OP_PAIRCOMMIT

[moonsettler]

OP_PAIRCOMMIT is the newest member of the LNhance family of opcodes. It provides limited vector commitment functionality in tapscript.

When evaluated, the OP_PAIRCOMMIT instruction:

• pops the top two values off the stack,
• takes the "PairCommit" tagged SHA256 hash of the stack elements,
• pushes the resulting commitment on the top of the stack.

Discussion: https://delvingbitcoin.org/t/op-paircommit-as-a-candidate-for-addition-to-lnhance/1216/12

[murchandamus]

This document has a few formatting issues, please make sure that the preamble matches the BIP 2 requirements and take a look at the rich diff to see whether it looks the way you intend.

Please note that the BIPs repository also accepts markdown files.

[moonsettler]

Switched back to markdown. Header now in BIP-2 format.

[moonsettler]

The original create date of OP_PAIRCOMMIT is 2024-03-15 this is the latest revision based on feedback from Anthony Towns.
https://gist.github.com/moonsettler/d7f1fb88e3e54ee7ecb6d69ff126433b/revisions
What date should go to the header?

[jonatack]

Added a discussion link to the PR description.

The original create date of OP_PAIRCOMMIT is 2024-03-15 this is the latest revision based on feedback from Anthony Towns.
gist.github.com/moonsettler/d7f1fb88e3e54ee7ecb6d69ff126433b/revisions
What date should go to the header?

Perhaps add a changelog with the revision based on Anthony Towns' feedback followed by the initial version. Or use the date of the current draft revision as your starting point.

[murchandamus]

According to BIP 2:

The Created header records the date that the BIP was assigned a number, […]

[murchandamus]

Has this proposal been sent to the mailing list?

[moonsettler]

Has this proposal been sent to the mailing list?

Not yet. Wanted to get it into an acceptable shape before I post it there.

Proposed to the mailing list, waiting for feedback.

[murchandamus]

I would like to see this proposal to get more review from other covenant researchers before it moves forward.

[murchandamus]

This paragraph seems to indicate that the OP_PAIRCOMMIT proposal would be especially useful in combination with these other opcodes. Could you perhaps clarify whether and how OP_PAIRCOMMIT is useful by itself in absence of the other three opcodes you mention here?

[moonsettler]

LNhance at it's core is CTV + CSFS. They together provide the core utility. IKEY is an optimization for not having to pay for the pubkey twice when the internal key can be used. PC is an optimization when CSFS has to commit to additional data required to recreate a spend script from an intermediate state, because OP_RETURN (to which CTV naturally commits to) is 4x more expensive in weight units for data availability.

PC could also be used by CHECKCONTRACTVERIFY to carry a complex state in the absence of CAT.

I don't think anyone would find PC useful enough to activate in isolation without the aforementioned other opcodes.
It can do general merkle tree style commitments that are not compatible with other merkle tree structures in bitcoin.

We probably will make a new BIP for LNhance that has these other BIPs as "Relies on".

[murchandamus]

I meant that these explanations should be part of the BIP, not just part of the conversation here in the pull request comments.

[moonsettler]

So should I link the "Use in LN-Symmetry" section from there? Or it needs better explanation?

[moonsettler]

Really leaning towards making a head BIP for LNhance and keeping the individual BIPs strictly limited to describing functionality without a lot of speculation on applications. (Which we would do after finalizing the individual ops.)

Is there a problem with this approach?

[murchandamus]

If you are creating multiple BIPs that only make sense together, it would be better to propose them as a single pull request. Since this is being proposed standalone, it should also provide its own raison d’être.

[moonsettler]

Maybe I should write that it was an explicit design goal that PAIRCOMMIT is pretty much completely useless on it's own with the current set of opcodes?

[rot13maxi]

what data needs to be available? how does PC solve that problem (does it stick the data in the witness and put a commitment somewhere covered by a signature? something else?)? Is this mechanism useful for things outside of LN-Symmetry?

[moonsettler]

The data that needs to be available for state n is:

state-n-recovery-data { settlement-n-hash or state-n-balance }

This is needed to reconstruct the whole script for the nth state address that the funds move to by the channel peer that only holds the latest state, so he can spend to the latest state.

edit:
Instead of an IF statement we could use different tap leaves (less optimal actually) and then merkle inclusion proof with sibling hashes would have to be known.

[moonsettler]

Is this mechanism useful for things outside of LN-Symmetry?

It was obviously our primary motivation, but I would not be surprised if other applications that use CSFS find a similar use for it.

[moonsettler]

One way to think about the 3 opcodes (CSFS, IKEY, PC) is we decompose a CSFS variant that can use 1 byte pubkey (internal key) and can commit to a vector of stack elements as message. They thus become more generally useful, but to a limited degree without additional opcodes.

Detailed introspection opcodes would also need vector commitments with CSFS, and as mentioned it would also be useful for CCV.

[moonsettler]

Updated. I believe we can resolve these conversations? @rot13maxi @murchandamus

[moonsettler]

It looks like we gonna have to amend the PAIRCOMMIT BIP with some new use cases.
Turns out within certain practical limitations any computational function can be proven out in the form of a merkle tree.
The root hash of the merkle tree represents the function the leaves represent the inputs and output.
Any 32 bit arithmetic function can certainly be proven out with this method.
CAT itself with a limited set of inputs or limited input sizes can be proven out.
At this point it's an open question if this enables new behaviors not enabled by taproot MAST itself?

Special thanks to: @JeremyRubin @Ademan @bigspider

edit:
Alternatively could consider imposing specific script limits that make PAIRCOMMIT explicitly less capable than MAST itself.

[Ademan]

I think I've changed my mind a bit. We were talking about computing a merkle tree for f(u32,u32) as if it was trivial but after a quick experiment it seems like that would take hundreds of years to compute (am I being dumb here?) Instead, you can compute mul(u32,u32) -> u32 using 3 mul(u16,u16)s which is feasible to compute. The witness size is worse, ~32 * 32 * 3 = 3072 instead of 32 * 64 * 1 = 2048, but computing the tree for mul(u16,u16) is feasible using a naive algorithm on commodity hardware.

The implication of this is that where a function can be decomposed into operations on smaller inputs, PAIRCOMMIT is massively more feasible to use than encoding things into a tap tree.

[bigspider]

I think I've changed my mind a bit. We were talking about computing a merkle tree for f(u32,u32) as if it was trivial but after a quick experiment it seems like that would take hundreds of years to compute (am I being dumb here?) Instead, you can compute mul(u32,u32) -> u32 using 3 mul(u16,u16)s which is feasible to compute. The witness size is worse, ~32 * 32 * 3 = 3072 instead of 32 * 64 * 1 = 2048, but computing the tree for mul(u16,u16) is feasible using a naive algorithm on commodity hardware.

Arithmetic and bitwise operations where inputs & outputs are small enough, can already be done in Script in cheaper ways. Merkle trees as lookup tables are only interesting for functions that are either extremely complex, or where preimages/images are larger than what Script can work with.
Note that you can already do small indexed lookup tables more efficiently by just hard-coding them in Script (that is: push the table on the stack and use OP_PICK to read its entries), and these techniques are widely used (e.g. in BitVM).

The implication of this is that where a function can be decomposed into operations on smaller inputs, PAIRCOMMIT is massively more feasible to use than encoding things into a tap tree.

I think the only substantial difference is that in a Script where you need several lookups, you can do it with Merkle trees, while you can only do a single lookup with a precomputed taptree.

[moonsettler]

Proving general computation

Merkle trees can be used to prove out computation where the root of the tree
represents the function and the leaves represent the inputs and output.
There are practical limits to the entropy space for the inputs as it needs
to be iterated over and hashed up.

Currently MAST trees can cover 128 bits of entropy space, which is well over
the practical limits to iterate over and merklize. Therefore we assume this
capability does not materially extend what computations are possible to prove
out in bitcoin script. While OP_PAIRCOMMIT is not limited to a height of 128,
that should not be practically feasible to utilize.

There is a way to reduce the size of the witness for proving out computation,
by eliminating the merkle path inclusion proofs, using OP_CHECKSIGFROMSTACK
together with OP_PAIRCOMMIT. This method involves deleted key assumptions,
most likely using MPC to create an enormous amount of signatures for the stack
elements representing the inputs and the output of the function.

Is this correct? Any suggestions? @Ademan @bigspider

[moonsettler]

The implication of this is that where a function can be decomposed into operations on smaller inputs, PAIRCOMMIT is massively more feasible to use than encoding things into a tap tree.

This is the main open question I believe. does it or does it not practically expand what we can already do?
For example using PC to emulate smolCAT and using traditional methods with lookup tables could make 32 bit or even 64 bit arithmetics more feasible?

edit:
Within the 32 bit realm we can already use OP_ADD, I see little practical diff between <0x1234> <0x5678> CAT and <0x12340000> <0x5678> ADD.
And it sounds like 64 bit smolCAT would be way too expensive to generate (and also to interact with trustlessly).

(actually the above examples are wrong, because internally bitcoin script uses little endian, but should convey the point)

[Ademan]

...

Arithmetic and bitwise operations where inputs & outputs are small enough, can already be done in Script in cheaper ways. Merkle trees as lookup tables are only interesting for functions that are either extremely complex, or where preimages/images are larger than what Script can work with. Note that you can already do small indexed lookup tables more efficiently by just hard-coding them in Script (that is: push the table on the stack and use OP_PICK to read its entries), and these techniques are widely used (e.g. in BitVM).

Even u16,u16 is quite a bit larger than I think is practical as a lookup table, but the efficiency for repeated operations is constant, obviously. The lookup table is less efficient for small numbers of operations (a u8,u8 table is 16k vs 1 u8,u8 proof is 0.4k) but the merkle tree loses quickly when those operations are repeated.

The implication of this is that where a function can be decomposed into operations on smaller inputs, PAIRCOMMIT is massively more feasible to use than encoding things into a tap tree.

I think the only substantial difference is that in a Script where you need several lookups, you can do it with Merkle trees, while you can only do a single lookup with a precomputed taptree.

Right, and the key point is these merkle trees and lookup tables rapidly become infeasible to compute as the input size grows, so multiple smaller lookups is significantly more useful.

EDIT: But your point is well taken that for smaller operations they can already be better accomplished by lookup tables.

[Ademan]

...
edit: Within the 32 bit realm we can already use OP_ADD, I see little practical diff between <0x1234> <0x5678> CAT and <0x12340000> <0x5678> ADD. And it sounds like 64 bit smolCAT would be way too expensive to generate (and also to interact with trustlessly).

(actually the above examples are wrong, because internally bitcoin script uses little endian, but should convey the point)

Yeah for arbitrary 8 byte strings smolCAT seems infeasible to compute the table or merkle tree for. After a bit of conversation on IRC it could probably be feasible for arbitrary f(b[4],b[4]) -> b[8] with a custom ASIC¹ or maybe a cluster of FPGAs in a span of ~a few years but that would not be very useful for the average person.

Bit shifts over 32 bit integers seems pretty feasible though, that's f(u32,u6)->u32 (maybe save some space by special casing shift = 0). it seems like my incredibly naive, unoptimized, single-core experiment could calculate that merkle tree in ~96 hours. Of course the proof is ~1.2k and users would likely need multiple, but the lookup table for that wouldn't fit in a block anyway so maybe something new is possible?

You can also separate positive and negative shifts, and maybe break it down into multiple rounds of shifts 1-3 or something (or 1k for a proof for a constant shift)

[1]: afaik existing ASICs operate on block headers so couldn't help

[moonsettler]

I think this BIP is already way more verbose than it was supposed to be.

It would be useful if we could reference it by a number.
Can we get a BIP number assigned? Not asking for a merge yet.