On Tue, Jan 20, 2015 at 11:17:44AM +0000, Michael Savage wrote: > Hi, > > I'm having trouble generating uniform random doubles in [0,1) with > arc4random. In games, the snippet: > > (double) arc4random() / (UINT32_MAX + 1.0) > > crops up multiple times, but that isn't utilising the full precision of > a double.
What precision does your program need? > If you do the equivalent with a 64bit random integer: > > (double) ((uint64_t) arc4random() << 32 | arc4random()) / (UINT64_MAX + > 1.0) > > then I can still see two issues. Firstly, we are now putting too much > precision into a double which is going to add slight bias when it > rounds. Secondly, `UINT64_MAX + 1.0` probably isn't going to give us > what we want because of floating point precision. > > Any ideas? The precision the program needs gives the number of randomness required. If double is suitable (i.e. if less than 53-bit precision is needed), then feed it with the needed amount of bits, as you do. Ex, to get 53 bit precision: lo = arc4random(); hi = arc4random() & ((1 << (53 - 32)) - 1); then use: (double)((uint64_t)hi << 32 | lo) / (1ULL << 53) HTH
