Difference between rounding modes in computational science?

Question

Are there any instances of scientific numerical problems where the choice of rounding mode matters?

There are usually a number of different rounding modes available: to $0$, away from $0$, to $\pm\infty$, nearest ties to even, nearest ties down. Apart from various currency-manipulation problems where rounding is mandated by problem details, does anyone know an example of where choice of rounding mode might actually matter?

A numerically stable algorithm would be insensitive to round-off errors and hence to choice of rounding mode, but are there any other issues that might be relevant?

Answer 1

Using the rounding modes, it's possible to implement interval arithmetic.

Suppose that two positive quantities $x$ and $y$ are represented by the intervals $[x_l,x_u]$ and $[y_l,y_u]$. The product is represented in interval arithmetic by $[x_l \otimes y_l, x_u \otimes y_u]$, where $x_l\otimes y_l$ should be rounded down and $x_u\otimes y_u$ should be rounded up. Control over rounding modes makes it easy to do this.

Varying the rounding mode and rerunning your code is also a quick test of your algorithm- if the answers change a lot then you know that the algorithm is quite sensitive to round-off errors (and that you probably need to rewrite it!)

Answer 2

The choice of rounding mode is governed by IEEE standards--Ref [1] mentions IEEE 754 and IEEE 854. Without the standards, one could find different results when porting code from one machine to another (see Goldberg [1]). Portability would be important if you want to be check if the different answer you get on another machine is due to a bug versus due to differences in implementation. The article also gives examples such as $x^2-y^2$ being more accurate than $(x+y)(x-y)$ when $x>>y$ or $y<<x$. In such cases, rounding error may become more important. As for compiler support, an example is the C99 standard which supports IEEE floating point arithmetic.

Note: some situations call for something more rigorous than rounding modes--these are problems involving the use of "arbitrary precision arithmetic. In the Wikipedia article on the topic, a number of applications are given, as well as appropriate libraries.

Reference:

[1] David Goldberg, "What every computer scientist should know about floating-point arithmetic", ACM Computing Surveys (CSUR), 23(1), 5-48, 1991. See http://dl.acm.org/ft_gateway.cfm?id=103163&type=pdf.