OP_ZKP updates
5 messages, 22 Jul to 14 Oct 2024. Started by Weiji Guo.
Weiji Guo (4), Weikeng Chen
Greetings, bitcoin developers. I am writing to update you about the OP_ZKP
proposal I initiated last year. This time, it concerns which ZKP scheme to use in
the underlying proving system. This email will contain the following four parts:
1, background, mainly about the initial proposal. You may skip it if you already
know about it.
2, high-level requirements for the ZKP scheme and the current priority regarding
selection. Also, a brief explanation of precluding trusted setup and FRI-based
schemes.
3, ideas and open issues regarding Inner Product Argument.
4, what-ifs
I should have studied all these further before sending this email, but I also want to
have something to talk about during Bitcoin Nashville. During the conf, you may
find me in the K14 booth for f2f talks.
———Background———
For those who haven't heard about this idea, here is the link to the earlier email
thread: https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2023-April/021592.html
Before proposing any BIP, we must explore existing ZKP schemes and discuss
how to use them with OP_ZKP within applications. That's why I took a detour to
develop potential applications to understand further how OP_ZKP could work in
real-world applications. That work concluded about two months ago; since then, I
have been working on OP_ZKP.
———High level requirements———
1, Security. The security assumptions should be minimal. Ideally, only the ECDLP
assumption is taken, leading directly to the Inner Product Argument over
secp256k1. However, an open issue still blocks IPA from working in OP_ZKP
(details will follow soon). We might consider pairing in a transparent setup setting,
but no trusted setup.
2, Small block size consumption. The proof should be both succinct and
concretely small. Being concretely small allows individual or small numbers of
proofs to be posted on-chain without incurring too many costs. Otherwise, they
will have to wait to be verified in a batch, should the scheme allow such. Arguably,
waiting for batching is not a good idea, as the transactions will have to be put on
hold, and there is no guarantee more will come. That said, we might revisit this
choice should we run out of candidate schemes.
FRI-based proofs are considered too big for 100- to 128-bit provable security. The
size is a few hundred kilobytes for a circuit of 2^20 size. Besides, it does not allow
batched verification (see the following requirement).
3, Batched verification is mandatory due to block size considerations. Should the
situation allow, we could replace the proof data (as transaction witnesses) of
thousands of OP_ZKP transactions with a single argument to save block space
(and transaction costs).
It also saves verification time, even without the benefits of saving block space.
4, Small verification key for deployment. It seems natural but is not the default
case for some schemes.
5, Aggregated proving. This is optional as it happens off-chain. However, it
effectively reduces proof size and verification time requirements for some non-
constant size proof schemes (we precluded trusted setup). Consider aggregating
many sub-circuits together with a constant or log-sized argument.
2^16 is a reasonable upper bound for a sub-circuit. Computing a block hash
takes about 60k constraints (Sha256 applied twice against an 80-byte block
header).
But 2^16 is still too large for FRI-based proof. Each FRI-query should cost 16^2
hashes, which is 8 kilobytes. There are dozens of queries to run to meet the target
security (preferably 128-bit, without further security conjectures). Interested
readers may refer to https://a16zcrypto.com/posts/article/17-misconceptions-
about-snarks/#section--11.
———Inner Product Argument———
Now, let's consider IPA-based solutions. IPA has a transparent setup, requires
only ECDLP assumption, can work on the secp256k1 curve, has a relatively small
proof size, and is capable of both batched verification and aggregated proving.
We can use the aggregated proving technique to address the issue of linear time
verification. The remaining open issue with IPA is that the size of the verification
key is linear to the circuit size.
Let me expand on the last two issues.
The linear verification time of IPA comes from the need to recompute the
Pedersen commitment base in each round of reduction (could be postponed but
still O(N)). We could adopt the aggregated proving technique to address this
issue and effectively reduce the complexity of the actual verification time.
Suppose there are M sub-circuits; each has a size bound of N (assuming N =
2^16). We could have an argument to combine the M inner products into one. This
argument is also made of IPA, thus having O(log(M)) size and O(M) verification
time. Then the total proof size is O(log(M) + log(N)), and verification time
becomes O(M)+O(N) rather than O(M*N) (there will be some extra costs per
aggregated proof, but let's skip it for now). This is how we could use IPA to
achieve succinctness and to deal with huge circuits (we need this as we might
recursively verify proofs of other schemes).
Linear verification key comes from the need for the verifier to use all the circuit
constants. It is at least accurate for BP, BP+, and even Halo, and it used to be
acceptable as the multi-scalar multiplication dominates the verification costs.
However, it becomes an issue with aggregated proving in terms of both
deployment size and verification time. When M is large enough, the aggregated
verification time to compute with these constants is non-trivial, even compared to
the MSM costs. The more significant issue is with the circuit deployment. We have
to save all these parameters on-chain.
To further expand on this, let me re-iterate the Sonic Arithmetization process. It is
R1CS-based with some pre-processing. Suppose there are N multiplication gates
such that:
A_i * B_i = C_i
for i in [0 - N-1], A_i, B_i, and C_i as field elements, and A, B, and C are circuit
witnesses of N-element vectors.
There are also Q linear constraints to capture the addition gates, circuit wiring,
and multiplied-by-constant constraints:
WL*A + WR*B + WO*C = K
where WL, WR, and WO are Q*N matrixes of field elements, and K is a Q-element
vector of field elements. WL, WR, WO, and K are circuit constants.
Although WL, WR, and WO are vastly sparse, they are at least O(N) size. The
verifier will have to compute on them during verification after generating a
challenge value in response to the prover's commitment to the witnesses. This
O(N) size prevents us from deploying verification keys on-chain.
The natural idea is to commit to those constants somehow and to offload the
commitment opening to the prover. Although the verifier still pays O(N) time to
verify, the deployment size is constant, and the opening size is only O(log(N))
instead of O(N), assuming an IPA opening. However, this no longer holds with
aggregated proving, as there is no guarantee that the aggregated constants
aWL, aWR, and aWO will be sparse. The complexity could become O(N^2).
The open issue here is to find a way to commit to WL, WR, and WO such that 1)
the verification key could be tiny, 2) the proof size to open the commitment is
acceptable, and 3) the commitments could be aggregated. If necessary, modify
the arithmetization further, ideally without introducing new security assumptions.
-- If we do, for example, introduce SXDH, then we can consider schemes like
Dory, which has a transparent setup and O(log(N)) verifier time.
———What-ifs———
What if the above open issue could be resolved? The resulting proving system
should work even for a Raspberry Pi 4 node. I ran some benchmarks on Raspberry
Pi 4 for some elliptical curve operations. The general conclusion is that it is about
ten times slower than my Mac Book Pro for a single thread. For a circuit of 2^16
size, the verification time with MBP is doubt 200~300 ms (inferred by
benchmarking MSM of this size); therefore, it would be about 2~3 seconds for
RP4. The aggregation might double or triple the time cost.
Now, for block verification, counted in batched verification, it should be
acceptable for RP4 to spend around 10 seconds verifying ALL OP_ZKP
transactions in a new block. Luckily, we won't have too many existing blocks full
of OP_ZKP transactions any time soon.
For transaction forwarding, a simple technique of pool-then-batched verification
could work for RP4. As its name suggests, an RP4 node could simply pool any
incoming OP_ZKP transactions in memory when it is already busy verifying some.
When it is done, the node retrieves all the pooled OP_ZKP transactions and
verifies them in a batch. This way, it only adds a few seconds of latency to
forwarding any OP_ZKP transactions.
What if the open issue cannot be resolved? We might consider Dory. It is
transparent, requires pairing, and has logarithmic proof size but concretely larger
than Bulletproofs (but Torus-based optimization might help here by a factor of 3;
check out here: https://youtu.be/i-uap69_1uw?t=4044 and https://eprint.iacr.org/
2007/429.pdf ). It also has logarithmic verification time, and batched verification
allows for saving block space and verification time. The community might need to
accept the SXDH assumption.
That's it. Thanks for reading. Any comments are welcome.
And again, see you in the K14 booth, Nashville.
Regards,
Weiji
I need to point out that Dory requires pairing, and therefore it cannot work with secp256k1? Please circle back. On Monday, July 22, 2024 at 9:16:18 AM UTC-5 Weiji Guo wrote:
Quoted text (398 lines)
> Greetings, bitcoin developers. I am writing to update you about the OP_ZKP > > proposal I initiated last year. This time, it concerns which ZKP scheme to > use in > > the underlying proving system. This email will contain the following four > parts: > > > 1, background, mainly about the initial proposal. You may skip it if you > already > > know about it. > > 2, high-level requirements for the ZKP scheme and the current priority > regarding > > selection. Also, a brief explanation of precluding trusted setup and > FRI-based > > schemes. > > 3, ideas and open issues regarding Inner Product Argument. > > 4, what-ifs > > > I should have studied all these further before sending this email, but I > also want to > > have something to talk about during Bitcoin Nashville. During the conf, > you may > > find me in the K14 booth for f2f talks. > > > ———Background——— > > > For those who haven't heard about this idea, here is the link to the > earlier email > > thread: > https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2023-April/021592.html > > > Before proposing any BIP, we must explore existing ZKP schemes and discuss > > how to use them with OP_ZKP within applications. That's why I took a > detour to > > develop potential applications to understand further how OP_ZKP could work > in > > real-world applications. That work concluded about two months ago; since > then, I > > have been working on OP_ZKP. > > > ———High level requirements——— > > > 1, Security. The security assumptions should be minimal. Ideally, only the > ECDLP > > assumption is taken, leading directly to the Inner Product Argument over > > secp256k1. However, an open issue still blocks IPA from working in OP_ZKP > > (details will follow soon). We might consider pairing in a transparent > setup setting, > > but no trusted setup. > > > 2, Small block size consumption. The proof should be both succinct and > > concretely small. Being concretely small allows individual or small > numbers of > > proofs to be posted on-chain without incurring too many costs. Otherwise, > they > > will have to wait to be verified in a batch, should the scheme allow such. > Arguably, > > waiting for batching is not a good idea, as the transactions will have to > be put on > > hold, and there is no guarantee more will come. That said, we might > revisit this > > choice should we run out of candidate schemes. > > > FRI-based proofs are considered too big for 100- to 128-bit provable > security. The > > size is a few hundred kilobytes for a circuit of 2^20 size. Besides, it > does not allow > > batched verification (see the following requirement). > > > 3, Batched verification is mandatory due to block size considerations. > Should the > > situation allow, we could replace the proof data (as transaction > witnesses) of > > thousands of OP_ZKP transactions with a single argument to save block > space > > (and transaction costs). > > > It also saves verification time, even without the benefits of saving block > space. > > > 4, Small verification key for deployment. It seems natural but is not the > default > > case for some schemes. > > > 5, Aggregated proving. This is optional as it happens off-chain. However, > it > > effectively reduces proof size and verification time requirements for some > non- > > constant size proof schemes (we precluded trusted setup). Consider > aggregating > > many sub-circuits together with a constant or log-sized argument. > > > 2^16 is a reasonable upper bound for a sub-circuit. Computing a block hash > > takes about 60k constraints (Sha256 applied twice against an 80-byte block > > header). > > > But 2^16 is still too large for FRI-based proof. Each FRI-query should > cost 16^2 > > hashes, which is 8 kilobytes. There are dozens of queries to run to meet > the target > > security (preferably 128-bit, without further security conjectures). > Interested > > readers may refer to > https://a16zcrypto.com/posts/article/17-misconceptions- > > about-snarks/#section--11. > > > ———Inner Product Argument——— > > > Now, let's consider IPA-based solutions. IPA has a transparent setup, > requires > > only ECDLP assumption, can work on the secp256k1 curve, has a relatively > small > > proof size, and is capable of both batched verification and aggregated > proving. > > We can use the aggregated proving technique to address the issue of linear > time > > verification. The remaining open issue with IPA is that the size of the > verification > > key is linear to the circuit size. > > > Let me expand on the last two issues. > > > The linear verification time of IPA comes from the need to recompute the > > Pedersen commitment base in each round of reduction (could be postponed > but > > still O(N)). We could adopt the aggregated proving technique to address > this > > issue and effectively reduce the complexity of the actual verification > time. > > Suppose there are M sub-circuits; each has a size bound of N (assuming N = > > 2^16). We could have an argument to combine the M inner products into one. > This > > argument is also made of IPA, thus having O(log(M)) size and O(M) > verification > > time. Then the total proof size is O(log(M) + log(N)), and verification > time > > becomes O(M)+O(N) rather than O(M*N) (there will be some extra costs per > > aggregated proof, but let's skip it for now). This is how we could use IPA > to > > achieve succinctness and to deal with huge circuits (we need this as we > might > > recursively verify proofs of other schemes). > > > Linear verification key comes from the need for the verifier to use all > the circuit > > constants. It is at least accurate for BP, BP+, and even Halo, and it used > to be > > acceptable as the multi-scalar multiplication dominates the verification > costs. > > However, it becomes an issue with aggregated proving in terms of both > > deployment size and verification time. When M is large enough, the > aggregated > > verification time to compute with these constants is non-trivial, even > compared to > > the MSM costs. The more significant issue is with the circuit deployment. > We have > > to save all these parameters on-chain. > > > To further expand on this, let me re-iterate the Sonic Arithmetization > process. It is > > R1CS-based with some pre-processing. Suppose there are N multiplication > gates > > such that: > > A_i * B_i = C_i > > for i in [0 - N-1], A_i, B_i, and C_i as field elements, and A, B, and C > are circuit > > witnesses of N-element vectors. > > > There are also Q linear constraints to capture the addition gates, circuit > wiring, > > and multiplied-by-constant constraints: > > WL*A + WR*B + WO*C = K > > where WL, WR, and WO are Q*N matrixes of field elements, and K is a > Q-element > > vector of field elements. WL, WR, WO, and K are circuit constants. > > > Although WL, WR, and WO are vastly sparse, they are at least O(N) size. > The > > verifier will have to compute on them during verification after generating > a > > challenge value in response to the prover's commitment to the witnesses. > This > > O(N) size prevents us from deploying verification keys on-chain. > > > The natural idea is to commit to those constants somehow and to offload > the > > commitment opening to the prover. Although the verifier still pays O(N) > time to > > verify, the deployment size is constant, and the opening size is only > O(log(N)) > > instead of O(N), assuming an IPA opening. However, this no longer holds > with > > aggregated proving, as there is no guarantee that the aggregated constants > > aWL, aWR, and aWO will be sparse. The complexity could become O(N^2). > > > The open issue here is to find a way to commit to WL, WR, and WO such that > 1) > > the verification key could be tiny, 2) the proof size to open the > commitment is > > acceptable, and 3) the commitments could be aggregated. If necessary, > modify > > the arithmetization further, ideally without introducing new security > assumptions. > > -- If we do, for example, introduce SXDH, then we can consider schemes > like > > Dory, which has a transparent setup and O(log(N)) verifier time. > > > ———What-ifs——— > > > What if the above open issue could be resolved? The resulting proving > system > > should work even for a Raspberry Pi 4 node. I ran some benchmarks on > Raspberry > > Pi 4 for some elliptical curve operations. The general conclusion is that > it is about > > ten times slower than my Mac Book Pro for a single thread. For a circuit > of 2^16 > > size, the verification time with MBP is doubt 200~300 ms (inferred by > > benchmarking MSM of this size); therefore, it would be about 2~3 seconds > for > > RP4. The aggregation might double or triple the time cost. > > > Now, for block verification, counted in batched verification, it should be > > acceptable for RP4 to spend around 10 seconds verifying ALL OP_ZKP > > transactions in a new block. Luckily, we won't have too many existing > blocks full > > of OP_ZKP transactions any time soon. > > > For transaction forwarding, a simple technique of pool-then-batched > verification > > could work for RP4. As its name suggests, an RP4 node could simply pool > any > > incoming OP_ZKP transactions in memory when it is already busy verifying > some. > > When it is done, the node retrieves all the pooled OP_ZKP transactions and > > verifies them in a batch. This way, it only adds a few seconds of latency > to > > forwarding any OP_ZKP transactions. > > > What if the open issue cannot be resolved? We might consider Dory. It is > > transparent, requires pairing, and has logarithmic proof size but > concretely larger > > than Bulletproofs (but Torus-based optimization might help here by a > factor of 3; > > check out here: https://youtu.be/i-uap69_1uw?t=4044 and > https://eprint.iacr.org/ > > 2007/429.pdf ). It also has logarithmic verification time, and batched > verification > > allows for saving block space and verification time. The community might > need to > > accept the SXDH assumption. > > > That's it. Thanks for reading. Any comments are welcome. > > And again, see you in the K14 booth, Nashville. > > > Regards, > > Weiji > > >
Yes, that's true. With Dory we will have to work on some pairing-friendly curve. Not secp256k1.
On Tuesday, July 23, 2024 at 3:01:59 AM UTC+8 Weikeng Chen wrote:
I need to point out that Dory requires pairing, and therefore it cannot work with secp256k1? Please circle back. On Monday, July 22, 2024 at 9:16:18 AM UTC-5 Weiji Guo wrote:
———What-ifs———
What if the open issue cannot be resolved? We might consider Dory. It is
transparent, requires pairing, and has logarithmic proof size but concretely larger
I believe I have found the solution to the open issue mentioned in the earlier email. It is just recursive verification. Instead of publishing each application circuit's verification key on-chain, we should have only one circuit that OP_ZKP will verify, which is a recursive verifier.
Interested readers are welcome to visit the GitHub org dedicated for OP_ZKP: https://github.com/opzkp
So far I have just put up the high level ideas here: https://github.com/opzkp/tea-horse. There are nothing else yet. But we will add stuff as we move on.
Regards, Weiji
On Tuesday, July 23, 2024 at 8:40:08 AM UTC+8 Weiji Guo wrote:
Quoted text (19 lines)
> Yes, that's true. With Dory we will have to work on some pairing-friendly > curve. Not secp256k1. > > On Tuesday, July 23, 2024 at 3:01:59 AM UTC+8 Weikeng Chen wrote: > > I need to point out that Dory requires pairing, and therefore it cannot > work with secp256k1? > Please circle back. > > On Monday, July 22, 2024 at 9:16:18 AM UTC-5 Weiji Guo wrote: > > ———What-ifs——— > > What if the open issue cannot be resolved? We might consider Dory. It is > > transparent, requires pairing, and has logarithmic proof size but > concretely larger > >
One more update before implementations:
1, Combine zkStark and aggregated IPA. Application circuits will be developed and proved in skStark, while the recursive verifier will be developed in aggregated IPA. This way we may achieve 6~7 KB proof size and around 1 second verification time for each OP_ZKP transaction.
2, There will be a threshold value T. That is, any bitcoin block may contain at most T OP_ZKP transactions, otherwise small miner such as Raspberry Pi 4 won't be able to verify the block quick enough, even as we implement the batched verification for aIPA.
3, Surpassing count T, we may have a block prove to recursively verify the proof of all the OP_ZKP transactions, generating a new proof for miners to verify. This time we are NOT limited to schemes with batched verification. Therefore we may consider for example, zkStark again, so that the verification will be very fast and the proof size is still acceptable.
We maintain a live doc in GH: https://github.com/opzkp/tea-horse Please let me know if you have any comments.
Regards, Weiji On Wednesday, August 28, 2024 at 11:35:55 PM UTC+8 Weiji Guo wrote:
Quoted text (37 lines)
> I believe I have found the solution to the open issue mentioned in the > earlier email. It is just recursive > verification. Instead of publishing each application circuit's > verification key on-chain, we should have > only one circuit that OP_ZKP will verify, which is a recursive verifier. > > Interested readers are welcome to visit the GitHub org dedicated for > OP_ZKP: https://github.com/opzkp > > So far I have just put up the high level ideas here: > https://github.com/opzkp/tea-horse. There are nothing > else yet. But we will add stuff as we move on. > > Regards, > Weiji > > On Tuesday, July 23, 2024 at 8:40:08 AM UTC+8 Weiji Guo wrote: > >> Yes, that's true. With Dory we will have to work on some pairing-friendly >> curve. Not secp256k1. >> >> On Tuesday, July 23, 2024 at 3:01:59 AM UTC+8 Weikeng Chen wrote: >> >> I need to point out that Dory requires pairing, and therefore it cannot >> work with secp256k1? >> Please circle back. >> >> On Monday, July 22, 2024 at 9:16:18 AM UTC-5 Weiji Guo wrote: >> >> ———What-ifs——— >> >> What if the open issue cannot be resolved? We might consider Dory. It is >> >> transparent, requires pairing, and has logarithmic proof size but >> concretely larger >> >>