Some computers can beat grandmasters at chess.
Others can compose music.
A few have even been programmed to work as therapists.
So perhaps it was only a matter of time before someone would write a program able to match the world's best art historians at telling the difference between masterpieces and high-quality fakes.
Some will find it hard to believe that a computer can eye a piece of art and detect the almost ineffable subtleties that together declare it the work of a master.
Art is, after all, more art than science.
And to be sure, the new program, created by a team at Dartmouth College, is not likely to end the divisiveness that has long characterized the field of art authentication. The new system's conclusions about who painted certain faces in a valuable Italian Renaissance painting have already started a minor squabble among experts.
But the achievement marks a milestone in the application of computational mathematics to the arcane field of artistic authorship. If nothing else, its creators said, it adds a new voice to the small circle of specialists who have dominated the endeavor for so long.
"It's not like we're going to step in and solve all the art world's controversies," said Hany Farid, the computer scientist who led the effort. "I see this as contributing to the dialogue."
The new method relies on a computational field known as wavelet statistics, which offers a mathematical means of detecting patterns within images.
Wavelet analysis is at the heart of so-called compression technology, which allows complex computer images to be stored digitally without taking up huge amounts of memory. But one aspect of wavelet analysis -- its ability to assess graphical elements on many scales at once -- makes it a potentially useful tool for art authentication.
Specifically, it can recognize big patterns, such as the sweep of a dark pencil stroke on white paper. Yet at the same time it can identify the almost invisible shadows cast by tiny ridge lines left by individual hairs in a painter's finest brush. The combination of such patterns on many scales adds up to a particular artist's unique style.
Farid had been using wavelet analysis to tell whether digital images have been tampered with -- a blossoming problem with the advent of software programs such as Photoshop, which allows people to alter photographs almost seamlessly. Then he and a colleague -- fellow Dartmouth computer scientist Daniel Rockmore -- visited the Metropolitan Museum of Art in New York.
Trying it out
The exhibit they saw -- about the 16th century Flemish artist Pieter Bruegel the Elder -- focused on the large number of drawings that had once been attributed to Bruegel but had later been determined to be either frauds or, at best, the innocent work of imitators.
"We thought, 'This would be great if [wavelets] worked for drawings and paintings,' " Farid said.
Working with graduate student Siwei Lyu, they derived a series of wavelet functions they predicted would allow a computer to detect the minor differences in sweep, style and technique that might separate a real Bruegel from a fake. The idea, Farid said, was to pick up on the little distortions that inevitably come up when someone tries to imitate another's work.
"Everyone knows about this because when we tried, as kids, to forge our parent's signature on notes from school, we all did it in a very jerky way, and it didn't look right," he said.
Those are the kinds of differences, though on a much finer level, that wavelet functions detect.
Then came the test.
They turned their program loose on eight authenticated Bruegels and five acknowledged imitations. The program correctly identified each as real or fake, the team reported in the Nov. 21 early edition of the Proceedings of the National Academy of Sciences.
"We're not entirely sure of what the connoisseurs look for," Rockmore said. "But that we can get some mathematical analogy to what people key in on in art is a fascinating finding."
To try something more challenging -- a complex painting with multiple layers of pigment and variations in brush strokes -- the team tested its system on "Madonna and Child With Saints," a large-canvas painting by the Italian early Renaissance painter Pietro Perugino.
It has long been presumed that several hands contributed to that work, as was the custom of the day, and virtually all experts today agree that the Madonna figure was by Perugino himself. But what of the others?
The wavelet analysis clumped three of the six faces in the painting as the product of the same hand: The Madonna and the two saints to the Madonna's right. The three remaining figures appeared to be by three different artists.
Hold on, says Laurence Kanter, curator in charge of the Robert Lehman collection at the Met in New York.