Deriving Decoder-Free Sparse Autoencoders from First Principles