Efficient Discrete Position Design for Movable Antenna Systems: Low Complexity and Robustness

What the math gives to ML

The paper's transferable contribution is the recognition that a discrete placement objective is a monotone submodular set function under a 2-system feasibility constraint, enabling a much cheaper greedy search with a provable 1/3 approximation rather than exhaustive or branch-and-bound enumeration. This structure applies whenever a neural network must select a diverse subset of tokens, patches, features, experts, inducing points, or memory entries and the utility exhibits diminishing returns. The most promising adaptation is a constrained token or patch selector whose marginal utility is recomputed greedily while enforcing a spacing or diversity constraint, with noisy minibatch utility estimates used to test robustness to imperfect information.