Tuesday, September 25, 2012

September 25, 2002

Ten years ago today, Lucent Technologies released the Final Report of our committee on possible misconduct by J. Hendrik Schön. The conclusion was clear: that there had been extensive misrepresentation and, by implication, that the entire body of work was suspect. Actually, the report had been largely finished for a while, but Lucent wanted a chance to have their lawyers and management check things over. This was a small price to pay to have the report publicly released in full. Remember that they had no legal obligation to release anything, but they released everything. The response to the report was very gratifying. Although there had been until then some holdouts who felt that the investigation might be some kind of witch-hunt, the extensive documentation that we provided of the many instances of misconduct assured that virtually no one maintained that belief afterwards. In spite of this sense of satisfaction and closure, however, we took no joy in delivering that verdict.

Tuesday, June 26, 2012

June 26, 2002: When did it start?

The second question most people ask with respect to Hendrik Schön is "When did he start faking data?"

(The first question is either "Why did he do it?" or "How could he possibly think he could get away with it?" Neither of these questions can be answered without a Vulcan mind meld.)

In our investigation committee, we did not seek to answer this question, only to pursue allegations that were brought to our attention, either directly or through Bell Labs. In fact, as the number of papers under suspicion grew from five to 25 or so, we realized we needed to close the door to new allegations, and we set an arbitrary cutoff date of June 20, 2002. We had plenty of material to work with for our report.

It turns out that a few days later we got another allegation that gave a strong hint that the faking began when Hendrik was in graduate school at the University of Konstanz, before anyone at Bell Labs knew anything about him.

To be honest, we were somewhat relieved that the new information came in after the cutoff, because investigating it would have involved us in a completely new class of materials and measurements, as well as new co-authors (obviously none from Bell Labs). We also did not have access to the digital data that was so useful for the other papers, and our authority to investigate this work would have been somewhat questionable, since we had been invited in by Bell Labs.

As we would later be told, Hendrik's thesis was rather unexceptional: lots of slow experimental work trying to accomplish a rather mundane task of introducing electrically active dopants into potential solar-cell materials. Reportedly, he was not having much success despite stead efforts, but at some point he suddenly reported dramatic success in doping the materials.

There were several different experiments on different materials with different dopants. Three figures from the papers are overlaid in this animated gif:

.

At least in the context of the other clear cases of copying, this looks pretty clear. Obviously not all the curves are the same, but some are, and the three figures are represented as coming from three different materials. However, the committee never examined this case in detail, so there may be important caveats.

The University of Konstanz decided in 2004 to rescind Hendrik's doctorate, deeming him "unworthy." This retroactive action struck me as rather odd. A court concluded in 2010 (amusing translation here; follow either link for a relatively recent picture) that the revocation was inappropriate, but in 2011 that decision was reversed.

In 2003, Jennifer Couzin reported in Science that "a committee at the University of Konstanz examining the work of disgraced physicist Jan Hendrik Schön found inconsistencies in several papers Schön published during his studies there, but no proof that he had deliberately manipulated data."

But I think we know he had already started down the path.

Friday, June 1, 2012

Confirming Expectations


A good experiment definitively distinguishes between alternative hypotheses. But things get tricky when a well established standard view is pitted against ill-defined alternatives.

My Focus story today describes measurements of density fluctuations in ultra-cold gases. Several experimental groups have been capturing and cooling bunches of about a thousand atoms above "atom chips" to study such things as Bose-Einstein condensates. In this case, Julien Armijo, a former member of a group at the Insitute d'Optique in Palaiseau, attributes some of the density fluctuation to quantum zero-point excitation of sound waves in the atomic cloud.

For an isolated oscillator, the signature of zero-point motion is conceptually straightforward: below a certain temperature the motions no longer decrease with temperature. What's left are the intrinsic quantum-mechanical oscillations, and the freezing temperature corresponds to the minimum quantum of energy needed to excite the oscillator. The obvious "null hypothesis" to be excluded would therefore be that the fluctuations continue to decrease toward zero with further cooling.

For the atomic cloud, however, the situation is much more subtle, because the sound waves have a continuous spectrum that extends to zero energy. This means that there are always some waves--the ones with the longest wavelength--that are excited no matter how low the temperature.

A further complication is that different wavelengths vary in their effect on the density fluctuations. A sophisticated theory says that the quantum contribution from the longest wavelengths does not add to the density fluctuations at all. In fact, this theory says that, at very long wavelengths, the fluctuations go away at zero temperature--exactly what one would expect if there were no quantum fluctuations!

As it turns out, though, the experiment measures fluctuations in individual pixels that are a few microns on a side, which corresponds to including waves with wavelengths on the same scale. The theory says that these waves will cause a measurable density fluctuation even at zero temperature.

But the same theory says that at nonzero temperatures, including shorter wavelengths will decrease the contribution of thermal excitations by exactly the same amount. The two terms get bigger as the pixels get smaller, but since they cancel anyway that doesn't affect the prediction.

This is messier than just looking for fluctuations that don't freeze out, isn't it?

It's rather difficult to choose a good null hypothesis where there are no zero-point motions. After all, everyone believes that, physically, the quantum fluctuations should be there, although different theoretical treatments may make slightly different predictions. So there is no particularly obvious way to choose a model where the quantum fluctuations are absent.

Armijo measures fluctuations that don't change with effective pixel size, just as the complete theory predicts. Of course, the measurements also agree with a theory that omits the size dependence of both the quantum and the thermal contributions. What they don't agree with, he emphasizes, is a model that includes only the thermal corrections, since these are no longer cancelled by the quantum term. It's not clear that anyone thinks this would be a credible model that needs to be excluded. (Setting Plank's constant to zero, a common way to "turn off" quantum effects, seems to make both corrections go away.)

What is clear is that the claim that this is a "direct observation of quantum phonon fluctuations" needs to be parsed quite carefully.

Tuesday, May 29, 2012

May 28, 2002: First look at internal Schön investigations

Just after Memorial Day in 2002, our committee finally received the records of multiple internal Bell Labs investigations of the work of Hendrik Schön. The frustrating delay of several weeks in getting these documents was caused in part by the need to be sure that we had legal indemnification from Lucent for our activities. That being accomplished, we were ready to begin in earnest, and the documentation was weighty and troubling.

We received a large three-ring binder with a dozen or so sections dealing with various issues. In addition, we got copies of email correspondence with and about Hendrik, including management discussions about various issues. (There was no easy way to verify how complete the email collection was, but there was a lot there.)

Since these documents, including internal company communications, were made available as part of our official committee activities, I can't share them in detail, even now. But they illustrated serious concern among the Bell Labs managers, going back at least to the summer of 2001. Hendrik was strongly encouraged, for example, to help other researchers to reproduce his extraordinary dielectric films, and was told that it was more important to solidify his existing work than to demonstrate new breakthroughs.

This concern was not in evidence, however, in various public statements about the work, notably the "self-aligned monolayer field-effect transistor," or SAMFET, described in the fall of 2001. For example, a Business Week article quoted his recently appointed immediate manager John Rogers as if he had witnessed the assembly: "The whole thing just happens in a beaker at a chemistry bench." A Lucent press release included endorsements of the profound potential from his then third-level manager Cherry Murray ("Although there may be no practical applications for a decade, it could lead to a new paradigm in electronics") and second-level manager Federico Capasso ("The molecular-scale transistors that we have developed may very well serve as the historical 'bookend' to the transistor legacy started by Bell Labs in 1947." Ironically the word "bookend" was more appropriate than he seemed to realize.). Capasso is also reported to have told an internal Bell Labs meeting that it was OK if the work was not completely correct because it was so exciting that it would stimulate further work that would clarify what was going on.

During the same period, Hendrik's original mentor, Bertram Batlogg, continued to extol the earlier experiments on single-crystal organic semiconductors, even as he strongly questioned Hendrik's description in an unpublished document of the way the insulating film had been optimized.

There is a profound lesson in the fact that all of these fine scientists were willing to put aside their own misgivings and to publicly endorse Hendrik's results. I can't seriously entertain the idea that any of them the knew the results were faked and were deliberately misleading people. Instead, I have to believe that they had managed to deceive themselves that the problems were mistakes or misunderstandings, and that the groundbreaking work would stand the test of time even after these problems were resolved.

They were all wrong. The scientific world would have been better served if they had been more willing to openly question the honesty of their sincere young colleague.

Thursday, May 24, 2012

May 24, 2002: What got the Schön investigation going?

Leading into the spring of 2002, there had been increasing discomfort with Hendrik Schön's stunning body of scientific results. The official investigation into his possible misconduct, however, only began when a group of nano-physicists noticed duplicated data in five papers. As described nicely in Eugenie Reich's book Plastic Fantastic, Lydia Sohn (then at Princeton), Paul McEuen, Leo Kouwenhoven, and Charleses Marcus and Lieber sent their findings to the Bell Labs management and to the journals where the articles were published at the beginning of May. It is interesting to speculate whether and how the investigation would have proceeded if it had been communicated less publicly.

During this period, a PowerPoint version of the slides was widely circulating in the nano community. The usual technique was to print out the graphs on transparencies and line them up, but I include them here as animated gifs for a change of pace. The figures show very similar curves, even though the data are represented as being taken from completely different samples. In principle this could happen once through gross negligence, by sending the wrong figure (with the wrong label!). In addition, however, the axes are sometimes differ in sign or by an integer multiplier, or they have some curves missing, which are much harder to understand. But the most damning evidence is that the small deviations from the curves are often very similar in different plots, even though this "noise" should vary each time the measurement is repeated.




For many researchers, seeing these images was all the evidence they needed that Schön had fabricated at least some of his data. Others held out hope that there was some innocent explanation, and looked to our "blue-ribbon panel" to resolve the issue. We on the committee felt a lot of pressure to get it right. But by this time ten years ago, we still hadn't yet gotten any detailed documentation from Lucent.

Wednesday, May 23, 2012

May 21, 2002: Investigating the Schön Affair

Ten years ago this week, Lucent Technologies announced that it had convened a panel to investigate possible scientific misconduct at Bell Labs. Interestingly, the story in the New York Times by Kenneth Chang featured the role of Bertram Batlogg, mentioning Hendrik Schön only in the final paragraph.

At the time, I thought this story missed the point, since it seemed clear that Schön was at the center of the problems. (Batlogg wasn't even an author on some of the papers in question.) But as Chang told me later, Batlogg had been highly visible when the spectacular "breakthroughs" were being announced. For this reason, Chang (who had covered the earlier work) thought that Batlogg's was the name that readers would be most likely to recognize, and not highlighting it would be a disservice to readers.

One of the biggest challenges to the committee (which I served on) was figuring out how to deal with Batlogg's role, neither assigning him primary responsibility nor minimizing his role.

Wednesday, May 5, 2010

Free to Choose

Today's New York Times features a bizarre op-ed by Charles Murray of the American Enterprise Institute, a think tank that describes its scholars as "committed to expanding liberty, increasing individual opportunity, and strengthening free enterprise."

Murray summarizes the results of head-to-head comparison of standardized-test performance for students in public and charter schools. The charter-school students, he admits, "generally had 'achievement growth rates that are comparable' to similar Milwaukee public-school students. This is just one of several evaluations of school choice programs that have failed to show major improvements in test scores, but the size and age of the Milwaukee program, combined with the rigor of the study, make these results hard to explain away."

"So let's not try to explain them away," Murray says.

This makes good sense. This kind of experiment has been done many times, as described by Diane Ravitch in her thought-provoking 2010 book, The Death and Life of the Great American School System: How Testing and Choice Are Undermining Education. The results are clear in their lack of clarity: some charter schools are excellent, others are terrible. The data don't support the vision that non-public schools are automatically superior, although some no doubt are. In addition, the data show no indication that public schools faced with competition respond by cleaning up their act, another frequent argument for school choice.

So do these conclusions shake the confidence of a school-choice advocate like Murray? Hardly: "Why not instead finally acknowledge that standardized test scores are a terrible way to decide whether one school is better than another?" he says.

Murray is correct, of course, that tests scores have serious flaws. Ravitch, who admits that she was once a strong proponent of both choice and testing, spends much of her book describing the problems with standardized tests. Especially troublesome are the tests that states devise to show that they are meeting the goals of the "No Child Left Behind" act. Ravitch laments both the limited range of skills being tested--essentially basic math and reading--as well as the distortions that inevitably occur when tools meant to monitor progress start to be used to enforce it.

Ravitch forcefully argues that school improvement is a hard slog, not achieved by silver bullets like charter schools or by extensive data collection like that promoted by the Obama administration's "Race to the Top." Instead of statistical analyses modeled on business practices, she advocates a rigorous (voluntary) national curriculum and on-the-ground assessments by professional educators, not business managers.

Murray acknowledges the failures of previous silver bullets: "whether the reform in question is vouchers, charter schools, increased school accountability, smaller class sizes, better pay for all teachers, bonuses for good teachers, firing of bad teachers — measured by changes in test scores, each has failed to live up to its hype." But he concludes is that the problem lies with testing, and that choice is still a social good, because it allows parents to choose schools whose teaching styles the parents find appropriate.

It will be interesting to see whether Murray's fellow school-choice advocates follow his recommendation to admit that there is no measurable benefit of charter schools, but policy should support them anyway on ideological grounds. Somehow I doubt it.