Act as a Machine Learning Researcher. The standard self-attention mechanism in Transformer models has a quadratic complexity O(n^2) with respect to sequence length. Critically analyze the efficiency of 'Sparse Attention' mechanisms (e.g., Longformer, BigBird) and 'Linear Attention' approximations (e.g., Performer, Linformer). Propose a novel hybrid attention mechanism that combines local sliding window attention with global token attention for long-document summarization, and explain the mathematical implications for the complexity and memory footprint.