An estimate of the minimal number of degrees of freedom or independent parameters needed to locally represent data or a solution manifold, often reflecting a manifold's local topological or geometric dimension rather than the ambient space.

Principle

Local structure can be approximated by a lower-dimensional parameterization: intrinsic dimension counts the independent directions of local variability that suffice to reconstruct local neighborhoods up to acceptable error.

Demonstration

Points sampled from a smooth two-dimensional surface embedded in R^3 have intrinsic dimension 2: locally a tangent plane with two coordinates parameterizes neighborhoods, even though the ambient dimension is 3.

Misapplication

Confusing intrinsic dimension with ambient dimension or treating a global linear dimensionality (PCA) as the intrinsic dimension of a strongly curved or multi-scale manifold; or reporting a single global number when local dimension varies.

Consequence

Identifying intrinsic dimension guides model selection, sampling density, choice of reduced coordinates, and regularization; correctly estimated intrinsic dimension can reduce model complexity and improve generalization.

Reversal

The reverse view is extrinsic or ambient dimension: counting all coordinates of the embedding space rather than the minimal local degrees of freedom, which may overstate necessary complexity.

Boundary

Intrinsic dimension is meaningful locally and for structures that behave like manifolds at the scale of interest; it excludes fractal or highly noisy datasets without scale separation, and multiple definitions (topological, Hausdorff, correlation) may disagree.

Semantic Tension

Tension arises between manifold (topological) dimension, fractal (Hausdorff) dimensions and algorithmic 'effective' dimensions (e.g., number of large PCA eigenvalues); each captures a different notion of 'dimension' under different assumptions.

Synthesis

Intrinsic dimension is the local minimal count of independent coordinates needed to describe variability on a data or solution manifold; it is a scale‑ and definition‑dependent quantity that informs dimensionality reduction and modeling choices.