By the end of this lesson, you should be able to:
- Explain why projection can create false nearest neighbors.
- Test whether a nearest-neighbor pair remains close after adding another delayed coordinate.
- Interpret the false-nearest-neighbor fraction as a function of embedding dimension.
- Explain why a low plateau suggests a sufficient practical embedding dimension.
Why embedding dimension matters
A reconstruction made with too few coordinates can force distant parts of the true attractor to overlap. Points that are unrelated in the underlying dynamics may therefore appear artificially close. These are called false nearest neighbors.
Unfolding with another coordinate
For each reconstructed state in dimension \(m\), identify a nearest neighbor and then add the next delayed coordinate. If the pair separates dramatically in dimension \(m+1\), the apparent closeness was mainly a projection artifact. A common relative test compares the added-coordinate separation with the original distance \(R_m\).
Choosing a sufficient dimension
As \(m\) increases, the fraction of false nearest neighbors should usually fall. A low plateau suggests that adding more coordinates is no longer removing many projection overlaps. This provides a practical diagnostic for a sufficient embedding dimension.
Finite-data caution
False-nearest-neighbor estimates depend on delay choice, sampling, noise, thresholds, temporal-neighbor exclusion, and record length. A low FNN fraction is evidence that the reconstruction is reasonably unfolded under the chosen settings, not a proof of a unique optimal embedding dimension.
What should stay with you
- False nearest neighbors arise when projection makes dynamically distant states appear close.
- Adding embedding coordinates can unfold those projection overlaps.
- A low FNN plateau is a practical indicator of a sufficient embedding dimension, but the result depends on finite-data and threshold choices.
What does a high false-nearest-neighbor percentage usually indicate?
With too few embedding coordinates, unrelated parts of the attractor can overlap in projection and appear artificially close.