Roughly how do permissionless blockchains, permissioned blockchains, and central databases compare on performance and structure?

Throughput and latency get better as you move from permissionless → permissioned → central DB; the blockchains' cost buys tolerance of untrusted writers, which the central DB has zero of.

The single idea behind the table: consensus is what costs you performance. A permissionless chain runs heavy consensus (Proof of Work) to tolerate many anonymous, untrusted writers, so it's slowest. A permissioned chain uses lighter Byzantine Fault Tolerance consensus among few known writers, so it's faster. A central DB does no consensus at all — fastest, but tolerates no untrusted writers because there's only one.

Byzantine Fault Tolerance (BFT) is the property of still agreeing on one correct answer while some participants are Byzantine — not merely crashed, but arbitrarily faulty: lying, sending different answers to different peers, colluding. (The name comes from the Byzantine Generals Problem, where generals besieging a city must agree on attack-or-retreat while some of them are traitors.) The classic algorithm, Practical BFT (PBFT), reaches agreement by having the known participants vote in rounds, and works as long as fewer than a third of them are malicious — cheap compared to Proof of Work, but it needs a known, fixed membership list, which is exactly why it fits permissioned chains and not permissionless ones.

Go deeper:

• Consensus (computer science) (Wikipedia) — why PoW vs BFT vs no-consensus is exactly what buys (or forfeits) tolerance of untrusted writers.