The effect on soundness error is negligible for our current field, but this introduces an assertion that could fail if we changed to a field with more elements in the "ambiguous" range.
* Specialize `InterpolationGate`
To cosets of subgroups of roots of unity. This way
- `InterpolationGate` needs fewer routed wires, bringing our minimum routed wires down from 28 to 25.
- The recursive `compute_evaluation` avoids some multiplications, saving 100~200 gates depending on `num_routed_wires`.
* Update test
* feedback
* 2 challenges, 28 routed wires
2 challenges gives certain checks approximately (field_bits - degree_bits) * 2 bits of security, so we maintain our target of 100 bits for circuits with 2^14 gates or fewer.
28 routed wires is the min for `InterpolationGate`. A lower number helps reduce proof sizes. We can go back to a high number if there's any strong reason to reduce our gate count (e.g. if we were trying to hit 2^12).
* Check FRI conjectured security
* Fix
For examlpe, if I change a test to use `ConstantArityBits(4, 5)`, I get
To efficiently perform FRI checks with an arity of 16, at least 152 wires are needed. Consider reducing arity.
* Automatically select FRI reduction arities
This way when a proof's degree changes, we won't need to manually update the `FriConfig`s of any recursive proofs on top of it.
For now I've added two methods of selecting arities. The first, `ConstantArityBits`, just applies a fixed reduciton arity until the degree has shrunk below a certain threshold. The second, `MinSize`, searches for the sequence of arities that minimizes proof size.
Note that this optimization is approximate -- e.g. it doesn't account for the effect of compression, and doesn't count some minor contributions to proof size, like the Merkle roots from the commit phase. It also assumes we're not using Merkle caps in serialized proofs, and that we're inferring one of the evaluations, even though we haven't made those changes yet.
I think we should generally use `ConstantArityBits` for proofs that we will recurse on, since using a single arity tends to be more recursion-friendly. We could use `MinSize` for generating final bridge proofs, since we won't do further recursion on top of those.
* Fix tests
* Feedback
* Refactor GMiMC code
Adds a sub-trait of `Field` called `GMiMCInterface`, which is similar to `PoseidonInterface`.
This lets us have different fields with different GMiMC constants in a type-safe way.
* Remove `Interface`
* Const generic for width
* Move some Field members to a Field64 subtrait
I.e. move anything specific to 64-bit fields.
Also, relatedly,
- Tweak a bunch of prover code to require `Field64`, since 64-bit stuff is used in a couple places, like the FRI proof-of-work
- Remove `bits()`, which was unused and assumed a 64-bit field
- Rename a couple methods to reflect that they're u64 variants
There are no functional changes.
* Field64 -> PrimeField
* Remove `exp_u32`, `kth_root_u32`
* PrimeField: PrimeField
* Move `to_canonical_biguint` as well
* Add back from_noncanonical_u128
* Disable ZK in large_config
Speeds up the tests from ~6m to ~1m (debug mode). `large_config` is crate-private so I don't think we need to worry about real users forgetting ZK, and I don't think ZK seems important in these tests, though we should probably have ZK enabled for a couple tests.
A couple tests need ZK or they fail; I added a TODO to look later.
This led to a few other changes:
- Fixed a bug where `trim` could truncate the final poly to a non-power-of-two length. This was improbable when ZK is on due to randomization.
- Gave a few methods access to the whole `CircuitConfig` vs `FriConfig` -- sort of necessary for the above fix, and I don't think there's much downside.
- Remove `cap_height` from `FriConfig` -- didn't really need it any more after giving more methods access to `CircuitConfig`, and having a single copy of the param feels cleaner/safer to me.
* PR feedback
It's just a wrapper around `Target`, which signifies that the wrapped `Target` has already been range checked. Should make it easier to audit code that expects bools.
I sort of "shifted" the loop in `fri_verifier_query_round` so that `fri_combine_initial` is called before the loop, and all `compute_evaluation` calls are in the loop (rather than the final one being outside). This lines up with my mental model of FRI, and I think it's more natural as it results in a loop with no branches, no `i - 1`s, and less state stored between iterations. Also added some comments etc.
Should be functionally equivalent to the old version.
* Bit of refactoring in FRI code
- Inline `OpeningSet[Target]` and their `verify` methods, as they had become fairly trivial wrappers
- Have the challenger observe the openings and generate alpha inside `verify_fri_proof`. Conceptually I think of it as part of the batch-FRI protocol, and it minimizes redundancy.
* Fix tests
No functional changes here. The biggest change was moving certain files into new directories like `plonk` and `iop` (for things like `Challenger` that could be used in STARKs or other IOPs). I also split a few files, renames, etc, but again nothing functional, so I don't think a careful review is necessary (just a sanity check).
We have two division methods: one "unsafe" one, which permits 0/0 = anything, and one "safe" one, for which 0/0 results in an unsatisfiable instance. The latter is slightly more expensive.
I switched a few calls over to safe division, where unsafe division didn't seem sound (or at least it wasn't obvious). For calls where unsafe division did seem sound, I added comments explaining why.
Closes#97.
