CS 6210: Matrix Computations
Review
Big ticket items
Linear algebra and calculus
Linear algebra background (abstract and concrete)
Vectors, spaces, subspaces, bases
Interpreting matrices: operators, mappings, quadratic forms
Canonical forms
Calculus with matrices
Sensitivity analysis and conditioning
Variational notation for derivatives
Optimization with quadratics
Lagrange multipliers and constraints
Matrix algebra
Ways to write matrix-matrix products
Blocked matrices and blocked algorithms
Graph structures: sparse, diagonal, triangular, Hessenberg, etc
LA structures: symmetric, skew, orthogonal, etc
Other structure: Toeplitz, Hankel, other special matrices
The big problems
\[\begin{aligned} Ax &= b \\ \mbox{minimize } \|Ax&-b\|^2 \\ Ax &= x \lambda \end{aligned}\]
The big factorizations
LU and company (\(LDL^T\) and Cholesky)
QR (economy and full)
SVD (economy and full)
Schur factorization
Symmetric eigendecomposition
Iterations
Iterative refinement
Stationary iterations (Jacobi, Gauss-Seidel, etc)
Krylov subspace definition
Approximation from a subspace and Galerkin
Characterization of CG and GMRES
Philosophical odds and ends
Identifying the right structure matters a lot
We need both algebra and analysis
When you don’t know what else to do... eigenvalues or SVD
I differentiate five expressions before breakfast!
What else?
There is a lot that I wish I could get to in a course like this. If it were a two semester course, perhaps I would! Three things come immediately to mind.
LA for data science (c.f. CS 6241)
Non-negative matrix factorizations
Tensors and tensor factorizations
More on factorization-based methods in stats/ML
The linear algebra of multivariate normals
Connections to convex optimization: active sets, quadratic programming, etc
Iterative methods (c.f. CS 6220)
More on multigrid and domain decomposition
More on other “data-sparse” matrices
More on elliptic PDEs, integral equations, etc
Eigensolvers
More on eigensolvers (especially iterative ones)
Much more on perturbation theory and sensitivity analysis
Matrix functions, and complex analysis connections
Connections to control theory
More on orthogonal polynomials
But there is always more to learn. If the course gave you a starting point to thinking about other corners of linear algebra that you care about for your research, then it was a success.
I enjoyed the class this semester. I hope you did as well.