Difference between revisions of "Numerical Errors"

From SourceWiki
Jump to navigation Jump to search
m (Protected "Numerics" [edit=sysop:move=sysop])
(No difference)

Revision as of 15:24, 10 November 2008

Unexpected things can happen when you use floating point numbers!

Introduction

Sooner or later, you're going to want to include real, aka floating-point, numbers in your programs. Beit for measured quantities, or parameterisations, but integers don't always cut it. Now this is all fine and as it should be. A word of warning, however. Programs containing floating point numbers can do very odd things indeed! Use them with care.

At core the reason that unexpected things happen is because you are sat in front of a digital computer, which is capable to storing and manipulating a finite number of discrete entities. This is contrasted against the infinite nature of real numbers. There are an inifinity of real numbers between o and 1, say, or indeed between 0.000001 and 0.0000011. Also there is no theoretical limit to the number of digits in a decimal fraction.

Digital computers must then approximate real numbers using a finite number of discrete ones, and it is this approximation which is the source of our surprises. Given the huge capacity of modern computers, it is easy to forget that this approximation. It's my goal in the sections below, to draw your attention to the areas in which ignoring this approximation will come back to bite you!

I didn't expect that!

Let's dive into our first example:

svn co http://source.ggy.bris.ac.uk/subversion-open/numerics/trunk ./numerics
cd numerics/examples/example1
make

There are two programs in this directory, one written in C, the other in Fortran. Let's run the C program first:

./surprising_c.exe

This program does some very simple arithmetic and yet we start to see odd things happening already! For example the trigonometric function tan can give us hugely different values depending upon whether we give it a single precision- (typically stored using 4 bytes) or a double precision- (8 byte) number. To be fair \tan(\pi/2) (radians) is undefined, and so it's hard to say what the right output should be. However, is does highlight that you need to be vigilant when using floating-point numbers and not to expect that your program's behaviour will always be benign.

Algorithms Matter

Deterministic, yes. Outputs, different