本文作者：小红猪小分队

Source：http://www.newscientist.com/article/mg21528771.100-the-algorithm-that-runs-the-world.html?full=true

Its services are called upon thousands of times a second to ensure the world’s business runs smoothly – but are its mathematics as dependable as we thought?

YOU MIGHT not have heard of the algorithm that runs the world. Few people have, though it can determine much that goes on in our day-to-day lives: the food we have to eat, our schedule at work, when the train will come to take us there. Somewhere, in some server basement right now, it is probably working on some aspect of your life tomorrow, next week, in a year’s time.

Perhaps ignorance of the algorithm’s workings is bliss. The door to Plato’s Academy in ancient Athens is said to have borne the legend “let no one ignorant of geometry enter”. That was easy enough to say back then, when geometry was firmly grounded in the three dimensions of space our brains were built to cope with. But the algorithm operates in altogether higher planes. Four, five, thousands or even many millions of dimensions: these are the unimaginable spaces the algorithm’s series of mathematical instructions was devised to probe.

Perhaps, though, we should try a little harder to get our heads round it. Because powerful though it undoubtedly is, the algorithm is running into a spot of bother. Its mathematical underpinnings, though not yet structurally unsound, are beginning to crumble at the edges. With so much resting on it, the algorithm may not be quite as dependable as it once seemed.

To understand what all this is about, we must first delve into the deep and surprising ways in which the abstractions of geometry describe the world around us. Ideas about such connections stretch at least as far back as Plato, who picked out five 3D geometric shapes, or polyhedra, whose perfect regularity he thought represented the essence of the cosmos. The tetrahedron, cube, octahedron and 20-sided icosahedron embodied the “elements” of fire, earth, air and water, and the 12-faced dodecahedron the shape of the universe itself.

Things have moved on a little since then. Theories of physics today regularly invoke strangely warped geometries unknown to Plato, or propose the existence of spatial dimensions beyond the immediately obvious three. Mathematicians, too, have reached for ever higher dimensions, extending ideas about polyhedra to mind-blowing “polytopes” with four, five or any number of dimensions.

A case in point is a law of polyhedra proposed in 1957 by the US mathematician Warren Hirsch. It stated that the maximum number of edges you have to traverse to get between two corners on any polyhedron is never greater than the number of its faces minus the number of dimensions in the problem, in this case three. The two opposite corners on a six-sided cube, for example, are separated by exactly three edges, and no pair of corners is four or more apart.

Hirsch’s rule holds true for all 3D polyhedra. But it has never been proved generally for shapes in higher dimensions. The expectation that it should translate has come largely through analogy with other geometrical rules that have proved similarly elastic (see “Edges, corners and faces“). When it comes to guaranteeing short routes between points on the surface of a 4D, 5D or 1000D shape, Hirsch’s rule has remained one of those niggling unsolved problems of mathematics – a mere conjecture.

How is this relevant? Because, for today’s mathematicians, dimensions are not just about space. True, the concept arose because we have three coordinates of location that can vary independently: up-down, left-right and forwards-backwards. Throw in time, and you have a fourth “dimension” that works very similarly, apart from the inexplicable fact that we can move through it in only one direction.

But beyond motion, we often encounter real-world situations where we can vary many more than four things independently. Suppose, for instance, you are making a sandwich for lunch. Your fridge contains 10 ingredients that can be used in varying quantities: cheese, chutney, tuna, tomatoes, eggs, butter, mustard, mayonnaise, lettuce, hummus. These ingredients are nothing other than the dimensions of a sandwich-making problem. This can be treated geometrically: combine your choice of ingredients in any particular way, and your completed snack is represented by a single point in a 10-dimensional space.

Brutish problems

In this multidimensional space, we are unlikely to have unlimited freedom of movement. There might be only two mouldering hunks of cheese in the fridge, for instance, or the merest of scrapings at the bottom of the mayonnaise jar. Our personal preferences might supply other, more subtle constraints to our sandwich-making problem: an eye on the calories, perhaps, or a desire not to mix tuna and hummus. Each of these constraints represents a boundary to our multidimensional space beyond which we cannot move. Our resources and preferences in effect construct a multidimensional polytope through which we must navigate towards our perfect sandwich.

In reality, the decision-making processes in our sandwich-making are liable to be a little haphazard; with just a few variables to consider, and mere gastric satisfaction riding on the outcome, that’s not such a problem. But in business, government and science, similar optimisation problems crop up everywhere and quickly morph into brutes with many thousands or even millions of variables and constraints. A fruit importer might have a 1000-dimensional problem to deal with, for instance, shipping bananas from five distribution centres storing varying numbers of fruit to 200 shops with different numbers in demand. How many items of fruit should be sent from which centres to which shops while minimising total transport costs?

A fund manager might similarly want to arrange a portfolio optimally to balance risk and expected return over a range of stocks; a railway timetabler to decide how best to roster staff and trains; or a factory or hospital manager to work out how to juggle finite machine resources or ward space. Each such problem can be depicted as a geometrical shape whose number of dimensions is the number of variables in the problem, and whose boundaries are delineated by whatever constraints there are (see diagram). In each case, we need to box our way through this polytope towards its optimal point.

This is the job of the algorithm.

Its full name is the simplex algorithm, and it emerged in the late 1940s from the work of the US mathematician George Dantzig, who had spent the second world war investigating ways to increase the logistical efficiency of the US air force. Dantzig was a pioneer in the field of what he called linear programming, which uses the mathematics of multidimensional polytopes to solve optimisation problems. One of the first insights he arrived at was that the optimum value of the “target function” – the thing we want to maximise or minimise, be that profit, travelling time or whatever – is guaranteed to lie at one of the corners of the polytope. This instantly makes things much more tractable: there are infinitely many points within any polytope, but only ever a finite number of corners.

If we have just a few dimensions and constraints to play with, this fact is all we need. We can feel our way along the edges of the polytope, testing the value of the target function at every corner until we find its sweet spot. But things rapidly escalate. Even just a 10-dimensional problem with 50 constraints – perhaps trying to assign a schedule of work to 10 people with different expertise and time constraints – may already land us with several billion corners to try out.

The simplex algorithm finds a quicker way through. Rather than randomly wandering along a polytope’s edges, it implements a “pivot rule” at each corner. Subtly different variations of this pivot rule exist in different implementations of the algorithm, but often it involves picking the edge along which the target function descends most steeply, thus ensuring each step takes us nearer the optimal value. When a corner is found where no further descent is possible, we know we have arrived at the optimal point.

Practical experience shows that the simplex method is generally a very slick problem-solver indeed, typically reaching an optimum solution after a number of pivots comparable to the number of dimensions in the problem. That means a likely maximum of a few hundred steps to solve a 50-dimensional problem, rather than billions with a suck-it-and-see approach. Such a running time is said to be “polynomial” or simply “P”, the benchmark for practical algorithms that have to run on finite processors in the real world.

Dantzig’s algorithm saw its first commercial application in 1952, when Abraham Charnes and William Cooper at what is now Carnegie Mellon University in Pittsburgh, Pennsylvania, teamed up with Robert Mellon at the Gulf Oil Company to discover how best to blend available stocks of four different petroleum products into an aviation fuel with an optimal octane level.

Since then the simplex algorithm has steadily conquered the world, embedded both in commercial optimisation packages and bespoke software products. Wherever anyone is trying to solve a large-scale optimisation problem, the chances are that some computer chip is humming away to its tune. “Probably tens or hundreds of thousands of calls of the simplex method are made every minute,” says Jacek Gondzio, an optimisation specialist at the University of Edinburgh, UK.

Yet even as its popularity grew in the 1950s and 1960s, the algorithm’s underpinnings were beginning to show signs of strain. To start with, its running time is polynomial only on average. In 1972, US mathematicians Victor Klee and George Minty reinforced this point by running the algorithm around some ingeniously deformed multidimensional “hypercubes”. Just as a square has four corners, and a cube eight, a hypercube in n dimensions has 2ⁿ corners. The wonky way Klee and Minty put their hypercubes together meant that the simplex algorithm had to run through all of these corners before landing on the optimal one. In just 41 dimensions, that leaves the algorithm with over a trillion edges to traverse.

A similar story holds for every variation of the algorithm’s pivot rule tried since Dantzig’s original design: however well it does in general, it always seems possible to concoct some awkward optimisation problems in which it performs poorly. The good news is that these pathological cases tend not to show up in practical applications – though exactly why this should be so remains unclear. “This behaviour eludes any rigorous mathematical explanation, but it certainly pleases practitioners,” says Gondzio.

Flashy pretenders

The fault was still enough to spur on researchers to find an alternative to the simplex method. The principal pretender came along in the 1970s and 1980s with the discovery of “interior point methods”, flashy algorithms which rather than feeling their way around a polytope’s surface drill a path through its core. They came with a genuine mathematical seal of approval – a guarantee always to run in polynomial time – and typically took fewer steps to reach the optimum point than the simplex method, rarely needing over 100 moves regardless of how many dimensions the problem had.

The discovery generated a lot of excitement, and for a while it seemed that the demise of Dantzig’s algorithm was on the cards. Yet it survived and even prospered. The trouble with interior point methods is that each step entails far more computation than a simplex pivot: instead of comparing a target function along a small number of edges, you must analyse all the possible directions within the polytope’s interior, a gigantic undertaking. For some huge industrial problems, this trade-off is worth it, but for by no means all. Gondzio estimates that between 80 and 90 per cent of today’s linear optimisation problems are still solved by some variant of the simplex algorithm. The same goes for a good few of the even more complex non-linear problems (see “Straight down the line“). “As a devoted interior-point researcher I have a huge respect for the simplex method,” says Gondzio. “I’m doing my best trying to compete.”

We would still dearly love to find something better: some new variant of the simplex algorithm that preserves all its advantages, but also invariably runs in polynomial time. For US mathematician and Fields medallist Steve Smale, writing in 1998, discovering such a “strongly polynomial” algorithm was one of 18 outstanding mathematical questions to be dealt with in the 21st century.

Yet finding such an algorithm may not now even be possible.

That is because the existence of such an improved, infallible algorithm depends on a more fundamental geometrical assumption – that a short enough path around the surface of a polytope between two corners actually exists. Yes, you’ve got it: the Hirsch conjecture.

The fates of the conjecture and the algorithm have always been intertwined. Hirsch was himself a pioneer in operational research and an early collaborator of Dantzig’s, and it was in a letter to Dantzig in 1957 musing about the efficiency of the algorithm that Hirsch first formulated his conjecture.

Until recently, little had happened to cast doubt on it. Klee proved it true for all 3D polyhedra in 1966, but had a hunch the same did not hold for higher-dimensional polytopes. In his later years, he made a habit of suggesting it as a problem to every freshly scrubbed researcher he ran across. In 2001 one of them, a young Spaniard called Francisco Santos, now at the University of Cantabria in Santander, took on the challenge.

As is the way of such puzzles, it took time. After almost a decade working on the problem, Santos was ready to announce his findings at a conference in Seattle in 2010. Last month, the resulting paper was published in the Annals of Mathematics (vol 176, p 383). In it, Santos describes a 43-dimensional polytope with 86 faces. According to Hirsch’s conjecture, the longest path across this shape would have (86-43) steps, that is, 43 steps. But Santos was able to establish conclusively that it contains a pair of corners at least 44 steps apart.

If only for a single special case, Hirsch’s conjecture had been proved false. “It settled a problem that we did not know how to approach for many decades,” says Gil Kalai of the Hebrew University of Jerusalem. “The entire proof is deep, complicated and very elegant. It is a great result.”

A great result, true, but decidedly bad news for the simplex algorithm. Since Santos’s first disproof, further Hirsch-defying polytopes have been found in dimensions as low as 20. The only known limit on the shortest distance between two points on a polytope’s surface is now contained in amathematical expression derived by Kalai and Daniel Kleitman of the Massachusetts Institute of Technology in 1992. This bound is much larger than the one the Hirsch conjecture would have provided, had it proved to be true. It is far too big, in fact, to guarantee a reasonable running time for the simplex method, whatever fancy new pivot rule we might dream up. If this is the best we can do, it may be that Smale’s goal of an idealised algorithm will remain forever out of reach – with potentially serious consequences for the future of optimisation.

All is not lost, however. A highly efficient variant of the simplex algorithm may still be possible if the so-called polynomial Hirsch conjecture is true. This would considerably tighten Kalai and Kleitman’s bound, guaranteeing that no polytopes have paths disproportionately long compared with their dimension and number of faces. A topic of interest before the plain-vanilla Hirsch conjecture melted away, the polynomial version has been attracting intense attention since Santos’s announcement, both as a deep geometrical conundrum and a promising place to sniff around for an optimally efficient optimisation procedure.

As yet, there is no conclusive sign that the polynomial conjecture can be proved either. “I am not confident at all,” says Kalai. Not that this puts him off. “What is exciting about this problem is that we do not know the answer.”

A lot could be riding on that answer. As the algorithm continues to hum away in those basements it is still, for the most part, telling us what we want to know in the time we want to know it. But its own fate is now more than ever in the hands of the mathematicians.

Richard Elwes is a visiting fellow at the University of Leeds, UK, and the author of Maths 1001: Absolutely everything that matters in mathematics(Quercus, 2010)