Re: [bitcoin-dev] Data structure for efficient proofs of non-inclusion

Re: [bitcoin-dev] Data structure for efficient proofs of non-inclusion

[Peter Todd]

[-- Attachment #1: Type: text/plain, Size: 6665 bytes --]

On Wed, Feb 14, 2018 at 09:09:18PM +0000, Daniel Robinson wrote:
> Hi Peter,
> 
> Thanks for the mind-expanding presentation on client-side validation at
> Chaincode last week.
> 

CCing bitcoin-dev because this is of general interest...

For background, Daniel is asking about my client-side verified single-use-seals
via proof-of-publication model, previously published here¹, which creates an
anti-double-spend primitive via a proof-of-publication, and many
proofs-of-non-publication.

tl;dr: A seal is closed by publishing a valid signature for the seal to a
ledger. The first signature is the valid one, so if Alice want to convince Bob
you she closed a seal correctly (e.g. to pay Bob), she has to supply Bob with a
proof that the signature _was_ published - a proof-of-publication - as well as
proof that prior to being published, no valid signature was previously
published - a proof-of-non-publication.

It's the proofs-of-non-publication that take up the most space.

> I'm trying to sketch out a system that would use some of those ideas, and
> am looking at what Merkle-tree-like structure for the "transaction root"
> would best allow efficient proofs of non-inclusion in a block. The simplest
> solution seems to just be a Merkle tree with the transactions sorted by
> public key, but it seems like having intermediate information about ranges
> higher up in the branches might help make non-inclusion proofs more
> compact. Other solutions I've seen like Patricia trees and sparse Merkle
> trees seem to be optimizing for easy recomputation on update, which doesn't
> seem to be necessary for at least this version of the idea.
>
> Are there any verifiable data structures you've found that improve on
> sorted Merkle trees with respect to proofs of non-inclusion?

So remember that the system I proposed¹ used sorted merkle trees only within a
block; for blocks themselves you mostly can't do any better than a linear list.
Though I think there may be some exceptions which deserves another email. :)

As you know, asymptotically merkle trees have excellent log2(n) proof size
growth. But as you correctly suggest, their high overhead in the small-n case
suggests that we can do better. In fact, Bram Cohen previously proposed² a "TXO
Bitfield" for the somewhat similar use-case of committing to the spentness of
outputs efficiently.


# Naive Analysis

So suppose at an intermediate node you commit to a simple bitfield where each
possible value under that node is represented by a single bit. Thus for m
values under that point in the tree, the marginal size of the non-inclusion
proof is m bits. By comparison a naive merkle tree built from a hash function
with k bits takes approximately k*log2(m) bits to prove non-inclusion. For an
rough, unsophisticated, analysis just solve:

    m = k * log2(m)

Apparently you can do this analytically, but as this analysis is only
approximate a much better idea is to just plot it on a graph: for k=256bits the
crossover point is roughly m=3000.


# Merkle Tree Structure

But what is the bitfield competing against exactly? Just saying "merkle tree"
isn't very precise. Most designs along these lines use something like a
merkelized patricia tree, where each bit of the key is compared individually
and each node is a two-way (left vs right) branch. Better yet is the radix
tree, where each inner node that has only one child is merged with its parent.

Regardless, the logic of these kinds of trees can be thought of a recursive
query, where each type of node has a `get(key)` operation that returns either a
value or None.

So let's define a new type of inner node that commits to a
merkle-mountain-range (MMR) tip and a 2^j element bitfield. `get(key)` is then
this pseudo-rust:

    fn get(self, prefix) -> Option<Value> {
        let idx = Int::from(prefix[0 .. j]);
        if self.bitfield[idx] {
            let mmr_idx = node.bitfield[0 .. idx].count_ones() - 1;
            Some(node.mmr[mmr_idx].get(prefix)
        } else {
            None
        }
    }

The hard part with this is choosing when to use a bitfield-based inner node
instead of a standard one. Assuming keys are randomly distributed, it makes
sense to do so when the bitfield table is partially empty, but how empty? It's
a trade-off between multiple parameters, including the size of
proofs-of-publication - although the latter may be OK to ignore as the number
of proof-of-non-publication needed should usually greatly outnumber
proofs-of-publication.

Question: is it OK for this choice to not be part of the deterministic
consensus? Is that even possible to enforce?


# Security

For a proof-of-publication to be secure, it must ensure that any attempt to
construct a false proof-of-non-publication will fail. In the pruning model,
that means that a proof-of-publication is simply the data necessary for the
proof-of-non-publication verifier to return false. Concretely:

    fn verify_pop(tree, key) -> bool {
        !verify_non_pop(tree, key)
    }

However note the subtle difference in trust model with respect to censorship
between the following two possible non-pop verifiers:

    fn verify_non_pop(tree, key) -> bool {
        !tree.exists(key)
    }

    fn verify_non_pop(tree, key) -> bool {
        match tree.get(key) {
            Some(value) => !verify_signature(value),
            None => true,
        }
    }


## False Positives

Note how if we use the second `verify_non_pop()` function shown above we can
also use probabilistic data structures such as bloom filters in place of a
bitfield. This works because a false-positive is acceptable, as it will still
fail signature verification (or sooner if the key is committed in the leaf
node).

For example, it's plausible that a compressed bloom filter would be more space
efficient than a bitfield, as the multiple hashing steps might use the bits in
the filter more efficiently. Investigating this further would be a good
research topic.


# References

1) "[bitcoin-dev] Scalable Semi-Trustless Asset Transfer via Single-Use-Seals and Proof-of-Publication",
   Peter Todd, Dec 5th 2017, https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-December/015350.html

2) "[bitcoin-dev] The TXO bitfield",
   Bram Cohen, Mar 31st 2017, https://lists.linuxfoundation.org/pipermail/bitcoin-dev/2017-March/013928.html

3) "Bloom filters",
    Wikipedia, Jan 27th 2018, https://en.wikipedia.org/w/index.php?title=Bloom_filter&oldid=822632093

-- 
https://petertodd.org 'peter'[:-1]@petertodd.org

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 614 bytes --]