[bitcoin-dev] Soft-forks and schnorr signature aggregation

[Anthony Towns <aj@erisian.com.au>]

Hello world,

There was a lot of discussion on Schnorr sigs and key and signature
aggregation at the recent core-dev-tech meeting (one relevant conversation
is transcribed at [0]).

Quick summary, with more background detail in the corresponding footnotes:
signature aggregation is awesome [1], and the possibility of soft-forking
in new opcodes via OP_RETURN_VALID opcodes (instead of OP_NOP) is also
awesome [2].

Unfortunately doing both of these together may turn out to be awful.

RETURN_VALID and Signature Aggregation
--------------------------------------

Bumping segwit script versions and redefining OP_NOP opcodes are
fairly straightforward to deal with even with signature aggregation,
the straightforward implementation of both combined is still a soft-fork.

RETURN_VALID, unfortunately, has a serious potential pitfall: any
aggregatable signature operations that occur after it have to go into
separate buckets.

As an example of why this is the case, imagine introducing a covenant
opcode that pulls a potentially complicated condition from the stack
(perhaps, "an output pays at least 50000 satoshi to address xyzzy"),
checks the condition against the transaction, and then pushes 1 (or 0)
back onto the stack indicating compliance with the covenant (or not).
You might then write a script allowing a single person to spend the coins
if they comply with the covenant, and allow breaking the covenant with
someone else's sign-off in addition. You could write this as:

  pubkey1 CHECKSIGVERIFY
  cond CHECKCOVENANT IFDUP NOTIF pubkey2 CHECKSIG ENDIF

If you pass the covenant, you supply "SIGHASHALL|BUCKET_1" and aggregate
the signature for pubkey1 into bucket1 and you're set; otherwise you supply
"SIGHASHALL|BUCKET_1 SIGHASHALL|BUCKET_1" and aggregate signatures for both
pubkey1 and pubkey2 into bucket1 and you're set. Great!

But this isn't a soft-fork: old nodes would see this script as:

  pubkey1 CHECKSIGVERIFY
  cond RETURN_VALID IFDUP NOTIF pubkey2 CHECKSIG ENDIF

which it would just interpret as:

  pubkey1 CHECKSIGVERIFY cond RETURN_VALID

which is fine if the covenant was passing; but no good if the covenant
didn't pass -- they'd be expecting the aggregted sig to just be for
pubkey1 when it's actually pubkey1+pubkey2, so old nodes would fail the
tx and new nodes would accept it, making it a hard fork.

Solution 0a / 0b
----------------

There are two obvious solutions here:

 0a) Just be very careful to ensure any aggregated signatures that
     are conditional on an redefined RETURN_VALID opcode go into later
     buckets, but be careful about having separate sets of buckets every
     time a soft-fork introduces a new redefined opcode. Probably very
     complicated to implement correctly, and essentially doubles the
     number of buckets you have to potentially deal with every time you
     soft fork in a new opcode.

 0b) Alternatively, forget about the hope that RETURN_VALID
     opcodes could be converted to anything, and just reserve OP_NOP
     opcodes and convert them to CHECK_foo_VERIFY opcodes just as we
     have been doing, and when we can't do that bump the segwit witness
     version for a whole new version of script. Or in twitter speak:
     "non-verify upgrades should be done with new script versions" [3]

I think with a little care we can actually salvage RETURN_VALID though!

Solution 1
----------

You don't actually have to write your scripts in ways that can cause
this problem, as long as you're careful. In particular, the problem only
occurs if you do aggregatable CHECKSIG operations after "RETURN_VALID"
-- if you do all the CHECKSIGs first, then all nodes will be checking
for the same signatures, and there's no problem. So you could rewrite
the script above as:

  pubkey1 CHECKSIGVERIFY
  IF pubkey2 CHECKSIG ENDIF
  cond CHECKCOVENANT OR

which is redeemable either by:

  sig1 0        [and covenant is met]
  sig1 1 sig2   [covenant is not checked]

The witness in this case is essentially committing to the execution path
that would have been taken in the first script by a fully validating node,
then the new script checks all the signatures, and then validates that the
committed execution path was in fact the one that was meant to be taken.

If people are clever enough to write scripts this way, I believe you
can make RETURN_VALID soft-fork safe simply by having every soft-forked
RETURN_VALID operation set a state flag that makes every subsequent
CHECKSIG operation require a non-aggregated sig.

The drawback of this approach is that if the script is complicated
(eg it has multiple IF conditions, some of which are nested), it may be
difficult to write the script to ensure the signatures are checked in the
same combination as the later logic actually requires -- you might have
to store the flag indicating whether you checked particular signatures
on the altstack, or use DUP and PICK/ROLL to organise it on the stack.

Solution 2
----------

We could make that simpler for script authors by making dedicated opcodes
to help with "do all the signatures first" and "check the committed
execution path against reality" steps. I think a reasonable approach
would be something like:

  0b01 pubkey2 pubkey1 2 CHECK_AGGSIG_VERIFY
  cond CHECKCOVENANT 0b10 CHECK_AGG_SIGNERS OR

which is redeemed either by:

  sighash1 0   [and passing the covenant cond]
  sighash2 sighash1 0b10

(I'm using the notation 0b10110 to express numbers as binary bitfields;
0b10110 = 22 eg)

That is, two new opcodes, namely:

CHECK_AGGSIG_VERIFY which takes from the stack:
    - N: a count of pubkeys
    - pubkey1..pubkeyN: N pubkeys
    - REQ: a bitmask of which pubkeys are required to sign
    - OPT: a bitmask of which optional pubkeys have signed
    - sighashes: M sighashes for the pubkeys corresponding to the set
      bits of (REQ|OPT)

  CHECK_AGGSIG_VERIFY fails if:
    - the stack doesn't have enough elements
    - the aggregated signature doesn't pass
    - a redefined RETURN_VALID opcode has already been seen
    - a previous CHECK_AGGSIG_VERIFY has already been seen in this script

  REQ|OPT is stored as state

CHECK_AGG_SIGNERS takes from the stack:
    - B: a bitmask of which pubkeys are being queried
  and it pushes to the stack 1 or 0 based on:
    - (REQ|OPT) & B == B ? 1 : 0

A possible way to make sure the "no agg sigs after an upgraded
RETURN_VALID" behaviour works right might be to have "RETURN_VALID"
fail if CHECK_AGGSIG_VERIFY hasn't already been seen. That way once you
redefine RETURN_VALID in a soft-fork, if you have a CHECK_AGGSIG_VERIFY
after a RETURN_VALID you've either already failed (because the
RETURN_VALID wasn't after a CHECK_AGGSIG_VERIFY), or you automatically
fail (because you've already seen a CHECK_AGGSIG_VERIFY).

There would be no need to make CHECKSIG, CHECKSIGVERIFY, CHECKMULTISIG
and CHECKMULTISIGVERIFY do signature aggregation in this case. They could
be left around to allow script authors to force non-aggregate signatures
or could be dropped entirely, I think.

This construct would let you do M-of-N aggregated multisig in a fairly
straightforward manner without needing an explicit opcode, eg:

  0 pubkey5 pubkey4 pubkey3 pubkey2 pubkey1 5 CHECK_AGGSIG_VERIFY
  0b10000 CHECK_AGG_SIGNERS
  0b01000 CHECK_AGG_SIGNERS ADD
  0b00100 CHECK_AGG_SIGNERS ADD
  0b00010 CHECK_AGG_SIGNERS ADD
  0b00001 CHECK_AGG_SIGNERS ADD
  3 NUMEQUAL

redeemable by, eg:

  0b10110 sighash5 sighash3 sighash2

and a single aggregate signature by the private keys corresponding to
pubkey{2,3,5}.

Of course, another way of getting M-of-N aggregated multisig is via MAST,
which brings us to another approach...

Solution 3
----------

All we're doing above is committing to an execution path and validating
signatures for that path before checking the path was the right one. But
MAST is a great way of committing to an execution path, so another
approach would just be "don't have alternative execution paths, just have
MAST and CHECK/VERIFY codes". Taking the example I've been running with,
that would be:

  branch1: 2 pubkey2 pubkey1 2 CHECKMULTISIG
  branch2: pubkey1 CHECKSIGVERIFY cond CHECKCOVENANT

So long as MAST is already supported when signature aggregation becomes
possible, that works fine. The drawback is MAST can end up with lots of
branches, eg the 3-of-5 multisig check has 10 branches:

  branch1: 3 pubkey3 pubkey2 pubkey1 3 CHECKMULTISIG
  branch2: 3 pubkey4 pubkey2 pubkey1 3 CHECKMULTISIG
  branch3: 3 pubkey5 pubkey2 pubkey1 3 CHECKMULTISIG
  branch4: 3 pubkey4 pubkey3 pubkey1 3 CHECKMULTISIG
  branch5: 3 pubkey5 pubkey3 pubkey1 3 CHECKMULTISIG
  branch6: 3 pubkey5 pubkey4 pubkey1 3 CHECKMULTISIG
  branch7: 3 pubkey4 pubkey3 pubkey2 3 CHECKMULTISIG
  branch8: 3 pubkey5 pubkey3 pubkey2 3 CHECKMULTISIG
  branch9: 3 pubkey5 pubkey4 pubkey2 3 CHECKMULTISIG
  branch10: 3 pubkey5 pubkey4 pubkey3 3 CHECKMULTISIG

while if you want, say, 6-of-11 multisig you get 462 branches, versus
just:

  0 pubkey11 pubkey10 pubkey9 pubkey8 pubkey7 pubkey6
    pubkey5 pubkey4 pubkey3 pubkey2 pubkey1 11 CHECK_AGGSIG_VERIFY
  0b10000000000 CHECK_AGG_SIGNERS
  0b01000000000 CHECK_AGG_SIGNERS ADD
  0b00100000000 CHECK_AGG_SIGNERS ADD
  0b00010000000 CHECK_AGG_SIGNERS ADD
  0b00001000000 CHECK_AGG_SIGNERS ADD
  0b00000100000 CHECK_AGG_SIGNERS ADD
  0b00000010000 CHECK_AGG_SIGNERS ADD
  0b00000001000 CHECK_AGG_SIGNERS ADD
  0b00000000100 CHECK_AGG_SIGNERS ADD
  0b00000000010 CHECK_AGG_SIGNERS ADD
  0b00000000001 CHECK_AGG_SIGNERS ADD
  6 NUMEQUAL

Provided doing lots of hashes to calculate merkle paths is cheaper than
publishing to the blockchain, MAST will likely still be better though:
you'd be doing 6 pubkeys and 9 steps in the merkle path for about 15*32
bytes in MAST, versus showing off all 11 pubkeys above for 11*(32+4)
bytes, and the above is roughly the worst case for m-of-11 multisig
via MAST.

If everyone's happy to use MAST, then it could be the only solution:
drop OP_IF and friends, and require all the CHECKSIG ops to occur before
any RETURN_VALID ops: since there's no branching, that's just a matter of
reordering your script a bit and should be pretty easy for script authors.

I think there's a couple of drawbacks to this approach that it shouldn't
be the only solution:

 a) we don't have a lot of experience with using MAST
 b) MAST is a bit more complicated than just dealing with branches in
    a script (probably solvable once (a) is no longer the case) 
 c) some useful scripts might be a bit cheaper expressed with 
    of branches and be better expressed without MAST

If other approaches than MAST are still desirable, then MAST works fine
in combination with either of the earlier solutions as far as I can see.

Summary
-------

I think something along the lines of solution 2 makes the most sense,
so I think a good approach for aggregate signatures is:

 - introduce a new segwit witness version, which I'll call v2 (but which
   might actually be v1 or v3 etc, of course)

 - v2 must support Schnorr signature verification.

 - v2 should have a "pay to public key (hash?)" witness format. direct
   signatures of the transaction via the corresponding private key should
   be aggregatable.

 - v2 should have a "pay to script hash" witness format: probably via
   taproot+MAST, possibly via graftroot as well

 - v2 should support MAST scripts: again, probably via taproot+MAST

 - v2 taproot shouldn't have a separate script version (ie,
   the pubkey shouldn't be P+H(P,version,scriptroot)), as signatures
   for later-versioned scripts couldn't be aggregated, so there's no
   advantage over bumping the segwit witness version

 - v2 scripts should have a CHECK_AGG_SIG_VERIFY opcode roughly as
   described above for aggregating signatures, along with CHECK_AGG_SIGNERS

 - CHECK{MULTI,}SIG{VERIFY,} in v2 scripts shouldn't support aggregated
   signatures, and possibly shouldn't be present at all?

 - v2 signers should be able to specify an aggregation bucket for each
   signature, perhaps in the range 0-7 or so?

 - v2 scripts should have a bunch of RETURN_VALID opcodes for future
   soft-forks, constrained so that CHECK_AGG_SIG_VERIFY doesn't appear
   after them. the currently disabled opcodes should be redefined as
   RETURN_VALID eg.

For soft-fork upgrades from that point:

 - introducing new opcodes just means redefining an RETURN_VALID opcode

 - introducing new sighash versions requires bumping the segwit witness
   version (to v3, etc)

 - if non-interactive half-signature aggregation isn't ready to go, it
   would likewise need a bump in the segwit witness version when
   introduced

I think it's worth considering bundling a hard-fork upgrade something
like:

 - ~5 years after v2 scripts are activated, existing p2pk/p2pkh UTXOs
   (either matching the pre-segwit templates or v0 segwit p2wpkh) can
   be spent via a v2-aggregated-signature (but not via taproot)
   [4]

 - core will maintain a config setting that allows users to prevent
   that hard fork from activating via UASF up until the next release
   after activation (probably with UASF-enforced miner-signalling that
   the hard-fork will not go ahead)

This is already very complicated of course, but note that there's still
*more* things that need to be considered for signature aggregation:

 - whether to use Bellare-Neven or muSig in the consensus-critical
   aggregation algorithm

 - whether to assign the aggregate sigs to inputs and plunk them in the
   witness data somewhere, or to add a new structure and commitment and
   worry about p2p impact

 - whether there are new sighash options that should go in at the same time

 - whether non-interactive half-sig aggregation can go in at the same time

That leads me to think that interactive signature aggregation is going to
take a lot of time and work, and it would make sense to do a v1-upgrade
that's "just" Schnorr (and taproot and MAST and re-enabling opcodes and
...) in the meantime. YMMV.

Cheers,
aj

[0] http://diyhpl.us/wiki/transcripts/bitcoin-core-dev-tech/2018-03-06-taproot-graftroot-etc/

[1] Signature aggregation:

    Signature aggregation is cool because it lets you post a transaction
    spending many inputs, but only providing a single 64 byte signature
    that proves authorisation by the holders of all the private keys
    for all the inputs. So the witnesses for your inputs might be:

     p2wpkh: pubkey1 SIGHASH_ALL
     p2wpkh: pubkey2 SIGHASH_ALL
     p2wsh: "3 pubkey1 pubkey3 pubkey4 3 CHECKMULTISIG" SIGHASH_ALL SIGHASH_ALL SIGHASH_ALL

    where instead of including full 65-byte signature for each CHECKSIG
    operation in each input witness, you just include the ~1-byte sighash,
    and provide a single 64-byte signature elsewhere, calculated either
    according to the Bellare-Neven algorithm, or the muSig algorithm.

    In the above case, that means going from about 500 witness bytes
    for 5 public keys and 5 signatures, to about 240 witness bytes for
    5 public keys and just 1 signature.

    A complication here is that because the signatures are aggregated,
    in order to validate *any* signature you have to be able to validate
    *every* signature.

    It's possible to limit that a bit, and have aggregation
    "buckets". This might be something you just choose when signing, eg:

     p2wpkh: pubkey1 SIGHASH_ALL|BUCKET_1
     p2wpkh: pubkey2 SIGHASH_ALL|BUCKET_2
     p2wsh: "3 pubkey1 pubkey3 pubkey4 3 CHECKMULTISIG" SIGHASH_ALL|BUCKET_1 SIGHASH_ALL|BUCKET_2 SIGHASH_ALL|BUCKET_2

     bucket1: 64 byte sig for (pubkey1, pubkey1)
     bucket2: 64 byte sig for (pubkey2, pubkey3, pubkey4)

    That way you get the choice to verify both of the pubkey1 signatures
    or all of the pubkey{2,3,4} signatures or all the signatures (or
    none of the signatures).

    This might be useful if the private key for pubkey1 is essentially
    offline, and can't easily participate in an interactive protocol
    -- with separate buckets the separate signatures can be generated
    independently at different times, while with only one bucket,
    everyone has to coordinate to produce the signature)

    (For clarity: each bucket corresponds to many CHECKSIG operations,
    but only contains a single 64-byte signature)

    Different buckets will also be necessary when dealing with new
    segwit script versions: if there are any aggregated signatures for
    v1 addresses that go into bucket X, then aggregate signatures for
    v2 addresses cannot go into bucket X, as that would prevent nodes
    that support v1 addresses but not v2 addresses from validating
    bucket X, which would prevent them from validating the v1 addresses
    corresponding to that bucket, which would make the v2 upgrade a hard
    fork rather than a soft fork. So each segwit version will need to
    introduce a new set of aggregation buckets, which in turn reduces
    the benefit you get from signature aggregation.

    Note that it's obviously fine to use an aggregated signature in
    response to CHECKSIGVERIFY or n-of-n CHECKMULTISIGVERIFY -- when
    processing the script you just assume it succeeds, relying on the
    fact that the aggregated signature will fail the entire transaction
    if there was a problem. However it's also fine to use an aggregated
    signature in response to CHECKSIG for most plausible scripts, since:

        sig key CHECKSIG

    can be treated as equivalent to

        sig DUP IF key CHECKSIGVERIFY OP_1 FI

    provided invalid signatures are supplied as a "false" value. So
    for the purpose of this email, I'll mostly be treating CHECKSIG and
    n-of-n CHECKMULTISIG as if they support aggregation.

[2] Soft-forks and RETURN_VALID:

    There are two approaches for soft-forking in new opcodes that are
    reasonably well understood:

    1) We can bump the segwit script version, introducing a new class of
       bc1 bech32 addresses, which behave however we like, but can't be
       validated at all by existing nodes. This has the downside that it
       effectively serialises upgrades.

    2) We can redefine OP_NOP opcodes as OP_CHECK_foo_VERIFY
       opcodes, along the same lines as OP_CHECKLOCKTIMEVERIFY or
       OP_CHECKSEQUENCEVERIFY. This has the downside that it's pretty
       restrictive in what new opcodes you can introduce.

    A third approach seems possible as well though, which would combine
    the benefits of both approaches: allowing any new opcode to be
    introduced, and allowing different opcodes to be introduced in
    concurrent soft-forks. Namely:

    3) If we introduce some RETURN_VALID opcodes (in script for a new
       segwit witness version), we can then redefine those as having any
       behaviour we might want, including ones that manipulate the stack,
       and have the change simply be a soft-fork. RETURN_VALID would
       force the script to immediately succeed, in contrast to OP_RETURN
       which forces the script to immediately fail.

[3] https://twitter.com/bramcohen/status/972205820275388416

[4] https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2018-January/015580.html