From Linearity to Non-Linearity: How Masked Autoencoders Capture Spatial Correlations