This page is for student questions and my answers related to HW#6. Questions are in black, answers in blue.
Can you explain why there is a 1t in the definition of MEMBER-NP in Question 2? Would the language still be NP-complete if this were left out?
It turns out that if we had, say, {(M,w,k): M can accept w in time
nk} as our definition, we would define a language that is
NP-hard (everything in NP reduces to it) but would not be in NP!
This should make sense once you have the correct proof and look at
it carefully. Remember that to be in NP, there must be a single
NTM with a single polynomial time bound that works for any input.
But an arbitrary language in NP might have an arbitarily large polynomial
bound on its machine's running time.
I have a question about problem 1 on the homework. Initially I thought that I could prove it in a way similar to the way that we proved that REACH is complete for NL, but then I realized that this was not a log-space reduction (lecture 17 p 12-14) since CompGraph has nodes on the order n (for each h<=n). Is this correct? If so, then I am struggling to figure out how I can proceed. Any hints, or am I missing something obvious?
What is the reduction that you think is not log-space? Do you see that the reduction from w to Compgraph(N,w), where N is a logspace NDTM, is a logspace reduction?
Hmm. I guess I'm not really sure what it means to be a log-space reduction. I was interpreting this to mean that f(w) = CompGraph(N,w) may require at least one state for every position h<= n in w, therefore it's polynomial space but not log space. DOes it instead mean that a turing machine that generates f(w) from w uses only O(log n) worktape bits (whereas the number of output bits would reflect the size of f(w))?
Yes, exactly. The output goes on a write-only tape that doesn't
count as space. You should go back over all the reductions that
are asserted in lecture to be log-space and see that they can be
computed with only O(log n) bits of worktape. Usually most of this
worktape is used to keep one or more pointers into the input or to
maintain the index variables for loops.
Actually in this problem
you probably don't want to follow the proof of
"REACH is NL-complete" too closely, though it's probably
possible to do it that way. It's easier to (a) prove that
BETWEEN is in NL, and (b) reduce some known NP-complete problem,
such as REACH, to BETWEEN. The latter reduction would have
to be logspace, of course.
For question 2, what is n when considering wether MEMBER-NP is in NP. In other words, you have w and t, and if t = n, it seems like you could solve the problem in linear time, by running M on w for t time, and returning positive if it works. If w = n, this makes more sense.
Here "n" is the total size of the input
For question 4: by "distance between every pair of distinct nodes" do you mean weighted distance, or just distance by number of nodes traversed in the path? Or are the edges all of unit length?
No, The graph is complete, each edge has a weight, and by "the distance between a pair of nodes" I meant the weight of the edge between those two nodes.
Also, I'm having trouble figuring out how the decision problem helps determine a polytime algorithm for questions 3,5,6. Am I just missing something from the lectures? Should we actually be able to find these algorithms, or just postulate there existence based on what we're given from the decision knowledge.
This is an algorithm-construction problem. You are given (by the assumption) a poly-time algorithm for the decision problem, and you have to figure out how to use this (up to poly-many times along with poly-many other operations) to solve the other problems.
In Homework 6, Question 2, can we refer to the running time of an arbitrary NP language? That is, in performing the reduction of an arbitrary NP language L to MEMBER-NP, can we pass as parameters the NTM that accepts L as well as the time in which that NTM accepts?
If A is in NP, you may assume that a machine _and_ a polynomial exist such that the machine can accept an input within the time bound iff the input is in A. If you define the reduction function in terms of the machine and the time bound, then it also exists, and this is what you need to show that the reduction exists.
Is the TSP decision problem mentioned in problems 5 and 6 the sane as the one taking G and length k as arguments as in problem 4?
Yes.
A followup to QA 6.5:
Can I say A ≤ B if A is in F(NP) and B is NP-complete?
It's munging the difference between F(C) and C for a complexity class C, but I feel like it's a fairly obvious thing...we can translate any F(C) function to a C language without changing the complexity class, can't we?
I think you should be careful to only use the ≤ operator and the property NP-complete on languages. If a function is in F(NP), its bitgraph language is in NP, but it isn't in NP. If what you're planning to do is legitimate, there is a legitimate way to phrase it.
Fact 21.5 discusses Hamilton circuits for UNdirected graphs; HW6.3 talks about HAMCIRCUIT for directed graphs. I wouldn't expect this to make a difference, as undirected is a special case of directed, but I wanted to make sure I could utilize Fact 21.5 with impunity in my proof of question 3...
For the homework problem you actually only care that directed-HC is in NP, which is pretty obvious. You may take that as given. Once you have that, as you say, the NP-completeness of directed-HC follows immediately from that for undirected, as you can reduce undirected-HC to directed-HC by what amounts to a type-cast function (replace each undirected edge by two directed edges).
I'm sure I missed this point in the class, so pardon me for asking this again: in some other material, I read about NP-complete problems and they talk about reduction in poly-time, but we always talked about log-space. It's still not clear to me how these two are corelated.
In 611 and in most other algorithms courses NP is defined
in terms of poly-time reductions. In this course it is
defined in terms of log-space reductions. The problems
proved to be NP-complete with either definition are also
NP-complete under the other, because in all cases we study
the reductions are the same function and only log-space is
needed to compute them.
It is likely that there are languages that are 611-NP-complete
but not 601-complete, but certainly this could only happen
if P were different from L, which we believe but can't prove.
If there is a "one-way function" h that is computable in P but
whose inverse is not computable in L, then h(SAT) is 611-NP-complete
but not 601-NP-complete.
I have a couple of questions about hw 6.
The solutions to 3-6 seem too easy and that is making me uncomfortable.
Questions 3 and 6 seem to have the same simple solution of [redacted]. This seems too easy, especially since it answers two questions.
Hey, you know, sometimes I have trouble coming up with original questions. You want to explain this pretty thoroughly, but that's really the only idea, yes.
4 obeys the triangle inequality so we know there is a 1/2 approx from lecture.
Hmm... I didn't see that, but actually the argument I had in mind was even simpler.
5 is just [redacted]
Yeah.
Is there something I am missing?
'Fraid not, I'll try to make #7 more challenging. I'm sure 611 asked harder questions on this material.
In the homework, question 2 may have a typo in it, where you say Member-MP instead of Member-NP... unless there's a problem called Member-MP, too?
That was a typo that I've now fixed, thanks.
Question 4 on the homework says that we need to show that under the provided assumptions, TSP has a 1-approximation algorithm (based on the definition of a n-approximation in lecture 21). I'm assuming that the definition in lecture 21 is the same as the one given in the spring lecture on approximation (i.e. abs(c(M(x)) - OPT(x))/max(OPT(x),c(M(x))) <= n).
In the spring lecture it says that this is only valid for values less than one. So a 1-approx. doesn't fit the definition. Do we just ignore that constraint in the definition? Assuming we do this, saying that we have a 1-approximation really isn't saying anything by this definition. For example, if an approximation approximates a solution of A and the optimal solution is B, and assuming both are positive integers, then we know that (abs(A-B)/max(A,B)) < 1 in all cases. It seems with the provided assumptions we can easily provide a 1/2-approximation. Maybe the definition in lecture 21 is different then the Spring version, and maybe I'm completely misunderstanding it. Please explain, clarify or verify that this is in fact what you are looking for.
Yes, I screwed up, I meant a 1/2-approximation using this definition. I've fixed the assignment page and I mentioned this in class on 22 April. As I also said then, different definitions of the approximation parameter ε are sometimes used -- in a given paper you may have to read carefully to find out which is meant.
Last modified 25 April 2004