All publications

On the Intrinsic Dimensions of Data in Kernel Learning

Rustem Takhanov

A finite-stage Cantor dust branches into a covering-grid diagram for kernel-metric dimension and a spectral schematic for approximation-width dimension.
Original conceptual illustrationView full size

A finite-stage Cantor dust illustrates a domain with nontrivial geometry. The paper distinguishes dimension measured by coverings in the kernel metric from effective dimension defined through approximation widths; these need not coincide on fractal domains. Research source

About this paper

How does the geometry of data influence kernel learning? Two notions of intrinsic dimension offer complementary answers.

The analysis connects kernel-induced Minkowski dimension, Kolmogorov widths, and eigenvalue decay. It derives learning guarantees, proposes a sample-based estimation method, and shows that effective and geometric dimensions can differ on fractal domains.

Poster & presentation

The AISTATS poster and presentation slides are available in the resource links above.

Preview the research poster

The preview loads from Google Drive when opened.

Citation

BibTeX
@inproceedings{takhanov2026intrinsicdimensions,
  title = {On the Intrinsic Dimensions of Data in Kernel Learning},
  author = {Takhanov, Rustem},
  year = {2026},
  booktitle = {AISTATS 2026},
  eprint = {2601.16139},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/pdf/2601.16139}
}

Related work