FAQ: Q: Which platform gives the most accurate results? A: This is not easy to say. [1] gives results for many platforms. In general, the lower the results the more accurate is the data processing. You'll see that different platforms give different results. Q: Why do the results depend on the platform ? A: The results indicate errors. Most of them are caused by the inherent inaccuracy of floating point numbers [2-4]. The errors depend on many factors, like representation of the number, accuracy, the math's library, accuracy of the cpu, etc. However, Matlab and Octave are using the double format (8-bytes) of the IEEE 754/854 standard for floating point numbers. Despite, the errors are not the same on all platforms ([3] explains the reasons). In any case, the following questions arise: Q: Who is responsible for this mess? A: Difficult to say. Several instances (Hardware manufacturer as well as software developers for operating systems, compilers, and math libraries) including the developers from Mathworks Inc. and OCTAVE are involved. All of them should put some effort into the next point. Q: What has to be done ? A: The programs of FLTTEST are one starting point. These can be used to identify differences. Then, each difference has to be investigated in detail. Decide which solution is the best (or explain why this is not possible). This solution should be implemented on each platform. (see also [2-4]) References: [1] http://www.dpmi.tu-graz.ac.at/~schloegl/matlab/fermat/ [2] http://HTTP.CS.Berkeley.EDU/~wkahan/ieee754status/ieee754.ps [3] http://www.validlab.com/goldberg/addendum.html [4] http://cch.loria.fr/documentation/IEEE754/ (c) by A. Schloegl 2002