2. In practice, the assumption above is not realistic; miners know that if they create supersized blocks, they will make mining more attractive, thus attracting more miners, thus increasing the difficulty, thus reducing their profits. So longer-term it is actually not beneficial for them to do this. So in all likelihood, big miners will not create supersized blocks, and all miners will get the same reward per block, without any superlinear advantage. Because of this, the original objection that the method in the OP encourages centralization of mining, has been refuted.
I already said my original argument wasn't true because it was only true in cases where larger blocks provide a higher reward per block to the miner. Your algorithm doesn't do this.
That's what I was saying here;
What I wrote would only apply to penalties that don't reduce the reward below the target block size reward.
I agree that big miners will not create supersized blocks, but it isn't because it attracts more miners, it's because mining supersized blocks confers a competitive advantage to the other miners that don't. The net effect is I can't imagine a situation in which any miner would create any supersized blocks because doing so would immediately damage their position in the market.