Solving clustered low-rank semidefinite programs arising from polynomial optimization
Abstract
We study a primal-dual interior point method specialized to clustered low-rank semidefinite programs requiring high precision numerics, which arise from certain multivariate polynomial (matrix) programs through sums-of-squares characterizations and sampling. We consider the interplay of sampling and symmetry reduction as well as a greedy method to obtain numerically good bases and sample points. We apply this to the computation of three-point bounds for the kissing number problem, for which we show a significant speedup. This allows for the computation of improved kissing number bounds in dimensions $11$ through $23$. The approach performs well for problems with bad numerical conditioning, which we show through new computations for the binary sphere packing problem.
- Publication:
-
arXiv e-prints
- Pub Date:
- February 2022
- DOI:
- 10.48550/arXiv.2202.12077
- arXiv:
- arXiv:2202.12077
- Bibcode:
- 2022arXiv220212077L
- Keywords:
-
- Mathematics - Optimization and Control;
- Mathematics - Metric Geometry;
- 90C22;
- 90C23;
- 52C17;
- 90-04
- E-Print:
- 28 pages