And the meaning of the recently-broken metric TSP approximation ratio

Andrew Stothers wrote in 2007 a first-year qualifying report for the University of Edinburgh PhD programme. In only eleven pages he presented the parts of the famous paper by Don Coppersmith and Shmuel Winograd (CW) that do enough to obtain for the exponent of matrix multiplication. In a few more pages he surveyed recent work on sufficient conditions for . Page 15, the last except for references, laid out “Possible Directions for Future Work” in five sentences, the fifth being:

Finally, it will also be possible to investigate bilinear algorithms further and use more complex means to determine whether or not “better” algorithms exist for matrix multiplication.

Today we wish to explain what means, why lower is always “better,” why we feel our field is better for the great efforts expended by Stothers and Virginia Vassilevska Williams, and why it is better when intellectual values guide our judgment. We also wish to defend ourselves and others for calling the results “breakthroughs,” comparing and contrasting to recent advances on algorithms for the Traveling Salesman Problem.

I (Ken) wrote a similar first-year qualifier in summer 1982 after my first year at Oxford—they are a staple of leading British universities. Like Andrew’s it was not obligated to have original research, just to demonstrate understanding and capability. One item found its way into a conference paper, but my 1986 dissertation ranged into various other topics.

What distinguishes Andrew’s Nov. 2010 dissertation is its singular focus on accomplishing the objective of his proposal. And doing it. The resulting paper with his advisor Sandy Davie has recently been submitted. Meanwhile in North America, Virginia had been thinking about the same problem since 2005. It may not be true as asserted here that “hundreds of people tried to improve” it after 1987, but we’ll guess about as many people thought about it. It’s enough to note that many researchers felt it worthy of pursuit, including Andrew’s advisor, and that progress took large concerted effort. The nature of this effort—including the use of computers—may be an important harbinger for science by itself.

What Does ω Mean?

The definition of is the infimum of all such that matrix multiplication has an algorithm that runs in time . This is in a classical model of computation where addition and multiplication of scalars have unit cost. Note that there might not be an algorithm that runs in time itself— may properly be a limit even if it turns out to have the least possible value 2. Currently all we know is that from Virginia’s work.

As many have noted, does not have immediate practical relevance because there is evidently a tradeoff with the constant hiding under the in . The tradeoff is so far steep enough that apparently only Volker Strassen’s relatively simple algorithm achieving an exponent of has wide use. So why do we care about values of less than that?

One of the central questions in physics is, why do the fundamental constants of nature have the values they do? The desire for why, in the face of evidence that the values could be arbitrary, has aroused such passion that string theorist David Gross channeled Winston Churchill in exhorting young physicists to “never, never, never give up.”

Now in mathematics it may seem nonsensical to ask, why do numbers like and have the values they do? But with , we are in a sense closer to physics: understanding it is akin to discovering a natural law, at least a law of information. And quantum mechanics has if anything magnified William Hamilton’s vision approaching 200 years ago that Nature computes with matrices, yes large ones. It is possible that detracts from a better statement of important laws, but knowing better brings us closer to them all the same.

Barriers and Breakthroughs

Hence also the interest in whether takes a value with known meaning, such as or (both falsified) or , or as some have mused, or , or as some have argued more likely for exponents, a logarithm of some higher and simpler number. Strassen himself has often stated his belief that is strictly greater than . If it has a value not previously seen in mathematics, then we could hope to discover new mathematical regularities as well; if it has a known value, then this may yield some more explanation about algorithms.

The value seemed a natural possible barrier, but the time interval from to in 1981–82 was very short. By contrast the “CW” bound of was not a natural-seeming number, but it withstood attempts to scale it for 23 years, including a year’s work from each of Virginia and Andrew. Reasons of intellectual judgment, expanded below, we feel are enough to justify calling their final results a “breakthrough.” But we offer a recent concrete case for comparison first.

The Salesman Always Rings 1.5

The Traveling Salesman Problem (TSP) seeks the shortest route to visit each of a given set of sites, coming back in a ring to the starting site. When the sites are in real space and “shortest” refers to Euclidean distance, finding the length of the absolute shortest route remains -hard even in the plane, as discussed here. However, Nicos Christofides in 1976 found a polynomial-time algorithm that finds a tour guaranteed to have total length at most . The method and proof apply to various other cases where the distances between sites obey inequalities that define some kind of metric, with the same bound.

This stood as a barrier for most of these cases, until this year. Shayan Oveis Gharan, Amit Saberi, and Mohit Singh (OSS) broke it for a wide subclass of these problems. Well they “broke” it by achieving a guaranteed ratio to the optimum of

Actually their original paper did not give a value—it merely proved the existence of an giving . The estimate for the magnitude of their breakthrough was supplied later.

Improvements and Predictions

It must be said that the OSS paper and its predecessor introduced techniques that were more novel to the problem than is evident for the new matrix-multiplication papers, and the predecessor won the SODA 2010 “Best Paper” award. However, a referee could have wondered, why bother moving heaven and earth in a restricted case to demonstrate an increment a thousand times smaller than “nano”? Indeed the version of OSS linked above still ends with “” on page 65.

What happened is that the change attracted interest on several continents, and the people in Europe who made a bankable improvement are not even in the euro zone. First Tobias Mömke and Ola Svensson of KTH in Stockholm obtained by using matchings in place of the more-complicated ingredients of OSS, also in time for FOCS 2011. Then Marcin Mucha of the University of Warsaw improved their analysis to obtain

It is noted by all that an even older paper by Michael Held and Richard Karp from 1970 had set a believable target for provable improvements of Christofides’ bound, namely a conjectured ratio of from linear programming relaxations. However, also has dueling conjectures, some supporting a believable target of . The real point of comparison with TSP is that suddenly there was progress on a barrier that had stood for much of the age of the field, and this attracted others to try for more.

We note that Markus Bläser has contended in a comment that the extension of CW used by both new papers has limitations, and we infer that some other experts concur. However, the paper by Vassilevska Williams in particular has two new ingredients: a framework for managing any tensor power as a base algorithm, and a computer implementation of the framework. She also employed an insight of Stothers to simplify the analysis. Though we can imagine a geometrical insight that higher powers bring returns that “shrink geometrically,” can we really constrain the likelihood that pursuing them might dislodge a different insight that changes the whole game? Is there a limitation theorem here? All we know is that the game has changed, and perhaps the game is afoot.

Worth and Values

The last issue we wish to raise is the reliability of judging the worth of past achievement by present assessment of what it may or may not lead to in the future. We have posted several times on surprises and the difficulty of predicting or guessing how things will go. Of course impetus into the future is necessarily part of claiming something now a “breakthrough.”

However, if we seek a reliable and consistent standard of judging worth, we should use a value that the community has understood for many years: intellectual substance and effort. This value, plus the simple salience of the goal, led our principals to invest the years of work in the first place. Depth of thinking is our gold standard, while applicability is our paper currency. The improvements of and on paper for amid other issues may be a poor return now, but the new computer-assisted vein to mine may parlay true value later.

Dick and I hope this explains why fundamental effort and easily-stated achievement, on a fundamental problem after nearly a quarter-century of stasis, elicited the reaction it has. We agree with, and have tried to extend, Timothy Gowers’ comments here. As for how these values can be invested, only time will be the teller.

Open Problems

Can we be more open about the value of pursuing problems?

What is ?

Suppose we know something special about two matrices and to be multiplied, something not so obvious like their being sparse. Suppose in particular that we have a tiny circuit that for any outputs the value , and similarly for , where those values come from a fixed finite set. Or suppose we know that and/or preserve a similar succinctness property from argument vectors to values—which[…]