I read your article and agree with the conclusions generally; however, I think DPOS+BFT finds a middle ground where it supports two points on the triangle at the same time.
DPOS BFT starts out with the high-latency, low-overhead “average” number of nodes assuming you value number of nodes on a log2 scale. Many clients can use this light-weight headers with time delay to validate the blockchain and eventually conclude blocks as final. However, it is also possible to run the high-overhead “average number of nodes” for those who care about low-latency finality and are willing to pay the price.
So the pertinent question becomes what is a sufficient number of validators. The value each additional validator brings decays while the cost grows with 2N in both latency and overhead depending upon your needs. Clearly 2 nodes is better than 1, and 4 is better than 2. On a log scale each incremental increase in validators implies sets of : 1, 2, 4, 8, 16, 32, 64, 128, etc. DPOS with 21 validators ranks between 4.39 on my decentralization scale while 128 validators ranks a 7. We can conclude that costs are growing N² while benefits are growing N (measuring benefit of decentralization on log2 scale).
Lastly there is the question of the quality of the individual participants. Assuming all participants are untrustworthy scum with 0 accountability forces you to accept longer latency to get equal confidence. Meanwhile, assuming all participants are highly trust worthy saints with off-chain accountability and liability and your latency for a reasonable person to trust a transaction confirmed falls.
Even though historical performance is no guarantee of future performance, you could probably gather statistics over the number of orphaned blocks to determine the transaction confirmation probability of the most recent block.