Scott Aaronson's Feedback on Doron Zeilberger's Opinion 76, Dated Jan. 11, 2008


Your delightful post on Knuth today led me to your essay on Avi
Wigderson's ICM lecture and the P vs. NP problem.  I'm not Avi (just a
former postdoc of his), but I wanted to take the liberty anyway of
responding to some of your points.

1. I think you misunderstood the point about NP problems being easy to
check.  No, checkability doesn't help explain why the NP-complete
problems are hard, nor does anyone think it does.  What it does help
explain is why the P versus NP problem *itself* is hard!  (If the
solutions took exponential time to check, then it would be trivial to
prove there's no polynomial-time searching algorithm.)

2. I don't think Avi's basic argument for P!=NP requires any
Penrose-style leap of faith about the specialness of human creativity.
 One can say: our experience has been that solving combinatorial
search problems is an extremely hit-or-miss business, and that when we
do strike a hit, it's because we have some special knowledge of the
problem structure (no doubt helped by a billion years of evolution).
One can therefore speculate that brute-force search will in general be
unavoidable for *any* intelligent entity (humans, AI bots, etc.).
This is not that far from some of the arguments for P!=NP you yourself
give.

3. I think theoretical computer scientists understand perfectly well
that polynomial vs. exponential is an idealization imposed on a finite
universe.  Like any other idealization, this one can only be justified
by its usefulness: leading to powerful new theorems, open problems,
and so on.  By that criterion, I would argue that it's already
justified itself a thousand times over.  Of course, someone who
disparages the achievements of TCS as "yet another reduction theorem"
(which is sort of like disparaging the achievements of physics as "yet
another particle") would naturally disagree with that assessment.

4. While it's true that RH, Goldbach, and P!=NP have proofs of size
O(1) each assuming they're provable at all, it's not true (as you say)
that "introducing a parameter n [is] completely misleading."  One can
ask, for example: is there a proof of RH with at most n symbols?  And
then the issue becomes whether the time needed to answer that question
grows polynomially or exponentially with n.  (Indeed, this is exactly
how Gödel phrased the P vs. NP question in his now-famous 1956 letter
to von Neumann.)

5. Personally, I'd be *thrilled* if automated techniques could give us
new insight into complexity questions like P vs. NP.  I even tried
myself to apply automated techniques at various points: for example, I
wrote a software package for computing decision-tree measures of
Boolean functions, and I tried to do a search for the smallest
arithmetic circuit computing the 4-by-4 Permanent.  But in both cases,
I was stymied by the astronomical size of the search space, which
grows not merely exponentially but doubly-exponentially with the input
size n.  So, do you have a concrete proposal for how we complexity
theorists should be enlisting computers to help us in the near future
(i.e. without waiting around for general-purpose AI-bots)?  I think
such a proposal would be met with considerable enthusiasm (in contrast
to, say, kvetching from the sidelines :-) ).

6. I completely agree that Avi needs to use a spell-checker.

All the best,
Scott Aaronson

Back to Opinion 76 of Doron Zeilberger

Doron Zeilberger's Homepage