I'll try to repeat the calculations with a different demand curve, to demonstrate my point. But this will take some time and Shabbat is in soon, so that will have to wait.
Let's assume the demand curve - the number of transactions demanded as a function of the fee, per 10 minutes - is d(p) = 27/(8000p^2). It's safe to have d(p)->infinity as p->0 because supply is bounded (if there was no bound on supply, we'd need a more realistic bound on demand to have meaningful results). The behavior below is the same for other reasonable demand curves, as long as demand diminishes superlinearly with p (sublinear decay is less reasonable economically, and results in very different dynamics).
We'll assume 4000 transactions go in a MB, and that T=1MB. So the penalty, as a function of the number n of transactions, is f(n) = max(n-4000,0)^2 / (4000*(8000-n)).
We'll also assume that transactions are in no particular rush - users will pay the minimal fee that gives them a good guarantee to have the tx accepted in reasonable time (where this time is long enough to include blocks from the different miner groups). So there is a specific fee p for which the tx demand clears with the average number of txs per block (the number of txs can change between blocks). It would have been more interesting to analyze what happens when probabilistic urgency premiums enter the scene, but that's not relevant to the issue of mining centralization.
Scenario 1: 100 1% miners.
Each miner reclaims 1% of the penalty. If the optimal strategy is to have n txs per block, resulting in a fee of p, then n=d(p) and the marginal penalty (derivative of f) at n, corrected for the reclaiming, must equal p (so that adding another transaction generates no net profit). In other words, 0.99f'(d(p)) = p. Solving this gives p = 0.7496 mBTC, n = 6007.
Penalty is 0.5053 BTC, so pool size is 50.53.
Miners get 4.5027 BTC per block (6007 * 0.0007496 from txs + 0.5053 collection - 0.5053 penalty).
6007 txs are included per block.
Scenario 2: One 90% miner and 10 1% miners.
The market clears with a tx fee of p, with the 90% miner including n0 txs per block and the 1% miners including n1 txs per block.
The average #txs/block must equal the demand, so 0.9n0 + 0.1n1 = d(p).
Every miner must have 0 marginal profit per additional transaction, correcting for reclaiming. So
0.1 f'(n0) = p
0.99 f'(n1) = p
Solving all of this results in:
n0 = 7251
n1 = 5943
p = 0.6885 mBTC (lower than in scenario 1)
Penalty paid by 1% miners: f(5943) = 0.4589 BTC
Penalty paid by 90% miner: f(7251) = 3.5294 BTC
Average penalty: 0.9*3.5294 + 0.1*0.4589 = 3.2223 BTC
Pool size: 322.23 BTC
Reward per block for 1% miner: 5943 * 0.0006885 + 3.2223 - 0.4589 = 6.8552 BTC (more than in scenario 1)
Reward per block for 90% miner: 7251 * 0.0006885 + 3.2223 - 3.5294 = 4.68521 BTC (less than 1% miners in this scenario; more than the miners in scenario 1).
Average number of txs per block: 0.9 * 7251 + 0.1 * 5943 = 7120, more than in scenario 1.
Miners are happy - big or small, they gain more rewards.
Users are happy - more of their transactions are included, at a lower fee.
Nodes are not happy - they have to deal with bigger blocks.
Exactly as with the previously discussed demand curve.
Over time, difficulty will go up, nullifying the extra mining reward; and whoever is in charge of placing the checks and balances, will make the function tighter (or hold on with making it looser), to keep the block sizes at the desired level.
There is another issue at play here - the ones who benefit the most from the big miner's supersized blocks, are the small miners. The big miner could threaten to stop creating supersized blocks if the small miners don't join and create supersized blocks themselves. Forming such a cartel is advantageous over not having supersized blocks at all - however, I think the big miner's bargaining position is weak, and small miners will prefer to call the bluff and mine small blocks, avoiding the penalty and enjoying the big miner's supersized blocks. This is classic tragedy of the commons, but in a sort of reverse way - usually, TotC is discussed in this context when the mining cartel wants to exclude txs, not include them.