Course Overview

Scientific computing is a broad term, and encompasses everything from a small script to plot a dataset for publication, to a high performance numerical PDE solver that is running on a supercomputer. Underlying and supporting all these implementations are two important elements, without which the field of scientific computing could not exist. The first is the development and publishing of general purpose algorithms for solving common problems, such as algorithms for the solution of ordinary differential equations (e.g. Euler or Runge-Kutta methods), or to minimise a non-linear function (e.g. Newtons method or Nelder–Mead). The second is the development of software frameworks for ease of implementation of these algorithms, on both desktop and high performance computers. The most important category of software frameworks is the ubiquitous Linear Algebra libraries for representing and operating on matrices and vectors. The BLAS APIs that were developed in the 1970s were hugely important in developing a common set of functions for linear algebra that could be backed by processor-specific implementations for high execution speed. After these came higher level libraries such as ATLAS and LAPACK, leading eventually to mass market linear algebra applications like MATLAB.

This course attempts to give an introduction to a useful set of algorithms for scientific computing, along with practical skills in implementation of these algorithms using linear algebra and other scientific libraries. On the algorithm side, we will focus on:

  • Linear algebra direct solvers and matrix decompositions
  • Sparse matrices and iterative solvers
  • Solution of Ordinary Differential Equations
  • Non-linear Optimisation
  • Bayesian Inference and MCMC sampling

For a two week course it is impractical to give a proper treatment of all of these topics. Therefor we will attempt to touch on some of the more important algorithms within each topic and provide links to further reading to broaden your knowledge.

On the software side, we will focus exclusively on implementation in Python, using the Numpy library for multidimensional arrays, and the Scipy library for both sparse matrices and many, many other general purpose scientific computing algorithms.