I tripple-checked this: the call of the function earlier in my code has an influence on the outcome of a second call (with different variables involved) later on in my code. Random seed is not changed within. I’m puzzled.
10
All PRNGs (Pseudo-Random Number Generators) produce a deterministic sequence of values. When started from the same state, and used in an identical fashion, they will produce the same sequence every time. However, the observed sequence can be affected by using the underlying PRNG for different purposes.
PRNGs are designed to be uniformly distributed over their full range. However, when you choose a sub-range (a,b), the result is usually not uniformly distributed because of the pigeonhole principle — if you distribute more pigeons than there are pigeonholes, some pigeonholes will necessarily have more pigeons than others unless the number of pigeons is an integer multiple of the number of holes.
As a concrete example, suppose you have a 3-bit generator that produces any of 8 values U in the range [0..7], and you want to generate X which has values in the range [0..4]. You might be tempted to use X ← U % 5, but that would give the mapping 0→0, 1→1, 2→2, 3→3, 4→4, 5→0, 6→1, and 7→2. Note that there are twice as many ways to get 0, 1, and 2 than there are to get 3 or 4, so X is not uniformly distributed. Other mapping functions could be used but the pigeonhole principle still applies, so different 1-to-1 mappings would shift around which values occur more frequently but wouldn’t alter that the results are non-uniform.
Because of this, a good library implementation of a sub-range generator will use some form of rejection sampling and discard some of the generated values to achieve uniformity. (See this post for more details.) As a result, most of the raw random numbers are used, but when the rejection criteria are met one (or more) raw values will be discarded and the PRNG state is advanced for each discard. This would affect the observed outcomes of subsequent uses of the underlying generator.
If reproducibility is an issue, one solution for maintaining synchronization across multiple random processes is to create independently seeded PRNG instances for each distinct random process. In other words, don’t use the same PRNG instance to generate uniform(a,b)’s that you use for uniform(c,d)’s.