Copying my reply to Arvind:
Thanks for writing the paper, it definitely captures a lot of issues with many kinds of PoS algorithms that I think are very important! Some comments:
> (Globally-Predictable Selfish Mining)
One issue with this section is that there are ways to make selfish mining unprofitable from an incentive point of view. Particularly, if a block is created that is not part of the main chain, we can penalize both that block and the block with the same height on the main chain by 1 coin, so a successful selfish mining attack even if successful would cost the attacker.
The more general principle behind this is that when developing incentives for a consensus algorithm, as a general rule, if you can tell that either A or B were faulty, but you can’t tell which one, then it’s best to penalize both A and B at least by some amount (a version of this insight could also be used to bring PoW’s selfish mining resistance arbitrarily close to 50%; a version that looks only at first-order “sister” blocks brings it to ~39% as I recall).
Another issue that I think you missed that is even more interesting is that the randomness itself can be strategically manipulated. See RANDAO beacon exploitability analysis, round 2 for my analysis on our earlier RANDAO proposal. The summary is that in RANDAO each block producer contributes to randomness used to select the immediately next block using data that they precommitted to, and so any block proposer can manipulate the randomness by not showing up, at the expense of sacrificing their ability to make that block. This seems expensive, but actually an attacker can simulate the paths where manipulating randomness lets them get more than 1 block of advantage, and altogether an attacker with infinite computing power can do chain reversion attacks with ~36% of total stake. I have a mathematical argument and a link to a simulation in that post that backs up the numbers; it’s surprisingly cool, showing a relationship between the probability a path favoring the attacker exists and limit ratios of diagonals of Pascal’s triangle.
Finally, here Beacon chain Casper mini-spec is our latest proposal. We aim to prove the security property that the algorithm’s security is essentially independent of randomness, because there are many validators participating simultaneously in each round and so it would be near-impossible for the attacker to “corrupt” even one round; a GHOST variant (not the naive one!) is used as the fork choice rule.
I think the thing to keep in mind re GHOST is that PoS is fundamentally different from PoW, in that in PoW making two blocks is twice as hard as making one block, but in PoS this is not the case; hence, it’s philosophically wrong to have a fork choice rule that “adds up” two different messages from the same validator that are in support of a message. Additionally, validators need to be selected well in advance so there are no exponential forking issues.