Nova: Recursive Zero-Knowledge Arguments

from Folding Schemes

Abhiram Kothapalli† Srinath Setty⋆ Ioanna Tzialla‡

† ⋆ ‡
Carnegie Mellon University Microsoft Research New York University

Abstract. We introduce a new approach to realize incrementally ver-

ifiable computation (IVC), in which the prover recursively proves the
correct execution of incremental computations of the form y = F (ℓ) (x),
where F is a (potentially non-deterministic) computation, x is the input,
y is the output, and ℓ > 0. Unlike prior approaches to realize IVC, our ap-
proach avoids succinct non-interactive arguments of knowledge (SNARKs)
entirely and arguments of knowledge in general. Instead, we introduce
and employ folding schemes, a weaker, simpler, and more efficiently-
realizable primitive, which reduces the task of checking two instances in
some relation to the task of checking a single instance. We construct a
folding scheme for a characterization of NP and show that it implies an
IVC scheme with improved efficiency characteristics: (1) the “recursion
overhead” (i.e., the number of steps that the prover proves in addition
to proving the execution of F ) is a constant and it is dominated by two
group scalar multiplications expressed as a circuit (this is the smallest
recursion overhead in the literature), and (2) the prover’s work at each
step is dominated by two multiexponentiations of size O(|F |), providing
the fastest prover in the literature. The size of a proof is O(|F |) group
elements, but we show that using a variant of an existing zkSNARK,
the prover can prove the knowledge of a valid proof succinctly and in
zero-knowledge with O(log |F |) group elements. Finally, our approach
neither requires a trusted setup nor FFTs, so it can be instantiated
efficiently with any cycles of elliptic curves where DLOG is hard.

1 Introduction

We revisit the problem of realizing incrementally-verifiable computation (IVC) [47]:

a cryptographic primitive that enables producing proofs of correct execution of
“long running” computations such that a verifier can efficiently verify the correct
execution of any prefix of the computation. IVC enables a wide variety of appli-
cations including verifiable delay functions [10, 50], succinct blockchains [14, 34],
and incrementally-verifiable versions of verifiable state machines [36, 42].
A well-known approach to construct IVC is to use succinct non-interactive
arguments of knowledge (SNARKs) for NP [26, 27, 33, 39]: at each incremental
step i, the prover produces a SNARK proving that it has applied F correctly to
the output of step i − 1 and that the SNARK verifier represented as a circuit
has accepted the SNARK from step i − 1 [7, 9]. However, it is well-known that
this approach is impractical [7, 21]. Alternatively, one can use SNARKs without
trusted setup [19, 23, 41, 44] but their verifiers are more expensive than those
of SNARKs with trusted setup, both asymptotically and concretely. Recent
works [11, 13, 17, 18] aim to address the inefficiency of SNARK-based IVC, with
an innovative approach: at each step, the verifier circuit “defers” expensive steps
in verifying a SNARK for NP instances (e.g., verifying polynomial evaluation
proofs) by accumulating those steps into a single instance that is later checked
efficiently. However, these works still require the prover to produce a SNARK at
each step and the verifier circuit to partially verify that SNARK.
We introduce a new approach that avoids SNARKs (and more generally
arguments of knowledge) entirely and relies purely on deferral to realize IVC. In
a nutshell, instead of accumulating expensive steps of verifying a SNARK for
NP instances, the verifier circuit in our approach accumulates the NP instances
themselves. We formalize this technique as a new and minimal primitive, which we
refer to as a folding scheme. A folding scheme is weaker, simpler, and far more ef-
ficient compared to arguments of knowledge including SNARKs. Indeed, realizing
IVC via folding schemes results in improved efficiency over prior work (Figure 2):
(1) the verifier circuit is constant-sized and its size is dominated by two group
scalar multiplications; this is the smallest verifier circuit in the literature (in the
context of recursive proof composition); and (2) the prover’s work at each step
is dominated by two multiexponentiations of size O(|F |), providing the fastest
prover in the literature, both asymptotically and concretely. Section 1.4 provides
a detailed comparison between our approach and prior work.

1.1 Folding Schemes

A folding scheme is defined with respect to an NP relation, and it is a protocol

between an untrusted prover and a verifier. Both entities hold two N -sized NP
instances, and the prover in addition holds purported witnesses for both instances.
The protocol enables the prover and the verifier to output a single N -sized NP
instance, which we refer to as a folded instance. Furthermore, the prover privately
outputs a purported witness to the folded instance using purported witnesses for
the original instances. Informally, a folding scheme guarantees that the folded
instance is satisfiable only if the original instances are satisfiable. A folding
scheme is said to be non-trivial if the verifier’s costs and the communication are
lower in the case where the verifier participates in the folding scheme and then
verifies a purported NP witness for the folded instance than the case where the
verifier verifies purported NP witnesses for each of the original instances.
Several existing techniques exhibit the two-to-one reduction pattern of folding
schemes. Examples include the sumcheck protocol [38] and the split-and-fold
techniques in inner product arguments [12]. Appendix A provides further details.

Remark 1 (Folding Schemes vs. SNARKs). SNARKs for NP [8, 26, 27, 33, 39]
trivially imply a folding scheme for NP: given two NP instances u1 and u2 and
the corresponding witnesses, the prover proves u1 by producing a SNARK. The
verifier checks that SNARK and then sets u2 to be the folded instance. However,

2
we construct a folding scheme for NP without relying on SNARKs (or more
generally arguments of knowledge). Specifically, our folding scheme is weaker
than any argument of knowledge (succinct or otherwise) because it merely reduces
the satisfiability of two NP instances to the satisfiability of a single NP instance.1

To design a folding scheme for NP, we start with a popular NP-complete

language that generalizes arithmetic circuit satisfiability: R1CS (Definition 10).
As we illustrate later, it is difficult to devise a folding scheme for R1CS. To
address this, we introduce a variant of R1CS, called relaxed R1CS, which, like
R1CS, not only characterizes NP, but, unlike R1CS, can support a folding scheme.
The following theorem captures the cryptographic and efficiency characteristics
of our folding scheme for relaxed R1CS.
Theorem 1. There exists a constant-round, public-coin, zero-knowledge folding
scheme for relaxed R1CS where for N -sized relaxed R1CS instances over a finite
field F with the same “structure” (i.e., R1CS coefficient matrices), the prover’s
work is Oλ (N ), and the verifier’s work and the communication are both Oλ (1),
assuming the existence of any additively-homomorphic commitment scheme that
provides Oλ (1)-sized commitments to N -sized vectors over F (e.g., Pedersen’s
commitments), where λ is the security parameter.
Because our folding scheme is public coin, it can be made non-interactive in the
random oracle model using the Fiat-Shamir transform [25], and be instantiated
(heuristically) in the standard model using a concrete hash function. We rely on
such a non-interactive folding scheme to construct IVC.

1.2 IVC from Non-Interactive Folding Schemes

We show how to realize IVC using a non-interactive version of our folding scheme
for relaxed R1CS. We refer to our construction as Nova.
Recall that an IVC is an argument of knowledge [33, 39]2 for incremental
computations of the form y = F (ℓ) (x), where F is a (possibly non-deterministic)
computation, ℓ > 0, x is a public input, and y is the public output. At each
incremental step, the IVC prover produces a proof that the step was computed
correctly and it has verified a proof for the prior step. In other words, at
each incremental step, the IVC prover produces a proof of satisfiability for
an augmented circuit that augments the circuit for F with a “verifier circuit”
that verifies the proof of the prior step. Recursively, the final proof proves the
correctness of the entire incremental computation. A key aspect of IVC is that
neither the IVC verifier’s work nor the IVC proof size depends on the number of
1
This work realizes IVC using our folding scheme. As IVC implies SNARKs (e.g.,
see [7]), one might wonder whether folding schemes are in general weaker than
SNARKs. However, existing constructions of IVC (including our own) rely on addi-
tional assumptions (§4.2), which the resulting IVC-based SNARK inherits.
2
An argument of knowledge for circuit satisfiability enables an untrusted polynomial-
time prover to prove to a verifier the knowledge of a witness w such that C(w, x) = y,
where C is a circuit, x is some public input, and y is some public output.

3
 steps in the incremental computation. In particular, the IVC verifier only verifies
the proof produced at the last step of the incremental computation.
In Nova, we consider incremental computations, where each step of the in-
cremental computation is expressed with R1CS (all the steps in the incremental
computation share the same R1CS coefficient matrices). At step i of the incre-
mental computation, as in other approaches to IVC, Nova’s prover proves that
the step i was computed correctly. Furthermore, at step i, instead of verifying a
proof for step i − 1 (as in traditional approaches to IVC), Nova’s approach treats
the computation at step i − 1 as an R1CS instance and folds that into a running
relaxed R1CS instance. Specifically, at each step, Nova’s prover proves that it has
performed the step’s computation and has folded its prior step represented as an
R1CS instance into a running relaxed R1CS instance. In other words, the circuit
satisfiability instance that the prover proves at each incremental step computes a
step of the incremental computation and includes a circuit for the computation
of the verifier in the non-interactive folding scheme for relaxed R1CS.
A distinctive aspect of Nova’s approach to IVC is that it achieves the smallest
“verifier circuit” in the literature. Since the verifier’s costs in the non-interactive
version of the folding scheme for relaxed R1CS is Oλ (1), the size of the com-
putation that Nova’s prover proves at each incremental step is ≈|F |, assuming
N -sized vectors are committed with an Oλ (1)-sized commitments (e.g., Pedersen’s
commitments). In particular, the verifier circuit in Nova is constant-sized and
its size is dominated by two group scalar multiplications. Furthermore, Nova’s
prover’s work at each step is dominated by two multiexponentiations of size ≈|F |.
Note that Nova’s prover does not perform any FFTs, so it can be instantiated
efficiently using any cycles of elliptic curves where DLOG is hard.
With the description thus far, the size of an IVC proof (which is a purported
witness for the running relaxed R1CS instance) is Oλ (|F |). Instead of sending
such a proof to a verifier, at any point in the incremental computation, Nova’s
prover can prove the knowledge of a satisfying witness to the running relaxed
R1CS instance in zero-knowledge with an Oλ (log |F |)-sized succinct proof using
a zkSNARK that we design by adapting Spartan [41]. The following theorem
summarizes our key result.

Theorem 2. For any incremental function where each step of the incremental
function applies a (non-deterministic) function F , there exists an IVC scheme
with the following efficiency characteristics, assuming N -sized vectors are com-
mitted with an Oλ (1)-sized commitments.

– IVC proof sizes are O(|F |) and the verifier’s work to verify them is Oλ (|F |).
The prover’s work at each incremental step is ≈|F |. Specifically, the prover’s
work at each step is dominated by two multiexponentiations of size ≈|F |.
– Succinct zero-knowledge proofs of valid IVC proofs are size Oλ (log |F |), and
the verifier’s work to verify them is either Oλ (log |F |) or Oλ (|F |) depending
on the commitment scheme for vectors. The prover’s work to produce this
succinct zero-knowledge proof is Oλ (|F |).

4
1.3 Implementation and Performance Evaluation
We implement Nova as a library in about 6,000 lines of Rust [3]. The library
is generic over a cycle of elliptic curves and a hash function (used internally as
the random oracle). The library provides candidate implementations with the
Pasta cycle of elliptic curves [4] and Poseidon [2, 29]. For the former, Nova relies
on pasta-msm [5], a high-performance library for computing multiexpoentiations
over the Pasta cycle of curves. Finally, the library accepts F (i.e., a step of the
incremental computation) as a bellperson gadget [1].

Recursion Overheads. We measure the size of Nova’s verifier circuit, as it

determines the recursion overhead : the number of additional constraints that the
prover must prove at each incremental step besides proving an invocation of F .
We find that Nova’s verifier circuit is ≈10,000 R1CS constraints. This is
the smallest verifier circuit in the literature and hence Nova incurs the lowest
recursion overhead. Specifically, Nova’s recursion overhead is > 10× lower than in
SNARK-based IVC [7] with state-of-the-art per-circuit trusted setup SNARK [30],
and over 100× smaller than with a SNARK without trusted setup [23]. Compared
to recent works, Nova’s recursion overhead is over 13× lower than Halo’s [13],
and over 4× lower than the scheme of Bünz et al. [17].

Performance of Nova. We experiment with Nova on an Azure Standard

F32s v2 VM (16 physical CPUs, 2.70 GHz Intel(R) Xeon(R) Platinum 8168, and
64 GB memory). In our experiments, we vary the number of constraints in F .
Our performance metrics are: the prover time, the verifier time, and proof sizes.
We measure these for Nova’s IVC scheme as well as its Spartan-based zkSNARK
to compress IVC proofs. Figure 1 depicts our results, and we find the following.
– The prover’s per-step cost to produce an IVC proof and compress it scale
sub-linearly with the size of F (since the cost is dominated by two multiex-
ponentiations, which scale sub-linearly due to the Pippenger algorithm and
parallelize better at larger sizes). When |F | ≈ 220 constraints, the prover’s
per-step cost to produce an IVC proof is ≈1µs/constraint. For the same F ,
the cost to produce a compressed IVC proof is ≈24µs/constraint.3
– Compressed IVC proofs are ≈ 8–9 KB and are significantly shorter than IVC
proofs (e.g., they are ≈7,400× shorter when |F | ≈ 220 constraints).
– Verifying a compressed proof is only ≈2× higher costs than verifying a
significantly longer IVC proof.

1.4 A More Detailed Comparison with Prior Work

Figure 2 compares Nova with prior approaches. Nova’s approach can be viewed
as taking Halo’s approach to the extreme. Specifically:
3
If the prover produces a compressed IVC proof every ≈24 steps, the prover incurs at
most 2× overhead to compress IVC proofs. Similarly, if the prover compresses its
IVC proof every ≈240 steps, the overhead drops to ≈20%.

5
 102 1054 104

verifier time (ms)

proof size (KB)
prover time (s)
1 103
10 102 103
0
10 101 102
−1
100
10 10 101
214 215 216 217 218 219 220 214 215 216 217 218 219 220 214 215 216 217 218 219 220
# constraints # constraints # constraints

IVC Scheme (per step) A zkSNARK of a Valid IVC Proof

Fig. 1: Performance of Nova as a function of |F |. See the text for details.

– At each incremental step, Halo’s verifier circuit verifies a “partial” SNARK.

This still requires Halo’s prover to perform |F |-sized FFTs and O(|F |) expo-
nentiations (i.e., not an |F |-sized multiexponentiation). Whereas, in Nova,
the verifier circuit folds an entire NP instance representing computation at
the prior step into a running relaxed R1CS instance. This only requires Nova’s
prover to commit to a satisfying assignment of an ≈|F |-sized circuit (which
computes F and performs the verifier’s computation in a folding scheme for
relaxed R1CS), so at each step, Nova’s prover only computes an O(|F |)-sized
multiexponentiation and does not compute any FFTs. So, Nova’s prover
incurs lower costs than Halo’s prover, both asymptotically and concretely.
– The verifier circuit in Halo is of size Oλ (log |F |) whereas in Nova, it is Oλ (1).
Concretely, the dominant operations in Halo’s circuit is O(log |F |) group
scalar multiplications, whereas in Nova, it is two group scalar multiplications.
– Halo and Nova have the same proof sizes Oλ (log |F |) and verifier time Oλ (|F |).
Bünz et al. [18] apply Halo’s approach to other polynomial commitment
schemes. Halo Infinite [11] generalizes the approach in Halo [13] to any homomor-
phic polynomial commitment scheme; they also obtain PCD (and hence IVC)
even when polynomial commitment schemes do not satisfy succinctness.
Bünz et al. [17] propose a variant of the approach in Halo, where they realize
PCD (and hence IVC) without relying on succinct arguments. Specifically, they
first devise a non-interactive argument of knowledge (NARK) for R1CS with
Oλ (N )-sized proofs and Oλ (N ) verification times for N -sized R1CS instances.
Then, they show that most of the NARK’s verifier’s computation can be deferred
by performing Oλ (1) work in the verifier circuit. For zero-knowledge, Nova relies
on zero-knowledge arguments with succinct proofs, whereas their approach does
not rely on succinct arguments. However, Nova’s approach has several conceptual
and efficiency advantages over the work of Bünz et al [17]:
– Nova introduces a new primitive called a folding scheme, which is conceptually
simpler and is easier to realize than prior notions such as (split) accumulation
schemes used in prior work [17, 18]. Furthermore, a folding scheme for NP
directly leads to IVC and is again easier to analyze than with prior notions.
– At each step, their prover performs an O(|F |)-sized FFT (which costs
O(|F | log |F |) operations over F ). Whereas, Nova does not perform any FFTs.
– Their prover’s work for multiexponentitions at each step and the size of their
verifier circuit are both higher than in Nova by ≈4×.

6
 “Verifier circuit” Prover Proof size Verifier assumptions
(dominant ops) (each step)
BCTV14 [7] with [30]† 3P O(C) FFT Oλ (1) Oλ (1) q-type
O(C) MSM
√ √ √
Spartan [41]-based IVC O( C) G O(C) MSM Oλ ( C) Oλ ( C) DLOG, RO
Fractal [23] Oλ (log2 C) F O(C) FFT Oλ (log2 C) Oλ (log2 C) RO
O(log2 C) H O(C) MHT
Halo [13] O(log C) G O(C) FFT Oλ (log C) Oλ (C) DLOG, RO
O(C) EXP
BCLMS [17]⋆ 8G O(C) FFT Oλ (C) Oλ (C) DLOG, RO
O(C) MSM
Nova (this work) 2G O(C) MSM Oλ (log C) Oλ (C) DLOG, RO
Nova (this work) 2 GT O(C) MSM Oλ (log C) Oλ (log C) SXDH, RO
†
Requires per-circuit trusted setup and is undesirable in practice
O(C) FFT: FFT over an O(C)-sized vector costing O(C log C) operations over F
O(C) MHT: Merkle tree over an O(C)-sized vector costing O(C) hash computations
O(C) EXP: O(C) exponentiations in a cryptographic group
O(C) MSM: O(C)-sized multi-exponentiation in a cryptographic group

Fig. 2: Asymptotic costs of Nova and its baselines to produce and verify a proof for an
incremental computation where each incremental step applies a function F . C denotes
the size of the computation at each incremental step, i.e., |F | + |CV |, where CV is the
“verifier circuit” in IVC. The “verifier circuit” column depicts the number of dominant
operations in CV , where P denotes a pairing in a pairing-friendly group, F denotes
the number of finite field operations, H denotes a hash computation, and G denotes a
scalar multiplication in a cryptographic group. The prover column depicts the cost to
the prover for each step of the incremental computation, and proof sizes and verifier
times refer respectively to the size of the proof of the incremental computation and the
associated verification times. For Nova’s proof sizes and verification times, we depict the
compressed proof sizes (otherwise, they are Oλ (C)) and the time to verify a compressed
proof (otherwise, they are Oλ (C)). Rows with RO require heuristically instantiating
the random oracle with a concrete hash function in the standard model.

– Proof sizes are Oλ (|F |) in their work, whereas in Nova, they are Oλ (log |F |).
We believe, in theory, they can also compress their proofs, using a succinct
argument, but unlike Nova, they do not specify how to do so in a concretely
efficient manner. Furthermore, using succinct arguments is inconsistent with
their goal of not employing them.
Concurrent work. In an update concurrent with this work, Bünz et al. [17] provide
an improved construction of their NARK for R1CS, which leads to an IVC that,
like Nova, avoids FFTs. Furthermore, they improve the size of the verifier circuit
by ≈2×, which is still larger than Nova’s verifier circuit by ≈2×. The per-step
computation of the prover remains 4× higher than Nova.

1.5 An Overview of the Rest of the Paper

Section 2 provides the necessary background. Section 3 formally defines folding
schemes and their properties. In Section 4, we introduce a variant of R1CS

7
called relaxed R1CS for which we provide a folding scheme satisfying Theorem 1.
Then, in Section 5, we use a non-interactive version of the folding scheme (§4.2)
to construct an IVC scheme and a scheme to compress IVC proofs satisfying
Theorem 2 by assuming the existence of a zkSNARK for relaxed R1CS with
logarithmic-sized proofs. Finally, in Section 6, we construct such a zkSNARK.

2 Preliminaries

Let F denote a finite field with |F | = 2Θ(λ) , where λ is the security parameter.
Let ∼
= denote computational indistinguishability with respect to a PPT adversary.
We globally assume that generator algorithms that produce public parameters
are additionally provided appropriate size bounds.

2.1 A Commitment Scheme for Vectors over F

We require a commitment scheme for vectors over F that is additively homo-

morphic and succinct. We formally define these two properties and others noted
below in Appendix F. Below, we define the syntax for commitment schemes.

Definition 1 (A Commitment Scheme for Vectors). A commitment scheme

for F m is a tuple of three protocols with the following interface.

– Gen(1λ , m) → pp: takes length parameter m; produces public parameters pp.

– Com(pp, v, r) → C: takes vector v ∈ F m and r ∈ F ; produces commitment C.
– Open(pp, C, v, r) → {0, 1}: verifies the opening of commitment C to v ∈ F m .

A commitment scheme satisfies hiding (the commitment reveals no information),

binding (a PPT adversary cannot open a commitment to two different values),
and succinctness (the commitment size is logarithmic in the opening size).

2.2 Non-Interactive Arguments of Knowledge

Definition 2 (Non-Interactive Argument of Knowledge). Consider a

relation R over public parameters, structure, instance, and witness tuples. A non-
interactive argument of knowledge for R consists of PPT algorithms (G, P, V)
and deterministic K, denoting the generator, the prover, the verifier and the
encoder respectively with the following interface.

– G(1λ ) → pp: On input security parameter λ, samples public parameters pp.

– K(pp, s) → (pk, vk): On input structure s, representing common structure
among instances, outputs the prover key pk and verifier key vk.
– P(pk, u, w) → π: On input instance u and witness w, outputs a proof π
proving that (pp, s, u, w) ∈ R.
– V(vk, u, π) → {0, 1}: Checks instance u given proof π.

8
An argument of knowledge satisfies completeness if for any PPT adversary A
pp ← G(1λ ),
 

 (s, (u, w)) ← A(pp), 

Pr  V(vk, u, π) = 1 (pp, s, u, w) ∈ R,
  = 1.

 (pk, vk) ← K(pp, s), 
π ← P(pk, u, w)
An argument of knowledge satisfies knowledge soundness if for all PPT adversaries
A there exists a PPT extractor E such that for all randomness ρ
pp ← G(1λ ),
 
 V(vk, u, π) = 1, (s, u, π) ← A(pp; ρ), 
Pr 
 (pp, s, u, w) ̸∈ R (pk, vk) ← K(pp, s),  = negl(λ).


w ← E(pp, ρ)
Definition 3 (Zero-Knowledge). An argument of knowledge (G, K, P, V) for
relation R satisfies zero-knowledge if there exists PPT simulator S such that for
all PPT adversaries A
   

 pp ← G(1λ ), 
   (pp, τ ) ← S(1λ ), 

(s, (u, w)) ← A(pp),  (s, (u, w)) ← A(pp), 

   
    
(pp, s, u, π) (pp, s, u, w) ∈ R, ∼
= (pp, s, u, π) (pp, s, u, w) ∈ R,
(pk, vk) ← K(pp, s),  (pk, vk) ← K(pp, s), 

    

 
   

π ← P(pk, u, w) π ← S(pp, u, τ )
   

Definition 4 (Succinctness). A non-interactive argument system is succinct

if the size of the proof π is polylogarithmic in the size of the witness w.

2.3 Incrementally Verifiable Computation

Incrementally verifiable computation (IVC) [47] enables verifiable computation
for repeated function application. Intuitively, for a function F , with initial input
z0 , an IVC scheme allows a prover to produce a proof Πi for the statement
zi = F (i) (z0 ) (i.e., i applications of F on input z0 ) given a proof Πi−1 for
the statement zi−1 = F (i−1) (z0 ). Formally, IVC schemes additionally permit
F to take auxiliary input ω. We recall the definition of IVC using notational
conventions of modern argument systems.
Definition 5 (IVC). An incrementally verifiable computation (IVC) scheme is
defined by PPT algorithms (G, P, V) and deterministic K denoting the generator,
the prover, the verifier, and the encoder respectively. An IVC scheme (G, K, P, V)
satisfies perfect completeness if for any PPT adversary A
pp ← G(1λ ),
 

 F, (i, z0 , zi , zi−1 , ωi−1 , Πi−1 ) ← A(pp), 

 (pk, vk) ← K(pp, F ), 
Pr  V(vk, i, z ,
0 iz , Πi ) = 1 =1

 zi = F (zi−1 , ωi−1 ), 

 V(vk, i − 1, z0 , zi−1 , Πi−1 ) = 1, 
Πi ← P(pk, i, z0 , zi ; zi−1 , ωi−1 , Πi−1 )

9
where F is a polynomial time computable function. Likewise, an IVC scheme
satisfies knowledge soundness if for any constant n ∈ N, and for all expected
polynomial time adversaries P ∗ , there exists an expected polynomial-time extractor
E such that
pp ← G(1λ ),
 
zn = z where
Pr  zi+1 ← F (zi , ωi ) (F, (z0 , z), Π) ← P ∗ (pp, r),  ≈
r
∀i ∈ {0, . . . , n − 1} (ω0 , . . . , ωn−1 ) ← E(pp, r)
pp ← G(1λ ),
 

Pr  V(vk, (n, z0 , z), Π) = 1 (F, (z0 , z), Π) ← P ∗ (pp, r), 

r
(pk, vk) ← K(pp, F )

where r denotes an arbitrarily long random tape.

An IVC scheme satisfies succinctness if the size of the IVC proof Π does not
grow with the number of applications n.

We note that in the definition above, the number of steps n is treated as a

fixed environment variable that characterizes the extractor. This model is required
for all known general recursive techniques as they rely on recursive extractors
that blowup polynomially for each additional recursive step [11, 14, 17, 18, 23].
Bitansky et al. [9] avoid such a restriction by making non-blackbox assumptions
about the extractors runtime with respect to that of the malicious prover. In any
case, there are no known attacks on arbitrary depth recursion.

3 Folding Schemes

This section formally defines folding schemes. Intuitively, a folding scheme for a
relation R is a protocol that reduces the task of checking two instances in R to
the task of checking a single instance in R.

Definition 6 (Folding Scheme). Consider a relation R over public parameters,

structure, instance, and witness tuples. A folding scheme for R consists of a
PPT generator algorithm G, a deterministic encoder algorithm K, and a pair of
PPT algorithms P and V denoting the prover and verifier respectively, with the
following interface:

– G(1λ ) → pp: On input security parameter λ, samples public parameters pp.

– K(pp, s) → (pk, vk): On input pp, and a common structure s between instances
to be folded, outputs a prover key pk and a verifier key vk.
– P(pk, (u1 , w1 ), (u2 , w2 )) → (u, w): On input instance-witness tuples (u1 , w1 )
and (u2 , w2 ) outputs a new instance-witness tuple (u, w) of the same size.
– V(vk, u1 , u2 ) → u: On input instances u1 and u2 , outputs a new instance u.

Let

(u, w) ← ⟨P(pk, w1 , w2 ), V(vk)⟩(u1 , u2 )

10
denote the the verifier’s output instance u and the prover’s output witness w from
the interaction of P and V on witnesses (w1 , w2 ), prover key pk, verifier key vk
and instances (u1 , u2 ). Likewise, let

tr = ⟨P(pk, w1 , w2 ), V(vk)⟩(u1 , u2 )

denote the corresponding interaction transcript. A folding scheme satisfies perfect

completeness if for all PPT adversaries A

pp ← G(1λ ),
 

 (s, (u1 , w1 ), (u2 , w2 )) ← A(pp), 

Pr  (pp, s, u, w) ∈ R (pp, s, u1 , w1 ), (pp, s, u2 , w2 ) ∈ R,
  = 1.

 (pk, vk) ← K(pp, s), 
(u, w) ← ⟨P(pk, w1 , w2 ), V(vk)⟩(u1 , u2 )

A folding scheme satisfies knowledge soundness if for any expected polynomial-time

adversary P ∗ there is an expected polynomial-time extractor E such that

pp ← G(1λ ),
 
(pp, s, u1 , w1 ) ∈ R,
Pr  (s, (u1 , u2 )) ← P ∗ (pp, ρ),  ≥
(pp, s, u2 , w2 ) ∈ R
(w1 , w2 ) ← E(pp, ρ)
pp ← G(1λ ),
 
 (s, (u1 , u2 )) ← P ∗ (pp, ρ), 
 (pp, s, u, w) ∈ R
Pr   − negl(λ)
(pk, vk) ← K(pp, s), 
(u, w) ← ⟨P ∗ (pk, ρ), V(vk)⟩(u1 , u2 )

where ρ denotes arbitrary input randomness for P ∗ . We call a transcript an

accepting transcript if P outputs a satisfying folded witness w for the folded
instance u. We consider a folding scheme non-trivial if the communication costs
and V’s computation are lower in the case where V participates in the folding
scheme and then checks a witness sent by P for the folded instance than the case
where V checks witnesses sent by P for each of the original instances.

Definition 7 (Non-Interactive). A folding scheme (G, K, P, V) is non-interactive

if the interaction between P and V consists of a single message from P to V.
This single message is denoted as an output of P, and an input to V.

Definition 8 (Zero-Knowledge). A folding scheme (G, K, P, V) satisfies zero-

knowledge for relation R if there exists a PPT simulator S such that for all PPT

11
adversaries A, and V ∗ , and input randomness ρ
 

 pp ← G(1λ ), 

 (s, (u1 , w1 ), (u2 , w2 )) ← A(pp),

 


tr (pk, vk) ← K(pp, s), ∼
=
 (pp, s, u1 , w1 ), (pp, s, u2 , w2 ) ∈ R,

 


tr = ⟨P(pk, w1 , w2 ), V ∗ (vk, ρ)⟩(u1 , u2 )

 

 

 pp ← G(1λ ), 

 (s, (u1 , w1 ), (u2 , w2 )) ← A(pp), 

 

tr (pp, s, u1 , w1 ), (pp, s, u2 , w2 ) ∈ R,
 (pk, vk) ← K(pp, s),

 


 ∗ 
V (vk,ρ)
tr ← S (pk, u1 , u2 )
 

Definition 9 (Public Coin). A folding scheme (G, K, P, V) is called public coin

if all the messages sent from V to P are sampled from a uniform distribution.

Typically, knowledge soundness is difficult to prove directly. To assist these

proofs, prior works employ the forking lemma [12], which abstracts away much
of the probabilistic reasoning. The original forking lemma shows that to prove
knowledge soundness it is sufficient to construct a PPT extractor that takes
as input a “tree” of accepting transcripts and outputs a satisfying witness.
However, in our setting, this extractor must additionally take as input the prover’s
output (i.e., the folded instance and witness) for each of these transcripts, which
contains information needed to reconstruct the original witness. So, we introduce
a small variant of the forking lemma that captures this modification.

Lemma 1 (Forking Lemma for Folding Schemes). Consider a (2µ + 1)-

move folding scheme Π = (G, K, P, V). Π satisfies knowledge soundness if there
exists a PPT X such that for all input instance pairs (u1 , u2 ), outputs satisfying
witnesses (w1 , w2 ) with probability 1 − negl(λ), given public parameters pp, a
structure s, and an (n1 , . . . , nµ )-tree of accepting transcripts and the corresponding
folded instance-witness pairs (u, w). This tree comprises of n1 transcripts (and the
corresponding instance-witness pairs) with fresh randomness in V’s first message;
and for each such transcript, n2 transcripts (and the corresponding instance-
witness
Qµ pairs) with fresh randomness in V’s second message; etc., for a total of
i=1 i leaves bounded by poly(λ).
n

Proof Intuition. A proof for our variant of the forking lemma is similar to that
of Bootle et al. [12]. We present a formal proof in Appendix E.

4 A Folding Scheme for NP

In this section, we describe a public-coin, zero-knowledge interactive folding
scheme for NP. We additionally discuss how to make it non-interactive. We
leverage the non-interactivity property to realize IVC in the next section, and
the zero-knowledge property to achieve zero-knowledge IVC proof compression.

12
4.1 A Public-Coin, Zero-Knowledge Folding Scheme

To design a folding scheme for NP, we need an NP-complete language. While

theoretically any NP-complete language is a viable candidate, we focus on R1CS,4
a popular algebraic representation that generalizes arithmetic circuit satisfiability.

Definition 10 (R1CS). Consider a finite field F . Let the public parameters

consist of size bounds m, n, ℓ ∈ N where m > ℓ. The R1CS structure consists of
sparse matrices A, B, C ∈ F m×m with at most n = Ω(m) non-zero entries in each
matrix. An instance x ∈ F ℓ consists of public inputs and outputs and is satisfied
by a witness W ∈ F m−ℓ−1 if (A · Z) ◦ (B · Z) = C · Z, where Z = (W, x, 1).

As we show in the next section, to realize IVC, we only need a folding

scheme that can fold two R1CS instances with the same R1CS matrices (A, B, C).
Specifically, given R1CS matrices (A, B, C), and two corresponding instance-
witness pairs (x1 , W1 ) and (x2 , W2 ), we would like to devise a scheme that reduces
the task of checking both instances into the task of checking a single new instance-
witness pair (x, W ) against the same R1CS matrices (A, B, C). Unfortunately, as
we illustrate now, it is difficult to devise a folding scheme for R1CS such that it
satisfies completeness, let alone knowledge soundness.

First Attempt. As R1CS is an algebraic system, the most direct approach

would be to take a random linear combination. Ignoring efficiency concerns,
suppose that the prover sends witnesses W1 and W2 in the first step. The verifier
responds with a random r ∈ F ; the prover and the verifier both compute

x ← x1 + r · x2
W ← W1 + r · W2 ,

and set the new instance-witness pair to be (x, W ). However, for non-trivial
Z1 = (W1 , x1 , 1) and Z2 = (W2 , x2 , 1), and Z = (W, x, 1), we roughly have that

AZ ◦ BZ = A(Z1 + r · Z2 ) ◦ B(Z1 + r · Z2 )
= AZ1 ◦ BZ1 + r · (AZ1 ◦ BZ2 + AZ2 ◦ BZ1 ) + r2 · (AZ2 ◦ BZ2 )
̸= CZ.

The failed attempt exposes three issues. First, we must account for an addi-
tional cross-term, r · (AZ1 ◦ BZ2 + AZ2 ◦ BZ1 ). Second, the terms excluding the
cross-term combine to produce a term that does not equal CZ:

AZ1 ◦ BZ1 + r2 · (AZ2 ◦ BZ2 ) = CZ1 + r2 · CZ2 ̸= CZ1 + r · CZ2 = CZ.

Third, we do not even have that Z = Z1 +r·Z2 because Z1 +r·Z2 = (W, x, 1+r·1).
4
R1CS is implicit in the QAPs formalism of GGPR [26], but it was made explicit in
subsequent work [43]; they refer to it as a “constraint system in quadratic form”.

13
Second Attempt. To handle the first issue, we introduce a “slack” (or error)
vector E ∈ F m which absorbs the cross terms generated by folding. To handle
the second and third issues, we introduce a scalar u, which absorbs an extra
factor of r in CZ1 + r2 · CZ2 and in Z = (W, x, 1 + r · 1). We refer to a variant of
R1CS with these additional terms as relaxed R1CS.
Definition 11 (Relaxed R1CS). Consider a finite field F . Let the public
parameters consist of size bounds m, n, ℓ ∈ N where m > ℓ. The relaxed R1CS
structure consists of sparse matrices A, B, C ∈ F m×m with at most n = Ω(m)
non-zero entries in each matrix. A relaxed R1CS instance consists of an error
vector E ∈ F m , a scalar u ∈ F , and public inputs and outputs x ∈ F ℓ . An instance
(E, u, x) is satisfied by a witness W ∈ F m−ℓ−1 if (A · Z) ◦ (B · Z) = u · (C · Z) + E,
where Z = (W, x, u).
Note that any R1CS instance can be expressed as a relaxed R1CS instance by
augmenting it with u = 1 and E = 0, so relaxed R1CS retains NP-completeness.
Building on the first attempt, the prover and verifier can now use E to
accumulate the cross-terms. In particular, for Zi = (Wi , xi , ui ), the prover and
verifier additionally compute

u ← u1 + r · u2
E ← E1 + r · (AZ1 ◦ BZ2 + AZ2 ◦ BZ1 − u1 CZ2 − u2 CZ1 ) + r2 · E2 ,

and set the new instance-witness pair to be ((E, u, x), W ). Conveniently, updating
u in this manner also keeps track of how the constant term in Z should be updated,
which motivates our choice to use u in Z = (W, x, u) rather than introducing a
new variable. Now, for Z = (W, x, u), and for random r ∈ F ,

AZ ◦ BZ = AZ1 ◦ BZ1 + r · (AZ1 ◦ BZ2 + AZ2 ◦ BZ1 ) + r2 · (AZ2 ◦ BZ2 )

= (u1 CZ1 + E1 ) + r · (AZ1 ◦ BZ2 + AZ2 ◦ BZ1 ) + r2 · (u2 CZ2 + E2 )
= (u1 + r · u2 ) · C(Z1 + rZ2 ) + E
= uCZ + E.

This implies that, for R1CS matrices (A, B, C), the folded witness W is a satisfying
witness for the folded instance (E, u, x) as promised. A few issues remain: in the
above scheme, the prover sends witnesses (W1 , W2 ) for the verifier to compute E.
As a result, the folding scheme is not non-trivial; it is also not zero-knowledge.

Final Protocol. To circumvent these issues, we use succinct and hiding additively
homomorphic commitments to W and E in the instance, and treat both W and E
as the witness. We refer to this variant of relaxed R1CS as committed relaxed R1CS.
Below, we describe a folding scheme for committed relaxed R1CS, where the
prover sends a single commitment to aid the verifier in computing commitments
to the folded witness (W, E).
Definition 12 (Committed Relaxed R1CS). Consider a finite field F and a
commitment scheme Com over F . Let the public parameters consist of size bounds

14
m, n, ℓ ∈ N where m > ℓ, and commitment parameters ppW and ppE for vectors of
size m and m−ℓ−1 respectively. The committed relaxed R1CS structure consists of
sparse matrices A, B, C ∈ F m×m with at most n = Ω(m) non-zero entries in each
matrix. A committed relaxed R1CS instance is a tuple (E, u, W , x), where E and
W are commitments, u ∈ F , and x ∈ F ℓ are public inputs and outputs. An instance
(E, u, W , x) is satisfied by a witness (E, rE , W, rW ) ∈ (F m , F , F m−ℓ−1 , F ) if E =
Com(ppE , E, rE ), W = Com(ppW , W, rW ), and (A · Z) ◦ (B · Z) = u · (C · Z) + E,
where Z = (W, x, u).

Construction 1 (A Folding Scheme for Committed Relaxed R1CS).

Consider a finite field F and a succinct, hiding, and homomorphic commitment
scheme Com over F . We define the generator and the encoder as follows.

– G(1λ ) → pp: output size bounds m, n, ℓ ∈ N, and commitment parameters

ppW and ppE for vectors of size m and m − ℓ − 1 respectively.
– K(pp, (A, B, C)) → (pk, vk): output pk ← (pp, (A, B, C)) and vk ← ⊥.

The verifier V takes two committed relaxed R1CS instances (E 1 , u1 , W 1 , x1 ) and

(E 2 , u2 , W 2 , x2 ). The prover P, in addition to the two instances, takes witnesses to
both instances, (E1 , rE1 , W1 , rW1 ) and (E2 , rE2 , W2 , rW2 ). Let Z1 = (W1 , x1 , u1 )
and Z2 = (W2 , x2 , u2 ). The prover and the verifier proceed as follows.

1. P: Send T := Com(ppE , T, rT ), where rT ←R F and with cross term

T = AZ1 ◦ BZ2 + AZ2 ◦ BZ1 − u1 · CZ2 − u2 · CZ1 .

2. V: Sample and send challenge r ←R F .

3. V, P: Output the folded instance (E, u, W , x) where

E ← E1 + r · T + r2 · E 2
u ← u1 + r · u2
W ← W1 + r · W2
x ← x1 + r · x2

4. P: Output the folded witness (E, rE , W, rW ), where

E ← E1 + r · T + r2 · E2
rE ← rE1 + r · rT + r2 · rE2
W ← W1 + r · W2
rW ← rW1 + r · rW2

Theorem 3 (A Folding Scheme for Committed Relaxed R1CS). Con-

struction 1 is a public-coin folding scheme for committed relaxed R1CS with
perfect completeness, knowledge soundness, and zero-knowledge.

15
Proof Intuition. With textbook algebra, we can show that if witnesses (E1 , rE1 ,
W1 , rW1 ) and (E2 , rE2 , W2 , rW2 ) are satisfying witnesses, then the folded witness
(E, rE , W, rW ) must be a satisfying witness. We prove knowledge soundness via
the forking lemma (Lemma 1) by showing that the extractor can produce the
initial witnesses given three accepting transcripts and the corresponding folded
witnesses. Specifically, the extractor uses all three transcripts to compute Ei and
rEi , and any two transcripts to compute Wi and rWi for i ∈ {1, 2}. The choice
of which two transcripts does not matter due to the binding property of the
commitment scheme. We present a formal proof in Appendix B.

4.2 Achieving Non-Interactivity via the Fiat-Shamir Transform

To design Nova’s IVC scheme, we require our folding scheme for committed
relaxed R1CS to be non-interactive in the standard model. To do so we first
achieve non-interactivity in the random oracle model using the (strong) Fiat-
Shamir transform [25]. Next, we heuristically instantiate the random oracle using
a cryptographic hash function. As a result, we can only heuristically argue the
security of the resulting non-interactive folding scheme. Note that all existing
IVC constructions in the standard model require instantiating the random oracle
with a cryptographic hash function [13, 17, 23, 47].

Construction 2 (A Non-Interactive Folding Scheme). We achieve non-

interactivity in the random oracle model using the strong Fiat-Shamir trans-
form [25]. Let ρ denote a random oracle sampled during parameter genera-
tion and provided to all parties. Let (G, K, P, V) represent our interactive fold-
ing scheme (Construction 1). We construct a non-interactive folding scheme
(G, K, P, V) as follows:

– G(1λ ): output pp ← G(1λ ).

– K(pp, (A, B, C)): vk ← ρ(pp, s) and pk ← (pp, (A, B, C), vk); output (vk, pk).
– P(pk, (u1 , w1 ), (u2 , w2 )): runs P(([Link], pk.(A, B, C)) to retrieve its first mes-
sage T , and sends T to V; computes r ← ρ(vk, u1 , u2 , T ), forwards this to P,
and outputs the resulting output.
– V(vk, u1 , u2 , T ): runs V with T as the message from the prover and with
randomness r ← ρ(vk, u1 , u2 , T ), and outputs the resulting output.

Assumption 1 (RO instantiation). Construction 2 is a non-interactive folding

scheme that satisfies completeness, knowledge soundness, and zero-knowledge in
the standard model when ρ is instantiated with a cryptographic hash function.

5 Nova: An IVC Scheme with Proof Compression

This section describes Nova, an IVC scheme designed from a non-interactive fold-
ing scheme, which when instantiated with any additively-homomorphic commit-
ment scheme with succinct commitments achieves the claimed efficiency (Lemma 4).
In addition, Nova incorporates an efficient zkSNARK to prove the knowledge

16
of valid IVC proofs succinctly and in zero-knowledge, providing a succinct,
zero-knowledge proof of knowledge of a valid IVC proof.
In Nova, at each incremental step, the prover folds a particular step of the
incremental computation (represented as a committed relaxed R1CS instance-
witness pair) into a running committed relaxed R1CS instance-witness pair. At
any step in the incremental computation, a valid “IVC proof”, in a nutshell, is
a satisfying witness of the running committed relaxed R1CS instance (which
an honest prover can compute by folding witnesses associated with each step
of the incremental computation) along with the running committed relaxed
R1CS instance. Furthermore, at any incremental step, Nova’s prover can prove
in zero-knowledge and with a succinct proof—using a variant of an existing
zkSNARK [41] (Section 6)—that it knows a valid IVC proof (i.e., a satisfying
witness) to the running committed relaxed R1CS instance (Construction 4).
Note that Nova is not a zero-knowledge IVC scheme, as that would additionally
require an IVC proof to be zero-knowledge (in Nova’s case, an IVC proof does
not hide witnesses associated with steps of the incremental computation). This
difference is immaterial in the context of a single prover since it can use Nova’s
auxiliary zkSNARK to provide a zero-knowledge proof of knowledge of a valid
IVC proof; we leave it to future work to achieve zero-knowledge IVC.

5.1 Constructing IVC from a Folding Scheme for NP

Recall that an IVC scheme allows a prover to show that zn = F (n) (z0 ) for some
count n, initial input z0 , and output zn . We now show how to construct an IVC
scheme for a non-deterministic, polynomial-time computable function F using our
non-interactive folding scheme for committed relaxed R1CS (Construction 2).5
In our construction, as in a SNARK-based IVC, the prover uses an augmented
function F ′ (Figure 3), which, in addition to invoking F , performs additional
bookkeeping to fold proofs of prior invocations of itself.
We first describe a simplified version of F ′ , to provide intuition. F ′ takes
as non-deterministic advice two committed relaxed R1CS instances ui and Ui .
Suppose that Ui represents the correct execution of invocations 1, . . . , i − 1 of F ′
so long as ui represents the correct execution of invocation i of F ′ . F ′ performs
two tasks. First, it executes a step of the incremental computation: instance ui
contains zi which F ′ uses to output zi+1 = F (zi ). Second, F ′ invokes the verifier
of the non-interactive folding scheme to fold the task of checking ui and Ui into
the task of checking a single instance Ui+1 . The IVC prover then computes a
new instance ui+1 which attests to the correct execution of invocation i + 1 of F ′ ,
thereby attesting that zi+1 = F (zi ) and Ui+1 is the result of folding ui and Ui .
Now, we have that Ui+1 represents the correct execution of invocations 1, . . . , i
of F ′ so long as ui+1 represents the correct execution of invocation i + 1 of F ′ .
The above description glossed over a subtle discrepancy: Because F ′ must
output the running instance Ui+1 for the next invocation to use, it is contained
5
While, in theory, we can use any folding scheme for NP, we specifically invoke our
construction for committed relaxed R1CS for a simpler presentation.

17
 i
z0
zi zi+1

ui F hi+1

?
Ui ui .x = Ui+1 ← hi+1 ← hash
hash(i, z0 , zi , Ui ) NIFS.V(Ui , ui ) (i + 1, z0 , zi+1 , Ui+1 )

Fig. 3: A simplified depiction of F ′ . F ′ represented as a committed relaxed R1CS

instance ui+1 encodes the statement that there exists (i, z0 , zi , ui , Ui , Ui+1 , ri , ri+1 , T )
such that ui .x = hash(vk, i, z0 , zi , Ui , ri ), hi+1 = hash(vk, i + 1, z0 , F (zi ), Ui+1 , ri+1 ),
Ui+1 = NIFS.V (vk, Ui , ui , T ), and that F ′ outputs hi+1 . The diagram omits depicting
vk, ωi , ri , ri+1 , and T .

in ui+1 .x (i.e., the public IO of ui+1 ). But, in the next iteration, F ′ must fold
ui+1 .x into Ui+1 .x, meaning that F ′ is stuck trying to squeeze Ui+1 into Ui+1 .x.
To handle this inconsistency, we modify F ′ to output a collision-resistant hash
of its public IO rather than producing it directly (this ensures that the public
IO of F ′ is a constant number of finite field elements). The next invocation of
F ′ then additionally takes the preimage of this hash as non-deterministic advice.
Note that the hash function takes an additional random input. We assume that
this provides a commitment scheme with hiding commitments.

Producing IVC Proofs. Let (u⊥ , w⊥ ) be the trivially satisfying instance-

witness pair, where E, W, and x are appropriately-sized zero vectors, rE = 0,
rW = 0, and E and W are commitments of E and W respectively.
Now, in iteration i + 1, the IVC prover runs F ′ and computes ui+1 and Ui+1
as well as the corresponding witnesses wi+1 and Wi+1 . Because ui+1 and Ui+1
together attest to the correctness of i+1 invocations of F ′ (which indirectly attests
to i + 1 invocations of F ) the IVC proof Πi+1 is ((Ui+1 , Wi+1 ), (ui+1 , wi+1 )).
Moreover, succinctness is maintained by the properties of the underlying folding
scheme. We formally describe our construction below.
Construction 3 (IVC). Let NIFS = (G, K, P, V) be the non-interactive folding
scheme for committed relaxed R1CS (Construction 2). Consider a polynomial-
time function F that takes non-deterministic input, and a cryptographic hash
function hash. We define our augmented function F ′ as follows (all arguments to
F ′ are taken as non-deterministic advice):

F ′ (vk, Ui , ui , (i, z0 , zi ), ωi , T , ri , ri+1 ) → x:

18
If i = 0, output hash(vk, i + 1, z0 , F (z0 , ω0 ), u⊥ , ri+1 ),
otherwise,
(1) check that ui .x = hash(vk, i, z0 , zi , Ui , ri ), where ui .x is the public IO of ui ,
(2) check that (ui .E, ui .u) = (u⊥ .E, 1),
(3) compute Ui+1 ← NIFS.V(vk, Ui , ui , T ), and
(4) output hash(vk, i + 1, z0 , F (zi , ωi ), Ui+1 , ri+1 ).

Because F ′ can be computed in polynomial time, it can be represented as

a committed relaxed R1CS structure. We assume that there is a deterministic
procedure, which we denote with AUGMENT, that takes as input a function F
and public parameters pp sampled by G of the IVC scheme, and outputs the
committed relaxed R1CS structure corresponding to F ′ , which we denote with
sF ′ .6 Let

(ui+1 , wi+1 ) ← trace(F ′ , (vk, Ui , ui , (i, z0 , zi ), ωi , T , ri , ri+1 ))

denote the satisfying committed relaxed R1CS instance-witness pair (ui+1 , wi+1 )
for the execution of F ′ , as a committed Relaxed R1CS with structure sF ′ , on
non-deterministic advice (vk, Ui , ui , (i, z0 , zi ), ωi , T , ri , ri+1 ). Note that trace is
a randomized algorithm that internally samples randomness to create hiding
commitments inside ui+1 . Additionally, note that trace sets ui+1 .E = u⊥ .E and
that ui+1 .u = 1.
We define the IVC scheme (G, K, P, V) as follows.

G(1λ ) → pp: Output NIFS.G(1λ ).

K(pp, F ) → (pk, vk):

(1) compute sF ′ ← AUGMENT(pp, F );

(2) compute (pkfs , vkfs ) ← NIFS.K(pp, sF ′ );
(3) output (pk, vk) ← ((pp, F, pkfs ), (pp, sF ′ , vkfs )).

P(pk, (i, z0 , zi ), ωi , Πi ) → Πi+1 :

(1) if i = 0, compute (Ui+1 , Wi+1 , T ) ← (u⊥ , w⊥ , u⊥ .E);

(2) otherwise, parse Πi as ((Ui , Wi ), (ui , wi ), ri ) and compute (Ui+1 , Wi+1 , T ) ←
NIFS.P(pk, (Ui , Wi ), (ui , wi ));
(3) sample ri+1 ∈ F randomly;
(4) compute (ui+1 , wi+1 ) ← trace(F ′ , (vk, Ui , ui , (i, z0 , zi ), ωi , T , ri , ri+1 )), and
(5) output Πi+1 ← ((Ui+1 , Wi+1 ), (ui+1 , wi+1 ), ri+1 ).

V(vk, (i, z0 , zi ), Πi ) → {0, 1}:

6
In practice, F ′ is implemented as R1CS in a tool such as bellperson [1]. We require
that the procedure that outputs R1CS matrices for F ′ to be deterministic. This is
the case in our implementation of Nova [3].

19
If i = 0, check that zi = z0 ;
otherwise,
(1) parse Πi as ((Ui , Wi ), (ui , wi ), ri ),
(2) check that ui .x = hash(vk, i, z0 , zi , Ui , ri ),
(3) check that (ui .E, u.u) = (u⊥ .E, 1), and
(4) check that Wi and wi are satisfying witnesses to Ui and ui respectively
using [Link] and [Link] ′ .
Lemma 2 (Completeness). Construction 3 is an IVC scheme that satisfies
completeness.
Proof Intuition. Given a satisfying IVC proof Πi = ((Ui , Wi ), (ui , wi ), ri ) suppose
that P outputs Πi+1 = ((Ui+1 , Wi+1 ), (ui+1 , wi+1 ), ri+1 ). Because Πi is a valid
IVC proof, (ui , wi ) and (Ui , Wi ) are satisfying instance-witness pairs. Because
(Ui+1 , Wi+1 ) is obtained by folding (ui , wi ) and (Ui , Wi ), it must be satisfying
by the folding scheme’s completeness. By construction, (ui+1 , wi+1 ) is satisfying
instance-witness pair that satisfies the IVC verifier’s auxiliary checks including
the ones that involve ri+1 . Thus, Πi+1 is satisfying. Appendix C provides a
formal proof.
Lemma 3 (Knowledge Soundness). Construction 3 is an IVC scheme that
satisfies knowledge soundness.
Proof Intuition. For function F , constant n, pp ← G(1λ ), and (pk, vk) ← K(pp, F ),
consider an adversary P ∗ that outputs (z0 , z, Π) such that V(vk, (n, z0 , z), Π) = 1
with probability ϵ. We construct an extractor E that with input (pp, z0 , z), outputs
(ω0 , . . . , ωn−1 ) such that by computing zi ← F (zi−1 , ωi−1 ) for all i ∈ {1, . . . , n}
we have that zn = z with probability ϵ − negl(λ). We show inductively that E
can construct an extractor Ei that outputs (zi , . . . , zn−1 ), (ωi , . . . , ωn−1 ), and Πi
such that for all j ∈ {i + 1, . . . , n}, zj = F (zj−1 , ωj−1 ), V(vk, i, z0 , zi , Πi ) = 1,
and zn = z with probability ϵ − negl(λ). Then, because in the base case when
i = 0, V checks that z0 = zi , it is sufficient for E to run E0 to retrieve values
(ω0 , . . . , ωn−1 ). Initially, En simply runs the assumed P ∗ to get a satisfying Πn .
Given extractor Ei that satisfies the inductive hypothesis, we can construct
extractor Ei−1 . Appendix C provides a formal proof.
Lemma 4 (Efficiency). When instantiated with the Pedersen commitment
scheme (Construction 6), we have that |F ′ | = |F | + o(2 · G + 2 · H + R), where
|F | denotes the number of R1CS constraints to encode a function F , G is the
number of constraints required to encode a group scalar multiplication, H is the
number of constraints required to encode hash, and R is the number of constraints
to encode the RO ρ.
Proof. On input instances U and u, NIFS.V computes E ← U.E + r · T + r2 · u.E
and W ← U.W + r · u.W . However, by construction, u.E = u⊥ .E = 0. So, NIFS.V
computes two group scalar multiplications, as it does not need to compute r2 ·u.E.
NIFS.V additionally invokes the RO once to obtain a random scalar. Finally, F ′
makes two additional calls to hash (details are in the description of F ′ ).

20
5.2 Compressing IVC Proofs with zkSNARKs

To prove a statement about an incremental computation, the prover can produce

an IVC proof using the construction in the prior section and send the IVC
proof to the verifier. However, this does not satisfy zero-knowledge (as the IVC
proof described in the prior section does not hide the prover’s non-deterministic
inputs) and succinctness (as the IVC proof size is linear in the size F ). In theory,
one can address this problem with any zkSNARK for NP. Specifically, P can
produce a zkSNARK proving that it knows Πi such that IVC verifier V accepts
for statement (i, z0 , zi ). Naturally, the proof sent to the verifier is succinct and
zero-knowledge due to the corresponding properties of the zkSNARK.
Unfortunately, employing an off-the-shelf zkSNARK makes the overall solu-
tion impractical as the zkSNARK prover must prove, among other things, the
knowledge of vectors whose commitments equal a particular value; this requires
encoding a linear number of group scalar multiplications in the programming
model of zkSNARKs (e.g., R1CS or circuits). To address this, we design a zk-
SNARK tailored for our particular purpose and we describe it in Section 6. Below,
we describe how to use a zkSNARK to prove the knowledge of a valid IVC proof.
Formally, we design a zkSNARK for the following relation.

Definition 13 (IVC Proof Validity Relation). Let IVC = (G, K, P, V) denote

the IVC scheme described in Construction 3. We define the relation RVIVC over
public parameter, structure, instance, and witness tuples as follows.
 
vk ← IVC.K(pp, F ),
RVIVC = (pp, F, (n, z0 , zn ), Πn )
IVC.V(vk, (i, z0 , zi ), Π) = 1

In a nutshell, we leverage the fact that Π contains two committed relaxed

R1CS instance-witness pairs. So, P first folds the instance-witness pairs (u, w)
and (U, W) in Π to produce a folded instance-witness pair (U′ , W′ ), using NIFS.P.
Next, P runs zkSNARK.P to prove that it knows a valid witness for U′ . One
caveat is that for zero-knowledge to hold, we need that Π is honestly randomized.
We capture this qualification formally in Theorem 4.

Construction 4 (A zkSNARK of a Valid IVC Proof ). Let IVC = (G, K, P, V)

denote the IVC scheme in Construction 3, let NIFS denote the non-interactive
folding scheme in Construction 2, and let hash denote a randomized crypto-
graphic hash function that provides hiding. Assume a zkSNARK (Definition 2),
zkSNARK, for committed relaxed R1CS that has the same public parameter gen-
erator algorithm. We construct a zkSNARK (G, K, P, V) for the relation RVIVC
(Definition 13) as follows.

G(1λ ) → pp:

Output pp ← zkSNARK.G(1λ )

K(pp, F ) → (pk, vk):

21
(1) Compute (pkIVC , vkIVC ) ← IVC.K(pp, F ).
(2) Compute sF ′ ← AUGMENT(pp, F ).
(3) Compute (pkzkSNARK , vkzkSNARK ) ← zkSNARK.K(pp, sF ′ ).
(4) Output pk ← ((pkIVC , pkzkSNARK ) and vk ← (vkIVC , vkzkSNARK )).

P(pk, (n, z0 , zn ), Πn )) → π:

If n = 0, output ⊥;
otherwise,
(1) parse Πn as ((Un , Wn ), (un , wn ), rn )
(2) compute (U′ , W′ , T n ) ← NIFS.P(pkIVC , ((Un , Wn ), (un , wn )))
(3) compute πU′ ← zkSNARK.P(pkzkSNARK , U′ , W′ )
(4) output (Un , un , rn , T n , πU′ ).

V(vk, (n, z0 , zn ), π) → {0, 1}:

If n = 0, check that z0 = zi ;
otherwise,
(1) parse π as (Un , un , rn , T n , πU′ ),
(2) check that un .x = hash(vkIVC , i, z0 , zn , Un , rn ),
(3) check that (u.E, u.u) = (u⊥ .E, 1),
(4) compute U′ ← NIFS.V(vkIVC , Un , un , T n ), and
(5) check that zkSNARK.V(vkzkSNARK , U′ , πU′ ) = 1.

Theorem 4 (A zkSNARK of a Valid IVC Proof ). Construction 4 is a

SNARK of a valid IVC proof produced by Construction 3. It provides zero-
knowledge if the adversary A is restricted to sample the instance-witness pair
((n, z0 , zn ), Πn ) with the following strategy, where A⋆ is an arbitrary PPT algo-
rithm:

A(pp) → (F, (n, z0 , zi ), Πn ):

1. (F, n, z0 , ω0 , . . . , ωn−1 ) ← A⋆ (pp)

2. (pkIVC , vkIVC ) ← IVC.K(pp, F )
3. Π0 ← ⊥
4. for i = 1 to n
(a) Πi ← IVC.P(pkIVC , (i, z0 , zi−1 ), ωi−1 , Πi−1 )
(b) zi ← F (zi−1 , ωi−1 )
5. Output (F, (n, z0 , zn ), Πn )

Proof Intuition. Completeness and knowledge soundness hold due to the com-
pleteness and knowledge soundness of the underlying zkSNARK and the non-
interactive folding scheme. Assuming the non-interactive folding scheme satisfies
succinctness (e.g., by using the Pedersen commitment scheme), succinctness holds

22
due to the fact that u, U, and T are succinct, and due to the succinctness of the
underling zkSNARK.
To prove zero-knowledge, we leverage the restriction on the adversarily strat-
egy to generate instance-witness pairs for the zkSNARK. In particular, we con-
struct a simulator S that first iteratively simulates (Ui , ui ) for all i ∈ {1, . . . , n}.
Specifically, given the simulated values (Ui , ui ), S first uses the simulator of the
non-interactive folding scheme to simulate T i . S then folds Ui and ui using T i to
produce Ui+1 . S simulates ui using the observation that all terms are randomized.
The only exception is un .x, which is set to hash(vkNIFS , n, z0 , zn , Un , rn ), where
rn ∈ F is sampled randomly. In the final round, S folds un and Un (again using
a simulated T n ) to produce an instance U′ , and then uses the simulator of the
zkSNARK to produce πU′ . S outputs (Un , un , rn , T n , πU′ ). We provide a formal
proof in Appendix D.

6 A zkSNARK for Committed Relaxed R1CS

As described in Section 5.2, Nova needs a zkSNARK for committed relaxed R1CS
to prove the knowledge of a valid IVC proof succinctly and in zero-knowledge.
This section presents such a zkSNARK by adapting Spartan [41]. We build on
Spartan [41] to avoid FFTs and a trusted setup.

6.1 Background
We assume familiarity with polynomials. We provide background in Appendix G.
Definition 14 (Polynomial Extension). Suppose f : {0, 1}ℓ → F is a function
that maps ℓ-bit strings to an element of F . A polynomial extension of f is a low-
degree ℓ-variate polynomial fe : F ℓ → F such that fe(x) = f (x) for all x ∈ {0, 1}ℓ .
A multilinear extension (MLE) of a function f : {0, 1}ℓ → F is a low-degree
polynomial extension where the extension is a multilinear polynomial.
Every function f : {0, 1}ℓ → F has a unique MLE, and conversely every
ℓ-variate multilinear polynomial over F extends a unique function mapping
{0, 1}ℓ → F . Below, we use fe to denote the unique MLE of f .
Lemma 5 (The Sum-Check Protocol [38]). For ℓ-variate polynomial G over
F with degree at most µ in each variable, there exists a public-coin interactive
proof
P protocol (known as the sum-check protocol) to reduce theℓ task of checking
x∈{0,1}ℓ G(x) = T to the task of checking G(r) = e for r ∈ F . The interaction
consists of a total of ℓ rounds, where in each round the verifier sends a single
element of F and the prover responds with µ + 1 elements of F .

6.2 A Polynomial IOP for Idealized Relaxed R1CS

Our exposition below is based on Spartan [41] and its recent recapitulation [37].
The theorem below and its proof is a verbatim adaptation of Spartan’s polynomial
IOP for R1CS to relaxed R1CS.

23
 Recall that an interactive proof (IP) [28] for a relation R is an interactive
protocol between a prover and a verifier where the prover proves the knowledge of
a witness w for a prescribed instance u such that (u, w) ∈ R. An interactive oracle
proof (IOP) [6, 40] generalizes interactive proofs where in each round the prover
may send an oracle (e.g., a string) and the verifier may query a previously-sent
oracle during the remainder of the protocol. A polynomial IOP [19] is an IOP in
which the oracle sent by the prover is a polynomial and the verifier may query for
an evaluation of the polynomial at a point in its domain. We consider a (minor)
variant of polynomial IOPs, where the verifier has oracle access to polynomials
in the R1CS structure and instance.
We first construct a polynomial IOP for an idealized version of relaxed
R1CS (Definition 15) where the instance contains a purported witness. We then
compile it into a zkSNARK for committed relaxed R1CS (Definition 12).

Definition 15 (Idealized Relaxed R1CS). Consider a finite field F . Let the

public parameters consist of size bounds m, n, ℓ ∈ N where m > ℓ. An idealized
relaxed R1CS structure consists of sparse matrices A, B, C ∈ F m×m with at most
n = Ω(m) non-zero entries in each matrix. An idealized relaxed R1CS instance
consists of an error vector E ∈ F m , a scalar u ∈ F , witness vector W ∈ F m ,
and public inputs and outputs x ∈ F ℓ . An instance (E, u, W, x) is satisfying if
(A · Z) ◦ (B · Z) = u · (C · Z) + E, where Z = (W, x, u).

Construction 5 (Polynomial IOP for Idealized Relaxed R1CS). Consider

an idealized relaxed R1CS statement φ consisting of public parameters (m, n, ℓ),
structure (A, B, C), and instance (E, u, W, x), Without loss of generality, we
assume that m and n are powers of 2 and that m = 2 · (ℓ + 1).
Let s = log m. We interpret the matrices A, B, C as functions with signature
{0, 1}log m × {0, 1}log m → F in a natural manner. In particular, an input in
{0, 1}log m × {0, 1}log m is interpreted as the binary representation of an index
(i, j) ∈ [m] × [m], where [m] := {1, . . . , m} and the function outputs (i, j)th entry
of the matrix. As such, let A, e B,
e and C e denote multilinear extensions of A, B,
and C interpreted as functions, so they are 2 log m-variate sparse multilinear
polynomials of size n. Similarly, we interpret E and W as functions with respective
signatures {0, 1}log m → F and {0, 1}log m−1 → F . Furthermore, let E e and W f
denote the multilinear extensions of E and W interpreted as functions, so they
are multilinear polynomials in log m and log m − 1 variables respectively.
As noted earlier, the verifier has an oracle access to the following polynomials:
A,
e B,e C,
e E,
e and W f . Additionally, the verifier reads u and x in entirety.
Let Z = (W, x, u). Similar to how we interpret matrices as functions, we
interpret Z and (x, u) as functions with the following respective signatures:
{0, 1}s → F and {0, 1}s−1 → F . Observe that the MLE Z e of Z satisfies

e 1 , . . . , Xs ) = (1 − X1 ) · W
Z(X ]
f (X2 , . . . , Xs ) + X1 · (x, u)(X2 , . . . , Xs ) (1)

Similar to [41, Theorem 4.1], checking if φ is satisfiable is equivalent, except

for a soundness error of log m/|F | over the choice of τ ∈ F s , to checking if the

24
following identity holds:
?
X
0= eq(τ,
e x) · F (x), (2)
x∈{0,1}s

where
   
X X
F (x) =  e y) · Z(y)
A(x, e · e y) · Z(y)
B(x, e −
y∈{0,1}s y∈{0,1}s
 
X
u · e y) · Z(y)
C(x, e + E(x)
e ,
y∈{0,1}s

e is the multilinear extension of eq : {0, 1}s × {0, 1}s → F where eq(x, e) = 1

and eq
if x = e and 0 otherwise.
That is, if φ is satisfiable, then Equation (2) holds with probability 1 over
the choice of τ , and if not, then Equation (2) holds with probability at most
O(log m/|F |) over the random choice of τ .
To compute the right-hand side in Equation (2), the prover and the verifier
apply the sum-check protocol to the following polynomial: g(x) := eq(τ, e x) · F (x)
From the verifier’s perspective, this reduces the task of computing the right-hand
side of Equation (2) to the task of evaluating g at a random input rx ∈ F s . Note
that the verifier
Qs can locally evaluate eq(τ, e rx ) in O(log m) field operations via
eq(τ, rx ) = i=1 (τi rx,i + (1 − τi )(1 − rx,i )). With eq(τ, r ) in hand, g(rx ) can
Px
e e
be computed in O(1) time given the four quantities: y∈{0,1}s A(r e x , y) · Z(y),
e
P P
y∈{0,1}s B(rx , y) · Z(y), y∈{0,1}s C(rx , y) · Z(y), and E(rx ).
e e e e e
The last quantity can be computed with a single query to polynomial E. e
Furthermore, the first three quantities can be computed by applying the sum-
check protocol three more times in parallel, once to each of the following three
polynomials (using the same random vector of field elements, ry ∈ F s , in each of
the three invocations): A(re x , y) · Z(y),
e e x , y) · Z(y),
B(r e e x , y) · Z(y).
and C(r e
To perform the verifier’s final check in each of these three invocations of the
sum-check protocol, it suffices for the verifier to evaluate each of the above three
polynomials at the random vector ry , which means it suffices for the verifier to
evaluate A(r
e x , ry ), B(r
e x , ry ), C(r
e x , ry ), and Z(r
e y ). The first three evaluations
can be obtained via the verifier’s assumed query access to (A, e B,
e C).
e Z(re y ) can
be computed (via Equation (1)) from a query to W f and from computing (x, ] u).
In summary, we have the following polynomial IOP.

1. V → P: τ ∈R F s
2. V ↔ P: run the sum-check protocol to reduce the check in Equation (2)
to checking if the following hold, where rx , ry are vectors in F s chosen at
random by the verifier over the course of the sum-check protocol:
? ? ?
– A(r
e x , ry ) = vA , B(r
e x , ry ) = vB , and C(r
e x , ry ) = vC ;

25
 ?
– E(r
e x) = vE ; and
?
– Z(ry ) = vZ .
e
3. V:
? ? ?
– check if A(r
e x , ry ) = vA , B(r
e x , ry ) = vB , and C(r
e x , ry ) = vC , with a query
to A,
e B,
e C e at (rx , ry );
?
– check if E(r
e x) = vE with an oracle query to E; e and
?
e y ) = vZ by checking if: vZ = (1 − ry [1]) · vW + ry [1] ·
– check if Z(r
]
(x, u)(ry [2..]), where ry [2..] refers to a slice of ry without the first element
of ry , and vW ← W f (ry [2..]) via an oracle query (see Equation (1)).

Theorem 5. Construction 5 is a polynomial IOP for idealized relaxed R1CS

defined over a finite field F , with the following parameters, where m denotes the
dimension of the R1CS matrices, and n denotes the number of non-zero entries in
the matrices: Soundness error is O(log m)/|F |; round complexity is O(log m); The
verifier has query access to 2 log m-variate multilinear polynomials A,
e B,
e Ce in the
structure, and (log m)-variate multilinear polynomial E, and (log m − 1)-variate
e
multilinear polynomial W f in the instance; the verifier issues a single query to
polynomials A, B, C, and W
e e e f , E,
e and otherwise performs O(log m) operations
over F ; the prover performs O(n) operations over F to compute its messages in
the polynomial IOP and to respond to the verifier’s queries to (W f , E,
e A,
e B,
e C).
e

Proof. Perfect completeness follows from perfect completeness of the sum-check

protocol and the fact that Equation (2) holds with probability 1 over the choice
of τ if φ is satisfiable. Applying a standard union bound to the soundness error
introduced by probabilistic check in Equation (2) with the soundness error of the
sum-check protocol [38], we conclude that the soundness error for the depicted
polynomial IOP as at most O(log m)/|F |.
The sum-check protocol is applied four times (although three of the invocations
occur in parallel and in practice combined into one [41]). In each invocation, the
polynomial to which the sum-check protocol is applied has degree at most 3 in each
variable, and the number of variables is s = log m. Hence, the round complexity
of the polynomial IOP is O(log m). Since each polynomial has degree at most 3
in each variable, the total communication cost is O(log m) field elements.
The claimed verifier runtime is immediate from the verifier’s runtime in
the sum-check protocol, and the fact that eq e can be evaluated at any input
(τ, rx ) ∈ F 2s in O(log m) field operations. As in Spartan [41], the prover’s work in
the polynomial IOP in O(n) operations over F using prior techniques [45, 51].

6.3 Compiling Polynomial IOPs to zkSNARKs

As in prior works [19, 22, 41], we compile our polynomial IOP into a zkSNARK
using a polynomial commitment scheme [32] and the Fiat-Shamir transform [25].

26
Interpreting commitments to vectors as polynomial commitments. It is
well known that commitments to m-sized vectors over F are commitments to log m-
variate multilinear polynomials represented with evaluations over {0, 1}m [35, 41,
49, 52]. Furthermore, there is a polynomial commitment scheme for log m-variate
multilinear polynomials if there exists an argument protocol to prove an inner
product computation between a committed vector and an m-sized public vector
((r1 , 1 − r1 ) ⊗ . . . ⊗ (rlog m , 1 − rlog m )), where r ∈ F log m is an evaluation point.
There are two candidate constructions in the literature. Note that the primary
difference between two schemes is in the verifier’s time.

1. PCBP . If the commitment scheme for vectors over F is Pedersen’s commit-

ments (Construction 6), as in prior work [49], Bulletproofs [16] provides
a suitable inner product argument protocol. The polynomial commitment
scheme here achieves the following efficiency characteristics, assuming the
hardness of the discrete logarithm problem. For a log m-variate multilinear
polynomial, committing takes Oλ (m) time to produce an Oλ (1)-sized com-
mitment; the prover incurs Oλ (m) costs to produce an evaluation proof of
size Oλ (log m) that can be verified in Oλ (m). Note that PCBP is a special
case of Hyrax’s polynomial commitment scheme [49].
2. PCDory . If vectors over F are committed with a two-tiered “matrix” commit-
ment (see for example, [20, 35]), which provides Oλ (1)-sized commitments to
m-sized vectors under the SXDH assumption. With this commitment scheme,
Dory [35] provides the necessary inner product argument. The polynomial
commitment here achieves the following efficiency characteristics, assuming
the hardness of SXDH. For a log m-variate multilinear polynomial, commit-
ting takes Oλ (m) time to produce an Oλ (1)-sized commitment; the prover
incurs Oλ (m) costs to produce an evaluation proof of size Oλ (log m) that
can be verified in Oλ (log m).

Polynomial commitments for sparse multilinear polynomials. In our

constructions below, we require polynomial commitment schemes that can ef-
ficiently handle sparse multilinear polynomials. Spartan [41, §7] (and its opti-
mization [44, §6]) provides a generic compiler to transform existing polynomial
commitment schemes for multilinear polynomials into those that can efficiently
handle sparse multilinear polynomials. Specifically, we apply [37, Theorem 5])
(which captures Spartan’s compiler in a generic manner) to PCBP and PCDory to
obtain their variants that can efficiently handle sparse multilinear polynomials;
we refer to them as “Sparse-PCBP ” and “Sparse-PCDory ” respectively.

Theorem 6 (A zkSNARK from PCBP ). Assuming the hardness of the discrete

logarithm problem, there exists a zkSNARK in the random oracle model for
committed relaxed R1CS with the following efficiency characteristics, where m
denotes the dimensions of R1CS matrices and n denotes the number of non-zero
entries in the matrices: The encoder runs in time Oλ (n); The prover runs in
time Oλ (n); The proof length is Oλ (log n); and the verifier runs in time Oλ (n).7
7
Appendix H describes a minor optimization and a corresponding Corollary.

27
Proof. For R1CS structure (A, B, C), we first have the encoder directly provide
(A,
e B,
e C)e in the prover key, and additionally provide sparse polynomial commit-
ments to A, e B,
e Ce using Sparse-PCBP in both the prover and verifier keys. Next, we
apply the compiler of [19] using PCBP to the polynomial IOP from Construction 5.
At a high level, this replaces all of the oracles provided to the verifier with PCBP
commitments, which the prover and verifier then use to simulate ideal queries to
a committed oracle. By [19, Theorem 6] this provides a public-coin honest-verifier
zero-knowledge interactive argument of knowledge. In particular, we can treat
the resulting protocol as an argument for committed relaxed R1CS because the
verifier is now provided with (polynomial) commitments to E and W . Applying
the Fiat-Shamir transform [25] achieves non-interactivity and zero-knowledge in
the random oracle model.
The claimed efficiency follows from the efficiency of the polynomial IOP,
PCBP , and Sparse-PCBP . In more detail, using Sparse-PCBP , the encoder takes
Oλ (n) time to create commitments 2 log m-variate sparse multilinear polynomials
A,
e B,
e C.
e The prover’s costs in the polynomial IOP is O(n). Furthermore, proving
the evaluations of two O(log m)-variate multilinear polynomials using PCBP , it
takes Oλ (m) time. And, to prove the evaluations of three 2 log m-variate sparse
multilinear polynomials of size n, using Sparse-PCBP , it takes Oλ (n) time. In
total, the prover time is Oλ (n). The proof length in the polynomial IOP is
O(log m), and the proof sizes in the polynomial evaluation proofs is Oλ (log n),
so the proof length is Oλ (log n). The verifier’s time in the polynomial IOP is
O(log m). In addition, it verifies five polynomial evaluations, which costs Oλ (n)
time: the two polynomial in the instance take Oλ (m) time using PCBP , and the
three polynomials in the structure takes Oλ (n) time using Sparse-PCBP . So, in
total, the verifier time is Oλ (n).

Corollary 1 (A zkSNARK from PCDory ). Assuming the hardness of the

SXDH problem, there exists a zkSNARK in the random oracle model for committed
relaxed R1CS with the following efficiency characteristics, where m denotes the
dimensions of R1CS matrices and n denotes the number of non-zero entries in
the matrices: The encoder runs in time Oλ (n); The prover runs in time Oλ (n);
The proof length is Oλ (log n); and the verifier runs in time Oλ (log n).

Acknowledgments

We thank Arasu Arun, Weidong Cui, Justin Drake, Bryan Parno, and Justin
Thaler for various conversations and helpful comments on a prior version of this
paper. We thank Andy Polyakov and Kelly Olson for help with pasta-msm. We
thank Chhi’mèd Künzang for helpful conversations and for PRs to the Nova
repository, including an adapter that enables transforming circuits written with
bellperson to a format that Nova uses. We thank Wilson Nguyen for pointing
out a discrepancy in the zero-knowledge property proven in a prior version
of Appendix D. Abhiram Kothapalli (while at Carnegie Mellon University)
was supported by a gift from Bosch, NSF Grant No. 1801369, and by the

28
CONIX Research Center, one of six centers in JUMP, a Semiconductor Research
Corporation (SRC) program sponsored by DARPA.

29
 References

[1] bellperson. [Link]

[2] neptune. [Link]
[3] Nova: Recursive SNARKs without trusted setup. [Link]
Microsoft/Nova
[4] Pasta curves. [Link] curves
[5] pasta-msm. [Link]
[6] Ben-Sasson, E., Chiesa, A., Spooner, N.: Interactive Oracle Proofs. In: TCC
(2016)
[7] Ben-Sasson, E., Chiesa, A., Tromer, E., Virza, M.: Scalable zero knowledge
via cycles of elliptic curves. In: CRYPTO (2014)
[8] Bitansky, N., Canetti, R., Chiesa, A., Tromer, E.: From extractable collision
resistance to succinct non-interactive arguments of knowledge, and back
again. In: ITCS (2012)
[9] Bitansky, N., Canetti, R., Chiesa, A., Tromer, E.: Recursive composition
and bootstrapping for SNARKs and proof-carrying data. In: STOC (2013)
[10] Boneh, D., Bünz, B., Fisch, B.: A survey of two verifiable delay functions.
Cryptology ePrint Archive, Report 2018/712 (2018)
[11] Boneh, D., Drake, J., Fisch, B., Gabizon, A.: Halo Infinite: Recursive zk-
SNARKs from any Additive Polynomial Commitment Scheme. Cryptology
ePrint Archive, Report 2020/1536 (2020)
[12] Bootle, J., Cerulli, A., Chaidos, P., Groth, J., Petit, C.: Efficient zero-
knowledge arguments for arithmetic circuits in the discrete log setting. In:
EUROCRYPT (2016)
[13] Bowe, S., Grigg, J., Hopwood, D.: Halo: Recursive proof composition without
a trusted setup. Cryptology ePrint Archive, Report 2019/1021 (2019)
[14] Bowe, S., Grigg, J., Hopwood, D.: Halo2 (2020), [Link]
halo2
[15] Brakerski, Z., Holmgren, J., Kalai, Y.T.: Non-interactive delegation and
batch np verification from standard computational assumptions. In: STOC.
pp. 474–482 (2017)
[16] Bünz, B., Bootle, J., Boneh, D., Poelstra, A., Wuille, P., Maxwell, G.:
Bulletproofs: Short proofs for confidential transactions and more. In: S&P
(2018)
[17] Bünz, B., Chiesa, A., Lin, W., Mishra, P., Spooner, N.: Proof-carrying data
without succinct arguments. Cryptology ePrint Archive, Report 2020/1618
(2020)
[18] Bünz, B., Chiesa, A., Mishra, P., Spooner, N.: Proof-carrying data from
accumulation schemes. In: TCC (2020)
[19] Bünz, B., Fisch, B., Szepieniec, A.: Transparent SNARKs from DARK
compilers. In: EUROCRYPT (2020)
[20] Bünz, B., Maller, M., Mishra, P., Vesely, N.: Proofs for inner pairing products
and applications. Cryptology ePrint Archive, Report 2019/1177 (2019)
[21] Chen, W., Chiesa, A., Dauterman, E., Ward, N.P.: Reducing participation
costs via incremental verification for ledger systems. Cryptology ePrint
Archive, Report 2020/1522 (2020)
[22] Chiesa, A., Hu, Y., Maller, M., Mishra, P., Vesely, N., Ward, N.: Marlin: Pre-
processing zkSNARKs with universal and updatable SRS. In: EUROCRYPT
(2020)
[23] Chiesa, A., Ojha, D., Spooner, N.: Fractal: Post-quantum and transparent
recursive proofs from holography. In: EUROCRYPT (2020)
[24] Drucker, A.: New limits to classical and quantum instance compression.
SIAM Journal of Computing 44(5), 144–3–1479 (2015)
[25] Fiat, A., Shamir, A.: How to prove yourself: Practical solutions to identifica-
tion and signature problems. In: CRYPTO. pp. 186–194 (1986)
[26] Gennaro, R., Gentry, C., Parno, B., Raykova, M.: Quadratic span programs
and succinct NIZKs without PCPs. In: EUROCRYPT (2013)
[27] Gentry, C., Wichs, D.: Separating succinct non-interactive arguments from
all falsifiable assumptions. In: STOC. pp. 99–108 (2011)
[28] Goldwasser, S., Micali, S., Rackoff, C.: The knowledge complexity of interac-
tive proof-systems. In: STOC (1985)
[29] Grassi, L., Khovratovich, D., Rechberger, C., Roy, A., Schofnegger, M.:
Poseidon: A new hash function for zero-knowledge proof systems. Cryptology
ePrint Archive, Paper 2019/458 (2019)
[30] Groth, J.: On the size of pairing-based non-interactive arguments. In: EU-
ROCRYPT (2016)
[31] Harnik, D., Naor, M.: On the compressibility of N P instances and crypto-
graphic applications. SIAM Journal of Computing 39(5), 1667–1713 (2010)
[32] Kate, A., Zaverucha, G.M., Goldberg, I.: Constant-size commitments to
polynomials and their applications. In: ASIACRYPT. pp. 177–194 (2010)
[33] Kilian, J.: A note on efficient zero-knowledge proofs and arguments (extended
abstract). In: STOC (1992)
[34] Labs, O.: Mina cryptocurrency (2020), [Link]
[35] Lee, J.: Dory: Efficient, transparent arguments for generalised inner products
and polynomial commitments. Cryptology ePrint Archive, Report 2020/1274
(2020)
[36] Lee, J., Nikitin, K., Setty, S.: Replicated state machines without replicated
execution. In: S&P (2020)
[37] Lee, J., Setty, S., Thaler, J., Wahby, R.: Linear-time zero-knowledge SNARKs
for R1CS. Cryptology ePrint Archive, Report 2021/030 (2021)
[38] Lund, C., Fortnow, L., Karloff, H., Nisan, N.: Algebraic methods for interac-
tive proof systems. In: FOCS (Oct 1990)
[39] Micali, S.: CS proofs. In: FOCS (1994)
[40] Reingold, O., Rothblum, G.N., Rothblum, R.D.: Constant-round interactive
proofs for delegating computation. In: STOC. pp. 49–62 (2016)
[41] Setty, S.: Spartan: Efficient and general-purpose zkSNARKs without trusted
setup. In: CRYPTO (2020)
[42] Setty, S., Angel, S., Gupta, T., Lee, J.: Proving the correct execution of
concurrent services in zero-knowledge. In: OSDI (Oct 2018)

31
[43] Setty, S., Braun, B., Vu, V., Blumberg, A.J., Parno, B., Walfish, M.: Resolv-
ing the conflict between generality and plausibility in verified computation.
In: EuroSys (Apr 2013)
[44] Setty, S., Lee, J.: Quarks: Quadruple-efficient transparent zkSNARKs. Cryp-
tology ePrint Archive, Report 2020/1275 (2020)
[45] Thaler, J.: Time-optimal interactive proofs for circuit evaluation. In:
CRYPTO (2013)
[46] Thaler, J.: Proofs, arguments, and zero-knowledge. [Link]
[Link]/jthaler/[Link] (2020)
[47] Valiant, P.: Incrementally verifiable computation or proofs of knowledge
imply time/space efficiency. In: TCC. pp. 552–576 (2008)
[48] Wahby, R.S., Ji, Y., Blumberg, A.J., Shelat, A., Thaler, J., Walfish, M.,
Wies, T.: Full accounting for verifiable outsourcing. In: CCS (2017)
[49] Wahby, R.S., Tzialla, I., Shelat, A., Thaler, J., Walfish, M.: Doubly-efficient
zkSNARKs without trusted setup. In: S&P (2018)
[50] Wesolowski, B.: Efficient verifiable delay functions. In: EUROCRYPT. pp.
379–407 (2019)
[51] Xie, T., Zhang, J., Zhang, Y., Papamanthou, C., Song, D.: Libra: Succinct
zero-knowledge proofs with optimal prover computation. In: CRYPTO (2019)
[52] Zhang, Y., Genkin, D., Katz, J., Papadopoulos, D., Papamanthou, C.: vSQL:
Verifying arbitrary SQL queries over dynamic outsourced databases. In: S&P
(2017)

32
 Supplementary Materials

A Additional Related Work

A folding scheme is reminiscent of the technique used in protocols such as

Bulletproofs [16], where the prover splits an N -sized inner product instance
into two N/2-sized inner product instances, and then the prover and the verifier
interactively combine the two N/2-sized instances into a single N/2-sized instance.
A folding scheme differs from this technique in that it targets folding NP relations.
Furthermore, our folding scheme for relaxed R1CS folds arbitrary NP instances.
A folding scheme is also reminiscent of the sum-check protocol [38] applied to
layered circuits [48, 49], where the prover and the verifier interactively combine
β ≥ 2 instances of depth-1 circuit satisfiability into a single depth-1 circuit
satisfiability instance, and the combined instance is proven using additional
invocations of the sum-check protocol. Once the protocol combines β instances
of depth-1 circuit satisfiability into a single depth-1 circuit satisfiability instance,
it is not entirely clear how to fold additional circuit satisfiability instances into
a single running instance. Whereas, with a folding scheme, the prover and the
verifier can continually fold NP instances into a single running instance.
Finally, a folding scheme is also related to other prior notions such as “batch
verification” [15, 40] and “instance compression” [24, 31]. Given β NP instances,
batch verification protocols aim to produce a succinct proof that all β instances are
satisfiable. In contrast, a folding scheme’s goal is to transform the satisfiability
of two NP instances into a single NP-instance. (AND)-instance compression
schemes efficiently compress β SAT instances into a single short string that
retains information about whether all instances are satisfiable. A folding scheme
differs from this notion in that compression is aided by an (untrusted) prover
who knows (purported) witnesses to the original β instances. Furthermore, unlike
instance compression, a folding scheme can be interactive.

B Proof of Theorem 3 (A Folding Scheme for Committed

Relaxed R1CS)

Lemma 6 (Completeness). Construction 1 is a folding scheme for committed

relaxed R1CS with perfect completeness.

Proof. Consider a finite field F , a succinct, hiding, and homomorphic commitment

scheme Com over F , and public parameters consisting of size bounds m, n, ℓ ∈ N
where m > ℓ, and commitment parameters ppW and ppE for vectors of size m
and m − ℓ − 1 respectively sampled by the generator. Consider an adversarially
chosen R1CS structure (A, B, C) ∈ F m×m and two committed relaxed R1CS
instances

φ1 = (E 1 , u1 , W 1 , x1 ) and φ2 = (E 2 , u2 , W 2 , x2 ).

33
Suppose that the prover P, in addition to the two instances, holds satisfying
witnesses to both instances, (E1 , rE1 , W1 , rW1 ) and (E2 , rE2 , W2 , rW2 ). Let Z1 =
(W1 , x1 , u1 ) and Z2 = (W2 , x2 , u2 ).
Now suppose that P and V compute a folded instance φ = (E, u, W , x), and
suppose that P computes a folded witness (E, rE , W, rW ), To prove completeness,
we must show that (E, W ) is a satisfying witness for instance φ. Let Z = (W, x, u).
For (E, W ) to be a satisfying witness, we must have the following:

AZ ◦ BZ = u · CZ + E (3)

and

E = Com(ppE , E, rE ) (4)
W = Com(ppW , W, rW ) (5)

It is easy to see that Equations (4) and (5) hold from the additive homomorphism
of the commitment scheme.
Thus, we focus on proving that Equation (3) holds. By construction, for
Equation (3) to hold, we must have for r ∈R F

A(Z1 + r · Z2 ) ◦ B(Z1 + r · Z2 ) = (u1 + r · u2 ) · C(Z1 + r · Z2 ) + E.

Distributing, we must have

AZ1 ◦ BZ1 + r(AZ1 ◦ BZ2 + AZ2 ◦ BZ1 ) + r2 (AZ2 ◦ BZ2 ) =

u1 · CZ1 + r(u1 · CZ2 + u2 CZ1 ) + r2 · u2 · CZ2 + E.
Aggregating by powers of r, we must have
(AZ1 ◦ BZ1 − u1 · CZ1 )+
r(AZ1 ◦ BZ2 + AZ2 ◦ BZ1 − u1 · CZ2 − u2 CZ1 )+
(6)
r2 (AZ2 ◦ BZ2 − u2 · CZ2 )
=E.
However, because W1 and W2 are satisfying witnesses, we have

AZ1 ◦ BZ1 − u1 · CZ1 = E1

AZ2 ◦ AZ2 − u2 · CZ2 = E2 .

Additionally, by construction we have

AZ1 ◦ BZ2 + AZ2 ◦ BZ1 − u1 · CZ2 − u2 · CZ1 = T.

Thus, by substitution, for Equation (6) to hold we must have

E1 + r · T + r2 · E2 = E,

which holds by construction.

34
Lemma 7 (Knowledge Soundness). Construction 1 is a knowledge sound
folding scheme for committed relaxed R1CS.

Proof. Consider a finite field F , a succinct, hiding, and homomorphic commitment

scheme Com over F , and public parameters consisting of size bounds m, n, ℓ ∈ N
where m > ℓ, and commitment parameters ppW and ppE for vectors of size m
and m − ℓ − 1 respectively sampled by the generator. Consider an adversarially
chosen R1CS structure (A, B, C) ∈ F m×m and two committed relaxed R1CS
instances

φ1 = (E 1 , u1 , W 1 , x1 ) and φ2 = (E 2 , u2 , W 2 , x2 ).

We prove knowledge soundness via the forking lemma (Lemma 1). That is,
we prove that there exists a PPT algorithm X such that when given public
parameters pp, structure (A, B, C), and a tree of accepting transcripts and the
corresponding folded instance-witness pairs outputs a satisfying witness with
probability 1 − negl(λ).
In more detail, suppose X is provided three transcripts (τ1 , τ2 , τ3 ) with the
same initial commitment T from the prover. Note that a transcript τi for i ∈
{1, 2, 3} additionally comes attached with an accepting witness τi .(W, E, rW , rE )
and the verifier’s randomness τi .r. Interpolating points (τ1 .r, τ1 .W ) and (τ2 .r, τ2 .W ),
X retrieves (W1 , W2 ) such that

W1 + τi .r · W2 = τi .W (7)

for i ∈ {1, 2}. Similarly interpolating points (τ1 .r, τ1 .E), (τ2 .r, τ2 .E), (τ3 .r, τ3 .E),
X retrieves (E1 , E2 ) and a cross-term T such that

E1 + τi .r · T + τi .r2 · E2 = τi .E (8)

for i ∈ {1, 2, 3}. Using the same approach, X can interpolate for rW1 , rW2 and
rE1 , rT , rE2 . We must argue that (W1 , E1 , rW1 , rE1 ) and (W2 , E2 , rW2 , rE2 ) are
indeed satisfying witnesses for φ1 and φ2 respectively.
We first show that the retrieved witness elements are valid openings to the
corresponding commitments in the instance. For i ∈ {1, 2}, because τi .W, τi .rW
is part of a satisfying witness, by construction,

Com(ppW , W1 , rW1 ) + τi .r · Com(ppW , W2 , rW2 )

= Com(ppW , W1 + τi .r · W2 , rW1 + τi .r · rW2 )
= Com(ppW , τi .W, τi .rW )
= W 1 + τi .r · W 2 .

Interpolating, we must have that

Com(ppW , W1 , rW1 ) = W 1 (9)

Com(ppW , W2 , rW2 ) = W 2 . (10)

35
 Similarly, for i ∈ {1, 2, 3}, because τi .E, τi .rE is part of a satisfying witness,
by construction, we must have

Com(ppE , E1 , rE1 ) + τi .r · Com(ppE , T, rT ) + τi .r2 · Com(ppE , E2 , rE2 )

= Com(ppE , E1 + τi .r · T + τi .r2 · E2 , rE1 + τi .r · rT + τi .r2 · rE2 )
= Com(ppE , τi .E, τi .rE )
= E 1 + τi .r · T + τi .r2 · E 2 .

Interpolating, we must have that

Com(ppE , E1 , rE1 ) = E 1
Com(ppE , E2 , rE2 ) = E 2

Next, we must show that (W1 , E1 ) and (W2 , E2 ) satisfy the relaxed R1CS
relation. To show this, we must first argue that Equation (7) holds for i = 3 as
well (i.e., W1 + τ3 .r · W2 = τ3 .W ). Indeed, Equations (9) and (10) imply that

Com(ppW , W1 + τ3 .r · W2 , rW1 + τ3 .r · rW2 )

= Com(ppW , W1 , rW1 ) + τ3 .r · Com(ppW , W2 , rW2 )
= W 1 + τ3 .r · W 2
= Com(ppW , τ3 .W, τ3 .W )

Thus, by the binding property of Com, we must additionally have that

W1 + τ3 .r · W2 = τ3 .W (11)

with probability 1 − negl(λ).

Because (τi .W, τi .E) is part of a satisfying witness, for i ∈ {1, 2, 3} we have

A(τi .Z) ◦ B(τi .Z) = u · C(τi .Z) + τi .E

where τi .Z = (τi .W, τi .x, τi .u). However, by Equations (7), (8), and (11) for
i ∈ {1, 2, 3}, this implies that with probability 1 − negl(λ)

A(Z1 + τi .r · Z2 ) ◦ B(Z1 + τi .r · Z2 ) =
(u1 + τi .r · u2 ) · C(Z1 + τi .r · Z2 ) + (E1 + τi .r · T + τi .r2 · E2 )

where Z1 = (W1 , x1 , u1 ) and Z2 = (W2 , x2 , u2 ). Expanding and interpolating, we

have that

AZ1 ◦ BZ1 = u1 · CZ1 + E1

AZ2 ◦ BZ2 = u2 · CZ2 + E2

with probability 1−negl(λ). Thus, (W1 , E1 , rW1 , rE1 ) and (W2 , E2 , rW2 , rE2 ) meet
all the requirements to be satisfying witnesses for φ1 and φ2 respectively with
probability 1 − negl(λ).

36
Lemma 8 (Zero-Knowledge). Construction 1 is a zero-knowledge folding
scheme for committed relaxed R1CS.
Proof. Intuitively, zero-knowledge holds because the prover only sends a single
hiding commitment. More formally, the simulator S samples random T ∈ F m and
r ∈ F and computes T = Com(ppE , T, r). Next, S derives the verifier’s challenge,
r, using ρ and outputs tr = (T , r). (Perfect) zero-knowledge holds from the
(perfect) hiding property of the underlying commitment scheme.

C Proofs for IVC from Folding Schemes (Construction 3)

Lemma 9 (IVC Completeness). Construction 3 is an IVC scheme that sat-
isfies perfect completeness.
Proof. Consider (F, i + 1, z0 , zi+1 ) and the corresponding inputs (zi , ωi ) such that

zi+1 = F (zi , ωi )

Let pp ← G(1λ ), and let (pk, vk) ← K(pp, F ). Now consider a proof Πi such that

V(vk, i, z0 , zi , Πi ) = 1

We must show that given

Πi+1 ← P(pk, (i, z0 , zi ), ωi , Πi )

that

V(vk, i + 1, z0 , zi+1 , Πi+1 ) = 1

with probability 1. We show this by induction on i.

Base Case (i = 0): Suppose the prover is provided Π0 such that

V(vk, 0, z0 , z0 , Π0 ) = 1.

By the base case of P and F ′ , we have

Π1 = ((u⊥ , w⊥ ), (u1 , w1 ), r1 )

for some (u1 , w1 ). By definition, the instance-witness pair (u⊥ , w⊥ ) satisfies F ′ .

Moreover, by construction, (u1 , w1 ) must also be satisfying. Additionally, by the
construction of F ′ , we have

u1 .x = hash(vk, 1, z0 , F (z0 , w0 ), u⊥ , r1 ).

And, by the construction of P, we have w1 .E = 0 and w1 .rE = 0 (implying that

u1 .E = 0), and u1 .u = 1. Therefore, we have

V(pp, 1, z0 , z1 , Π1 ) = 1.

37
Inductive Step (i ≥ 1): Assume that for

Πi = ((Ui , Wi ), (ui , wi ), ri )

we have that

V(vk, i, z0 , zi , Πi ) = 1

and suppose that

Πi+1 = ((Ui+1 , Wi+1 ), (ui+1 , wi+1 ), ri+1 ) ← P(pk, (i, z0 , zi ), ωi , Πi ).

By the construction of P, we have that

(Ui+1 , Wi+1 , T ) ← NIFS.P(pk, (Ui , Wi ), (ui , wi )).

Thus, by the completeness of the underlying folding scheme, and the premise
that (ui , wi ) and (Ui , Wi ) are satisfying instance-witness pairs, we have that
(Ui+1 , Wi+1 ) is a satisfying instance-witness pair. Additionally, by the premise,
we have that ui .x = hash(vk, i, z0 , zi , Ui , ri ), ui .E = 0, and ui .u = 1. Therefore, P
can construct a satisfying instance-witness pair (ui+1 , wi+1 ) that represents the
correct execution of F ′ on input (U, u, (i, z0 , zi ), ωi , T , ri , ri+1 ), where ri+1 ∈R F .
By construction, this particular input implies that

ui+1 .x = hash(vk, ui .i + 1, ui .z0 , F (zi , wi ), NIFS.V(vk, Ui , ui , T ), ri+1 )

(12)
= hash(vk, ui .i + 1, ui .z0 , zi+1 , Ui+1 , ri+1 )
by the correctness of the underlying folding scheme. As in the base case, by the
construction of P, we have wi+1 .E = 0, wi+1 .rE = 0 (implying that ui+1 .E = 0)
and ui+1 .u = 1. Thus, by Equation (12) we have

V(vk, i + 1, z0 , zi+1 , Πi+1 ) = 1.

Lemma 10 (IVC Knowledge Soundness). Construction 3 is an IVC scheme

that satisfies knowledge soundness.
Proof. Our approach is inspired by a recursive extraction technique described by
Bünz et al [17]. Let pp ← G(1λ ). Consider an expected polynomial-time adversary
P ∗ that outputs a function F on input pp, and let (pk, vk) ← K(pp, F ). Suppose
that, for a constant n, P ∗ additionally outputs (z0 , z, Π) such that

V(vk, n, z0 , z, Π) = 1

with probability ϵ. We must construct an expected polynomial-time extractor E

that with input (pp, z0 , z), outputs (ω0 , . . . , ωn−1 ) such that by computing

zi ← F (zi−1 , ωi−1 )

38
we have that zn = z with probability ϵ − negl(λ).
We show inductively that E can construct an expected polynomial-time
extractor Ei (pp) that outputs ((zi , . . . , zn−1 ), (ωi , . . . , ωn−1 ), Πi ) such that for all
j ∈ {i + 1, . . . , n},
zj = F (zj−1 , ωj−1 )
and
V(vk, i, z0 , zi , Πi ) = 1 (13)
for zn = z with probability ϵ − negl(λ). Then, because in the base case when i = 0,
V checks that z0 = zi , the values (ω0 , . . . , ωn−1 ) retrieved by E0 (pp) are such that
computing zi+1 = F (zi , ωi ) for all i ≥ 1 gives zn = z. At a high level, to construct
an extractor Ei−1 , we first assume the existence of Ei that satisfies the inductive
hypothesis. We then use Ei (pp) to construct an adversary for the non-interactive
folding scheme (which we denote as P ei−1 ). This in turn guarantees an extractor
for the non-interactive folding scheme, which we denote as Eei−1 . We then use
Eei−1 to construct Ei−1 that satisfies the inductive hypothesis.
In the base case, for i = n, let En (pp; ρ) output (⊥, ⊥, Πn ) where Πn is the
output of P ∗ (pp; ρ). By the premise, En succeeds with probability ϵ in expected
polynomial-time.
For i ≥ 1, suppose E can construct an expected polynomial-time extractor Ei
that outputs ((zi , . . . , zn−1 ), (ωi , . . . , ωn−1 )), and Πi that satisfies the inductive
hypothesis. To construct an extractor Ei−1 , E first constructs an adversary P ei−1
for the non-interactive folding scheme as follows:

P
ei−1 (pp; ρ):

(1) Let ((zi , . . . , zn−1 ), (ωi , . . . , ωn−1 ), Πi ) ← Ei (pp; ρ).

(2) Parse Πi as ((Ui , Wi ), (ui , wi ), ri ).
(3) Parse wi to retrieve (Ui−1 , ui−1 , T i−1 , ri−1 ).
(4) Output (Ui−1 , ui−1 ) and ((Ui , Wi ), T i−1 ).
By the inductive hypothesis, we have that V(vk, i, z0 , zi , Πi ) = 1, where Πi ←
Ei (pp) with probability ϵ − negl(λ). Therefore, by the the verifier’s checks we
have that (ui , wi ) and (Ui , Wi ) are satisfying instance-witness pairs, and that
ui .x = hash(vk, i, z0 , zi , Ui , ri ).

Because V ensures that (ui .E, ui .u) = (0, 1) we have that wi is indeed a satisfying
assignment for F ′ (and not just a trivially satisfying witness). Then, by the
construction of F ′ and the binding property of the hash function, we have that
Ui = NIFS.V(vk, Ui−1 , ui−1 , T i−1 )

with probability ϵ − negl(λ). Thus, P ei−1 succeeds in producing an accepting

folded instance-witness pair (Ui , Wi ), for instances (Ui−1 , ui−1 ), with probability
ϵ − negl(λ) in expected polynomial-time.

39
 Then, by the knowledge soundness of the underlying non-interactive folding
scheme (Assumption 1) there exists an extractor Eei−1 that outputs (wi−1 , Wi−1 )
such that (ui−1 , wi−1 ) and (Ui−1 , Wi−1 ) satisfy F ′ with probability ϵ − negl(λ)
in expected polynomial-time.
Given an expected polynomial-time P ei−1 and an expected polynomial-time
Ei−1 , E constructs an expected polynomial time Ei−1 as follows
e

Ei−1 (pp; ρ):

(1) ((Ui−1 , ui−1 ), (Ui , Wi ), T i−1 ) ← P ei−1 (pp; ρ)

(2) Retrieve ((zi , . . . , zn−1 ), (ωi , . . . , ωn−1 ), Πi , ri−1 ) from the internal state of
P
ei−1 .
(3) Parse Πi .wi to retrieve zi−1 and ωi−1
(4) Let (wi−1 , Wi−1 ) ← Eei−1 (pp).
(5) Let Πi−1 ← ((Ui−1 , Wi−1 ), (ui−1 , wi−1 ), ri−1 ).
(6) Output ((zi−1 , . . . , zn−1 ), (ωi−1 , . . . , ωn−1 ), Πi−1 ).

We first reason that the output (zi−1 , . . . , zn−1 ), and (ωi−1 , . . . , ωn−1 ) are valid.
By the inductive hypothesis, we already have that for all j ∈ {i + 1, . . . , n},

zj = F (zj−1 , ωj−1 ),

and that V(vk, i, z0 , zi , ((Ui , Wi ), (ui , wi ), ri )) = 1 with probability ϵ − negl(λ).

Because V additionally checks that

ui .x = hash(vk, i, z0 , zi , Ui , ri ) (14)

by the construction of F ′ and the binding property of the hash function, we have

F (zi−1 , ωi−1 ) = zi

with probability ϵ − negl(λ). Next, we argue that Πi−1 is valid. Because (ui , wi )
satisfies F ′ , and (Ui−1 , ui−1 ) were retrieved from wi , by the binding property of
the hash function, and by Equation (14), we have that

ui−1 .x = hash(vk, i − 1, z0 , zi−1 , Ui−1 , ri−1 )

(ui−1 .E, ui−1 .u) = (0, 1)

Additionally, in the case where i = 1, by the base case check of F ′ , we have that
zi−1 = z0 . Because Eei−1 succeeds with probability ϵ − negl(λ), we have that

V(vk, i − 1, z0 , zi−1 , Πi−1 ) = 1

with probability ϵ − negl(λ).

40
D Proof of Theorem 4 (A zkSNARK of a Valid IVC
Proof )

Proof. Completeness and knowledge soundness hold due to the completeness and
knowledge soundness of the underlying zero-knowledge succinct non-interactive ar-
gument of knowledge (zkSNARK) and the non-interactive folding scheme (NIFS).
Succinctness holds due to the succinctness of the commitment scheme underlying
the non-interactive folding scheme, the succinctness of the hash function, and the
succinctness of the underlying non-interactive argument. To prove zero-knowledge,
we first construct a simulator S as follows. Let NIFS.S and zkSNARK.S denote the
simulators for the underlying non-interactive folding scheme and the zkSNARK
respectively. The simulator is given as input the statement proven by an IVC
proof (n, z0 , zn ) along with public parameters pp. For i ∈ {1, . . . , n−1}, let zi = ⊥.

S(pp, (n, z0 , zn ), τ ) → π:

If n = 0,
output ⊥;
otherwise,
(1) for i ∈ {0, . . . , n − 1}
′
(ri+1 , ri+1 ) ←R F
if i = 0
U1 ← u⊥
u1 ← (hash(vk, 1, z0 , z1 , U1 , r1 ), (u⊥ .E, 1, Com(pp, 0, r1′ )))
otherwise
T i ← NIFS.S(ppNIFS , Ui , ui )
Ui+1 ← NIFS.V(vkNIFS , Ui , ui , T i )
′
ui+1 ← (hash(vk, i+1, z0 , zi+1 , Ui+1 , ri+1 ), (u⊥ .E, 1, Com(pp, 0, ri+1 )));
(2) T n ← NIFS.S(ppNIFS , Un , un )
(3) πU′ ← zkSNARK.S(ppzkSNARK , NIFS.V(vkNIFS , Un , un , T n ), τ )
(4) output (Un , un , rn , T n , πU′ ).

If n = 0, both the simulator and the honest prover output ⊥. If n > 0, we

reason that the pair (U, u) is indistinguishable from the pair produced by an
honest prover. We do so by showing that if (Ui , ui ) is indistinguishable then
(Ui+1 , ui+1 ) is indistinguishable.

– Suppose that i = 1. By the definition of IVC.P, U1 = u⊥ in both the real

and simulated settings. Additionally, we have that u1 .E = u⊥ .E and u.u = 1
in both the real and simulated setting. Due to the hiding property of the
hash function, we have that the simulated hash of the public IO, hash(vk, i +
1, z0 , zi+1 , Ui+1 , ri+1 ), is indistinguishable from the one generated honestly.
Due to the hiding property of the commitment scheme, we additionally
′
have that the simulated commitment to the witness, Com(pp, 0, ri+1 ), is
indistinguishable from the one generated honestly.

41
 – For some i ∈ {1, . . . , n − 1}, suppose now that (Ui , ui ) is indistinguish-
able in the real and the simulated settings. Due to the zero-knowledge
property of the underlying non-interactive folding scheme, we have that
T i ← NIFS.S(ppNIFS , Ui , ui ) is indistinguishable in the real and the simu-
lated settings. Thus, we have that Ui+1 ← NIFS.V(vk, Ui , ui , T i ) is indis-
tinguishable in the real and the simulated settings. Then, we have that
hash(vkNIFS , i + 1, z0 , zi+1 , Ui+1 , ri+1 ) is indistinguishable in the real and the
simulated setting. The remaining pieces of ui+1 are indistinguishable in the
real and the simulated setting for the same reasons as the base case. Thus,
we have that (Ui+1 , ui+1 ) is indistinguishable in the real and the simulated
settings.

Furthermore, in both the real and simulated settings, Un and un satisfy the
verifier’s hash check that un .x = hash(vk, n, z0 , zn , Un , rn+1 ). In the simulated
setting, this holds because the simulator uses the provided values of (n, z0 , zn )
when computing un .x.
Now, because (Un , un ) are indistinguishable from that produced by honest
IVC.P, we have that T n produced by NIFS.S(pp, Un , un ) is indistinguishable from
that produced by P due to the zero-knowledge property of the underlying non-
interactive folding scheme. This means that U′ generated by NIFS.V(vk, Un , un , T )
is indistinguishable from that produced honestly. This means that πU′ is indistin-
guishable in the real and the simulated setting due to the zero-knowledge property
of the underlying zkSNARK. Thus, the simulated zkSNARK (Un , un , T n , πU′ ) is
indistinguishable from the one generated honestly.

E Proof of Lemma 1 (Forking Lemma for Folding

Schemes)

We provide a proof for our variant of the forking lemma by adapting the proof of
the forking lemma from Bootle et al. [12].

Lemma. Let F denote a finite field. Consider a (2µ + 1)-move folding scheme
Π = (G, K, P, V). Π satisfies knowledge soundness if there exists a PPT X such
that for all input instance pairs (u1 , u2 ), outputs satisfying witnesses (w1 , w2 )
with probability 1 − negl(λ), given public parameters pp, a structure s, and an
(n1 , . . . , nµ )-tree of accepting transcripts and the corresponding folded instance-
witness pairs (u, w). This tree comprises of n1 transcripts (and the corresponding
instance-witness pairs) with fresh randomness in V’s first message; and for each
such transcript, n2 transcripts (and the corresponding instance-witness Qµ pairs)
with fresh randomness in V’s second message; etc., for a total of i=1 ni leaves
bounded by poly(λ).

Proof. Consider a (2µ + 1)-move folding scheme (G, K, P, V) for relation R with
randomness sampled uniformly from F where |F | ≥ 2λ . Suppose there exists a
PPT X such that for arbitrary input instances (u1 , u2 ), outputs the corresponding
satisfying witnesses (w1 , w2 ) with probability 1 − negl(λ), given public parameters

42
pp, a structure s, and an (n1 , . . . , nµ )-tree of accepting
Qµ transcripts and the
corresponding folded instance-witness pairs where i=1 ni . Consider an arbitrary
expected polynomial-time adversary P ∗ , and pp ← G(1λ ). Let ϵ be the probability
that, given (u1 , u2 ) ← P ∗ (pp, ρ), P ∗ (pp, ρ) succeeds in producing a valid witness w
for the folded instance u. To prove that (G, K, P, V) satisfies knowledge-soundness
we will show that there exists a corresponding extractor E that outputs a valid
witness pair (w1 , w2 ) for instances (u1 , u2 ) given public parameters pp and the
prover randomness ρ with probability ϵ − negl(λ).
We first construct an extractor E as follows:

E(pp, ρ) → (w1 , w2 ):
1. Compute (s, (u1 , u2 )) ← P ∗ (pp, ρ)
2. Compute tree ← T (1), where the function T is defined below.
3. If tree is not a valid (n1 , . . . , nµ )-tree (i.e. there are collisions in the verifier’s
randomness) return ⊥.
4. Output (w1 , w2 ) ← X (pp, s, u1 , u2 , tree).
T (i) → tree:
1. Sample the verifier’s randomness for round i; denote it as ri .
2. If i = µ + 1, compute
(u, w) ← ⟨P ∗ (ρ), V(u1 , u2 )⟩
and let tr be the corresponding transcript. If (u, w) ∈ R, output {(tr, (u, w))}.
Otherwise output ⊥.
3. With fixed verifier randomness (r1 , . . . , ri ), compute tree ← T (i + 1) once. If
tree = ⊥ output ⊥.
4. With fixed verifier randomness (r1 , . . . , ri ), repeatedly run T (i + 1) until
ni − 1 additional lists of accepting transcripts are acquired. Append all the
results to tree.
5. Output tree.
Let E1 denote the event that T (1) outputs tree ̸= ⊥ in less than T time
steps (we specify T later). Given E1 , let E2 denote the event that the resulting tree
is a valid (n1 , . . . , nµ )-tree (i.e. there are no collisions in the verifier’s randomness).
Given E1 and E2 , let E3 denote the event that X succeeds with tree as input.
Then, we have that E succeeds with probability
PE = Pr [E3 ] · Pr [E2 ] · Pr [E1 ].
We now compute each of these probabilities. To compute Pr [E1 ], we observe that
T (1) fails (i.e. returns ⊥) only when its first call to T (2) fails. Likewise, T (2)
fails only when its first call to T (3) fails. Chaining these assertions, we have that
T (1) fails with probability (1 − ϵ), which, by assumption, is the probability that
T (µ + 1) fails. Thus, the expected number of times T (i) calls T (i + 1) is
(ni − 1)
1 + Pr [First call to T (i + 1) succeeds] ·
Pr [T (i + 1) returns tree ̸= ⊥]

43
 (ni − 1)
=1+ϵ·
ϵ
= ni .
Qµ
Hence, the total runtime is expected to be t = O( i=1 ni ) which is bounded
above by poly(λ) by assumption. Moreover, by Markov’s inequality, we have that
T runs for time longer than T > t with probability Tt . Thus, we have that

t
Pr[E1 ] = (1 − ) · ϵ.
T
Given E1 we have that T runs in at most T time-steps. This ensures that there are
at most T random challenges produced for the verifier, and that the probability
2
of collision is at most T|F | . Thus, we have that

T2
Pr[E2 ] = 1 − .
|F |

Finally, given E1 and E2 , we have that Pr[E3 ] = 1 − negl(λ) by assumption.

Consolidating our calculations we have that E succeeds with probability
 T2  t
PE = (1 − negl(λ)) · 1 − · 1− ·ϵ
|F | T
p
Setting T = 3
|F | (and assuming that T ≥ t) we have that
 1   t 
PE = (1 − negl(λ)) · 1 − p · 1 − p ·ϵ
3
|F | 3
|F |
= ϵ − negl(λ)

F A Commitment Scheme for Vectors

We first recall the syntax for commitment schemes for vectors introduced in
Section 2.1 and then define the required binding and hiding properties. Next,
we define the additionally required additively homomorphic and succinctness
properties. For the sake of concreteness, we reproduce Pedersen’s commitment
scheme, which satisfies all of the required properties.

Definition 16 (Commitment Scheme). A commitment scheme for F m is a

tuple of three protocols with the following syntax that satisfy the two properties
listed below:
– pp ← Gen(1λ , m): produces public parameters pp.
– C ← Com(pp, x, r): takes as input x ∈ F m and r ∈ F ; produces a public
commitment C.

44
 – b ← Open(pp, C, x, r): verifies the opening of commitment C to x ∈ F m and
r ∈ F ; outputs b ∈ {0, 1}.

(1) Binding. For any PPT adversary A, the following probability is negl(λ):

pp ← Gen(1λ , m),
 
 b0 = b1 = 1, (C, x0 ∈ F m , x1 ∈ F m , r0 ∈ F , r1 ∈ F ) ← A(pp), 
Pr 
 x0 ̸= x1

b0 ← Open(pp, C, x0 , r0 ), 
b1 ← Open(pp, C, x1 , r1 )

(2) Hiding. For all PPT adversaries A = (A0 , A1 ), the following probability is
negl(λ):  
(x0 , x1 , st) ← A0 (pp),
1  b ←R {0, 1}, r ←R F , 
− Pr 
 b = b̄ C ← Com(pp, xb , r), 

2
b̄ ← A1 (st, C)
If hiding holds for all adversaries, then the commitment is statistically hiding.

Definition 17 (Additively Homomorphic). A commitment scheme for vec-

tors over F m , (Gen, Com, Open), is additively homomorphic if for all public pa-
rameters pp produced from Gen(1λ , m), and for any x1 , x2 ∈ F m and for any
r1 , r2 ∈ F , Com(pp, x1 , r1 ) + Com(pp, x2 , r2 ) = Com(pp, x1 + x2 , r1 + r2 ).

Definition 18 (Succinctness). A commitment scheme for vectors over F m ,

(Gen, Com, Open), provides succinct commitments if for all public parameters
pp produced from Gen(1λ , m), and any x ∈ F m and r ∈ F , |Com(pp, x, r)| =
Oλ (polylog(|x|)).

Construction 6 (Pedersen Commitment). For group G such that |G| ≥ 2λ

with scalar field F and where the discrete logarithm problem is hard, the Pedersen
commitment scheme for vectors over F is defined as follows:

– pp ← Gen(1λ , m): Sample g ←R Gm , h ←RQG. Output (g, h).

– C ← Com(pp, x ∈ F m , r ∈ F ): Output hr · i∈{0,...,m} gixi .

Lemma 11 (Pedersen Commitment). The Pedersen commitment scheme is

hiding, binding, additively homomorphic, and succinct.

G Background on Polynomials

This section is adapted from prior work [37, 41, 46]. We recall a few basic facts
about polynomials.

– A polynomial g over F is an expression consisting of a sum of monomials

where each monomial is the product of a constant and powers of one or more
variables; all arithmetic is performed over F .

45
 – The degree of a monomial is the sum of the exponents of variables in the
monomial; the (total) degree of a polynomial g is the maximum degree of
any monomial in g. Furthermore, the degree of a polynomial g in a particular
variable xi is the maximum exponent that xi takes in any g’s monomials.
– A multivariate polynomial is a polynomial with more than one variable. A
multivariate polynomial is called a multilinear polynomial if the degree of
the polynomial in each variable is at most one.
– A multivariate polynomial g over a finite field F is called low-degree if the
degree of g in each variable is bounded above by a constant.

H An Optimization to Theorem 6
In Theorem 6, given that the verifier’s runtime to verify an evaluation is at least
linear in the polynomial size, the verifier does not gain by querying commitments
to structure polynomials A,e B,
e C.
e Rather, the verifier can directly evaluate them,
which not only simplifies the protocol (as one does not need Sparse-PCBP ), but
also improves costs, both asymptotically and concretely.

Corollary 2. Assuming the hardness of the discrete logarithm problem, there

exists a non-interactive zero-knowledge argument of knowledge for committed
relaxed R1CS with the following efficiency characteristics, where m denotes the
dimensions of the R1CS coefficient matrices and n denotes the number of non-zero
entries in the matrices: The encoder runs in time Oλ (m); The prover runs in
time O(n) + Oλ (m); The proof length is Oλ (log m); and the verifier runs in time
n + Oλ (m).

46