Optimal Harmonic Selection

Background: The scalability of Dataset Distillation (DD) measures how well performance improves as the size of the synthesized dataset increases, which is crucial for capturing global data characteristics.

Question / Future Work: While the proposed method improves scalability by operating in the frequency domain, further research could investigate how to optimally allocate synthetic data size ($M$) to capture the most relevant harmonics (i.e., finding a more dynamic or data-driven method for selecting the top-$k$ frequencies instead of using a fixed heuristic).