The RISC Zero STARK Protocol

The implementation in code for the RISC Zero STARK prover can be seen here. In this document, we present an overview to the RISC Zero STARK protocol, as well as a sequence diagram and a detailed description below. The STARK by Hand explainer and the RISC Zero ZKP Whitepaper are good companions to this document.

Overview

RISC Zero's receipts are built on the shoulders of several recent advances in the world of zero-knowledge cryptography. The core of the proof system is STARK-based, implementing DEEP-ALI & FRI. This proof system is used to generate zero-knowledge validity proofs for RISC Zero's RISC-V circuit and RISC Zero's recursion circuit. Users may also be interested in reading about the RISC Zero Groth16 Circuit, which enables on-chain verification.

At a high level, the design of the RISC Zero STARK protocol is very similar to the system described in ethSTARK, and the system implemented in Winterfell.

Setup Phase

The protocol includes a two-part setup phase; the first setup happens once per zkVM version, and the second setup establishes the Image ID for a given RISC-V binary file.

Part 1: Circuit Setup

This setup is transparent and establishes the public parameters for the prover & verifier. These public parameters include the number and length of the trace columns, the choice of hash function and Merklization structure, as well as a full enumeration of the constraints that are to be enforced.

Part 2: Program Setup

This phase establishes an Image ID, which is determined transparently from a RISC-V binary file and the circuit parameters. The Image ID is constructed by loading the RISC-V binary file into the zkVM memory, and then recording a Merkle snapshot of the full machine state. This setup can be repeated by anyone with access to the binary file, in order to confirm the correctness of the Image ID.

Main Trace & Auxiliary Trace

After the setup phase, the Prover executes the binary in the zkVM, computes a Low-Degree Extension on each column, and commits the Extended Main Execution Trace. Then, the prover computes and commits the Extended Auxiliary Execution Trace which depends on verifier randomness.

Compared to ethSTARK, our protocol adds an additional round of interaction to support constraints beyond basic AIR constraints. Using constraints that may span both the main trace and the auxiliary trace, we proceed with DEEP-ALI & FRI as described in ethSTARK. Adding an Auxiliary Execution Trace enables various enhancements, relative to a Vanilla STARK protocol. These enhancements are described well in From AIRs to RAPs.

We use this Auxiliary Execution Trace to support:

1. A permutation argument for memory verification
   The permutation argument is currently implemented as a grand product accumulator argument, as in PLONK. We plan to change this to a log derivative accumulator argument in the next version of the circuit.
   Here, operations corresponding to memory are committed to the main trace both in the original ordering and the permuted ordering, and grand product accumulators are committed in the auxiliary trace.
2. A lookup argument for range checks
   The lookup argument is currently implemented using the approach described in PLOOKUP. We plan to change this to a log derivative accumulator argument in the next version of the circuit.
   Here, the tables and the witness are committed in the main trace, and grand product accumulators are committed in the auxiliary trace.
3. A big integer accelerator to enable fast cryptographic operations
   The bigint accelerator implements multiplication of a and b by asking the host to provide the product c as non-deterministic advice. Then, the verifier provides randomness r, and the constraints enforce that when a, b, and c are interpreted as polynomials, a(r) * b(r) == c(r).
   Here, a, b, and c are committed in the main trace, and the evaluations at r are committed in the auxiliary trace.

DEEP-ALI & FRI

The rest of the protocol implements with DEEP-ALI & FRI as described in EthSTARK. We describe this in more detail below, and refer readers to the ZKP Whitepaper for a more formal description of the protocol.

Sequence Diagram

Detailed Step-by-Step Description

In this section, we elaborate on the sequence diagram above. For a more formal articulation of the protocol, refer to the ZKP Whitepaper.

Extended Main Execution Trace

• The Prover runs a computation in order to generate an Execution Trace.
  • The trace is organized into columns, and the columns are categorized as control columns, data columns, and auxiliary/accum columns.
    • The control columns handle system initialization and shutdown, the initial program code to load into memory before execution, and other control signals that don't depend on the program execution.
    • The data columns contain the input and the computation data, both of which are private. These columns are committed in two orderings:
      • in order of program execution, and
      • re-ordered by register first and clock cycle second. The re-ordered columns allow for efficient validation of RISC-V memory operations.
    • The auxiliary/accum columns are used for a permutation argument, a lookup argument, and a big integer accelerator circuit.
  • After computing the data columns and auxiliary/accum columns, the Prover adds some random noise to the end of those columns in order to ensure that the protocol is zero-knowledge.
• The Prover encodes the trace as follows:
  • The Prover converts each column into a polynomial using an iNTT. We'll refer to these as Trace Polynomials, denoted $P_i(x)$.
  • The Prover evaluates the data polynomials and the control polynomials over an expanded domain. The evaluations of the data polynomials and the control polynomials over this larger domain is called the Extended Main Execution Trace.
  • The Prover commits the Extended Main Execution Trace into two separate Merkle Trees, sending the roots to the Verifier.

Extended Auxiliary Execution Trace

• Using the transcript-thus-far as an entropy-source, we choose some random extension field elements, using a SHA-2 CRNG.
• Then, the Prover uses the randomness to generate the auxiliary/accum columns. The Prover computes the Low-Degree Extension of the auxiliary columns to form the Extended Auxiliary Execution Trace.
• The Prover commits the Extended Auxiliary Execution Trace to a Merkle tree and sends the Merkle root to the Verifier.
• Using the transcript-thus-far as an entropy-source, we choose a random constraint mixing parameter $\alpha$, using a SHA-2 CRNG.

DEEP-ALI (part 1)

• The Prover uses the constraint mixing parameter, the Trace Polynomials, and the Rule Checking Polynomials to construct a few Low Degree Validity Polynomials. The details are as follows:
  • By writing $k$ publicly known Rule Checking Polynomials, $R_0, R_1, ..., R_{k-1}$, in terms of the private Trace Polynomials, the Prover generates $k$ Constraint Polynomials, $C_j(x)$.
    • The key point about these polynomials is that for each of the $k$ rules and each input $z$ that's associated with the trace, $C_j(z)$ will return 0 if the trace "passes the test," so to speak.
  • Using the constraint mixing parameter $\alpha$, the Prover combines the Constraint Polynomials, $C_j$ into a single Mixed Constraint Polynomial, $C$, by computing $C(x)=\alpha^0C_0(x)+\ldots+\alpha^{k-1}C_{k-1}(x).$
    • Note that if each $C_j$ returns 0 at some point $z$, then $C$ will also return 0 at $z$.
  • Using a publicly known Zeros Polynomial, the Prover computes the High Degree Validity Polynomial, $V(x)=\frac{C(x)}{Z(x)}$.
    • The Zeros Polynomial $Z(x)$ is a divisor of any honest construction of $C(x)$. In other words, an honest prover will construct $V(x)$ to be a polynomial of lower degree than $C(x)$. We call $V$ "high degree" relative to the Trace Polynomials, $P_i$.
  • The Prover splits the High Degree Validity Polynomial into 4 Low Degree Validity Polynomials, $v_0(x), v_1(x), ..., v_3$.
  • The Prover evaluates the Low Degree Validity Polynomials, encodes them in a Merkle Tree, and sends the Merkle root to the Verifier.
  • We use Fiat-Shamir to choose an out-of-domain evaluation point, $z$.

DEEP-ALI (part 2)

• The Verifier would like to check the asserted relation between $C$, $Z$, and $V$ at the DEEP Test Point, $z$. Namely, the Verifier would like to confirm that $V(z)Z(z)=C(z)$.
  • The Prover sends the evaluations of each $v_i$ at $z$, which allows the Verifier to compute $V(z)$.
  • Computing $C(z)$ is slightly more complicated. Because rule checks can check relationships across multiple columns and multiple clock cycles, evaluating $C(z)$ requires numerous evaluations of the form $P_i(\omega^nz)$ for varying columns $i$ and cycles $n$. The Prover sends these necessary evaluations of each $P_i$ to allow the Verifier to evaluate $C(z)$. We refer to the necessary evaluations $P_i(\omega^nz)$ as the taps of $P_i$ at $z$.
  • The Verifier can now check $V(z)Z(z)=C(z)$.
  • Although these asserted evaluations have no associated Merkle branches, the DEEP technique offers an alternative to the usual Merkle proof.
• The Prover constructs the DEEP polynomials using the taps:
  • Denoting the taps of $P_i$ at $z$ as $(x_1,P_i(x_1)),\ldots,(x_n,P_i(x_n))$, the Prover constructs the DEEP polynomial $P'_i(x)=\frac{P_i(x)-\overline{P_i}(x)}{(x-x_1)\ldots(x-x_n)}$ where $\overline{P_i}(x)$ is the polynomial formed by interpolating the taps of $P_i$. The Prover computes $P'_i$, runs an iNTT on the result, and sends the coefficients of $P'_i$ to the Verifier. Using this technique, the Prover constructs and sends a DEEP polynomial for each $P_i$ and each $v_i$.
• At this point, the claim of trace validity has been reduced to the claim that each of the DEEP polynomials is actually a low-degree polynomial. To conclude the proof, the Prover mixes the DEEP polynomials into the FRI Polynomial using a DEEP mixing parameter and use the FRI protocol to show that the FRI Polynomial is a low-degree polynomial.

The FRI Protocol

• We omit the details of the DEEP-ALI & FRI for brevity.

Thanks for reading! If you have questions or feedback, we'd love to hear from you on Discord or Twitter.