home *** CD-ROM | disk | FTP | other *** search
- Newsgroups: comp.theory
- Path: sparky!uunet!news.smith.edu!orourke
- From: orourke@sophia.smith.edu (Joseph O'Rourke)
- Subject: Re: Silence on P=NP
- Message-ID: <1992Nov23.193912.9159@sophia.smith.edu>
- Organization: Smith College, Northampton, MA, US
- References: <1992Nov21.222755.18863@sophia.smith.edu> <1992Nov23.091136.7123@fwi.uva.nl> <By6nDC.6Gp@dcs.ed.ac.uk>
- Date: Mon, 23 Nov 1992 19:39:12 GMT
- Lines: 55
-
- I received this reply to my posting from David Shallcross at Bellcore:
-
- From davids@bellcore.com Mon Nov 23 12:44:35 1992
- Date: Mon, 23 Nov 92 12:43:17 -0500
- From: davids@bellcore.com (David Shallcross)
-
- Subject: Re: Silence on P=NP
- Newsgroups: comp.theory
- In-Reply-To: <1992Nov21.222755.18863@sophia.smith.edu>
- Organization: Bellcore, Morristown NJ
-
- In article <1992Nov21.222755.18863@sophia.smith.edu> you write:
- >The silence on the Swart & Gismondi proof of P=NP is deafening.
- >Will no one venture an opinion? My knowledge of linear programming
- >is insufficient to the task.
-
- Well, I have problems posting from here, but I read through the papers,
- and found some fundamental errors. Basically, he works with the set
- of two-sided permutation operators on matrices f(X) = P'XP, where P
- is some permutation matrix and P' is the transpose of P. Since f is
- linear in X, we can write this as f(x) = Tx, where x is just X
- rearranged into a vector. He further rearranges T into a vector t, and
- (the usually combinatorial optimization - integer programming thing)
- gives a system F1 of linear inequalities such that the set of such t's
- is the set of integer vectors that satisfy F1. (Let C1 be the set of
- real vectors that satisfy F1.(Gismondi and Swart never actually define
- C1.)) What we would like, in order to easily solve lots of NP-hard problems,
- is a polynomial description of the convex hull C0 of such t's. What they do,
- up to some gaps in their proofs that I cannot quite follow, is give a set C2,
- containing C0 and contained in C1, by adding new variables, deducing
- new constraints, and projecting the new variables back out.
-
- Their main theorem, which might be correct, is that any point in C2
- either lies in C0, or is a convex combination of a point in C1 and a
- point in C0. This is not very surprising, because, for any set A0
- contained in any other set A1 (in R^n), the union of A0 with the
- interior of A1 also has this property.
-
- The proof of Corollary 1 is clearly invalid, however. It claims that
- if we intersect C2 with some additional linear constraints (that a
- graph Ga is mapped to a graph Gb by f) to get C3, and C3 is nonempty,
- then C3 must contain an integer point (and so there must an isomorphism
- between Ga and Gb). The argument for this is that any point in C3 is the
- convex combination of an integer point in C0 and some other point in C1,
- which is true, but there is no reason that either of those two points
- need also satisfy the added linear constraints.
-
- The later two papers more or less act as if C0=C2, in which case all
- their wonderous results would be true. Basically, there are
- misunderstandings of the theory of integer programming.
-
- Post this if you like.
-
- Your once-student,
- David Shallcross
-