|« Previous: Lecture Summaries||Next: Lecture 2 Summary »|
Brief overview of the huge field of numerical methods, and outline of the small portion that this course will cover. Key new concerns in numerical analysis, which don't appear in more abstract mathematics, are (i) performance (traditionally, arithmetic counts, but now memory access often dominates) and (ii) accuracy (both floating-point roundoff errors and also convergence of intrinsic approximations in the algorithms).
Some discussion of how large matrices arise in practice, and gave a simple example of the discrete Laplacian matrix arising from a discretized version of Poisson's equation. Noted that large matrices in practice often have special structure, e.g. sparseness and symmetry, and it is very important to exploit this structure to make their solution practical.
Jumped right into a canonical dense-matrix direct-solver algorithm that we will use to illustrate some performance and accuracy concerns: Gaussian elimination. Briefly reviewed the basic algorithm, and used Trefethen's "graphical" trick to quickly estimate the number of additions+multiplications as roughly 2m3/3 for m×m problems. Regarding accuracy, one key question is how roundoff errors propagate in this algorithm, which turns out to be a very difficult and partially unsolved problem discussed in Trefethen chapter 20; another question is what to do with pivots that are nearly zero, which treated naively lead to roundoff disasters and lead to the solution of partial pivoting. We will return to both of these topics later in the course. Regarding performance, there are three key questions that we will return to in lecture 2: (0) how expensive is this in practice, (1) is counting arithmetic operations enough to predict performance, and (2) can one do better than Gaussian elimination?
The classic way to analyze performance is operation counts; from last time that flop count (real additions+multiplications) for Gaussian elimination is 2m3/3 for m×m problems. Show that this means 1000×1000 problems are now routine, but that 106×106 or larger problems (as commonly arise for PDEs) will require us to take advantage of some special structure.