public inbox for bitcoindev@googlegroups.com
 help / color / mirror / Atom feed
From: John Newbery <john@johnnewbery.com>
To: Bitcoin Protocol Discussion <bitcoin-dev@lists.linuxfoundation.org>
Subject: Re: [bitcoin-dev] Revisiting squaredness tiebreaker for R point in BIP340
Date: Fri, 21 Aug 2020 09:50:49 +0100	[thread overview]
Message-ID: <CAFmfg2uL-gCV+1e1rf8+ywSUcEGTA01PP4jm+8kYub7R9E9hNw@mail.gmail.com> (raw)
In-Reply-To: <g7VgKBKlUmgMkNo6Nq1OkyilwW6Z4WbqeSZHXfvh_LDYmry5R7uGbavyEHtv2xr_0KBevL057TxSdlo_tS0_ucBXYNm8MuTu_jgyrGO9jpo=@wuille.net>

[-- Attachment #1: Type: text/plain, Size: 8865 bytes --]

Pieter,

Thanks for the illuminating write-up. There seem to be two questions here,
one technical and one process:

1. Is changing to even tie-breaker for R the correct choice technically? I
can't comment on the performance characteristics of using a square/even
tie-breaker and I'll assume the numbers you give are correct. An enormous
benefit that you don't mention (but Nadav and Lloyd do) is that
standardizing to a single tie-breaker for R points and public keys is much
simpler to explain and much easier for implementers and developers to
understand. I've explained the taproot proposals to many people through the
optech workshops and bitdevs meetups, and people are invariably confused by
which type of tie-breaker to use where. Absent a large performance benefit
for having different tiebreakers, I think this alone is good reason to
standardize to one tie-breaker.

2. Is it too late in the process to change? No. We're building things to
last years, hopefully decades. We should measure a hundred times and cut
once. A benefit of the long lead time of taproot is that as we get more
information, we can improve the proposal. Let's do that here. Nadav and
Lloyd have both written alternative implementations of taproot and are
happy to make this change. Presumably if this was going to cause serious
pain for any other implementer/developer they would have raised objections
by now.

Summary: We should change the proposal and implementation to use even
tie-breakers everywhere.

John #notoquadraticresiduetiebreakers Newbery

On Wed, Aug 12, 2020 at 7:49 PM Pieter Wuille via bitcoin-dev <
bitcoin-dev@lists.linuxfoundation.org> wrote:

> Hello all,
>
> The current BIP340 draft[1] uses two different tiebreakers for conveying
> the Y coordinate of points: for the R point inside signatures squaredness
> is used, while for public keys evenness is used. Originally both used
> squaredness, but it was changed[2] for public keys after observing this
> results in additional complexity for compatibility with existing systems.
>
> The reason for choosing squaredness as tiebreaker was performance: in
> non-batch signature validation, the recomputed R point must be verified to
> have the correct sign, to guarantee consistency with batch validation.
> Whether the Y coordinate is square can be computed directly in Jacobian
> coordinates, while determining evenness requires a conversion to affine
> coordinates first.
>
> This argument of course relies on the assumption that determining whether
> the Y coordinate is square can be done more efficiently than a conversion
> to affine coordinates. It appears now that this assumption is incorrect,
> and the justification for picking the squaredness tiebreaking doesn't
> really exist. As it comes with other trade-offs (it slows down signing, and
> is a less conventional choice), it would seem that we should reconsider the
> option of having the R point use the evenness tiebreaker (like public keys).
>
> It is late in the process, but I feel I owe this explanation so that at
> least the possibility of changing can be discussed with all information. On
> the upside, this was discovered in the context of looking into a cool
> improvement to libsecp256k1[5], which makes things faster in general, but
> specifically benefits the evenness variant.
>
>
> # 1. What happened?
>
> Computing squaredness is done through the Jacobi symbol (same inventor,
> but unrelated to Jacobian coordinates). Computing evenness requires
> converting points to affine coordinates first, and that needs a modular
> inverse. The assumption that Jacobi symbols are faster to compute than
> inverses was based on:
>
> * A (possibly) mistaken belief about the theory: fast algorithms for both
> Jacobi symbols and inverses are internally based on variants of the same
> extended GCD algorithm[3]. Since an inverse needs to extract a full big
> integer out of the transition steps made in the extgcd algorithm, while the
> Jacobi symbol just extracts a single bit, it had seemed that any advances
> applicable to one would be applicable to the other, but inverses would
> always need additional work on top. It appears however that a class of
> extgcd algorithms exists (LSB based ones) that cannot be used for Jacobi
> calculations without losing efficiency. Recent developments[4] and a
> proposed implementation in libsecp256k1[5] by Peter Dettman show that using
> this, inverses in some cases can in fact be faster than Jacobi symbols.
>
> * A broken benchmark. This belief was incorrectly confirmed by a broken
> benchmark[6] in libsecp256k1 for the libgmp-based Jacobi symbol calculation
> and modular inverse. The benchmark was repeatedly testing the same constant
> input, which apparently was around 2.5x faster than the average speed. It
> is a variable-time algorithm, so a good variation of inputs matters. This
> mistake had me (and probably others) convinced for years that Jacobi
> symbols were amazingly fast, while in reality they were always very close
> in performance to inverses.
>
>
> # 2. What is the actual impact of picking evenness instead?
>
> It is hard to make very generic statements here, as BIP340 will hopefully
> be used for a long time, and hardware advancements and algorithmic
> improvements may change the balance. That said, performance on current
> hardware with optimized algorithms is the best approximation we have.
>
> The numbers below give the expected performance change from squareness to
> evenness, for single BIP340 validation, and for signing. Positive numbers
> mean evenness is faster. Batch validation is not impacted at all.
>
> In the short term, for block validation in Bitcoin Core, the numbers for
> master-nogmp are probably the most relevant (as Bitcoin Core uses
> libsecp256k1 without libgmp, to reduce consensus-critical dependencies).
> If/when [5] gets merged, safegcd-nogmp will be what matters. On a longer
> time scale, the gmp numbers may be more relevant, as the Jacobi
> implementation there is certainly closer to the state of the art.
>
> * i7-7820HQ: (verify) (sign)
>   - master-nogmp: -0.3% +16.1%
>   - safegcd-nogmp: +6.7% +17.1%
>   - master-gmp: +0.6% +7.7%
>   - safegcd-gmp: +1.6% +8.6%
>
> * Cortex-A53: (verify) (sign)
>   - master-nogmp: -0.3% +15.7%
>   - safegcd-nogmp: +7.5% +16.9%
>   - master-gmp: +0.3% +4.1%
>   - safegcd-gmp: 0.0% +3.5%
>
> * EPYC 7742: (verify) (sign)
>   - master-nogmp: -0.3% +16.8%
>   - safegcd-nogmp: +8.6% +18.4%
>   - master-gmp: 0.0% +7.4%
>   - safegcd-gmp: +2.3% +7.8%
>
> In well optimized cryptographic code speedups as large as a couple percent
> are difficult to come by, so we would usually consider changes of this
> magnitude relevant. Note however that while the percentages for signing
> speed are larger, they are not what is unexpected here. The choice for the
> square tiebreaker was intended to improve verification speed at the cost of
> signing speed. As it turns out that it doesn't actually benefit
> verification speed, this is a bad trade-off.
>
>
> # 3. How big a change is it
>
> * In the BIP:
>   - Changing both invocations of `has_square_y` to `has_even_y`.
>   - Changing the `lift_x_square_y` invocation to `lift_x_even_y`.
>   - Applying the same change to the test vector generation code, and the
> resulting test vectors.
> * In the libsecp256k1:
>   - An 8-line patch to the proposed BIP340 implementation[7]: see [8]
> * In Bitcoin Core:
>   - Similarly small changes to the Python test reimplementation[9]
> * Duplicating these changes in other draft implementations that may
> already exist.
> * Review for all the above.
>
>
> # 4. Conclusion
>
> We discovered that the justification for using squaredness tiebreakers in
> BIP340 is based on a misunderstanding, and recent developments show that it
> may in fact be a somewhat worse choice than the alternative. It is a
> relatively simple change to address this, but that has be weighed against
> the impact of changing the standard at this stage.
>
> Thoughts?
>
>
> # 5. References
>
>   [1]
> https://github.com/bitcoin/bips/blob/master/bip-0340.mediawiki#design
>   [2]
> https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2020-February/017639.html
>   [3] https://en.wikipedia.org/wiki/Extended_Euclidean_algorithm
>   [4] https://gcd.cr.yp.to/safegcd-20190413.pdf
>   [5] https://github.com/bitcoin-core/secp256k1/pull/767
>   [6] https://github.com/bitcoin-core/secp256k1/pull/797
>   [7] https://github.com/bitcoin-core/secp256k1/pull/558
>   [8]
> https://github.com/sipa/secp256k1/commit/822311ca230a48d2c373f3e48b91b2a59e1371d6
>   [9] https://github.com/bitcoin/bitcoin/pull/17977
>
>
> Cheers,
>
> --
> Pieter
>
> _______________________________________________
> bitcoin-dev mailing list
> bitcoin-dev@lists.linuxfoundation.org
> https://lists.linuxfoundation.org/mailman/listinfo/bitcoin-dev
>

[-- Attachment #2: Type: text/html, Size: 10658 bytes --]

  parent reply	other threads:[~2020-08-21  9:14 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2020-08-12 18:49 [bitcoin-dev] Revisiting squaredness tiebreaker for R point in BIP340 Pieter Wuille
2020-08-12 20:19 ` Nadav Kohen
2020-08-13  5:31   ` Lloyd Fournier
2020-08-19 23:16 ` Pieter Wuille
2020-08-21  8:50 ` John Newbery [this message]
2020-08-27  1:10   ` Pieter Wuille

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=CAFmfg2uL-gCV+1e1rf8+ywSUcEGTA01PP4jm+8kYub7R9E9hNw@mail.gmail.com \
    --to=john@johnnewbery.com \
    --cc=bitcoin-dev@lists.linuxfoundation.org \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox