PCG has had no peer review of its author's claims at all, and some of those claims are dubious. PRNGs are serious things and require serious deliberation.
I'm confused about why the author didn't just resubmit it elsewhere, but to me PCG as an example raises a lot of questions about the value of peer review today, and about how it is conducted (not that I think it does or doesn't have value, but it raises a lot of questions about it). At least from the author's perspective it seems the concerns were about presentation rather than substance. Speaking from personal experience, peer review can be quixotic and I put little weight on it not appearing in a peer-reviewed journal, as opposed to any specific criticisms that might have been raised in a review (given the attention to PCG at this point, if there were serious criticisms of it it seems a failing on the part of the journal not to address these).
So your argument is that rejection from peer-review counts as peer-review?
> I'm confused about why the author didn't just resubmit it elsewhere [...] concerns were about presentation rather than substance
As far as I understand, the PCG paper is double the size it should be to get accepted by a serious journal. It contains a lot of unnecessary explanations of basic facts, making it unnecessarily tiresome to read for an expert. Presentation matters, because it stands between you and the substance, and you have to be able to get to the substance as fast as possible to review the paper.
> [...] (given the attention to PCG at this point, if there were serious criticisms of it it seems a failing on the part of the journal not to address these).
Well, the paper was rejected before it could come to that point...
Presentation does matter, but is the least important part and the most changeable. In this case it is also the part that has nothing to do with the actual performance of PCG.
I think she should have resubmitted it, but I think it says volumes about the problems of peer review that PCG is a widely known algorithm, widely discussed, and yet here we are discussing the merits of a particular presentation of it, as perceived by a couple of unknown persons.
I'd much rather have this HN discussion focused on open assertions of problems with PCG and rebuttals.
> I think she should have resubmitted it, but I think it says volumes about the problems of peer review that PCG is a widely known algorithm, widely discussed
PCG is only "widely discussed" because its inventor has personally and very heavily promoted it.
We only have the author's side of the story regarding the review. That it wasn't resubmitted for peer review is a huge red flag. The author has also publicly dismissed (what appears to me to be) valid and well-founded criticism of the algorithm, and made some unsubstantiated and one-sided charges about the academic peer review process. This has more than a faint odor of quackery about it, and that's not good about something as fundamentally important as a PRNG.
That’s not true? BigCrush and PractRand results have been (trivially) reproduced, performance is also easy to verify. What claims are dubious? Everything about PCG strikes me as serious and deliberate.
On the one hand, PRNGs aren't a very popular research subject AFAIK. The Blackman/Vigna PRNGs (which are the other recently popular PRNG family) likewise only got papers on Arxiv, AFAIK; so it's not clear to which peer-reviewed PRNG are you comparing PCGs. We are talking about non-crypto PRNGs here after all, which mostly get tested just experimentally, to check their statistical quality.
On the other hand, there is a really weird claim associated with the PCG family: that it is challenging to predict and that it thus "provides many of the benefits provided by CSPRNGs". This is how the PCG designer ends the abstract of her PCG paper:
> These functions can be designed to make it difficult for an adversary to discover the generator’s internal state by examining its output, and thus make it challenging to predict. This property, coupled with the ability to easily switch between random streams, provides many of the benefits provided by cryptographically secure generators without the overheads usually associated with those generators.
That's baloney. PCGs are not CSPRNGs, but she's deceptively trying to make it sound like they kind of are (have the same benefits). Anyway, a relatively efficient attack has been demonstrated this year: https://hal.inria.fr/hal-02700791/
Another issue with the "challenging predictability" design goal is that it seems to come at the cost of necessary trade-offs at the expense of speed and/or statistical quality, even though it is of dubious utility (considering that no real guarantee is given, and that a relatively efficient attack has recently been demonstrated): https://github.com/numpy/numpy/issues/16313#issuecomment-726...
My point was that PRNGs like the PCG family don't get peer-reviewed anyway (if they're not meant as CSPRNGs), although the way PCGs were presented is problematic because the author lays claim to "many of the benefits" of CSPRNGs, which would require peer-review for a start to be taken seriously.
Also, to be clear I suppose that using a PCG is a fine choice for, e.g., a simulation; just don't rely on it being "challenging" to predict.
