From file: lecture1.html
-
There have been many proofs that were claimed
to be correct for the four color theorem. Well..
Is
this one correct? I certainly can't tell at this point. So,
(attempt to) shed some light.
-
This is an approach to computer verification of mathematical proofs.
One such approach has lent credence to the four color theorem's proof.
If you are into verification, perhaps you would like to make an excelle
presentation of this.
From file: lecture2.html
-
VLSI layout. Perhaps compare simple approaches to VLSI layout. For
example, local improvement versus finding partitions and recursing
versus laying out by the first two eigenvectors. I would do a
"simplified layout" where one simply places nodes on a grid and
attempts to minimize the total wirelength. Another complication
is that in VLSI, one actually routes "nets", i.e., hyperedges
that consist of more than two "endpoints" that should be connected
perhaps by a Steiner tree (which again is difficult.) I can give
some suggestions or not. (One suspects that local improvement should be done
afterwards on any approach.)
Try it on the networks at Charles
Alpert's Benchmark. Or if you want to restrict yourself
to graphs, try these
-
Theoretical. A generalization of planarity is the notion of minor
exclusion. That is, a planar graph excludes $K_5$. There is
a separator theorem for any graph that excludes $K_h$ where
the size is $O(h^{3/2} \sqrt{n})$ by Alon, Seymour,
Thomas. . One can modify this construction as suggested
here
and get $O(h \sqrt{n \log n}).$ Can you get $O(h \sqrt{n})$?
Warning: this may not be too easy, but understanding the previous
is fine as well.
-
There is a lovely proof of the planar separator theorem using maximum
flow, understand and give a presentation of this proof. Speculate
on extending this to higher genus surfaces.
From file: lecture3.html
-
Try it! Use the eigenvalue method and compare it for finding
sparse cuts in these.
You can compare to METIS. Probably, you will get killed (and be slow)
compared to METIS. Explain why.
-
Evilness. Construct a family of examples that makes an eigenvalue
approach or METIS perform extremely badly (is not within any constant
factor of optimal.) Generate them, and show me. Try to understand
why.
-
Theory. Read and understand this paper. Perhaps
make some progress on the main conjecture.
-
Understand and extend, experiment with ideas in this paper
on rounding spectral cuts.
From file: lecture4.html
-
See lecture3.
From file: lecture5.html
-
See lecture3.
From file: lecture6.html
-
Read and understand and present
this paper. There is a followup paper by Hagerup that
is perhaps easier to understand.
-
Understand and present the state of the art in predecessors. Perhaps start
here.
From file: lecture7.html
-
Probably impossible but improve the following here.
If no improvement, you fail. Well, just kidding, make a presentation
of it and its predecessor (Karger, Benzcur.)
From file: lecture8.html
-
Make a presentation of $O(\log n)$ amortized complexity implementation
of dynamic trees.
From file: lecture9.html
-
Use the flow based algorithm to write code for image
segmentation. Compare to previous results. Perhaps start
here
to get graphs of appropriate edge weights from various
images and for comparison.
-
Read, understand, make a presentation of
Orlin's strongly polynomial time algorithm for min cost flow.
From file: lecture10.html
- Present a survey on the state of the art regarding the Hirsch
conjecture (the diameter of the simplex is $n-d.$).
From file: lecture12.html
- Use these ideas in lecture 12 to code a VLSI routing algorithm to minimize the
maximum congestion in any track. Extra points for dealing with
multiway nets. Again, see
Charles
Alpert's Benchmark for examples.
From file: lecture13.html
- Understand and discuss whether the $\sqrt{n}$ iterations per drop
in potential is necessary. What makes it necessary. Perhaps, start
here where they suggest $n^{.25}$ may happen reasonably often.
My intuition is that the bad case for the primal step and
the bad case for the dual may be better than what it seems.
That is, the primal only makes a small drop if there is a single
large dimension which restricts how far you can go. Perhaps a
different direction is ok. Similarly, the Cauchy-Swartz limit
in the dual step seems good when there are a few large directions. Do
I have a sign error?
- Can you explain why this algorithm gets exponential convergence
where the experts algorithm gets polynomial convergence for two
person zero sum games. That is the
latter takes $O(1/\epsilon^2)$ iterations to get within
$(1+\epsilon)$ whereas the former takes $O(\log (1/\epsilon))$ along
with various dependencies on $n$ and $m$. See the surveys by Goemans
and Arora on the two methods on the course webpage.