About a couple of side-topics
I believe we speak about "Argument" when a "Proof" of soundness depends on computational assumptions: so Arguments' soundness is weaker than Proofs' one, but often strong enough considering that current applications of cryptography always rely on computational hardness; and the counterpart is that, for example, NP Arguments can be ZK is a stronger way than NP Proofs (Statistically ZK vs Computationally ZK).
"of Knowledge" suffix is used (for both Proofs or Arguments) when the prover holds information that can be extracted efficiently via a "special setup" (special in an analogous way the Simulator is special).
Or try to think it in this way: "soundness flavour" and "knowing something (or not)" are two orthogonal properties, so you could have a "cartesian product" of combinations:
- ZK Proof
- ZK Argument
- ZK Proof of Knowledge
- ZK Argument of Knowledge (the "ARK" in ZK-SNARK)
1 and 3 have stronger(*) soundness, 2 and 4 only computational; for 3 and 4 an Extractor exist (that's the recognized way to prove Knowledge) which can efficiently get the information the prover holds (and it doesn't break ZK property in an analogous way the Simulator doesn't break soundness... halting/rewinding and all that stuff..)
so an Argument, which is weaker than a proof from a soundness perspective, can have a stronger Zero Knowledge level of security (statistical) than a proof (computational)? I mean, ok... but how come? Can't see why that would be possible.
If you get the point that "soundness" and "knowledge" are two distinct properties, you would also get that a third distinct property is "zero-knowledgeness", which could depend on former in a way that's not what you are expecting.
I'm only a geek and an avid reader, so I guess you could easily receive many better explanations than mine, however I want to suggest you chapter 4 of Oded Goldreich's Foundations of Cryptography Vol.1 ...really insightful...
(*) not only on statistical "level", in Proofs soundness is proved by logical derivation from context and axioms, so the common "classical/standard" demonstrations