Skip to content

Commit

Permalink
Update soundness-notes.md (#1050)
Browse files Browse the repository at this point in the history
  • Loading branch information
BorkBorked committed Jan 15, 2024
1 parent 9f8231e commit f9f8754
Showing 1 changed file with 5 additions and 5 deletions.
10 changes: 5 additions & 5 deletions notes/soundness-notes.md
Original file line number Diff line number Diff line change
Expand Up @@ -41,7 +41,7 @@ will be unable to demonstrate an opening of the commitment.


NOTE: the notation used when displaying a Lurk commitment is `(comm <field-element>)`, but the actual value represented
is just a field element tagged with the `Comm` type -- not a list begining with the symbol `comm`. This representation
is just a field element tagged with the `Comm` type -- not a list beginning with the symbol `comm`. This representation
is chosen because the `comm` built-in creates commitments from numbers.

```
Expand All @@ -59,7 +59,7 @@ the prover must know the original secret (as well as the original value).
=> 169
```

Here, we see that the output is exactly as expected when calling the secret funcion. 169 is the sum of the squares (4,
Here, we see that the output is exactly as expected when calling the secret function. 169 is the sum of the squares (4,
16, 49, 100) of the input. However, we learn nothing *else* about the secret function even when we can fully inspect the
input and output of evaluation.

Expand All @@ -74,16 +74,16 @@ It follows from the above that Lurk programs are deterministic. For every termin
correct eventual output. Future versions of Lurk will introduce a non-deterministic option -- in which a single input
may evaluate to more than one correct output.

In either event, whenever `x` evalutes to `y` (whether `y` is unique or otherwise), it is possible to construct a proof
In either event, whenever `x` evaluates to `y` (whether `y` is unique or otherwise), it is possible to construct a proof
that convinces a verifier that this is true. We claim that the circuit for the Lurk [reduction
step](reduction-notes.md) is complete in this sense, but will not otherwise address the claim in this document.

## Soundness

Our concern here is that even if it is true that we can prove that `x` evalutes to `y` whenever this is true, it might
Our concern here is that even if it is true that we can prove that `x` evaluates to `y` whenever this is true, it might
also be the case that we can generate a proof that `x` evaluates to `z` when Lurk's evaluation rules do *not* specify
that this is true. In the case of the deterministic subset of Lurk (which is all that is implemented in the
current/initial release), it *cannot* be the case that `x` evalutes to both `y` and `z` -- so any instance of such a
current/initial release), it *cannot* be the case that `x` evaluates to both `y` and `z` -- so any instance of such a
proof would constitute *ipso facto* proof that Lurk proofs are unsound. This would be a show-stopping bug that needs to
be addressed. We frame the discussion in this way to make clear that Lurk auditors must be satisfied that no two such
'contradictory proofs' are obtainable if Lurk is to be declared sound (as we claim).
Expand Down

1 comment on commit f9f8754

@github-actions
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Benchmarks

Table of Contents

Overview

This benchmark report shows the Fibonacci GPU benchmark.
NVIDIA L4
Intel(R) Xeon(R) CPU @ 2.20GHz
125.78 GB RAM
Workflow run: https://github.com/lurk-lab/lurk-rs/actions/runs/7529548607

Benchmark Results

LEM Fibonacci Prove - rc = 100

fib-ref=9f8231ec309611f76a4f5301df2e7f87464b1564 fib-ref=f9f87545a3ffa0cf921fa2557021a7b9040b7ad5
num-100 1.75 s (✅ 1.00x) 1.74 s (✅ 1.01x faster)
num-200 3.37 s (✅ 1.00x) 3.36 s (✅ 1.00x faster)

LEM Fibonacci Prove - rc = 600

fib-ref=9f8231ec309611f76a4f5301df2e7f87464b1564 fib-ref=f9f87545a3ffa0cf921fa2557021a7b9040b7ad5
num-100 2.03 s (✅ 1.00x) 2.02 s (✅ 1.01x faster)
num-200 3.40 s (✅ 1.00x) 3.38 s (✅ 1.00x faster)

Made with criterion-table

Please sign in to comment.