extension-bolt: simple taproot channels (feature 80/81)

[Roasbeef]

This PR puts forth two concepts:

1. The concept of an "extension BOLT", which is a single documentation extension to the base BOLT spec that references existing "base" BOLTs with new slightly modified functionality. This presents an alternative to littering the main set of BOLTs with a series of "if" statements, which can be somewhat unwieldy for larger changes, and also harder to review/parse.
2. A new set set of feature bit, channel type, funding output changes, commitment changes, and HTLC script changes under the umbrella of "simple taproot channels", a.k.a the minimal amount of changes needed to get us to "taprooty level 1".

The extensions described in this document have purposefully excluded any gossip related changes, as the there doesn't yet appear to be a predominant direction we'd all like to head in (nu nu gossip vs kick the can and add schnorr).

Most of the changes here described are pretty routine: use musig2 when relevant, and create simple tapscript trees to fold in areas where the script has multiple conditional paths. The main consideration with musig2 is ofc: how to handle nonces. This document takes a very conservative stance, and simply proposes that all nonces be 100% ephemeral, and forgotten, even after a connection has been dropped. This has some non-obvious implications w.r.t the retransmission flow. Beyond that, it's mostly: piggy back the nonce set of nonces (4 public nonces total, since there're "two" messages) on a message to avoid having to add additional round trips.

The other "new" thing this adds is the generation/existence of a NUMs point, which is used to ensure that certain paths can only be spent via the script spend path (like the to remote output for the remote party, as this inherits anchor outputs semantics).

This is still marked as draft, as it's just barely to the point of being readable, and still has a lot of clean ups to be done w.r.t notation, clarify, wording, and full specification.

[Roasbeef]

Some things that came up in meatspace discussions:

• Need to ensure we specify co-op close interaction re needing to check diff types of co-op close transactions (remote party trimmed and output, etc)
• Maybe we should remove the NUMs point for the to_remote output, and just use the musig2 funding key there: in the best interest of the local party to just never sign for that
• We may need to actually send the next nonce in the commit_sig message (only one of them?) to ensure that after a commitment dance, both parties are able to immediately send a sig.

[Crypt-iQ]

I think the commit_sig should contain the sender's "remote nonce" and the revoke_and_ack contain the sender's "local nonce".

Also since funding_locked will be sent repeatedly with scid-alias when that is merged and deployed, then there should probably be language to define that the nonces are only sent the first time?

[instagibbs]

let's try to pick naming conventions for nonces that doesn't make me cry over the asymmetry

[ZmnSCPxj]

Some points:

This interacts with the 2-of-3 goal of @moneyball . If one participant uses a 2-of-3 and owns ALL 3 keys, then it is fine and we can just have MuSig2 with both channel endpoints. But the 2-of-3 goal is that one channel endpoint is really a nodelet-like setup: there is one sub-participant with 2 keys and another "server" participant with 1 key, a la GreenWallet. This requires composable MuSig2. Now I think composable MuSig2, if it can be proven safe, just requires two Rs just like normal non-composable MuSig2, but we probably need to go pester the MuSig2 authors --- I think they wrote up how composable MuSig2 would work, but only internally and they never actually published the details. This is important because we may need to have variable number of Rs, not just two.

--

This interacts with VLS as well @ksedgwic . The nonce r behind R = r * G needs to be retained, since we pre-send the R and on the actual signing MUCH later we use the r. VLS cannot have the host store r since exfiltration of the r together with a complete R, s implies exfiltration of the private key. But this is a per-channel state and constrained devices might not have enough space for each channel. What could be done would be to put the per-channel states into a Merkle tree and have the VLS constrained device store only the Merkle tree root, then every time the r has to be rotated (at each reconnect or at each signing event) update the Merkle tree root in the constrained persistent storage --- it has to be persistent since the host could keep the connection alive while power-cycling the VLS device. You also have to be careful of "UPDATE IN PLACE is a poison apple" and replicated storage. Now as mentioned we cannot store r directly so what the VLS has to do would be something like: generate random scalar q, compute Q = q * G, compute r = ECDH(k, Q) where k is the private key held by the signer device, and then store Q on the host so it can recover r later without the host being able to recover r as well.

A similar technique may also be useful for the server in the 2-of-3 of @moneyball; rather than maintain a state for each channel of each client, the client could store the per-channel Q that the server generates and uses ECDH to get r and then R = r * G for each channel.

[ZmnSCPxj]

So I talked to @jonasnick, and as I understand it, we can work with just two Rs even in the composition case, probably will also work in FROST, maybe. This should be safe but we do not have a written out proof, because it seems the proof is complicated. So at least it looks like we do not need a variable number of Rs, just two from each side should work.

[Roasbeef]

Re recursive musig2: I'm gonna give the implementation a shot (outside the LN context, just the musig-within-musig) just to double check my assumptions re not needing to modify the (revised) nonce exchange flow.

[antonilol]

i made a pull request on this pull request with script fixes

Roasbeef#1

[antonilol]

Maybe we should remove the NUMs point for the to_remote output, and just use the musig2 funding key there: in the best interest of the local party to just never sign for that

Why not the revocation key? When i publish an old state, the remote party can claim my output and htlcs with the key path, but not his own output, and also has to wait a block. If we set the internal key to the revocation key it will give the remote party more privacy, nobody on chain can see which outputs were to local and to remote (and htlcs if they are swept along). It will also give more consistency with other output as they also have the revocation key as internal key.

it will also be cheaper (or get a higher fee rate with the same amount of sats), this only requires a signature from a taptweaked revocation key (65) instead of a signature (65), the script (36) and the controlblock (34) (incl length prefix)
this will save 70 wu (17.5 vB) (keep in mind this is only applies to revoked commitments, for normal force closes we want to enforce the 1 OP_CSV)

[antonilol]

#995 (comment) makes it invisible for outside observers to identify the to_remote output in case of a revoked commitment. if there are some htlcs on it that are long expired and the second stage is broadcasted (like in the fee siphoning attack), the funds go to the local delayed pubkey + relative timelock. outside observers can now see which output was the to_local one, just search the output of an htlc 2nd stage tx in the commitment transaction.

example ctx: 15c262aeaa0c5a44e9e5f25dd6ad51b4162ec4e23668d568dc2c6ad98ae31023 (testnet)

the transaction with the expired htlc reveals the to_local output. (it is already revealed by the script, but this wouldnt be the case with a revoked taproot ctx)

this can be fixed by tweaking the local delayed pubkey with the hash of vout of the htlc on the ctx (something you can see on-chain, to make restoring backups easier) and some secret (so only you can do this, not outside observers). this secret can be static across commitments and stored in a static channel backup (this is one way i came up with, but there are of course more ways to change this key and still make restoring backups easy enough)

EDIT: no secret is needed, instead a taptweak like tweak can be done. everywhere where a local delayed pubkey is used, it is tweaked with sha256(pubkey || output index) (or a tagged hash) where output index refers to the output index of the output on the commitment transaction, this way there are no duplicates because there can't be two outputs at the same index.

for clarity: htlc outputs that send funds to the local delayed pubkey use a tweaked local delayed pubkey where the output index of the htlc output on the commitment transaction is used, not the htlc success or timeout tx

[Crypt-iQ]

for clarity: htlc outputs that send funds to the local delayed pubkey use a tweaked local delayed pubkey where the output index of the htlc output on the commitment transaction is used, not the htlc success or timeout tx

this would preserve privacy, but you'd also need to do this for the to_local and the local anchor output since if those are claimed, the delayed pubkey is also leaked. if the user doesn't claim their anchor, then only the counter-party would be able to claim their anchor (thereby leaking the local delayed pubkey) rather than anybody with the ability to watch the chain after the 16 CSV has elapsed. The keys could all be tweaked, but then perhaps there is more UTXO bloat if the anchors aren't claimed

[antonilol]

for clarity: htlc outputs that send funds to the local delayed pubkey use a tweaked local delayed pubkey where the output index of the htlc output on the commitment transaction is used, not the htlc success or timeout tx

this would preserve privacy, but you'd also need to do this for the to_local and the local anchor output since if those are claimed, the delayed pubkey is also leaked. if the user doesn't claim their anchor, then only the counter-party would be able to claim their anchor (thereby leaking the local delayed pubkey) rather than anybody with the ability to watch the chain after the 16 CSV has elapsed. The keys could all be tweaked, but then perhaps there is more UTXO bloat if the anchors aren't claimed

hmmm true, so it is either privacy, with no key reuse or no utxo set bloat.

btw another idea about anchors and less utxo set bloat:
with non taproot channels, anchors can always be claimed, because funding keys are used
with taproot, funding keys can also be used
party A and B both have a public key, the funding key becomes P = A * H(H(A || B) || A) + B (musig2 keyagg)
the anchor pubkey for A is A * H(H(A || B) || A) and for B is just its public key B
if one of the anchors is spent outside observers can calculate the other anchor because A = P - B and B = P - A (P = funding key)
the to_local and to_remote can also use these keys to ensure that if neither of the anchors is spent they can be cleaned up but that will cost some privacy

[Crypt-iQ]

P = A * H(H(A || B) || A) + B (musig2 keyagg) the anchor pubkey for A is A * H(H(A || B) || A) and for B is just its public key B if one of the anchors is spent outside observers can calculate the other anchor because A = P - B and B = P - A (P = funding key) the to_local and to_remote can also use these keys to ensure that if neither of the anchors is spent they can be cleaned up but that will cost some privacy

The KeyAgg routine here specifies some tweaking so the aggregation above may not always be the same (https://github.com/jonasnick/bips/blob/musig2/bip-musig2.mediawiki#key-aggregation). I think in your example an observer has a 50% chance of identifying B's funding_pubkey since it isn't tweaked. I am not sure if knowledge of the funding_pubkey actually gives an observer anything as I think these would have sort of an ephemeral nature as they are generated for the funding flow. A user could generate them outside of the funding context (say in their bitcoind wallet and regularly use the key for receiving/sending payments) and use them in the funding flow, but I don't see why a user would do that

[antonilol]

P = A * H(H(A || B) || A) + B (musig2 keyagg) the anchor pubkey for A is A * H(H(A || B) || A) and for B is just its public key B if one of the anchors is spent outside observers can calculate the other anchor because A = P - B and B = P - A (P = funding key) the to_local and to_remote can also use these keys to ensure that if neither of the anchors is spent they can be cleaned up but that will cost some privacy

The KeyAgg routine here specifies some tweaking so the aggregation above may not always be the same (https://github.com/jonasnick/bips/blob/musig2/bip-musig2.mediawiki#key-aggregation). I think in your example an observer has a 50% chance of identifying B's funding_pubkey since it isn't tweaked. I am not sure if knowledge of the funding_pubkey actually gives an observer anything as I think these would have sort of an ephemeral nature as they are generated for the funding flow. A user could generate them outside of the funding context (say in their bitcoind wallet and regularly use the key for receiving/sending payments) and use them in the funding flow, but I don't see why a user would do that

if this is a problem B can tweak the key before using it without A even knowing (also A has to do this because the lexicographically smaller key is tweaked), but i dont think it is, lnd uses a separate bip32 tree for this (separate from the wallet)

(btw without taproot funding pubkeys were revealed every time a channel was closed)

[antonilol]

The KeyAgg routine here specifies some tweaking so the aggregation above may not always be the same (https://github.com/jonasnick/bips/blob/musig2/bip-musig2.mediawiki#key-aggregation).

afaict the algorithm in this bip is generalized for 32 byte pubkeys and more than 2 signers, the 'simple' musig2 with the pubkey's with the parity bit known looks like this equation i used, btw i got it here https://github.com/t-bast/lightning-docs/blob/master/schnorr.md#musig2

[instagibbs]

Some old comments I forgot to submit

Most recent comment is noting that partial sigs are 32 bytes, so this needs explicit defining somewhere, since signature types seem to assume 64(may have missed it).

[antonilol]

#995 (comment)

P = A * H(H(A || B) || A) + B (musig2 keyagg) the anchor pubkey for A is A * H(H(A || B) || A) and for B is just its public key B if one of the anchors is spent outside observers can calculate the other anchor because A = P - B and B = P - A (P = funding key)

nvm this wouldn't work because keys are only revealed when swept without signature

to make this problem somewhat easier i suggest to remove the to_remote_anchor when a to_remote exists, and no longer 1 OP_CSV the to_remote output. the remote party can fee bump using his to_remote output. the only scenario where a to_remote_anchor would be needed is when there is at least 1 non dust htlc attached and the remote party has no (or below dust limit) to_remote output. this can happen when sending a non dust htlc directly after channel opening (the remote party always needs to have the channel reserve as balance, but doesn't have this just after opening)

now that the to_remote_anchor is out of the way (except for 1 case), things get easier
the to_local_anchor's internal pubkey will be local_delayed_pubkey, it is revealed after the csv delay when swept by the local party

this special case can of course be seen from both sides:

• the local party who opens a channel, sends an htlc an force closes. both parties need an anchor output here. the remote party has no to_remote output because he has no balance and cant fee bump with that. the to_remote_anchor will have the remote_htlc_pubkey as internal key because it is revealed when expired (a second sig is needed there for 2nd stage), and fulfilled
• (i switch local and remote here) the local party who just got a channel opened by the remote party and sent an htlc. the local party force closes. there is no to_local to reveal the anchor key, so the to_local_anchor's internal key will be the local_htlc_pubkey, the remote party doesn't need an anchor because he has a to_remote output

even more rare: revocation

no anchor keys are revealed here because with the revocation key the taproot key path is used. i don't know to make anchor sweepable in this case

long story short:

• extension-bolt: simple taproot channels (feature 80/81) #995 (comment) wont work
• only a to_remote_anchor when no to_remote exists but the is at least 1 htlc, its interal pubkey is the htlc pubkey
• anchor use the delayed pubkey as internal key
• if no to_local output exists the htlc pubkey is used

Questions/feedback welcome!

[Roasbeef]

Aight, I just pushed two major updates to the BOLT

1. Adding new TLVs to enable the new RBF closing protocol to be compatible with taproot channels:
   • The shutdown message is mostly unchanged w.r.t the new TLV that was added for the legacy closing flow.
   • Both sides send shutdown which includes their closee_nonce.
   • When one party goes to sign a new update, they include a JIT nonce along side closing_complete.
     • This uses the same TLV structure as the way commit sig works here.
   • When a party responds with closing_sig, they send their next_closee_nonce, as the closer just consumed their prior one by sending closing_complete.
2. A new next_local_nonces TLV to be compatible with splicing and taproot channels:
   • I tried to make this somewhat backwards compatible, to reflect lnd's current implementation.
   • We add a new map of txid:public_nonce in the channel_ready message. For implementations that don't yet support splicing, there'll only ever be a single element her, and it maps to all zeroes.
   • I'm not as familiar with the latest draft of splicing, so I sort of tried to reason about the semantics based on my existing knowledge. Feedback here is much appreciated!

[sstone]

We've tested simple taproot channels against https://github.com/Roasbeef/lnd/tree/prod-taproot-chans and we can

• open channels
• create and pay HTLCs

For these tests to work we had to:

• modify lnd's commitment weight for taproot channels (which is not specified yet) and set it to 960 instead of 968 (because we expect the tx output count to be encoded on 1 byte instead of 3).
• implement the single nonce TLV fallback when the map is missing for revoke_and_ack

Closing does not work yet though, as lnd does not seem to understand closing_complete (or I could not configure it properly ?), but we implemented the changes described in #995 (comment) and I believe that there's everything we need for option_simple_close but I proposed changes to some of the descriptions.

[sstone]

Suggested change

	- MUST extract the partial signature (first 32 bytes) and the sender's next
	closee nonce (remaining 66 bytes)
	- MUST extract the partial signature (first 32 bytes) and the sender's JIT
	nonce (remaining 66 bytes)

[Roasbeef]

Here the distinct terminology was used on purpose. I used closee/closer here, as the roles are always non ambiguous. If you're sending closing_complete, then you're the closer.

I think we should agree on what terminology we'll use here. I adopted the closee/closer based on the initial comment @t-bast left outlining a protocol (adapted it a bit as you see here).

[sstone]

It's the use of "next" that is wrong, but I guess that just removing it could be enough.

[sstone]

Suggested change

	- 66 bytes: The sender's next closee nonce for potential RBF iterations
	- 66 bytes: The sender's JIT nonce used to generate the partial signature.

[sstone]

Same here, what about The sender's closee nonce used to generate the partial signature ?

[sstone]

Suggested change

	signature with the next closee nonce
	signature with the JIT nonce it was generated with

[sstone]

Remove ? The receiver does not need to store these, only the sender must remember them for when they receive closing_sig

[Roasbeef]

Agree, will remove.

[sstone]

Suggested change

	the `shutdown` message or the previous `PartialSigWithNonce` in
	`closing_complete`
	the `shutdown` message or the `next_closee_nonce` in
	the previous `closing_sig`

[sstone]

Suggested change

	- The receiver's own closer nonce (locally generated)
	- The receiver's own closer nonce used to generate their `closing_complete`

[sstone]

Suggested change

	With this JIT nonce approach for coop close, an implementation only needs to
	store (in memory) the current closee nonce for the remote and local party. When
	either side is ready to sign, it'll generate a new closer nonce, and then that
	along with the sig.
	With this JIT nonce approach for coop close, an implementation only needs to
	store (in memory) the current closee nonce for the remote and local party, and
	the closer nonces they used in their `closing_complete` message. When
	either side is ready to sign, it'll generate new closer nonces, and send them
	along with the partial signatures.

[sstone]

We should reuse the same nonce map TLV that is attached to channel_reestablish, it is easier to use with splices.

[Roasbeef]

modify lnd's commitment weight for taproot channels (which is not specified yet) and set it to 960 instead of 968 (because we expect the tx output count to be encoded on 1 byte instead of 3).

Typically we use the most conservative values for the weight estimates. One example is assuming the largest possible push for op codes like OP_CSV.

[Roasbeef]

Closing does not work yet though, as lnd does not seem to understand closing_complete (or I could not configure it properly ?),

The RBF support is behind an additional build flag: --protocol.rbf-coop-close. For y'all's tests, did you run with this active?

[Roasbeef]

implement the single nonce TLV fallback when the map is missing for revoke_and_ack

You mean reading the existing filed if the newer nonce map isn't used?

[sstone]

implement the single nonce TLV fallback when the map is missing for revoke_and_ack

You mean reading the existing filed if the newer nonce map isn't used?

Yes.

[t-bast]

@Roasbeef @sstone opened a PR to clarify what we think must change to match the latest discussions in Roasbeef#8

This PR also adds support for dual-funded taproot channels and introduces the TLVs used for splicing.

[sstone]

modify lnd's commitment weight for taproot channels (which is not specified yet) and set it to 960 instead of 968 (because we expect the tx output count to be encoded on 1 byte instead of 3).

Typically we use the most conservative values for the weight estimates. One example is assuming the largest possible push for op codes like OP_CSV.

Using 3 bytes to encode the number of tx outputs is reasonable (though it seems that the BOLTs sometimes use 1 byte) and we can use 968 bytes for the commitment weight of taproot transactions.

[ZmnSCPxj-jr]

Now as mentioned we cannot store r directly so what the VLS has to do would be something like: generate random scalar q, compute Q = q * G, compute r = ECDH(k, Q) where k is the private key held by the signer device, and then store Q on the host so it can recover r later without the host being able to recover r as well.

This is broken, actually, because the host can generate its own Q = q * G, and tell the VLS the Q that the host generated (as the VLS is memoryless, it cannot remember whether it generated that q), and then just get r = ECDH(q, K) where K is the public key of the VLS. Instead, we should have a r' random number that is stored openly on the host, and then on the VLS, use r = HMAC(k, r'). Even if the host changes around the r', if random oracle model holds, they should not be able to predict HMAC(k, r'). I think.

[Roasbeef]

@t-bast thx for the PR, I just gave some feedback there.

[MPins]

The text says that n_A_L is sent in the open_channel message, but in the diagram (taproot_channel_open.jpg) it is sent in the funding_created message.