Optimization with rank-structured matrices
Event details
| Date | 19.11.2024 |
| Hour | 16:15 › 17:15 |
| Speaker | Prof. Martin Skovgaard Andersen - DTU Orbit, Denmark |
| Location | |
| Category | Conferences - Seminars |
| Event Language | English |
Computational Mathematics Seminar
Abstract :
Simple optimization methods commonly suffer from slow convergence near a minimizer, and yet such methods are often a de facto choice when dealing with large-scale problems with a large number of variables and/or large quantities of data. This talk presents some examples of how this issue can be addressed using rank-structured matrices as building blocks. Such matrices have a low-rank structure that admits a storage-efficient representation and fast algebraic operations (e.g., matrix-vector multiplication, factorization, etc.), and they arise frequently in engineering and data science, e.g., in applications that involve kernel functions. Examples include semiseparable (SS), hierarchical off-diagonal low rank (HODLR), and hierarchical semiseparable (HSS) matrices, all of which allow for linear or quasi-linear time and space complexity in key operations.
Abstract :
Simple optimization methods commonly suffer from slow convergence near a minimizer, and yet such methods are often a de facto choice when dealing with large-scale problems with a large number of variables and/or large quantities of data. This talk presents some examples of how this issue can be addressed using rank-structured matrices as building blocks. Such matrices have a low-rank structure that admits a storage-efficient representation and fast algebraic operations (e.g., matrix-vector multiplication, factorization, etc.), and they arise frequently in engineering and data science, e.g., in applications that involve kernel functions. Examples include semiseparable (SS), hierarchical off-diagonal low rank (HODLR), and hierarchical semiseparable (HSS) matrices, all of which allow for linear or quasi-linear time and space complexity in key operations.
Practical information
- Informed public
- Free
Organizer
- Prof. Daniel Kressner
Contact
- Prof. Daniel Kressner