from data and generalize to unseen examples, often framed as empirical risk minimization [3]. The paper’s central contribution is a provable separation between what it terms the “oblivious” and “adaptive” models [1][2]. Under the oblivious model, the optimal ℓ∞ error is attainable in near-linear time using approximately k log d samples [1][2]. In the adaptive model, however, any algorithm requires at least roughly k^2 samples to achieve the same bound [1][2]. This finding establishes a contrast with the standard ℓ2 setting, where approximately k log d samples suffice even for adaptive sparse recovery [1][2]. The distinction hinges on how the signal is presented. In the oblivious case, a single measurement matrix works for each fixed signal, while the adaptive case demands a matrix that works for all signals simultaneously [2]. The result quantifies a fundamental cost of universality in this high-dimensional inference problem. The research also includes a preliminary examination of a partially-adaptive model, where nontrivial variable selection guarantees remain possible with approximately k log d measurements [1][2]. The study contributes to the theoretical understanding of problem-solving in computational statistics, a domain where specialized techniques are developed to overcome formal, fact-based obstacles [4]. The work was submitted on February 18, 2026, and last revised on June 22, 2026 [1].