Lurk – Language for Recursive ZK-SNARKs Inspired by Common Lisp and Scheme

yes, zk-SNARKs are useful because you can prove you executed a program without revealing inputs or outputs.

you can't send encrypted inputs to a zk-SNARK program and get encrypted outputs out, though. you're confusing this with homomorphic encryption.

iirc, with Lurk you could, say, convince someone you'd correctly solved a difficult NP-complete problem, by writing a Lurk program that validates a solution to that problem, and publishing the output (Solved) along with the proof steps without revealing your answer.

you could, theoretically, write an homomorphic encryption library in Lurk, and then when people send you their encrypted inputs you can send them encrypted outputs along with a proof you faithfully evaluated the FHE circuit.

this is actually missing from FHE - a malicious evaluator can do whatever the hell they want with your input data, or make up outputs completely, they just can't see what's inside.

that seems pretty neat to me actually, I wonder if the Lurk people are thinking about this. niche but it'd have its uses.

Close: the so-called 'public inputs' must be revealed because these are checked by the verifier. What's neat is that many (even 'most') of the so-called 'private inputs' need not be revealed. This lets you prove statements that have the form, "I know of secret values x, y, z, etc. such that the following arithmetic relationship holds: [statement including public and private inputs]." One of Lurk's contributions is that this 'statement' can express arbitrary computation, including anything requiring unbounded recursion. That's what makes it Turing complete. A traditional SNARK circuit is expressive enough to describe any computation, but the limitation on requiring a fixed size means you have to know (and are limited by) an upper bound on the size of the computation in advance.

> Is my understanding correct that lurk would be used instead of something like circom where circom can create simple arithmetic zk circuits, whereas lurk can create circuits that can describe arbitrary computations?

Yes, exactly. You can construct an arbitrary program (say `((lambda (x) (* x x)) public-input)`) and pass that whole expression to the verifier. The prover can then provide the claimed result, which will also be a public input to the verifier, along with a proof that the 'input' evaluates (according to Lurk semantics) to the 'output'. Note that both what I called 'input' and 'output' will be public-inputs used by the verifier.

> So lurk programs can point to previously proven (and published?) hashes to make provable claims about their calculations?

Yes, that is one use case. Take the example above: the whole lambda expression I gave as an input example will actually be represented as a 'hash' (a content-address for the complete expression). But the 'interesting' part of the function doesn't actually have to have been given to the verifier. Instead, the prover might have committed to this function previously. The prover then 'opens' the commitment on the provided input. This is possible because the verifier can construct the 'hash' corresponding to the 'input' without knowing the internal hashes of the entire expression. The verifier might only know she is verifying a statement of the form `(<secret function> public-input)`. But she still knows that the claimed output is indeed an evaluation of the pre-committed secret function because she trusts the cryptographic proof.

We have an example project dealing with exactly this situation. The examples need a lot of work, but you can see the idea here: https://github.com/lurk-lang/lurk-rs/tree/master/fcomm

This WIP commit adapts the example to use a more explicit form of commitments and should make clearer what's going on: https://github.com/lurk-lang/lurk-rs/pull/112/files#diff-d6b...

The initial example took advantage of the fact that all compound Lurk data is produced by hashing anyway.

> Is this like "merkle-izing" computation?

You merkle-ize the recipe/program. The computation itself is performed normally (by the prover), but only needs to happen once. Thereafter any number of verifiers can 'check the prover's work' and see that it was performed correctly. Both the input and output to the computation proven can be merkle roots — so the amount of private versus public data can be tuned to the needs of the application.

> Sorry I'm just really trying to wrap my head around this.

It's a lot to grok and takes a while (but not too bad if you're really interested). If you follow us on Twitter (https://twitter.com/LurkLab) you'll see when we publish some blog posts in which I hope to walk through examples like this in a lot more detail. That should give a very concrete sense for what is possible and what is involved from a programmer perpsective.

The original proof-of-concept was written in Common Lisp, and development of core language features often happens in that version. It's more flexible for rapid development, and the 'API' can be taken it at a glance a bit more simply. Having multiple implementations also has the benefit of keeping us honest (at the cost of some hopefully-ephemeral drift and maintenance).

On the one hand, Lurk allows expression of recursive computations (like in the cited example). This doesn't necessarily demand 'verifying previous proofs' (in proofs). However, recursive SNARKs (where proofs of proofs are verified ad infinitum) are useful for actually producing a single succinct proof of a potentially unbounded recursive computation.

So there is a practical relationship between the two recursive aspects, but they're not strictly the same thing. This is indeed a bit of a confusing subtlety.