This is a rust library for building R1CS gadgets over prime fields, which are useful in SNARKs and other argument systems.
An R1CS instance is defined by three matrices, A
, B
and C
. These encode the following NP-complete decision problem: does there exist a witness vector w
such that Aw ∘ Bw = Cw
?
A gadget for some R1CS instance takes a set of inputs, which are a subset of the witness vector. If the given inputs are valid, it extends the input set into a complete witness vector which satisfies the R1CS instance.
The goal of this library is to make SNARK programming easy. To that end, we support a broad set of features, including some fairly high-level abstractions:
- Basic operations on field elements, such as multiplication, division, and comparisons
- Type-safe boolean operations, such as
GadgetBuilder::and
andGadgetBuilder::bitwise_and
- Type-safe binary operations, such as
GadgetBuilder::binary_sum
GadgetBuilder::assert_permutation
, which efficiently verifies a permutation using an AS-Waksman network- Methods for sorting lists of expressions, such as
GadgetBuilder::sort_ascending
- Methods for working with Merkle trees, such as
GadgetBuilder::merkle_tree_root
- Common cryptographic constructions such as Merkle-Damgård, Davies-Meyer, and Sponge functions
- R1CS-friendly primitives like MiMC, Poseidon and Rescue
Field
is a trait representing prime fields. An Element<F>
is an element of the prime field F
.
A Wire
is an element of the witness vector. An Expression<F>
is a linear combination of wires.
A BooleanWire
is a Wire
which has been constrained in such a way that it can only equal 0 or 1. Similarly, a BooleanExpression<F>
is an Expression<F>
which has been so constrained.
A BinaryWire
is a vector of BooleanWire
s. Similarly, a BinaryExpression<F>
is a vector of BooleanExpression<F>
s.
Here's a simple gadget which computes the cube of a BN128 field element:
// Create a gadget which takes a single input, x, and computes x*x*x.
let mut builder = GadgetBuilder::<Bn128>::new();
let x = builder.wire();
let x_exp = Expression::from(x);
let x_squared = builder.product(&x_exp, &x_exp);
let x_cubed = builder.product(&x_squared, &x_exp);
let gadget = builder.build();
// This structure maps wires to their (field element) values. Since
// x is our input, we will assign it a value before executing the
// gadget. Other wires will be computed by the gadget.
let mut values = values!(x => 5u8.into());
// Execute the gadget and assert that all constraints were satisfied.
let constraints_satisfied = gadget.execute(&mut values);
assert!(constraints_satisfied);
// Check the result.
assert_eq!(Element::from(125u8), x_cubed.evaluate(&values));
This can also be done more succinctly with builder.exp(x_exp, 3)
, which performs exponentiation by squaring.
You can define a custom field by implementing the Field
trait. As an example, here's the definition of Bn128
which was referenced above:
pub struct Bn128 {}
impl Field for Bn128 {
fn order() -> BigUint {
BigUint::from_str(
"21888242871839275222246405745257275088548364400416034343698204186575808495617"
).unwrap()
}
}
Suppose we wanted to hash a vector of Expression
s. One approach would be to take a block cipher like MiMC, transform it into a one-way compression function using the Davies-Meyer construction, and transform that into a hash function using the Merkle-Damgård construction. We could do that like so:
fn hash<F: Field>(
builder: &mut GadgetBuilder<F>,
blocks: &[Expression<F>]
) -> Expression<F> {
let cipher = MiMCBlockCipher::default();
let compress = DaviesMeyer::new(cipher);
let hash = MerkleDamgard::new_defaults(compress);
hash.hash(builder, blocks)
}
To verify that two lists are permutations of one another, you can use assert_permutation
. This is implemented using AS-Waksman permutation networks, which permute n
items using roughly n log_2(n) - n
switches. Each switch involves two constraints: one "is boolean" check, and one constraint for routing.
Permutation networks make it easy to implement sorting gadgets, which we provide in the form of sort_ascending
and sort_descending
.
Suppose we wish to compute the multiplicative inverse of a field element x
. While this is possible to do in a deterministic arithmetic circuit, it is prohibitively expensive. What we can do instead is have the user compute x_inv = 1 / x
, provide the result as a witness element, and add a constraint in the R1CS instance to verify that x * x_inv = 1
.
GadgetBuilder
supports such non-deterministic computations via its generator
method, which can be used like so:
fn inverse<F: Field>(builder: &mut GadgetBuilder<F>, x: Expression<F>) -> Expression<F> {
// Create a new witness element for x_inv.
let x_inv = builder.wire();
// Add the constraint x * x_inv = 1.
builder.assert_product(&x, &Expression::from(x_inv),
&Expression::one());
// Non-deterministically generate x_inv = 1 / x.
builder.generator(
x.dependencies(),
move |values: &mut WireValues<F>| {
let x_value = x.evaluate(values);
let x_inv_value = x_value.multiplicative_inverse();
values.set(x_inv, x_inv_value);
},
);
// Output x_inv.
x_inv.into()
}
This is roughly equivalent to the built-in GadgetBuilder::inverse
method, with slight modifications for readability.
The r1cs-zkinterface crate can be used to export these gadgets to the standard zkinterface format.
There is also a direct backend for bellman via the r1cs-bellman crate.
This code has not been thoroughly reviewed or tested, and should not be used in any production systems.