5
$\begingroup$

I am a CS person so please excuse the hand-waving.

Given a set of machine-represented proofs, each different (but not necessarily proving a different thing), what sort of information-theoretic statements could we make about these?

I would define the information "density" of a proof by simply losslessly compressing it using some standard means (like lz4 or rle or whatever is suited to this domain) and comparing the compressed size to the input size. The compression ratio.

For example: if we compress them all, would they all have similar compression ratios (meaning non-redundant/repeating information content)?

Given a set of different proofs, all proving the same thing (say like the prime number theorem), what is the most dense? The least dense?

Are there useful proofs that are not "information dense", meaning they have long repeating sequences or repeating structure that is easy to compress. These would look tedious to a human but would be discover-able by a machine.

Are there any papers that have looked at this? What area is this? My googling has failed me.

$\endgroup$
1

1 Answer 1

3
$\begingroup$

First, beware: this can only be done meaningfully within a fixed "language of proof". If you try to compare across different systems, you can get wildly different results. There is a whole domain of Proof Complexity which has lots of results about "expressivity". Adding features to your language (like let x = <big-expression> in <other-expression>) can make a huge difference. Other subtler changes can still result in exponential differences in proof length.

Similarly, which compressor you use matters. Kolmogorov Complexity relies crucially on Turing-completeness to obtain a 'stable' measure of length. Normal compressors can be used to approximate it, and give surprising good results in practice. Paul Vitanyi has written a variety of papers on that topic.

Regarding a specific example, like the prime number theorem, an obvious question becomes: what results do you consider to be 'background' and which are part of the proof? The stable answer, again relying on Kolmogorov Complexity, which I explain in Understanding expression simplification [ISSAC 2004], is to in fact consider the proof length to contain all the previous results needed to express the proof as part of its length. Then, to compare two proofs, you subtract their lengths, which takes care of 'quotienting out' all of the material that is common. However, if the two proofs take wildly different routes through mathematics, you might need to use finer concepts such as relative complexity, i.e. you compress the proofs on their own (but with the full library), and concatenated. See the works of Vitanyi for details and variants.

I would say that we currently don't really too much about this topic. We're still in the stage where we're finding uniform ways of writing down mathematical proofs. While the existing libraries are getting large, they are still not quite large enough to do convincing data mining, although that hasn't stopped many from doing so already. But they are aware that their results are very system-relative.

Some interesting early results are available on decently large databases: some use the whole of the arxiv as their data source. I quite like Discovering Mathematical Objects of Interest—A Study of Mathematical Notations for example.

$\endgroup$
1
  • $\begingroup$ The Kolmogorov complexity was the concept I was looking for (versus "compression"), thanks! I have actually heard of this before in the context of cellular automata, but I forget exactly how. Now let me attempt that paper.. $\endgroup$
    – user318904
    Aug 13, 2020 at 20:58

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.