Derive the computational complexity (in Big O notation) of the standard self-attention mechanism in Transformer models with respect to sequence length N and embedding dimension d. Then, analyze the mathematical basis of efficient attention variants like FlashAttention or Linear Attention. Explain how these methods reduce the memory footprint from O(N^2) to O(N) by tiling or kernel fusion, and discuss the trade-offs in numerical precision or approximation error.