Talk:MIXMAX generator

The printable version is no longer supported and may have rendering errors. Please update your browser bookmarks and please use the default browser print function instead.

TestU01 statistical tests

@Intgr: As I understand you would like to know if MIXMAX has passed statistical tests. The U01 test was performed for the MIXMAX generator. The results of these tests are published in the article already refereed, K. Savvidy (2015) "The MIXMAX Random Number Generator". Comp.Phys.Communic. 196: 161–165. The article is also available in an open-access site: http://arxiv.org/abs/1403.5355. The Table 1 of the article represents the necessary data. In the last column of the Table 1 one can see that the MIXMAX generators of the dimension N bigger than N=88 are passing the BigCrush suite of tests (from TestU01). The default dimension recommended to the users is N=256. Sincerely George Savvidy (George Savvidy 19:30, 21 April 2016 (UTC))

@Savvidy: Well, it's not about "me wanting to know", but about properly satisfying the WP:Verifiability and WP:No original research policies. I read that arXiv copy, that's why I removed the claim.

What particularly annoyed me, the sources cited in the article didn't support the original claims that "it is one of only a few generators in general use which pass all statistical tests". Like I mention in my edit summary, there's no end of CSPRNGs and stream ciphers that pass every conceivable statistical test, so it's provably false. Coupled with the COI editing, I get the impression that this is just promotion of this research paper. -- intgr [talk] 08:05, 22 April 2016 (UTC)[reply]

@Intgr: Would this [[1]] help shed some more light in this topic ? I understand that there's (probably) not such a thing as a perfect RNG in our imperfect physical universe, but perhaps this one looks very promising comparing to others. Anyways, I (also) understand your objection; Wikipedia has standards of objectivity and of being an encyclopedia for the layman, not for researchers. But then, one can find highly technical articles, say: Dirac_delta_function. So, maybe Wikipedia isn't concerned about the layman only, but also for students, and other people with technical knowledge that are looking for a short introduction and a list of bibliography/references. Having said that, I'd kindly suggest to Dr @Savvidy: to consider writing a more detailed article in Scholarpedia (example: [[2]]) as well, since that encyclopedia's articles are created by researchers for researchers. Mlliarm (talk) 23:58, 26 July 2017 (UTC)[reply]

@Mlliarm: I had a look at the slides, but they don't look at all convincing to me. First, the author seems to be very partial, i.e. he wants to show MIXMAX is good and for this purpose he compares it with random number generators with well-known defects (Mersenne twister). Second, the PractRand screenshots in the slides do not show tests for long sequences (say, a few TB of random data. PractRand goes by default up to 32TB). Third, speed and memory usage DO matter for random number generators, and they are never mentioned!!! One cannot disregard such considerations: even the lousiest LCG can become excellent if you only use a few of the highest bits and/or if for each call you also compute but not return (ie skip) some values (to uncorrelated the output, RANLUX-style). MIXMAX claims a very long period, but this also implies high memory usage (because it has to store the state), which for MIXMAX must be at least 15KB; this is quite a lot, as common RNG's such as xoroshift* etc. only use ~128/256 bits. Also, no info is given about speed. If we disregard speed/memory we could just always use a good cipher in counter mode (AES, ChaCha...) and be done once and for all. L0rents (talk) 12:15, 25 November 2017 (UTC)[reply]

@L0rents: Thanks for the very informative reply. Mlliarm (talk) 10:36, 4 December 2017 (UTC)[reply]

MIXMAX passes statistical tests, including TestU01, Dieharder, GJRand and the last I heard also Pracrand with many terabytes of data consumed. The memory use of MIXMAX is considerably less than you suggest: in the officially recommended version, i.e. N=17 which has 17*8+8=144 byte state. The other comment is that CSPRNG are for various reasons not suitable for scientific calculations. The reason is performance but also absence of theoretical guarantees (unlike MIXMAX), and lastly also seeding convenience. This is why several important physics MC packages, such as Geant4, ROOT, and Pythia are adding or making MIXMAX the default. I hope this helps. Kotika98 (talk) 05:42, 21 December 2017 (UTC)[reply]

Is this niche random number generator notable?

I don't think so... Gnuish (talk) 05:38, 21 June 2019 (UTC)[reply]

@Gnuish, well...it is being used in the ROOT data library. Is CERN notable enough for you? Mlliarm (talk) 06:07, 17 January 2022 (UTC)[reply]

Thank you for that. Yes, CERN is notable. ROOT is notable. But, for example, MixMax is not mentioned in ROOT's own Wikipedia article. It's just a single algorithm implemented by a minor subroutine in a very large and very varied library. That's not normally notable, unless there is something more relevant about this algorithm. Also note that the article was created and mostly evolved by the creator of the algorithm (George Saviddy, see [3]), and this is the only article that they have ever contributed to in Wikipedia (Special:Contributions/Savvidy)! I have minor expertise in random number generators (mostly on the cryptographic side) and so far this one isn't standing out in the same way that e.g. Linear congruential generator is worth an encyclopedia article. Perhaps where MinMax should properly appear (without having its own article) is as an entry in the table of implementations of linear congruential generators in that article? Gnuish (talk) 00:40, 26 January 2022 (UTC)[reply]