Feat : added prefold approach

[ocdbytes]

ZK-WHIR Prefolding Protocol (Mixed Arity)

This document describes the full prefolding protocol used to reduce over-sized polynomials to the target arity before entering the main ZK-WHIR flow, while preserving zero-knowledge and soundness.

---

Phase 0 — Commitment

Each polynomial (native and over-sized) is committed independently at its original arity.

For a polynomial

$$ f : {0,1}^L \to \mathbb{F} $$

the prover samples ZK preprocessing data.

• Masking polynomial

$$ msk $$

• Helper polynomials

$$ \hat g_0, \hat g_1, \ldots, \hat g_L, M $$

All helper polynomials are $\ell$-variate, with $\ell < L$.

The prover commits to the masked polynomial

$$ \hat f = f + msk $$

using Interleaved Reed–Solomon (IRS):

• Evaluate on an RS domain
• Arrange evaluations into a matrix
• Merkle-commit to the matrix

Helper polynomials

$$ [M, \hat g_1, \ldots, \hat g_L] $$

are batch-committed via a separate helper-WHIR configuration.

The IRS interleaving depth is

$$ k = 2^d $$

where

$$ d = L - n_{\min} $$

Each Merkle opening reveals a coset of $k$ evaluations, which is exactly what is required for folding in Phase 2.

---

Phase 1 — ZK Blinding Setup

This phase is shared across all groups (native and prefold) and establishes the Fiat–Shamir challenges that bind the protocol.

Sample

$$ \beta \leftarrow \mathbb{F} $$

from the transcript.

Blinding Polynomial

For each polynomial, construct

$$ g(x) = \hat g_0(x_0,\ldots,x_{\ell-1}) + \sum_i \beta^i \cdot x_{i-1} \cdot \hat g_i(x_0,\ldots,x_{\ell-1}) $$

• $g$ is a uniformly random $L$-variate polynomial
• For $i > \ell$, fresh variable slots are used so that $g$ spans all $L$ variables

The prover sends

$$ g(a_i) $$

for each constraint point $a_i$.

Blinded Polynomial

Sample mixing scalar

$$ \rho \leftarrow \mathbb{F} $$

Define

$$ P = \rho \cdot f + g $$

Since $g$ is uniform over $\mathbb{F}^{2^L}$, $P$ is a perfect one-time pad on $\rho \cdot f$.

Modified evaluations are computed as

$$ \tilde v_{i,j} = \rho \cdot v_{i,j} + g_j(a_i) $$

Both prover and verifier can compute these values.

---

Phase 2 — Prefold

This phase is executed for each group with arity $L > n_{\min}$, starting from the highest arity.

---

Step 2.1 — Random Linear Combination (RLC)

For polynomials $P_1,\ldots,P_N$ in the group, combine

$$ P_{\mathrm{comb}} = \sum_j \lambda_j \cdot P_j $$

Combine constraints

$$ C(x) = \sum_i c_i \cdot eq(a_i, x) $$

Compute the claimed sum

$$ S = \sum_i c_i \cdot \left( \sum_j \lambda_j \cdot \tilde v_{i,j} \right) $$

---

Step 2.2 — Partial Sumcheck

Prove

$$ \sum_{x \in {0,1}^L} P_{\mathrm{comb}}(x), C(x) = S $$

Run only

$$ d = L - n_{\min} $$

rounds of sumcheck.

For each round:

• Prover sends a degree-2 polynomial $s_t(X)$
• Verifier checks

$$ s_t(0) + s_t(1) = S_{t-1} $$

and samples randomness $r_t$.

After $d$ rounds:

• Fold randomness

$$ r = (r_1,\ldots,r_d) $$

• Reduced claim $v'$

If the group has no constraints, $r$ is sampled directly from Fiat–Shamir.

---

Step 2.3 — Fold

Define the folded polynomial

$$ P'(x_0,\ldots,x_{n-1}) = P_{\mathrm{comb}}(x_0,\ldots,x_{n-1}, r_1,\ldots,r_d) $$

Each coefficient of $P'$ is the multilinear evaluation of a block of $2^d$ coefficients of $P_{\mathrm{comb}}$.

---

Step 2.4 — Reveal Folded Polynomial

The prover sends all coefficients of $P'$ into the transcript.

Since

$$ P' = \rho \cdot \mathrm{fold}(f, r) + \mathrm{fold}(g, r) $$

and $\mathrm{fold}(g,r)$ is uniformly random, this perfectly hides

$$ \rho \cdot \mathrm{fold}(f,r) $$

while binding the prover to $P'$.

---

Step 2.5 — Binding Equation Check

Split each evaluation point

$$ a_i = (a_i^{\mathrm{low}}, a_i^{\mathrm{high}}) $$

where

$$ a_i^{\mathrm{low}} \in {0,1}^n \qquad a_i^{\mathrm{high}} \in {0,1}^d $$

Using

$$ eq(a_i,(y,z)) = eq(a_i^{\mathrm{low}},y)\cdot eq(a_i^{\mathrm{high}},z) $$

the verifier checks

$$ v' = \sum_i c_i \cdot eq(a_i^{\mathrm{high}}, r) \cdot P'(a_i^{\mathrm{low}}) $$

Any incorrect evaluation causes this check to fail.

---

Step 2.6 — STIR Consistency Check

This step ensures $P'$ is the fold of the committed polynomial.

The verifier opens $\hat f$ at $q$ random IRS query points.

Each query reveals a coset

$$ {\hat f(\gamma_0),\ldots,\hat f(\gamma_{k-1})} $$

The prover provides helper evaluations, allowing the verifier to compute

$$ h(\gamma) = m(\gamma,\rho) + \sum_i \beta^i \cdot \gamma^{2^{i-1}} \cdot \hat g_i(\mathrm{pow}(\gamma)) $$

and

$$ L(\gamma) = \rho \cdot \hat f(\gamma) + h(\gamma) $$

By construction, $L(\gamma) = P(\gamma)$ on the evaluation domain.

The verifier:

1. Verifies helper polynomials via a sub-WHIR proof
2. Folds each coset using $r$ to obtain $V_\alpha$
3. Checks

$$ V_\alpha = P'(\mathrm{pow}(\alpha)) $$

A mismatch at any query implies cheating with overwhelming probability (Schwartz–Zippel).

---

Phase 3 — Main ZK-WHIR

After all prefold groups are verified, the remaining native-arity polynomials enter the standard ZK-WHIR protocol at arity $n_{\min}$.

This includes the full sumcheck, STIR folding, and final coefficient checks.

Unified API

The low-level prefold API requires the caller to manually create configs, sample preprocessings, commit, and wire up group inputs. The new batch_prove_zk / batch_verify_zk API abstracts all of this — the caller just provides polynomials, weights, and evaluations grouped by arity:

// Prover
let groups = vec![
    ProverInput::new(vec![&poly_10var], weights_10, evals_10),
    ProverInput::new(vec![&poly_12var], weights_12, evals_12),
];
let (point, evals) = config.batch_prove_zk(&mut prover, &params, &groups, &mut rng);

// Verifier
let claims: Vec<VerifierInput<_>> = groups.iter().map(|g| g.to_verifier_input()).collect();
let result = config.batch_verify_zk(&mut verifier, &params, &claims);

The library automatically identifies the native group, creates prefold configs for higher arities, samples ZK preprocessing, commits, and routes to the appropriate prove/verify path (falling back to plain prove_zk when all groups share the same arity).

--

Summary

• Over-sized polynomials are folded down to the target arity
• Zero-knowledge is preserved via polynomial blinding
• Partial sumcheck derives folding randomness
• STIR binds folded polynomials to their commitments
• Native ZK-WHIR runs unchanged after prefolding

This achieves sound, ZK-secure batching across mixed arities.