I’ve been looking into fixing some of the sonar issues. One highlighted a test
of the MiddleSquareWeylSequence which just executes a code path but has no
assertions. This is for the default seeding (i.e. a seed less than the correct
seed length). Currently the generator uses the same method as all the others
for array seeds. It fills it with random numbers using the fillState method:
public MiddleSquareWeylSequence(long[] seed) {
if (seed.length < SEED_SIZE) {
final long[] tmp = new long[SEED_SIZE];
fillState(tmp, seed);
setSeedInternal(tmp);
} else {
setSeedInternal(seed);
}
}
When I added an assertion it shows that a zero length seed requires over 2^29
cycles to produce non zero long output. This is 2^58 int outputs, which is the
primary output of the generator.
This is because the default seeding fillState method results in a low
complexity increment for the Weyl sequence. The full seed is actually [0, 0, 2]
and the increment ends up as 3 (as it is forced to be odd). This has only 2
bits set in the 64 bit increment and a single 0 to 1 transition of the bit
state. FYI increments used in the Weyl sequence in SplittableRandom recommend
at least 24 bit state transitions in the 64 bit increment.
I suggest changing the constructor to use a default seed that contains a good
Weyl increment. This seed is overwritten with the input seed values if present.
Something like this:
public MiddleSquareWeylSequence(long[] seed) {
if (seed.length < SEED_SIZE) {
// Default high quality seed
final long[] tmp = {0x012de1babb3c4104L, 0xc8161b4202294965L,
0xb5ad4eceda1ce2a9L};
for (int i = 0; i < seed.length; i++) {
tmp[i] = seed[i];
}
setSeedInternal(tmp);
} else {
setSeedInternal(seed);
}
}
This ensures that the generator cannot be created with a bad Weyl increment
unless it is explicitly passed in via the seed as the increment is the final
position in the input seed array.
Note that this generator requires special seeding in commons-rng-simple module
that creates generators. It would seem appropriate to make a better effort to
avoid a bad generator in the core component too.
Alex
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
