What was already known
Two-sided length preservation needs order row sparsity at linear embedding dimension. Earlier one-sided lower bounds ruled out constant sparsity under additional structural assumptions.
AI involvement
The authors credit the Gemini 4 Colosseum harness with most technical proofs and derivations. They independently proved several structural lemmas, edited the presentation and checked the mathematical steps.
- AI disclosure and author checksAI disclosure and author checks
Importance
The construction shows that preventing collapse alone permits substantially sparser sketches at the same embedding dimension.
Technical verification
The authors report checking the mathematical arguments and auditing the correspondence between the Lean statements and both results.
Independent checking
Checked the mathematical steps and audited the formal definitions and theorem statements.
- Checker
- Paper authors
- Scope
- SparseStack can sketch to dimensions using nonzero entries per row and keeps the smallest singular value bounded below on each fixed r-dimensional subspace with constant success probability. A separate lower bound needs nonzeros per row for isotropic sketches satisfying the paper’s row-support overlap condition.
- Basis
- Reported by the linked checker
Formal checking
Checked the upper and lower bounds in the published Lean development.
- Checker
- Lean 4 kernel
- Scope
- SparseStack can sketch to dimensions using nonzero entries per row and keeps the smallest singular value bounded below on each fixed r-dimensional subspace with constant success probability. A separate lower bound needs nonzeros per row for isotropic sketches satisfying the paper’s row-support overlap condition.
- Basis
- Reported by the linked checker
- Lean development and stated scopeLean development and stated scope
Human understanding
The disclosure records author contributions and checking. A human account of the model-generated parts of the proof is undocumented.
Other useful sources
Correction and publication history
Reviewed the primary result and its documented AI and checking scope.
Read this published revision