Barriers 2

Later today, I'll be giving a talk at the 2nd Barriers Workshop in Princeton.

Here's my attempt to survey data structure lower bounds in 70 minutes: PDF, PPSX.

Update: Based on comments, I'm publishing the slides as PPSX (which can be viewed with a free viewer) and PDF. I will try to convert my other talks to these formats when I have time.

IOI: A Medium Problem

Here is another, medium-level problem from the IOI. (Parental advisory: this is not quite as easy as it may sound!)

I think of a number between 1 and N. You want to guess the secret number by asking repeatedly about values in 1 to N. After your second guess, I will always reply "hotter" or "colder", indicating whether your recent guess was closer or farther from the secret compared to the previous one.

You have lg N + O(1) questions.

***

The solution to the first problem I mentioned can be found in the comments. Bill Hesse solved the open question that I had posed. He has a neat example showing that the space should be N^1.5log₂3 bits, up to lower order terms. It is very nice to know the answer.

A very elegant solution to the second problem was posted in the comments by a user named Radu (do you want to disclose your last name?). Indeed, this is simpler than the one I had. However, the one I had worked even when the numbers in the array were arbitrary (i.e. you could not afford to sort them in linear time). I plan to post it soon if commenters don't find it.

IOI: Another Hard Problem

You are given a matrix A[1..N][1..M] that contains a permutation of the numbers {1, ..., NM}. You are also given W≤N and H≤M. The goal is to find that rectangle A[i ... i+W][j ... j+H] which has the lowest possible median.

Running time: O(NM).

***

This could have been another hard problem at the IOI, but it was decided to give maximum score to solutions running in O(NM lg(NM)). This is considerably easier to obtain (but still interesting).

In the style of what Terry Tao tried for the IMO, I am opening this blog post as a discussion thread to try to solve the problem collectively ("poly-CS" ?). You are encouraged to post any suggestions, half-baked ideas, trivial observations, etc – with the goal to jointly reach a solution. If you think about the problem individually and find the solution, please don't post, as it will ruin the fun of the collective effort. Following this rule, I will not engage in the discussion myself.

I did not encourage a discussion for the first problem, since it was the kind of problem that only required one critical observation to solve. This second problem requires several ideas, and in fact I can see two very different solutions.

IOI: The Hard Problem

The International Olympiad in Informatics (IOI 2010) is taking part this week at the University of Waterloo, Canada.

The olympiad often features a hard problem, which is intended to be solved by a handful of contestants. This year, the problem was solved by about 6 people. Read the problem below and give it a shot! :)

I will describe the problem in both TCS and IOI fashion.

Asymptotic version. You are given an unweighted, undirected graph on N vertices. Some sqrt(N) vertices are designated as "hubs". You have to encode the pairwise distances between all hubs and all vertices in O(N^1.5) bits of space.

The encoder and decoder may run in polynomial time. Of course, the decoder does not see the original graph; it receives the output of the encoder and must output the explicit distances between any hub and any other vertex. (This list of explicit distances takes O(N^1.5lg N) bits.)

Non-asymptotic version. You are given a graph on 1000 nodes and 36 designated hubs. You have to encode the distances between all hubs and all vertices in 70,000 bits of space.

The non-asymptotic version is a bit harder, since you have to pay more attention to some details.

The research version. Prove or disprove that the distances can be encoded using (1+o(1)) N^1.5 bits of space. I don't know the answer to this question (but I find the question intriguing.)

A taxonomy of range query problems

In this post, I will try to enumerate the range query problems that I know about. Let me know if I'm missing anything.

The query. Say you have n points in the plane, and you query for the points in an axis-parallel rectangle. What could we mean by "query"?

• existential range queries: Is there any point in the rectangle?
• counting queries: How many points are there in the rectangle?
• reporting queries: List the points in the rectangle. Unlike the previous cases, the query time is now broken into two components: it is usually given as f(n) + k*g(n), where k is the number of output points.

Now let's assume the points have some number associated to them (a weight or a priority). Then one could ask the following natural queries:

• weighted counting: What is the total weight of the points inside?
• range min (max) query
• range median query. (Possible generalizations: selection or quantiles.)
• top-k reporting: Report just the top k guys, by priority (for k given). One may demand the output to be sorted. More stringently, one may ask the query algorithm to enumerate points sorted by priority, in time g(n) per point, until the user says "stop."

The number associated to a point can also be a "color". For instance, points could represent Irish pubs / Belgian bars / etc, and we may only want one point of each type in our output. Then the queries become:

• colored counting: How many distinct colors are in the rectangle?
• colored reporting: List the distinct colors in the rectangle (possibly with one example point from each color).
• top-k colored reporting: If the colors are sorted by priorities (e.g. I prefer points of color 1 over points of color 2), one can then ask for the top-k distinct colors inside the rectangle.

Dynamism. The problem could be:

• static: Preprocess the point set and then answer queries.
• dynamic: Insert and delete from the point set.
• incremental / decremental: We only insert or delete.
• offline: The sequence of operations is known in advance. This is enough for many applications to algorithms.
• parametric / kinetic. I confess ignorance with respect to these.

Orthogonal range queries. The setting from above works in any number of dimensions d≥1: the data set consists of n points in d-dimensional space and the query is a box [a₁, b₁]×···×[a_d, b_d]. This setup is usually called "orthogonal range queries".

We can consider the following important restrictions on the query:

• dominance queries: the box is [0, b₁]×···×[0, b_d]. In other words, we are asking for the points dominated, coordinate-wise, by a point (b₁, ..., b_d).
  
  
• k-sided queries: exactly 2d-k values in (a₁, a₂, ..., a_d) are zero. For instance, a 3-sided query in 2D is a rectangle with one side on the x axis. Dominance queries are the special case of d-sided queries.

The universe. The coordinates of the points and queries can come from the following sets:

• general universe. In the standard RAM model, we assume that the coordinates are integers that fit in a machine word.
  
  
• rank space: the coordinates are from {1, 2, ..., n}. One can reduce any static problem to rank space by running 2d predecessor queries. Most problems can be shown to be at least as hard as predecessor search, so their complexity is precisely: "optimal cost of predecessor search" + "optimal cost for the problem in rank space". In other words, for most problems it is sufficient to solve them in rank space.
  
  
• dense universe: the points are exactly the points of the grid [1, n₁]×···×[1, n_d] where n₁n₂···n_d = n. In 1D this is the same as rank space, but in 2 or more dimensions the problems are very different. (To my knowledgeable readers: Is there a standard name for this case? For counting queries people call this "the partial sums problem", but how about e.g. min queries?)

For dynamic problems, the "rank space" changes when a new coordinate value is inserted. Thus, a rank-space solution must support a "insert value" operation that increments all coordinate values after a given one, creating space for a newly inserted point. (I have heard the term "list space" for this. Should we just use "rank space"?)

Stabbing. So far, our data set consisted of points and the queries asked for points in a given rectangle. Conversely, one can consider a data set of rectangles; the query is a point and asks about the rectangles containing that point ("stabbed" by it). This problem is important, among others, in routing: we can have rules for packets coming from some range of IP addresses and going to some other range of IP addresses.

The notion of rank space, and all query combinations still make sense. For instance, interval max stabbing is the following problem: given a set of interval (in 1D) with priorities, return the interval of highest priority stabbed by a query point.

Note that the rectangles in the input can intersect! If we ask that the rectangles be disjoint, or more stringently, be a partition of the space, we obtain the point location problem.

Rectangle-rectangle queries. So far we looked at containment relations between rectangles and points. More generally, the data set can consist of rectangles, and the query can also be a rectangle. Then one can ask:

• intersection queries: analyze the set of input rectangles that intersect the query rectangle.
• containment queries: analyze the set of rectangles that contain / are-contained-by the query.

Two important special cases arise when the rectangles degenerate to line segments:

• orthogonal segment intersection: Given a set of horizontal segments, find the ones that intersect a vertical query segment.
  
  
• orthogonal ray shooting: Given a set of horizontal segments, find the lowest one immediately above a query point. In other words, consider a vertical ray leaving from the point and find the first segment it intersects. (This is the min segment intersection query, where the priority of each horizontal segment is its y coordinate.)

More complex geometry. Of course, our ranges need not be orthogonal. One can consider:

• balls
• arbitrary lines
• half spaces
• simplices (e.g 2D triangles).

In non-orthogonal geometry, the concept of rank space disappears. However, most other combinations are possible. For instance, one could ask about the points in a query range; the ranges stabbed by a query point; the ranges intersecting a query range; the first rangeintersected by a ray; etc. We can ask existential, counting, or reporting questions, on ranges that can have weights or priorities.

SODA

Being on the SODA PC is excruciating. Too many submissions are at the level of a hard exercise – things that David Karger would assign in Advanced Algorithms or Randomized Algorithms as homework. And since I'm a problem solver by nature, I cannot resist solving the problems before (in lieu of) reading the papers...

I am left wondering how life on the ICS PC is like. The fact that the PC members can solve the problem in half an hour is promised to not be a negative – so is there any point in engaging the technical side of your brain during the review process?

Which brings me to graph algorithms, a field I have engaged in sporadically. This field strikes me as problem solving par excellence. You think about the problem, you pull a rabbit out of the hat, and run around naked screaming "Εὕρηκα!" Yet most reviewers (which, I will assume, write graph-theory papers in the same way) cannot help commenting on the lack of new techniques in the solution. I interpret this as a code for "We didn't care to think about the problem, we just read over your solution and remarked that it solved the problem by looking at some edges in some order, then at some trees and some vertices."

(I should say I have no sour grapes about this... Graph algorithms in not my main field and all my papers on the topic are in STOC/FOCS. Yet I am amazed by how much the field sabotages itself with this "techniques thing.")

By contrast, I once came across a reviewer that actually wore his problem-solver hat while reading my paper. The review read "I spent days to figure out the interplay and indispensability of the data structures and techniques used. But after fully understanding the entire picture and the interplay of all the data structures, I felt very satisfied. It is indeed a very beautiful piece of algorithmic art. I congratulate the authors for such a novel algorithm." This is possibly the best review I ever got – complemented, of course, by some "no new techniques but important result" reviews.

Which brings me to the main trouble with TCS, in my opinion. If only we could stop inventing a ton of artificial problems every year, and we actually cared about solving some age-old clear-cut problems – then it would be no surprise to find reviewers that have actually thought carefully about your hard problem.

3SUM Puzzle

A puzzle courtesy of Mohan Paturi:

Given a set S of n numbers, the 3SUM problem asks whether there exist a,b,c ∈ S such that a+b+c=0. It is generally believed that this should take around n² time.

But let's assume |S+S| = t ≪ n². Show that the problem can be solved in O(t lg²n) time.

Any ideas for a better bound?