What the study found
The study introduces INK-ESTIMATE, an algorithm that incrementally computes estimates of ridge leverage scores for large-scale kernel ridge regression (KRR). The authors report that it maintains a small sketch of the kernel matrix, uses a single pass over the matrix, and works with a fixed, small space budget.
Why the authors say this matters
The authors say this matters because large kernel matrices are hard to store, and ridge leverage score-based sampling can give strong reconstruction guarantees for Nystrom approximations. The study suggests that sequentially updating these estimates may make KRR more practical while preserving guarantees about approximation quality and statistical risk.
What the researchers tested
The researchers studied KRR in a sequential setting and developed INK-ESTIMATE, which incrementally updates estimates of ridge leverage scores from a small sketch of the kernel matrix. The sketch is updated without needing access to previously seen columns, and the method is designed around the effective dimension of the kernel matrix.
What worked and what didn't
According to the abstract, the sketch update works in a single pass and does not require revisiting earlier columns. The method provides strong approximation guarantees for the distance between the true kernel matrix and its approximation, as well as for the statistical risk of the approximate KRR solution at any time. The abstract does not describe any failures or cases where the method did not work.
What to keep in mind
The summary provided does not include numerical results, experimental comparisons, or dataset details. The abstract also does not describe limitations beyond noting that exact ridge leverage scores are as difficult to compute as a KRR solution.
Key points
- INK-ESTIMATE incrementally computes ridge leverage score estimates for sequential kernel ridge regression.
- The method maintains a small sketch of the kernel matrix and uses a fixed, small space budget.
- It can update without access to previously seen columns, so a single pass is sufficient.
- The abstract says the method has strong approximation guarantees for kernel-matrix distance and statistical risk.
- The summary does not report numerical experiments, comparisons, or specific limitations.
Disclosure
- Research title:
- Sequential ridge leverage score estimates enable Nystrom KRR
- Authors:
- Daniele Calandriello, Alessandro Lazaric, Michal Vaľko
- Institutions:
- Institut national de recherche en sciences et technologies du numérique
- Publication date:
- 2026-04-22
- OpenAlex record:
- View
Get the weekly research newsletter
Stay current with peer-reviewed research without reading academic papers — one filtered digest, every Friday.