I submitted an article to Code Project yesterday, Simple Random Number Generation, describing a small C# class called SimpleRNG that uses George Marsaglia’s WMC algorithm. The article was posted around 5 PM (central US time) and comments started pouring in right away. I didn’t expect any feedback on a Friday afternoon or Saturday morning. But as I write this post, there have been 580 page views and 11 comments.
There have been three basic questions raised in the comments.
- Why not just use the random number generator that comes with .NET?
- Is this code suitable for cryptography?
- Is this code suitable for Monte Carlo applications?
Why not use the built-in generator? For many applications, the simplest thing would be to use the .NET random number generator. But there are instances where this might not be best. There are questions about the statistical quality of the .NET generator; I’ll get to that in a minute. The primary advantages I see to the SimpleRNG class are transparency and portability.
By transparency I mean that the internal state of the generator is simple and easy to access. When you’re trying to reproduce a result, say while debugging, it’s convenient to have full access to the internal state of the random generator. If you’re using your own generator, you can see everything. You can even temporarily change it: for debugging, it may be convenient to temporarily have the “random” generator return a very regular, predictable sequence.
By portability I do not necessarily mean moving the code between operating systems. The primary application I have in mind is moving the algorithm between languages. For example, in my work we often have prototype code written in R that needs to be rewritten in C++ for efficiency. If the code involves random number generation, the output of the prototype and the rewrite cannot be directly compared, only compared on average. Then you have to judge whether the differences are to be expected or whether they indicate a bug. But if both the R and the C++ code use the same RNG algorithm and the same seed, the results may be directly comparable. (They still may not be directly comparable due to other factors, but at least this way the results are often comparable.)
As for cryptography, no, SimpleRNG is not appropriate for cryptography.
As for Monte Carlo applications, not all Monte Carlo applications are created equal. Some applications do not require high quality random number generators. Or more accurately, different applications require different kinds of quality. Some random number generators break down when used for high-dimensional integration. I suspect SimpleRNG is appropriate for moderate dimensions. I use the Mersenne Twister generator for numerical integration. However, SimpleRNG is faster and much simpler; the MT generator has a very large internal state.
Someone commented on the CodeProject article that the random number generator in .NET is not appropriate for Monte Carlo simulation because it does not pass Marsaglia’s DIEHARD tests while SimpleRNG does. I don’t know what algorithm the .NET generator uses, so I can’t comment on its quality. Before I’d use it in statistical applications, I’d want to find out.