Please submit...

1. Papers to MASSIVE 2011. This will be a SoCG satellite workshop (think Paris in June...). Publication here does not hinder publication anywhere else.
   
   
2. Algorithmic problems for the Balkan Olympiad (BOI 2011). I am chairing the Scientific Committee, and we need your help in putting together an interesting set of problems. If your problem is selected for the competition, you will be invited to be a member of the Scientific Committee, and you get a free trip to the Olympiad if you accept (think Bistrița in July...). Anybody can submit problems, but I think the free trip only applies if you're starting somewhere in Romania and you are a Romanian/EU^(?) citizen. Hence the call for problems is in Romanian; sorry.
   
   
3. Comments to other blogs. This blog is now on full moderation after I had to delete a batch of comments that did not score particularly high on the sanity scale. Don't worry, the main point of this is not censorship; nasty but coherent comments will continue to be accepted.

Complexity Theory

If you're a student thinking of going into complexity theory, take a good look around and ask yourself: "Do I really want to be in a field with this amount of groupthink?" [1,2,3,4,5, and last but certainly not least 6]

Back in the day, I was myself saved from a career in complexity by Omer Reingold's logspace connectivity.

FOCS 2010

I am back from FOCS 2010, where I gave two talks:

• a tutorial on data structure lower bounds: PPSX, PDF
• a regular conference talk on distance oracles: PPSX, PDF (for this paper coauthored with Liam Roditty).

This paper caused some amount of fun in certain "inner circles", which I think I should share with a wider audience. If you read the paper, the algorithms repeatedly grow balls (aka shortest path trees) around vertices of the graph. After obsessing about growing these balls for more than a year, I found it natural to name the paper "How to Grow Your Balls". At least it allowed me to begin various talks by telling the audience that, "This is a topic of great economic importance; I receive email about it almost every day."

The committee took its role as Cerberus at the gates of Theory very seriously, and refused to accept the paper under the original title. In the process, many jokes were made, at least some of which transpired outside the PC room. I won't post them here, as the PC probably intends to keep some decorum. (But if you get a chance, ask your friend on the PC about that proposed proceedings cover.)

So the title changed, but I stretched the joke a bit further by titling the tutorial "How to Grow Your Lower Bounds".

This episode confirmed two things that I already knew. First, this community insists on taking itself way too seriously. For comparison, the title certainly triggered some PR alarms inside AT&T. But it made it through, perhaps because I told people that it's in the interest of AT&T Labs to prove publicly that it's a real research lab: it lets researchers be researchers, which includes being silly at times. Given the reputation of PR departments in American telecoms, it's a bit sad to find that the theory community has a higher bar on form.

Secondly, I should turn it down a notch with this blog. Several people said they had expected me to withdraw the paper rather than change the title. I found it quite amazing that people had such an image of me. Perhaps it is time to emphasize that I'm doing my research because I think it's important and cool. The horsing around is, well, not quite the main message.

Problem solving versus new techniques

This is a guest post by Mikkel Thorup:

--------------------------

I think there is nothing more inhibiting for problem solving than referees looking for new general techniques.

When I go to STOC/FOCS, I hope to see some nice solutions to important problems and some new general techniques. I am not interested in semi-new techniques for semi-important problems. A paper winning both categories is a wonderful but rare event.

Thus I propose a max-evaluation rather than a sum. If the strength of a paper is that it solves an important problem, then speculations on the generality of the approach are of secondary importance. Conversely, if the strength of the paper is some new general techniques, then I can forgive that it doesn't solve anything new and important.

One of the nice things about TCS is that we have problems that are important, not just as internal technical challenges, but because of their relation to computing. At the end of the day, we hope that our techniques will end up solving important problems.

Important problems should be solved whatever way comes natural. It may be deep problem specific understanding, and it may build on previous techniques. Why would we be disappointed if an old problem got solved by a surprising reuse of an old technique?

Elegant reuse of old techniques is not a bad thing. After all, when we develop new techniques, we are hoping they will be reused, and from a teaching perspective, it is great to discover that the same general technique applies to more diverse problems. The last thing we want is to encourage authors to hide reuse. Of course, reuse should be highly non-obvious for a strong conference, but for established open problems, non-obviousness will normally be self-evident.

Retrieval-Only Dictionaries

We saw two cool applications of dictionaries without membership; now it's time to construct them. Remember that we are given a set S, where each element x∈S has some associated data[x], a k-bit value. We want a data structure of O(nk) bits which retrieves data[x] for any x∈S and may return garbage when queried for x∉S.

A conceptually simple solution is the "Bloomier filter" of [Chazelle, Kilian, Rubinfeld, Tal SODA'04]. This is based on the power of two choices, so you should first go back and review my old post giving an encoding analysis of cuckoo hashing.

Standard cuckoo hashing has two arrays A[1..2n], B[1..2n] storing keys, and places a key either at A[h(x)] or B[g(x)]. Instead of this, our arrays A and B will store k-bit values (O(nk) bits in total), and the query retrieve-data(x) will return A[h(x)] xor B[g(x)].

The question is whether we can set up the values in A and B such that any query x∈S returns data[x] correctly. This is a question about the feasibility of a linear system with n equations (one per key) and 4n variables (one per array entry).

Consider a connected component in the bipartite graph induced by cuckoo hashing. If this component is acyclic, we can fix A and B easily. Take an arbitrary node and make it "zero"; then explore the tree by DFS (or BFS). Each new node (an entry in A or B) has a forced value, since the edge advancing to it must return some data[x] and the parent node has been fixed already. As the component is acyclic, there is only one constraint on every new node, so there are no conflicts.

On the other hand, if a component has a cycle, we are out of luck. Remark that if we xor all cycle nodes by some Δ, the answers are unchanged, since the Δ's cancel out on each edge. So a cycle of length k must output k independent data values, but has only k-1 degrees of freedom.

Fortunately, one can prove the following about the cuckoo hashing graph:

• the graph is acyclic with some constant probability. Thus, the construction algorithm can rehash until it finds an acyclic graph, taking O(n) time in expectation.
  
  
• the total length of all cycles is O(lg n) with high probability. Thus we can make the graph acyclic by storing O(lg n) special elements in a stash. This gives construction time O(n) w.h.p., but the query algorithm is slightly more complicated (for instance, it can handle the stash by a small hash table on the side).

These statements fall out naturally from the encoding analysis of cuckoo hashing. A cycle of length k allows a saving of roughly k bits in the encoding: we can write the k keys on the cycle (klg n bits) plus the k hash codes (klg(2n) bits) instead of 2k hash codes (2klg(2n) bits).

Further remarks. Above, I ignored the space to store the hash functions h and g. You have to believe me that there exist families of hash functions representable in O(n^ε) space, which can be evaluated in constant time, and make cuckoo hashing work.

A very interesting goal is to obtain retrieval dictionaries with close to kn bits. As far as I know, the state of the art is given by [Pagh-Dietzfelbinger ICALP'08] and [Porat].

Static 1D Range Reporting

Method 4 for implementing van Emde Boas with linear space, described in my last post, is due to [Alstrup, Brodal, Rauhe: STOC'01]. They worked on static range reporting in 1 dimension: preprocess a set of integers S, and answer query(a,b) = report all points in S ∩ [a,b]. This is easier than predecessor search: you can first find the predecessor of a and then output points in order until you exceed b. Using van Emde Boas, we would achieve a linear-space data structure with query time O(lglg u + k), where k is the number of points to be reported.

Alstrup, Brodal, and Rauhe showed the following surprising result:

Static 1D range reporting can be solved with O(n) space and O(1+k) query time.

I like this theorem a lot, since it is so easy to describe to anybody with minimal background in Computer Science, yet the result is not obvious. I have used it many times to answer questions like, "Tell me a surprising recent result from data structures."

The solution. We need a way to find some (arbitrary) key from S ∩ [a,b] in constant time. Once we have that, we can walk left and right in an ordered list until we go outside the interval.

Let's first see how to do this with O(n lg u) space; this was described by [Miltersen, Nisan, Safra, Wigderson: STOC'95]. Of course, we build the trie representing the set. Given the query [a,b] let us look at the lowest common ancestor (LCA) of a and b. Note that LCA(a,b) is a simple mathematical function of the integers a and b, and can be computed in constant time. (The height of the LCA is the most significant set bit in a xor b.)

• if LCA(a,b) is a branching node, look at the two descendant branching nodes. If the interval [a,b] is nonempty, it must contain either the max in the tree of the left child, or the min in the tree of the right child.
  
  
• if LCA(a,b) is an active node, go to its lowest branching ancestor, and do something like the the above.
  
  
• if LCA(a,b) is not an active node, the interval [a,b] is certainly empty!

Thus, it suffices to find the lowest branching ancestor of LCA(a,b) assuming that LCA(a,b) is active. This is significantly easier than predecessor search, which needs the lowest branching ancestor of an arbitrary node.

The proposal of [Miltersen et al.] is to store all O(n lg u) active nodes in a hash table, with pointers to their lowest branching ancestors.

As in my last post, the technique of [Alstrup et al.] to achieve O(n) space is: store only O(n √lg u) active nodes, and store them in a retrieval-only dictionary with O(lglg u) bits per node. We store the following active nodes:

1. active nodes at depth i·√lg u ;
2. active nodes less than √lg u levels below a branching node.

We first look up LCA(a,b) in the dictionary. If the lowest branching ancestor is less than √lg u levels above, LCA(a,b) is in the dictionary and we find the ancestor. If not, we truncate the depth of the LCA to a multiple of √lg u , and look up the ancestor at that depth. If [a,b] is nonempty, that ancestor must be an active node and it will point us to a branching ancestor.

vEB Space: Method 4

In the previous post I described 3 ways of making the "van Emde Boas data structure" take linear space. I use quotes since there is no unique vEB structure, but rather a family of data structures inspired by the FOCS'75 paper of van Emde Boas. By the way, if you're curious who van Emde Boas is, here is a portrait found on his webpage.

In this post, I will describe a 4th method. You'd be excused for asking what the point is, so let me quickly mention that this technique has a great application (1D range reporting, which I will discuss in another post) and it introduces a nice tools you should know.

Here is a particularly simple variant of vEB, introduced by Willard as the "y-fast tree". Remember from the last post that the trie representing the set has n-1 branching nodes connected by 2n-1 "active" paths; if we know the lowest branching ancestor of the query, we can find the predecessor in constant time. Willard's approach is to store a hash table with all O(n lg u) active nodes in the trie; for each node, we store a pointer to its lowest branching ancestor. Then, we can binary search for the height of the lowest active ancestor of the query, and follow a pointer to the lowest branching node above. As the trie height is O(lg u), this search takes O(lglg u) look-ups in the hash table.

Of course, we can reduce the space from O(n lg u) to O(n) by bucketing. But let's try something else. We could break the binary search into two phases:

1. Find v, the lowest active ancestor of the query at some depth of the form i·√lg u (binary search on i). Say v is on the path u→w (where u, w are branching nodes). If w is not an ancestor of the query, return u.
   
   
2. Otherwise, the lowest branching ancestor of the query is found at some depth in [ i·√lg u , (i+1)√lg u ]. Binary search to find the lowest active ancestor in this range, and follow a pointer to the lowest active ancestor.

With this modification, we only need to store O(n √lg u ) active nodes in the hash table! To support step 1., we need active nodes at depths i·√lg u. To support step 2., we need active nodes whose lowest branching ancestor is only ≤ √lg u levels above. All other active nodes can be ignored.

You could bring the space down to O(n lg^εu) by breaking the search into more segments. But to bring the space down to linear, we use heavier machinery:

Retrieval-only dictionaries. Say we want a dictionary ("hash table") that stores a set of n keys from the universe [u], where each key has k bits of associated data. The dictionary supports two operations:

• membership: is x in the set?
• retrieval: assuming x is in the set, return data[x].

If we want to support both operations, the smallest space we can hope for is log(u choose n) + nk ≈ n(lg u + k) bits: the data structure needs to encode the set itself, and the data.

Somewhat counterintuitively, dictionaries that only support retrieval (without membership queries) are in fact useful. (The definition of such a dictionary is that retrieve(x) may return garbage if x is not in the set.)

Retrieval-only dictionaries can be implemented using only O(nk) bits. I will describe this in the next post, but I hope it is believable enough.

When is a retrieval-only dictionary helpful? When we can verify answers in some other way. Remember the data structure with space O(n √lg u ) from above. We will store branching nodes in a real hash table (there are only n-1 of them). But observe the following about the O(n √lg u ) active nodes that we store:

1. We only need k=O(lglg u) bits of associated data. Instead of storing a pointer to the lowest branching ancestor, we can just store the height difference (a number between 1 and lg u). This is effectively a pointer: we can compute the branching ancestor by zeroing out so many bits of the node.
   
   
2. We only need to store them in a retrieval-only dictionary. Say we query some node v and find a height difference δ to the lowest branching ancestor. We can verify whether v was real by looking up the δ-levels-up ancestor of v in the hash table of branching nodes, and checking that v lies on one of the two paths descending from this branching node.

Therefore, the dictionary of active nodes only requires O(n √lg u · lglg u) bits, which is o(n) words of space! This superlinear number of nodes take negligible space compared to the branching nodes.