bellperson 0.24.1

zk-SNARK library
Documentation
use std::convert::TryFrom;
use std::ops::{AddAssign, MulAssign};

use blstrs::Compress;
use ff::{Field, PrimeField};
use group::{prime::PrimeCurveAffine, Curve};
use log::{debug, info};
use rayon::prelude::*;
use serde::Serialize;

use super::{
    commit,
    commit::{VKey, WKey},
    compress, inner_product,
    poly::DensePolynomial,
    structured_scalar_power,
    transcript::Transcript,
    AggregateProof, AggregateProofAndInstance, GipaProof, KZGOpening, ProverSRS,
    ProverSRSInputAggregation, TippMippProof,
};
use crate::groth16::{aggregate::AggregateVersion, multiscalar::*, Proof};
use crate::SynthesisError;
use pairing::{Engine, MultiMillerLoop};

/// Aggregate `n` zkSnark proofs, where `n` must be a power of two.
/// WARNING: transcript_include represents everything that should be included in
/// the transcript from outside the boundary of this function. This is especially
/// relevant for ALL public inputs of ALL individual proofs. In the regular case,
/// one should input ALL public inputs from ALL proofs aggregated. However, IF ALL the
/// public inputs are **fixed, and public before the aggregation time**, then there is
/// no need to hash those. The reason we specify this extra assumption is because hashing
/// the public inputs from the decoded form can take quite some time depending on the
/// number of proofs and public inputs (+100ms in our case). In the case of Filecoin, the only
/// non-fixed part of the public inputs are the challenges derived from a seed. Even though this
/// seed comes from a random beeacon, we are hashing this as a safety precaution.
pub fn aggregate_proofs<E>(
    srs: &ProverSRS<E>,
    transcript_include: &[u8],
    proofs: &[Proof<E>],
    version: AggregateVersion,
) -> Result<AggregateProof<E>, SynthesisError>
where
    E: MultiMillerLoop + std::fmt::Debug,
    E::Fr: Serialize,
    <E::Fr as PrimeField>::Repr: Send + Sync,
    <E as Engine>::Gt: Compress + Serialize,
    E::G1: Serialize,
    E::G1Affine: Serialize,
    E::G2Affine: Serialize,
{
    info!("aggregate_proofs [version {}]", version);
    if proofs.len() < 2 {
        return Err(SynthesisError::MalformedProofs(
            "aggregating less than 2 proofs is not allowed".to_string(),
        ));
    }
    if !proofs.len().is_power_of_two() {
        return Err(SynthesisError::NonPowerOfTwo);
    }

    if !srs.has_correct_len(proofs.len()) {
        return Err(SynthesisError::MalformedSrs);
    }
    // We first commit to A B and C - these commitments are what the verifier
    // will use later to verify the TIPP and MIPP proofs
    par! {
        let a = proofs.iter().map(|proof| proof.a).collect::<Vec<_>>(),
        let b = proofs.iter().map(|proof| proof.b).collect::<Vec<_>>(),
        let c = proofs.iter().map(|proof| proof.c).collect::<Vec<_>>()
    };

    // A and B are committed together in this scheme
    // we need to take the reference so the macro doesn't consume the value
    // first
    let refa = &a;
    let refb = &b;
    let refc = &c;
    try_par! {
        let com_ab = commit::pair::<E>(&srs.vkey, &srs.wkey, refa, refb),
        let com_c = commit::single_g1::<E>(&srs.vkey, refc)
    };

    let hcom = Transcript::<E>::new("hcom")
        .write(&com_ab)
        .write(&com_c)
        .into_challenge();

    // Derive a random scalar to perform a linear combination of proofs
    let r = Transcript::<E>::new("random-r")
        .write(&hcom)
        .write(&transcript_include)
        .into_challenge();

    // 1,r, r^2, r^3, r^4 ...
    let r_vec: Vec<E::Fr> = structured_scalar_power(proofs.len(), &*r);
    // 1,r^-1, r^-2, r^-3
    let r_inv = r_vec
        .par_iter()
        .map(|ri| ri.invert().unwrap())
        .collect::<Vec<_>>();

    // B^{r}
    let b_r = b
        .par_iter()
        .zip(r_vec.par_iter())
        .map(|(bi, ri)| (bi.to_curve() * ri).to_affine())
        .collect::<Vec<_>>();
    let refb_r = &b_r;
    let refr_vec = &r_vec;
    try_par! {
        // compute A * B^r for the verifier
        let ip_ab = inner_product::pairing::<E>(refa, refb_r),
        // compute C^r for the verifier
        let agg_c = inner_product::multiexponentiation::<E::G1Affine>(refc, refr_vec)
    };

    // w^{r^{-1}}
    let wkey_r_inv = srs.wkey.scale(&r_inv)?;

    // we prove tipp and mipp using the same recursive loop
    let tmipp = prove_tipp_mipp::<E>(
        srs,
        &a,
        &b_r,
        &c,
        &wkey_r_inv,
        &r_vec,
        &ip_ab,
        &agg_c,
        &hcom,
        version,
    )?;
    debug_assert!({
        let computed_com_ab = commit::pair::<E>(&srs.vkey, &wkey_r_inv, &a, &b_r).unwrap();
        com_ab == computed_com_ab
    });

    Ok(AggregateProof {
        com_ab,
        com_c,
        ip_ab,
        agg_c,
        tmipp,
    })
}

pub fn aggregate_proofs_and_instances<E: Engine + std::fmt::Debug>(
    srs: &ProverSRSInputAggregation<E>,
    transcript_include: &[u8],
    statements: &[Vec<E::Fr>],
    proofs: &[Proof<E>],
    version: AggregateVersion,
) -> Result<AggregateProofAndInstance<E>, SynthesisError>
where
    E: MultiMillerLoop + std::fmt::Debug,
    E::Fr: Serialize,
    <E::Fr as PrimeField>::Repr: Send + Sync,
    <E as Engine>::Gt: Compress + Serialize,
    E::G1: Serialize,
    E::G1Affine: Serialize,
    E::G2Affine: Serialize,
{
    if statements.len() < 2 {
        return Err(SynthesisError::MalformedProofs(
            "aggregating less than 2 proofs is not allowed".to_string(),
        ));
    }
    assert!(statements.len() > 1);
    let n = statements[0].len();
    for s in statements {
        if s.len() != n {
            return Err(SynthesisError::MalformedProofs(
                "all statements must be equally sized".to_string(),
            ));
        }
    }
    let mut com_f: Vec<E::G1> = Vec::new();
    let mut com_w0: Vec<E::G1> = Vec::new();
    let mut com_wd: Vec<E::G1> = Vec::new();
    let mut f_eval: Vec<E::Fr> = Vec::new();
    let mut f_eval_proof: Vec<E::G1> = Vec::new();

    let mut poly_f: Vec<Vec<E::Fr>> = Vec::new();

    for j in 0..n / 2 {
        let mut poly_f_j: Vec<E::Fr> = Vec::new();
        let mut poly_w0: Vec<E::Fr> = Vec::new();
        for i in 0..proofs.len() {
            poly_f_j.push(statements[i][j]);
            if i < (proofs.len() - 1) {
                poly_w0.push(statements[i + 1][j]);
            }
        }

        let poly_f_cp = poly_f_j.clone();
        let poly_f_cp_2 = poly_f_j.clone();

        let calc_multiscalar = |left: &[E::Fr], table, len: usize| {
            let getter = |i: usize| -> <E::Fr as PrimeField>::Repr { left[i].to_repr() };
            par_multiscalar::<_, E::G1Affine>(
                &ScalarList::Getter(getter, len),
                table,
                std::mem::size_of::<<E::Fr as PrimeField>::Repr>() * 8,
            )
        };

        // compute F
        let com_f_j = calc_multiscalar(&poly_f_j, &srs.g_alpha_powers_table, proofs.len());

        // check that a0 is the zero coefficient of F
        let com_w0_j = calc_multiscalar(&poly_w0, &srs.g_alpha_powers_table, proofs.len() - 1);

        // Check F^x^(n - d) exists i.e. that F is bounded
        let com_wd_j = calc_multiscalar(&poly_f_cp, &srs.g_alpha_powers_end_table, proofs.len());

        com_f.push(com_f_j);
        com_w0.push(-com_w0_j);
        com_wd.push(-com_wd_j);
        poly_f.push(poly_f_cp_2);
    }

    let transcript_new = Transcript::<E>::new("transcript-with-coms")
        .write(&com_f)
        .write(&com_w0)
        .write(&com_wd)
        .write(&transcript_include)
        .into_bytes();

    let pi_agg = aggregate_proofs(srs, &transcript_new, proofs, version).unwrap();

    let hcom = Transcript::<E>::new("hcom")
        .write(&pi_agg.com_ab)
        .write(&pi_agg.com_c)
        .into_challenge();

    // Random linear combination of proofs
    let r = Transcript::<E>::new("random-r")
        .write(&hcom)
        .write(&transcript_new)
        .into_challenge();

    for poly_f_j in poly_f {
        let mut poly_eval = E::Fr::zero();
        let mut r_pow = E::Fr::one();
        for poly_f_j_i in &poly_f_j {
            poly_eval += *poly_f_j_i * r_pow;
            r_pow *= &*r;
        }
        f_eval.push(poly_eval);

        f_eval_proof.push(
            create_kzg_opening_for_instance::<E>(
                &srs.g_alpha_powers_table,
                DensePolynomial::from_coeffs(poly_f_j.clone()),
                poly_eval,
                &*r,
            )
            .unwrap(),
        );
    }

    Ok(AggregateProofAndInstance {
        num_inputs: u32::try_from(n).expect("too many statements"),
        pi_agg,
        com_f,
        com_w0,
        com_wd,
        f_eval,
        f_eval_proof,
    })
}

/// Proves a TIPP relation between A and B as well as a MIPP relation with C and
/// r. Commitment keys must be of size of A, B and C. In the context of Groth16
/// aggregation, we have that B = B^r and wkey is scaled by r^{-1}. The
/// commitment key v is used to commit to A and C recursively in GIPA such that
/// only one KZG proof is needed for v. In the original paper version, since the
/// challenges of GIPA would be different, two KZG proofs would be needed.
#[allow(clippy::too_many_arguments)]
fn prove_tipp_mipp<E>(
    srs: &ProverSRS<E>,
    a: &[E::G1Affine],
    b: &[E::G2Affine],
    c: &[E::G1Affine],
    wkey: &WKey<E>, // scaled key w^r^-1
    r_vec: &[E::Fr],
    ip_ab: &<E as Engine>::Gt,
    agg_c: &E::G1,
    hcom: &E::Fr,
    version: AggregateVersion,
) -> Result<TippMippProof<E>, SynthesisError>
where
    E: MultiMillerLoop,
    E::Fr: Serialize,
    <E::Fr as PrimeField>::Repr: Send + Sync,
    <E as Engine>::Gt: Compress + Serialize,
    E::G1: Serialize,
    E::G1Affine: Serialize,
    E::G2Affine: Serialize,
{
    let r_shift = r_vec[1];
    // Run GIPA
    let (proof, mut challenges, mut challenges_inv, extra_challenge) =
        gipa_tipp_mipp::<E>(a, b, c, &srs.vkey, wkey, r_vec, ip_ab, agg_c, hcom, version)?;

    // Prove final commitment keys are wellformed
    // we reverse the transcript so the polynomial in kzg opening is constructed
    // correctly - the formula indicates x_{l-j}. Also for deriving KZG
    // challenge point, input must be the last challenge.
    challenges.reverse();
    challenges_inv.reverse();
    let r_inverse = r_shift.invert().unwrap();

    // KZG challenge point
    let z = match version {
        AggregateVersion::V1 => Transcript::<E>::new("random-z")
            .write(&challenges[0])
            .write(&proof.final_vkey.0)
            .write(&proof.final_vkey.1)
            .write(&proof.final_wkey.0)
            .write(&proof.final_wkey.1)
            .into_challenge(),
        AggregateVersion::V2 => Transcript::<E>::new("random-z")
            .write(&extra_challenge)
            .write(&proof.final_vkey.0)
            .write(&proof.final_vkey.1)
            .write(&proof.final_wkey.0)
            .write(&proof.final_wkey.1)
            .write(&proof.final_a)
            .write(&proof.final_b)
            .write(&proof.final_c)
            .into_challenge(),
    };

    // Complete KZG proofs
    par! {
        let vkey_opening = prove_commitment_v(
            &srs.h_alpha_powers_table,
            &srs.h_beta_powers_table,
            srs.n,
            &challenges_inv,
            &z,
        ),
        let wkey_opening = prove_commitment_w(
            &srs.g_alpha_powers_table,
            &srs.g_beta_powers_table,
            srs.n,
            &challenges,
            &r_inverse,
            &z,
        )
    };

    Ok(TippMippProof {
        gipa: proof,
        vkey_opening: vkey_opening?,
        wkey_opening: wkey_opening?,
    })
}

/// gipa_tipp_mipp peforms the recursion of the GIPA protocol for TIPP and MIPP.
/// It returns a proof containing all intermdiate committed values, as well as
/// the challenges generated necessary to do the polynomial commitment proof
/// later in TIPP.
/// The extra challenge at the end is used to bridge the two proofs mipp/tipp
/// and the KZG checks.
#[allow(
    clippy::many_single_char_names,
    clippy::type_complexity,
    clippy::too_many_arguments
)]
fn gipa_tipp_mipp<E>(
    a: &[E::G1Affine],
    b: &[E::G2Affine],
    c: &[E::G1Affine],
    vkey: &VKey<E>,
    wkey: &WKey<E>, // scaled key w^r^-1
    r: &[E::Fr],
    ip_ab: &<E as Engine>::Gt,
    agg_c: &E::G1,
    hcom: &E::Fr,
    version: AggregateVersion,
) -> Result<(GipaProof<E>, Vec<E::Fr>, Vec<E::Fr>, E::Fr), SynthesisError>
where
    E: MultiMillerLoop,
    E::Fr: Serialize,
    <E::Fr as PrimeField>::Repr: Sync,
    <E as Engine>::Gt: Serialize,
    E::G1: Serialize,
    E::G1Affine: Serialize,
    E::G2Affine: Serialize,
{
    // the values of vectors A and B rescaled at each step of the loop
    let (mut m_a, mut m_b) = (a.to_vec(), b.to_vec());
    // the values of vectors C and r rescaled at each step of the loop
    let (mut m_c, mut m_r) = (c.to_vec(), r.to_vec());
    // the values of the commitment keys rescaled at each step of the loop
    let (mut vkey, mut wkey) = (vkey.clone(), wkey.clone());

    // storing the values for including in the proof
    let mut comms_ab = Vec::new();
    let mut comms_c = Vec::new();
    let mut z_ab = Vec::new();
    let mut z_c = Vec::new();
    let mut challenges: Vec<E::Fr> = Vec::new();
    let mut challenges_inv: Vec<E::Fr> = Vec::new();

    let mut c_inv: E::Fr = *Transcript::<E>::new("gipa-0")
        .write(hcom)
        .write(ip_ab)
        .write(agg_c)
        .write(&r[1])
        .into_challenge();
    let mut c = c_inv.invert().unwrap();

    let mut i = 0;

    while m_a.len() > 1 {
        // recursive step
        // Recurse with problem of half size
        let split = m_a.len() / 2;

        // TIPP ///
        let (a_left, a_right) = m_a.split_at_mut(split);
        let (b_left, b_right) = m_b.split_at_mut(split);
        // MIPP ///
        // c[:n']   c[n':]
        let (c_left, c_right) = m_c.split_at_mut(split);
        // r[:n']   r[:n']
        let (r_left, r_right) = m_r.split_at_mut(split);

        let (vk_left, vk_right) = vkey.split(split);
        let (wk_left, wk_right) = wkey.split(split);

        // since we do this in parallel we take reference first so it can be
        // moved within the macro's rayon scope.
        let (rvk_left, rvk_right) = (&vk_left, &vk_right);
        let (rwk_left, rwk_right) = (&wk_left, &wk_right);
        let (ra_left, ra_right) = (&a_left, &a_right);
        let (rb_left, rb_right) = (&b_left, &b_right);
        let (rc_left, rc_right) = (&c_left, &c_right);
        let (rr_left, rr_right) = (&r_left, &r_right);
        // See section 3.3 for paper version with equivalent names
        try_par! {
            // TIPP part
            let tab_l = commit::pair::<E>(rvk_left, rwk_right, ra_right, rb_left),
            let tab_r = commit::pair::<E>(rvk_right, rwk_left, ra_left, rb_right),
            // \prod e(A_right,B_left)
            let zab_l = inner_product::pairing::<E>(ra_right, rb_left),
            let zab_r = inner_product::pairing::<E>(ra_left, rb_right),

            // MIPP part
            // z_l = c[n':] ^ r[:n']
            let zc_l = inner_product::multiexponentiation::<E::G1Affine>(rc_right, rr_left),
            // Z_r = c[:n'] ^ r[n':]
            let zc_r = inner_product::multiexponentiation::<E::G1Affine>(rc_left, rr_right),
            // u_l = c[n':] * v[:n']
            let tuc_l = commit::single_g1::<E>(rvk_left, rc_right),
            // u_r = c[:n'] * v[n':]
            let tuc_r = commit::single_g1::<E>(rvk_right, rc_left)
        };

        // Fiat-Shamir challenge
        // combine both TIPP and MIPP transcript
        if i == 0 {
            match version {
                AggregateVersion::V1 => {
                    // already generated c_inv and c outside of the loop
                }
                AggregateVersion::V2 => {
                    // in this version we do fiat shamir with the first inputs
                    c_inv = *Transcript::<E>::new("gipa-0")
                        .write(&c_inv)
                        .write(&zab_l)
                        .write(&zab_r)
                        .write(&zc_l)
                        .write(&zc_r)
                        .write(&tab_l.0)
                        .write(&tab_l.1)
                        .write(&tab_r.0)
                        .write(&tab_r.1)
                        .write(&tuc_l.0)
                        .write(&tuc_l.1)
                        .write(&tuc_r.0)
                        .write(&tuc_r.1)
                        .into_challenge();

                    c = c_inv.invert().unwrap();
                }
            };
        } else {
            c_inv = *Transcript::<E>::new(&format!("gipa-{}", i))
                .write(&c_inv)
                .write(&zab_l)
                .write(&zab_r)
                .write(&zc_l)
                .write(&zc_r)
                .write(&tab_l.0)
                .write(&tab_l.1)
                .write(&tab_r.0)
                .write(&tab_r.1)
                .write(&tuc_l.0)
                .write(&tuc_l.1)
                .write(&tuc_r.0)
                .write(&tuc_r.1)
                .into_challenge();

            // Optimization for multiexponentiation to rescale G2 elements with
            // 128-bit challenge Swap 'c' and 'c_inv' since can't control bit size
            // of c_inv
            c = c_inv.invert().unwrap();
        }
        info!("prover: challenge {} -> {:?}", i, c);

        // Set up values for next step of recursion
        // A[:n'] + A[n':] ^ x
        compress(&mut m_a, split, &c);
        // B[:n'] + B[n':] ^ x^-1
        compress(&mut m_b, split, &c_inv);

        // c[:n'] + c[n':]^x
        compress(&mut m_c, split, &c);
        r_left
            .par_iter_mut()
            .zip(r_right.par_iter_mut())
            .for_each(|(r_l, r_r)| {
                // r[:n'] + r[n':]^x^-1
                r_r.mul_assign(&c_inv);
                r_l.add_assign(*r_r);
            });
        let len = r_left.len();
        m_r.resize(len, E::Fr::zero()); // shrink to new size

        // v_left + v_right^x^-1
        vkey = vk_left.compress(&vk_right, &c_inv)?;
        // w_left + w_right^x
        wkey = wk_left.compress(&wk_right, &c)?;

        comms_ab.push((tab_l, tab_r));
        comms_c.push((tuc_l, tuc_r));
        z_ab.push((zab_l, zab_r));
        z_c.push((zc_l, zc_r));
        challenges.push(c);
        challenges_inv.push(c_inv);

        i += 1;
    }

    assert!(m_a.len() == 1 && m_b.len() == 1);
    assert!(m_c.len() == 1 && m_r.len() == 1);
    assert!(vkey.a.len() == 1 && vkey.b.len() == 1);
    assert!(wkey.a.len() == 1 && wkey.b.len() == 1);

    let (final_a, final_b, final_c) = (m_a[0], m_b[0], m_c[0]);
    let (final_vkey, final_wkey) = (vkey.first(), wkey.first());
    let (final_zab_l, final_zab_r) = z_ab.last().unwrap();
    let (final_zc_l, final_zc_r) = z_c.last().unwrap();
    let (final_tab_l, final_tab_r) = comms_ab.last().unwrap();
    let (final_tuc_l, final_tuc_r) = comms_c.last().unwrap();
    // This extra challenge is simply done to make the bridge between the
    // MIPP/TIPP proofs and the KZG proofs, but is not used in TIPP/MIPP.
    let extra_challenge = *Transcript::<E>::new("gipa-extra-link")
        .write(&challenges.last().unwrap())
        .write(&final_a)
        .write(&final_b)
        .write(&final_c)
        .write(&final_zab_l)
        .write(&final_zab_r)
        .write(&final_zc_l)
        .write(&final_zc_r)
        .write(&final_tab_l.0)
        .write(&final_tab_l.1)
        .write(&final_tab_r.0)
        .write(&final_tab_r.1)
        .write(&final_tuc_l.0)
        .write(&final_tuc_l.1)
        .write(&final_tuc_r.0)
        .write(&final_tuc_r.1)
        .into_challenge();
    debug!("prover: extra challenge {:?}", extra_challenge);
    Ok((
        GipaProof {
            nproofs: a.len() as u32, // TODO: ensure u32
            comms_ab,
            comms_c,
            z_ab,
            z_c,
            final_a,
            final_b,
            final_c,
            final_vkey,
            final_wkey,
        },
        challenges,
        challenges_inv,
        extra_challenge,
    ))
}

fn prove_commitment_v<G>(
    srs_powers_alpha_table: &dyn MultiscalarPrecomp<G>,
    srs_powers_beta_table: &dyn MultiscalarPrecomp<G>,
    n: usize,
    transcript: &[G::Scalar],
    kzg_challenge: &G::Scalar,
) -> Result<KZGOpening<G>, SynthesisError>
where
    G: PrimeCurveAffine,
    <G::Scalar as PrimeField>::Repr: Send + Sync,
{
    // f_v
    let vkey_poly = DensePolynomial::from_coeffs(polynomial_coefficients_from_transcript(
        transcript,
        &G::Scalar::one(),
    ));

    // f_v(z)
    let vkey_poly_z = polynomial_evaluation_product_form_from_transcript(
        transcript,
        kzg_challenge,
        &G::Scalar::one(),
    );

    create_kzg_opening(
        srs_powers_alpha_table,
        srs_powers_beta_table,
        n,
        vkey_poly,
        vkey_poly_z,
        kzg_challenge,
    )
}

fn prove_commitment_w<G>(
    srs_powers_alpha_table: &dyn MultiscalarPrecomp<G>,
    srs_powers_beta_table: &dyn MultiscalarPrecomp<G>,
    n: usize,
    transcript: &[G::Scalar],
    r_shift: &G::Scalar,
    kzg_challenge: &G::Scalar,
) -> Result<KZGOpening<G>, SynthesisError>
where
    G: PrimeCurveAffine,
    <G::Scalar as PrimeField>::Repr: Send + Sync,
{
    // this computes f(X) = \prod (1 + x (rX)^{2^j})
    let mut fcoeffs = polynomial_coefficients_from_transcript(transcript, r_shift);
    // this computes f_w(X) = X^n * f(X) - it simply shifts all coefficients to by n
    let mut fwcoeffs = vec![G::Scalar::zero(); n];
    fwcoeffs.append(&mut fcoeffs);
    let fw = DensePolynomial::from_coeffs(fwcoeffs);

    par! {
        // this computes f(z)
        let fz = polynomial_evaluation_product_form_from_transcript(transcript, kzg_challenge, r_shift),
        // this computes the "shift" z^n
        let zn = kzg_challenge.pow_vartime(&[n as u64])
    };
    // this computes f_w(z) by multiplying by zn
    let mut fwz = fz;
    fwz.mul_assign(&zn);

    create_kzg_opening(
        srs_powers_alpha_table,
        srs_powers_beta_table,
        2 * n, // here we have twice the coefficients size
        fw,
        fwz,
        kzg_challenge,
    )
}

fn create_kzg_opening_for_instance<E: Engine>(
    srs_powers_alpha_table: &dyn MultiscalarPrecomp<E::G1Affine>, // h^alpha^i
    poly: DensePolynomial<E::Fr>,
    eval_poly: E::Fr,
    kzg_challenge: &E::Fr,
) -> Result<E::G1, SynthesisError> {
    let neg_kzg_challenge = -*kzg_challenge;

    // f_v(X) - f_v(z) / (X - z)
    let quotient_polynomial = &(&poly - &DensePolynomial::from_coeffs(vec![eval_poly]))
        / &(DensePolynomial::from_coeffs(vec![neg_kzg_challenge, E::Fr::one()]));

    let quotient_polynomial_coeffs = quotient_polynomial.into_coeffs();

    //// Compute the proof that F opens to f_v(z) at z as g^{quotient_polynomial(alpha)}.
    let pi_open = {
        let getter =
            |i: usize| -> <E::Fr as PrimeField>::Repr { quotient_polynomial_coeffs[i].to_repr() };

        par_multiscalar::<_, E::G1Affine>(
            &ScalarList::Getter(getter, quotient_polynomial_coeffs.len()),
            srs_powers_alpha_table,
            std::mem::size_of::<<E::Fr as PrimeField>::Repr>() * 8,
        )
    };

    Ok(pi_open)
}

/// Returns the KZG opening proof for the given commitment key. Specifically, it
/// returns $g^{f(alpha) - f(z) / (alpha - z)}$ for $a$ and $b$.
fn create_kzg_opening<G>(
    srs_powers_alpha_table: &dyn MultiscalarPrecomp<G>, // h^alpha^i
    srs_powers_beta_table: &dyn MultiscalarPrecomp<G>,  // h^beta^i
    srs_powers_len: usize,
    poly: DensePolynomial<G::Scalar>,
    eval_poly: G::Scalar,
    kzg_challenge: &G::Scalar,
) -> Result<KZGOpening<G>, SynthesisError>
where
    G: PrimeCurveAffine,
    <G::Scalar as PrimeField>::Repr: Send + Sync,
{
    let neg_kzg_challenge = -*kzg_challenge;

    if poly.coeffs().len() != srs_powers_len {
        return Err(SynthesisError::MalformedSrs);
    }

    // f_v(X) - f_v(z) / (X - z)
    let quotient_polynomial = &(&poly - &DensePolynomial::from_coeffs(vec![eval_poly]))
        / &(DensePolynomial::from_coeffs(vec![neg_kzg_challenge, G::Scalar::one()]));

    let quotient_polynomial_coeffs = quotient_polynomial.into_coeffs();

    // multiexponentiation inner_product, inlined to optimize
    let quotient_polynomial_coeffs_len = quotient_polynomial_coeffs.len();
    let getter = |i: usize| -> <G::Scalar as PrimeField>::Repr {
        if i >= quotient_polynomial_coeffs_len {
            return G::Scalar::zero().to_repr();
        }
        quotient_polynomial_coeffs[i].to_repr()
    };

    // we do one proof over h^a and one proof over h^b (or g^a and g^b depending
    // on the curve we are on). that's the extra cost of the commitment scheme
    // used which is compatible with Groth16 CRS insteaf of the original paper
    // of Bunz'19
    Ok(rayon::join(
        || {
            par_multiscalar::<_, G>(
                &ScalarList::Getter(getter, srs_powers_len),
                srs_powers_alpha_table,
                std::mem::size_of::<<G::Scalar as PrimeField>::Repr>() * 8,
            )
            .to_affine()
        },
        || {
            par_multiscalar::<_, G>(
                &ScalarList::Getter(getter, srs_powers_len),
                srs_powers_beta_table,
                std::mem::size_of::<<G::Scalar as PrimeField>::Repr>() * 8,
            )
            .to_affine()
        },
    ))
}

/// It returns the evaluation of the polynomial $\prod (1 + x_{l-j}(rX)^{2j}$ at
/// the point z, where transcript contains the reversed order of all challenges (the x).
/// THe challenges must be in reversed order for the correct evaluation of the
/// polynomial in O(logn)
pub(super) fn polynomial_evaluation_product_form_from_transcript<F: Field>(
    transcript: &[F],
    z: &F,
    r_shift: &F,
) -> F {
    // this is the term (rz) that will get squared at each step to produce the
    // $(rz)^{2j}$ of the formula
    let mut power_zr = *z;
    power_zr.mul_assign(r_shift);

    let one = F::one();

    let mut res = one + (transcript[0] * power_zr);
    for x in &transcript[1..] {
        power_zr = power_zr.square();
        res.mul_assign(one + (*x * power_zr));
    }

    res
}

// Compute the coefficients of the polynomial $\prod_{j=0}^{l-1} (1 + x_{l-j}(rX)^{2j})$
// It does this in logarithmic time directly; here is an example with 2
// challenges:
//
//     We wish to compute $(1+x_1ra)(1+x_0(ra)^2) = 1 +  x_1ra + x_0(ra)^2 + x_0x_1(ra)^3$
//     Algorithm: $c_{-1} = [1]$; $c_j = c_{i-1} \| (x_{l-j} * c_{i-1})$; $r = r*r$
//     $c_0 = c_{-1} \| (x_1 * r * c_{-1}) = [1] \| [rx_1] = [1, rx_1]$, $r = r^2$
//     $c_1 = c_0 \| (x_0 * r^2c_0) = [1, rx_1] \| [x_0r^2, x_0x_1r^3] = [1, x_1r, x_0r^2, x_0x_1r^3]$
//     which is equivalent to $f(a) = 1 + x_1ra + x_0(ra)^2 + x_0x_1r^2a^3$
//
// This method expects the coefficients in reverse order so transcript[i] =
// x_{l-j}.
// f(Y) = Y^n * \prod (1 + x_{l-j-1} (r_shiftY^{2^j}))
fn polynomial_coefficients_from_transcript<F: Field>(transcript: &[F], r_shift: &F) -> Vec<F> {
    let mut coefficients = vec![F::one()];
    let mut power_2_r = *r_shift;

    for (i, x) in transcript.iter().enumerate() {
        let n = coefficients.len();
        if i > 0 {
            power_2_r = power_2_r.square();
        }
        for j in 0..n {
            let coeff = coefficients[j] * (*x * power_2_r);
            coefficients.push(coeff);
        }
    }

    coefficients
}