A manifold is a lower-dimensional shape embedded in a higher-dimensional space that looks simple locally. A bent line can form a circle, while Earth’s surface is a two-dimensional manifold in three-dimensional space. section 7.1.4
The manifold assumption treats real-world high-dimensional data as lying on or near a lower-dimensional surface. PCA assumes that surface is linear; nonlinear manifold learning can instead preserve local structure and unroll shapes such as the Swiss roll. section 7.1.4
In machine learning, the distinction matters because:
- PCA can flatten nonlinear structure without keeping similar points nearby. section 7.1.4
- A nonlinear projection can unroll a two-dimensional manifold so that a linear model separates colored point classes. section 7.1.4
Figure 7.5 Comparison of PCA to a more complex manifold learning algorithm on the Swiss Roll data set.

Compare PCA’s flattened Swiss-roll representation with a nonlinear manifold-learning projection. section 7.1.4



