Phsctanu 1 Re

editorial
No man is an island
The financial crisis underlines the need for new economic models — models that can only be built by
following a truly interdisciplinary approach.
When the Kiel Institute for the World interdisciplinary research). Goldenfeld’s Models of the financial markets as
Economy was founded in 1914, it was given statement is a beautifully succinct definition complex systems do already exist, under
the graceful name Königliches Institut für of the tipping point at which a simple system the banner of ‘econophysics’, but these
Seeverkehr und Weltwirtschaft — Royal becomes ‘complex’. For the financial markets, certainly need much further refinement.
Institute for Maritime Transport and the tipping point has been passed — the US There is also help to be had from other
the World Economy. The emphasis on housing market hits trouble, and suddenly fields that deal with complex systems —
maritime transport, however, was soon Iceland is on the brink of bankruptcy. Even including sociology, epidemiology, ecology,
deemed outdated and dropped after twenty more sobering is that, despite active research engineering and brain research. Different
years, along with the reference to the King in the field (keeping many a physicist busy), as their specific objectives and systems of
of Prussia. the crisis, in this form and to this extent, was investigation are, the means of reaching a
The original name is a reminder of the role not predicted — or even thought possible. fuller understanding of the mechanisms at
that maritime navigation played in building What has gone wrong with economic work are often quite similar between such
a truly global economy. Yet those relatively theory? This question is discussed by fields (as even some of the mechanisms
small trade networks have little in common Thomas Lux, a member of the scientific themselves may be). Data collection and
with their modern counterparts: today, more advisory board of the Kiel Institute of handling, robustness of networks against
goods and information are transported at an World Economics, and Frank Westerhoff in failure and attacks, the need for early
incomparably faster pace; many more players their Commentary on page 2 of this issue. detection and prevention of failure and,
interact, in ever more sophisticated ways. But They argue that the distinction made in not least, the fact that different networks
this increased connectivity has not only made natural sciences between phenomena on interact with each other — these are topics
the system more efficient and effective, it has the micro-scale of single constituents and relevant across all disciplines dealing with
also made it more difficult to understand — the macro-scale of the whole system has not complex systems.
to the point that it is unpredictable, as the reached mainstream economic theories. In The solution to the financial crisis will not
current financial crisis plainly shows. economic models, the focus has been on a come through the application of physics alone,
“Complexity starts when causality breaks lone actor, a representative ‘Robinson Crusoe’, although physicists have much to contribute.
down”, said physicist Nigel Goldenfeld which carelessly neglects the reality of many Better models with more predictive power will
of the University of Illinois at a recent heterogeneous agents interacting strongly arise only through close collaboration with
conference on complex systems (organized and giving rise to emergent phenomena. economists who understand the details of the
by the National Academies Keck Futures Economics, say Lux and Westerhoff, has a lot markets and through strong interaction within
Initiative, a programme aiming to catalyse to learn from the natural sciences. the wider complex-systems community. ❐
2009 and all that

For the new year, we have a new look.
This is the first issue of Volume 5 of Nature Nature Photonics and Nature Geoscience — has The editorial team of Nature Physics also
Physics, and with it comes a new look for our been redesigned, refreshing the templates that has a new look. We’re pleased to welcome
pages (or PDFs, if you’re reading online). Every we have used since our launch in 2005. The David Gevaux as the fifth member of the
section of the journal — and of its sister titles, result, we hope you’ll agree, is a look that is expanded team, taking the growing journal
Nature Materials, Nature Nanotechnology, modern, effective and still distinctively ours. into its fourth year of publishing. ❐
The editorial team: Alison Wright, Edmund Gerstner, May Chiao, Andreas Trabesinger and David Gevaux.
nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 1
© 2009 Macmillan Publishers Limited. All rights reserved

editorial
No man is an island
The financial crisis underlines the need for new economic models — models that can only be built by
following a truly interdisciplinary approach.
When the Kiel Institute for the World interdisciplinary research). Goldenfeld’s Models of the financial markets as
Economy was founded in 1914, it was given statement is a beautifully succinct definition complex systems do already exist, under
the graceful name Königliches Institut für of the tipping point at which a simple system the banner of ‘econophysics’, but these
Seeverkehr und Weltwirtschaft — Royal becomes ‘complex’. For the financial markets, certainly need much further refinement.
Institute for Maritime Transport and the tipping point has been passed — the US There is also help to be had from other
the World Economy. The emphasis on housing market hits trouble, and suddenly fields that deal with complex systems —
maritime transport, however, was soon Iceland is on the brink of bankruptcy. Even including sociology, epidemiology, ecology,
deemed outdated and dropped after twenty more sobering is that, despite active research engineering and brain research. Different
years, along with the reference to the King in the field (keeping many a physicist busy), as their specific objectives and systems of
of Prussia. the crisis, in this form and to this extent, was investigation are, the means of reaching a
The original name is a reminder of the role not predicted — or even thought possible. fuller understanding of the mechanisms at
that maritime navigation played in building What has gone wrong with economic work are often quite similar between such
a truly global economy. Yet those relatively theory? This question is discussed by fields (as even some of the mechanisms
small trade networks have little in common Thomas Lux, a member of the scientific themselves may be). Data collection and
with their modern counterparts: today, more advisory board of the Kiel Institute of handling, robustness of networks against
goods and information are transported at an World Economics, and Frank Westerhoff in failure and attacks, the need for early
incomparably faster pace; many more players their Commentary on page 2 of this issue. detection and prevention of failure and,
interact, in ever more sophisticated ways. But They argue that the distinction made in not least, the fact that different networks
this increased connectivity has not only made natural sciences between phenomena on interact with each other — these are topics
the system more efficient and effective, it has the micro-scale of single constituents and relevant across all disciplines dealing with
also made it more difficult to understand — the macro-scale of the whole system has not complex systems.
to the point that it is unpredictable, as the reached mainstream economic theories. In The solution to the financial crisis will not
current financial crisis plainly shows. economic models, the focus has been on a come through the application of physics alone,
“Complexity starts when causality breaks lone actor, a representative ‘Robinson Crusoe’, although physicists have much to contribute.
down”, said physicist Nigel Goldenfeld which carelessly neglects the reality of many Better models with more predictive power will
of the University of Illinois at a recent heterogeneous agents interacting strongly arise only through close collaboration with
conference on complex systems (organized and giving rise to emergent phenomena. economists who understand the details of the
by the National Academies Keck Futures Economics, say Lux and Westerhoff, has a lot markets and through strong interaction within
Initiative, a programme aiming to catalyse to learn from the natural sciences. the wider complex-systems community. ❐
2009 and all that

For the new year, we have a new look.
This is the first issue of Volume 5 of Nature Nature Photonics and Nature Geoscience — has The editorial team of Nature Physics also
Physics, and with it comes a new look for our been redesigned, refreshing the templates that has a new look. We’re pleased to welcome
pages (or PDFs, if you’re reading online). Every we have used since our launch in 2005. The David Gevaux as the fifth member of the
section of the journal — and of its sister titles, result, we hope you’ll agree, is a look that is expanded team, taking the growing journal
Nature Materials, Nature Nanotechnology, modern, effective and still distinctively ours. into its fourth year of publishing. ❐
The editorial team: Alison Wright, Edmund Gerstner, May Chiao, Andreas Trabesinger and David Gevaux.

commentary
Economics crisis
Thomas Lux and Frank Westerhoff
Economic theory failed to envisage even the possibility of a financial crisis like the present one. A new
foundation is needed that takes into account the interplay between heterogeneous agents.
O
nce viewed as mystic monetary is alive, and was well — at least until the pervasively in economics education that
engineer, Alan Greenspan, former current financial crisis started to unfold. their basic tenets are taken for granted
Chairman of the US Federal Although the principles outlined above as the principles ruling the real world,
Reserve, has been re-cast as irresponsible are still the basis of most contemporary despite all of the anomalies and exceptions
villain, one who laid the ground for the scholarly activity in economics, there are discovered in empirical research.
present worldwide financial catastrophe. other trends. These include innovative For instance, it would be hard
Asked whether his ‘ideology’ had pushed work in ‘behavioural economics’ and to find supporting evidence for the
him to make decisions he now regrets, experimental work with human subjects — firmly held belief that more derivative
Greenspan confessed1 that he would have recognized in the award of the 2002 instruments — which should allow agents
deemed impossible the ongoing disruptions Nobel Prize to Daniel Kahneman and to insure themselves better against the
to the financial system and that “his belief Vernon L. Smith — which have revealed stochastic wheels of fortune — lead to a
in deregulation had been shaken”. a plethora of behavioural patterns that better allocation of resources and thus
He is not the only one who has been contradict the assumption of perfectly an increase in market efficiency. This
taken entirely by surprise. Most economists rational behaviour. assertion is based entirely on the benefits
did not in any way foresee the depth of of contingent claims in the textbook
the current crisis, or even consider it To the shocked public, it looks general-equilibrium model. Derived in
possible. Even those who warned2 of over- as if some Dr Strangelove the abstract, the efficiency gain through
exuberance in the US housing market did derivatives is only a hypothesis, yet this is
not have any clue about the impending on Wall Street had pushed not how economists are used to thinking
meltdown, which, to the shocked public, of such theorems: it is the mathematical
looks as if some Dr Strangelove on Wall
the button on a financial proof within the model economy that is
Street had pushed the button on a financial Doomsday Device. considered its validation, rather than any
Doomsday Device. Greenspan’s supposed empirical evidence.
ideology certainly coincided widely with However, these developments still occupy A glance at real-life operations in
that of mainstream economists who believe only a marginal position. The widespread derivative markets easily shows why the
in the self-regulating forces of unrestrained perception within our profession is that theory fails: instead of hedging away risk,
financial markets, the ‘efficiency’ of behavioural research delivers a curious set of many market participants use derivatives
asset-price formation, and the increased anomalies or exceptions that lack coherence, in an ‘anomalous’ way, to build up
efficiency in risk allocation and sharing and whose impact gets washed out in the speculative positions so as to profit from
through the introduction of ever more aggregate. In contrast, the mainstream higher returns, as long as the downside
financial instruments. paradigm is seen as a more solid and risk does not materialize. The near disaster
All of this is just the finance version consistent framework. Economic policy brought about in the late 1990s by the
of that textbook economic paradigm, advice, particularly in financial economics, collapse of notorious hedge fund Long-
‘homo economicus’, who has unlimited will therefore typically be based on a set of Term Capital Management (intellectually
insightfulness and capability of deliberation axioms and hypotheses derived ultimately based on modern derivative theory) should
(economists typically speak of ‘rationality’). from the Robinson Crusoe scenario. As have raised some doubts. If that was not
This admirable person manages his the prevailing financial crisis cannot be compelling enough, the present crisis
financial affairs as a side-aspect of his explained using these standard tools, should constitute its ultimate rejection.
utility maximization problem, taking into economic theory basically offers policy The dominance of the rational-agent
account all potential future happenings makers little guidance about what to do in paradigm is intimately intertwined with
with the correct probabilities. As there the current situation. an even more cumbersome ‘conceptual
is only one way to be perfectly rational, A major problem is that despite many reductionism’. As there can only be one
this agent is usually the lone actor in refinements, this is not at all a system based way to act fully rationally, everyone
economic models — a representative on and confirmed by empirical research should display exactly the same behaviour.
Robinson Crusoe. (as the naive believer in ‘positive science’ Therefore, a representative agent would be
Of course, this Crusoe has been often might expect). The vision (or ideology) sufficient. Taking both aspects together,
derided as a straw-man illustration of encapsulated in the mainstream approach the typical format of current economic
the dominant paradigm, criticized by is of a more ‘pre-analytical’ nature and is models is that of a single household or firm
non-mainstream economists, unbelieving supported mainly by elegant but idealistic maximizing its utility or profit over a finite
natural scientists and a similarly models of the economy. Perfect rationality or infinite lifespan. Technically, this is a
unbelieving public. Still, the straw man and optimizing behaviour are used so dynamic programming problem.
2 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

commentary
This methodological preference excludes macroeconomics has been entirely reduced of financial markets. Currently, similarly
the study of interaction among economic to microeconomic theory in the past few simple stochastic models are being
agents. However, most of what is relevant decades by insisting on representative developed in the study of the distribution
and interesting in economic life has to do rational agents. That the overall system of income and wealth7, and some
with the interplay and connection between is different from its parts is plainly economists have even taken this approach
diverse economic actors. The current crisis incomprehensible from the viewpoint of to macroeconomic models8.
is a perfect example of the importance of the ruling school of thought.
interactions at various levels. It was the Economics has thus, by its methodology, Economics should take stock
interaction between highly connected tied its own hands and prevented the of the experience of the
international financial markets that has analysis of vital aspects of economic
generated the spillover from the US systems. For example, despite the recent natural sciences in handling
subprime-mortgage problem to other layers surge of research in network theory, the
of the financial system. now apparent linkages between banks have
complex systems with strong
Securitization of credit risks enabled received scant attention. In the few papers interactions.
lenders to sell various parts of their that have been published, the analyses are
mortgage portfolios to other financial of a static nature based on equilibrium The apparent systemic vulnerability
institutions thus creating new links with concepts and do not easily lend themselves of our globalized financial markets has
these buyers as well as, indirectly, among to empirical applications. The even smaller brought to the fore another carelessly
them. Other new asset classes, such as number of studies using empirical data or neglected facet of economic interactions.
credit default swaps, added additional realistic models come from authors with Most economic problems are emergent
new links between formerly unconnected a background in physics5. Unfortunately, phenomena of complex societies that
entities. This gives us a glimpse of how the study of anything at a systemic level require a systemic perspective. A new
financial innovations have increased has been defined away from economics by micro-foundation based on interactions
the degree of connectivity within the the insistence on micro-foundations that would be the missing macro counterpart
financial system. It is well known that simply set the macro sphere equal to the to the microeconomic regularities
highly connected systems might be ‘robust microscopic base unit. revealed in behavioural economics. To
yet fragile’3, but such important aspects What could be the way out of this develop a proper perspective on systemic
have been out of reach of the mainstream dilemma? In our view, a change of phenomena, economics as a science should
approach to economics. methodological orientation in economics take stock of the experience of the natural
In fact, the ubiquitous notion of is needed, to take into account the ‘more sciences in handling complex systems with
‘systemic risk’ signals that current events is different’ paradigm. On the one hand, strong interactions. A partial reorientation
concern the financial system at an aggregate economists need to take seriously the in modelling principles and more
level. For natural scientists, the distinction various deviations from ‘rationality’ methodological flexibility would enable
between micro-level phenomena and those revealed by behavioural research. On the us to tackle more directly those problems
originating on a macro-, system-wide scale other hand, however, to avoid getting lost that seem to be most vital in our large,
from the interaction of microscopic units in a patchwork of behavioural biases and globalized economic systems. ❐
is clear. The overall systemic features of anomalies, a new empirically based type of
the crisis would be seen as an emergent micro-foundation is necessary — one that Thomas Lux is in the Department of Economics,
phenomenon of the dispersed micro- stresses more the links between boundedly University of Kiel, Olshausenstraβe 40, D‑24118
activity. To reduce these macro events to rational agents rather than the agents’ Kiel, Germany, and is a member of the research
the outcome of the decision process of a internal processes. It would, therefore, group ‘Risks in the Banking Sector’ of the Kiel
single agent seems to be missing the point. also not be enough to replace the current Institute for the World Economy.
paradigm by a representative ‘non-rational’ e‑mail: lux@bwl.uni‑kiel.de
Economic theory offers actor (as has sometimes been done in Frank Westerhoff is in the Department
policy makers little guidance recent literature). of Economics, University of Bamberg,
The experience of the natural Feldkirchenstraβe 21, D‑96045 Bamberg, Germany.
about what to do in the sciences in coping with complex e‑mail: frank.westerhoff@uni‑bamberg.de
systems would suggest a parsimonious
current situation. stochastic approach. Because agents References
1. Andrews, E. L. Greenspan concedes error on regulation.
As with systemic risk, the notion of in large economic systems will display The New York Times online 23 October 2008.
coordination failure (a term often used heterogeneity in terms of their different 2. Shiller, R. Irrational Exuberance 2nd edn (Princeton Univ.
to characterize the endogenous nature micro motives, degrees of deliberation and Press, 2005).
of economic slumps and recessions) information-processing capabilities, one 3. Watts, D. J. Proc. Natl Acad. Sci. USA 99, 5766–5771 (2002).
in itself encapsulates a perspective of might hope that this variability of human 4. Anderson, P. W. Science 177, 393–396 (1972).
5. Iori, G., Jafarey, S. & Padilla, F. J. Econ. Behav. Organ.
the ‘more is different’ paradigm4, an behaviour can be quantified in a tractable
61, 525–542 (2006).
involuntary negative collective outcome way using statistical laws. Ongoing work 6. Lux, T. in Handbook on Financial Economics (eds Schenk-Hoppé, K.
of a system of dispersed activity. Rather inspired by statistical physics shows that & Hens, T.) (Elsevier, in the press).
than looking for the explanation in a relatively simple models with plausible 7. Chatterjee, A., Yarlagadda, S. & Chakrabarti, B. (eds) Econophysics
particularly odd case of a microscopic behavioural rules have the potential to of Wealth Distributions (Springer, 2005).
dynamic programming problem, it would replicate key empirical regularities of 8. Aoki, M. & Yoshikawa, H. Reconstructing Macroeconomics:
A Perspective from Statistical Physics and Combinatorial Stochastic
seem much more plausible to investigate financial markets6. In these models, direct
Processes (Cambridge Univ. Press, 2007).
the ‘logic of collective activity’ on the and indirect local and global interactions
macroscopic scale. However, due to the between market participants are important Acknowledgement
conceptual reductionist philosophy, ingredients in understanding the dynamics The authors are grateful to Mishael Milakovic for inspiring discussions.

thesis
Know the unknown

The human mind isn’t attuned to statistical Ord and colleagues argue that if lower-tier
argument. Suppose a person selected at journals received the same scrutiny as the
More quantitative
random on a US street tests positive for most prominent, this number would rise to
HIV: what’s the chance that they really are thinking about the around 10−3, which can be taken as a crude
infected with the virus? Know that the test is estimate of the chance that a published
very accurate, yielding only a 0.0001 fraction
nature of our ignorance scientific argument will turn out to be
of false positives. Given this number, it is in order. significantly incorrect.
comes as a surprise to most people — What about arguments for LHC safety?
physicists too, I suspect — that this random Certainly it seems hard to believe that
individual, after testing positive, actually has potentially catastrophic risks linked to arguments given so far are less likely than
only a 50% chance of really being infected. experiments at the Large Hadron Collider, this to be seriously incorrect. There are
We’re lured into error by that or LHC (Nature Phys. 4, 431; 2008). just too many unknowns in the relevant
impressively small number, which leads us, Careful studies carried out by a range of physics, and in current limits on our
almost without thinking, to conclude ‘very acknowledged experts, using conservative understanding of that physics, to have
unlikely’. A little care with the conscious estimates, arrived at a probability of the such certainty. Is there less than a 10−3
mind says that we need to include the order of 10−9 per year for a dangerous event. chance of major revisions to quantum
fact that our ‘random’ individual, before But as several authors have since pointed chromodynamics in the next century?
testing, has only a 1 in 10,000 chance of out, one has to be very careful in accepting The theory is impressive, a monumental
being HIV-positive, this being the general such estimates at face value, even if they intellectual achievement, and the chance
US infection rate. Of 10,000 people tested, have been carried out by the best experts may be small, but that small?
on average, one will be truly infected, as using great caution. Ord and colleagues’ overall point is that
their test will almost certainly show. But In short, as Toby Ord and colleagues if an argument has a small but significant
of the other 9,999 without the virus, on suggest (http://arxiv.org/abs/0810.5515), chance of being mistaken, then it becomes
average one will also turn up a false positive. conclusions of very small probabilities questionable whether one can really
Hence, half the positives will be false (the inevitably put extremely strong accept its conclusion that an event has a
results change markedly if the person requirements on the accuracy of the comparatively much smaller chance of
isn’t selected ‘at random’, but belongs to a argument; even a very small chance of error happening. But a concrete conclusion to
group already known to be at high risk for can undermine the argument’s conclusions. that effect isn’t obvious either. The quantity
HIV infection.) Their logic is simple and persuasive. about which we know the very least is the
Our susceptibility to these kinds of error Suppose one gives an argument A for some probability P(X; not A) for an event to take
has been documented in many experiments. conclusion X, finding a probability P. This place, given that the argument we’ve just
A survey several years ago found that as probability is not simply P(X) — that is, the made about it is wrong. Were this of order
many as 40% of doctors got the problem probability for the event to happen. Rather, 10−3, then arguments for LHC safety might
wrong, even though they’d received it is the conditional probability, P(X; A), for only give an overall chance of danger at 10−6
specialist training in such reasoning. But the event to occur given that the argument per year. But estimating P(X; not A) means
the difficulty of such problems isn’t special is correct. The full probability for the taking a step into the ‘unknown unknowns’,
to medicine; it reflects an issue we all have event X has to be augmented by a term made famous by former US Secretary of
when trying to think clearly about very equal to the product P(X; not A)P(not A); Defence Donald Rumsfeld. We may be okay
unlikely events, and one that warrants that is, the probability for X to occur given as long as the chance of trouble is quite
considerable attention. that the argument is incorrect, multiplied small even if our arguments are wrong,
This is most obvious, perhaps, from by the probability that the argument is although this is something of a leap of faith.
the current economic meltdown, fuelled indeed incorrect. Ultimately, I’m not entirely sure what to
in large part by ignorance or denial on the This may seem pedantic, but the authors make of this argument, except that it raises
part of many financial actors, including give some empirical weight to it. As they an exceedingly important point that is very
both banks and regulators, of the large risks note, scientific papers published in high- easy to overlook. If there is one constant in
associated with the US housing bubble and profile journals represent arguments that the history of science and human thought,
the proliferation of financial products that have been developed carefully by skilled it is that we’re not very skilled at assessing
amplified and spread these risks globally. researchers. Retracting a paper affects our own certainty and the chance that we’ll
But the issue is far deeper as well, given the a scientist’s reputation. It is never done be hugely surprised in the near future.
great many areas in which public policy, lightly, and usually only in circumstances We have some idea of what we know, and
and private action, requires an accurate when the primary conclusion has been very little idea of what we don’t. More
assessment of the risks associated with seriously undermined by later findings. quantitative thinking about the nature of
rare events. Nevertheless, using data from the our ignorance is certainly in order. ❐
Some months ago I explored the MEDLINE database, earlier researchers
ongoing discussion regarding the found a retraction rate of the order 10−4. Mark Buchanan

books & arts
The autist aesthete

The Strangest extension, all the fundamental building expositions of quantum electrodynamics.
Man: The Hidden blocks of matter. “Dirac did it all before me”, Feynman,
The prediction of the Dirac equation never a man to hide his own light,
Life of Paul Dirac, that the electron must have a positively insistently repeated.
Quantum Genius charged partner — that antimatter Farmelo, a fellow of London’s Science
existed — was one that Dirac himself was Museum, has a firm grasp of the science.
by Graham Farmelo slow to embrace, through “pure cowardice”, He is naturally aided in his narrative by a
he noted later. But Carl Anderson’s stellar cast of supporting characters, but also
FAbeR & FAbeR:
discovery of the positron in cosmic by his own fine line in thumbnail portraits.
2009. 511pp. £22.50
rays in the autumn of 1932 sealed Rutherford, for example, is perfectly
Dirac’s unsought fame, and seemed the portrayed as “a portly fifty-two-year-old
consummation of the quantum theorists’ with a walrus moustache [and] a bay
mastery of their new reality. Together window of a belly... who spoke more loudly
with Erwin Schrödinger, Dirac received than anyone else.”
T
he 1920s and 1930s were physics’ the 1933 Nobel Prize in Physics “for the Dirac does not lend himself so easily
heroic age. In the lecture halls and discovery of new productive forms of to such bold brushstrokes. But Farmelo
seminar rooms of central Europe, a atomic theory”. Characteristically, he succeeds in making clear that Dirac was
cast of semi-legendary characters — Bohr, accepted only on Rutherford’s advice more than the desiccated calculating
Heisenberg, Pauli, Schrödinger et al. — that refusing would garner him even machine of scientific anecdote. He
stormed the citadels of classical physics more publicity. was intensely loyal to those colleagues
and brought them tumbling down. One The same obsessive aesthetic drive he esteemed — Werner Heisenberg,
man seemed oddly unmoved by these epic later brought Dirac darker hours. His for example, whom he defended
upheavals, although he was as much their brilliance eclipsed by a new generation against charges of Nazi sympathies, or
protagonist as any other: Paul Dirac, a man of theorists such as Richard Feynman, he Pyotr Kapitsa, whose enforced detention
with a credible claim to be the greatest spent much of his last decades trying, and in the Soviet Union when on leave
theoretical physicist Britain has produced failing, to bring clarity and beauty to the from Cambridge Dirac, in a rare public
since Isaac Newton. unruly infinities and renormalizations of intervention, unsuccessfully attempted
Part of Dirac’s detachment was modern quantum field theory. That was to end.
down to geography: the University of bitter irony indeed, as Dirac’s work on To the surprise of many, Dirac married.
Cambridge, then a hotbed of experimental the electron had laid the foundation for His union with Eugene Wigner’s ebullient
physics under the leadership of Feynman and others’ ultimately successful sister, Margit, was a union of unequals if
Ernest Rutherford, was in theoretical
terms a relative backwater, far from
on our bookSHeLf
the fervid salons of Göttingen, Zürich Discrete symmetries and cp Violation: From
and Copenhagen where the quantum experiment to theory
revolution was being hatched. Mostly,
however, it was just Dirac. Notoriously by Marco S. Sozzi
reticent and almost pathologically shy, OxFORd UNIVeRsItY pRess: 2008. 592pp. £55
his impassivity perplexed even those With emphasis on experiments and the challenges
who revered his brilliance. “This Dirac”, involved in the discovery process, the book
Niels Bohr grumbled on Dirac’s first visit to provides a thorough discussion of parity, charge
Copenhagen in 1926, “seems to know a lot conjugation, time reversal and cp symmetry in
about physics, but he never says anything.” atomic, nuclear and particle physics.
Such anecdotes are a recurring motif of
Graham Farmelo’s new biography of Dirac, Quantum: einstein, Bohr and the Great Debate
which appears 25 years after Dirac’s death about the nature of reality
in 1984. Dirac had little faith in words:
“I was taught not to start a sentence until by Manjit Kumar
I knew how to finish it”, he once told IcON bOOks Ltd: 2008. 480pp. £20
Bohr. What he lacked in rhetorical art, he An exhaustive narrative history on the fierce
made up for in mathematics. His quest debate that took place in the first half of the
for elegance in physical laws brought twentieth century over the fundamental meaning
him in 1928 his undisputedly greatest of quantum mechanics, with particular focus on
triumph — the equation, described by the lives and personalities of those involved in
Frank Wilczek as the most “achingly its development.
beautiful” in all of physics, with which
he synthesized relativity and quantum
theory to characterize the electron and, by

books & arts
ever there was one. But it was steadfast, if blamed for an unhappy childhood and for modern reader, whatever its scientific
not without turbulence. That is perhaps the suicide of his elder brother, Felix, in worth: Paul Dirac was autistic.
the most unexpected facet of a life that 1925 — an event that evidently affected Whether that is a particularly revealing
supplies other unimagined nuggets — him more deeply than his own words conclusion, or indeed whether another
Dirac’s lifelong love of Mickey Mouse could ever express. ‘outlier’ thus diagnosed adds significantly
cartoons, for example, or his touching But Farmelo sees the answer to Dirac’s to arguments about the nature of genius, is
latter-day reverence for the American oddities more in nature than in nurture. perhaps debatable. Farmelo is wise to relegate
starlet Cher. He lays out the evidence: Dirac’s uncanny the issue to a postscript. Ultimately, the
So what made Dirac the man he was? ability to see the world in entirely abstract answer is neither here nor there in a book
Attempting to answer such questions forms; his total absence of guile — that is a rewarding account of an enigmatic
risks the descent into psychobabble, “childlike, but never childish”, as one man and an intriguing era of physics. ❐
but Farmelo makes a worthy attempt at Cambridge contemporary described it; his
offering a diagnosis. It is clear that Dirac seeming inability to express his emotions revieweD by ricHarD webb
harboured a lifelong resentment of his or empathize with others. The conclusion
father, whose domineering personality he he reaches positively screams itself to the Richard Webb is a features editor for New Scientist.
Lucky for some

New York’s Museum of
exHIbItION
Modern Art (MoMA) is

celebrating its acquisition of
13 recent compositions by
Jasper Johns, with a ‘focus’
exhibition that draws on its
extensive collection of the
US artist’s work.
The newly acquired
© Jasper Johns / Licenced by VaGa, new york, ny

prints, created in 2001 and
all untitled, share a
mathematical theme. Each features the
catenary — the shape formed by a flexible
chain hanging under gravity between
its two supported ends, and rendered
mathematically in terms of the hyperbolic
cosine. Despite its familiarity throughout
centuries of architecture, the recognition of
the catenary form dates only to the second
half of the seventeenth century: as St Paul’s
cathedral was rebuilt from the ashes of
the Great Fire of London, Robert Hooke
declared it the optimal shape of an arch
(although, rather unhelpfully, he only
published an encrypted solution in his the complexity of its creation, built up in Through his use of popular iconography,
lifetime). The word ‘catenary’ — rather encaustic (a wax-based paint) layered over Johns is often recognized — alongside
than the original ‘catenaria’, from the Latin other materials such as newsprint on fabric Robert Rauschenberg and others
catena for ‘chain’ — seems to have first been mounted on plywood. Untitled (pictured) such as French artist Yves Klein — as
used in a 1788 letter from Thomas Jefferson was created in acrylic over aquatint a ‘neo-Dadaist’, in reference to the
to Thomas Paine. and etching. earlier Dada movement that included
Johns is known for his use of images. Alongside the flags, numbers have Marcel Duchamp, Francis Picabia and
In the work pictured here, one of the also been a prevalent motif in Johns’ Man Ray. The MoMA display of its new
untitled 13, the catenary is partnered with work. Among the 85 prints, drawings and prints, in the wider context provided by
impressions of an old-fashioned family painting from MoMA’s collection that form the museum’s collection of Johns’ work,
portrait, a spiral galaxy, and the image the current exhibition are two variations provides, according to exhibition curator
that features most famously in Johns’ of a portfolio entitled 0–9. Dating from Deborah Wye, “a focus that demonstrates
repertoire, the flag. His iconic work of 1963, one is printed in a range of colours his investigatory process especially well.”
1954–55, Flag, is also held by MoMA. This on white paper, the other in grey ink Focus: Jasper Johns is at the Museum of
seemingly straightforward rendering of on brown; both sets were made from a Modern Art, 11 West 53 Street, New York
the star-spangled banner — although with single lithographic stone, Johns having until 16 February 2009. ❐
48 stars, as it predates the admission of reworked its surface after the printing of
Alaska and Hawaii to the Union — belies each numeral. aLiSon wriGHT

books & arts
ever there was one. But it was steadfast, if blamed for an unhappy childhood and for modern reader, whatever its scientific
not without turbulence. That is perhaps the suicide of his elder brother, Felix, in worth: Paul Dirac was autistic.
the most unexpected facet of a life that 1925 — an event that evidently affected Whether that is a particularly revealing
supplies other unimagined nuggets — him more deeply than his own words conclusion, or indeed whether another
Dirac’s lifelong love of Mickey Mouse could ever express. ‘outlier’ thus diagnosed adds significantly
cartoons, for example, or his touching But Farmelo sees the answer to Dirac’s to arguments about the nature of genius, is
latter-day reverence for the American oddities more in nature than in nurture. perhaps debatable. Farmelo is wise to relegate
starlet Cher. He lays out the evidence: Dirac’s uncanny the issue to a postscript. Ultimately, the
So what made Dirac the man he was? ability to see the world in entirely abstract answer is neither here nor there in a book
Attempting to answer such questions forms; his total absence of guile — that is a rewarding account of an enigmatic
risks the descent into psychobabble, “childlike, but never childish”, as one man and an intriguing era of physics. ❐
but Farmelo makes a worthy attempt at Cambridge contemporary described it; his
offering a diagnosis. It is clear that Dirac seeming inability to express his emotions revieweD by ricHarD webb
harboured a lifelong resentment of his or empathize with others. The conclusion
father, whose domineering personality he he reaches positively screams itself to the Richard Webb is a features editor for New Scientist.
Lucky for some

New York’s Museum of
exHIbItION
Modern Art (MoMA) is

celebrating its acquisition of
13 recent compositions by
Jasper Johns, with a ‘focus’
exhibition that draws on its
extensive collection of the
US artist’s work.
The newly acquired
© Jasper Johns / Licenced by VaGa, new york, ny

prints, created in 2001 and
all untitled, share a
mathematical theme. Each features the
catenary — the shape formed by a flexible
chain hanging under gravity between
its two supported ends, and rendered
mathematically in terms of the hyperbolic
cosine. Despite its familiarity throughout
centuries of architecture, the recognition of
the catenary form dates only to the second
half of the seventeenth century: as St Paul’s
cathedral was rebuilt from the ashes of
the Great Fire of London, Robert Hooke
declared it the optimal shape of an arch
(although, rather unhelpfully, he only
published an encrypted solution in his the complexity of its creation, built up in Through his use of popular iconography,
lifetime). The word ‘catenary’ — rather encaustic (a wax-based paint) layered over Johns is often recognized — alongside
than the original ‘catenaria’, from the Latin other materials such as newsprint on fabric Robert Rauschenberg and others
catena for ‘chain’ — seems to have first been mounted on plywood. Untitled (pictured) such as French artist Yves Klein — as
used in a 1788 letter from Thomas Jefferson was created in acrylic over aquatint a ‘neo-Dadaist’, in reference to the
to Thomas Paine. and etching. earlier Dada movement that included
Johns is known for his use of images. Alongside the flags, numbers have Marcel Duchamp, Francis Picabia and
In the work pictured here, one of the also been a prevalent motif in Johns’ Man Ray. The MoMA display of its new
untitled 13, the catenary is partnered with work. Among the 85 prints, drawings and prints, in the wider context provided by
impressions of an old-fashioned family painting from MoMA’s collection that form the museum’s collection of Johns’ work,
portrait, a spiral galaxy, and the image the current exhibition are two variations provides, according to exhibition curator
that features most famously in Johns’ of a portfolio entitled 0–9. Dating from Deborah Wye, “a focus that demonstrates
repertoire, the flag. His iconic work of 1963, one is printed in a range of colours his investigatory process especially well.”
1954–55, Flag, is also held by MoMA. This on white paper, the other in grey ink Focus: Jasper Johns is at the Museum of
seemingly straightforward rendering of on brown; both sets were made from a Modern Art, 11 West 53 Street, New York
the star-spangled banner — although with single lithographic stone, Johns having until 16 February 2009. ❐
48 stars, as it predates the admission of reworked its surface after the printing of
Alaska and Hawaii to the Union — belies each numeral. aLiSon wriGHT

research highlights
Frustration control since been extended into five dimensions weak measurement. If another null result
Nature 456, 898–903 (2008) with the invocation of supersymmetry. A is obtained, the system returns to its initial
whole class of black-hole solutions now state. In the authors’ experiment, this
In geometrically frustrated systems, exists, complemented by further solutions happened 70% of the time.
lattice structure prevents simultaneous generated through applying a so-called
minimization of all local interactions — squashing transformation to the geometry.
antiferromagnetically interacting spins, Cristian Stelea and colleagues have
close to extinction
Nano Lett. doi:10.1021/nl801735y (2008)
for example, cannot be arranged on a been pondering the squashed Gödel black
triangular lattice such that any two of them hole (introduced by Shuang-Qing Wu,
are antiparallel (and thus the pair-wise Phys. Rev. Lett. 100, 121301; 2008) and find
energy minimal). As a result, frustrated that it has some curious properties. For
systems exhibit complex phases of matter. starters, squashing doesn’t necessarily do
Geometric frustration is seen in a wide away with the upper bound on the black-
range of physical and biological systems, hole entropy of the black hole that exists in
and has also been produced artificially. the unsquashed case. Moreover, in lifting
But Yilong Han and colleagues present a the scenario up to ten dimensions, they find
platform that allows a wider exploration of that the T-dual geometry of the squashed
the phenomenon than has been possible black hole is a rotating Kaluza–Klein
so far. black-hole-like spacetime.
Han et al. devised a system of
micrometre-size colloidal spheres, closely
packed between parallel plates that are little
Damage undone
Phys. Rev. Lett. 101, 200401 (2008)
more than a particle’s diameter apart. To
© istockphoto / Felix Moeckel

maximize the free volume, neighbouring The collapse of a wavefunction is one of
spheres move towards opposite walls. But the most intriguing and misunderstood
when arranged in a triangular lattice, the elements of quantum physics. By looking
system gets frustrated in the same manner into the box, we ensure that Schrödinger’s
as an antiferromagnet does. As the diameter cat is either dead or alive — until that
of the spheres can be tuned by controlling point, it is said to be both. Nadav Katz and
their temperature, the dynamical co-workers show in experiment that this
properties of frustration can be studied — ‘collapse’ can be reversed, provided that we
something that is not possible in other look into the box in the right way.
artificial systems. When it comes to observing quantum
behaviour, superconducting circuits are Single quantum emitters, such as single
much better than cats. The state of the molecules and quantum dots, have been
Black-hole squash circuit, one of two possibilities, can be successfully studied using fluorescence
Phys. Rev. D 78, 124006 (2008)
investigated using a ‘weak’ measurement. spectroscopy, but are much harder to
This might not be the question that keeps A null result provides only partial study in the absence of fluorescence.
you awake at night, but what happens when information — and it is this imperfection on An alternative is to use extinction
a charged black hole, embedded in the which the reversal of wavefunction collapse spectroscopy — detecting the extinction
Gödel universe, gets squashed? relies — we know only the likelihood of of a laser beam caused by a single quantum
Kurt Gödel came up with his exact the system being in each state, although emitter in its path — but the small signals
solution to Einstein’s equations in 1949. The one is more likely than the other. Katz et al. can be very difficult to detect on top of a
original version worked in four dimensions, apply a microwave pulse to reverse these noisy background. Philipp Kukura and
but the solution, or ‘Gödel universe’, has probabilities and then repeat the same colleagues show how the situation is
improved by a careful optimization of the
experimental conditions.
protein interference Appl. Phys. Lett. 93, 223904 (2008) The main control ‘knob’ in the
experiments is the interference between
Protein microarrays, which enable the detection of many different proteins in a solution the electric field of the excitation beam and
at once, have revolutionized molecular biology. They work by detecting changes in the the field scattered by the quantum object.
photoluminescence of protein spots that have been designed to fluoresce when they It turns out that the relative phase between
bind to specific ‘target’ proteins. Variations in the fluorescent response from one spot to the two signals is strongly dependent on the
another, however, limit their accuracy and efficiency. phase of the laser beam, which acquires a
Xuefeng Wang and colleagues instead take a non-fluorescent approach. Unlike the phase shift of ±π/2 when propagating from
traditional spots of protein on a flat substrate, their microarrays consist of a uniform the focus to infinity (the detector). Thus, by
layer of protein grown on a patterned film of 140-nm-deep silicon dioxide spots moving the sample away from the focus in
surrounded by a 77-nm-deep silicon dioxide film. The thickness of the spots and the a controlled fashion, Kukura et al. realize a
surrounding film were chosen so that the optical reflectance of each region when significant enhancement of the extinction
coated with a certain thickness of protein is the same. But when a target binds to this signal, which allows them unprecedented
layer, interference effects due to the increased thickness cause the reflectance of the access to photophysical properties such as
spots to increase and of the surrounding film to decrease. photobleaching and blinking in the absence
of fluorescence.

news & views
QUANTUM TOMOGRAPHY
Measured measurement
A method for characterizing quantum measurement devices completes the suite of ‘tomography techniques’,
which should enable us to learn all there is to know about a given quantum-physics experiment.
Markus Aspelmeyer
A
typical day in the quantum-optics lab observations5,6. Recently, this procedure has the number of detected photons. Their
course: two gifted and enthusiastic also been used to quantify the dynamics experimental reconstruction reveals the
third-year students have just of a quantum state via quantum process striking difference between the measurement
proved quantum theory to be wrong. Its tomography7–9, which has become a powerful operators that need to be used to correctly
predictions do not match the experimental analysis tool for quantum-information describe these two systems. But it also
observation. When repeated measurements processing. The calculation of experimental provides a direct quantitative comparison
confirm the results, one of the two already expectation values, however, requires an between the two. This is a particularly
starts developing faint hopes for a Nobel accurate knowledge of the detector used for relevant feature as measurement procedures
Prize — until an instructor enters the scene the tomography measurements. are becoming increasingly important in the
and explains to the students where their To realize quantum-detector tomography, context of quantum metrology and quantum-
measurement had gone wrong. One of the Lundeen et al. have been turning the information processing. One could even
wave plates involved in the measurement of tables. Instead of performing calibrated envisage a calibrated measurement apparatus
photon polarization was labelled incorrectly measurements on identically prepared that is optimized to a specifically tailored
by a previous user. Quantum theory has been quantum systems, they use a well-calibrated quantum circuit. From a purely practical
saved, once again. But still, the experience quantum system as a meter for measuring point of view, the detector tomography
highlights a general problem: how can we the measurement apparatus. The idea is the demonstrated by Lundeen et al. provides a
be sure that a certain measurement device same as above: as experimental outcomes direct operational approach to comparing
used in the laboratory ‘behaves’ according to are determined by both the system and the and calibrating the performance of different
our expectations? Or, to put it the other way measurement device, accurate knowledge photon detectors — without the need for a
around, given a real measurement device, of the system can be used to reconstruct the specific model of the apparatus.
how do we best describe its effect in the density matrix of the measurement apparatus Finally, the advent of quantum-detector
framework of quantum theory? Jeff Lundeen that best describes the experimentally tomography also raises an interesting
and colleagues1 now provide an experimental observed values. To accommodate philosophical question. Does the completion
solution to answer this question: by experimental imperfections (that could lead of the ‘triad’ of state, process and detector
measuring the measurement apparatus. to a non-physical reconstructed density tomography mean that we can indeed
In quantum physics, each experiment matrix), they make use of so-called convex fully specify, from a quantum-theory
resembles a specific ‘observational situation’2, optimization to find the optimal physical perspective, an experiment without any
in which the wavefunction of the experiment representation of the unknown detector. additional assumptions? It is hard to avoid
contains the probabilities of all possible Their experiment is geared towards being caught in a circular argument; state
outcomes for all possible measurements. A characterizing single-photon detectors, tomography requires perfect knowledge
© istockphoto / AmAndA Rohde

complete description of a given quantum- one of the core ingredients in modern about the detector, and process and detector
physics experiment therefore boils down to quantum-optics and photon-based quantum- tomography rely on exact knowledge of the
knowing its wavefunction, which typically information experiments. For these detectors, state — a real chicken-or-egg dilemma.
comprises knowledge about the produced a calibrated quantum meter is naturally What we can
state, its dynamics and about the performed available in the form of coherent states, say is that,
measurement. In the language of modern which describe to a very good
quantum theory, this knowledge is described approximation the states of laser
by (positive semidefinite) operators on a radiation10. A complete set of
Hilbert space. Every working quantum tomographic measurements
physicist is then left with the is provided by probing even if it includes
question: which operators the detector with minor assumptions, the
best represent the system and coherent now-completed triad illustrates
measurement at hand? The system the remarkable feature of quantum physics
side has been covered by quantum- that it does not make sense to divide the
state tomography3,4. There, a sequence states of world into quantum states, quantum
of calibrated measurements on identically different amplitudes. processes and quantum detectors. Once we
prepared states enables the reconstruction Lundeen et al.1 compare two types have decided on the experiment — that is,
of the full density matrix (which provides of detectors: a commercial avalanche once the apparatus is set up — we can obtain
complete knowledge of the quantum system). photodiode that is sensitive to single photons a full description of its overall wavefunction
Using maximum-likelihood estimation but cannot tell how many photons participate and hence on the probabilities of all possible
methods eventually recovers the operator in a detection event, and a home-made outcomes. And it is completely up to us
that comes closest to the experimental single-photon detector that can resolve what to call state, process or detection in

news & views
the experiment. Lundeen and colleagues Austrian Academy of Sciences, Boltzmanngasse 3, 4. James, D. F. V., Kwiat, P. G., Munro, W. J. & White, A. G.
have strikingly demonstrated that quantum A‑1090 Vienna, Austria. Phys. Rev. A 64, 052312 (2001).
tomography serves as the ideal method for e‑mail: markus.aspelmeyer@quantum.at 5. Hradil, Z. Phys. Rev. A 55, R1561–R1564 (1997).
6. Banaszek, K., D’Ariano, G. M., Paris, M. G. A. & Sacchi, M. F.
translating our assembled experiment into Phys. Rev. A 61, 010304 (1999).
References
the language of quantum theory. ❐ 1. Lundeen, J. S. et al. Nature Phys. 5, 27–30 (2009).
7. Chuang I. L. & Nielsen, M. A. J. Mod. Opt. 44, 2455–2467 (1997).
8. Poyatos J. F., Cirac, J. I. & Zoller, P. Phys. Rev. Lett.
2. Heisenberg, W. Der Teil und das Ganze (Piper, Munich, 1969).
78, 390–393 (1997).
Markus Aspelmeyer is at the Institute for 3. Paris, M. G. A. & Rehacek, J. (eds) Quantum State Estimation
9. Lobino, M. et al. Science 322, 563–566 (2008).
Quantum Optics and Quantum Information, (Lecture Notes in Physics Vol. 649) (Springer, 2004). 10. Glauber, R. J. Phys. Rev. Lett. 10, 84–86 (1963).
SPINTRONICS
Electric spin orchestra

Localized electron spins can be manipulated electrically through electric-dipole spin resonance. The ensemble of
mechanisms involved has now been brought under the baton of a unifying theoretical description.
Yasuhiro Tokura
C
entral to any applications in defined GaAs double-dot system subject demonstrated in the EDSR experiment
spintronics and spin-based quantum- to a static magnetic field1–3. The second conducted by Laird and colleagues3.
information processing is the ability crucial ingredient was the application of an The basic mechanism there is similar
to control the spin degree of freedom and, oscillating electric field to induce transitions to the inhomogeneous magnetic-field
in particular, transitions and superpositions between spin-up and spin-down states in case of Pioro-Ladrière et al.2. However,
between ‘spin-up’ and ‘spin-down’ states one of the dots (Fig. 1a) and the subsequent the inhomogeneous magnetic field that
of electrons. One way of achieving this selective charge detection via the Pauli the electron spins experience originates
is through electron spin resonance: an spin-blockade phenomenon — essentially, intrinsically from the so-called ‘Overhauser
oscillating magnetic field is applied with electrons can only tunnel when the electron field’, an effective magnetic field caused by
a frequency tuned to the Zeeman energy spins in the double-dot system are in an fluctuations of the nuclear spins (Fig. 1b).
splitting of localized states, induced by anti-parallel configuration. The probability The unifying approach outlined by
an additional static magnetic field. For for the electrically induced spin-flip Rashba4 is able to describe the SO, magnetic
practical purposes, though, the use of transition, P(t), is an oscillating function of and hyperfine-mediated EDSR phenomena
electric fields would be more desirable as the burst time (the duration of the applied in quantum dots. Using a semi-classical
their high-frequency, on-chip localized oscillating electric field), a phenomenon mean-field theory, the presented theoretical
generation is more readily achieved than known as Rabi oscillations. Since the charge framework describes the spin dynamics
for magnetic fields. However, unlike transfer through the double-dot system is of quantum dot electrons in a sea of
magnetic fields, which couple directly to directly related to P(t) owing to the spin- nuclear spins. The mean-field character
the spin degree of freedom, electric fields blockade mechanism, a carefully executed of the theoretical description essentially
couple directly only to the charge. There transport measurement should therefore means neglecting the quantum nature of
are, nevertheless, indirect ways to ‘entangle’ reveal signatures of spin resonance. the nuclear spins and focusing solely on
charge and spin such that electron spins The third crucial prerequisite for the the contribution of the single-site pair
can be controlled by electric fields. This observation of EDSR is a mechanism that correlation function of the nuclear angular
was recently shown in three independent mediates the coupling between the applied momenta. This assumption can be justified
experimental demonstrations1–3 of electric field and the electrons’ spin degree by the fact that the interaction between
electric-dipole spin resonance (EDSR) — of freedom. The three experiments reported the nuclear spins is weak and higher order
the electrical analogue to electron spin in refs. 1–3 relied on entirely different correlations can be neglected for a large
resonance — in quantum dots. Remarkably, mechanisms. In the experiment by Novack number of nuclei.
it was found that the indirect interaction and co-workers1, the mediating interaction Although the charge detection process
between the applied oscillating electric was based on spin–orbit (SO) coupling, in the experimental EDSR observations
field and the localized electron spins in the most ‘traditional’ way to entangle spin is in principle sensitive to the spin-flip
these experiments is mediated by three and charge in semiconductor spintronics. probability, P(t), in reality, the charge
very different mechanisms. Writing in An alternative coupling mechanism can be detection requires an integration over a
Physical Review B, Emmanuel Rashba now realized by the use of an inhomogeneous large number of electric-field bursts and
provides a unifying theoretical description magnetic field, as observed by Pioro-Ladrière thus, the measured signal, W(t), contains
of the three observed types of EDSR, and and co-workers, in an experiment where information about P(t) averaged over
also describes the crucial role of nuclear a slanted magnetic field was explicitly many pulses. The averaging process covers
spins in determining their properties4. generated on-chip by a micro-magnet 2. timescales exceeding the nuclear-spin
The first main ingredient in the three A third, more intricate, way to entangle diffusion time. W(t) therefore represents
experimental demonstrations of EDSR was the spin and charge degrees of freedom an average of all possible nuclear-
the sensitive ability to create, manipulate is via the hyperfine interaction between spin configurations. In the theoretical
and detect single electron spins localized the electron spin and the nuclear spins description, this corresponds to a Gaussian
in each of the dots in an electrostatically of the atoms in the quantum dots, as integration over the longitudinal and

news & views
the experiment. Lundeen and colleagues Austrian Academy of Sciences, Boltzmanngasse 3, 4. James, D. F. V., Kwiat, P. G., Munro, W. J. & White, A. G.
have strikingly demonstrated that quantum A‑1090 Vienna, Austria. Phys. Rev. A 64, 052312 (2001).
tomography serves as the ideal method for e‑mail: markus.aspelmeyer@quantum.at 5. Hradil, Z. Phys. Rev. A 55, R1561–R1564 (1997).
6. Banaszek, K., D’Ariano, G. M., Paris, M. G. A. & Sacchi, M. F.
translating our assembled experiment into Phys. Rev. A 61, 010304 (1999).
References
the language of quantum theory. ❐ 1. Lundeen, J. S. et al. Nature Phys. 5, 27–30 (2009).
7. Chuang I. L. & Nielsen, M. A. J. Mod. Opt. 44, 2455–2467 (1997).
8. Poyatos J. F., Cirac, J. I. & Zoller, P. Phys. Rev. Lett.
2. Heisenberg, W. Der Teil und das Ganze (Piper, Munich, 1969).
78, 390–393 (1997).
Markus Aspelmeyer is at the Institute for 3. Paris, M. G. A. & Rehacek, J. (eds) Quantum State Estimation
9. Lobino, M. et al. Science 322, 563–566 (2008).
Quantum Optics and Quantum Information, (Lecture Notes in Physics Vol. 649) (Springer, 2004). 10. Glauber, R. J. Phys. Rev. Lett. 10, 84–86 (1963).
SPINTRONICS
Electric spin orchestra

Localized electron spins can be manipulated electrically through electric-dipole spin resonance. The ensemble of
mechanisms involved has now been brought under the baton of a unifying theoretical description.
Yasuhiro Tokura
C
entral to any applications in defined GaAs double-dot system subject demonstrated in the EDSR experiment
spintronics and spin-based quantum- to a static magnetic field1–3. The second conducted by Laird and colleagues3.
information processing is the ability crucial ingredient was the application of an The basic mechanism there is similar
to control the spin degree of freedom and, oscillating electric field to induce transitions to the inhomogeneous magnetic-field
in particular, transitions and superpositions between spin-up and spin-down states in case of Pioro-Ladrière et al.2. However,
between ‘spin-up’ and ‘spin-down’ states one of the dots (Fig. 1a) and the subsequent the inhomogeneous magnetic field that
of electrons. One way of achieving this selective charge detection via the Pauli the electron spins experience originates
is through electron spin resonance: an spin-blockade phenomenon — essentially, intrinsically from the so-called ‘Overhauser
oscillating magnetic field is applied with electrons can only tunnel when the electron field’, an effective magnetic field caused by
a frequency tuned to the Zeeman energy spins in the double-dot system are in an fluctuations of the nuclear spins (Fig. 1b).
splitting of localized states, induced by anti-parallel configuration. The probability The unifying approach outlined by
an additional static magnetic field. For for the electrically induced spin-flip Rashba4 is able to describe the SO, magnetic
practical purposes, though, the use of transition, P(t), is an oscillating function of and hyperfine-mediated EDSR phenomena
electric fields would be more desirable as the burst time (the duration of the applied in quantum dots. Using a semi-classical
their high-frequency, on-chip localized oscillating electric field), a phenomenon mean-field theory, the presented theoretical
generation is more readily achieved than known as Rabi oscillations. Since the charge framework describes the spin dynamics
for magnetic fields. However, unlike transfer through the double-dot system is of quantum dot electrons in a sea of
magnetic fields, which couple directly to directly related to P(t) owing to the spin- nuclear spins. The mean-field character
the spin degree of freedom, electric fields blockade mechanism, a carefully executed of the theoretical description essentially
couple directly only to the charge. There transport measurement should therefore means neglecting the quantum nature of
are, nevertheless, indirect ways to ‘entangle’ reveal signatures of spin resonance. the nuclear spins and focusing solely on
charge and spin such that electron spins The third crucial prerequisite for the the contribution of the single-site pair
can be controlled by electric fields. This observation of EDSR is a mechanism that correlation function of the nuclear angular
was recently shown in three independent mediates the coupling between the applied momenta. This assumption can be justified
experimental demonstrations1–3 of electric field and the electrons’ spin degree by the fact that the interaction between
electric-dipole spin resonance (EDSR) — of freedom. The three experiments reported the nuclear spins is weak and higher order
the electrical analogue to electron spin in refs. 1–3 relied on entirely different correlations can be neglected for a large
resonance — in quantum dots. Remarkably, mechanisms. In the experiment by Novack number of nuclei.
it was found that the indirect interaction and co-workers1, the mediating interaction Although the charge detection process
between the applied oscillating electric was based on spin–orbit (SO) coupling, in the experimental EDSR observations
field and the localized electron spins in the most ‘traditional’ way to entangle spin is in principle sensitive to the spin-flip
these experiments is mediated by three and charge in semiconductor spintronics. probability, P(t), in reality, the charge
very different mechanisms. Writing in An alternative coupling mechanism can be detection requires an integration over a
Physical Review B, Emmanuel Rashba now realized by the use of an inhomogeneous large number of electric-field bursts and
provides a unifying theoretical description magnetic field, as observed by Pioro-Ladrière thus, the measured signal, W(t), contains
of the three observed types of EDSR, and and co-workers, in an experiment where information about P(t) averaged over
also describes the crucial role of nuclear a slanted magnetic field was explicitly many pulses. The averaging process covers
spins in determining their properties4. generated on-chip by a micro-magnet 2. timescales exceeding the nuclear-spin
The first main ingredient in the three A third, more intricate, way to entangle diffusion time. W(t) therefore represents
experimental demonstrations of EDSR was the spin and charge degrees of freedom an average of all possible nuclear-
the sensitive ability to create, manipulate is via the hyperfine interaction between spin configurations. In the theoretical
and detect single electron spins localized the electron spin and the nuclear spins description, this corresponds to a Gaussian
in each of the dots in an electrostatically of the atoms in the quantum dots, as integration over the longitudinal and

news & views
a b E c
1
W(t)
B0 0
Burst time, t
Figure 1 | Electrically manipulating spins in quantum dots. a, Band diagram of the experimental quantum dot system. An applied oscillating electric field induces
spin-flips via EDSR, which manifest as oscillations in the transport characteristics owing to the spin-blockade phenomenon. b, A localized electron spin interacts
with a nuclear spin bath, leading to decoherence. c, Time evolution of the probability W(t) of the dot’s electron spin being in the ‘up’ state when the mediating
mechanism is due to SO/inhomogeneous magnetic field (solid line) or a gradient of the Overhauser field (dashed line).
transverse components of the nuclear spins’ important. As a characteristic signature nuclear spins. One is to manipulate the
magnetization with respect to the externally differentiating it from the SO and magnetic nuclear spin magnetization and to reduce
applied static magnetic field. cases, W(t) was theoretically shown to its fluctuations by a controlled interaction
The details of the averaging process have a monotonic increase with burst with electron spins7,8. The dynamics in such
turn out to have direct consequences for time (dashed line in Fig. 1c.) in agreement a composite system would be an exciting
experimental observations. For example, with the experimental observations by field of future research. The other possible,
analytical formulas for W(t) in the Laird et al.3. but technologically challenging, avenue is
asymptotic limit for the SO coupling and The presented mean-field theoretical to switch to other systems, such as p-type
magnetic-field-gradient mediated EDSR approach seems to explain the GaAs, Si/Ge or carbon nanotubes, which
scenarios show oscillatory behaviour 4, a experimentally observed EDSR transport have much weaker interaction strengths
manifestation of Rabi oscillations observed characteristics in W(t) very well. However, with the nuclear spins9. ❐
in experiments1,5. More importantly, the there are still challenges that remain to be
theory shows that W(t) is damped at a rate addressed. One important aspect relates Yasuhiro Tokura is at the Basic Research
inversely proportional to the square-root to the quantum nature of the nuclear Laboratories, NTT Corporation and Quantum
of the burst time, t, with a universal phase spins, which was not considered fully in Spin Information Project, ICORP–JST, Atsugi‑shi,
shift with respect to the expected sinusoidal the presented theoretical framework, but 243‑0198, Japan.
squared oscillation (solid line in Fig. 1c). would become important for large external e‑mail: tokura@nttbrl.jp
This characteristic damping behaviour has magnetic fields6. In order to clarify the
References
also been observed experimentally 5. Its importance of hyperfine interactions with 1. Nowack, K. C., Koppens, F. H. L., Nazarov, Y. V. &
origin is explained by the theory in terms the nuclear spins on the electron spins’ Vandersypen, L. M. K. Science 318, 1430–1433 (2007).
of ‘decoherence’ owing to fluctuations in coherence, further experimental and 2. Pioro-Ladrière, M. et al. Nature Phys. 4, 776–779 (2008).
only the longitudinal component of the theoretical studies are needed. Theoretical 3. Laird, E. A. et al. Phys. Rev. Lett. 99, 246601 (2007).
nuclear spins’ magnetization. analyses along these lines would be 4. Rashba, E. I. Phys. Rev. B 78, 195, 302 (2008).
5. Koppens F. H. L. et al. Phys. Rev. Lett. 99, 106803 (2007).
In contrast, for the case of an challenging because the hyperfine coupling
6. Coish, W. A., Loss, D., Yuzbashyan, E. A. & Altshuler, B. L.
Overhauser-field-mediated interaction strength is not uniform over the nuclear J. Appl. Phys. 101, 081715 (2007).
between the electron spin and an applied spins involved (about 105–106) owing 7. J. Baugh, Y. Kitamura, K. Ono & S. Tarucha, Phys. Rev. Lett.
electric field, fluctuations in both the to the localized nature of electron wave 99, 096804 (2007).
longitudinal and the transverse components functions. There seem to be at least two 8. J. R. Petta et al. Phys. Rev. Lett. 100, 067601 (2008).
of the nuclear spin magnetization are ways to avoid decoherence due to the 9. Schroer, M. D. & Petta, J. R. Nature Phys. 4, 516–518 (2008).
CONDENSED MATTER
How do your crystals grow?

More than 100 years ago, Wilhelm Ostwald predicted that crystalline structures would grow from the melt via a
series of unstable states — now this cascade has been observed directly in an inorganic semiconductor.
Simon J. L. Billinge
C
rystals nucleate out of molten studied for centuries. On page 68 of this stable state — an observation that supports
material and then grow to make issue1, Chung and colleagues show for the Ostwald’s ‘rule of stages’.
the beautiful, perfect, final product. first time the formation of a single crystallite Below the freezing-temperature, groups
How this happens is surprisingly complex through a cascade of intermediate forms of atoms spontaneously congregate and
and difficult to understand despite being on the way to the final thermodynamically separate from the melt, rather like clusters

news & views
a b E c
1
W(t)
B0 0
Burst time, t
Figure 1 | Electrically manipulating spins in quantum dots. a, Band diagram of the experimental quantum dot system. An applied oscillating electric field induces
spin-flips via EDSR, which manifest as oscillations in the transport characteristics owing to the spin-blockade phenomenon. b, A localized electron spin interacts
with a nuclear spin bath, leading to decoherence. c, Time evolution of the probability W(t) of the dot’s electron spin being in the ‘up’ state when the mediating
mechanism is due to SO/inhomogeneous magnetic field (solid line) or a gradient of the Overhauser field (dashed line).
transverse components of the nuclear spins’ important. As a characteristic signature nuclear spins. One is to manipulate the
magnetization with respect to the externally differentiating it from the SO and magnetic nuclear spin magnetization and to reduce
applied static magnetic field. cases, W(t) was theoretically shown to its fluctuations by a controlled interaction
The details of the averaging process have a monotonic increase with burst with electron spins7,8. The dynamics in such
turn out to have direct consequences for time (dashed line in Fig. 1c.) in agreement a composite system would be an exciting
experimental observations. For example, with the experimental observations by field of future research. The other possible,
analytical formulas for W(t) in the Laird et al.3. but technologically challenging, avenue is
asymptotic limit for the SO coupling and The presented mean-field theoretical to switch to other systems, such as p-type
magnetic-field-gradient mediated EDSR approach seems to explain the GaAs, Si/Ge or carbon nanotubes, which
scenarios show oscillatory behaviour 4, a experimentally observed EDSR transport have much weaker interaction strengths
manifestation of Rabi oscillations observed characteristics in W(t) very well. However, with the nuclear spins9. ❐
in experiments1,5. More importantly, the there are still challenges that remain to be
theory shows that W(t) is damped at a rate addressed. One important aspect relates Yasuhiro Tokura is at the Basic Research
inversely proportional to the square-root to the quantum nature of the nuclear Laboratories, NTT Corporation and Quantum
of the burst time, t, with a universal phase spins, which was not considered fully in Spin Information Project, ICORP–JST, Atsugi‑shi,
shift with respect to the expected sinusoidal the presented theoretical framework, but 243‑0198, Japan.
squared oscillation (solid line in Fig. 1c). would become important for large external e‑mail: tokura@nttbrl.jp
This characteristic damping behaviour has magnetic fields6. In order to clarify the
References
also been observed experimentally 5. Its importance of hyperfine interactions with 1. Nowack, K. C., Koppens, F. H. L., Nazarov, Y. V. &
origin is explained by the theory in terms the nuclear spins on the electron spins’ Vandersypen, L. M. K. Science 318, 1430–1433 (2007).
of ‘decoherence’ owing to fluctuations in coherence, further experimental and 2. Pioro-Ladrière, M. et al. Nature Phys. 4, 776–779 (2008).
only the longitudinal component of the theoretical studies are needed. Theoretical 3. Laird, E. A. et al. Phys. Rev. Lett. 99, 246601 (2007).
nuclear spins’ magnetization. analyses along these lines would be 4. Rashba, E. I. Phys. Rev. B 78, 195, 302 (2008).
5. Koppens F. H. L. et al. Phys. Rev. Lett. 99, 106803 (2007).
In contrast, for the case of an challenging because the hyperfine coupling
6. Coish, W. A., Loss, D., Yuzbashyan, E. A. & Altshuler, B. L.
Overhauser-field-mediated interaction strength is not uniform over the nuclear J. Appl. Phys. 101, 081715 (2007).
between the electron spin and an applied spins involved (about 105–106) owing 7. J. Baugh, Y. Kitamura, K. Ono & S. Tarucha, Phys. Rev. Lett.
electric field, fluctuations in both the to the localized nature of electron wave 99, 096804 (2007).
longitudinal and the transverse components functions. There seem to be at least two 8. J. R. Petta et al. Phys. Rev. Lett. 100, 067601 (2008).
of the nuclear spin magnetization are ways to avoid decoherence due to the 9. Schroer, M. D. & Petta, J. R. Nature Phys. 4, 516–518 (2008).
CONDENSED MATTER
How do your crystals grow?

More than 100 years ago, Wilhelm Ostwald predicted that crystalline structures would grow from the melt via a
series of unstable states — now this cascade has been observed directly in an inorganic semiconductor.
Simon J. L. Billinge
C
rystals nucleate out of molten studied for centuries. On page 68 of this stable state — an observation that supports
material and then grow to make issue1, Chung and colleagues show for the Ostwald’s ‘rule of stages’.
the beautiful, perfect, final product. first time the formation of a single crystallite Below the freezing-temperature, groups
How this happens is surprisingly complex through a cascade of intermediate forms of atoms spontaneously congregate and
and difficult to understand despite being on the way to the final thermodynamically separate from the melt, rather like clusters

news & views
5 nm
Figure 1 | It happens in stages. This ‘movie’ of a single crystallite as it cascades through different states towards its stable form (left to right) was created by
Chung et al.1 from high-resolution electron micrographs. Atomic displacements are shown taking place over a period of four minutes.
of people conversing at a crowded party. seeking previously unknown crystal forms, So where do we go from here? The
As a group forms, others join and the for example in pharmaceutical science. experimental results are certainly
group grows, but it is still unstable — it But the rule is quite an extrapolation from tantalizingly supportive of Ostwald’s
can dissolve and dissipate again — until it observations — is it right? It has taken prescient insight: we now have a movie
reaches a certain size, the critical radius. 111 years to be able to follow the cascade of a single crystallite cascading through
At this point, everyone at the party realizes as it takes place in a single crystallite, intermediate states on its way to the stable
that this is the place to be and the group but Chung et al.1 present the first direct form. Many questions remain, however.
grows rapidly, limited only by factors such experimental evidence for a process like this Is it a universal law or a rare fortuitous
as the ability of the crowd to diffuse over actually occurring in inorganic materials. occurrence in this particular material
and join it. The critical radius is a tipping Why did it take 111 years? The system? Crystallization is a highly complex
point where the bulk energy of the more experimental difficulties are formidable. process and no-one has found any sound
stable crystalline core overcomes the extra The nuclei form and dissolve very rapidly theoretical basis for such a general
cost of forming an interface. For large in the melt: how can we capture the principle as Ostwald’s rule6,7. Is the same
clusters the bulk energy always wins, but dynamic process at all, let alone follow an behaviour also observed from the melt,
below the critical radius the cost of making evolution through multiple intermediate as Ostwald originally envisaged it? Is the
the interface is greater, resulting in an states? Also, many thousands of nuclei are result experiment-dependent? Studying
energetic barrier to crystal formation, even forming at any given time: how can we a crystallite at the very edge of a thin
in the undercooled or supersaturated state follow the time-dependent behaviour of a section in vacuum under an intense, highly
where, thermodynamically, the crystal single nucleus out of the whole ensemble? interacting, electron beam is necessary to
is preferred. Atoms must be thermally A number of creative experimental insights make the experiment a success, but these are
excited to overcome the energy barrier, a led Chung et al. to the solution. not exactly optimal conditions for studying
stochastic process. First, they chose to study near-equilibrium behaviour.
It is worth pondering whether there are recrystallization from an amorphous state, Finally, and most tantalizing of all,
any intermediate states that can be accessed where the dynamics of the process can be Ostwald’s prediction was very precise: the
that allow one to circumvent the high slowed down to experimentally accessible cascade is from the least stable accessible
activation barrier. These are not the most timescales. Indeed, the timescale can be crystalline form to the most stable. In this
thermodynamically stable states, but they controlled to some extent with temperature, experiment, the existence of intermediate
can be accessed by thermal excitation with provided the temperature stays above that states is established, but Ostwald’s full
greater probability because the activation for recrystallization and below that for law of stages remains unverified. We have
barrier to form them is lower. Once the melting. Second, they used a local probe observed the crowd dynamics at the packed
intermediate state is formed, there may of structure, electron diffraction, which party, but the social dynamics that give rise
be a pathway to the thermodynamically enables individual nuclei to be isolated and to this behaviour are still a distant, barely
stable state with a lower activation barrier. followed with time. It is quite difficult to get intelligible, conversation. ❐
Kinetically, this will then be the preferred quantitative 3D-structure solutions using
path for the transformation to take. electrons (although recent progress in this Simon J. L. Billinge is in the Department of Applied
In 1897, Wilhelm Ostwald postulated regard has been rapid4), but quantitative Physics and Applied Mathematics, Columbia
just such a process for crystal formation2. determination of the lattice symmetry and University, New York, New York 10027, USA and
Crystals are notorious for having multiple parameters enables structural forms to be the Condensed Matter Physics and Materials
forms with nearly equal energy; this is the differentiated (Fig. 1) — all that is required Science Department, Brookhaven National
reason for polymorphism and structural here. The authors were richly rewarded Laboratory, Upton, New York 11973, USA.
phase transitions at moderate temperatures. with images and diffraction patterns of e‑mail: sb2896@columbia.edu
Based on observations of the crystallization a cascade of structural forms in LiFePO4 References
of a number of systems from supercooled evident before the stable crystal structure 1. Chung, S.-Y., Kim, Y.-M., Kim, J.-G. & Kim, Y.-J. Nature Phys.
melts and supersaturated solutions, Ostwald was finally reached (with a wink to history: 5, 68–73 (2009).
postulated that intermediate states form polymorphism itself was discovered in 2. Ostwald, W. Z. Z. Phys. Chem. 22, 289–330 (1897).
3. Threlfall, T. Org. Process Res. Dev. 7, 1017–1027 (2003).
a cascade from the least stable available phosphate/arsenate salts by Mitscherlisch5
4. Zou, X. & Hovmöller, S. Acta Crystallogr. A 64, 149–160 (2008).
crystalline form to the most stable. This in 1822). As a bonus, additional 5. Mitscherlich, E. Ann. Chim. Phys. 19, 350–419 (1822).
became known as the ‘Ostwald rule of information was also revealed about the 6. Bernstein, J. Polymorphism in Molecular Crystals (Oxford Univ.
stages’ (refs 2,3). For many years this rule mechanism of the transformations between Press, 2002).
has been a guiding principle for chemists intermediate states. 7. ten Wolde, P. R. & Frenkel, D. Science 277, 1975–1978 (1997).

news & views
HIGH-TEMPERATURE SUPERCONDUCTORS
Vortices wiggled and dragged

The ability to manipulate an individual superconducting vortex represents a powerful tool for studying the
dynamics of vortices and the superconductors that support them. It could also lead to the development of a new
class of fluxon-based electronics.
Charles Reichhardt
W
hen a sufficiently strong a b c
MFM tip
magnetic field is applied to a
superconductor, some of the
field can pierce it through the generation
of magnetic vortices, each of which
contains a quantized amount of magnetic Defect
flux. Although the superconducting state
of the material outside each vortex is
maintained (and destroyed within each
vortex), the interaction of vortices with
a current passing through the material Vortex
Superconductor
can cause them to move, dissipating
energy and thereby generating a source
Figure 1 | The ability to grab the end of an individual vortex in a superconductor using the tip of a
of electrical resistance. The consequent
magnetic force microscope (MFM) provides the opportunity to study many more vortex phenomena
breakdown in the lossless transmission
than is possible by conventional bulk measurement techniques. a, To demonstrate one such
of electrical current creates problems for
possibility, Auslaender et al.1 use this technique to measure the anisotropy of the pinning force
applications in which superconductors
as a vortex is dragged through a random distribution of point-like defects. The technique should
are used. But studying the dynamics of
allow the study of other phenomena such as b, the behaviour of a pinned vortex as it is peeled
vortices to develop new ways in which to
from a line-defect created by ion-beam damage, and c, the interaction of multiple vortices twisted
minimize, or even exploit, their effects
around each other.
is not easy. On page 35 of this issue1,
Auslaender et al. show that they can grab
hold of a single vortex using the tip of a When this tip is brought close to a sample, an elastic string? How does a vortex de-pin
magnetic force microscope (MFM), and, the presence of magnetic features at the from different defect topologies, such as a
by observing its response as they drag it surface of the sample will cause it to be line-like pinning site generated with ion
back and forth through a superconductor, deflected. By measuring the deflection of irradiation (as in Fig. 1b)? Does the vortex
can probe both the dynamic behaviour the tip as it is scanned across a surface, begin to unzip from top of the defect like a
of the vortex and the properties of the images of these features can be constructed. zipper, or form a break-away loop vortex2?
superconductor itself. This is the basic principle of magnetic force There is also the question of whether
A vortex in a superconductor is a line- (and, indeed, conventional atomic force) vortices in the glass or liquid phase can
like structure that extends all the way from microscopy. But the MFM tip will also exert entangle around each other, as in a glassy
one side of a superconducting sample to a force on these features, the magnitude of polymer system, or whether they can freely
the other. The most effective means found which can be controlled by the distance of cut through each other 3. It should now be
so far for limiting the movement of these the tip from the surface. Auslaender et al. possible to answer these questions by, for
vortices (as a result of thermal fluctuations find that they can use this force to grab example, dragging a vortex off a line defect
or the passage of an electrical current) is to hold of, and drag, the portion of the vortex or artificially wrapping one vortex around
introduce defects into a superconductor’s line nearest to the surface of the sample, another (Fig. 1c).
microscopic structure, which serve to while the remainder of the vortex deeper Already, the authors observe that the
‘pin’ the vortices in place. Such pinning in the bulk remains pinned (Fig. 1a). pinning force on the vortices they probe
is not perfect, but attempts to understand Moreover, they find that by ‘wiggling’ is anisotropic, which may mean that
and better control the processes involved the MFM sideways (perpendicular to the the defects in the sample responsible
have been limited by the fact that it has dragging direction), they can move the for pinning the vortices are clustered
only been possible to study them through vortex over a much larger distance than is oxygen defects. And because the vortices
bulk measurements of the average possible without such wiggling — similar couple to inhomogeneities in the sample,
behaviour of thousands of vortices across in effect to trying to disentangle a power moving individual vortices may also be
a sample. cable from a mass of other cables behind useful for shedding light on the intrinsic
Auslaender et al.1 address this limitation your computer. superconductivity mechanism underlying
by developing a technique that enables The ability to manipulate individual high-temperature materials. For example,
them to both image and manipulate a vortices should help answer a number of by measuring the force required to drag
single vortex using the probe of an MFM, open questions that could not be resolved a vortex, it may be possible to determine
which consists of a very sharp magnetic tip with bulk measurements. For example, to whether the vortex is moving over stripe,
attached to the end of a flexible cantilever. what extent does a vortex line behave like checkerboard or fluctuating charge-ordered

news & views
structures of the type predicted in Individual particle manipulation role of heterogeneities in determining the
some theories4. has been widely developed as a tool in sample behaviour. ❐
The ability to manipulate individual systems with larger-scale objects, such as
vortices might also lead to the realization in biological matter where DNA strands Charles Reichhardt is in the Los Alamos National
of new types of flux-line-based devices, are mechanically manipulated7, or in Laboratory, T‑13, MS B213, Los Alamos,
called fluxtronics. There have been soft-matter systems where grabbing and New Mexico 87545, USA.
proposals for devices using vortices shaking a single colloid can be used as e‑mail: charlesr@cnls.lanl.gov
as the elementary units in classical a microrheological probe8. The work of
logic devices5, as well as proposals to Auslaender et al.1 points to the feasibility References
1. Auslaender, O. M. et al. Nature Phys. 5, 35–39 (2009).
manipulate individual spins or charges of using many of these ideas on a much
2. Kafri, Y., Nelson, D. R. & Polkovnikov, A. Phys. Rev. B
by coupling the vortex to these objects smaller scale, not only for vortex systems, 76, 144501 (2007).
and then manipulating the vortex line6. but also for a wider class of solid states. 3. Olson Reichhardt, C. J. & Hastings, M. B. Phys. Rev. Lett.
Similar techniques for moving individual One can imagine using multiple probe 92, 157002 (2004).
line-like objects could be used in other tips to create a nanorheological probe for 4. Kivelson, S. A. et al. Rev. Mod. Phys. 75, 1201–1241 (2003).
systems, such as dragging or shaking vortex or quantum glasses. These local 5. Hastings, M. B., Olson Reichhardt, C. J. & Reichhardt, C.
Phys. Rev. Lett. 90, 247004 (2003).
individual domain walls in magnets, or probe techniques may usher in a new era
6. Berciu, M., Rappoport, T. G. & Janko, B. Nature
moving dislocation lines in materials where the individual manipulation of 435, 71–75 (2005).
to create specific patterns that will quantum objects could be used to explore 7. Bustamante, C., Bryant, Z. & Smither, S. B. Nature
enhance the mechanical properties of local and non-local responses, quantum 421, 423–427 (2003).
the material. entanglement, many-body effects and the 8. Levine, A. J. & Lubensky, T. C. Phys. Rev. Lett. 85, 1774–1777 (2000).
MATTER WAVES
Cloaking matters
on its head: what type of barrier will act
on a particle with a specific energy to
recreate the cloaking effect. Thinking in
two dimensions, they propose a series of
rings, each with a different potential, that
together mask whatever is within from a
probing matter wave.
The idea of a cloak for matter waves
is not new, however, construction via the
approach published previously Shuang
Zhang and colleagues (Phys. Rev. Lett.
100, 123002; 2008) is likely to be
difficult in practice. Greenleaf et al.
hope that their concept will offer an
easier path.
Its simplicity comes at a price — the
cloaking is not perfect — but even this
drawback has its advantage: particles at
an energy at which the device doesn’t
work can be trapped within the cloak.
Such traps have already proved a boon for
GettY
fundamental research on atomic physics,

and a novel approach is always likely to
The idea of an invisibility cloak — a particles and those of waves, such as come in handy.
device capable of bending light, keeping light, it is not outrageous to wonder — as The matter cloak may be a long way
whatever is inside safe from prying Allan Greenleaf and colleagues do from fruition, but it is a further example of
eyes — is one that was always destined to (Phys. Rev. Lett. 101, 220404; 2008) — how optical phenomena can be translated
capture the public imagination. Over the whether these same cloaking ideas can to matter waves. The idea complements
past few years it has developed from sci-fi be applied to matter. proposals to use graphene as a
mainstay to experimental fact (albeit How does a particle react to a Veselago’s lens for electron beams and
only under certain conditions), and the potential barrier of a certain size and ‘extraordinary’ transmission of rubidium
excitement in the popular press has just shape? It is a standard question posed atoms through an array of sub-de Broglie-
kept growing. to undergraduate students when wavelength slits.
As quantum mechanics blurs the introducing Schrödinger’s equation. But
boundaries between the properties of Greenleaf et al. have turned this problem DAVID GEVAUx

news & views
structures of the type predicted in Individual particle manipulation role of heterogeneities in determining the
some theories4. has been widely developed as a tool in sample behaviour. ❐
The ability to manipulate individual systems with larger-scale objects, such as
vortices might also lead to the realization in biological matter where DNA strands Charles Reichhardt is in the Los Alamos National
of new types of flux-line-based devices, are mechanically manipulated7, or in Laboratory, T‑13, MS B213, Los Alamos,
called fluxtronics. There have been soft-matter systems where grabbing and New Mexico 87545, USA.
proposals for devices using vortices shaking a single colloid can be used as e‑mail: charlesr@cnls.lanl.gov
as the elementary units in classical a microrheological probe8. The work of
logic devices5, as well as proposals to Auslaender et al.1 points to the feasibility References
1. Auslaender, O. M. et al. Nature Phys. 5, 35–39 (2009).
manipulate individual spins or charges of using many of these ideas on a much
2. Kafri, Y., Nelson, D. R. & Polkovnikov, A. Phys. Rev. B
by coupling the vortex to these objects smaller scale, not only for vortex systems, 76, 144501 (2007).
and then manipulating the vortex line6. but also for a wider class of solid states. 3. Olson Reichhardt, C. J. & Hastings, M. B. Phys. Rev. Lett.
Similar techniques for moving individual One can imagine using multiple probe 92, 157002 (2004).
line-like objects could be used in other tips to create a nanorheological probe for 4. Kivelson, S. A. et al. Rev. Mod. Phys. 75, 1201–1241 (2003).
systems, such as dragging or shaking vortex or quantum glasses. These local 5. Hastings, M. B., Olson Reichhardt, C. J. & Reichhardt, C.
Phys. Rev. Lett. 90, 247004 (2003).
individual domain walls in magnets, or probe techniques may usher in a new era
6. Berciu, M., Rappoport, T. G. & Janko, B. Nature
moving dislocation lines in materials where the individual manipulation of 435, 71–75 (2005).
to create specific patterns that will quantum objects could be used to explore 7. Bustamante, C., Bryant, Z. & Smither, S. B. Nature
enhance the mechanical properties of local and non-local responses, quantum 421, 423–427 (2003).
the material. entanglement, many-body effects and the 8. Levine, A. J. & Lubensky, T. C. Phys. Rev. Lett. 85, 1774–1777 (2000).
MATTER WAVES
Cloaking matters
on its head: what type of barrier will act
on a particle with a specific energy to
recreate the cloaking effect. Thinking in
two dimensions, they propose a series of
rings, each with a different potential, that
together mask whatever is within from a
probing matter wave.
The idea of a cloak for matter waves
is not new, however, construction via the
approach published previously Shuang
Zhang and colleagues (Phys. Rev. Lett.
100, 123002; 2008) is likely to be
difficult in practice. Greenleaf et al.
hope that their concept will offer an
easier path.
Its simplicity comes at a price — the
cloaking is not perfect — but even this
drawback has its advantage: particles at
an energy at which the device doesn’t
work can be trapped within the cloak.
Such traps have already proved a boon for
GettY
fundamental research on atomic physics,

and a novel approach is always likely to
The idea of an invisibility cloak — a particles and those of waves, such as come in handy.
device capable of bending light, keeping light, it is not outrageous to wonder — as The matter cloak may be a long way
whatever is inside safe from prying Allan Greenleaf and colleagues do from fruition, but it is a further example of
eyes — is one that was always destined to (Phys. Rev. Lett. 101, 220404; 2008) — how optical phenomena can be translated
capture the public imagination. Over the whether these same cloaking ideas can to matter waves. The idea complements
past few years it has developed from sci-fi be applied to matter. proposals to use graphene as a
mainstay to experimental fact (albeit How does a particle react to a Veselago’s lens for electron beams and
only under certain conditions), and the potential barrier of a certain size and ‘extraordinary’ transmission of rubidium
excitement in the popular press has just shape? It is a standard question posed atoms through an array of sub-de Broglie-
kept growing. to undergraduate students when wavelength slits.
As quantum mechanics blurs the introducing Schrödinger’s equation. But
boundaries between the properties of Greenleaf et al. have turned this problem DAVID GEVAUx

news & views
x-RAY IMAGING
Caught in a spin
An algorithm that enables a protein’s molecular structure to be determined from the faintest of diffraction patterns
could increase the potential of next-generation X-ray sources.
Keith A. Nugent
T
he next few years will see the new
X-ray free-electron lasers switched
on, with the first being the Linac
Coherent Light Source in California.
These facilities are the latest stage in the
development of X-ray sources, which has
seen the available coherent flux increase
by an order of magnitude every eighteen
months for the past twenty years, opening
scientific areas previously inaccessible to
experimental investigation. But although
the advent of such facilities is significant,
several challenges need to be overcome
for them to reach their full potential.
On page 64 of this issue1, Fung and
colleagues report an innovative approach to
the problem of orienting exceptionally weak
diffraction patterns, and argue that their
approach overcomes another obstacle in
the quest to image single biomolecules with
X-ray lasers. Figure 1 | Electron density isosurface of a chignolin protein molecule, reconstructed from 72,000
The ability to use X-ray crystallography simulated randomly oriented diffraction patterns (Reproduced from Fung et al.1).
to determine the structure of complex
biomolecules at the atomic level is one of
the greatest triumphs of X-ray science. As There are questions, for example, of diffraction patterns where the number of
the name implies, crystallography requires whether any molecule, let alone a fragile scattering events required is around three
that the molecule of interest is capable of biomolecule, could survive in the beam orders of magnitude lower than those
being grown into a well-ordered crystal. As of an XFEL for long enough to allow obtained using electron diffraction.
well as being fundamental to the technique, collection of a useful signal, and also Fung et al.1 propose a promising strategy
this ensures that the X-ray dose, and whether the calculated structure would to bridge this gap in the experimental
therefore potential for radiation damage, is bear any resemblance to its original protocol. The key to their approach involves
distributed over many identical copies of structure. Chapman et al. have already recognizing that a molecule has only three
a molecule. Unfortunately, this imposes a shown that it is indeed possible to degrees of rotational freedom, and so its
severe restriction on the types of molecule reconstruct the image of an object from diffraction patterns must occupy a three-
whose structure can be determined, as light scattered in the moments before dimensional subset of the set of all possible
many important molecules — most notably its obliteration4. But this was for a flat intensity measurements. They then use
the entire class of membrane proteins — do structure whose orientation was well Bayesian statistics to classify this subspace,
not readily form crystals. defined and held steady in the incident and hence determine the orientation of
In 2000, Neutze et al. proposed that beam. Repeating the feat for a complex, the molecules making up the diffraction
the advent of the X-ray free-electron laser three-dimensional molecule that is both patterns. Phasing of the diffraction data
(XFEL) could provide a solution2. Free- falling and rotating in free space (making and recovery of the molecular structure can
electron lasers produce incredibly intense it impossible to know its orientation) is a then follow.
bursts of coherent radiation. Combining this much more challenging task. The authors demonstrate their approach
with recently developed coherent diffractive Methods for controlling the orientation by recovering the structure from 72,000
imaging techniques3, Neutze et al. suggested of molecules have been proposed5 and may simulated diffraction patterns generated
that it might be possible to calculate go some way to solving this problem. And for a chignolin protein molecule. They
molecular structure from the diffraction techniques for orienting weak diffraction estimate that their method will work
signal of a single molecule dropped patterns obtained from single particles for scattered photon fluxes as low
through the beam of an XFEL. As well as have been developed in the context of as 10–2 photons per detector pixel —
avoiding the need for crystals, the apparent electron microscopy. But these patterns are corresponding, the authors estimate, to
simplicity of this approach has great appeal. much stronger than those anticipated from the expected diffraction signal generated
In practice, though, things turn out to be a FEL experiment — the single molecule from of a single XFEL pulse by an 8-kDa
more complicated. X-ray experiments will need to manage molecule. If the approach by Fung et al.

news & views
turns out to be experimentally achievable, complicate their analysis. The technique Keith A. Nugent is in the School of Physics,
it will offer a very straightforward route also assumes that the X-ray scattering is well University of Melbourne, Cnr Tin Alley & Swanston
to the direct XFEL-based recovery described by established diffraction physics. Street, Parkville, VIC 3010, Australia.
of the structures of even relatively Whether this will hold at the high energies e‑mail:keithan@unimelb.edu.au
small biomolecules. and extreme intensities generated by an
References
What the authors have not yet shown XFEL remains to be seen. Indeed, there are 1. Fung, R., Shneerson, V., Saldin, D. K. & Ourmazd, A. Nature Phys.
is the ability to realize a full recovery of already indications that new descriptions of 5, 64–67 (2009).
orientation in three-dimensions, which they the diffraction physics may be needed6. But 2. Neutze, R., Wouts, R., van der Spoel, D., Weckert, E. & Hajdu, J.
argue is a limitation set by the computing these are challenges for the future — for the Nature 406, 752–757 (2000).
power available to them rather than by present, there is every reason to expect that 3. Miao, J. W., Charalambous, P., Kirz, J. & Sayre, D. Nature
400, 342–344 (1999).
fundamental limitations in their approach. the approach provided by Fung et al. has
4. Chapman, H. N. et al. Nature Phys. 2, 839–843 (2006).
This represents an important unresolved removed yet another obstacle to realizing 5. Wu, J. S. & Spence, J. C. H. J. Opt. Soc. Am. A
question for further work. It is possible that the full potential of the next generation of 22, 1453–1459 (2005).
internal motions within the molecule will XFEL sources. ❐ 6. Sorokin, A. A. et al. Phys. Rev. Lett. 99, 213002 (2007).
ERRATUM
On our bookshelf
Nature Physics 4, 827 (2008); published online: 3 November 2008; corrected after print: 14 December 2008.
In the version of this article originally published, the book Bose‑Einstein Condensation in Dilute Gases was attributed to the wrong
publisher; it should have read Cambridge University Press. This information has been corrected in the HTML and PDF versions.

news & views
erratum
On our bookshelf
Nature Physics 4, 827 (2008); published online: 3 November 2008; corrected after print: 14 December 2008.
In the version of this article originally published, the book Bose-Einstein Condensation in Dilute Gases was attributed to the wrong
publisher; it should have read Cambridge University Press. This information has been corrected in the HTML and PDF versions.

PROGRESS ARTICLE
PUBLISHED ONLINE: 2 JANUARY 2009 DOI: 10.1038/NPHYS1157
Measurement-based quantum computation

H. J. Briegel1,2 *, D. E. Browne3 , W. Dür1,2 , R. Raussendorf4 and M. Van den Nest2,5
Quantum computation offers a promising new kind of information processing, where the non-classical features of quantum
mechanics are harnessed and exploited. A number of models of quantum computation exist. These models have been shown to
be formally equivalent, but their underlying elementary concepts and the requirements for their practical realization can differ
significantly. A particularly exciting paradigm is that of measurement-based quantum computation, where the processing of
quantum information takes place by rounds of simple measurements on qubits prepared in a highly entangled state. We review
recent developments in measurement-based quantum computation with a view to both fundamental and practical issues, in
particular the power of quantum computation, the protection against noise (fault tolerance) and steps towards experimental
realization. Finally, we highlight a number of connections between this field and other branches of physics and mathematics.
Q
uantum computation is a promising and fruitful area progress article—universal quantum computation can be achieved
of research, and impressive theoretical and experimental with single-qubit measurements alone.
achievements have been reported in recent years. At More specifically, in one-way quantum computation, the system
the same time, many fundamental questions remain unanswered. is first prepared in a highly entangled quantum state, the two-
Realizing a large-scale computational device with the technology dimensional (2D)-cluster state20 (see Box 1), independently of the
available in the foreseeable future remains a challenge, and the
full range of applications for a working quantum computer is Box 1 | Graph states.
still unknown. A number of quantum algorithms are known for
particular problems, including factoring and simulation of other The cluster state20 belongs to a family of highly entangled multi-
quantum systems, but discovering new quantum algorithms that particle quantum states, which can be efficiently parameterized
outperform classical ones remains a great challenge. On a more by mathematical graphs. These are the so-called ‘graph states’21 .
fundamental level, we are still missing a good understanding of A graph G = (V ,E) is a set of N vertices (V ), together with a set of
where the border in computational power between classical and edges E ⊆ [V ]2 , which connect the vertices in an arbitrary way.
quantum resources lies. To every such graph, we associate a specific N -qubit quantum
Even the very notion of what makes a quantum computer and state |Gi. The graph state |Gi is obtained by preparing all N
what it should be capable of doing, is not entirely understood. The qubits in the +1 √ eigenstate of the Pauli operator, σx , namely
latter point is highlighted by the existence of different models for |+i = (|0i + |1i)/ 2, and by applying two-qubit phase gates
quantum computation, including the quantum circuit (or network) UPG = diag(1,1,1,−1) between all pairs of qubits connected
model1,2 , adiabatic quantum computation3 , the quantum Turing (i,j)
by an edge: |Gi = {i,j}∈E UPG |+i⊗N . Equivalently, |Gi can be
Q
machine4,5 and measurement-based models such as teleportation- defined asNa simultaneous fixed point of the correlation operators
based approaches6–9 as well as the one-way quantum computer10–12 . Kj = σx(j) {i,j}∈E σz(i) , which are entirely determined by the graph.
Indeed, there seem to be many ways to exploit nature for quantum
information processing. a b
z x z
As the features of these models differ significantly, some x
computational schemes may lend themselves more than others
providing insight into certain aspects of quantum computation
and to overcome challenges in their experimental realization. The c z
new paradigm of ‘measurement-based quantum computation6–19 ’ z
(MQC), with the ‘one-way quantum computer’ and the z x z
teleportation-based model as the most prominent examples, is
particularly promising in these respects, and provides a new z
conceptual framework in which these experimental and theoretical
challenges can be faced. Whereas in, for example, the circuit
model quantum information is processed by coherent unitary
evolutions (quantum gates), in MQC the processing of quantum Figure B1 | Three examples of graph states.
information takes place by carrying out sequences of adaptive a, Greenberger–Horne–Zeilinger state. b, 1D cluster state. c, 2D cluster
measurements. Moreover, whereas in the teleportation-based state (which is a universal resource for MQC). X = σx and Z = σx (Pauli
model joint (that is, entangling) measurements are used, in the matrices composing one of the Kj ).
one-way quantum computer—which will be the focus of this
1 Institute
for Theoretical Physics, University of Innsbruck, Technikerstraße 25, A-6020 Innsbruck, Austria, 2 Institute for Quantum Optics and Quantum
Information of the Austrian Academy of Sciences, Technikerstraße 21a, A-6020 Innsbruck, Austria, 3 Department of Physics and Astronomy, University
College London, Gower Street, London WC1E 6BT, UK, 4 University of British Columbia, Department of Physics and Astronomy, 6224 Agricultural Road,
Vancouver, British Columbia V6T 1Z1, Canada, 5 Max-Planck-Institut für Quantenoptik, Hans-Kopfermann-Str. 1, D-85748 Garching, Germany.
*e-mail: hans.briegel@uibk.ac.at.
NATURE PHYSICS | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 19

PROGRESS ARTICLE NATURE PHYSICS DOI: 10.1038/NPHYS1157
quantum algorithm that is to be implemented—one thus calls the addressing and lattice-wide entangling operations would enable
cluster state a ‘universal resource’. In a second step, the qubits in the large-scale one-way quantum computation to be realized in an
system are measured individually, in a certain order and basis, and it optical lattice system.
is this measurement pattern that specifies the entire algorithm (see Proof-of-principle experiments of a one-way quantum
Box 2). The quantum algorithm thereby corresponds, in an explicit computer with few qubits have already been carried out in
sense, to a processing of quantum correlations.
The one-way quantum computer is equipped with a remarkable Box 2 | The one-way quantum computer.
feature, namely that the entire resource for the computation is
provided by the entangled cluster state in which the system is a t0 t1 t2
initialized. This implies, in particular, that the computational
power of such a quantum computer can be traced back entirely
M M M
to the properties of its entangled resource state, thereby offering
a focused way of thinking about the nature and strength of
quantum computation. Moreover, the problem of an experimental
realization of a quantum computer is now reduced to the
preparation of a specific multi-particle state and the ability to
carry out single-qubit measurements, offering practical advantages
for certain physical set-ups. Finally, a fruitful marriage of ideas
from MQC and topological error correction was recently achieved,
paving the ground for a scalable computational device that operates b
in a noisy environment. Information flow
The computational scheme of the one-way quantum computer
was introduced in ref. 10. This work has inspired numerous studies
into MQC, both theoretical and experimental in nature. Apart
from offering an alternative approach towards realizing quantum
computation, today MQC has become an interdisciplinary field
of research, relating entanglement theory, graph theory, topology,
computational complexity, logic and statistical physics. Here, Quantum gate
we discuss a selection of recent results in MQC, to illustrate
the vigour and diversity of research in this field. We will
consider MQC in the sense of the one-way quantum computer,
but we emphasize that there are other measurement-based
Measurements: in Z direction
approaches to quantum computation—in particular teleportation- in X direction
based models—as cited above. in X–Y plane
Experimental proposals and achievements In contrast to the quantum circuit model, where quantum
Apart from its useful conceptual status as an alternative model of computations are implemented by unitary operations, in the
quantum computation, MQC can have practical advantages over one-way quantum computer, information is processed by
the standard circuit model in a variety of different physical settings, sequences of single-qubit measurements10 . These measurements
from optical lattices and single photons to spatially separated matter are carried out on a universal resource state—the 2D-cluster
qubits. In an optical lattice, cold atoms are kept in a standing- state20 —which does not depend on the algorithm to be
wave potential created by counter-propagating laser fields. The implemented. A one-way quantum computation proceeds as
potential minima create a lattice of sites in which individual atoms follows (see a and b). (1) A classical input is provided, which
can be trapped, storing quantum information in their long-lived specifies the data and the program. (2) A 2D-cluster state |C i
internal states (see Fig. 1a). Tuning the polarization of the trapping of sufficiently large size is prepared. The cluster state serves as
lasers can induce entangling interactions22,23 between neighbouring the resource for the computation. (3) A sequence of adaptive
atoms across the array and create a cluster state across the whole one-qubit measurements M (see a) is implemented on certain
lattice20,24,25 . In recent years, there has been substantial experimental qubits in the cluster. In each step of the computation, the
progress in the trapping, cooling and manipulation of ultracold measurement bases (see b) depend on the program and on
atomic gases in optical lattices in one, two and three dimensions the outcomes of previous measurements. A simple classical
(see, for example ref. 26). In particular, the creation of a Mott computer is used to compute which measurement directions
insulator state with a crystal-like arrangement of single atoms in the have to be chosen in every step. (4) After the measurements, the
lattice27,28 , and the realization of controlled entangling collisions22,25 state of the system has the form |ξ α i|ψoutα
i, where α indexes the
have been key achievements for the coherent control of matter on collection of measurement outcomes of the different branches of
the atomic level in these systems. Recent experiments use exchange the computation. The states |ψoutα
i in all branches are equal to the
interactions in double-well potentials to create arrays of robust Bell desired output state up to a local (Pauli) operation; the measured
pairs, which could be used as an alternative way to create cluster qubits are in a product state |ξ α i, which also depends on the
states in the lattice29,30,36 . measurement outcomes. The one-way quantum computer is
A remaining obstacle to the implementation of MQC in universal: even though the results of the measurements in every
these systems is that the lattice spacing is typically of the step of the computation are random, any quantum computation
order of the wavelength of the trapping light, too small for can deterministically be realized. The temporal ordering of the
individual atoms to be addressed and measured. However, recent measurements has an important role and has been formalized,
progress in the creation of lattices with wider spacing31 , sorting for example in refs 11,15,17. For different perspectives and recent
atoms in periodic potentials32 , proposed methods for single-site reviews on MQC, see refs 14,18,104–106. Part b reprinted with
addressing in tighter lattices33 , as well as new methods achieving permission from ref. 10 © 2001 APS.
subwavelength resolution34,35 , are promising. Combining single-site
20 NATURE PHYSICS | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

NATURE PHYSICS DOI: 10.1038/NPHYS1157 PROGRESS ARTICLE
a the gate is successful; for other ‘failure outcomes’, the photon states
are measured. In a standard quantum circuit, these failure outcomes
would be very damaging, because the measurement would destroy
the coherence of the state, disrupting the computation.
The one-way quantum computation model provides a way
to scalable quantum computation with such non-deterministic
gates37–41 , too. The non-deterministic gates are used to create the
cluster state, which can be done offline and stochastically (see
Fig. 1b). Once the cluster state has been created, one can then
proceed deterministically with the one-way quantum computation.
Furthermore, a polarizing beam splitter provides a simple non-
deterministic entangling gate for efficient cluster state generation42 .
For the above linear optical schemes to be truly scalable, single-
photon detectors and generators with extremely high efficiency
are required, beyond the capabilities of current experiments.
However, demonstration experiments of the key components
b
have been achieved, using post-selection43–49 . Data are kept only
when every detector fires, enabling loss or inefficiency errors
to be discounted. These achievements have demonstrated the
Non-deterministic
controlled-Z gate
basic principles of one-way quantum computation, with recent
Success
experiments including simple algorithms and active feedforward of
measurement results46,49 .
In addition to all-optical approaches, hybrid optical–matter
schemes38 seem increasingly promising. In these schemes, mat-
ter qubits—such as atoms, quantum dots or diamond nitrogen–
vacancy centres—are kept isolated from one another, for instance
Failure in separate cavities. Entangling operations can be achieved non-
deterministically, through the emission of photons, entangled with
the qubits, and linear optical entangling gates50,51 . Owing to the
non-deterministic nature of the entangling gates, the one-way
model seems again a natural approach to scalable quantum com-
Figure 1 | Practical implementations One-way quantum computation can putation either through non-deterministic strategies39,53 , repeat-
be implemented in a variety of physical systems, including systems that until-success strategies52 or so-called ‘broker–client’ approaches54 .
offer massively parallel operations, such as atoms in optical lattices, and Such approaches carry significant advantages: individual qubits
systems where two-qubit quantum gates are not deterministic, such as are isolated from one another, reducing correlated error, and the
photons with linear optics. a, Cold neutral atoms are trapped in an optical modularity of the approach facilitates scaling up to many qubits.
lattice, a standing-wave potential formed from counter-propagating laser Scaling up these approaches would seem to require optical
fields that holds atoms in a three-dimensional array (reproduced networks of increasingly complicated switching circuits—
from ref. 26). State-dependent entangling operations can be realized in a potential barrier to their implementation. However, by
parallel across the lattice, by tuning the trapping fields22 . As was exploiting a phenomenon known as percolation, scalable
experimentally demonstrated in ref. 25, this can be used to generate, in few quantum computation with non-deterministic gates is possible
steps, a cluster state over the entire lattice (see also ref. 24). If the problem even with simple non-switching optical circuits41 . Also in
of addressing single lattice sites could be solved (see discussion in the main these implementations, the simplicity and uniformity of the
text), this would open the way for a large-scale implementation of one-way entanglement structure of cluster/graph states21 (see Fig. B1) is
quantum computation. b, One-way quantum computation enables scalable essential for this approach to succeed.
quantum computation also with non-deterministic entangling logic gates. In addition to these studies, there has been a series of proposals
With probability p < 1, the gate is achieved as required, but otherwise the on ways to implement one-way quantum computation (or to
participating qubits are measured. In a circuit-based approach, this would create cluster states) in solid-state systems, using superconducting
destroy the coherence of the computation state, disrupting the circuits55 and quantum dots56,57 .
computation. Cluster states can, however, be efficiently built with such
operations, because a successful entangling operation increases the Topological protection and fault-tolerant computation
number of cluster-state qubits. A failure, while reducing the number of In real physical systems, decoherence tends to make quantum
entangled qubits, leaves the remaining qubits in an intact cluster state. systems behave more classically. One could therefore expect
Strategies can be adopted to enable the efficient generation of cluster that decoherence would threaten any computational advantage
states of any size. possessed by a quantum computer. However, the effects of
decoherence can be counteracted by quantum error correction58 .
which qubits are represented by photons. A photon can encode In fact, arbitrarily large quantum computations can be carried
a qubit, for example, in its polarization, or in spatial degree out with arbitrary accuracy, provided the error level of the
of freedom. The generation and detection of single photons is elementary components of the quantum computer is below a
making great advances and single-qubit operations can be achieved certain threshold. This important result is called the threshold
with precision through interferometers or wave plates. However, theorem of quantum computation59–62 .
the deterministic two-qubit gates, which would enable universal Recent work has been dedicated to bringing fault tolerance closer
quantum computation, cannot be achieved with interferometric to experimental reality. Proofs of fault tolerance have been derived
techniques (that is, with linear optics) alone. By adding photon for error models showing the characteristic features of realistic
counters to interferometric networks, non-deterministic entangling physical systems such as long-range correlated errors63 . Moreover,
gates can, however, be achieved7 . For some measurement outcomes, a very high threshold of 3% (that is, on average, one gate in thirty

a b
X X
X X
X X
Syndrome qubit
Code qubit
c d
Code surface
e
Control qubit Hole
Ze XC
Defect
Target qubit
Zm m
e Xe
XC
Xm
Xt
m ‘Simulated time’
Figure 2 | Topological fault tolerance from 3D cluster states. a, Elementary cell of the cluster lattice (reprinted from ref. 73 © 2006 Elsevier). b, A single
2D layer of the cluster. If the syndrome qubits are measured in the X-basis, the code qubits are projected into a surface code76 (code lattice indicated by
dashed lines). c, A surface code with electric and magnetic holes, pairwise forming encoded electric and magnetic qubits, respectively. Strings supporting
the encoded Pauli operators are also shown (parts b, c reprinted from ref. 75 © 2007 APS). d, 3D cluster for a CNOT gate between two encoded qubits. The
gate is implemented by a monodromy between world lines of holes in the code surface. The holes evolve in ‘simulated time’ (the third cluster dimension).
Also shown is the string corresponding to an encoded Pauli operator X on the control qubit and its evolution from the initial to the final code surface.
is allowed to fail) has been obtained for a method using offline state combines the universality already found in the 2D counterpart
preparation and post-selection within the circuit model64 . with the topological error-correction capabilities of the toric code74 .
What is the status of fault tolerance in MQC? Results have so far Error correction is directly built into the cluster lattice and yields a
been obtained for the one-way quantum computer. The existence threshold of 6.7 × 10−3 (ref. 75), for a model with probabilistic gate
of a non-zero threshold was first proven by reduction to the circuit errors, including imperfect preparation of the cluster state.
model65–67 . Subsequent developments evolved along two lines. First, Topologically protected quantum gates are carried out by
after it was realized37,42,68 that the one-way quantum computer may measuring some regions of qubits in the Z -basis, which effectively
be advantageously combined with the Knill–Laflamme–Milburn removes the qubits from the state. The remaining cluster, the qubits
(KLM) scheme of optical quantum computation7 , fault-tolerant of which are measured in the X - and X ± Y -basis, thereby attains
schemes using photons were developed69–71 . The dominant sources a non-trivial topology in which fault-tolerant quantum gates can
of error in this setting are photon loss and gate inaccuracies. The be encoded. A topological method of fault tolerance can then be
constraint of short-range interaction and arrangement of qubits achieved77 . We shall explain this method in some detail. First, we
in a 2D lattice—a characteristic feature of the initial one-way map the 3D cluster state to a surface code76 propagating in time.
quantum computer—is not relevant for photons. In ref. 69, both We consider a 3D cluster with an elementary cell as in Fig. 2a,
photon loss and gate inaccuracies were taken into account, yielding and single out one spatial direction on the cluster as ‘simulated
a trade-off curve between the two respective thresholds. Fault- time’. As a first step, we consider a perpendicular 2D slice of this
tolerant optical computation is possible, for example, for a gate cluster, as shown in Fig. 2b. The qubits are subdivided into code
error rate of 10−4 and photon loss rate of 3 × 10−3 (the stability and syndrome qubits. As can be easily verified, measurement of
against the main error source of photon loss is discussed in ref. 71). the syndrome qubits in the X -basis projects the code qubits into
With non-unit efficiencies ηS and ηD of photon creation and a surface code state. In a 3D cluster consisting of many linked 2D
detection as the only imperfections, the very high threshold of slices, measurement of the code qubits results in teleportation of
ηS ηD > 2/3 was established72 . the encoded state from one slice to the next (plus local Hadamard
A second line of research73 kept the geometric constraint of gates), and measurement of the syndrome qubits amounts to
nearest-neighbour interaction, which is a realistic scenario for measurement of the surface code stabilizer.
stationary qubits. To achieve fault tolerance, it is then advantageous Now note that we can modify the code surface of Fig. 2a
to increase the lattice dimension from two to three. A 3D cluster by changing measured observables from X to Z . For example, if

we measure a syndrome qubit on plaquette p in the Z N-basis, we As an important example, the 2D-cluster state is a universal
choose to not read out the surface code stabilizer Bp = f ∈∂p Zf of resource in this sense. A universal measurement-based quantum
the corresponding plaquette p. If we measure a code qubit on edge e computer is then identified as a device that enables universal
in the Z -basis, then we destroy syndrome information. N Specifically, state preparation by local operations alone79,80 . This implies that,
the eigenvalues of the surface code stabilizers As = f |s∈∂f Xf on sites whenever any given type of entanglement—as quantified by an
adjacent to e will become undefined. These elementary operations appropriate entanglement measure81 —is to be generated from the
lead to techniques for manipulating the code surface. First, we can resource state, it must already be present in the resource itself, as
punch holes into it. Holes come in two types, electric and magnetic. single-qubit operations cannot add entanglement to the system.
An electric hole is a site s where the condition hAs i = 1 is not imposed Using this intuition, it can be shown that every such universal
on the code space. Two electric holes support an electric encoded resource must exhibit maximal (scaling of) entanglement with
qubit. Similarly, a magnetic hole is a plaquette p with hBp i = 1 not respect to essentially all types of entanglement79,80 . The 2D-cluster
imposed on the code space. Two magnetic holes support a magnetic state provides a key example of such a maximally entangled resource
qubit. Now, the encoded Pauli operators Z e , X e and Z m , X m for state, but we emphasize that the entanglement criteria hold for every
these encoded qubits can be described in geometric terms. Namely, possible universal state preparator.
(E) for the electric qubit on {s,s0 }, Z e is a tensor product of Pauli This insight can be used to develop a systematic framework
operators Z along a string stretching from s to s0 . X e is a tensor to investigate which states are universal state preparator resources
product of Pauli operators X along a string looping around either s for MQC, and in particular to obtain no-go results. Following
or s0 . (M) for the magnetic qubit on {p,p0 } is just the same, with the this approach, it can be shown that, for example, n-particle
roles of Z and X interchanged; see Fig. 2c. To protect these encoded 1D-cluster states, Greenberger–Horne–Zeilinger states, W-states,
qubits against harmful errors, the holes are enlarged from one site Dicke states and certain ground states of strongly correlated 1D
or plaquette to extended connected sets of sites or plaquettes. spin systems, are not universal resources79 . Note that many of
By slightly shifting the locations of Z -measurements from the above states are considered to be highly entangled. However,
one cluster slice to the next, the holes in the code surface can in each case there is at least one type of entanglement that is
be moved and fused. This gives rise to encoded unitary gates non-maximal, implying that these states cannot be universal state
and measurements, respectively. As an example, consider the preparator resources.
topological encoded CNOT gate shown in Fig. 2d, between a States that do not violate any of the entanglement criteria include
magnetic control and electric target qubit. It is realized by ‘moving’ the graph states21 associated with various types of regular 2D lattice
one of the electric holes around one of the magnetic holes. (triangular, hexagonal, kagome), as well as lattices with a high
Again, by ‘moving’, we mean slightly shifting the locations of degree of defects (up to about 41% corresponding to the classical
Z -measurements in consecutive cluster slices. It is important to site-percolation threshold in 2D); in fact, it can be proven that such
note that when the holes are moved, the above rules (E), (M) states are universal in the same way as the 2D-cluster states79,82 .
continue to apply. Thus, the strings corresponding to the encoded The notion of ‘universality’ in MQC used in the preceding
Pauli operators are dragged along. We can now easily verify the discussion is the strongest one possible, and can be relaxed in
four conjugation relations for the CNOT gate in a topological several ways. Most importantly, one may study an altogether
manner, by sliding the operator strings forward along the hole world different concept of universality, where the goal is not to prepare
lines. One example, Xc −→ Xc ⊗ Xt (t : target, electric; c: control, arbitrary quantum states as outputs. In contrast, it is only
magnetic) is shown in Fig. 2d; the other three are similar. Note that required that a universal measurement-based quantum computer
fault-tolerant CNOT gates between two electric or two magnetic is capable of (efficiently) reproducing the classical output of any
encoded qubits can also be carried out but are more complicated; quantum computation implemented on a standard gate array
they require fusion of holes75 . quantum computer. Even though it seems difficult to formulate
Finally, by way of the described mapping between the 3D cluster entanglement-based criteria for this form of universality, one
state and a 2D surface code changing in time, there exists a circuit can study it from a constructive perspective: in refs 19,85 it is
variant of fault-tolerant cluster-state computation that requires shown how such universal resources for MQC—beyond the 2D-
only a 2D lattice of qubits. In addition, this 2D variant requires cluster states—can systematically be constructed. The underlying
only translation-invariant nearest-neighbour interaction. It yields structure of this approach can be described mathematically using
a threshold value of 7.5×10−3 (ref. 77), which is the highest known the language of matrix-product states or projected entangled
threshold for a 2D local architecture (see also ref. 78). This scenario pairs18,86,87 . The main conclusion of this investigation is that a
is suited for realization in, for example, optical lattices, but also number of extremal entanglement features (such as maximal
arrays of superconducting qubits and ion traps. localizable entanglement) exhibited by 2D-cluster states no longer
have to be present in universal resources if only classical outputs
Entanglement as a resource for computational power are considered19,85 . At present, it is not clear whether the latter
In MQC, universal quantum computation is realized by carrying form of universality is fundamentally distinct from universal state
out sequences of single-qubit measurements on a system that preparation under relaxed conditions, such as encodings36,79,83,84 .
has initially been prepared in a 2D-cluster state. As individual This issue is now under investigation.
measurements can only destroy entanglement, the entire compu- When a state is identified as not being a universal resource,
tational power of the one-way quantum computer is carried by the this does not necessarily mean that it could not be used for
entanglement structure of its resource state. Can we understand and a specific quantum computational task for which it may still
quantify what are the essential features that make the 2D-cluster outperform classical computers. This naturally leads to a study
state, and possibly other states, ‘universal resources’, which are of classical simulation of MQC, where we ask which resource
capable of enabling arbitrary measurement-based quantum com- states do not offer any computational speed-up with respect to
putations? There has been recent progress on this issue, but some classical computation. This question is closely related to major
important questions remain. investigations in condensed-matter theory, where one studies
A first feature that is to be emphasized, is that there exist several under which conditions quantum systems can efficiently be
natural notions of ‘universality’79 . In its strongest form, universality described and simulated.
is defined as the capability of generating every possible quantum Recently, several techniques have been developed to tackle
state from the resource by means of single-qubit operations. classical simulatability of MQC (refs 88–91). For example, for

most states that have been identified above as not being universal, How does the expression (1) enable us to connect this model
it has been shown that efficient classical simulation is possible. with MQC? To see this, simply consider an MQC with the state
More precisely, for such states it is possible to efficiently and |ψL i as a resource state. Then, according to (1), the (computation
exactly compute the outcome probabilities of any sequence of of the) partition function corresponds to a specific measurement
single-qubit measurements. Further examples of simulatable states pattern on the resource state |ψL i. In this way, a connection is
include the toric code states92 and the ground states of several 1D drawn between concepts from MQC and statistical physics.
spin systems93 . The techniques invoked to obtain these results are This simple connection opens the possibility to obtain a cross-
again centred around entanglement; for example, the entanglement fertilization between statistical mechanics and MQC. For example,
measure ‘entanglement width’ can be used to identify efficiently one finds a notable relation between the solvability of the Ising
simulatable states91,94 . model on a lattice L and the computational power of an MQC
Here, we have described two—in some sense complementary— operating on a resource state |ψL i. This brings us back to the
investigations regarding the origin of the quantum computational issue of the power of quantum computation, and the possibility
power. Even though these investigations were carried out within (or impossibility) of an efficient classical simulation. The central
the specific framework of the one-way computer, the insights are quantities to be considered in this context are overlaps between
of general relevance, because the different quantum computational the resource state and product states: these quantities need to be
models can simulate one another efficiently and are thus, in this computed to determine with which probabilities the outcomes of
complexity-theoretic sense, equivalent. local measurements occur, and exactly those overlaps are identified
Although significant progress has been obtained in the issues with the partition function ZL . Equation (1) now implies that any
of universality and simulatability, these matters are far from being model where the partition function can efficiently be computed,
fully understood. Most importantly, it is at present not known leads to a corresponding MQC that offers no computational
whether a universal quantum computer disallows efficient classical advantage over classical devices, and vice versa. For example, the
simulation—in other words, whether quantum computers are truly solvability of the 2D Ising model without magnetic fields implies
(exponentially) ‘more powerful’ than classical ones. Nevertheless, that MQC on the toric code state74 can be efficiently simulated92 . In
we believe that the recent studies provide the first steps to turn, the efficient classical simulation of MQC on stabilizer states
understanding this important but difficult question. with bounded tree width94 , as demonstrated in ref. 91, yields a novel
classical algorithm to efficiently calculate the partition function on
MQC and classical statistical mechanics (inhomogeneous) q-state spin models on tree-like-graphs with or
We have already seen that the study of the principles of MQC turns without magnetic fields95,96 .
out to be connected to different fields, such as entanglement theory, The relation (1) can also be exploited in a different sense. For
topology and graph theory. We discuss now how some of the central example, in ref. 95, the universality of the 2D-cluster states was used
questions raised in the study of MQC are related to notions of to prove that the 2D Ising model with magnetic fields is ‘complete’
the statistical physics of classical spin systems, such as the Ising in the sense that the partition function of any q-state spin model
model and the Potts model97 . Such spin models were introduced on an arbitrary lattice can be expressed as a special instance of the
in the context of (anti-)ferromagnetism, but they seem to have partition function of the 2D Ising model (in a complex parameter
widespread applications not only in physics, but also, for example, regime). We believe that several other interesting applications
in optimization theory and biology98 . can be found. For one, these connections enable one to phrase
Let us illustrate these connections by considering the example statistical physics problems naturally in a quantum mechanical
of the Ising model in the presence of an external field. In the Ising setting87,92,95,96,99–102 . This may open a new path towards, for
model, one considers a large lattice L (for example, a 2D square example, quantum algorithms for problems in this area.
lattice) of (classical) spins sa = ±1. The spins interact according to Finally, further connections between MQC and other fields have
the Hamiltonian function been established, for example to decidability of formal languages in
X X mathematical logic103 .
HL ({sa }) = − Jab sa sb − ha sa .
ha,bi a Outlook
The couplings Jab and ha determine the strength of the pairwise The discovery of the one-way quantum computer has opened new
interaction and the external field, respectively. The lattice L may experimental avenues towards the realization of quantum compu-
be arbitrary, in the sense that lattices of arbitrary dimension— tation in the laboratory. At the same time it has challenged the
and inPfact arbitrary graphs—are possible. The partition function traditional view of the very nature of quantum computation itself.
ZL = {sa } e−HL ({sa })/(kB T ) , where kB is the Boltzmann constant and For the future, we see several open problems and challenges.
T is the temperature, is a central quantity in this context, and On the experimental side, one of the main challenges is to realize
from it other relevant system properties such as free energy or large-scale quantum computation in the laboratory, beyond proof-
magnetization can be derived. of-principle demonstrations. For measurement-based schemes, we
The connection between MQC and the Ising model is given believe that optical lattices are still one of the most suitable
by a mapping of the Ising model on an arbitrary lattice L to candidates that enable us to create large-scale cluster states with
an MQC model where the entangled resource is determined by high efficiency. However, the problem of addressing single sites in
the geometry of L, the lattice—and thus to fully implement one-way computation—
O remains unsolved so far, even though there are new and
ZL ∼ = hψL | |αi i. (1) encouraging developments. Recent progress with photonic one-way
i quantum computation is exciting, and few-qubit applications, such
This expression states that the partition function ZL is identified as in the context of the quantum repeater, are conceivable; however,
with a quantum mechanical amplitude that is obtained as an for a scalable set-up that could go much beyond proof-of-principle
overlap between two quantum states92,95,96,99 . The multi-particle experiments, new single-photon sources with higher efficiency
entangled state |ψL i (ref. 95) encodes the interaction
N pattern and is are needed. In the meantime, a variety of new and promising
a graph state21 (see Fig. B1). The product state i |αi i contains no proposals for one-way quantum computation using hybrid systems
entanglement and specifies interaction strengths and local magnetic have been put forward, which combine the advantages of different
fields as well as the temperature of the model. physical implementations. It remains to be seen which system will

become practical in the long run. Apart from engineering issues, the 20. Briegel, H. J. & Raussendorf, R. Persistent entanglement in arrays of
capabilities of the system to naturally accommodate quantum error interacting particles. Phys. Rev. Lett. 86, 910–913 (2001).
correction (that is, on the hardware level), and therefore to facilitate 21. Hein, M. et al. in Quantum Computers, Algorithms and Chaos Vol. 162
Int. School of Physics Enrico Fermi (eds Casati, G et al.) 115–218
fault-tolerant operation, will have an important role. (IOS Press, 2006).
On the theoretical side, the study of fault-tolerant schemes and 22. Jaksch, D., Briegel, H. J., Cirac, J. I., Gardiner, C. W. & Zoller, P.
the search for new quantum algorithms will remain central issues, Entanglement of atoms via cold controlled collisions. Phys. Rev. Lett.
related to the fundamental issue of universality, classical simulation 82, 1975–1978 (1999).
and the role of entanglement. In the context of measurement-based 23. Jaksch, D. & Zoller, P. The cold atom Hubbard toolbox. Ann. Phys. 315,
52–79 (2005).
computation, a deeper understanding of universality (that is, the 24. Briegel, H. J., Raussendorf, R. & Schenzle, A. in Laser Physics at the Limits
properties that make an entangled state a universal resource) (eds Figger, H. et al.) 236–261 (Springer, 2002).
will help us to find resource states that are tailored to specific 25. Mandel, O. et al. Controlled collisions for multi-particle entanglement of
physical systems. It will also provide a basis to improve schemes for optically trapped atoms. Nature 425, 937–940 (2003).
fault-tolerant computation, for example by choosing more robust 26. Bloch, I. Ultracold quantum gases in optical lattices. Nature Phys. 1,
23–30 (2005).
states or by reducing the physical overhead. A deeper understanding 27. Greiner, M., Mandel, O., Esslinger, T., Hansch, T. W. & Bloch, I. Quantum
of efficient classical simulation, on the other hand, will narrow phase transition from a superfluid to a mott insulator in a gas of ultracold
down the set of interesting quantum algorithms. A promising atoms. Nature 415, 39–44 (2002).
strategy to find new quantum algorithms is to connect quantum 28. Jaksch, D., Bruder, C., Cirac, J. I., Gardiner, C. W. & Zoller, P. Cold bosonic
computation with other fields. An example of such a connection atoms in optical lattices. Phys. Rev. Lett. 81, 3108–3111 (1998).
29. Anderlini, M. et al. Controlled exchange interaction between pairs of neutral
(with classical statistical mechanics) has been given here, but there atoms in an optical lattice. Nature 448, 452–455 (2007).
seem to be many more. 30. Trotzky, S. et al. Time-resolved observation and control of superexchange
In conclusion, it seems that the conceptual framework of the interactions with ultracold atoms in optical lattices. Science 319,
one-way model continues to be an attractive alternative plat- 295–299 (2008).
form for experimental and theoretical investigations of quantum 31. Nelson, K. D., Li., X. & Weiss, D. S. Imaging single atoms in a
three-dimensional array. Nature Phys. 3, 556–560 (2007).
computation and its ramifications. 32. Miroshnychenko, Y. et al. An atom-sorting machine. Nature 442, 151 (2006).
33. Cho, J. Addressing individual atoms in optical lattices with standing-wave
Received 2 January 2008; accepted 18 November 2008; driving fields. Phys. Rev. Lett. 99, 020502 (2007).
published online 2 January 2009 34. Gorshkov, A. V., Jiang, L., Greiner, M., Zoller, P. & Lukin, M. D. Coherent
quantum optical control with sub-wavelength resolution. Phys. Rev. Lett. 100,
093005 (2008).
References 35. Yi, W., Daley, A. J., Pupillo, G. & Zoller, P. State-dependent addressable
1. Deutsch, D. Quantum computational networks. Proc. R. Soc. A 425, subwavelength lattices with cold atoms. New J. Phys. 10, 073015 (2008).
73–90 (1989). 36. Vaucher, B., Nunnenkamp, A. & Jaksch, D. Creation of robust entangled
2. Barenco, A. et al. Elementary gates for quantum computation. Phys. Rev. A states and new resources for measurement-based quantum computation
52, 3457–3467 (1995). using optical superlattices. New J. Phys. 10, 023005 (2008).
3. Farhi, E. et al. A quantum adiabatic evolution algorithm applied to random 37. Nielsen, M. A. Optical quantum computation using cluster states. Phys. Rev.
instances of an NP-complete problem. Science 292, 472–476 (2001). Lett. 93, 040503 (2004).
4. Deutsch, D. Quantum-theory the Church-Turing principle and the universal 38. Barrett, S. D. & Kok, P. Efficient high-fidelity quantum computation using
quantum computer. Proc. R. Soc. A 400, 97–117 (1985). matter qubits and linear optics. Phys. Rev. A 71, 060310 (2005).
5. Bernstein, E. & Vazirani, U. Proc. 25th Annual ACM Symp. on the Theory of 39. Duan, L. M. & Raussendorf, R. Efficient quantum computation with
Computing 11–20 (ACM, 1993). probabilistic quantum gates. Phys. Rev. Lett. 95, 080503 (2005).
6. Gottesman, D. & Chuang, I. L. Demonstrating the viability of universal 40. Gross, D., Kieling, K. & Eisert, J. Potential and limits to cluster state quantum
quantum computation using teleportation and single-qubit operations. computing using probabilistic gates. Phys. Rev. A 74, 042343 (2006).
Nature 402, 390–393 (1999). 41. Kieling, K., Rudolph, T. & Eisert, J. Percolation renormalization and quantum
7. Knill, E., Laflamme, R. & Milburn, G. J. A scheme for efficient quantum computing with non-deterministic gates. Phys. Rev. Lett. 99, 130501 (2007).
42. Browne, D. E. & Rudolph, T. Resource-efficient linear optical quantum
computation with linear optics. Nature 409, 46–52 (2001).
computation. Phys. Rev. Lett. 95, 010501 (2005).
8. Nielsen, M. A. Quantum computation by measurement and quantum
43. Walther, P. et al. Experimental one-way quantum computing. Nature 434,
memory. Phys. Lett. A 308, 96–100 (2003).
169–176 (2005).
9. Leung, D. W. Quantum computation by measurements. Int. J. Quantum 44. Kiesel, N. et al. Experimental analysis of a four-qubit photon cluster state.
Inform. 2, 33–43 (2004). Phys. Rev. Lett. 95, 210502 (2005).
10. Raussendorf, R. & Briegel, H. J. A one-way quantum computer. Phys. Rev. Lett. 45. Lu., C. Y. et al. Experimental entanglement of six photons in graph states.
86, 5188–5191 (2001). Nature Phys. 3, 91–95 (2007).
11. Raussendorf, R. & Briegel, H. J. Computational model underlying the one-way 46. Prevedel, R. et al. High-speed linear optics quantum computing using active
quantum computer. Quantum Inform. Comput. 2, 443–486 (2002). feed-forward. Nature 445, 65–69 (2007).
12. Raussendorf, R., Browne, D. E. & Briegel, H. J. Measurement-based quantum 47. Vallone, G., Pomarico, E., De Martini, F., Mataloni, P. & Berardi, V.
computation using cluster states. Phys. Rev. A 68, 022312 (2003). Realization and characterization of a 2-photon 4-qubit linear cluster state.
13. Aliferis, P. & Leung, D. W. Computation by measurements: A unifying Phys. Rev. Lett. 98, 180502 (2007).
picture. Phys. Rev. A 70, 062314 (2004). 48. Chen, K. et al. Experimental realization of one-way quantum computing with
14. Perdrix, S. & Jorrand, P. Measurement-based quantum turing machines and two-photon four-qubit cluster states. Phys. Rev. Lett. 99, 120503 (2007).
their universality. Preprint at <http://arxiv.org/abs/quant-ph/0404146> 49. Vallone, G., Pomarico, E., De Martini, F. & Mataloni, P. Active one-way
(2004). quantum computation with 2-photon 4-qubit cluster states. Phys. Rev. Lett.
15. Danos, V., Kashefi, E. & Panangaden, P. The measurement calculus. 100, 160502 (2008).
J. Association Comput. Machinery 54, 8 (2007). 50. Moehring, D. L. et al. Entanglement of single-atom quantum bits at a distance.
Nature 449, 68–71 (2007).
16. Childs, A. M., Leung, D. W. & Nielsen, M. A. Unified derivations of
51. Laurat, J. et al. Towards experimental entanglement connection with atomic
measurement-based schemes for quantum computation. Phys. Rev. A 71,
ensembles in the single excitation regime. New J. Phys. 9, 207 (2007).
032318 (2005). 52. Lim, Y. L., Beige, A. & Kwek, L. C. Repeat-until-success linear optics
17. Browne, D. E., Kashefi, E., Mhalla, M. & Perdrix, S. Generalized flow and distributed quantum computing. Phys. Rev. Lett. 95, 030505 (2005).
determinism in measurement-based quantum computation. New J. Phys. 9, 53. Kieling, K., Gross, D. & Eisert, J. Cluster state preparation using gates
250 (2007). operating at arbitrary success probabilities. New J. Phys. 9, 200 (2007).
18. Verstraete, F. & Cirac, J. I. Valence-bond states for quantum computation. 54. Benjamin, S. C., Browne, D. E., Fitzsimons, J. & Morton, J. J. L. Brokered
Phys. Rev. A 70, 060302 (2004). graph-state quantum computation. New J. Phys. 8, 141 (2006).
19. Gross, D. & Eisert, J. Novel schemes for measurement-based quantum 55. Tanamoto, T., Liu, Y. X., Fujita, S., Hu, X. & Nori, F. Producing cluster states
computation. Phys. Rev. Lett. 98, 220503 (2007). in charge qubits and flux qubits. Phys. Rev. Lett. 97, 230501 (2006).

56. Borhani, M. & Loss, D. Cluster states from Heisenberg interaction. Phys. Rev. 84. Tame, M. S., Paternostro, M. & Kim, M. S. One-way quantum computing in
A 71, 034308 (2005). a decoherence-free subspace. New J. Phys. 9, 201 (2007).
57. Weinstein, Y. S., Hellberg, C. S. & Levy, J. Quantum-dot cluster-state 85. Gross, D., Eisert, J., Schuch, N. & Perez-Garcia, D. Measurement-based
computing with encoded qubits. Phys. Rev. A 72, 020304(R) (2005). quantum computation beyond the one-way model. Phys. Rev. A 76,
58. Shor, P. W. Proc. 37th Annual Symp. Foundations of Computer Science 56–65 052315 (2007).
(IEEE Computer Society Press, 1996). 86. Fannes, M., Nachtergaele, B. & Werner, R. F. Finitely correlated states on
59. Knill, E., Laflamme, R. & Zurek, W. H. Resilient quantum computation: Error quantum spin chains. Commun. Math. Phys. 144, 443–490 (1992).
models and thresholds. Proc. R. Soc. A 454, 365–384 (1998). 87. Verstraete, F., Wolf, M. M., Perez-Garcia, D. & Cirac, J. I. Criticality the
60. Aharonov, D. & Ben-Or, M. Proc. Twenty-Ninth Annual ACM Symp. on area law and the computational power of projected entangled pair states.
Theory of Computing 176–188 (El Paso, 1997). Phys. Rev. Lett. 96, 220601 (2006).
61. Gottesman, D. Stabilizer Codes and Quantum Error Correction. Thesis, 88. Markov, I. & Shi, Y. Simulating quantum computation by contracting tensor
Caltech (1997). networks. SIAM J. Comput. 38, 963–981 (2008).
62. Aliferis, P., Gottesman, D. & Preskill, J. Quantum accuracy threshold 89. Jozsa, R. On the simulation of quantum circuits. Preprint at
for concatenated distance-3 codes. Quantum Information Comput. 6, <http://arxiv.org/abs/quant-ph/0603163> (2006).
97–165 (2006). 90. Yoran, N. & Short, A. J. Classical simulation of limited-width cluster-state
63. Aharonov, D., Kitaev, A. & Preskill, J. Fault-tolerant quantum computation quantum computation. Phys. Rev. Lett. 96, 170503 (2006).
with long-range correlated noise. Phys. Rev. Lett. 96, 050504 (2006). 91. Shi, Y.-Y., Duan, L.-M. & Vidal, G. Classical simulation of quantum
64. Knill, E. Quantum computing with realistically noisy devices. Nature 434, many-body systems with a tree tensor network. Phys. Rev. A 74, 022320 (2006).
39–44 (2005). 92. Bravyi, S. & Raussendorf, R. Measurement-based quantum computation with
65. Raussendorf, R. Measurement-Based Quantum Computation with Cluster the toric code states. Phys. Rev. A 76, 022304 (2007).
States. Thesis, Ludwig-Maximilians-Univ. Munich (2003). 93. Verstraete, F. & Cirac, J. I. Matrix product states represent ground states
66. Nielsen, M. A. & Dawson, C. M. Fault-tolerant quantum computation with faithfully. Phys. Rev. B 73, 094423 (2006).
cluster states. Phys. Rev. A 71, 042323 (2005). 94. Van den Nest, M., Dür, W., Vidal, G. & Briegel, H. J. Classical simulation
67. Aliferis, P. & Leung, D. W. Computation by measurements: A unifying versus universality in measurement-based quantum computation. Phys. Rev. A
picture. Phys. Rev. A 70, 062314 (2004). 75, 012337 (2007).
68. Mor, T. & Yoran, N. Methods for scalable optical quantum computation. 95. Van den Nest, M., Dür, W. & Briegel, H. J. Completeness of the classical
Phys. Rev. Lett. 97, 090501 (2006). 2D Ising model and universal quantum computation. Phys. Rev. Lett. 100,
69. Dawson, C. M., Haselgrove, H. L. & Nielsen, M. A. Noise thresholds for 110501 (2008).
optical quantum computers. Phys. Rev. Lett. 96, 020501 (2006). 96. Van den Nest, M., Dür, W. & Briegel, H. J. Classical spin models and the
70. Dawson, C. M., Haselgrove, H. L. & Nielsen, M. A. Noise thresholds for quantum-stabilizer formalism. Phys. Rev. Lett. 98, 117207 (2007).
optical cluster-state quantum computation. Phys. Rev. A 73, 052306 (2006). 97. Wu, F. Y. The Potts model. Rev. Mod. Phys. 54, 235–268 (1982).
71. Varnava, M., Browne, D. E. & Rudolph, T. Loss tolerance in one-way 98. Mezard, M., Parisi, G. & Virasoro, M. A. Spin Glass Theory and Beyond
quantum computation via counterfactual error correction. Phys. Rev. Lett. 97, (Word Scientific, 1987).
120501 (2006). 99. Bombin, H. & Martin-Delgado, M. Statistical mechanical models and
72. Varnava, M., Browne, D. E. & Rudolph, T. How good must single photon topological color codes. Phys. Rev. A 77, 042322 (2008).
sources and detectors be for efficient linear optical quantum computation? 100. Lidar, D. A. On the quantum computational complexity of the Ising spin glass
Phys. Rev. Lett. 100, 060502 (2008). partition function and of knot invariants. New J. Phys. 6, 167 (2004).
73. Raussendorf, R., Harrington, J. & Goyal, K. A fault-tolerant one-way quantum 101. Somma, R. D., Batista, C. D. & Ortiz, G. A quantum approach to classical
computer. Ann. Phys. 321, 2242–2270 (2006). statistical mechanics. Preprint at <http://arxiv.org/abs/quant-ph/0609216>
74. Kitaev, A. Y. Fault-tolerant quantum computation by anyons. Ann. Phys. 303, (2006).
2–30 (2003). 102. Aharonov, D., Arad, I., Eban, E. & Landau, Z. Polynomial quantum algorithms
75. Raussendorf, R., Harrington, J. & Goyal, K. Topological fault-tolerance in for additive approximations of the Potts model and other points of the Tutte
cluster state quantum computation. New J. Phys. 9, 199 (2007). plane. Preprint at <http://arxiv.org/abs/quant-ph/0702008> (2007).
76. Bravyi, S. & Kitaev, A. Universal quantum computation with ideal Clifford 103. Van den Nest, M. & Briegel, H. J. Measurement-based quantum computation
gates and noisy ancillas. Phys. Rev. A 71, 022316 (2005). and undecidable logic. Foundations Phys. 38, 448–457 (2008).
77. Raussendorf, R. & Harrington, J. Fault-tolerant quantum computation with 104. Nielsen, M. A. Cluster-state quantum computation. Rep. Math. Phys. 57,
high threshold in two dimensions. Phys. Rev. Lett. 98, 190504 (2007). 147–161 (2006).
78. Svore, K. M., DiVincenzo, D. P. & Terhal, B. M. Noise threshold for a 105. Jozsa, R. An introduction to measurement based quantum computation.
fault-tolerant two-dimensional lattice architecture. Quant. Inf. Comp. 7, Preprint at <http://arxiv.org/abs/quant-ph/0508124> (2005).
297–318 (2007). 106. Browne, D. E. & Briegel, H. J. in Lectures on Quantum Information (eds Bruss,
79. Van den Nest, M., Dür, W., Miyake, A. & Briegel, H. J. Fundamentals of D. & Leuchs, G.) 359–371 (Wiley–VCH, 2006).
universality in one-way quantum computation. New J. Phys. 9, 204 (2007).
80. Van den Nest, M., Miyake, A., Dür, W. & Briegel, H. J. Universal resources Acknowledgements
for measurement-based quantum computation. Phys. Rev. Lett. 97,
We thank I. Bloch for helpful comments on the manuscript. The work was supported by
150504 (2006).
the European Union (QICS,OLAQUI,SCALA), EPSRC’s QIPIRC programme, the
81. Vidal, G. Entanglement monotones. J. Mod. Opt. 47, 355–376 (2000).
NSERC and the Austrian Science Foundation.
82. Browne, D. E. et al. Phase transition of computational power in the
resource states for one-way quantum computation. New J. Phys. 10,
023010 (2008). Additional information
83. Bartlett, S. D. & Rudolph, T. Simple nearest-neighbor two-body Hamiltonian Reprints and permissions information is available online at http://npg.nature.com/
system for which the ground state is a universal resource for quantum reprintsandpermissions. Correspondence and requests for materials should be addressed
computation. Phys. Rev. A 74, 040302(R) (2006). to H.J.B.

LETTERS
PUBLISHED ONLINE: 16 NOVEMBER 2008 DOI: 10.1038/NPHYS1133
Tomography of quantum detectors

J. S. Lundeen1 *, A. Feito2,3 , H. Coldenstrodt-Ronge1 , K. L. Pregnell2,3 , Ch. Silberhorn4 , T. C. Ralph5 ,
J. Eisert2,3 , M. B. Plenio2,3 and I. A. Walmsley1 *
Measurement connects the world of quantum phenomena state ρ, the probability pn,ρ of obtaining detection outcome n is
to the world of classical events. It has both a passive
role—in observing quantum systems—and an active one, in pn,ρ = tr[ρπn ], (1)
preparing quantum states and controlling them. In view of
the central status of measurement in quantum mechanics, where {πn } is the detector POVM. In state tomography, an
it is surprising that there is no general recipe for designing unknown ρ is characterized by carrying out a set of known
a detector that measures a given observable1 . Compounding measurements, each on many identical copies of the state to
this, the characterization of existing detectors is typically estimate pn . From this estimate, one can invert equation (1) to find
based on partial calibrations or elaborate models. Thus, ρ. The interchangeability of ρ and πn in equation (1) shows that
experimental specification (that is, tomography) of a detector detector tomography has a dual role to state tomography. Now,
is of fundamental and practical importance. Here, we present measuring a set of known probe states {ρ} enables us to characterize
the realization of quantum detector tomography2–4 . We identify an unknown detector, and thus find {πn }. For these operators to
the positive-operator-valued measure describing the detector, describe a physical measurement apparatus, they must be positive
with no ancillary assumptions. This result completes the triad, semi-definite, πn ≥ 0, and n πn = I , ensuring positive probabilities
P
state5–11 , process12–17 and detector tomography, required to that add up to one. In addition, the operators {ρ} must be chosen to
fully specify an experiment. We characterize an avalanche be tomographically complete, that is, form a basis for the operator
photodiode and a photon-number-resolving detector capable of space of πn .
detecting up to eight photons18 . This creates a new set of tools In the specific case of optical detectors, lasers provide us with an
for accurately detecting and preparing non-classical light. ideal tomographic probe: the coherent state |αi. By transforming
The reduction of the quantum state of a system by measurement, the magnitude |α| through attenuation (for example, with a beam
as postulated by von Neumann, is now generally accepted splitter) and the phase arg(α) by optical delay, we can create a
to be a limiting case of a more general theory of quantum tomographically complete set of probe states {|αihα|}. Remarkably,
measurement that involves state reduction on an extended with coherent state probes, the measured statistics are themselves
Hilbert space encompassing the system and an (possibly fictional) a full representation of the detector. The measurement outcomes
ancilla. However, even within this general theory, it is not sample the function
known how to incorporate the complete chain of apparatus 1 1
components in a derivation of the actual measurement: as Qn (α) = hα|πn |αi = pn,α . (2)
Braginsky has written, ‘‘the Schrödinger equation cannot tell us π π
the connection between the design of the measuring device and This is, in fact, a definition of the well-known Q-function
the nature of the measurement’’1 . Measurement is increasingly representation of the POVM element2 . As Qn (α) contains the same
becoming a driving component in quantum technologies such information as the element πn itself, estimation of this function
as super-resolution metrology19 , Heisenberg-limited sensitivity20 constitutes detector tomography. Predictions of the detection
and quantum computing21 . Input states and dynamical processes probabilities for arbitrary input states can then be calculated directly
are accepted as resources for quantum technologies and therefore from the Q-function representation. Unfortunately, experimental
the techniques of quantum state tomography5–11,22 and quantum errors and statistical fluctuations can cause a simple fit to the
process tomography12–17,23 have been developed to measure them. Q-function to be consistent with unphysical POVM elements.
A distinct omission is that of the experimental tomography of Owing to this, we ultimately wish to directly find the POVM
detectors, which would enable more accurate classification of elements {πn } that are closest to the measured statistics, while
measurement types, objective comparison of competing devices and constraining them to be physical.
precise design of new detectors. This omission is even more striking We now turn to the description of the experimental realization,
given that the tomography of states and processes are predicated on shown in Fig. 1 (see the Methods section). The first detector was
a well-characterized detector. Here, we extend previous theoretical a commercial single-photon counting module based on a silicon
descriptions of detector tomography2–4 to include regularization avalanche photodiode (APD). It has two detection outcomes, either
and to accommodate the classical uncertainties of the experimental outputting an electronic pulse (1 click) or not (0 clicks). Past
apparatus. We apply this theory to the characterization of two evaluation of the detector has shown that the 1-click outcome
quantum detectors. is mainly associated with the arrival of one or more photons,
Characterizing a detector consists of determining its correspond- although dark counts and afterpulsing can also create this outcome.
ing positive-operator-valued measure (POVM). Given an input The 0-click event is mainly associated with vacuum at the input
1 Clarendon Laboratory, Oxford University, Parks Road, Oxford, OX1 3PU, UK, 2 Institute for Mathematical Sciences, Imperial College London, SW7 2PG, UK,
3 QOLS, The Blackett Laboratory, Imperial College London, Prince Consort Road, SW7 2BW, UK, 4 Max-Planck Research Group for Optics, Information and
Photonics, 91058 Erlangen, Germany, 5 Department of Physics, University of Queensland, Brisbane, QLD 4072, Australia.
*e-mail: j.lundeen1@physics.ox.ac.uk; walmsley@physics.ox.ac.uk.

LETTERS NATURE PHYSICS DOI: 10.1038/NPHYS1133
NDF FC 1.0
Unknown
detector 1.2
Click
λ /2 P
pn,α
0.5
1.0 No click
Laser Power meter
T~1 n=0
0.8 0
0 5 10 15 20
8
pn,α
|α |2
Figure 1 | The experimental set-up. A half-wave plate (λ/2) and 0.6
Glan–Thompson polarizer (P) are used to vary the amplitude of the probe 1 7
0.4 2 3 4 6
coherent state, which is subsequently attenuated by neutral density filters 5
(NDF) and coupled into a fibre (FC). See the Methods section for 0.2
more details. 0
0 10 20 30 40 50 60
|α |2
or photons lost owing to non-unit efficiency of the photodiode.
Having only two outcomes, this detector cannot directly measure
the incoming photon number if it is above one. The second Figure 2 | The detector tomography data. The outcome statistics (red
detector circumvents this by splitting the incoming pulse into many dots) are measured as a function of the coherent state magnitude |α|2 and
spatially or temporally separate bins, making unlikely the presence form an estimate of pn,α for each detector outcome n (number of clicks).
of more than one photon per bin. Subsequently all of the bins As they are proportional to the Q-function Qn (α) for each outcome, the
are detected with two APDs. Photon-number resolution results by statistics directly fully characterize the detector. The main plot corresponds
summing the number of 1-click outcomes from all of the bins. This to the TMD with nine outcomes and the inset corresponds to the APD. The
time-multiplexed detector (TMD) is not commercially available but vertical statistical error is too small to be seen. From the reconstructed
can be constructed with standard tools18 . Ours has eight bins in POVM elements {πn }, we generate the corresponding probabilities
total (four time bins in each of two output fibres) and thus nine pn,α = hα|πn |αi (blue curves).
outcomes—from zero to eight clicks, making it capable of detecting
up to eight photons. The added complexity and greater number in this case, there exists a dual optimization problem in which
of outcomes of this detector provide a more challenging test for the solution coincides with the original problem. Thus, the dual
detector tomography. problem provides a certificate of optimality that we use to verify
For both detectors, we first allowed the phase of α to drift. our solution. This procedure identifies the optimal POVM for this
We observed no variation in the outcome frequencies, as expected cost function.
from a detector without a phase reference. This simplifies the The measured statistics for each detector outcome (that is,
experimental procedure, requiring us to control only the magnitude the number of clicks) are shown in Fig. 2 for the TMD and
of α (as has been done for tomography of a single photon24 ). A for the APD. The distributions (equivalent to the Q-function
detector with no observed phase dependence will be described by Qn (α) of the detector) show smooth profiles and distinct photon
POVM elements diagonal in the number basis, number ranges of sensitivity for increasing number of clicks in
the detector. Figure 3 shows the diagonals (the off-diagonals are
∞
X zero for these phase-insensitive detectors) of the POVMs that
πn = θk(n) |kihk|, (3)
k=0
result from optimization of equation (4) (see the Methods section
for g (Π )). Note that πn , being the POVM element for n clicks,
simplifying henceforth the reconstruction of πn . shows nearly zero amplitude for detecting less than n photons,
For a POVM set {πn } containing only diagonal matrices as in exhibiting essentially no dark counts. Prominent in an otherwise
equation (3) that are each truncated at a number state M , we can smooth distribution, this sharp feature provides the detector with
rewrite equation (2) as a matrix equation, its discriminatory power: n clicks guarantees that there were at least
n photons in the input pulse. To assess the performance of the
P = FΠ . tomography, we find the difference (yellow bars in Fig. 3) between
the estimated POVM elements πnrec and a previously developed
For an N -outcome detector, PD×N contains all of the mea- simple theoretical model of a √ TMD, πnteo √(ref. 26; see the Methods
sured statistics, FD×M contains the D probe states α, α1, ... , αD section). The fidelity F = tr(( πnteo πnrec πnteo )1/2 )2 ≥ 98.7% for all
and ΠM ×N contains the unknown POVM set (matrix sub- n, indicating excellent agreement between the two.
scripts are the matrix dimensions). For a coherent state probe, To visualize the action of the detector, in the special case of
Fi,k = (|αi |2k exp(−|αi |2 )/k!). This can easily be reformulated for a optical detectors one can plot a Wigner function of each of the
probe in a mixed state, as was done to model the laser technical noise reconstructed POVM elements, Wn (α, α ∗ ). The response of the
(see the Methods section). The physical POVM consistent with the detector to an input state
R with Wigner function Wψ is proportional
data can be estimated through the following optimization problem: to the overlap, pn,ψ = Wn Wψ dαdα ∗ . We focus on the one-click
Wigner function W1 (α,α ∗ ) for the APD (Fig. 4a) and the TMD
min ||P − F Π ||2 + g (Π ) ,

(Fig. 4b). An APD detector is sometimes regarded as a ‘single
photon detector’, but here we can see the marked difference between
N −1
X the two Wigner functions. Instead, it is the TMD that has a fidelity
subject to πn ≥ 0, πn = I , (4) of 98% with a single photon (having experienced a 52.2% loss).
n=0
Conversely, the APD Wigner function extends to α 1, having
where the 2-norm of a matrix A is defined as ||A||2 = ( i,j |Ai,j |2 )1/2 .
P
significant overlap with photon number states >1. Therefore, to use
We regularize the inversion by means of an extra constraint: the an APD as a ‘single-photon detector’ one must make the ancillary
specification of a convex quadratic function g , independent of the assumption that the input beam has insignificant components
type of detector, that smooths fluctuations with respect to photon containing more than one photon. Despite their differences, both
number from element to element of the POVM. This is a convex Wigner functions have negative values near the origin, indicating
quadratic optimization problem, and hence also a semi-definite the absence of a classical optical analogue. Consequently, these are
problem25 which can be efficiently solved numerically. Moreover, both fundamentally quantum detectors.

NATURE PHYSICS DOI: 10.1038/NPHYS1133 LETTERS
a 1.0 a b
n=0 n=1 n=2
0.8 W1 (α , α∗) W1 (α , α∗)
2 2
0.6
θ i(n)
0.4
0.2 α∗ α∗
–5
5 5
0
0 10 20 0 10 20 0 10 20 30
α α
i i i
1.0 –2 –2
n=3 n=4 n=5
0.8
0.6 Figure 4 | The Wigner functions of the ‘one click’ detector outcomes.
θ i(n)
a,b, From the diagonal elements of π1 for the APD (a) and TMD (b) one can
0.4
generate the Wigner function representing their measurement of the
0.2 optical mode.
0
0 10 20 0 10 20 0 10 20 30
i i i
which are critical for quantum information processing using
1.0 the electromagnetic field as the information carrier27,28 . As
n=6 n=7 n=8 superconducting and semiconductor photon number counters are
0.8
developed, tomography could be used as an objective benchmark
0.6 to compare competing devices. Moreover, for one of these
θ i(n)
0.4 photon number counters, only an incomplete and empirical model

is available29 , making detector tomography the best option to
0.2
completely determine its action. We expect detector tomography
0 will become the standard for the adequate calibration of all
0 10 20 0 10 20 0 10 20 30
i i i measurement and state preparation devices.
b No click Click
1.0 Methods
Experimental set-up. Pulses of a mode-locked laser travel through a half-wave
0.8
plate (λ/2) and a Glan–Thompson polarizer with which we varied their
0.6 amplitude α. We subsequently sent the pulses through a beam splitter (T = 95%).
θ i(n)
The reflected beam travelled through three neutral-density (that is, spectrally
0.4 flat) filters before being coupled into a single-mode fibre. The attenuation from
0.2 all elements, the reflection off the beam splitter, each of the filters, and fibre
coupling, were measured individually with a calibrated power meter, resulting in
0 a total attenuation γ . This power meter was then placed in the transmission port
0 10 200 10 20
i i of the beam splitter so that the magnitude of α for the probe state in the fibre
was found from P, the measured time-averaged power and the pulse rate R via
|α|2 = γ Pλ/(2π R~c). For each value of α, we recorded the number of times each
Figure 3 | Optimal physical POVMs. a,b, The diagonals of the detection outcome occurred in J trials (that is, laser pulses), which provides an
reconstructed POVMs represented in the photon-number basis for the estimate of pn,α .
photon-number-resolving TMD (a) and the binary APD detector (b). The
Source of light and technical noise. The input states were generated by a
TMD POVM elements were obtained up to basis state |60ih60| (therefore
mode-locked Ti:sapphire laser with centre wavelength λ and a full-width at
M = 60), but are shown only up to |30ih30| for display purposes. The APD half-maximum bandwidth of 1λ specifically chosen for each detector. It was
POVM elements are shown in full. Stacked on top of each θi(n) , cavity dumped to reduce its repetition rate R to be compatible with tested
|θi(n)(rec) − θi(n)(teo) | is shown in yellow, where n is the number of clicks, and detectors. Long-term drift of the intensity over 1 million pulses was <0.5%. To
rec and teo are the reconstructed and theoretical diagonals of POVM characterize it, a NIST-calibrated Coherent FieldMaxII-TO power meter was
used (systematic error of 5%). In the case of the APD detector (a Perkin Elmer
element, πn . The theoretical TMD and APD models are described in the SPCM-AQR-13-FC), we set λ = 780±1 nm, 1λ = 20 nm, and chose the appropriate
Methods section. rate R = 1.4975 ± 0.0005 kHz, J = 1472967 and γ = (5.66 ± 0.08) × 10−9 . For
the TMD detector, we set λ = 789 ± 1 nm, 1λ = 26 nm, R = 76.169 ± 0.001 kHz,
As quantum technologies advance, detectors are becoming more J = 38,084 and γ = (8.51 ± 0.11) × 10−9 . We now evaluate the importance to
our tomography of the technical noise found at some level in all lasers. Our
complex, making a black-box approach to their characterization laser randomly varies in energy between subsequent pulses with a standard
an important tool. Identifying the exact operations of detectors deviation of 1.88% ± 0.02% of |α|2 . Attenuated to the single-photon level, as in
will benefit precision tasks, such as state tomography or metrology. this experiment, one might expect the inherent large fractional uncertainty in the
By eliminating assumptions, full characterization enables more coherent state to render this technical noise insignificant. We test this expectation
√
by modelling the pulse distribution as a Gaussian fα (β) = e−(β−α) /(2σ ) /(σ 2π )
2 2
flexible design and use of detectors, be they noisy, nonlinear,

centred around α in phase space, with a variance approximately equal to that
inefficient or operating outside their normal range. With precise measured, σ 2 = 0.0004|α|4 . Each probe state is then best described by a mixture of
characterization, we can ask precise quantitative questions about coherent states,
our power to prepare non-classical states or herald quantum Z
operations21 . This opens a path for the experimental study ρhαi = d2 β|βihβ|fα (β)
of yet unexplored concepts such as the non-classicality of ∞
detectors. For optical detector tomography, a promising avenue
X
= El,m,α |lihm|,
for research will be to transfer well-established techniques from l,m=0
homodyne tomography (for example, balanced noise-reduction, where

direct measurement of the Wigner function or pattern functions). Z
Now that it is well characterized, the photon counter also 1 −(β−α)2 /(2σ 2 )
β l+m e−β
2
Ej,m,α = √ √ dβ.
provides a unique tool for carrying out non-Gaussian operations, σ 2π l!m!

The detection probability for outcome n is then 8. Banaszek, K., D’Ariano, G. M., Paris, M. G. A. & Sacchi, M. F.
∞
X Maximum-likelihood estimation of the density matrix. Phys. Rev. A 61,
phαi,n = Ek,k,α θk(n) . 010304(R) (2000).
k=0 9. White, A. G., James, D. F. V., Munro, W. J. & Kwiat, P. G. Exploring Hilbert
Comparing our analysis done with pure input states |αihα| to that done with mixed space: Accurate characterization of quantum information. Phys. Rev. A 65,
states ρhαi , we find the difference between the POVMs obtained was negligible. 012301 (2002).
For example 10. Ourjoumtsev, A., Jeong, H., Tualle-Brouri, R. & Grangier, P. Generation of
optical ‘Schrödinger cats’ from photon number states. Nature 448,
784–786 (2007).
||Πpure − Πmixed ||2
≤ 0.7% 11. Neergaard-Nielsen, J. S., Melholt Nielsen, B., Hettich, C., Mølmer, K. &
||Πmixed ||2 Polzik, E. S. Generation of a superposition of odd photon number states for
quantum information networks. Phys. Rev. Lett. 97, 083604 (2006).
and the largest relative difference between any two θk(n) coming from a mixed-state 12. Chuang, I. L. & Nielsen, M. A. Prescription for experimental determination of
or a pure-state derivation was 1.3%. Furthermore, the reconstructed probability the dynamics of a quantum black box. J. Mod. Opt. 44, 2455–2467 (1997).
distributions are so close that they are indistinguishable on the scale of Fig. 1. This 13. Poyatos, J. F., Cirac, J. I. & Zoller, P. Complete characterization of a
reinforces our earlier expectation that technical noise in the laser will be negligible quantum process: The two-bit quantum gate. Phys. Rev. Lett. 78,
when using single-photon-level coherent states. This differs from homodyne 390–393 (1997).
tomography where technical noise can shift a strong local oscillator to a nearly 14. Altepeter, J. B. et al. Ancilla-assisted quantum process tomography. Phys. Rev.
orthogonal state. Lett. 90, 193601 (2003).
15. D’Ariano, G. M. & Maccone, L. Measuring quantum optical Hamiltonians.
Discussion of regularization. Care has to be taken that the optimization problem Phys. Rev. Lett. 80, 5465–5468 (1998).
is well conditioned to find the true POVM of the detector. In finding the 16. Nielsen, M. A., Knill, E. & Laflamme, R. Complete quantum teleportation
number basis representation, we are deconvolving a coherent state from our using nuclear magnetic resonance. Nature 396, 52–55 (1998).
statistics, which is intrinsically an ill-conditioned problem. Similar issues of 17. Mitchell, M. W., Ellenor, C. W., Schneider, S. & Steinberg, A. M. Diagnosis,
conditioning have been discussed in the context of state and process tomography, prescription and prognosis of a Bell-state filter by quantum process
see, for example, refs 30,31. Owing to a large ratio between the largest and tomography. Phys. Rev. Lett. 91, 120402 (2003).
smallest singular values of the matrices defining the quadratic problem, small 18. Achilles, D., Silberhorn, Ch., Sliwa, C., Banaszek, K. & Walmsley, I. A.
fluctuations in the probability distribution can result in large variations for the Fiber-assisted detection with photon number resolution. Opt. Lett. 28,
reconstructed POVM. This can result in operators that closely approximate 2387–2389 (2003).
the outcome statistics and yet contain errant spikes in their distribution in 19. Resch, K. J. et al. Time-reversal and super-resolving phase measurements.
photon number. To suppress this effect, we penalize the difference θk(n) − θk+1 (n)
Phys. Rev. Lett. 98, 223601 (2007).
(independent of the shape of the POVM) by using the regularization g = y S with 20. Higgins, B. L., Berry, D. W., Bartlett, S. D., Wiseman, H. M. & Pryde, G. J.
S = k,n [θk(n) − θk+1(n) 2
P
] . This is motivated by the fact that any realistic detector Entanglement-free Heisenberg-limited phase estimation. Nature 450,
will have a finite efficiency η, which necessitates a smooth θk distribution:
(n)
393–396 (2007).
if G(r) is the probability of registering
P r photons and H (q) is the probability 21. Knill, E., Laflamme, R. & Milburn, G. J. A scheme for efficient quantum
that q were present then, G(r) = q ( qr )ηr (1 − η)q−r H (q). Consequently, if computation with linear optics. Nature 409, 46–52 (2001).
θk 6= 0 then θk+1 , θk+2 and so on cannot be zero, but will follow some smooth 22. Dunn, T., Walmsley, I. A. & Mukamel, S. Experimental determination of the
distribution. As we do not assume any knowledge about the precise loss of quantum-mechanical state of a molecular vibrational mode using fluorescence
our detector, we simply choose an arbitrary value for y. Varying y by three tomography. Phys. Rev. Lett. 74, 884–887 (1995).
orders of magnitude hardly affects the exact value of the estimated POVM, 23. O’Brien, J. L. et al. Quantum process tomography of a controlled-NOT gate.
changing it by only 10%. Furthermore, the regularization g = y S also proves Phys. Rev. Lett. 90, 193601 (2004).
to be robust to noise up to δ = 0.2 (varying α → α(1 + δ) across {|αihα|} 24. Lvovsky, A. I. et al. Quantum state reconstruction of the single-photon Fock
with a Gaussian distribution for δ). This shows that the regularization’s main state. Phys. Rev. Lett. 87, 050402 (2001).
effect is to suppress the ill-conditioning and noise while leaving the POVM 25. Boyd, S. & Vandenberghe, L. Convex Optimization (Cambridge Univ.
fitting unaffected. Press, 2004).
26. Achilles, D. et al. Photon-number-resolving detection using
The theoretical model of the detector. Detector tomography does not make use time-multiplexing. J. Mod. Opt. 51, 1499–1515 (2004).
of any physical model on the functioning of the detector. To verify the success of 27. Eisert, J., Scheel, S. & Plenio, M. B. Distilling Gaussian states with Gaussian
this approach, we have compared the outcome of the estimation to those POVM operations is impossible. Phys. Rev. Lett. 89, 137903 (2002).
elements obtained from a theoretical model of the APD and TMD (ref. 26). The 28. Browne, D. E., Eisert, J., Scheel, S. & Plenio, M. B. Driving non-Gaussian to
APD is treated as a binary detector with a loss of 43.2%. The theoretical TMD Gaussian states with linear optics. Phys. Rev. A 67, 062320 (2003).
assumes no dark counts, three sequential beam splitters with experimentally 29. Kardynal, B. E., Yuan, Z. L. & Shields, A. J. An avalanche-photodiode-based
inferred reflectivities, 50.18%, 50.60% and 41.92%, and an overall loss of 52.2% photon-number-resolving detector. Nature Photon. 2, 425–428 (2008).
(that best fits the data), followed by two perfect APDs. 30. Boulant, N., Havel, T. F., Pravia, M. A. & Cory, D. G. Robust method for
estimating the Lindblad operators of a dissipative quantum process from
Received 16 June 2008; accepted 8 October 2008; measurements of the density operator at multiple time points. Phys. Rev. A 67,
published online 16 November 2008 042322 (2003).
31. Jezek, M., Fiurasek, J. & Hradil, Z. Quantum inference of states and processes.
Phys. Rev. A 68, 012305 (2003).
References
1. Braginsky, V. R. & Khalili, F. Ya. Quantum Measurement 38 (Cambridge Univ.
Press, 1992).
Acknowledgements
This work has been supported by the EU integrated project QAP and STREP COMPAS,
2. Luis, A. & Sanchez-Soto, L. L. Complete characterization of arbitrary quantum
EPSRC grants EP/C546237/1 and QIP-IRC, the Royal Society, Microsoft Research and the
measurement processes. Phys. Rev. Lett. 83, 3573–3576 (1999).
EURYI Award Scheme. H.C.-R. has been supported by the European Commission under
3. Fiurasek, J. Maximum-likelihood estimation of quantum measurement. Phys.
the Marie Curie Programme and by the Heinz-Durr Programme of the Studienstiftung
Rev. A 64, 024102 (2001).
des dt. Volkes.
4. D’Ariano, G. M., Maccone, L. & Lo Presti, P. Quantum calibration of
measurement instrumentation. Phys. Rev. Lett. 93, 250407 (2004).
5. Vogel, K. & Risken, H. Determination of quasiprobability distributions in Author contributions
terms of probability distributions for the rotated quadrature phase. Phys. Rev. J.S.L., H.C-R. and I.A.W. contributed to the concept of the experiment and its design,
A 40, 2847–2849 (1989). as well as to laboratory measurements and data analysis. A.F., K.L.P., M.B.P. and J.E.
6. Smithey, D. T., Beck, M., Raymer, M. G. & Faridani, A. Measurement of the contributed modelling and data analysis. Ch. S. and T.C.R. contributed to the conception
Wigner distribution and the density matrix of a light mode using optical of the project and to its planning.
homodyne tomography: Application to squeezed states and the vacuum. Phys.
Rev. Lett. 70, 1244–1247 (1993). Additional information
7. Banaszek, K., Radzewicz, C., Wodkiewicz, K. & Krasinski, J. S. Direct Reprints and permissions information is available online at http://npg.nature.com/
measurement of the Wigner function by photon counting. Phys. Rev. A 60, reprintsandpermissions. Correspondence and requests for materials should be addressed
674–677 (1999). to J.S.L. or I.A.W.

LETTERS
Linear temperature dependence of resistivity and

change in the Fermi surface at the pseudogap
critical point of a high-Tc superconductor
R. Daou1 *, Nicolas Doiron-Leyraud1 *, David LeBoeuf1 , S. Y. Li1 , Francis Laliberté1 ,
Olivier Cyr-Choinière1 , Y. J. Jo2 , L. Balicas2 , J.-Q. Yan3 , J.-S. Zhou3 , J. B. Goodenough3 and
Louis Taillefer1,4†
A fundamental question for high-temperature superconductors Nd-LSCO

is the nature of the pseudogap phase, which lies between the
Mott insulator at zero doping and the Fermi liquid at high
doping p (refs 1,2). Here we report on the behaviour of charge 200
carriers near the zero-temperature onset of this phase, namely p = 0.20
at the critical doping p∗ , where the pseudogap temperature T*
ρ (μΩ cm)
T ∗ goes to zero, accessed by investigating a material in
which superconductivity can be fully suppressed by a steady Tmin
magnetic field. Just below p∗ , the normal-state resistivity 100
and Hall coefficient of La1.6−x Nd0.4 Srx CuO4 are found to rise p = 0.24
simultaneously as the temperature drops below T ∗ , suggesting
a change in the Fermi surface with a large associated drop in
conductivity. At p∗ , the resistivity shows a linear temperature
dependence as the temperature approaches zero, a typical 0
signature of a quantum critical point3 . These findings impose 0 50 100 150
new constraints on the mechanisms responsible for inelastic T (K)
scattering and Fermi-surface transformation in theories of the
pseudogap phase1,4–8 . Figure 1 | Normal-state resistivity. In-plane electrical resistivity ρ(T) of
At low hole doping p, high-transition-temperature (high-Tc ) Nd-LSCO as a function of temperature, at p = 0.20 and 0.24, measured in a
superconductors are doped Mott insulators, strongly correlated magnetic field strong enough to fully suppress superconductivity (see
metals characterized by a low carrier density n equal to the Supplementary Information). The black line is a linear fit to the p = 0.20
concentration of doped holes. Indeed, Hall-effect measurements data between 80 and 300 K. Below a temperature T ∗ = 80 K, ρ(T) deviates
on La2−x Srx CuO4 (LSCO) at x = p < 0.05 yield a Hall number from its linear-T behaviour at high temperature and develops a pronounced
nH ≡ V /eRH equal to p at low temperature9 , where RH is the Hall upturn at low temperature, with a minimum at Tmin = 37 K. By contrast,
coefficient, e is the electron charge and V is the volume per Cu ρ(T) at p = 0.24 shows no upturn down to the lowest temperature.
atom. At high doping, however, these materials are Fermi liquids,
metals characterized by a well-defined coherent three-dimensional
Fermi surface10 and a resistivity ρ that grows quadratically with
temperature11 : ρ ∼ T 2 . In this regime, the Fermi surface is a large maximal Tc makes it possible to suppress superconductivity entirely
cylinder containing 1 + p holes10 , so the carrier density is high, with a steady magnetic field.
given by n = 1 + p. At p ≈ 0.25, low-temperature measurements on In Fig. 1, we show the normal-state resistivity ρ(T ) of Nd-LSCO
Tl2 Ba2 CuO6+y yield nH = 1 + p (ref. 12). These findings naturally at a doping p = 0.20. Above a temperature T ∗ = 80 K, ρ(T ) shows
beg the following question: How do the electrons in copper oxide the linear temperature dependence characteristic of all hole-doped
superconductors go from one state to the other? copper oxides. Below this temperature, it deviates upwards and
This is intimately tied to the question of the nature of the develops an upturn visible even in zero field (see Supplementary
‘pseudogap phase’, this enigmatic region of the doping phase Information, Fig. S1), with a minimum at Tmin = 37 K > Tc = 20 K,
diagram present in all high-Tc superconductors below a crossover in excellent agreement with early data in zero field13 . By applying
temperature T ∗ (ref. 2). Here we investigate the T = 0 onset of a magnetic field of 35 T, we were able to track the upturn in
this pseudogap phase by measuring the transport properties of ρ(T ) down to 1 K, thus revealing a pronounced rise at low
La1.6−x Nd0.4 Srx CuO4 (Nd-LSCO), a material whose relatively low temperature (Fig. 1).
1 Département de physique and RQMP, Université de Sherbrooke, Sherbrooke, Québec J1K 2R1, Canada, 2 National High Magnetic Field Laboratory,
Florida State University, Tallahassee, Florida 32310-3706, USA, 3 Texas Materials Institute, University of Texas at Austin, Austin, Texas 78712, USA,
4 Canadian Institute for Advanced Research, Toronto, Ontario M5G 1Z8, Canada. *These authors contributed equally to this work.
† e-mail: louis.taillefer@physique.usherbrooke.ca.

a 400 b 50 160
Nd-LSCO Nd-LSCO
p = 0.20 p = 0.24 7 Nd-LSCO
300 T*
40
ρab (μΩ cm)
ρab (μΩ cm)
ρc (mΩ cm)
200 6
Temperature (K)
30 80
100
~T Tch
5
0 20
1 10 100 0 25 Tm
T (K) T (K)
Tc
Figure 2 | In-plane and out-of-plane resistivities at low temperature. 0
a, Semilog plot of the in-plane resistivity ρab (T) of Nd-LSCO at p = 0.20. 0.1 0.2 p* pc 0.3
The black line is a linear fit above 80 K. This shows that, after a rapid rise, Hole doping
the in-plane resistivity saturates at low temperature, in contrast to the
weak logarithmic divergence observed in LSCO at p < 0.16 (ref. 29). Figure 3 | Phase diagram. Temperature–doping phase diagram of Nd-LSCO
b, Temperature dependence of the normal-state electrical resistivity of showing the superconducting phase below Tc (open black circles) and the
Nd-LSCO at p = 0.24 ≈ p∗ , in the low-temperature regime. Both ρab and the pseudogap region delineated by the crossover temperature T ∗ (blue
out-of-plane resistivity ρc show a linear temperature dependence down to squares). Also shown is the region where static magnetism is observed
the lowest measured temperature. below Tm (red circles) and charge order is detected below Tch (black
diamonds and green circles). These onset temperatures are respectively
defined as the temperature below which (1) the resistance is zero, (2) the
The absence of magneto-resistance (see Supplementary Infor- in-plane resistivity ρab (T) deviates from its linear dependence at high
mation, Fig. S1) implies that the magnetic field simply serves to temperature, (3) an internal magnetic field is detected by zero-field muon
remove superconductivity and reveal the unaltered behaviour of the spin relaxation and (4) charge order is detected by either X-ray diffraction
underlying normal state down to T ≈ 0. The evolution with tem- or nuclear quadrupole resonance (NQR). All lines are guides to the eye.
perature is perfectly smooth, indicating a crossover as opposed to Values of Tc and error bars are given in Supplementary Information. T ∗ is
a transition. Most significantly, ρ(T ) saturates at low temperature obtained from a fit to the ρab (T) data of ref. 20 for p = 0.12 and 0.15, and
(see Fig. 2a). This shows that the ground state is a metal and not those reported here for p = 0.20 and 0.24 (see Fig. 1 and Supplementary
an insulator, and that T ∗ therefore marks the onset of a crossover Information, Fig. S2, and Supplementary Information for fits and error bars).
from one metallic state to another. Note that the loss of conductivity The blue line p = 0.20 is made to end at p = 0.24, thereby defining the
is substantial, by a factor of approximately ρ0 /ρ(T → 0) = 5.8, critical doping where T ∗ goes to zero as p∗ = 0.24. Experimentally, this
where ρ0 = 245 µ cm is the resistivity measured at T = 1 K and point must lie in the range 0.20 < p∗ 6 0.24, because ρ(T) remains linear
ρ(T → 0) = 42 µ cm is the value extrapolated linearly to T = 0 down to the lowest temperature at p = 0.24 (see Fig. 2b). Tm is obtained
from above T ∗ . from the muon spin relaxation measurements of ref. 31. The red line is
We identify T ∗ as the onset of the pseudogap phase, following made to end below p = 0.20, as no static magnetism was detected at
the standard definition: the temperature below which the in-plane p = 0.20 down to T = 2 K. Tch is obtained from hard-X-ray diffraction on
resistivity ρab (T ) starts to deviate from its linear-T behaviour Nd-LSCO (filled black diamonds and error bars; ref. 22) and from resonant
at high temperature2,14 . (Note that the deviation can be either soft-X-ray diffraction on Eu-LSCO (open black diamonds and error bars;
upwards, as in LSCO, or downwards, as in YBa2 Cu3 Oy (ref. 14), ref. 23). The onset of charge order has been found to coincide with the
depending on the relative magnitude of inelastic and elastic wipe-out anomaly in NQR at TNQR , reproduced here from ref. 21 (error bars
(disorder) scattering at T ∗ ; in YBa2 Cu3 Oy , the copper oxide quoted therein) for Nd-LSCO (filled green circles) and Eu-LSCO (open
material with the lowest disorder scattering, the loss of inelastic green circles).
scattering below T ∗ is a much larger relative effect than in LSCO,
hence the drop in ρab (T ).) In Fig. 3, we plot T ∗ as a function
of doping in a p–T phase diagram. Note that the magnitude of By comparison, at p = 0.20, the magnitude of RH at T → 0 yields
T ∗ in Nd-LSCO is comparable to that found in other hole-doped nH = 0.3 ±0.05. The change in the Hall number at T → 0 between
copper oxides, pointing to a common origin (see Supplementary p = 0.24 and p = 0.20 is therefore 1nH = 1.0±0.2 hole per Cu atom.
Information, Fig. S2 for a comparison with LSCO). If the Hall number is interpreted as a carrier density, these values are
In Fig. 4, we present the Hall coefficient RH (T ) measured on the consistent with a crossover from a metal with a large hole-like Fermi
same crystal (with p = 0.20), and compare it directly with ρ(T ). surface at p∗ (where n = 1+p) to a metal with a low density of holes
Both coefficients are seen to rise simultaneously, with Tmin the below p∗ (where n ≈ p).
coincident onset of their respective upturns. This is strong evidence In contrast to p = 0.20, the electrical resistivity at p = 0.24 shown
that the cause of both upturns is a modification of the Fermi surface. in Fig. 2b shows a monotonic temperature dependence down to 1 K,
Let us now look at a slightly higher doping. Figures 1 and 4 linear as T → 0. The absence of any anomaly demonstrates that
respectively show ρ(T ) and RH (T ) measured on a second crystal, T ∗ = 0 at that doping. Therefore, the critical doping p∗ where the
with p = 0.24. The low-temperature behaviour has changed: ρ(T ) pseudogap line ends is located between p = 0.20 and 0.24, inside the
shows no sign of an upturn and RH (T ) is now constant below region where superconductivity exists in zero field. For definiteness,
25 K, extrapolating to RH = +0.45 ± 0.05 mm3 C−1 as T → 0. The in Fig. 3 we set it at p∗ = 0.24, although it could be slightly lower.
corresponding Hall number is nH = 1.3 ± 0.15, in good agreement As shown in Fig. 2b, not only is the in-plane resistivity
with the carrier density n = 1 + p = 1.24 expected for a large ρab (T ) linear as T → 0 at p = 0.24, but so is the out-of-
Fermi cylinder, and quantitatively consistent with measurements plane resistivity ρc (T ). Moreover, the fact that RH (T ) is flat at
on Tl2 Ba2 CuO6+y at p = 0.26 and T → 0, where nH = 1.3 (ref. 12). low temperature implies that the cotangent of the Hall angle,

200 TNQR = 60 ± 6 K (ref. 21) was shown to coincide with the onset of
Nd-LSCO charge order measured via hard-X-ray diffraction, at Tch = 62±5 K
2
(ref. 22) (see Fig. 3). Direct evidence of a charge modulation via
resonant soft-X-ray diffraction was reported recently for the closely
p = 0.20 related material Eu-LSCO, with Tch = 70±10 K at p = 0.15 (ref. 23),
R H (mm3 C ¬1)
TNQR whereas TNQR = 60 ± 6 K (ref. 21) in Eu-LSCO at p = 0.16 (see
ρ (μΩ cm)
100 Fig. 3). Clearly, the upturn in ρ(T ) is correlated with the onset of
1 charge order in these two materials. While the correlation between
p = 0.20
TNQR and Tmin has been noted previously20 , the mechanism causing
the upturn in ρ(T ) remained unclear. Our data show that the
mechanism is a change in Fermi surface, and the positive rise in
p = 0.24
RH (T ) imposes a strong constraint on the topology of the resulting
Fermi surface. An additional constraint comes from the fact that
0 0
0 25 50 75 100 RH (T ) drops to negative values near p = 1/8, not only in Nd-LSCO
T (K) (ref. 13) and other materials with ‘stripe’ order24,25 , but also in
YBa2 Cu3 Oy (ref. 26).
Figure 4 | Normal-state Hall coefficient. Hall coefficient RH (T) of Recent calculations of the Fermi-surface reconstruction caused
Nd-LSCO as a function of temperature for p = 0.20 and 0.24, measured in by stripe order are consistent with a negative RH near p = 1/8 in
a magnetic field of 15 T. Below 12 K, the 0.20 data are in 33 T, a magnetic that spin stripes tend to generate an electron pocket in the Fermi
field strong enough to fully suppress superconductivity (see Supplementary surface27 . Interestingly, charge stripes do not27 , and this might
Information). The dashed blue horizontal line is the value of RH calculated explain the positive rise in RH seen at higher doping, provided that
for a large cylindrical Fermi surface enclosing 1 + p holes, namely RH = V/e stripe order involves predominantly charge order at high doping (in
(1 + p), at p = 0.24. At p = 0.20, the rise in RH (T) at low temperature line with the fact that charge order sets in at a higher temperature
signals a modification of this large Fermi surface. The upturn is seen to than spin order20,21 ).
coincide with a simultaneous upturn in ρ(T) (reproduced in black from In the other kind of theory of the pseudogap phase, the
Fig. 1) and with the onset of charge order at TNQR as detected by NQR critical point reflects a T = 0 transition from small hole pockets,
(see the text and ref. 21). characteristic of a doped Mott insulator, to a large hole pocket,
without symmetry breaking4,5 . Recent work suggests that the
quasiparticle scattering rate above such a critical point may indeed
cotθH (T ) ∼ ρab (T )/RH (T ), is also linear at low temperature. We grow linearly with temperature28 . Although calculations are needed
infer that a single anomalous scattering process dominates the to confirm this, a change in carrier density from n ≈ p to n = 1 + p
electron–electron correlations at low temperature at p∗ (or just would seem natural in this kind of scenario. However, it is more
above). This shows that the Fermi-liquid behaviour observed at difficult to see what could cause the negative values of RH (T → 0)
p = 0.3 (in LSCO), where ρab (T ) ∼ T 2 below T ≈ 50 K (ref. 11), near p = 1/8. It seems that stripe order or fluctuations would have
breaks down just before the onset of the pseudogap phase at p∗ . to be invoked as a secondary instability inside the pseudogap phase,
This kind of ‘non-Fermi-liquid’ behaviour, whereby ρ(T ) ∼ T with an onset in doping that would be essentially simultaneous with
as T → 0, has typically been observed in heavy-fermion metals p∗ in the case of Nd-LSCO.
at the quantum critical point where the onset temperature for We end by comparing our results qualitatively with those of
antiferromagnetic order goes to zero3 . It is also consistent with the previous high-field studies on LSCO. The resistivity shows very
marginal-Fermi-liquid description of cuprates15 . similar features at high temperature: linear T above T ∗ (ref. 14) and
In summary, our experimental findings offer compelling an upturn below T ∗ (ref. 29). The Hall coefficient of LSCO (ref. 30),
evidence that the pseudogap phase ends at a T = 0 critical point on the other hand, has a more subtle and complex evolution with
p∗ located below the onset of superconductivity (at pc ≈ 0.27), in doping than that presented here for Nd-LSCO, which makes it
agreement with previous but more indirect evidence from other harder to pinpoint p∗ using the same criteria as we have used above.
hole-doped copper oxides16 . Moreover, they impose two strong Nonetheless, it seems likely that the same fundamental mechanisms
new constraints on theories of the pseudogap phase: (1) its onset are responsible for both the linear-T resistivity and the resistivity
below p∗ modifies the large Fermi surface characteristic of the upturns, and for the onset of the pseudogap at T ∗ , in both LSCO
overdoped metallic state; (2) quasiparticle scattering at p∗ is linear and Nd-LSCO.
in temperature as T → 0.
The existence of a quantum critical point is consistent with two Methods
kinds of theory of the pseudogap phase. The first kind invokes Single crystals of La2−y−x Ndy Srx CuO4 (Nd-LSCO) were grown with a Nd content
the onset of an order, with some associated broken symmetry6–8 . y = 0.4 using a travelling-float-zone technique and cut from boules with nominal
Sr concentrations x = 0.20 and 0.25. The actual doping p of each crystal was
Because T ∗ marks a crossover and not a sharp transition, this estimated from its Tc and ρ(250 K) values compared with published data, giving
order is presumably short range or fluctuating. In the electron- p = 0.20±0.005 and 0.24±0.005, respectively. The resistivity ρ and Hall coefficient
doped copper oxides, for example, the pseudogap phase has RH were measured at the National High Magnetic Field Laboratory (NHMFL)
been interpreted as a fluctuating precursor of the long-range in Tallahassee in steady magnetic fields up to 35 T and in Sherbrooke in steady
fields up to 15 T. The field was always applied along the c axis. Neither ρ nor RH
antiferromagnetic order that sets in at lower temperature17 , and
showed any field dependence up to the highest fields. More details are available in
the signatures of the pseudogap critical point in transport are Supplementary Information.
similar to those found here: a linear-T resistivity as T → 0 (ref. 18)
and a sharp change in RH (T = 0) (ref. 19). For Nd-LSCO and Received 5 June 2008; accepted 24 September 2008;
LSCO, an analogous scenario would be ‘stripe’ fluctuations, as a published online 2 November 2008
precursor to the static spin and charge modulations observed at
lower temperature20 . Note that in Nd-LSCO at p = 0.20 the onset References
1. Lee, P. A., Nagaosa, N. & Wen, X.-G. Doping a Mott insulator: Physics of
of the upturn in ρ(T ) and RH (T ) at Tmin = 37 K coincides with the high-temperature superconductivity. Rev. Mod. Phys. 78, 17–85 (2006).
loss of NQR intensity at TNQR = 40 ± 6 K (ref. 21) (see Fig. 4). In 2. Timusk, T. & Statt, B. The pseudogap in high-temperature superconductors:
Nd-LSCO at p = 0.15, this so-called ‘wipe-out’ anomaly in NQR at An experimental survey. Rep. Prog. Phys. 62, 61–122 (1999).

3. v Löhneysen, H., Rosch, A., Vojta, M. & Wölfle, P. Fermi-liquid instabilities at 22. Niemöller, T. et al. Charge stripes seen with X-rays in La1.45 Nd0.4 Sr0.15 CuO4 .
magnetic quantum phase transitions. Rev. Mod. Phys. 79, 1015–1075 (2007). Eur. Phys. J. B 12, 509–513 (1999).
4. Yang, K.-Y., Rice, T. M. & Zhang, F.-C. Phenomenological theory of the 23. Fink, J. et al. Charge order in La1.8−x Eu0.2 Srx CuO4 studied by resonant soft
pseudogap state. Phys. Rev. B 73, 174501 (2006). X-ray diffraction. Preprint at <http://arxiv.org/abs/0805.4352> (2008).
5. Haule, K. & Kotliar, G. Avoided criticality in near-optimally doped 24. Adachi, T., Noji, T. & Koike, Y. Crystal growth, transport properties, and
high-temperature superconductors. Phys. Rev. B 76, 192503 (2007). crystal structure of the single-crystal La2−x Bax CuO4 (x = 0.11). Phys. Rev. B 64,
6. Kivelson, S. A. et al. How to detect fluctuating stripes in the high-temperature 144524 (2001).
superconductors. Rev. Mod. Phys. 75, 1201–1241 (2003). 25. Takeshita, N et al. Giant anisotropic pressure effect on superconductivity
7. Chakravarty, S., Laughlin, R. B., Morr, D. K. & Nayak, C. Hidden order in the within the CuO2 plane of La1.64 Eu0.2 Sr0.16 CuO4 : Strain control of stripe
cuprates. Phys. Rev. B 63, 094503 (2001). criticality. J. Phys. Soc. Jpn. 73, 1123–1126 (2004).
8. Varma, C. Non-Fermi-liquid states and pairing instability of a general model 26. LeBoeuf, D. et al. Electron pockets in the Fermi surface of hole-doped high-Tc
of copper oxide metals. Phys. Rev. B 55, 14554–14580 (1997). superconductors. Nature 450, 533–536 (2007).
9. Ando, Y. et al. Evolution of the Hall coefficient and the peculiar electronic 27. Millis, A. J. & Norman, M. R. Antiphase stripe order as the origin of
structure of the cuprate superconductors. Phys. Rev. Lett. 92, 197001 (2004). electron pockets observed in 1/8-hole-doped cuprates. Phys. Rev. B 76,
10. Hussey, N. E. et al. Observation of a coherent three-dimensional Fermi surface 220503 (2007).
in a high-transition temperature superconductor. Nature 425, 814–817 (2003). 28. Ossadnik, M., Honerkamp, C., Rice, T. M. & Sigrist, M. Breakdown of Landau
11. Nakamae, S. et al. Electronic ground state of heavily-overdoped theory in overdoped cuprates near the onset of superconductivity. Preprint at
non-superconducting La2−x Srx CuO4 . Phys. Rev. B 68, 100502 (2003). <http://arxiv.org/abs/0805.3489> (2008).
12. Mackenzie, A. P. et al. Normal-state magnetotransport in superconducting 29. Boebinger, G. S. et al. Insulator-to-metal crossover in the normal
Tl2 Ba2 CuO6+y to millikelvin temperatures. Phys. Rev. B 53, 5848–5855 (1996). state of La2−x Srx CuO4 near optimum doping. Phys. Rev. Lett. 77,
13. Nakamura, Y. & Uchida, S. Anisotropic transport properties of single-crystal 5417–5420 (1996).
La2−y−x Ndy Srx CuO4 : Effect of the structural phase transition. Phys. Rev. B 46, 30. Balakirev, F. F. et al. Fermi surface reconstruction at optimum doping in
5841–5844 (1992). high-Tc superconductors. Preprint at <http://arxiv.org/abs/0710.4612>
14. Ando, Y. et al. Electronic phase diagram of high-Tc cuprate superconductors (2007).
from a mapping of the in-plane resistivity curvature. Phys. Rev. Lett. 93, 31. Nachumi, B. et al. Muon spin relaxation study of the stripe phase order in
267001 (2004). La1.6−x Nd0.4 Srx CuO4 and related 214 cuprates. Phys. Rev. B 58,
15. Varma, C. M. et al. Phenomenology of the normal state of Cu–O 8760-8772 (1998).
high-temperature superconductors. Phys. Rev. Lett. 63, 1996–1999 (1989).
16. Tallon, J. L. & Loram, J. W. The doping dependence of T∗ —what is the real Acknowledgements
high-Tc phase diagram? Physica C 349, 53–68 (2001). We thank K. Behnia, A. Chubukov, P. Coleman, Y.B. Kim, S.A. Kivelson, G. Kotliar,
17. Motoyama, E. M. et al. Spin correlations in the electron-doped K. Haule, G.G. Lonzarich, A.J. Millis, M.R. Norman, C. Proust, T.M. Rice, S. Sachdev,
high-transition-temperature superconductor Nd2−x Cex CuO4±δ . Nature 445, T. Senthil, H. Takagi and A.-M.S. Tremblay for discussions, and J. Corbin for his
186–189 (2007). assistance with the experiments. L.T. acknowledges support from the Canadian Institute
18. Fournier, P. et al. Insulator–metal crossover near optimal doping in for Advanced Research and funding from NSERC, FQRNT, and a Canada Research
Pr2−x Cex CuO4 : anomalous normal-state low-temperature resistivity. Phys. Chair. L.B. was supported by NHMFL-UCGP and Y.J.J. by the NHMFL-Schuller fellow
Rev. Lett. 81, 4720–4723 (1998). program. J.S.Z. and J.B.G. were supported by an NSF grant. The NHMFL is supported
19. Dagan, Y. et al. Evidence for a quantum phase transition in Pr2−x Cex CuO4−δ by an NSF grant and the State of Florida.
from transport measurements. Phys. Rev. Lett. 92, 167001 (2004).
20. Ichikawa, N. et al. Local magnetic order vs superconductivity in a layered Additional information
cuprate. Phys. Rev. Lett. 85, 1738–1741 (2000). Supplementary Information accompanies this paper on www.nature.com/naturephysics.
21. Hunt, A. W. et al. Glassy slowing of stripe modulation in Reprints and permissions information is available online at http://npg.nature.com/
(La,Eu,Nd)2−x (Sr,Ba)x CuO4 : A 63 Cu and 139 La NQR study down to 350 mK. reprintsandpermissions. Correspondence and requests for materials should be addressed
Phys. Rev. B 64, 134525 (2001). to L.T.

LETTERS
Mechanics of individual isolated vortices in a

cuprate superconductor
Ophir M. Auslaender1 *, Lan Luan1 , Eric W.J. Straver1 , Jennifer E. Hoffman1,2 , Nicholas C. Koshnick1 ,
Eli Zeldov1,3 , Douglas A. Bonn4 , Ruixing Liang4 , Walter N. Hardy4 and Kathryn A. Moler1
Superconductors often contain quantized microscopic the platelet-shaped sample with the c axis along ẑ and the a axis ∼9◦
whirlpools of electrons, called vortices, that can be modelled from x̂ (Fig. 1a).
as one-dimensional elastic objects1 . Vortices are a diverse MFM uses a sharp magnetic tip on a flexible cantilever. Our
area of study for condensed matter because of the interplay tip magnetization provided an attractive tip–vortex force, F. While
between thermal fluctuations, vortex–vortex interactions and rastering in the x–y plane, parallel to the sample surface, we
the interaction of the vortex core with the three-dimensional measured local variations in the resonant frequency of the cantilever
disorder landscape2–5 . Although vortex matter has been to determine ∂Fz /∂z (ref. 20). We deliberately used the lateral
studied extensively1,6,7 , the static and dynamic properties of components of F, Fx x̂ + Fy ŷ ≡ Flat , for vortex manipulation.
an individual vortex have not. Here, we use magnetic force At low temperature (T ≈ 5 K), we observed no vortex motion up
microscopy (MFM) to image and manipulate individual vortices to our largest lateral force, 20 pN. At higher temperatures, pinning
in a detwinned YBa2 Cu3 O6.991 single crystal, directly measuring is reduced, and we could reduce the tip–sample distance z to tune
the interaction of a moving vortex with the local disorder from non-invasive imaging to manipulation. Figure 1 shows typical
potential. We find an unexpected and marked enhancement examples for individual, well-isolated, vortices. The similarity of
of the response of a vortex to pulling when we wiggle it the behaviour of vortices pinned at different locations indicates the
transversely. In addition, we find enhanced vortex pinning uniformity of the pinning landscape in this sample.
anisotropy that suggests clustering of oxygen vacancies in our Figure 1d shows one of our two main findings: ‘vortex wiggling’.
sample and demonstrates the power of MFM to probe vortex Adding an alternating transverse force enhances vortex dragging
structure and microscopic defects that cause pinning. markedly. The wiggling occurs because of the rastering associated
A superconducting vortex is characterized by two length scales: with imaging: between incremental steps forward along the ‘slow’
the nanoscale core size and the much larger magnetic penetration scan direction, we raster the tip left and right along the ‘fast’ scan
depth, λ. Pinning can occur when the core is co-located with a defect axis. If Flat ≡ |Flat | is small, this creates an image of a stationary
that locally suppresses superconductivity. λ determines the decay vortex (Fig. 1a,c,e). When Flat is large, the vortex moves as the tip
length for the currents encircling the core and the elastic properties passes over it (Fig. 1b,d,f). Although this motion is substantial in
of a vortex. Here, we probe the dynamics of an individual driven the quasi-one-dimensional (1D) scans along the fast axis (Fig. 1f),
vortex, which is especially interesting when pinning and elasticity it is much larger along the slow axis (Fig. 1d).
compete. Furthermore, we establish a dragged vortex as a probe that We study wiggling further in Fig. 2. Figure 2a shows a scan that
extends deep into the bulk of the sample to interact with defects far we acquired after imaging the same area over and over, resulting
from the surface, circumventing disadvantages of other local-probe in an enhanced wiggling effect. Line scans along ±x̂ from that
techniques that give information only about the immediate vicinity image (Fig. 2b) do not differ from Fig. 1f. To study the vortex after
of the surface8 . Previous strategies for manipulating single vortices completing the scan in Fig. 2a, we moved the tip back and forth over
in superconductors usually applied forces relatively delocalized on the centre of the vortex along ±ŷ to obtain Fig. 2c,d. Details of the
the scale of λ (refs 9–11). We use magnetic force microscopy motion along the fast axis (Fig. 2b) are similar to motion along what
(MFM) to combine imaging and vortex manipulation with a level was the slow axis (Fig. 2c,d)—a vortex moves in jerks, reminiscent
of control far beyond what has been demonstrated before12,13 . This of avalanches5 —but the total distance moved differs substantially.
capability may enable testing vortex entanglement14 and schemes Moreover, Fig. 2c,d shows that although a vortex moves very freely
for quantum computation15,16 . between the initial and final positions in Fig. 2a, it does not readily
Our sample is ideal for studying the interplay between pinning move outside this range. In fact, we were never able to permanently
and elasticity, with weak, well-controlled, pinning and fairly rigid drag a vortex very far from its original location, contrary to the
vortices. In YBa2 Cu3 O6.991 (YBCO), superconductivity arises in case in thin films12 . This tethering suggests that each vortex was
CuO2 planes, parallel to the a and b axes, and in Cu–O chains, pinned along its full length across the 40-µm-thick crystal, and that
along the b axis. In pristine samples, such as ours, oxygen we observed the vortex stretching.
vacancies in the chains are the dominant source of pinning17–19 . A vortex presumably stops moving where elasticity and pinning
The orthorhombic crystal structure gives rise to penetration depth balance Flat . Qualitatively, wiggling helps segments of the vortex
anisotropy, enabling us to determine the orientation of the crystal to depin, facilitating the extra motion along the slow axis.
axes in situ (see Supplementary Information, Fig. S1). We mounted Confirming this is Fig. 2 with the enhancement of the effect by
1 GeballeLaboratory for Advanced Materials, Stanford University, Stanford, California 94305, USA, 2 Department of Physics, Harvard University,
Cambridge, Massachusetts 02138, USA, 3 Department of Condensed Matter Physics, Weizmann Institute of Science, Rehovot 76100, Israel, 4 Department
of Physics and Astronomy, University of British Columbia, Vancouver, British Columbia V6T 1Z1, Canada. *e-mail: ophir@stanford.edu.

a b a Span:340 pN μm –1
c d
z ||c axis
x 2 2
#1
1 1
#6
Slow scan direction →
Scan position (μm)
Scan position (μm)

#2
c y (μm) d y (μm) 0 0
#3
4 4
Slow scan direction →

←Slow scan direction
#4
#5
2 x (μm) 2 x (μm) –1 –1
–3 –2 –3 –2
–2 –2
Scan Scan
Fast scan direction → Fast scan direction → direction direction
–3 –3
Span:140 pN μm –1 Fast scan direction → 0 200 0 200
b
200 ∂Fz /∂z
e f
40 pN μm –1
(pN μm –1)
100 pN μm –1
0
–1.0 –0.5 0 0.5 1.0
Scan position (μm)
–4 –3 –2 –4 –3 –2
x (μm) x (μm) Figure 2 | MFM image and line scans at T = 20 K, showing how wiggling
enhances dragging along the slow direction. a, Scan at z = 80 nm
Figure 1 | MFM imaging and manipulation of individual vortices in YBCO max ≈ 20 pN), acquired after repeated imaging, reversing the slow scan
(Flat
at T = 22.3 K. a,b, Schematic diagrams, not drawn to scale, showing an direction for each new scan, which enhances the wiggling effect. Scan
MFM tip (triangles) that attracts a vortex (thick lines) in a sample with directions are denoted on the frame; dashed lines show the trajectories of
randomly distributed point pinning sites (dots): at ‘surveillance’ height, the the line scans in b–d. Inset: Scan with similar parameters at 5.2 K, where
applied force Flat is too weak to move the vortex (a); at manipulation height, vortices were immobile. b, Line scans from a along the horizontal dashed
the vortex moves right and then left, as the tip rasters over it (b). Here, we line. Arrows show the scan direction. c,d, Immediately after the scan in a,
illustrate what happens in a scan along +x̂, as indicated by the arrow. Also we scanned back and forth along the vertical dashed line at z = 80 nm ((c)
shown are three configurations of the vortex, previously dragged along −x̂: scans down, (d) scans up). Indices: Order of line-scan acquisition.
the blue line on the left illustrates the vortex before the tip drags it to the
right, the dashed blue line shows an intermediate configuration as it follows
the tip and the green line on the right shows the final configuration, after moves until pinning and the growing elastic force balance Flat ,
the tip moves away. c,d, MFM scans for two different scan heights (colour presumably when Flat is maximal, enabling us to treat this as a
scale gives ∂Fz /∂z; fast scan and slow scan directions are indicated on the static problem. The dependence of w, the distance dragged along
frame): z = 420 nm (maximum applied lateral force Flat max ≈ 6 pN), not low max
the fast axis, on Flat ≡ max(Flat ) is shown in Fig. 3. We have
enough to perturb vortices at this temperature (c); z = 170 nm constructed a model based on weak collective pinning (WCP) for a
max ≈ 12 pN), low enough to drag the vortices significantly (d). Inset: Scan
(Flat single vortex1 to explain this data (see Supplementary Information,
at 5.2 K, showing a stationary vortex at tip height and force comparable to Discussion S2). WCP assumes that pinning is only by the collective,
those in the main panel. e, Line cut through the data in c along the dashed cumulative, effect of many pinning sites, each too weak to pin a
line, showing the signal from a stationary vortex (blue). Overlapping it is a vortex on its own, and that vortices are elastic strings23 , as described
line cut from the reverse scan (green). f, Line cut through the data in d in Ginzburg–Landau theory1 . A characteristic length, Lc , emerges
along the dashed line, showing a typical signal from a dragged vortex. The from the competition between pinning and elasticity: a vortex takes
right arrow shows the data acquired with the tip moving along +x̂, as in d; advantage of pinning by bending on a length L Lc , but it cannot
the left arrow shows the data acquired with the tip moving back along −x̂. bend for L Lc . In WCP, the vortex is broken into elastically
coupled Lc segments, each pinned by a characteristic force, Fp .
We assume that each vortex is initially along the c axis, that only
adding wiggling cycles (further test in Supplementary Information, the top portion interacts directly with the tip and that Lc λab
Fig. S2). Wiggling is reminiscent of ‘vortex shaking’, used to (where λab is the in-plane penetration depth). Each Lc segment is
accelerate equilibration in vortex matter by oscillating a magnetic subject to elastic forces from neighbouring segments and a pinning
field perpendicular to the applied magnetic field generating the force up to Fp , modelled as static friction. We find that the top
vortices6,21,22 , and may be the mechanism for it. The models portion of a vortex depins when Flat max
> Fp Λ/Lc (Λ of order λab )
presented below describe motion without wiggling, as we observe and that:
along the fast axis. Our single-vortex data should be amenable to
more advanced and quantitative theoretical modelling to describe F max Flat
max
− F1
w ∼ lat , (1)
the wiggling. 2Fp ε⊥ /Lc
The behaviour along the fast axis can be analysed as individual
one-dimensional scans. As the tip approaches a vortex, Flat increases max
for Flat Fp Λ/Lc (see Supplementary Information, Discus-
until, if it overcomes pinning, the vortex moves. The vortex then sion S3). Here, F1 /Fp ≈ 2Λ/Lc and ε⊥ is the vortex line tension.

a b z (nm)
505 250 140
2w = 465 nm
Along y^
10 3 500
400
∂Fz /∂z
5k
200 Along x^
(pN μm –1)
∂Fzmax /∂z (pN μm –1) 5k 0.5 μm
∂Fzmax /∂z
∂Fz /∂z
400 pN μm –1
7k 200
10 2
10 k
w (nm)
20 k 0 300
Fast scan position
30 k Position
10 1 35 k z=0 200
39.5 k
42 k
10 0 100
400 600 900 1,350
z+h 0 (nm) 0
5 10 15
max
Flat (pN)
Figure 3 | Force dependence of the dragged distance, w, along the fast axis. a, Force calibration: magnitude of peaks from vortices,
max(∂Fz /∂z) ≡ ∂Fzmax /∂z, versus scan height at various temperatures. Clearly the temperature dependence in this range is weak. The solid line is a fit to
max(∂Fz /∂z) = (m̃Φ0 /π)/(z + h0 )3 (h0 = 358 ± 5 nm, m̃ = 32 ± 1 nAm, Supplementary Information, Discussion S1). The excellent fit enables us to use
Fzmax = (m̃Φ0 /2π)/(z + h0 )2 . The maximum applied lateral force is given by Flat max = αF max . For a wide range of tip shapes 0.3 < α < 0.4 (ref. 30). We set
z
α = 0.35, adding at most 25% systematic error to Flat max . Inset: Single line from a scan at z = 65 nm, T = 5.2 K, showing the peak height for an immobile
vortex (here, max(∂Fz /∂z) = 365 pN µm−1 ). b, Distance moved by the vortex along the fast direction versus z (top axis) acquired at T = 25 K versus Flat max
(bottom axis). In addition to the maximum w (filled circles), we plot the distribution (diamonds and crosses), which shows the stochasticity of the vortex
motion. Other errors are not shown. Dashed lines are fits to equation (1) (along x̂:F1 = 6.3 ± 0.4 pN, Fp ε⊥ /Lc = 272 ± 12p N2 µm−1 ; along ŷ:
F1 = 4.1 ± 0.7 pN, Fp ε⊥ /Lc = 212 ± 17 pN2 µm−1 ). Inset: Subsequent line scans showing the jerky nature of the vortex motion and how w was extracted from
the difference of the vortex positions in pairs of subsequent line scans.
a b
0.5 μm
0.5 μm
Slow scan direction ←
Fast scan direction ←

Span:280 pN μm –1
Span:280 pN μm –1
e
b axis
b axis
a axis a axis
Fast scan direction ← Slow scan direction ←
c b axis d b axis
120° 60° 120° 60°
le gle
150° 150° 30°
an
30°
g
an
can
an
a axis
s
a axis
c
Fast
Fast s
180°
0.5 1.0 1.5 0.5 1.0 1.5 w (μm)
w (μm)
210°z = 80 nm 330° T = 20 K
210° 330°
z = 130 nm T = 15 K
240° 300° T = 10240°
K 300°
z = 230 nm 270° 270°
Figure 4 | Images and analysis showing the anisotropy of dragging when the fast axis is along the a axis and when it is along the b axis. a,b, Vortex
max ≈ 20 pN). Note the erratic nature of the motion and the abrupt snap-in to the
motion when the fast scan is along x̂ (a) and ŷ (b) (T = 20 K, z = 80 nm, Flat
tip on approaching the vortex along the slow axis, apparent as a sharp onset of the signal. Insets: Images of immobile vortices obtained with the same tip,
scan height and scan directions at T = 5.2 K. c,d, Distance moved along the fast axis, w, measured near the maximum lateral force, versus scan angle
measured from x̂. Bars denote 70% confidence intervals. Dashed lines show the fit described in the text, with the oxygen vacancy cluster size 2Rb as a free
max ≈ 20, 15, 10 pN, respectively). In d, z = 80 nm, for T = 20, 15, 10 K.
parameter. Fit result: 2Rb /ξab = 0.7. In c, T = 20 K at z = 80 nm, 130 nm, 230 nm (Flat
e, Schematic diagram of vortex core meandering across the crystal in the presence of point defects clustered along the b axis.

max
Figure 3b shows a reasonable fit for large Flat , where the exact shape Methods
of the top portion of the vortex and the fine details of the pinning Frequency-modulated MFM (ref. 20). We oscillate the cantilever along ẑ and
landscape are not important. measure the resonant frequency, f0 , which shifts by 1f in a force gradient.
Assuming small oscillation amplitude, ∂Fz /∂z ≈ −2c1f /f0 (where c is the
Figure 3b also highlights our second main finding—in-plane
cantilever spring constant, see Supplementary Information, Table S1).
anisotropy due to the crystal: we can drag a vortex farther along the
fast axis if it is along the YBCO b axis than if it is along the a axis. Sample details. The platelet-shaped single crystal was grown from flux in a
This is also apparent in Fig. 4. Figure 4c,d shows the dependence of BaZrO3 crucible for high purity and crystallinity17 . The (001) surfaces were free
w on the fast scan angle (see Supplementary Information, Fig. S3, of visible inclusions. Mechanical detwinning was followed by annealing to oxygen
for w here). We found a similar effect using different tips (see content 7 − δ = 6.991, implying Tc ≈ 88 K (ref. 19). The sample was stored at
room temperature for a few years. The finite a–b anisotropy enabled us to identify
Supplementary Information, Table S1) and for every vortex that crystal orientation by determining the directions along which vortex spacings
we probed. were extremal (see Supplementary Information, Fig. S1). The ratio between these
Equation (1) explains the anisotropy qualitatively. The pene- spacings is ζ = 1.3, corroborating previous results on similar samples25 .
tration depth imparts in-plane anisotropy to ε⊥ (refs 1,24), im-
plying that it is easier to tilt a vortex towards the b axis. In Received 4 July 2008; accepted 16 September 2008;
addition, because of the in-plane anisotropy of the vortex core published online 16 November 2008
radius, even for point pinning, the effective pinning potential is
less steep along the b axis, implying a smaller depinning force. References
Equation (1) thus naturally explains the weak dependence of the 1. Blatter, G., Feigel’man, M. V., Geshkenbein, V. B., Larkin, A. I. &
max max
anisotropy on Flat and T , seen in Fig. 4c,d. When Flat F1 , Vinokur, V. M. Vortices in high-temperature superconductors. Rev. Mod.
max Phys. 66, 1125–1388 (1994).
the only dependence on Flat is an overall scale. The same is
2. Fisher, M. P. A. & Lee, D. H. Correspondence between two-dimensional
true for the T dependence, because both the superconducting
bosons and a bulk superconductor in a magnetic field. Phys. Rev. B 39,
parameters and the defect structure depend only weakly on T in 2756–2759 (1989).
our range25 . 3. Nelson, D. R. & Vinokur, V. M. Boson localization and correlated pinning of
For quantitative analysis, we have determined the angular superconducting vortex arrays. Phys. Rev. B 48, 13060–13097 (1993).
dependence of w ∝ Lc /ε⊥ Fp (see Supplementary Information, 4. Ertaş, D. & Kardar, M. Anisotropic scaling in depinning of a flux line. Phys.
Discussion S4), which depends on the anisotropy factor ζ . Fitting Rev. Lett. 73, 1703–1706 (1994).
5. Sethna, J. P., Dahmen, K. A. & Myers, C. R. Crackling noise. Nature 410,
all five data sets in Fig. 4c,d, we find ζ = 1.6 (see Supplementary 242–250 (2001).
Information, Fig. S5), in clear disagreement with our direct 6. Avraham, N. et al. ‘Inverse’ melting of a vortex lattice. Nature 411,
measurement (see Supplementary Information, Fig. S1) and the 451–454 (2001).
known value25 , ζ = 1.3. We surmise that there is an extra source 7. Pardo, F., de la Cruz, F., Gammel, P. L., Bucher, E. & Bishop, D. J. Observation
of anisotropy. of smectic and moving-Bragg-glass phases in flowing vortex lattices. Nature
396, 348–350 (1998).
A likely source for extra pinning anisotropy is nanoscale 8. Hoffman, J. E. et al. Imaging quasiparticle interference in Bi2 Sr2 CaCu2 O8+δ .
clustering of oxygen vacancies along the Cu–O chains26,27 Science 297, 1148–1151 (2002).
(Fig. 4e). We modified our model (see Supplementary Information, 9. Park, G. S., Cunningham, C. E., Cabrera, B. & Huber, M. E. Vortex pinning
Discussion S5), imposing ζ = 1.3, and fit the data in Fig. 4c,d force in a superconducting niobium strip. Phys. Rev. Lett. 68,
for the cluster size. We find clusters of order 10 vacancies along 1920–1922 (1992).
10. Breitwisch, M. & Finnemore, D. K. Pinning of a single Abrikosov vortex in
the Cu–O chains (see Supplementary Information, Discussion superconducting Nb thin films using artificially induced pinning sites. Phys.
S5), large but reasonable for the sample’s length of time at room Rev. B 62, 671–677 (2000).
temperature, where oxygen vacancies migrate slowly and cluster28 . 11. Gardner, B. W. et al. Manipulation of single vortices in YBa2 Cu3 O6.354 with a
It is also possible that non-Ginzburg–Landau physics affects the locally applied magnetic field. Appl. Phys. Lett. 80, 1010–1012 (2002).
core structure29 , changing details of the pinning-force anisotropy. 12. Moser, A., Hug, H. J., Stiefel, B. & Güntherodt, H.-J. Low temperature
magnetic force microscopy on YBa2 Cu3 O7−δ thin films. J. Magn. Magn. Mater.
These results demonstrate that single-vortex manipulation is 190, 114–123 (1998).
a local probe of both the core and the defect structure on a 13. Roseman, M. & Grütter, P. Magnetic imaging and dissipation force microscopy
scale down to the core size. For example, in samples without of vortices on superconducting Nb films. Appl. Surf. Sci. 188, 416–420 (2002).
intrinsic a–b anisotropy and with known defect structure, MFM 14. Olson Reichhardt, C. J. & Hastings, M. B. Do vortices entangle? Phys. Rev. Lett.
could be used to probe the intrinsic structure of the vortex 92, 157002 (2004).
15. Berciu, M., Rappoport, T. G. & Boldizsár, J. Manipulating spin and charge in
core itself.
magnetic semiconductors using superconducting vortices. Nature 435,
Despite the fact that YBCO is one of the most studied 71–75 (2005).
superconductors, our data reveal major surprises about the 16. Weeks, C., Rosenberg, G., Seradjeh, B. & Franz, M. Anyons in a weakly
behaviour of individual vortices. A model based on weak collective interacting system. Nature Phys. 3, 796–801 (2007).
pinning for a single vortex quantitatively describes quasistatic 17. Liang, R., Bonn, D. A. & Hardy, W. N. Growth of high quality YBCO single
aspects of vortex motion with the incorporation of anisotropy crystals using BaZrO3 crucibles. Physica C 304, 105–111 (1998).
18. Liang, R. X., Bonn, D. A. & Hardy, W. N. Preparation and X-ray
in the local microscopic pinning. This demonstrates that single- characterization of highly ordered ortho-II phase YBa2 Cu3 O6.50 single crystals.
vortex manipulation is a local probe of the structure of both Physica C 336, 57–62 (2000).
the vortex and the pinning defects. Further work is required 19. Liang, R., Bonn, D. A. & Hardy, W. N. Evaluation of CuO2 plane hole doping
to describe the dynamic aspects of individual vortex motion in YBa2 Cu3 O6+x single crystals. Phys. Rev. B 73, 180505 (2006).
that we revealed: its stochastic nature and the marked effect of 20. Albrecht, T. R., Grütter, P., Horne, D. & Rugar, D. Frequency-modulation
detection using high-Q cantilevers for enhanced force microscope sensitivity.
transverse wiggling. In particular: how do the dynamics alter the J. Appl. Phys. 69, 668–673 (1991).
effective pinning landscape? How do they affect the mechanical 21. Willemin, M. et al. Strong shift of the irreversibility line in high-Tc
properties? Practically, wiggling is an important tool for future superconductors upon vortex shaking with an oscillating magnetic field. Phys.
experiments that require pulling vortices long distances, for Rev. B 58, R5940–R5943 (1998).
example, in the study of vortex entanglement14 . Our results show 22. Brandt, E. H. & Mikitik, G. P. Why an ac magnetic field shifts the irreversibility
line in type-II superconductors. Phys. Rev. Lett. 89, 027002 (2002).
the utility of local force probes for accessing the pinning properties 23. Brandt, E. H. Large range of validity of linear elasticity of the vortex lattice in
and mechanical behaviour of individual vortices, the collective high-Tc superconductors. Phys. Rev. Lett. 69, 1105–1108 (1992).
behaviour of which is of great importance for the properties 24. Burlachkov, L. I. Structure of vortex lattice in biaxial superconductors.
of superconductors. Europhys. Lett. 8, 673–677 (1989).

25. Pereg-Barnea, T. et al. Absolute values of the London penetration depth in Acknowledgements
YBa2 Cu3 O6+y measured by zero field ESR spectroscopy on Gd doped single Research supported by DOE Contract No. DE-AC02-76SF00515, AFOSR grant
crystals. Phys. Rev. B 69, 184513 (2004). FA50-05-1-0290, the Packard Foundation and the US–Israel Binational Science
26. Erb, A. et al. Experimental evidence for fast cluster formation of chain oxygen Foundation (BSF). The experimental set up was developed with extensive help from
vacancies in YBa2 Cu3 O7−δ as the origin of the fishtail anomaly. Solid State D. Rugar. We acknowledge useful discussions with J. R. Kirtley, H. Bluhm, G. P. Mikitik,
Commun. 112, 245–249 (1999). E. H. Brandt and V. B. Geshkenbein.
27. Nishizaki, T., Shibata, K., Maki, M. & Kobayashi, N Vortex phase transition
and oxygen vacancy in YBa2 Cu3 Oy single crystals. J. Low Temp. Phys. 131,
Author contributions
Conception and design: O.M.A., L.L., E.W.J.S., J.E.H., N.C.K. and K.A.M. Experiments:
931–940 (2003).
O.M.A. and L.L. Analysis: O.M.A. and L.L. with ideas developed with E.Z. and K.A.M.
28. Liang, R. et al. Preparation and characterization of homogeneous YBCO single Sample growth: D.A.B., R.L. and W.N.H. Manuscript preparation: O.M.A., L.L., E.Z. and
crystals with doping level near the SC-AFM boundary. Physica C 383, K.A.M., with input from all co-authors.
1–7 (2002).
29. Sonier, J. E. et al. Expansion of the vortex cores in YBa2 Cu3 O6.95 at low Additional information
magnetic fields. Phys. Rev. B 59, R729–R732 (1999). Supplementary Information accompanies this paper on www.nature.com/naturephysics.
30. Wadas, A., Fritz, O., Hug, H. J. & Güntherodt, H.-J. Magnetic force Reprints and permissions information is available online at http://npg.nature.com/
microscopy signal of flux line above a semiinfinite reprintsandpermissions. Correspondence and requests for materials should be addressed
type-II-superconductor—theoretical approach. Z. Phys. B 88, 317–320 (1992). to O.M.A.

LETTERS
Fermi-surface-dependent superconducting gap

in C6Ca
K. Sugawara1 , T. Sato1 * and T. Takahashi1,2
The discovery of superconductivity in C6 Yb and C6 Ca (ref. 1)

has activated fierce debates on whether it is described FC
0
within the conventional Bardeen–Cooper–Schrieffer scheme
or some other exotic mechanisms are involved, because the C6Ca
superconducting transition temperature (Tc ) is significantly –0.2
Moment (e.m.u. cm –3)

higher than that of the alkali-metal graphite intercalation 1 mm
compounds intensively studied in the 1980s (refs 2–4). –0.4
The key to understand the mechanism of superconductivity Tc= 11.5 K
lies in the superconducting energy gap associated with the –0.6
formation of superconducting pairs. Here, we report the first
direct observation of a superconducting gap in C6 Ca by high- –0.8
resolution angle-resolved photoemission spectroscopy. We ZFC H = 10 Oe
found that the superconducting gap of 1.8–2.0 meV opens on
–1.0
the intercalant Fermi surface, whereas the gap is very small
or absent on the Fermi surface derived from the π∗ band of
graphene layers. These experimental results unambiguously 6 8 10 12 14
establish that the interlayer band has an essential role for the Temperature (K)
high-Tc superconductivity in C6 Ca.
Although the thermodynamic5 and magnetic properties6 of Figure 1 | Characterization of C6 Ca single crystal. Magnetic susceptibility
C6 Ca infer the isotropic s-wave superconducting order parameter of C6 Ca as a function of temperature for zero field cooling (ZFC) and field
within the conventional Bardeen–Cooper–Schrieffer theory, cooling (FC) at 10 Oe. Inset: Photograph of a C6 Ca single crystal used for
substantial deviations and anomalies beyond the conventional ARPES measurements.
framework have been reported by tunnelling spectroscopy7,8
and electron spin resonance experiments9 , proposing the cooling at 10 Oe. The sharp drop in the zero-field-cooling curve
strong-coupling superconductivity, the anisotropic gap and/or confirms that the crystal exhibits superconductivity with the onset
the multi-component gaps. The discovery of ‘high-Tc ’ graphite Tc = 11.5 K.
intercalation compounds has stimulated several theoretical Figure 2a,b shows the experimentally determined valence-band
proposals for the superconducting mechanism, ranging from structure along two high-symmetry lines 0K and 0M. Although the
the electronic pairing10 to the coupling to lattice vibrations overall band structure looks similar between graphite and C6 Ca,
(phonons)11–15 . Recent first-principles calculations11–15 have the band dispersion of C6 Ca is shifted as a whole towards high
suggested that the electron coupling to the out-of-plane carbon and binding energy by 1.0–1.5 eV, indicating that electrons are certainly
intercalant phonons is responsible for achieving higher Tc values, doped into pristine graphite by Ca intercalation. A small electron
although a clear demonstration to experimentally distinguish pocket that appears near the Fermi level (EF ) at the K point in
various models has not yet been made. This key issue as well C6 Ca is ascribed to the C 2pπ∗ band, which is above EF in pristine
as the discrepancy in the nature of the superconducting gap graphite and is shifted down below EF by electron doping16–18 . It
among experiments urgently requires further precise investigations is noted that the overall valence-band dispersion basically follows
on the momentum and Fermi-surface dependence of the the periodicity of the Brillouin zone for graphite rather than
superconducting gap. Although angle-resolved photoemission that for C6 Ca, possibly owing to the relatively weaker periodic
spectroscopy (ARPES) is most suitable for these investigations, potential of C6 Ca with the R3̄m symmetry than the potential
few such experiments have been carried out because of the small of bulk graphite. Figure 2c shows the Fermi surface of C6 Ca
superconducting gap as well as the lack of high-quality single determined by ARPES, compared with that of graphite (Fig. 2d).
crystals. We used a high-quality C6 Ca single crystal grown from The Fermi-surface topology of C6 Ca is drastically different from
kish graphite (artificially grown single-crystalline graphite). We that of graphite. In C6 Ca, we find two different Fermi surfaces, a
have succeeded for the first time in observing the interlayer band circular Fermi surface at the 0 point and a triangular Fermi surface
and the Fermi-surface-sheet-dependent superconducting gap and at the K point, whereas only one tiny Fermi surface is observed
electron–phonon coupling. at the K(H) point in graphite19 . The triangular Fermi surface at
Figure 1 shows the temperature dependence of the magnetic the K point originates in the π∗ band and is well explained in
susceptibility of the C6 Ca sample for zero field cooling and field terms of the simple rigid shift of the chemical potential due to the
1 Departmentof Physics, Tohoku University, Sendai 980-8578, Japan, 2 WPI Research Center, Advanced Institute for Material Research, Tohoku University,
Sendai 980-8577, Japan. *e-mail: t-sato@arpes.phys.tohoku.ac.jp.

a b
EF EF
1 1
2 2
EF EF
Binding energy (eV)

Binding energy (eV)
4 4
8 8 High
12 12
16 Low
16
M K Γ M K M (L) Γ (A) K (H)
Wave vector Wave vector
c d
B′ M (L)
B M
K (H)
1.0 1.0
K
A
Γ (A)
ky(Å–1)
ky(Å–1)
M High
0 Γ A′ 0
K
–1.0 –1.0
Low
–1.0 0 1.0 –1.0 0 1.0
kx(Å–1) kx(Å–1)
Figure 2 | Comparison of electronic structure and Fermi surface between C6 Ca and graphite. a,b, Valence-band structure of C6 Ca (a) and graphite (b),
determined by plotting the ARPES spectral intensity as a function of wave vector and binding energy. Expansion near EF is also shown in the upper panel.
c,d, Fermi surfaces for C6 Ca (c) and graphite (d), determined by plotting the ARPES spectral intensity integrated within ±10 meV with respect to EF and
folding with the hexagonal symmetry. Solid red and white lines in c are surface Brillouin zones of graphene and C6 Ca, respectively.
electron doping20 . In contrast, the small Fermi surface observed at on the triangular Fermi surface (point B). The energy position of
the 0 point is not explained by the simple rigid shift of graphite the 170 meV dip is similar to that of the dispersion kink observed
bands or the folding of the π∗ band due to the superstructure of in graphite21 and C6 Ca (ref. 22). We find that the energy position
Ca intercalants, because the shape and size of the Fermi surface of the dip does not vary in the momentum region around the kF
are distinctly different from those of the π∗ Fermi surface. On the point. This is not explained by the presence of another band located
other hand, the band-structure calculations10–15 have predicted an around the dip energy, but is reminiscent of the strong many-body
extra band at the 0 point in C6 Ca named the interlayer band, effects observed in high-Tc copper oxides23,24 and Ba0.67 K0.33 BiO3
which has a sizable contribution from the intercalant states and (ref. 25). In fact, as seen in Fig. 3a, the Eliashberg function
produces a small free-electron-like spherical Fermi surface at the α 2 F (ω) calculated for C6 Ca (ref. 11) shows dominant peaks at
0 point. Considering a good agreement in the location and the 50–75 meV and 160–180 meV originating in the out-of-plane and
shape as well as the superconducting-gap size (discussed later the in-plane vibrations of carbon atoms, respectively, in good
in detail), we assign the small Fermi surface observed at the 0 agreement with the anomalies in the ARPES spectra. This suggests
point to the interlayer band in C6 Ca. To our best knowledge, this that electrons on the circular Fermi surface at the 0 point are
is the first direct observation of the interlayer band in graphite strongly coupled to the out-of-plane vibrations, whereas those on
intercalation compounds. the triangular Fermi surface at the K point are coupled dominantly
Figure 3a shows ARPES spectra near EF of C6 Ca at 6 K measured with the in-plane phonons. This experimentally demonstrates that
at two representative Fermi vectors (kF ) on the circular Fermi the interlayer electrons strongly interact with the surrounding
surface at the 0 point and the triangular Fermi surface at the K lattice unlike in the free-electron picture, as predicted from
point, respectively. The spectral intensity shows a sudden drop at first-principles calculations13,14 . The reason π∗ electrons at the
80 meV on the circular Fermi surface (point A), whereas a small K point are only weakly coupled to the out-of-plane phonons
but clear dip is observed at 170 meV in the spectrum measured may be the antisymmetry of the wave function with respect to

a b
C6Ca
Tc = 11.5 K
Intensity (arb. units)

A
T=
6K
C6Ca 15 K B
Tc = 11.5 K
10 5 EF –5
Binding energy (meV)
A
c
1.0 T=6K
C (xy) C (z) A

A′
B
Ca B′
α 2F(ω)
0.5 (xy) B
Δ = 1.8 meV
Γ = 0.5 meV A
0.4 0.3 0.2 0.1 EF 10 5 EF –5

Binding energy (eV) Binding energy (meV)
Figure 3 | Electron–phonon coupling and superconducting gap in C6 Ca. a, ARPES spectra of C6 Ca measured at 6 K on (A) the circular Fermi surface and
(B) the triangular Fermi surface (points A and B in Fig. 2c, respectively), compared with the calculated Eliashberg function α 2 F(ω) as a function of binding
energy11 . Filled triangles denote the characteristic energy scales where the spectrum shows a distinct anomaly. b, ARPES spectra of C6 Ca in the vicinity of
EF at points A and B, measured at temperatures below/above Tc (6 K and 15 K). c, Top: Comparison of ARPES spectrum at 6 K among points A, A0 , B and B0 .
Bottom: Numerical fit (black curve) to the 6 K spectrum at point A (red circles) with the Dynes function26 multiplied by the Fermi–Dirac function at T = 6 K
convoluted with a Gaussian taking into account the instrumental resolution.
the graphene layer11 . On the other hand, the interlayer band with that of point A (B). This suggests that the anisotropy
forming the circular Fermi surface at the 0 point is expected of the superconducting gap is very small in both circular and
to be strongly coupled to the out-of-plane phonons owing to triangular Fermi surfaces, although the gap size is markedly
the symmetry14 . A spectral dip related to the coupling with the different between these two Fermi surfaces. To estimate the size
Ca vibrations is not clearly observed in Fig. 3a, possibly because of the superconducting gap, we numerically fit the spectrum using
the dip is broad (20–30 meV) compared with the Ca phonon the Dynes function26 . The representative result for point A is
energy (15 meV). shown in Fig. 3c. The estimated gap size ∆ at 6 K is 1.8 ± 0.2
Figure 3b shows ultrahigh-resolution ARPES spectra near EF and 2.0 ± 0.2 meV at points A and A0 , respectively. The reduced
of C6 Ca measured at the same two points as in Fig. 3a at gap value of 2∆(0)/kB Tc determined by assuming the mean-field
temperatures below/above Tc . At point A, the midpoint of the temperature dependence is 4.1 ± 0.5, indicating intermediate- to
leading edge in the spectrum at 6 K is shifted towards high binding strong-coupling character. The first-principles calculation13 has
energy by approximately 0.8 meV with respect to EF , whereas the predicted a similar gap value. On the other hand, at points B and
spectrum at 15 K has the midpoint at EF . This clearly indicates B0 , the estimated superconducting-gap size is 0.2 ± 0.2 meV. These
the opening/closing of a superconducting gap as a function of results suggest that the deviation from the single isotropic s-wave
temperature. On the other hand, at point B, the midpoint of gap function in the tunnelling density of states7 , interpreted in terms
the leading edge is located at around EF even at 6 K, and the of the anisotropic s-wave gap, is not explained by the anisotropy of
spectrum shows a simple temperature dependence to obey the the gap within the interlayer Fermi surface at the 0 point, but is
Fermi–Dirac function, suggesting the absence or the very small actually influenced by a smaller superconducting gap of other Fermi
magnitude of the superconducting gap. This marked difference is surfaces, most likely the π∗ Fermi surface. The reported deviation
better illustrated in Fig. 3c, where the spectra at 6 K at points A of the specific heat27 from the single isotropic s-wave behaviour13,15
and B are directly compared. To clarify the possible anisotropy of would also be explained by this Fermi-surface dependence of the
the gap in each Fermi-surface sheet, we also plot the 6 K spectra superconducting gap.
measured at points A0 and B0 on the same Fermi surfaces but on The present results demonstrate that the coupling of electrons
different high symmetry lines (Fig. 2c), because the anisotropy, in the interlayer band with the out-of-plane phonons is essential
if it exists, should be largest between points A and A0 (B and to realize superconductivity in C6 Ca, whereas simple electron
B0 ). We note that the ARPES technique basically observes the doping into the π∗ band of graphene layers may not lead to
gap averaged over kz (k perpendicular to the surface), so that it superconductivity at such a high temperature. The next challenge
measures the anisotropy of the gap in the kx –ky plane. As seen is to clarify the role of intercalant (Ca) phonons, which have been
from Fig. 3c, the leading edge at point A0 (B0 ) almost coincides theoretically proposed to enhance the Tc value11–15 .

Methods 13. Sanna, A. et al. Anisotropic gap of superconducting CaC6 : A first-principles
High-quality C6 Ca single crystals were grown by heating kish graphite mixed with density functional calculation. Phys. Rev. B 75, 020511(R) (2007).
molten lithium–calcium alloy at around 350 ◦ C for 6 days under high vacuum28 . 14. Boeri, L. et al. Electron–phonon interaction in graphite intercalation
The typical size of the crystals was 3 × 3 × 0.5 mm3 . ARPES measurements were compounds. Phys. Rev. B 76, 064510 (2007).
carried out using an SES-2002 spectrometer with a high-flux discharge lamp and 15. Mazin, I. I et al. Unresolved problems in superconductivity of CaC6 . Physica C
a toroidal grating monochromator at Tohoku University. As the sample is very 460–462, 116–120 (2007).
unstable in the atmosphere, almost all of the steps such as sample mounting were 16. Takahashi, T. et al. Electronic band structure of C8 K studied by
done in a glove box filled with Ar gas. The superconducting volume fraction angle-resolved ultraviolet photoelectron spectroscopy. J. Phys. Soc. Jpn 55,
is estimated to be 90 ± 10% at 5 K, higher than previous reports1,28 . The X-ray 3498–3502 (1986).
diffraction measurement confirmed that inclusion of C6 Li is at most 5%. The 17. Molodtsov, S. L. et al. Folded bands in photoemission spectra of La-graphite
crystal was cleaved in situ along the (0001) plane in an ultrahigh vacuum better intercalation compounds. Phys. Rev. B 67, 115105 (2003).
than 2 × 10−11 torr to obtain a clean surface for ARPES measurements. The He IIα 18. McChesney, J. L. et al. Massive enhancement of electron–phonon coupling in
(40.814 eV) resonance line was used to excite photoelectrons. The energy and doped graphene by an electronic singularity. Preprint at
angular (momentum) resolutions were set at 4.5 meV and 0.2◦ (0.01 Å−1 ), except <http://arxiv.org/abs/0705.3264> (2007).
for in the measurement of the wide valence-band region (1E = 16 meV). The 19. Sugawara, K. et al. Fermi surface and edge-localized states in graphite studied
temperature of the samples was controlled to remain stationary within ±0.1 K by high-resolution angle-resolved photoemission spectroscopy. Phys. Rev. B
during measurements by monitoring with a silicon diode sensor embedded below 73, 045124 (2006).
the sample. The Fermi level of the sample was referenced to that of a gold film 20. Tatar, R. C. & Rabii, S. Electronic properties of graphite: A unified theoretical
deposited onto the sample substrate. No degradation of the sample surface was study. Phys. Rev. B 25, 4126–4141 (1982).
observed during the measurements. We checked the reproducibility of the data by 21. Sugawara, K. et al. Anomalous quasiparticle lifetime and strong
measuring several samples. electron–phonon coupling in graphite. Phys. Rev. Lett. 98, 036801 (2007).
22. Valla, T. et al. Anisotropic electron–phonon coupling and dynamical
Received 8 April 2008; accepted 8 October 2008; nesting on the graphene sheets in CaC6 . Preprint at
<http://arxiv.org/abs/0803.0254> (2008).
published online 9 November 2008 23. Damascelli, A., Hussain, Z. & Shen, Z. -X. Angle-resolved photoemission
References studies of the cuprate superconductors. Rev. Mod. Phys. 75, 473–541 (2003).
24. Campuzano, J. C., Norman, M. R. & Randeria, M. in Physics of Conventional
1. Weller, T. E., Ellerby, M., Saxena, S. S., Smith, R. P. & Skipper, N. T.
and Unconventional Superconductors Vol. 2 (eds Bennemann, K. H. &
Superconductivity in the intercalated graphite compounds C6 Yb and C6 Ca.
Ketterson, J. B.) 167–273 (Springer, 2004).
Nature Phys. 1, 39–41 (2005).
25. Chainani, A. et al. Electron–phonon coupling induced pseudogap and the
2. Dresselhaus, M. S. & Dresselhaus, G. Intercalation compounds of graphite.
superconducting transition in Ba0.67 K0.33 BiO3 . Phys. Rev. B 64,
Adv. Phys. 51, 1–186 (2002).
180509(R) (2001).
3. Hannay, N. B. et al. Superconductivity in graphitic compounds. Phys. Rev.
26. Dynes, R. C., Narayanamurti, V. & Garno, J. P. Direct measurement of
Lett. 14, 225–226 (1965).
quasiparticle-lifetime broadening in a strong-coupled superconductor. Phys.
4. Belash, I. T., Bronnikov, A. D., Zharikov, O. V. & Palnichenko, A. V. On the
Rev. Lett. 41, 1509–1512 (1978).
superconductivity of graphite intercalation compounds with sodium. Solid
27. Kim, J. S., Kremer, R. K., Boeri, L. & Razavi, F. S. Specific heat of the
State Commun. 64, 1445–1447 (1987).
Ca-intercalated superconductor CaC6 . Phys. Rev. Lett. 96, 217002 (2006).
5. Sutherland, M. et al. Bulk evidence for single-gap s-wave superconductivity in
28. Emery, N. et al. Superconductivity of bulk CaC6 . Phys. Rev. Lett. 95,
the intercalated graphite superconductor C6 Yb. Phys. Rev. Lett. 98,
087003 (2005).
067003 (2007).
6. Lamura, G. et al. Experimental evidence of s-wave superconductivity in bulk
CaC6 . Phys. Rev. Lett. 96, 107008 (2006). Acknowledgements
7. Bergeal, N. et al. Scanning tunnelling spectroscopy on the novel We thank H. Suematsu for providing high-quality kish graphite. We also thank
superconductor CaC6 . Phys. Rev. Lett. 97, 077003 (2006). K. Tanigaki, K. Sato, T. Rachi and K. Kadowaki for their help in the sample preparation.
8. Kurter, C. et al. Large energy gaps in CaC6 from tunnelling spectroscopy: This work was supported by grants from JSPS, JST-CREST and MEXT of Japan.
Possible evidence of strong-coupling superconductivity. Phys. Rev. B 76,
220502(R) (2007). Author contributions
9. Murànui, F., Urbanik, G., Kataev, V. & Büchner, B. Electron spin dynamics of K.S. and T.S. shared responsibility for all aspects of this project including instrument
the superconductor CaC6 probed by ESR. Phys. Rev. B 77, 024507 (2008). construction, data collection, analysis and sample growth. T.T. contributed to the project
10. Csanyi, G. et al. The role of the interlayer state in the electronic structure of planning and managed the whole project.
superconducting graphite intercalated compounds. Nature Phys. 1,
42–45 (2005).
11. Calandra, M. & Mauri, F. Theoretical explanation of superconductibity in Additional information
C6 Ca. Phys. Rev. Lett. 95, 237002 (2005). Reprints and permissions information is available online at http://npg.nature.com/
12. Mazin, I. I. Intercalant-driven superconductivity in YbC6 and CaC6 . Phys. Rev. reprintsandpermissions. Correspondence and requests for materials should be addressed
Lett. 95, 227001 (2005). to T.S.

LETTERS
Thermal-transport measurements in a quantum

spin-liquid state of the frustrated triangular
magnet κ-(BEDT-TTF)2Cu2(CN)3
Minoru Yamashita1 *, Norihito Nakata1 , Yuichi Kasahara1,2 , Takahiko Sasaki2 , Naoki Yoneyama2 ,
Norio Kobayashi2 , Satoshi Fujimoto1 , Takasada Shibauchi1 and Yuji Matsuda1
The notion of quantum spin-liquids (QSLs), antiferromagnets 2.5

with quantum fluctuation-driven disordered ground states, is A
now firmly established in one-dimensional (1D) spin systems
as well as in their ladder cousins. The spin-1/2 organic 2.0
insulator κ-(bis(ethylenedithio)-tetrathiafulvalene)2 Cu2 (CN)3
(κ-(BEDT-TTF)2 Cu2 (CN)3 ; ref. 1) with a 2D triangular lattice
structure is very likely to be the first experimental realization
1.5
κ (W K –1 m –1)
of this exotic state in D ≥ 2. Of crucial importance is to unveil
the nature of the low-lying elementary spin excitations2,3 , B
particularly the presence/absence of a ‘spin gap’, which
will provide vital information on the universality class of 1.0
this putative QSL. Here, we report on our thermal-transport
measurements carried out down to 80 mK. We find, rather t t
unexpectedly, unambiguous evidence for the absence of a t’
gapless excitation, which sharply contradicts recent reports of 0.5
heat capacity measurements4 . The low-energy physics of this
intriguing system needs be reinterpreted in light of the present
results indicating a spin-gapped QSL phase. 0
In antiferromagnetically coupled spin systems, geometrical 0 2 4 6 8 10
T (K)
frustrations enhance quantum fluctuations. Largely triggered
by the proposal of the resonating-valence-bond theory for
S = 1/2 degrees of freedom residing on a frustrated two- Figure 1 | Temperature dependence of the in-plane thermal conductivity
dimensional (2D) triangular lattice5–7 and its possible appli- below 10 K. κ(T) in zero field for two different single crystals of deuterated
cation to high-Tc cuprates with a doped 2D square lattice8,9 , κ-(BEDT-TTF)2 Cu2 (CN)3 (sample A and sample B) measured in a 3 He
realizing/detecting QSLs in 2D systems has been a long- cryostat (black for sample A and green for sample B) and dilution
sought goal. Recently, discoveries of QSL states on S = 1/2 refrigerator (blue for sample A and light green for sample B). As the
triangular lattices have been reported in organic compounds, temperature is lowered, κ(T) decreases and exhibits a broad hump starting
κ-(BEDT-TTF)2 Cu2 (CN)3 (Fig. 1, inset)1,10,11 , C2 H5 (CH3 )3 Sb to increase at around T ∗ ' 6 K. Inset: The crystal structure of a
[Pd(1,3-dithiole-2-thione-4,5-dithiolate)2 ]2 (ref. 12) and 3 He thin two-dimensional BEDT-TTF layer of κ-(BEDT-TTF)2 Cu2 (CN)3 viewed
film on graphite13 . In particular, the NMR spectrum of along the long axes of BEDT-TTF molecules. Pairs of BEDT-TTF molecules
κ-(BEDT-TTF)2 Cu2 (CN)3 exhibits no signs of magnetic ordering form dimers arranged in a triangular lattice in terms of transfer integrals t
down to ∼30 mK, which is some four orders of magnitude below and t0 between the dimers. The ratio of transfer integrals is nearly unity11
the exchange coupling J ∼ 250 K (refs 1,11). These findings aroused and the spin-1/2 nearly isotropic triangular lattice is realized1 .
great interest because it is generally believed that whereas a QSL
state is realized in the strongly frustrated S = 1/2 2D kagome To understand the nature of novel QSL states, knowledge
lattice14 , which can be viewed as corner-sharing triangles, the on the structure of the low-lying excitation spectrum in the
classical magnetically ordered state is stable in the less frustrated zero-temperature limit, particularly the absence/presence of a spin
isotropic Heisenberg triangular lattice15,16 . Several ideas, such as gap, is indispensable, bearing immediate implications on the spin
a Hubbard model with a moderate onsite repulsion17 , a ring correlations of the ground state, as well as on the quantum
exchange model18 and one-dimensionalization by a slight distortion numbers carried by each elementary excitation. For instance in 1D,
from the isotropic triangular lattice19,20 , have been put forth half-integer spin Heisenberg chains feature a massless spectrum,
to explain the absence of the long-range magnetic ordering in which enables proliferation of low-energy spinon excitations,
κ-(BEDT-TTF)2 Cu2 (CN)3 . Nevertheless, the origin for the QSL whereas such excitations are suppressed in the integer spin case,
state remains unresolved. which has a massive spectrum21 .
1 Department of Physics, Kyoto University, Kyoto 606-8502, Japan, 2 Institute for Materials Research, Tohoku University, Sendai 980-8577, Japan.
*e-mail: yamashitaminoru@scphys.kyoto-u.ac.jp.

0.14
A 10 T
0.12 Clean limit
0.10 B 0T 0.10
κ /T (W K –2 m –1)
κ /T (W K –2 m –1)
0.08
0.06
0.04 Dirty
A 0T
0.01
0.02
0 0 0.03 0.06 0.09

0 0.03 0.06 0.09
T 2(K2)
T 2(K2)
Figure 2 | Thermal conductivity in the low-temperature region. Thermal Figure 3 | Comparison between the data and the theory based on the
conductivity divided by temperature plotted as a function of T 2 below gapless QSL with a spinon Fermi surface. κ/T data (sample A) in zero
300 mK in zero field (blue for sample A and green for sample B) and at field (blue) plotted together with expected dependence of equation (1). The
µ0 H = 10 T (red, sample A) applied perpendicular to the basal plane. green line is for the clean limit (1/τ = 0) and brown for a dirty case with the
Convex and non-T 2 dependent κ/T is observed for both crystals. κ/T of mean free path as short as 10a, where a(' 0.8 nm) is the lattice parameter
sample A at 10 T shows a nearly parallel shift from that in zero field. It is of the triangular lattice.
immediately obvious that κ/T for all data vanishes as extrapolating to
T → 0 K, indicating the absence of the gapless fermionic excitations. This is suggested as a possible source of the anomaly at T ∗ and warrant
in sharp contrast to the specific-heat measurements, which claim the further studies.
presence of gapless excitations4 . The thermal conductivity at µ0 H = 0 and 10 T in the
low-temperature regime (T < 300 mK) is shown in Fig. 2. A
As it is not possible to directly probe the microscopic spin striking deviation of κ/T from a T 2 dependence is observed for
structure using neutron scattering owing to the compound’s both samples; both curves exhibit a convex trend. At such low
organic nature, thermodynamic measurements must be adopted to temperatures, the mean free path of phonons is as long as the
unveil the low-lying excitation of κ-(BEDT-TTF)2 Cu2 (CN)3 . Very crystal size and κph /T has a T 2 dependence, which has indeed
recent specific-heat measurements of κ-(BEDT-TTF)2 Cu2 (CN)3 been reported in a similar compound κ-(BEDT-TTF)2 Cu(NCS)2
show a large linear temperature-dependent contribution, (ref. 28). Therefore, the observed non-T 2 dependence, together
γ ∼ 15 mJ K−2 mol−1 (ref. 4), which suggests the presence of with the fact that κ is enhanced by magnetic field, definitely
gapless excitations, similar to the electronic specific heat in metals. indicates the substantial contribution of κmag in κ even in this
This observation provides strong support for several theoretical T range.
models, including a QSL with gapless ‘spinons’, which, like its The results shown in Fig. 2 provide key information on the
1D predecessors are (fermionic) elementary excitations that carry elementary excitations from the QSL state of κ-(BEDT-TTF)2 Cu2
spin-1/2 and zero charge2,3 , which are to be compared with (CN)3 . Most importantly, it is extremely improbable from the
conventional (bosonic) magnons that carry spin-1. However, it is experimental data that κ/T in the T → 0 K regime has a finite
premature to conclude that the QSL in κ-(BEDT-TTF)2 Cu2 (CN)3 residual value for data of both samples in zero field and that
is gapless from these measurements because the specific-heat data of sample A under 10 T. (Indeed, a simple extrapolation of both
are plagued by a very large nuclear Schottky contribution below 1 K data in zero field even gives a negative intersect.) These results
(ref. 4), which would necessarily lead to ambiguity. Incorporation of lead us to conclude that κ/T vanishes at T = 0 K. It should be
a probe that is free from such a contamination is strongly required22 . stressed that the vanishing κ/T immediately indicates the absence
As pointed out in ref. 3, thermal conductivity (κ) measurements of low-lying fermionic excitations, in sharp contrast to the finite γ
are highly advantageous as probes of elementary excitations in term reported in the heat capacity measurements4 . We believe that
QSLs, because κ is sensitive exclusively to itinerant excitations and the heat capacity measurements incorrectly suggest the presence of
is totally insensitive to localized entities such as are responsible gapless excitation, possibly owing to the large Schottky contribution
for Schottky contributions. The heat is carried primarily by at low temperatures.
acoustic phonons (κph ) and magnetic contributions (κmag ). Indeed, The present conclusion is reinforced by comparing the data with
a large magnetic contribution to the heat current is observed in the thermal conductivity calculated by assuming a spinon Fermi
low-dimensional spin systems23,24 . surface with gapless excitations3 , which is given as
As shown in Fig. 1, the thermal conductivity exhibits an
#−1
unusual behaviour characterized by a hump structure around
"
κ
2/3
~ kB T mA 1 1
T ∗ ' 6 K. A similar hump is observed in the magnetic part of the = 2 + 2 , (1)
specific heat4 and NMR relaxation rate1,10 around T ∗ , although T k B F kB τ d
no structural transition has been detected. These results obviously
indicate that κmag occupies a substantial portion in κ. Various where F is the Fermi energy, m is the electron mass, A is the
scenarios, such as a crossover to a QSL state4 , a phase transition unit cell area of the layer, d is the interlayer distance and τ is the
associated with the pairing of spinons2 , spin-chirality ordering25 , impurity scattering time. Estimating F = J as in 1D spin systems29 ,
Z2 vortex formation26 and exciton condensation27 , have been we compare our result with equation (1) as shown in Fig. 3. It is

10 ¬1 8 the spin gap, we try to fit the data to
κ mag(W K ¬1 m ¬1)
10 –1 6 κ = αexp(−∆/kB T ) + βT 3 ,
10 ¬2
4 as shown in an Arrhenius plot in Fig. 4. The best fit for the 0
n
(10) T data gives ∆ = 0.46 (0.38) K and a β-value that implies
2 that κph is roughly 1/4 of the total κ at 100 mK. We note that the
10 ¬3
κ (W K ¬1 m ¬1)
amplitude of ∆ is little affected by the choice of κph (see Fig. 4). As

0
0.1 0.2 0.3 the Arrhenius-type behaviour is observed in only one order range
T (K) of κ, the estimation of the gap size may have a large ambiguity.
10 –2
Nevertheless, we can safely conclude that the estimated gap value
from Fig. 4 is strikingly small compared with J (∆ ∼ J /500) and
insensitive to magnetic fields.
This field insensitivity is consistent with a theory of a gapped QSL
(ref. 7) with a finite energy gap for both magnetic and non-magnetic
excitations. On the other hand, the tiny gap value may alternatively
be attributed to a proximity to a quantum critical point of Z2 spin-
10 –3
2 4 6 8 10 12 14
liquid27 , or as a result of a slight anisotropy of J (ref. 19). However, at
1/T (K ¬1) present, the origin of the spin gap is an open question. It is tempting
to associate the extremely small gap value with kB T ∗ ( J ) (instead
Figure 4 | An Arrhenius plot of the thermal conductivity in the of to J itself), which may be a characteristic temperature of the QSL
low-temperature region. Inset: logκmag (sample A) in zero field plotted of κ-(BEDT-TTF)2 Cu2 (CN)3 . In any case, our low-temperature
against logT estimated for several values of κph (= 0 (blue), 1/2 (green) thermal-transport measurements demonstrate that the fermionic
and 3/4 (orange) of κ at 100 mK) assuming a cubic temperature spinons, if present, will experience an instability in this system,
dependence of κph . In the same figure, the exponent n = (dlogκ/dlogT) which will generate a small gap in the spin excitation spectrum.
obtained by assuming a power-law dependence of κ ∝ T n is also plotted
Methods
(κph = 0 (blue), 1/2 (green), 3/4 (orange) of κ at 100 mK). Main panel: An
κ-(BEDT-TTF)2 Cu2 (CN)3 single crystals were grown by the electrochemical
Arrhenius plot of κ of sample A (κ ∝ exp(−∆/kB T)) in zero field and at method. The thermal conductivity was measured by a standard steady-state
µ0 H = 10 T. Activation-type behaviour can be seen for both data. The method with a one-heater–two-thermometer configuration in 3 He and dilution
dashed–dotted and dotted lines are fits to equation (1) with refrigerators. The thermal current was applied within the 2D plane and the
α = 0.12 W K−1 m−1 , β = 0.51 W K−4 m−1 and ∆ = 0.46 K for 0 T, and with magnetic field was applied perpendicular to the plane. We have measured several
deuterated and non-deuterated crystals and observed no significant sample
α = 0.081 W K−1 m−1 , β = 0.64 W K−4 m−1 and ∆ = 0.38 K for 10 T, dependence. It has been reported that in superconductors, thermal decoupling
respectively. The green and orange show the same data as in the inset and between the electron and phonon conductions can be caused by the poor contacts
the fittings for αe−∆/kB T give α = 0.087 W K−1 m−1 , ∆ = 0.45 K and at very low temperatures30 . It could be argued that such a decoupling may occur
α = 0.048 W K−1 m−1 , ∆ = 0.43 K, respectively. in the phonons and the spinons, and may lead to apparent absence of finite κ/T
at T → 0 K. However, we note that this is inconsistent with the observed increase
of κ with H (shown in Supplementary Information). Because the magnetic field
evident that equation (1) yields κ/T that increases with decreasing decreases the number of spinons, κ should decrease with H owing to the further
T for both clean and dirty cases and is opposite to the observation. reduction of the coupling. Moreover, we measured the thermal conductivity on the
Moreover, to obtain the same magnitude of κ/T in this model at samples with the contact resistance ranging from 1 to 20 and found no serious
the lowest temperature, we need to assume that the mean free path difference in the thermal conductivity at low temperatures.
is only a few times longer than the lattice constant a. However,
such a large concentration of the impurity is highly unlikely in this
Received 19 May 2008; accepted 13 October 2008;
clean system1 . Thus, the theory based on a gapless fermionic spinon
published online 23 November 2008
picture is incompatible with the present results, although it may be References
applicable to other systems. 1. Shimizu, Y., Miyagawa, K., Kanoda, K., Maesato, M. & Saito, G. Spin liquid
Having established the absence of the low-lying fermionic state in an organic Mott insulator with a triangular lattice. Phys. Rev. Lett. 91,
excitations, we turn to a more detailed analysis of the T dependence 107001 (2003).
of κ. As seen in the inset of Fig. 4, where logκmag is plotted against 2. Lee, S.-S., Lee, P. A. & Senthil, T. Amperean pairing instability in the U(1) spin
liquid state with fermi surface and application to κ-(BEDT-TTF)2 Cu2 (CN)3 .
logT , κmag does not show a power-law dependence on T . As the Phys. Rev. Lett. 98, 067006 (2007).
precise value of κph is unknown, κmag (= κ − κph ) is estimated for 3. Nave, C. P. & Lee, P. A. Transport properties of a spinon Fermi surface
several values of κph . For each case, no linear relation is observed in coupled to a U(1) gauge field. Phys. Rev. B 76, 235124 (2007).
this log–log plot. It should be noted that when κph is increased, the 4. Yamashita, S. et al. Thermodynamic properties of a spin-1/2 spin-liquid state
nonlinearity becomes more pronounced. This is also manifested by in a κ-type organic salt. Nature Phys. 4, 459–462 (2008).
5. Anderson, P.W. Resonating valence bonds: A new kind of insulator? Mater.
the index n = (dlogκ/dlogT ) plotted in the same figure where n Res. Bull. 8, 153–160 (1973).
increases steeply with decreasing temperature, and subtracting κph 6. Kalmeyer, V & Laughlin, R. B. Equivalence of the resonating-valence-bond
from the observed κ even enhances the non-power-law behaviour. and fractional quantum hall states. Phys. Rev. Lett. 59, 2095–2098 (1987).
Thus, in spite of the ambiguity for estimating κph , we can safely 7. Moessner, R. & Sondhi, S. L. Resonating valence bond phase in the triangular
conclude that κmag at low temperatures does not exhibit a power-law lattice quantum dimer model. Phys. Rev. Lett. 86, 1881–1884 (2001).
8. Anderson, P.W. The resonating valence bond state in La2 CuO4 and
temperature dependence. These results place further constraints on superconductivity. Science 235, 1196–1198 (1987).
the theoretical description of the excitation spectrum; for example, 9. Powell, B. J. & McKenzie, R. H. Symmetry of the superconducting order
the nodal excitations that may be expected in systems with an parameter in frustrated systems determined by the spatial anisotropy of spin
anisotropic gap structure2 which will give rise to a power-law correlations. Phys. Rev. Lett. 98, 027005 (2007).
dependence of κ on T , in analogy to the quasiparticles in d-wave 10. Shimizu, Y., Miyagawa, K., Kanoda, K., Maesato, M. & Saito, G. Emergence of
inhomogeneous moments from spin liquid in the triangular-lattice Mott
superconductors, are also absent. insulator κ-(ET)2 Cu2 (CN)3 . Phys. Rev. B 73, 140407(R) (2006).
The absence of the gapless excitation implies the presence of a 11. Zheng, W., Singh, R. R. P., McKenzie, R. H. & Coldea, R. Temperature
spin gap in the excitation spectrum. To estimate the magnitude of dependence of the magnetic susceptibility for triangular-lattice

antiferromagnets with spatially anisotropic exchange constants. Phys. Rev. B 25. Baskaran, G. Novel local symmetries and chiral-symmetry-broken phases in
71, 134422 (2005). S = 12 triangular-lattice Heisenberg model. Phys. Rev. Lett. 63,
12. Itou, T., Oyamada, A., Maegawa, S., Tamura, M. & Kato, R. Quantum spin 2524–2527 (1989).
liquid in the spin-1/2 triangular antiferromagnet EtMe3 Sb[Pd(dmit)2 ]2 . Phys. 26. Kawamura, H. & Miyashita, S. Phase transition of the two-dimensional
Rev. B 77, 104413 (2008). Heisenberg antiferromagnet on the triangular lattice. J. Phys. Soc. Jpn. 53,
13. Masutomi, R., Karaki, Y. & Ishimoto, H. Gapless spin liquid behavior in 4138–4154 (1984).
two-dimensional solid 3 He. Phys. Rev. Lett. 92, 025301 (2004). 27. Qi, Y. & Sachdev, S. Insulator–metal transition on the triangular lattice. Phys.
14. Waldtmann, Ch. et al. First excitations of the spin 1/2 Heisenberg Rev. B 77, 165112 (2008).
antiferromagnet on the kagome lattice. Eur. Phys. J. B 2, 501–507 (1998). 28. Belin, S., Behnia, K. & Deluzet, A. Heat conduction in
15. Huse, D. A. & Elser, V. Simple variational wave functions for κ-(BEDT-TTF)2 Cu(NCS)2 . Phys. Rev. Lett. 81, 4728–4731 (1998).
two-dimensional Heisenberg spin- 12 antiferromagnets. Phys. Rev. Lett. 60, 29. Luther, A. & Peschel, I. Calculation of critical exponents in two dimensions
2531–2534 (1988). from quantum field theory in one-dimension. Phys. Rev. B 12,
16. Capriotti, L., Trumper, A. E. & Sorella, S. Long-range Néel order in the 3908–3917 (1975).
triangular Heisenberg model. Phys. Rev. Lett. 82, 3899–3902 (1999). 30. Smith, M. F., Paglione, J., Walker, M. B. & Taillefer, L. Origin of anomalous
17. Morita, H., Watanabe, S. & Imada, M. Nonmagnetic insulating states near the low-temperature downturns in the thermal conductivity of cuprates. Phys.
Mott transitions on lattices with geometrical frustration and implications for Rev. B 71, 014506 (2005).
κ-(ET)2 Cu2 (CN)3 . J. Phys. Soc. Jpn. 71, 2109–2112 (2002).
18. Misguich, G., Lhuillier, C., Bernu, B. & Waldtmann, C. Spin-liquid phase of
the multiple-spin exchange Hamiltonian on the triangular lattice. Phys. Rev. B
Acknowledgements
60, 1064–1074 (1999). We would like to thank G. Baskaran, L. N. Bulaevskii, Y. Hayashi, N. Kawakami,
19. Yunoki, S. & Sorella, S. Two spin liquid phases in the spatially anisotropic H. Kawamura, Y. Nakazawa, S. Sachdev, A. Tanaka and S. Watanabe for valuable
triangular Heisenberg model. Phys. Rev. B 74, 014408 (2006). discussion. This work was supported by Grants-in-Aid (No. 20224008 and No.
20. Hayashi, Y. & Ogata, M. Possibility of gapless spin liquid state by 20840026) from JSPS and a Grant-in-Aid for the Global COE Program ‘The Next
one-dimensionalization. J. Phys. Soc. Jpn. 76, 053705 (2007). Generation of Physics, Spun from Universality and Emergence’ from MEXT of Japan.
21. Haldane, F. D. M. Nonlinear field theory of large-spin Heisenberg
antiferromagnets: Semiclassically quantized solitons of the one-dimensional Author contributions
easy-axis Néel state. Phys. Rev. Lett. 50, 1153–1156 (1983). M.Y., N.N., Y.K., T. Shibauchi and Y.M. carried out measurements, data analysis and
22. Ramirez, A. P. Quantum spin liquids: A flood or a trickle?. Nature Phys. 4, discussion. T. Sasaki, N.Y. and N.K. prepared the samples. S.F. gave theoretical advice.
442–443 (2008).
23. Sales, B. C., Lumsden, M. D., Nagler, S. E., Mandrus, D. & Jin, R. Magnetic
field enhancement of heat transport in the 2D Heisenberg antiferromagnet Additional information
K2 V3 O8 . Phys. Rev. Lett. 88, 095901 (2002). Supplementary Information accompanies this paper on www.nature.com/naturephysics.
24. Li, S. Y., Taillefer, L., Wang, C. H. & Chen, X. H. Ballistic magnon transport Reprints and permissions information is available online at http://npg.nature.com/
and phonon scattering in the antiferromagnet Nd2 CuO4 . Phys. Rev. Lett. 95, reprintsandpermissions. Correspondence and requests for materials should be addressed
156603 (2005). to M.Y.

ARTICLES
Superconducting nanocircuits for topologically

protected qubits
Sergey Gladchenko1 , David Olaya1 , Eva Dupont-Ferrier1 , Benoit Douçot2 , Lev B. Ioffe1
and Michael E. Gershenson1 *
For successful realization of a quantum computer, its building blocks—the individual qubits—should be simultaneously
scalable and sufficiently protected from environmental noise. Recently, a novel approach to the protection of superconducting
qubits has been proposed. The idea is to prevent errors at the hardware level, by building a fault-free logical qubit from
‘faulty’ physical qubits with properly engineered interactions between them. The decoupling of such a topologically protected
logical qubit from local noises is expected to grow exponentially with the number of physical qubits. Here, we report on
proof-of-concept experiments with a prototype device that consists of twelve physical qubits made of nanoscale Josephson
junctions. We observed that owing to properly tuned quantum fluctuations, this qubit is protected against magnetic flux
variations well beyond linear order, in agreement with theoretical predictions. These results suggest that topologically
protected superconducting qubits are feasible.
Design of the protected qubit
F
or implementation of quantum correction codes with realistic
redundancy, the decoherence time of a qubit, τd , should be at The prototype device studied here involves three levels of
least 104 –105 times longer than the time of a single operation, organization. Level 1 is constituted by the ‘faulty’ qubits shown as
τ0 (refs 1,2). Unfortunately, all existing scalable qubits fall short rhombi in Fig. 1. These Josephson elements with an effective energy
of this criterion. For instance, at present, all superconducting
qubits demonstrate error rates of ε ∼ 10−2 or larger3–6 . The main VR = E2R cos(2φ)
reason for a relatively short decoherence time of superconducting
qubits is their strong interaction with uncontrollable degrees of are implemented as superconducting loops interrupted by four
freedom in their environment, that is, ‘noises’ such as critical Josephson junctions and threaded by the magnetic flux ΦR = Φ0 /2
current fluctuations, charge noise, flux noise and quasiparticle (refs 21–23). Here, φ = πΦR /Φ0 is the phase difference across
poisoning7,8 . Recent experimental and theoretical studies indicate each element; Φ0 is the superconducting flux quantum. An
that these noises are local, that is, they can be represented as a individual Josephson junction is characterized by the Josephson
sum of operators acting on individual qubits. Because of the local energy EJ = hI c /2e and the charging energy EC = e 2 /2CJ ; here,
nature of noises, the task of large-scale quantum computation is Ic and CJ are the critical current of the Josephson junction
considered to be realistic9,10 . Several approaches to the elimination and its capacitance, respectively24 . If EC is negligible, each
of noise-induced decoherence have been proposed so far; these rhombus has two degenerate √ classical states separated by the
approaches are based on improving materials used for qubit energy barrier 2E2R ≈ 4( 2 − 1)EJ (see Fig. 1a). These states
fabrication11,12 and decoupling of a qubit from local noises. In correspond to the phase difference φ = ±π/2 across the rhombus
particular, the experiments13–16 have shown that the qubit–noise (the supercurrents circulate in a rhombus either clockwise
coupling can be reduced by tuning the qubit control parameters. or anticlockwise). A non-zero (but small) charging energy
However, this tuning results only in a linear-order decoupling, causes rare tunnelling events between these states and removes
which is insufficient for large-scale quantum computations. the degeneracy, resulting in √ the symmetric and antisymmetric
Recently, an alternative approach to qubit protection has been superpositions: √ |+iR = (1/ 2)(|φ = −π/2i + |φ = π/2i) and
proposed17–20 . This approach is based on topological protection; |−iR = (1/ 2)(|φ = −π/2i − |φ = π/2i). These low-energy
it can be viewed as a hardware implementation of the software states of a rhombus√ are separated by a small energy gap
error correction protocols. The idea is to build a fault-free logical t ≈ EJ3/4 EC1/4 exp(−1.6 EJ /EC ). A single rhombus is not protected
qubit from ‘faulty’ physical qubits with carefully tuned interaction against local noises. Noise deforms the potential VR (φ) and
between them. Such interactions result in a highly entangled induces fluctuations of 2E1R , the energy difference between states
collective state (similar to the one used in error correction schemes) |φ = π/2i and |φ = −π/2i, which leads to dephasing. The deformed
protected by non-trivial symmetries from all local noises. The potential can be represented as VR (φ) ≈ E2R cos(2φ) + E1R sin(φ),
advantage of the topological protection is that it eliminates the need where the second term describes both the effect of noise (the
for high-fidelity ancilla-state preparation and measurements at the time-dependent part of E1R ) and an asymmetry of the rhombus
lowest level of error correction, thereby decreasing enormously (the time-independent part of E1R ).
the required redundancy. Here, we describe experiments with To suppress dephasing, the ‘cos(2φ)’ Josephson elements are
prototypes of topologically protected qubits that demonstrate the combined in a chain at level 2. The simplest two-element chain
viability of this approach. (Fig. 1b) connects a superconducting island to a current lead, the
1 Department of Physics and Astronomy, Rutgers University, 136 Frelinghuysen Road, Piscataway, New Jersey 08854, USA, 2 Laboratoire de Physique
Théorique et Hautes Energies, CNRS UMR 7589, Universités Paris 6 et 7, 4 place Jussieu, 75005 Paris, France. *e-mail: gersh@physics.rutgers.edu.

NATURE PHYSICS DOI: 10.1038/NPHYS1151 ARTICLES
a ΦR because the energy barrier between the states, 2E2 , prevents a
φ =π VR(φ)
Φ0 single rhombus from flipping its phase by π. However, a pair
of rhombi can flip simultaneously with the tunnelling amplitude
I¬〉R t2R . Remarkably, these flips induced by quantum fluctuations
Φ R = Φ0/2 = 2E2R
t help suppress dephasing. Indeed, let us assume that the local
I+〉R noise changes the energy difference between states |φ = π/2i and
¬π/2 π/2 φ |φ = −π/2i of the first and second rhombi by δE1 and δE2 ,
respectively. Because the states of individual rhombi enter the
b V(ϕ ) logical states symmetrically, the noise does not affect the energy
of these states in the first order of the perturbation theory. The
ϕA = 0 ϕ Δ3 I¬〉 Δ2 difference in energies of the logical states appears only in the second
2E2
δ order, δE01 = (δE1 δE2 /∆2 ), where ∆2 = 2t2R is the gap that separates
C→∞ C0 I+〉 the logical states from the first excited state (see Fig. 1b). Thus, the
0 π ϕ = ∑ φi chain is protected from dephasing provided ∆2 is sufficiently large.
i
In a longer chain of N physical qubits, the effect of a local noise is
c expected to be suppressed up to the N th order of the perturbation
theory18,20 . For example, the effective magnetic flux, which affects
the logical states of the N -element chain, is greatly suppressed with
ϕA = 0 ϕB = ϕ
respect to the flux noise δ Φi = Φi − Φ0 /2 (refs 18,20):
N −1
δ Φ i EJ
Y
δ Φeff = δ Φi . (1)
C→∞ i
Φ0 ∆i
C0
Here, ∆i is the excitation energy associated with changing the num-
Figure 1 | Protected qubit based on ‘cos(2φ)’ Josephson elements. ber of Cooper pairs by one on the superconducting ‘island’ shared
a, The building blocks of the protected qubit are ‘cos(2φ)’ Josephson by the ith and (i+1)th rhombi. The minimum value of ∆i = ∆2 cor-
elements (rhombi) implemented as superconducting loops interrupted by responds to the excitation energy of the central island (this central
four Josephson junctions (red crosses) and threaded by the magnetic flux island is shared among three chains in Fig. 1c). Note that δ Φi = 0 in
ΦR = Φ0 /2. The Josephson energy of this ‘faulty’ physical qubit, a single rhombus results in δ Φeff = 0 and, thus, the cos(2ϕ)-periodic
VR = E2R cos2φ, is doubly periodic in the phase difference φ across the energy of the whole chain. Equation (1) shows that for an efficient
rhombus. The black dashed lines show VR (φ) in the classical limit noise decoupling, the energy gap ∆2 between the logical states
(EC → 0); quantum fluctuations ‘smear’ the cusps at φ = nπ and result in and the first excited state of the qubit should be large (this is
tunnelling between the states |φ = π/2i and |φ = −π/2i. b, The simplest also required for decreasing the qubit operation time τ0 > ~/∆2 ).
protected qubit is a chain of two cos(2φ) elements, which connects the The decoupling from other types of local noise is described by a
rightmost ‘island’ to a large superconducting lead. Two degenerate similar formula with EJ being replaced by a relevant energy scale
classical states at ϕ − ϕA = 0, π in the effective potential of the chain, (for example, EC for the charge noise). This decoupling is expected
V(ϕ) = −E2 cos2ϕ, are shown by solid red lines. A finite probability of to be even more efficient because these energies are smaller than EJ .
tunnelling between the states |ϕ = 0i and |ϕ = πi removes the degeneracy The protection is maximized at level 3 by choosing an optimal
√ in a small energy splitting δ between the logical states of the
and results interconnection between the elements and tuning the parameters
qubit, (1/ 2)(|0i ± |πi). The first and second excited levels are separated of individual Josephson junctions. To realize large values of both
from this (almost) degenerate doublet by gaps ∆2 and ∆3 , respectively. E2 and 12 , several conflicting requirements should be reconciled.
c, Connection of several rhombi chains in parallel increases the depth of the Indeed, an increase of the number of rhombi in a chain improves
effective potential V(ϕ) and suppresses the transitions between the qubit’s the protection from noise. However, the energy barrier, which has
logical states. to be high for protection, diminishes with N : 2E2 ≈ π2 E2R /2N
in the quasiclassical limit (C0 → ∞) (ref. 25). This conflict can
phase of which is fixed owing to a large capacitance C0 (we choose be resolved by parallel connection of several chains (see Fig. 1c).
ϕA = 0 in Fig. 1b). The logical variableP
is the superconducting phase Similarly, quantum fluctuations help to establish a global coherent
difference across the chain, ϕ − ϕA = i φi , which is the sum of the state across the chain, but, if excessively strong, they suppress the
phase differences across individual rhombi. For an even number Josephson energy barrier between the logical states of a chain.
of rhombi in a chain, ϕ is either 0 or π. The double periodicity of Numerical simulations (see Supplementary Information) show that
the energies of rhombi leads to the double periodicity of the chain for each qubit architecture (M chains of N rhombi each), an
energy V (ϕ) = −E2 cos(2ϕ), where E2 = ~I2 /4e is determined by the optimal protection is realized for a certain value of the ratio EJ /EC
critical current of the chain, I2 . that controls the strength of quantum fluctuations. For example, for
Quantum information is encoded in the (almost) degenerate the tested design (M = 3, N = 4), sizable values of both E2 and ∆2
ground states of the chain. For a two-rhombi chain, these states in can be realized for 3 < EJ /EC < 5.
the quasiclassical limit (C0 → 0) can be represented as
Protection testing: the idea of the experiment
1 The experiments described below were designed to test the key
|ϕ = 0i = √ (|φ = −π/2i|φ = π/2i + |φ = π/2i|φ = −π/2i),
2 theoretical prediction (1), namely the decoupling from local noises
well beyond the linear order. This test exploits the fact that
1 equation (1) describes the decoupling of a qubit from both time-
|ϕ = πi = √ (|φ = −π/2i|φ = −π/2i + |φ = π/2i|φ = π/2i). dependent flux noises and time-independent flux variations. This
2
enabled us to experimentally test the protection in our arrays by
The non-local logical states are protected from both relaxation studying their response to the static deviations of magnetic flux
and dephasing induced by local noises. The energy relaxation from its optimal value Φ0 /2. For a single rhombus, the amplitude
(transitions between states |ϕ = 0i and |ϕ = πi) is suppressed of the first harmonic, E1R , in the expression for its potential

ARTICLES NATURE PHYSICS DOI: 10.1038/NPHYS1151
a b
c
L
2 μm
ΦR
ϕB d
Cg ΦL CID Cg
ΦL
L
ϕA
CID
L 20 μm
Figure 2 | The prototype of a protected superconducting qubit. a–d, Schematic design (a,c) and micrographs (b,d) of the device. Josephson junctions are
formed at intersections of aluminium strips (the tunnel barrier is shown in blue in a). b shows micrographs of the array: three chains of rhombi connected
in parallel to increase E2 . To probe V(ϕAB ), the array is included in one ‘arm’ of a superconducting loop; another arm of the loop contains two larger
Josephson junctions (c and d). The magnetic flux ΦR through each rhombus of an area of 1 µm2 controls the effective Josephson energy of the rhombi. The
SQUID-like device is protected from external high-frequency noise and non-equilibrium quasiparticles generated outside the device by two meander-type
inductances L(∼ 2 × 10−8 H). To ensure the ‘classical’ behaviour of larger Josephson junctions, the SQUID-like device is shunted by an inter-digitized
capacitor CID ∼ 3 × 10−14 F. The ends of the inductors are connected to the external circuit, which is used for the generation of current pulses.
Table 1 | Parameters of the studied devices.
Device Area (µm2 ) RN (k) EC (K) EJ /EC
1 0.165 × 0.165 4.78 0.68 2.2

2 0.153 × 0.153 3.27 0.79 2.7
3 0.150 × 0.180 2.82 0.69 3.7
4 0.173 × 0.173 2.43 0.62 4.7
5 0.180 × 0.180 2.49 0.57 5.0
The values of RN have been extracted from the total resistance of devices driven in the normal state by the current I > ISW . The Josephson energy EJ for individual Josephson junctions has been determined
from RN using the Ambegaokar–Baratoff formula; the Coulomb energy EC was estimated from the area of the junctions (CJ = (area) × 50 fF µm−2 ; refs 24,31).
VR (φ) ≈ E2R cos(2φ)+E1R sin(φ) is proportional to δ Φ = Φ − Φ0 /2. the chains can be significantly varied without affecting the phase
Topological protection of a rhombi array would imply that the difference across individual rhombi.
ratio E1 /E2 in the expression for its energy, V (ϕ) ≈ −E2 cos(2ϕ) −
E1 cos(ϕ), is exponentially reduced in comparison with that for Device response to static perturbations
a single rhombus: E1 ∼ (δ Φ )N . The experiment was designed to The V (ϕAB ) dependence of this underdamped Josephson device was
measure the dependence E1 (δ Φ ) and to demonstrate that with studied by measuring the probability of switching of the device into
proper tuning of quantum fluctuations, even a relatively small the resistive state by 1-ms-long current pulses for different values
rhombi array is protected well beyond linear order. of the external magnetic field B (for details, see Supplementary
The tested superconducting quantum interference device Information). Figure 3 shows the dependence of the switching
(SQUID)-like circuit (see Fig. 2) included the rhombi array and current ISW (defined as the pulse amplitude that switches the device
two larger (0.3 µm × 0.3 µm) Josephson junctions. The details of into the resistive state with probability 0.5) on the magnetic field
device design and fabrication are provided in the Methods section for a device with the ratio EJ /EC = 2.7. The range of the magnetic
and Supplementary Information. We have tested several devices field in Fig. 3a corresponds to the magnetic flux through a single
(see Table 1) with the normal-state resistance RN = 2.4–5 k and rhombus, ΦR , varying between −Φ0 /2 and Φ0 /2. The switching
the ratio EJ /EC = 2–5 for the individual Al–Al2 O3 –Al Josephson current oscillates with the magnetic flux ΦL through the loop of
junctions in the rhombi. The phase difference ϕAB across the array the SQUID-type device with an area of ∼110 µm2 . The period
was controlled by varying the magnetic flux ΦL in the SQUID loop. of oscillations, 1ΦL , depends on ΦR : 1ΦL = Φ0 for all values
Because of a large difference between the areas of individual rhombi of ΦR except for ΦR ≈ (n + 1/2)Φ0 , where the period is halved
(1 µm2 ) and the SQUID loop (110 µm2 ), the phase difference across (1ΦL = Φ0 /2). Note that in the quasiclassical case (EJ /EC > 20),

a 280 b 280
240 240
Φ R = ¬Φ0 /2 Φ R = Φ0 /2
Isw (nA)
Isw (nA)
200 200
160 160
¬10 ¬5 0 5 10 ¬10 ¬5 0 5 10
B (G) B (G)
c d
200 200
150
Isw (nA)
100
Isw (nA)
100
Φ R = Φ0 /2 Φ R = Φ0 /2
50
0
¬100
6 7 8 9 10 3 4 5 6 7
B (G) B (G)
Figure 3 | Coherent transport of pairs of Cooper pairs. a, Oscillations of the switching current ISW (B), measured for device 2 at T = 50 mK, for the
magnetic field corresponding to the flux through a single rhombus, ΦR , ranging from −Φ0 /2 to Φ0 /2. b, Simulation of the first harmonic of ISW oscillations
(with a period of 1ΦL = Φ0 ). The calculations neglect quantum fluctuations and assume that the EJ values for individual Josephson junctions are randomly
scattered within a ±20% interval. c, Comparison of the data and simulations in the protected regime (ΦR ≈ Φ0 /2). The red curve, which represents the
experimental data, shows that ISW oscillates with the period 1ΦL = Φ0 /2 when ΦR ≈ Φ0 /2. The green curve (shifted down by 70 nA for clarity) shows the
first harmonic of ISW oscillations (1ΦL = Φ0 ), which is strongly suppressed over a relatively wide range of magnetic fields around ΦR = Φ0 /2. The
calculated first harmonic is shown as the blue curve shifted down by 170 nA. d, The experimental data (red) and the first harmonic of oscillations (green,
shifted down by 130 nA) for a single two-rhombi chain.
such a period halving has recently been observed26 . The beatings where ∆2 and ∆3 are the energy gaps that separate two almost
of oscillations with 1ΦL = Φ0 , observed at ΦR /Φ0 ≈ ±1/8, ±1/4 degenerate low-energy levels from the first and second excited
and ±3/8, are due to the flux quantization in the intermediate-size levels, respectively (see Fig. 1b). Equation (2) translates into
loops between adjacent rhombi chains (the area of these loops is E1 ≈ (40δ ΦR /Φ0 )4 EJ for the array shown in Fig. 3c. Thus, the
four times greater than the area of a single rhombus). Below, we theory predicts that the term cos(ϕAB ) should be suppressed over
focus on the regime ΦR ≈ (±1/2)Φ0 , where the states |φ = π/2i the range δ ΦR = ±0.025Φ0 , in agreement with the experimental
and |φ = −π/2i of individual rhombi are almost degenerate and the data. Observation of the cos(2ϕAB ) oscillations of ISW together
rhombi array is expected to be protected from noise. with the vanishing cos(ϕAB ) oscillations provides an essential
The oscillations of ISW with the period 1ΦL = Φ0 vanish near test for the proper strength of quantum fluctuations. These
ΦR = ±Φ0 /2 (Fig. 3c). In this regime, the effective Josephson experimental results also indicate that the scattering of Josephson
energy of a rhombus, VR = E2R cos(2φ), becomes small and junction parameters in the studied devices is relatively small:
the supercurrent of single Cooper pairs is blocked by quantum only the (axially) symmetric rhombi contribute to the order of
fluctuations. Comparison between the data obtained for the 4 × 3 protection, N .
rhombi array (Fig. 3c) and a single two-rhombi chain (Fig. 3d) The height of the energy barrier that separates the states of the
shows that the oscillations with the period 1ΦL = Φ0 and, thus, the rhombi array, 2E2 , can be found from the amplitude I2 = 4eE2 /~
term E1 cos(ϕ) in the device energy, are suppressed for the array of the oscillations of switching current with period ∆ΦL = Φ0 /2.
over a much wider range of magnetic fields. This observation is Figure 4 shows that the values of 2E2 /EC measured for devices
in good agreement with equation (1). Indeed, if the suppression is with different values of EJ /EC are in good agreement with our
due to quantum fluctuation, the term cos(ϕAB ) responsible for the numerical simulations (for simulation details, see Supplementary
oscillations with the period 1ΦL = Φ0 should appear in the energy Information). The observed agreement verifies the validity of
of a four-rhombi chain only in the fourth order in flux deviations theoretical assumptions, which have also been used in the
from the optimal values ΦR = (n+1/2)Φ0 . The direct computation calculations of ∆2 (see Fig. 4). The gap ∆2 is the smallest of two
of the effective Josephson coupling for the whole array in the fourth gaps: the gap 2t2R ∼ t 2 /EJ is associated with the excitations
√ inside
order in the perturbation theory gives the chain with ϕ being fixed, whereas the other gap (∼ 32E2 EC )
is associated with the fluctuations of ϕ around its classical value.
δ ΦR
4
1 In the quasiclassical case (EJ /EC 1), ∆2 coincides with 2t2R
E1 ≈ 8.6 EJ , (2)
Φ0 ∆2 ∆23 (ref. 19). However, the realization of a sizable ∆2 requires not-

1.2 with the gate voltage Vg . The amplitude of oscillations, 1ISW ,

in the protected regime ΦR ≈ Φ0 /2 is in good agreement with
0.6 our calculations of the dependence 1I2 (EJ /EC ) (note that no
fitting parameters are involved in this comparison). The period
0.8 of oscillations, which corresponds to charging of the central
‘island’ with charge 2e, is approximately the same for both
0.4
ΦR = 0 (Fig. 5a) and ΦR = Φ0 /2 (Fig. 5b). This is expected for
2E2 /EC
Δ2 (K)
relatively long (1 ms) current pulses used in these measurements:
although the transport of single Cooper pairs is suppressed by
0.4
0.2 quantum fluctuations in this regime, there is still a considerable
probability of tunnelling of a Cooper pair to/from the island over
a long timescale.
0 0 Discussion and summary
2 3 4 5 6 The reported data suggest that the topological protection can be
EJ /EC realized in a Josephson circuit with a properly tuned ratio EJ /EC .
Our results indicate that (1) the scattering of Josephson junction
Figure 4 | Characteristic energies 2E2 and ∆2 for the devices with parameters and individual rhombi areas can be made sufficiently
different values of EJ /EC . The experimental points show the potential small for the realization of symmetry-protected superconducting
barrier 2E2 between the states |0i and |πi of rhombi arrays calculated from qubits, and (2) our theoretical model captures all of the essential
the measured amplitude of the 1ΦL = Φ0 /2 oscillations of switching features of real devices. Further reduction of the Josephson junction
current, I2 = 4eE2 /h̄ (the parameters of individual Josephson junctions in dimensions (larger EC ), with a simultaneous increase of the
the studied devices are listed in Table 1). The horizontal error bars transparency of the tunnel barrier (larger EJ ), will enable an increase
correspond to the uncertainty in EJ /EC estimated from the scattering of of the operational temperature of protected qubits.
areas of individual Josephson junctions. The dependencies of 2E2 /EC and Vanishing of the ISW oscillations with the period 1ΦL = Φ0 in the
∆2 on EJ /EC (green and blue curves, respectively) show the results of protected regime suggests that the supercurrent of single Cooper
numerical calculations for a 4 × 3 rhombi array (for details, see pairs, I1 ≡ 2eE1 /~, is blocked by quantum fluctuations. In other
Supplementary Information). Note that no fitting parameters are involved words, single Cooper pairs become localized in the array when the
in the comparison between the calculations and experimental data. effective Josephson energy of a rhombus, VR = E2R cos(2πΦR /Φ0 ),
becomes small. Note that in the charge basis, two states of a
too-large values of EJ /EC and, thus, numerical simulations (see logical qubit differ by the parity of the number of Cooper pairs.
Supplementary Information). Figure 4 shows that large values of Observation of the oscillations with the period 1ΦL = Φ0 /2 in
both E2 and ∆2 required for protection can be realized for the the protected regime indicates that the supercurrent is carried by
studied qubit design within the range EJ /EC ∼ 3–5. correlated pairs of Cooper pairs with charge 4e (refs 18–20). To the
Another probe of quantum fluctuations in the studied array best of our knowledge, this is the first observation of the coherent
is provided by the measurements of the effect of the gate voltage transport of pairs of Cooper pairs in a small-size rhombi array in
on the switching current. In the absence of quantum fluctuations, the quantum regime. The persistence of this phenomenon in larger
the critical current of the device coincides with the critical current arrays would imply the appearance of a new thermodynamic phase
of three rhombi chains connecting the central strip to one of characterized by hcos2ϕi 6= 0 in the absence of conventional Cooper
the leads. Quantum fluctuations, which result in tunnelling of pair coherence. This phase can be regarded as ‘superconductor
the phase of the central strip between 0 and π, reduce ISW . nematic’; the global state of the whole array in this phase is
The offset charge 1q = Vg /Cg induced by the gate modulates characterized by the Z2 topological order parameter (which takes
the phase of the tunnelling amplitudes tisl → tisl exp(i1q1φ), values either +1 or −1) (refs 18–20).
affecting the interference of processes with 1ϕ = ±π. Thus, The rhombi array studied here can be used as a qubit protected
in the presence of quantum fluctuations, we might expect to in its ‘idle’ state from all local noises in the fourth order if one
observe modulation of ISW by Vg . Indeed, Fig. 5a,b shows that for of the leads is replaced with an island disconnected from the
the device with EJ /EC = 4.7, the switching probability oscillates outside world. To carry out a generic gate operation, the qubit
a b c
ΦR = 0 400 Φ R = Φ0 /2
0.8
1.5
0.6
240
0.7
390
Probability
Probability
ΔIsw (nA)
Φ R = Φ0 /2
Isw (nA)
Isw (nA)
0.5 1.0
0.6
238
380 0.4 0.5
0.5
370 0.3 236
0.4 0
0 0.2 0.4 0.6 0.8 1.0 0 0.2 0.4 0.6 0.8 1.0 0 2 4 6 8
Vg (V) Vg (V) EJ /EC
Figure 5 | Gate voltage dependence of the switching current. a,b, The probability of switching into the resistive state as a function of the gate voltage for
the device with EJ /EC ∼ 4.7, measured with a fixed amplitude of current pulses. This probability can be directly translated into the value of the switching
current shown on the right vertical axes. c, The amplitude of modulation of the switching current, 1ISW , measured for two devices with different values of
EJ /EC in the regime ΦR = Φ/2. The vertical error bars reflect the uncertainty in ISW measurements; the horizontal error bars correspond to the uncertainty
in EJ /EC estimated from the scattering of areas of individual Josephson junctions. The experimental data are in good agreement with the result of the
numerical calculation (the blue curve); no fitting parameters are involved in this comparison.

protection has to be temporarily removed. This will result in a 6. Ithier, G. et al. Decoherence in a superconducting quantum bit circuit.
faster decoherence, which is equivalent to a loss of the operation Phys. Rev. B 72, 134519 (2005).
precision. Note, however, that the time required for a single 7. Faoro, L. & Ioffe, L. B. Quantum two level systems and Kondo-like traps as
possible sources of decoherence in superconducting qubits. Phys. Rev. Lett. 96,
operation (τ0 ∼ ~/∆2 ∼ 1 ns) is expected to be about three orders 047001 (2006).
of magnitude shorter then the decoherence time of the unprotected 8. Faoro, L. & Ioffe, L. B. Microscopic origin of critical current fluctuations in
qubit. In addition, only a small fraction of qubits are addressed large, small, and ultra-small area Josephson junctions. Phys. Rev. B 75,
at a given moment of time, whereas all other qubits involved in 132505 (2007).
9. Aharonov, D., Kitaev, A. & Preskill, J. Fault-tolerant quantum computing with
computation remain in the protected state. This strongly decreases
long-range correlated noise. Phys. Rev. Lett. 96, 050504 (2006).
the effective error rate of the whole logical qubit. Thus, the 10. Aliferis, P., Gottesman, D. & Preskill, J. Quantum accuracy threshold for
possibility to ‘turn off’ decoherence in the idle state markedly concatenated distance-3 codes. Quantum Inform. Comput. 6, 97–165 (2006).
reduces the redundancy required for the implementation of error 11. Oh, S. et al. Elimination of two level fluctuators in superconducting quantum
correction schemes1,2 . bits by an epitaxial tunnel barrier. Phys. Rev. B 74, 100502(R) (2006).
12. Martinis, J. M. et al. Decoherence in Josephson qubits from dielectric loss.
Moreover, the present qubit design enables a large (although Phys. Rev. Lett. 95, 210503 (2005).
incomplete) set of operations in the protected state. The idea 13. Vion, D. et al. Manipulating the quantum state of an electrical circuit. Science
of these operations (for a different protected qubit design) was 296, 886–889 (2002).
discussed in ref. 27. The essential ingredient of this scheme is a 14. Chioresku, I., Nakamura, Y., Harmans, C. J. P. M. & Mooij, J. E. Coherent
variable coupling of a protected qubit to a large inductor L quantum dynamics of a superconducting flux qubit. Science 299,
1869–1871 (2003).
(~/2e)2 EJ−1 . Such a coupling can be provided by a chain of SQUIDs 15. Pashkin, Yu. A. et al. Quantum oscillations in two coupled charge qubits.
(ref. 28). In the idle state of the qubit, each SQUID is frustrated by Nature 421, 823–826 (2003).
the magnetic flux Φ0 /2, which results in zero coupling. The current 16. Wallraff, A. et al. Strong coupling of a single photon to a superconducting
pulse in the control line shifts the flux in SQUIDs away from its ‘idle’ qubit using circuit quantum electrodynamics. Nature 431, 162–167 (2004).
value (Φ0 /2) and couples the qubit to the inductor for a short time 17. Kitaev, A.Yu. Fault-tolerant quantum computation by anyons. Ann. Phys. 303,
2–30 (2003).
∼τ0 . The effect of this coupling is equivalent to the fault-tolerant 18. Ioffe, L. B. & Feigelman, M. V. Possible realization of an ideal quantum
R(π/4) = exp(iπ/4σ z ) unitary operation29 . The two-qubit gates can computer in Josephson junction array. Phys. Rev. B 66, 224503 (2002).
be implemented similarly. Together with the measurements and 19. Doucot, B., Feigelman, M. V. & Ioffe, L. B. Topological order in the insulating
R(π/8) rotations, these operations are sufficient for an arbitrary Josephson junction arrays. Phys. Rev. Lett. 90, 107003 (2003).
quantum computation29,30 . Although R(π/8) rotation cannot be 20. Doucot, B., Feigelman, M. V., Ioffe, L. B. & Ioselevich, A. S. Protected qubits
and Chern-Simons theories in Josephson junction arrays. Phys. Rev. B 71,
implemented in a fault-tolerant manner, the error rates acceptable 024505 (2005).
for this operation are remarkably large (∼0.5) provided that all 21. Blatter, G., Geshkenbein, V. B. & Ioffe, L. B. Design aspects of
other operations are fault tolerant (ref. 30). We conclude that the superconducting-phase quantum bits. Phys. Rev. B 63, 174511 (2001).
protected qubit design discussed here might provide us with an 22. Doucot, B. & Vidal, J. Pairing of cooper pairs in a fully frustrated
almost fault-tolerant scheme of quantum computation. Josephson-junction chain. Phys. Rev. Lett. 88, 227005 (2002).
23. Protopopov, I. V. & Feigelman, M. V. Anomalous periodicity of supercurrent
in long frustrated Josephson-junction rhombi chains. Phys. Rev. B 70,
Methods 184519 (2004).
The Coulomb energy of individual Josephson junctions in rhombi, EC , should be 24. Tinkham, M. Introduction to Superconductivity 2nd edn (Dover, 1996).
at least 0.5–1 K to satisfy the requirement kB T ∆2 < 0.5EC in our experiments 25. Doucot, B. & Ioffe, L. B. Voltage–current curves for small Josephson junction
conducted at T ∼ 30–50 mK. Taking into account that the specific capacitance of arrays: Semiclassical treatment. Phys. Rev. B 76, 214507 (2007).
Al2 O3 -based tunnel barriers is ∼50 fF µm−2 (ref. 31), the latter condition translates 26. Pop, I. M. et al. Measurement of the current-phase relation in Josephson
into submicrometre in-plane dimensions of individual Josephson junctions. The junction rhombi chains. Phys. Rev. B 78, 104504 (2008).
devices were fabricated using multi-angle electron-gun deposition of Al films 27. Kitaev, A. Protected qubit based on a superconducting current mirror.
through a nanoscale lift-off mask. Josephson junctions were formed between the Preprint at <http://arxiv.org/abs/cond-mat/0609441> (2006).
aluminium strips of a well-controlled width overlapping at the right angle. To 28. Corlevi, S., Guichard, W., Hekking, F. & Haviland, D. B. Phase-charge duality
suppress quantum fluctuations in two large (0.3 × 0.3 µm2 ) Josephson junctions, of a Josephson junction in a fluctuating electromagnetic environment.
they were shunted by an inter-digitized capacitor CID ∼ 3 × 10−14 F. Phys. Rev. Lett. 97, 096802 (2006).
For protection from external high-frequency noise and non-equilibrium 29. Gottesman, D., Kitaev, A. & Preskill, J. Encoding a qubit in an oscillator.
quasiparticles, the device was flanked by two meander-type inductances L. The Phys. Rev. A 64, 012310 (2001).
thickness of the horizontal Al strips in the meanders shown in Fig. 2d is a factor 30. Bravyi, S. & Kitaev, A. Universal quantum computation with ideal Clifford
of 2–3 smaller than that in the vertical strips. Because of the superconducting gates and noisy ancillas. Phys. Rev. A 71, 022316 (2005).
gap variations within each meander’s segment (the critical temperature in Al 31. Lichtenberger, A. W. et al. Fabrication of Nb/Al–Al2 O3 /Nb junctions with
films substantially depends on their thickness in the range of ∼10–30 nm), the extremely low leakage currents. IEEE Trans. Magn. 25, 1247–1250 (1989).
meanders ‘trap’ non-equilibrium quasiparticles generated outside the device32,33 . 32. Lutchyn, R. M., Glazman, L. I. & Larkin, A. I. Kinetics of the superconducting
The charging effects in the device were probed by applying the gate voltage Vg to charge qubit in the presence of a quasiparticle. Phys. Rev. B 74, 064515 (2006).
the gate capacitor Cg ∼ 2 × 10−18 F connected to the central ‘island’ common to 33. Aumentado, J., Keller, M. W., Martinis, J. M. & Devoret, M. H.
all chains. Nonequilibrium quasiparticles and 2e periodicity in single-Cooper-pair
transistors. Phys. Rev. Lett. 92, 066802 (2004).
Received 2 January 2008; accepted 27 October 2008;
published online 30 November 2008 Acknowledgements
References We thank J. Sanchez and S. Pereverzev for help with experiments and O. Buisson,
W. Guichard, L. Faoro, M. Feigelman, B. Pannetier and V. Schmidt for stimulating
1. Steane, A. M. Overhead and noise threshold of fault-tolerant quantum error
discussions. The work at Rutgers University was supported in part by the NSF grant
correction. Phys. Rev. A 68, 042322 (2003).
ECS-0608842 and the Rutgers Academic Excellence Fund. The work at the University
2. Knill, E. Quantum computing with realistically noisy devices. Nature 434,
of Paris has been partly supported by the ANR grant ANR-06BLAN-0218-01.
39–44 (2005).
3. Steffen, M. et al. Measurements of the entanglement of two superconducting
qubits via state tomography. Science 313, 1423–1425 (2006). Additional information
4. Yoshihara, F. et al. Decoherence of flux qubits due to 1/f flux noise. Supplementary Information accompanies this paper on www.nature.com/naturephysics.
Phys. Rev. Lett. 97, 167001 (2006). Reprints and permissions information is available online at http://npg.nature.com/
5. Schreier, J. A. et al. Suppressing charge noise decoherence in superconducting reprintsandpermissions. Correspondence and requests for materials should be
charge qubits. Phys. Rev. B 77, 180502(R) (2008). addressed to M.E.G.

ARTICLES
Size and mobility of excitons in (6, 5)

carbon nanotubes
Larry Lüer1 *, Sajjad Hoseinkhani1 , Dario Polli2 , Jared Crochet3 , Tobias Hertel3
and Guglielmo Lanzani2
Knowledge of excited-state dynamics in carbon nanotubes is determinant for their prospective use in optoelectronic
applications. It is known that primary photoexcitations are quasi-one-dimensional excitons, the electron–hole correlation
length (‘exciton size’) of which corresponds to a finite volume in the phase space. This volume can be directly measured by
nonlinear spectroscopy provided the time resolution is short enough for probing before population relaxation. Here, we report
on the experimental determination of exciton size and mobility in (6, 5) carbon nanotubes. The samples are sodium cholate
suspensions of nanotubes (produced by the CoMoCat method) obtained by density-gradient ultracentrifugation. By using
sub-15 fs near-infrared pulses to measure the nascent bleach of the lowest exciton resonance, we estimate the exciton size to
be 2.0 ± 0.7 nm. Exciton–exciton annihilation in our samples is found to be rather inefficient so that many excitons can coexist
on a single nanotube.
O
wing to their size and their unique mechanical, chemical (Ex1 ) manifold to the ground state (GS) and the creation of
and electronic properties, carbon nanotubes (CNTs) are excitons in the second manifold (Ex2 ) as short-lived intermediates
considered to be promising candidates for a variety of of lifetime 1/k21 . From intensity-dependent exciton decay, the
electronic and optical applications. In semiconducting CNTs, the annihilation coefficient ka is obtained. Unfortunately, mobilities
lowest-lying photoexcitations are strongly bound one-dimensional obtained from these methods differ by orders of magnitude and
excitons with binding energies—depending on chirality and it is not clear whether variable sample character and quality or
diameter—of up to about 0.5 eV (ref. 1). Such excitations are systematic differences in the way that mobilities were measured
described as bound electron–hole pairs with two characteristic can account for the discrepancies. Pump–probe anisotropy was
coordinates, the centre-of-mass position (RCM ) and the relative shown to depolarize within a few picoseconds6 , which was
electron–hole distance or correlation length (ξe ). The latter is rationalized by a high exciton mobility leading to an exciton
directly linked to the electronic structure of the material and its diffusion length in excess of the typical nanotube curvature radius.
physical properties, such as screening, Coulomb attraction, binding Fluorescence quenching, on the other hand, yields a rather short
energy, exchange interaction and confinement of wavefunctions. exciton diffusion length of about 90 nm (ref. 7) and diffusion
For this reason, ξe is the key figure of merit for a better constant of Dexc = 0.4 cm2 s−1 . If exciton–exciton annihilation
understanding of optical, optoelectronic and photonic properties is limited by one-dimensional diffusion, then the annihilation
and also for validation of existing theories on excitonic effects coefficient attains the time-dependent form ka = ka∗ t −1/2 ; in this
in CNTs. Specifically, if ξe is comparable to the lattice constant, case, exciton annihilation and fluorescence√ quenching results
then the exciton is of the tightly bound Frenkel type, typical can be related to each other through ka∗ = 32Dexc /π (ref. 10,
of molecular solids. In contrast, a value of ξe much larger also see Supplementary Information). By using the reported
than the lattice constant supports the Wannier–Mott picture, value for the annihilation constant ka ≈ 300 nm ps−1 (ref. 8), the
typical of covalent semiconductors. The existence of excitons obtained diffusion coefficient is about 90 cm2 s−1 (at t = 1 ps).
in CNTs has been invoked theoretically2 and has recently been In summary, there have been many attempts to evaluate the
confirmed experimentally1 ; however, many of their properties exciton diffusion constant but a unifying interpretation of these
remain controversial. In particular, the size of excitons, as expressed experiments remains elusive, presumably owing to the variety of
by the average electron–hole separation in the Wannier–Mott techniques and samples used in these investigations, leaving many
envelope function, has been predicted theoretically3–5 but was never questions unanswered.
measured experimentally. Different theories report values for this Here, we present the first experimental determination of the
critical parameter ranging from about 1 to 2 nm. exciton size in semiconducting CNTs by probing the lowest
Not only exciton size, but also exciton mobility in CNTs energetic (E11 ) excitonic transition in (6, 5) tubes with sub-15 fs
is currently debated. Exciton mobility has been explored temporal resolution. According to the phase space filling model11 ,
by pump–probe depolarization spectroscopy6 , using stepwise the measured reduction in oscillator strength can be directly related
quenching of single molecule fluorescence7 and through an to the exciton size ξe . In addition, we use the experimentally
8,9
investigation of exciton–exciton annihilation (‘Auger’) kinetics
ka
. determined exciton densities to quantify exciton annihilation
In the last case, the bimolecular interaction 2Ex1 −→ Ex2 + kinetics, which in turn yield the exciton diffusion constant and
k21
GS −→ Ex1 + GS causes a rapid decay of excitons in the first diffusion length.
1 National
Laboratory for Ultrafast and Ultraintense Optical Science, INFM-CNR, Dipartimento di Fisica, Politecnico di Milano, Italy, 2 Italian Institute of
Technology and Dipartimento di Fisica, Politecnico di Milano, P.za L. da Vinci 32, 20133 Milano, Italy, 3 Institut für Physikalische Chemie, Universität
Würzburg, Am Hubland, 97074 Würzburg, Germany. *e-mail: larry.luer@fisi.polimi.it.

a 1,060 b 0.2
ΔA 200 fs
1,040 0.01 0 500 fs
Probe wavelength (nm) 2,000 fs
0
ΔAani, normalized
1,020 –0.2
–0.02
1,000 –0.4
–0.04
–0.06 –0.6
980
–0.08 –0.8
960
–1.0
940
0 1 2 3 940 960 980 1,000 1,020 1,040
Pump¬probe delay (ps) Probe wavelength (nm)
c d
0.5
0
0.4
ΔAani, normalized
0.3
r
–0.5
0.2
Probe wavelengths:
992 nm 0.1
1,005 nm
–1.0 960 nm
0
0 1 2 3 0 10 20
Pump¬probe delay (ps) Pump¬probe delay (ps)
Figure 1 | Two-dimensional pump–probe spectroscopy. a, Two-dimensional pump–probe spectrum in the E11 bleaching band after pumping with 15 fs
pulses centred at 960 nm (pump energy 8.3 nJ). The signal strength 1A is given by the false colour scale on the right. b, Spectra, as obtained by
normalizing vertical cuts through a at fixed delay times (see key). c, Kinetics, obtained by normalizing horizontal cuts through a at fixed wavelengths
(see key). d, Time-dependent anisotropy r(t) in the E11 band after pumping the E22 band with 70 fs pulses.
A two-dimensional pump–probe spectrum for resonant where the indices k and ⊥ refer to parallel and perpendicular
excitation of the E11 transition at 992 nm is shown in Fig. 1a, polarization of pump and probe beams, respectively. We found
showing the pump-induced differential absorption 1A(λpr ,t ) as r ≈ 0.3 for 0 < t < 40 ps, see Fig. 1d. This is close to the value
a function of probe wavelength λpr and pump–probe delay t . It was r = 0.4 expected for a sample containing randomly oriented
obtained with parallel polarization of pump and probe beams. At nanotubes. The lack of anisotropy decay contrasts with previous
t = 0, the transient spectrum is characterized by reduced absorption results in bundles, which are possibly affected by inter-tube
within the E11 band because population is transferred from the exciton migration.
ground state to the excited state (ground-state bleaching). The A power dependence of optical transients for resonant excitation
maximum ground-state bleaching is found at the peak of the of the E11 transition is shown in Fig. 2a. The pump fluence Ipu
E11 band of the (6, 5) tubes at 992 nm. Normalized pump–probe was varied over two orders of magnitude from 0.5 to 50 µJ cm−2 .
spectra for three different delay times, obtained from vertical cuts At the highest pump fluence, the differential absorbance reaches a
through Fig. 1a, are shown in Fig. 1b. We find that these spectra maximum of 1A992 nm = −0.14. Comparison with the ground-state
are practically identical from 990 to 1,050 nm, demonstrating absorbance of AGS,992 nm = 0.69 shows that this is still far from
that the observed transition does not exhibit a time-dependent the saturation of the electronic transition. In Fig. 2b, we compare
shift. Note that the shoulder at 1,030 nm is probably due to the kinetics of normalized cross-correlations for different pump
bleaching of the (7, 5) tubes, which are present in small amounts. fluences and find significant changes only for pump intensities
The spectra differ in the shorter-wavelength range, where we exceeding 5.8 µJ cm−2 . The traces at low pump fluences around
observe a positive 1A signal caused by photoinduced absorption t = 0 do not show a fast decay on a 100 fs timescale, demonstrating
at t = 2 ps (ref. 12). A horizontal cut through the data in Fig. 1a that the pump–probe cross-correlation is much shorter than any
yields decay kinetics at specific wavelengths as shown in Fig. 1c. excitation dynamics under these conditions.
Comparison of transients at different probe wavelengths reveals The degree to which an exciton is localized can most readily be
that the kinetics does not change significantly between 992 and accessed through the corresponding distribution of electron (fe (k))
1,005 nm. In contrast, we do find a faster decay on the blue side of and hole (fh (k)) oscillator strengths, which in the case of resonantly
the bleaching band, around 960 nm, caused by delayed formation generated excitons is related to Ψ̃exc (k), the Fourier transform of the
of the blue-shifted photoinduced absorption band12 . Note that exciton relative motion orbital wavefunction Ψexc (x) by13 :
a coherent artefact around zero time delay has been omitted.
Both, transient spectra and kinetics suggest that the contribution N 2
fe (k) = fh (k) = Ψ̃exc (k) ,
of photoinduced absorption is negligible at the maximum of 2
the ground-state bleaching band as well as on its red wing. The
anisotropy of the transient bleaching signal in the E11 transition where N is the density of excitons per unit length. The exclusion
is shown in Fig. 1d, given as r = (1Ak − 1A⊥ )/(1Ak + 21A⊥ ), principle for fermions blocks transitions into final states that are

already occupied. The effect of this ‘phase space filling’ is an overall a

reduction of oscillator strength on exciton creation13 : 0 0.5
∂f X 5.8
fe (k) + fh (k) Ψ̃exc (k)/Ψexc (x = 0).

=− (1) 18.3
f k
–0.05 29
ΔA ani
The relative motion exciton wavefunction in one dimension can be 46.4 μJ cm–2
approximated by a Gaussian function14 , Ex2
1 ∗ –1/2
–0.10 k 21 k t N1
2 a
√ −1
q
Ψexc (x) = a π exp(−x 2 /(2ξe2 )), Ex1
hν pump k 10t α 1 ∗ –1/2
k t N1
2 a
where ξe is the electron–hole correlation length, which can –0.15 GS

be identified with the exciton size. With this wavefunction, 0 0.5 1.0 1.5 2.0
equation (1) becomes Pump¬probe delay (ps)
δf b
= −N γ ξe , γ ≈ 2.05. (2) 0
f
Pump fluences:
Here, γ is a wavefunction-dependent proportionality constant (see 46.4 μJ cm–2
Supplementary Information). The relative reduction in oscillator
ΔA ani, normalized
29
strength is equal to the relative transient ground-state bleaching, 18.3
–0.5
δf /f = 1A/A, available directly from experiment. Equation (2) is 5.8
of the simple and instructive ‘saturation’ form, 1A/A = −N /NS , 0.5
where NS−1 = γ ξe , the saturation density, is inversely proportional
to the exciton size. The exciton density per unit length in our
experiment is related to the macroscopic exciton density per unit –1.0
area, n, through N = n/η(6,5) , where η(6,5) is the total length of
all (6, 5) tubes in the light path normalized to its cross-section,
as obtained from η(6,5) = ξC nC . Here, ξC−1 = 0.01 nm−1 is the C 0 0.5 1.0 1.5 2.0
atom density per unit length of a (6, 5) tube, and nC = A/σC is Pump¬probe delay (ps)
the macroscopic density of carbon atoms per unit area. This can
be evaluated by measuring the natural ground-state absorption Figure 2 | Pump intensity dependence of pump–probe kinetics. a, Kinetic
A (defined through the transmission T by A = − ln T ) using traces of differential absorption 1A in the first excitonic transition E11 of
the natural absorption cross-section σC = 7 × 10−18 cm2 obtained (6, 5) CNTs as a function of pump intensity (solid lines). Pump pulses: 12 fs
by Zheng et al.15 (and private communication (2008); see also duration, centre wavelength 960 nm. Probe wavelength is 992 nm. Parallel
Supplementary Information). Finally, at low excitation levels where polarization of pump and probe beams. Dashed lines: fits according to the
two-photon processes are negligible, the macroscopic density of kinetic model given in the inset and discussed in the text. Inset: Kinetic
excitons per unit area is equal to the number of absorbed photons model that has been used to reproduce the decay traces. b, As in a, but
per unit area, n = Iabs . Inserting these relations into equation (2), normalized curves. Dashed lines: numerical fits, normalized to the
and considering the anisotropy of the exciton distribution induced maximum; solid lines: experimental curves, normalized to the respective
by the linearly polarized pump field, finally yields fits for t > 200 fs.
1Aani = −ra γ σC ξe /ξC Iabs , Fig. 1. First, we fit numerical solutions of the rate equations
dN1 /dt = −k10 t α N1 − ka N12 + k21 N2 and dN2 /dt = (1/2)ka N12 −
with ra ≈ 1.6 being the correction factor to obtain the differential k21 N2 to our data, where N1 and N2 are the first (Ex1 ) and
absorption of the hypothetical isotropic distribution (see second (Ex2 ) exciton densities per unit length, respectively,
Supplementary Information). k1 = k10 t α is the dispersive first-order decay with exponent
The experimental differential absorption as a function of the α, k21 = 2.3 × 1013 s−1 is the rate constant for the first-order
absorbed light intensity Iabs is shown in Fig. 3a. We find linear decay from the second to the first exciton sub-bands16 and
behaviour at low Iabs , whereas saturation is observed at higher ka = ka∗ t −1/2 is the exciton annihilation rate constant (see the
intensities. We repeated the experiment on different spots on inset in Fig. 2a and Supplementary Information for details).
the sample, to check for sample inhomogeneity, and also on We note that the t −1/2 dependence of the annihilation process
a different sample containing small bundles of nanotubes; all is caused by one-dimensional diffusion under the assumption
experiments (demonstrated by different colours) yield similar of immediate annihilation on contact. The presence of strong
results. A magnification of the linear region is shown in Fig. 3b. A Coulomb interaction enables us to reject a multi-collision model
linear regression yields a slope of −rγ σC ξe /ξC = −4.5 × 10−15 cm2 . (see also Supplementary Information). As shown by the dashed lines
From this, we calculate an exciton size of ξe = 2.0 ± 0.7 nm, using in Fig. 2, this model enables us to describe the measured dynamics at
a Gaussian envelope function (see Supplementary Information all pump intensities by a single set of adjustable parameters. From
for other envelope functions). This value is in fair agreement this we obtain a time-dependent exciton annihilation constant of
with theoretical predictions4 . We note that the exciton size can ka (t ) = 1.0 × 107 nm s−1/2 t−1/2 . The excellent agreement of this
be influenced by a screening effect of the surrounding medium; simple model with experimental data lends strong support to the
when comparing isolated nanotubes with small bundles, we do not assertion of one-dimensional diffusion in this system. It enables
observe a significant extra screening effect caused by bundling. us to obtain both the annihilation coefficient and the diffusion
In the following, we discuss how the exciton annihilation constant from the same measurement. At t = 1 ps, the annihilation
constant can be obtained from the intensity-dependent data in rate becomes ka (1 ps) = 10 nm ps−1 . This is more than one order

a does not generate bleaching of the E11 resonance. Although we
Sample 1 cannot conclusively assign this process, we propose two alternatives.
0 Sample 2 no. 1
Sample 2 no. 2 The effect may be explained by sequential excitation of an Ex1
state by a second pump photon, generating higher excited states.
Alternatively, a true two-photon absorption is also possible, and
can be expected at the peak powers that we use (see estimate in
ΔAani
–0.1 Supplementary Information). Two-photon absorption with 1.24 eV

photons does not necessarily create transient bleaching because
2.48 eV may be the resonance energy for a gerade–gerade transition
of a different tube type. Photons that are ‘consumed’ by the
two-photon process cannot contribute to linear absorption.
–0.2 Our value ξe = 2.0 ± 0.7 nm for the exciton size in (6, 5) CNTs
is in agreement with theoretical predictions. As it is much larger
0 1 2 than the lattice constant in CNTs, it supports a posteriori the
Iabs (1014 cm–2) conjecture that excitons in semiconducting CNTs can be described
by a Wannier–Mott picture. In isolated tubes, we find a low
b exciton annihilation rate, resulting in an exciton diffusion length
0 Sample 1 of below 10 nm. This is an order of magnitude less than the
Sample 2 no. 1
Sample 2 no. 2
previously reported value, obtained by fluorescence quenching7 .
–0.02 The difference can be understood considering that the decay
of the exciton population is strongly dispersive12 with more
than half of the initial population decaying within 4 ps, whereas
–0.04
typical fluorescence decay times are in the range of 20–200 ps
ΔAani
Slope = 4.5 × 10–15 cm2

(ref. 17). Consequently, fluorescence probes predominantly the low
–0.06 population of long-lived excitons, and so the diffusion length of
90 nm is typical for long-lived excitons. In contrast, differential
–0.08
absorption is a direct probe for exciton population, and thus the
short diffusion length that we find is that of the short-lived excitons,
being the most abundant ones. This finding is in favour of the
–0.10 prospective use of CNTs in optoelectronics. Moreover, knowledge
0 1 2 2
of exciton mobility is crucial for the use of CNTs as biosensors, and
Iabs (1013 cm–2)
our results suggest that the right length scale for surface quenching
of optical properties is below 10 nm.
Figure 3 | Extraction of exciton size from the intensity-dependent bleach.
a, Maximum differential absorption from Fig. 1 as a function of the Methods
absorbed pump intensity, Iabs . The experiment was done with different The study presented here was carried out using CNTs from the CoMoCat
samples, as indicated. No. 2: same sample, but different spot and double procedure. Colloidal CoMoCat/sodium cholate suspensions were obtained by
spot size. The error bars reflect the uncertainty in the incident light density-gradient ultracentrifugation. Semiconducting (6, 5) CNTs from this
process were dispersed in a xerogel, where they exhibited negligible inter-tube
intensity. b, Differential absorption as a function of Iabs , for the measured
interaction. Details of the processing method can be found in a previous paper18
points in the linear regime of Fig. 2a. The solid line is a linear regression and in Supplementary Information.
through the origin. Near-infrared light pulses in the 900–1,100 nm wavelength region were
generated by a home-built non-collinear optical parametric amplifier (ref. 19)
of magnitude less than obtained in previous studies. The diffusion and compressed to nearly transform-limited 15 fs duration by a fused-silica
prism pair. We used a degenerate pump–probe set-up equipped with fast spectral
constant that we obtain is D ≈ 0.1 cm2 s−1 , of the same order acquisition using a computer-controlled optical multichannel analyser enabling
7
of magnitude as the value obtained by fluorescence √ quenching . single-shot detection at the full 1 kHz repetition rate. This set-up enables us to
The exciton diffusion length as given by Lexc = Dτe , where τe record two-dimensional differential transmission maps as a function of probe
is the non-radiative decay time, τe = 4 ps (see Supplementary wavelength and delay in a few-minute measurement time with high sensitivity.
Information), is found to be Lexc = 6 nm. This value might depend Pump and probe pulses pass through the same focusing mirrors and thus have an
on local imperfections acting like trapping sites, and not be equal spot size of about 60 µm, as measured by the ‘razor blade method’ (see also
Supplementary Information). Instrumental response function and time zero were
of general validity. Yet we highlight here that in our samples verified daily by measuring the cross-correlation of the pump and probe pulses.
many excitons can coexist on a single nanotube, owing to the More details of the experimental set-up have been published elsewhere20 .
small diffusion length. Note that an order-of-magnitude guess
for Lexc can be obtained directly from Fig. 2: it is found that Received 21 January 2008; accepted 24 October 2008;
at a maximum bleach of 1A = −0.08, which corresponds to a published online 23 November 2008
phase space occupation of about 12%, exciton decay is enhanced
by about 30% when compared with low-intensity decay. It can References
1. Wang, F., Dukovic, G., Brus, L. E. & Heinz, T. F. The optical resonances in
immediately be concluded that only 30% of the excitons move carbon nanotubes arise from excitons. Science 308, 838–841 (2005).
more than four times their size. So the diffusion length from 2. Ando, T. Excitons in carbon nanotubes. J. Phys. Soc. Jpn 66, 1066–1073 (1997).
this simple calculation must be below 8 nm, in accordance with 3. Perebeinos, V., Tersoff, J. & Avouris, P. Scaling of excitons in carbon
our result. nanotubes. Phys. Rev. Lett. 92, 257402 (2004).
Finally, we address the sublinear dependence of the initial 4. Tretiak, S. et al. Excitons and Peierls distortion in conjugated carbon
bleaching at high pump intensity. As mentioned above, it cannot nanotubes. Nano Lett. 7, 86–92 (2007).
5. Chang, E., Bussi, G., Ruini, A. & Molinari, E. Excitons in carbon nanotubes:
be caused by complete phase space filling because in this case, An ab initio symmetry-based approach. Phys. Rev. Lett. 92, 196401 (2004).
saturation should occur at 1A ≈ −0.69 and not at 1A ≈ −0.2 6. Korovyanko, O. J., Sheng, C.-X., Vardeny, Z.V., Dalton, A. B. &
as observed. To correctly account for this effect, we included a Baughman, R. H. Ultrafast spectroscopy of excitons in single-walled carbon
‘parasitic’ absorption process that is quadratic in pump power but nanotubes. Phys. Rev. Lett. 92, 017403 (2004).

7. Cognet, L. et al. Stepwise quenching of exciton fluorescence in 18. Crochet, J., Clemens, M. & Hertel, T. Quantum yield heterogeneities of
carbon nanotubes by single-molecule reactions. Science 316, aqueous single-wall carbon nanotube suspensions. J. Am. Chem. Soc. 129,
1465–1468 (2007). 8058–8059 (2007).
8. Wang, F., Dukovic, G., Knoesel, E., Brus, L. E. & Heinz, T. F. Observation of 19. Manzoni, C., Polli, D. & Cerullo, G. Two-color pump-probe system broadly
rapid Auger recombination in optically excited semiconducting carbon tunable over the visible and the near infrared with sub-30 fs temporal
nanotubes. Phys. Rev. B 70, 241403(R) (2004). resolution. Rev. Sci. Instrum. 77, 023103 (2006).
9. Ma, Y.-Z., Valkunas, 1 L., Dexheimer, S. L., Bachilo, S. M. & Fleming, G. R. 20. Polli, D., Lüer, L. & Cerullo, G. High-time-resolution pump-probe system
Femtosecond spectroscopy of optical excitations in single-walled carbon with broadband detection for the study of time-domain vibrational dynamics.
nanotubes: Evidence for exciton–exciton annihilation. Phys. Rev. Lett. 94, Rev. Sci. Instrum. 78, 103108 (2007).
157402 (2005).
10. Goesele, U. M. Reaction kinetics and diffusion in condensed matter. Prog. Acknowledgements
React. Kinetics 13, 63–161 (1984). This work was financially supported by the European Commission through the Human
11. Greene, B. I., Orenstein, J. & Schmitt-Rink, S. All-optical nonlinearities in Potential Programme (Marie-Curie RTN BIMORE, Grant No. MRTN-CT-2006-035859)
organics. Science 247, 679–687 (1990). and by the National Science Foundation (NSF DMR-0606505). We acknowledge the
12. Zhu, Z. et al. Pump-probe spectroscopy of exciton dynamics in (6, 5) carbon financial support of the project FIRB SYNERGY.
nanotubes. J. Phys. Chem. C 111, 3831–2825 (2007).
13. Schmitt-Rink, S., Chemla, D. S. & Miller, D. A. B. Theory of transient Author contributions
excitonic optical nonlinearities in semiconductor quantum-well structures. All authors contributed significantly to this work. In detail, project planning by G.L. and
Phys. Rev. B 32, 6601–6609 (1985). T.H., sample preparation and characterization by J.C. and T.H., measurement set-up
14. Capaz, R. B., Spataru, C. D., Beigi, S. I. & Louie, S. G. Diameter and chirality and automation by D.P. and L.L., experiments by L.L. and S.H., data evaluation by L.L.,
dependence of exciton properties in carbon nanotubes. Phys. Rev. B 74, G.L. and S.H., manuscript writing by L.L., G.L., T.H. and D.P.
121401, R (2006).
15. Zheng, M. & Diner, A. Solution redox chemistry of carbon nanotubes. J. Am. Additional information
Chem. Soc. 126, 15490–15494 (2004). Supplementary Information accompanies this paper on www.nature.com/naturephysics.
16. Manzoni, C. et al. Intersubband exciton relaxation dynamics in single-walled Reprints and permissions information is available online at http://npg.nature.com/
carbon nanotubes. Phys. Rev. Lett. 94, 207401 (2005). reprintsandpermissions. Correspondence and requests for materials should be
17. Hagen, A. et al. Phys. Rev. Lett. 95, 197401 (2005). addressed to L.L.

ARTICLES
Atomic and molecular signatures for

charged-particle ionization
Ola Al-Hagan1 , Christian Kaiser2 , Don Madison1 * and Andrew James Murray2 *
The way in which atoms and molecules are ionized by the impact of charged particles has important consequences for the
behaviour of many physical systems, from gas lasers to astrophysical plasmas. Much of our understanding of this process
has come from ionization measurements of the energy and angular distribution of electrons ejected in the same plane as
the trajectory of the incident ionizing beam. Such studies suggest that the mechanisms governing the ionization of atoms
and molecules are essentially the same. But by measuring the electrons ejected from a gas in a plane perpendicular to the
incident beam, we show this is not always the case. Experiments and quantum mechanical calculations enable us to construct a
remarkably accurate classical picture of the physics of charged-particle ionization. This model predicts that the differences in
ionization behaviour arise in molecules that do not have nuclei at their centres of mass.
Experimental arrangement and results
T
he most sophisticated experiments being carried out at
present measure the ionization probability as a function of Here, we report an investigation of ionization in a plane
the outgoing projectile and ejected electron momenta1–4 . perpendicular to the incident beam direction (ψ = 90◦ in Fig. 1)
These measurements, called differential cross-sections (DSCs), using electrons as projectiles and atomic helium and molecular
provide very sensitive tests for theory. Theoretical models for hydrogen as targets. Figure 2 shows the experimental data for
low to intermediate energies (where the ionization probability electron impact ionization of He and H2 in the perpendicular
is highest) must consider many factors, including distortions plane, where the outgoing electron energies are Ea = Eb = 10 eV
in the wavefunctions describing the projectile and target, target (the incident electron energy is 44.6 eV for He and 35.6 eV for
polarization due to the Coulomb interaction between the H2 ). As neutral He and H2 have an equivalent number of protons
incident projectile, nucleus and bound electrons, exchange effects, and electrons, these results markedly contrast the difference in
multiple scattering and post-collision interactions between particles distribution of the constituents that make up these atomic and
emerging from the reaction. The most sophisticated theories molecular targets. For He, three peaks are observed as a function of
include all of these processes, and compare well to experimental the angle ϕ = θa +θb (see Fig. 1), with a large central peak at ϕ = 180◦
data for atomic targets such as hydrogen5–7 , helium8 , the noble (two electrons leaving back-to-back) and clearly resolved smaller
gases9–13 and alkali and alkali-earth metals14,15 . peaks at ϕ ∼ 90◦ ,270◦ (the three-lobe atomic helium structure has
A common experimental arrangement is to fix the scattered previously been observed for different kinematics17 ). Similar to He,
projectile energy and angular location, and then measure the we find peaks in the vicinity of ϕ ∼ 90◦ ,270◦ for ionization of H2 .
probability that the ejected electron emerges at different angles However, instead of a maximum for back-to-back scattering as in
in a plane determined by the initial and final momentum of He, we find a minimum at ϕ = 180◦ . This difference must be due to
the projectile (called the coplanar scattering plane, ψ = 0◦ in either the nuclear configuration of H2 compared with He, or to the
Fig. 1). These measurements show that there is a large probability different bound-state electron momentum distributions. The data
for ejecting the target electron in the direction of the projectile clearly show the sensitivity of measurements in this geometry. It
momentum change (this is for ionization of s states; for p states, should be noted that experimental results in a coplanar geometry
this peak may split into two lobes centred on the direction of in this energy regime are very similar for both targets, which means
momentum change), and a smaller probability that the electron the marked differences between atoms and molecules seen here
is ejected opposite to this direction16 . In the first feature, the are not observable in the usual coplanar geometry adopted by
ejected electron moves in a direction that conserves momentum most researchers.
for the projectile-ejected electron system, and so this is attributed
to a classical binary collision between these two particles, which Possible types of collision
is then called the binary peak. The second feature is attributed It is instructive to consider classically how the projectile–target
to a binary collision sending the atomic electron in the direction interaction can produce ionization into the perpendicular plane.
of momentum transfer, followed by an elastic 180◦ backscattering First, consider only binary collisions between the projectile and the
from the nucleus. This second feature containing a double collision target electrons, ignoring the nuclei. If we look at the 180◦ case,
process is called the recoil peak because the nucleus must recoil where there is a large difference between atoms and molecules,
to conserve momentum. Binary and recoil peaks are the dominant the two final-state electrons have equal energies and are moving
features in all ionizing collisions, and are found for all projectiles in opposite directions such that the net final-state momentum is
and for all atomic and molecular targets. zero. This means that, in the initial state, the bound-state electron
1 Departmentof Physics, Missouri University of Science and Technology, Rolla, Missouri 65409, USA, 2 School of Physics and Astronomy, University of
Manchester, Manchester M13 9PL, UK. *e-mail: madison@mst.edu; Andrew.Murray@manchester.ac.uk.

Perpendicular plane
x Helium
Incident Inter 1.0
me
dia
Detection plane Ea = Eb = 10 eV
electron te
geo 0.8
me Ea
try
0.6 Type (c)
Coplanar geometry ψ θa Plane wave
Detected collision Projectile electron
θb electrons 0.4
DCS (normalized)
Eb 0.2 Type (b) 3DW Type (b)
Detection plane
0
Hydrogen
Figure 1 | The experimental geometry. A plane is defined by the detected 2.0
Plane wave Ea = Eb = 10 eV
electrons. The incident-electron gun can move from a coplanar geometry Projectile electron
1.5
(ψ = 0◦ ) to the perpendicular plane (ψ = 90◦ ), where the angle ϕ = θa + θb
is defined. A common point between all planes occurs when θa = θb = π /2. 1.0 Type (b) Type (b)
momentum kbd would need to be opposite to the initial projectile 0.5

3DW
momentum kin , as shown in Fig. 3a. For back-to-back final-state 0
electron measurements, this is the only process leading to ionization 50 100 150 200 250 300
into the perpendicular plane that does not involve the nucleus. φ (°)
If we include the nucleus in our model, it is possible for
the projectile electron to enter the perpendicular plane by first Figure 2 | Experimental and theoretical DCS data in the perpendicular
undergoing a small-impact-parameter elastic collision with the plane for He and H2 targets, normalized to unity at the experimental
nucleus, followed by a classical binary collision with the atomic maximum. The outgoing energies were Ea = Eb = 10 eV in both cases. The
electron, so that both electrons emerge in the perpendicular plane. results show the significant differences between ionizing atomic and
In this case, a binary collision will tend to cause the two electrons molecular targets, and contrast the effects of using plane and distorted
to emerge at a relative angle of ϕ ∼ 90◦ owing to their equal waves to describe the projectile electron. Error bars in the DCS indicate the
mass (Fig. 3b). For spherically symmetric targets, scattering into statistical variation measured over a series of sweeps of the analysers
this plane must be symmetric around kin , resulting in peaks at around the detection plane. Horizontal error bars show the estimated
ϕ ∼ 90◦ ,270◦ as seen for both He and H2 . angular response of the spectrometer due to the analyser entrance
For this process to produce back-to-back (ϕ ∼ 180◦ ) electrons, apertures and the incident electron beam pencil angle. The type of collision
we must have an extra scattering from the nucleus. As small process noted in this figure is described in Fig. 3.
impact parameters are required to bring the projectile into the
perpendicular plane (of the order of 0.5a0 , where a0 ∼ 0.53 nm wave is elastic scattering from the target. Consequently, (χin ) is a
is the Bohr radius), one of the electrons (ka ) may also re-scatter wavefunction representing elastic scattering from a neutral target
from the nucleus so as to emerge in a direction opposite the other and (χscat , χej ) are wavefunctions representing elastic scattering
electron (kb ) after the binary collision occurs (Fig. 3c). As this from an ion.
happens on either side of the nucleus with equal probability, a As current experiments using molecules do not determine the
peak centred at ϕ = 180◦ results. Other second- and higher-order orientation of the molecule at the time of ionization, an average over
processes involving nuclear scattering may also occur. all orientations must be made. We approximate this by calculating
an elastic scattering potential for molecules obtained by averaging
Quantum mechanical calculation the bound-state electron charge density over all orientations, and
Although these simple classical pictures are very appealing, atomic by averaging the two nuclei over all orientations. For H2 , the two
and molecular ionization is fundamentally a quantum mechanical protons are separated by 1.4a0 , and so we approximate averaging
process. It is however possible to use quantum mechanics to test over all molecular orientations by assuming the nuclear charge of
these classical ideas. We can calculate the probability of these +2 is uniformly distributed on a spherical shell with a radius of 0.7a0
processes occurring quantum mechanically by evaluating a quantity (see Fig. 4). The approximate form of the T-matrix we use is called
called the T-matrix, which, in a three-body approximation, is the three-body distorted wave (3DW) model19,20 . The 3DW results
given by: for ionization of He and H2 are compared to experiment in Fig. 2.
As the data are not absolute, experiment and theory are each scaled
T = χscat χej Cscat–ej V |ϕbound χin i .

to unity at their highest value. The agreement between experiment
| {z } | {z } and theory is clearly very good.
Final state Initial state
The T-matrix is an integral involving the initial and final states of Using quantum mechanics to identify collision types
the system and the interaction between the projectile and target (V ). The key objective of this work is to understand the underlying
The initial state consists of the incoming projectile wavefunction physical effects producing both similarities and differences between
(χin ) and the bound-state wavefunction for the atomic or molecular atomic and molecular targets. There are two components to the
electron (ϕbound ) (we use numerical Hartree–Fock wavefunctions theoretical calculation—the bound-state wavefunctions (φbound )
for either an atom or molecule). The final state consists of and wavefunctions describing the electrons in the continuum (χ). If
the scattered projectile wavefunction (χscat ), the ejected electron the mechanism in Fig. 3a is the main contributor to the maximum
wavefunction (χej ) and the Coulomb interaction between the at ϕ ∼ 180◦ for He and also produces the minimum in H2 , this
scattered projectile and ejected electron (Cscat–ej ) (we use a form first must be due to the bound-state wavefunctions, as they contain the
proposed by Ward and Macek18 ). For the calculations presented initial-state momentum distributions kbd of the bound electrons.
here, the wavefunctions for the free particles (χ) are called distorted To determine the importance of the momentum distribution in
waves. Distorted waves are solutions of the Schrödinger equation H2 , we repeated the H2 calculation by replacing the molecular
for a spherically symmetric potential representing either the atom H2 wavefunction (φH2 ) with a He wavefunction (φHe ) (leaving
or molecule. The important physics contained in the distorted everything else unchanged for the molecule). The result of these

a
kin φ = 180°
H2 He
ka 1.4a0
+2e
+e +e
Nucleus
Electron cloud
kbd
kb Momentum
Perpendicular plane balance Spherical averaging
kin
Nuclear shell
b
φ = 90°, 270°
ka 1. Nuclear +2e
scattering
kb +2e shell
2. Binary
Perpendicular plane collision
kin Figure 4 | Averaging of the electronic and nuclear structure of the targets
c as experimental constraints mean that the orientation of the molecule
φ = 180°
ka cannot be determined. For H2 , the nuclear charge is distributed on a thin
shell of diameter 1.4 Bohr radii, whereas for He the charge is concentrated
3. Nuclear 1. Nuclear
scattering scattering
at the centre of the target.
kb
2. Binary b
Nuclear distribution causes the difference
Perpendicular plane collision The most striking observation is the difference seen at ϕ = 180◦ .
Impact parameter b The source of this feature must lie in the difference between elastic
kin scattering wavefunctions for atoms and molecules, as distorted
d Nuclear shell
waves give a peak for He and a minimum for H2 . Given that
φ = 90°, 270° these wavefunctions contain elastic scattering from the bound-state
1. Nuclear electrons as well as elastic scattering from the nuclei, we investigated
ka
scattering the electronic and nuclear contributions individually and found the
important difference lies in the treatment of the nuclei (described
H2
kb as a point charge for He and a thin spherical shell of charge for H2 ,
2. Binary collision as discussed above).
Perpendicular plane
To investigate the importance of the size of the nuclear shell
for H2 , calculations were repeated with the shell size reducing
Figure 3 | Different mechanisms that may lead to ionization in the from R = 0.7a0 to a point charge (while keeping the electronic
perpendicular plane. a, The only mechanism that can occur without component unchanged). The marked changes in the predicted
nuclear scattering. b, The effect of nuclear scattering followed by a binary results are seen in Fig. 5. As the shell diameter decreases, the
collision, leading to peaks at φ ∼ 90◦ ,270◦ . c, The triple scattering process minimum at ϕ = 180◦ becomes deepest at 0.5a0 , after which a
that leads to a central peak at φ ∼ 180◦ for targets that have a nucleus at maximum appears, which is largest for a point charge (R = 0.0a0 ).
the centre of mass. d, The effect of distributing the nuclear charge on a thin The H2 results are then very similar to that for He when the
shell, in which case the mechanism in c, cannot occur. nuclear charge is concentrated at a single point, indicating that the
distance between the nuclei has a critical role in determining the
calculations again produced a minimum at ϕ = 180◦ . This clearly ionization probability for back-to-back scattering. We also carried
indicates that the mechanism in Fig. 3a is not the primary source of out a similar calculation for He. In this case, we replaced the
the differences between He and H2 at this angle. point charge with a charge of +2 on a sphere of increasing size
The mechanisms shown in Fig. 3b,c both require elastic while leaving everything else the same. Again, we found that the
scattering from the target. As noted above, the free-particle maximum at 180◦ quickly developed into a minimum. It is therefore
distorted waves we use are elastic scattering wavefunctions from clear that the 180◦ minimum stems from the separated nuclei in the
the target. A wavefunction that does not contain elastic scattering molecule.
from the target is a free-particle plane wave. Consequently, we The effect of shell size on the peak at ϕ = 180◦ indicates this
can determine the effect of elastic scattering of the projectile from feature is dominated by the processes shown in Fig. 3c,d. If classical
the target by replacing (χin ,χscat ) by plane waves. The results of Rutherford scattering theory is used to equate the impact parameter
these calculations are also shown in Fig. 2, where it is seen that b with scattering angle, it is found that b = 0.4a0 − 0.6a0 is required
elastic scattering of the projectile from the target produces the to elastically scatter into the perpendicular plane for the present
peaks near 90 ◦ and 270◦ for both H2 and He. The 90 ◦ and kinematics. For He, the projectile (or ejected) electron following
270◦ peaks for He have been observed previously for different the binary collision is then close to the nucleus so it has a strong
kinematics17 , and Zhang et al.21 carried out a detailed study for attraction to the point nucleus, and preferentially backscatters
ionization of He into the perpendicular plane. They proposed this elastically as in Fig. 3c. For H2 , the electrons are mostly inside
method for identifying the physical mechanisms of the collision. the spherical shell at the time of the binary collision (Fig. 3d),
We hence conclude that the mechanism in Fig. 3b is responsible so experience no attractive force that would produce a peak at
for these outlying structures in both atomic He and molecular H2 ϕ = 180◦ . For binary collisions within the shell, the Coulomb force
targets and speculate that these features are generic for atomic and from the nuclei is zero, so the electrons will then leave at a mutual
molecular targets. angle ϕ ∼ 90◦ .

2.0 inside a spherical shell of charge. This spherical shell of charge

R = 0a0 R = 0a0 Expt resulted from averaging over all molecular orientations. The present
R=0
R = 0.2
results suggest that a minimum would be found for any diatomic
1.5
DCS (normalized)
R = 0.3 molecule or perhaps for any molecule that does not have a nucleus
R = 0.3a0 R = 0.5
R = 0.2a0
R = 0.7
located at the centre of mass. Conversely, molecules that have a
1.0 nucleus at the centre of mass might then be expected to act in a
similar way to an atom. To test this conjecture, experimental results
R = 0.3a0
R = 0.7a0 for ionization of the 1πg state of CO2 in the perpendicular plane are
0.5
presented in Fig. 6, where it is seen that we find a broad maximum
R = 0.7a0
R = 0.5a0
at 180◦ instead of a minimum. Consequently, these results strongly
0 suggest that molecules that have no nuclei at the centre of mass will
50 100 150 200 250 300
φ (°)
have a minimum for back-to-back scattering and molecules that
have a nucleus at the centre of mass will have a maximum, as is
found for atoms.
Figure 5 | Change in the calculated ionization DCS for H2 in the
perpendicular plane as a function of the size of the spherically averaged Received 23 April 2008; accepted 11 October 2008;
nuclear shell, normalized to unity at the experimental maximum. Error published online 16 November 2008
bars are as described in Fig. 2.
References
1. Dürr, M., Dimopoulou, C., Najjari, B., Dorn, A. & Ullrich, J.
Three-dimensional images for electron-impact single ionization of He:
1.0 Complete and comprehensive (e,2e) benchmark data. Phys. Rev. Lett. 96,
243202 (2006).
DCS (normalized)
0.8 2. Casagrande, E. M. S. et al. New coplanar (e,2e) experiments for the ionisation
of He and Ar atoms. J. Electron Spectrosc. Relat. Phenom. 161, 27–30 (2007).
4.4a0
0.6 3. Lower, J. R., Bellm, S. & Weigold, E. An improved double-toroidal
C
O O spectrometer for gas phase (e,2e) studies. Rev. Sci. Instrum. 78, 111301 (2007).
0.4 4. Deharak, B. A. & Martin, N. L. S. An out-of-plane (e, 2e) spectrometer using a
Spherical average
movable electron gun. Meas. Sci. Tech. 19, 015604 (2008).
Molecular CO2 5. Bray, I. & Stelbovics, A. Explicit demonstration of the convergence of the
0.2 Oxygen C
Ea = Eb = 7.5 eV close-coupling method for a Coulomb three-body problem. Phys. Rev. Lett. 69,
nuclear shell 53–56 (1992).
0 6. Rescigno, T. N., Baertschy, M., Isaacs, W. A. & McCurdy, C. W. Collisional
50 100 150 200 250 300
breakup in a quantum system of three charged particles. Science 286,
φ (°)
2474–2479 (1999).
7. Colgan, J., Pindzola, M. S., Robicheaux, F. J., Griffin, D. C. & Baertschy, M.
Figure 6 | DCS for ionization of CO2 in the perpendicular plane
Time-dependent close-coupling calculations of the triple-differential cross
normalized to unity at the experimental maximum. A maximum is seen at section for electron-impact ionization of hydrogen. Phys. Rev. A 65,
φ = 180◦ , as found for He, in contrast to the results from H2 . This is 042721 (2002).
attributed to the carbon nucleus being at the centre of mass of the 8. Bray, I., Fursa, D. V. & McCarthy, I. E. Convergent close-coupling method for
molecule, so that spherical averaging produces an oxygen nuclear shell electron scattering on helium. J. Phys. B 27, L421–L425 (1994).
9. Murray, A. J. & Read, F. H. Evolution from the coplanar to the perpendicular
with an extra nuclear target at the centre of the molecule. The scattering
plane geometry of helium (e,2e) differential cross sections symmetric in
process shown in Fig. 3c can then take place from this carbon nucleus. Error scattering angle and energy. Phys. Rev. A 47, 3724–3732 (1993).
bars are as described in Fig. 2. 10. Murray, A. J. & Read, F. H. Low energy (e, 2e) differential cross section
measurements on neon from the coplanar to the perpendicular plane
geometry. J. Phys. B 33, L297–L302 (2000).
The question arises of why there is a peak at 180◦ for He resulting 11. Rasch, J., Whelan, C. T., Allan, R. J., Lucey, S. P. & Walters, H. R. J. Strong
from backscattering from the nucleus when all scattering angles interference effects in the triple differential cross section of neutral-atom
should be equally likely. This is of course true only if all impact targets. Phys. Rev. A 56, 1379–1383 (1997).
parameters are equally likely. However, for the present kinematics 12. Stelbovics, A. T., Bray, I., Fursa, D. V. & Bartschat, K. Electron-impact
following the binary collision, the electrons have energy ∼10 eV ionization of helium for equal-energy-sharing kinematics. Phys. Rev. A 71,
052716 (2005).
and an impact parameter between 0.4a0 and 0.6a0 . Again using
13. Milne-Brownlie, D. S., Foster, M., Gao, J., Lohmann, B. & Madison, D. H.
Rutherford scattering from a point charge, the resulting scattering Young-type interference in (e, 2e) ionization of H2 . Phys. Rev. Lett. 96,
angles range from 150◦ to 160◦ for these values. As there is an equal 233201 (2006).
probability of left and right scattering, the resulting signal will be 14. Murray, A. J. (e, 2e) ionization studies of alkaline-earth-metal and
distributed between ±150◦ centred around ϕ = 180◦ , as is observed. alkali-earth-metal targets: Na, Mg, K, and Ca, from near threshold to beyond
intermediate energies. Phys. Rev. A 72, 062711 (2005).
Although these classical descriptions of ionization aid in
15. Srivastava, M. K., Kumar, R. C. & Srivastava, R. Coplanar doubly symmetric
explaining the differences between He and H2 and highlight (e, 2e) process on sodium and potassium. Phys. Rev. A 74, 064701 (2006).
the importance of the nuclear configuration, the fully quantum 16. Schulz, M. et al. Three-dimensional imaging of atomic four-body processes.
mechanical calculation is of course needed to accurately describe Nature 422, 48–50 (2003).
the data. It is intriguing to note that the nuclear configuration has 17. Murray, A. J., Woolf, M. B. J. & Read, F. H. Results from symmetric and
such a crucial role in the observed structures in the perpendicular non-symmetric energy sharing (e, 2e) experiments in the perpendicular plane.
J. Phys. B 25, 3021–3036 (1992).
plane, whereas it has almost no role for the binary and recoil peaks 18. Ward, S. J. & Macek, J. H. Wave functions for continuum states of charged
seen in a coplanar geometry (where most previous measurements fragments. Phys. Rev. A 49, 1049–1056 (1994).
have been made). 19. Gao, J., Peacher, J. L. & Madison, D. H. An elementary method for calculating
orientation-averaged fully differential electron-impact ionization cross
Generalizing the model to larger molecules sections for molecules. J. Chem. Phys 123, 204302 (2005).
20. Gao, J., Madison, D. H. & Peacher, J. L. Distorted wave born and three-body
The minimum in the cross-section for 180◦ scattering in the distorted wave born approximation calculations for the fully differential cross
perpendicular plane found for H2 has been attributed here to the section for electron impact ionization of nitrogen molecules. J. Chem. Phys.
fact that the binary collisions are taking place in a force-free region 123, 204314 (2005).

21. Zhang, X., Whelan, C. T. & Walters, H. R. J. Energy sharing (e,2e) Author contributions
collisions-ionisation of helium in the perpendicular plane. J. Phys. B 23, O.A.-H. and D.M. carried out the theoretical calculations for this work, whereas C.K.
L173–L178 (1990). and A.M. carried out the experimental investigations that are presented.
Additional information
Acknowledgements Reprints and permissions information is available online at http://npg.nature.com/
We thank the EPSRC (UK) for supporting these experiments, and the NSF for support reprintsandpermissions. Correspondence and requests for materials should be addressed
of the theoretical work under grant No. PHY-0757749. to D.M. or A.J.M.

ARTICLES
Structure from fleeting illumination of faint

spinning objects in flight
Russell Fung, Valentin Shneerson, Dilano K. Saldin and Abbas Ourmazd *
Moves are afoot to illuminate particles in flight with powerful X-ray bursts, to determine the structure of single molecules,
viruses and nanoparticles. This would circumvent important limitations of current techniques, including the need to condense
molecules into pure crystals. Proposals to reconstruct the molecular structure from diffraction ‘snapshots’ of unknown
orientation, however, require ∼1,000 times more signal than available from next-generation sources. Using a new approach,
we demonstrate the recovery of the structure of a weakly scattering macromolecule at the anticipated next-generation X-ray
source intensities. Our work closes a critical gap in determining the structure of single molecules and nanoparticles by X-ray
methods, and opens the way to reconstructing the structure of spinning, or randomly oriented objects at extremely low
signal levels.
T
he X-ray signal scattered by a single molecule is so faint that
crystals containing at least 108 molecules have had to be used
to determine molecular structure. Diffraction-quality crystals
are hard to produce, and complicate retrieval of the information of
interest: the structure of the molecule. Ideally, one would like to 1
Photons/pixel
dispense with the need for crystals. This has led to proposals to use
powerful next-generation X-ray sources, such as X-ray free-electron
d = 1.8 Å
lasers (XFELs), to determine the structure of individual (that is,
not crystallized) macromolecules and nanoparticles1–7 . A train of
identical objects would be successively exposed to powerful X-ray
pulses, and diffraction ‘snapshots’ collected from single objects of
unknown orientation. The diffraction patterns would be oriented
relative to each other and used to reconstruct the three-dimensional
(3D) diffracted intensity distribution (the diffraction volume) in 0
reciprocal space. In general, the object structure can then be 0.15 0.30 0.43 0.55
determined through iterative ‘phasing algorithms’8–11 . q (Å–1)
The key difficulties with this approach stem from the small
number of photons scattered by a single molecule. A 500 kD Figure 1 | Diffracted intensity. Number of photons scattered versus wave
biological molecule in the beam of a next-generation XFEL focused vector q by the test molecule chignolin. Solid curve: Noise-free signal
to 0.1 µm diameter, for example, scatters ∼4 × 10−2 photons averaged over all pixels at magnitude q, scaled to give a mean photon count
per pulse into a pixel at 1.8 Å resolution12 (Fig. 1). Determining of 4 × 10−2 per pixel at 1.8 Å resolution. Red circles: Actual signal along a
the orientations of the individual low-signal diffraction patterns typical wave vector q with shot noise. The signal level corresponds to that
and hence reconstructing the diffraction volume by proposed expected from a 500 kD biological molecule exposed to a single pulse of an
approaches require 1,000 times more signal than available12 . XFEL. The higher-intensity, information-poor regions at small wave vector
To circumvent this difficulty, suggestions have been made first are not shown.
to orientationally classify and average the individual diffraction
patterns in each class to improve the signal-to-noise ratio, and principle, and also closes a key conceptual and algorithmic gap in
then determine the orientation of each class. The per-shot dose planned ‘single-molecule’ experiments.
needed to ‘classify’ exceeds the available photon flux by two We begin by outlining the conceptual framework of our
orders of magnitude12,13 . These problems stem, in essence, from approach, first non-mathematically, then in more detail. This is
reliance on the very limited information available in one, or a few followed by demonstrating the ability to orient simulated low-
low-signal-to-noise diffraction patterns. No algorithm capable of signal diffraction patterns from a small test molecule, chignolin14
reconstructing the structure of such faint objects to high resolution (Protein Data Bank designation: 1UAO) down to a scattered mean
has been demonstrated. photon count (MPC) of 4 × 10−2 per diffraction pattern pixel
Here, we present a new approach, which exploits the correlations at 1.8 Å with shot noise—the signal-to-noise level expected for a
in the entire scattered photon ensemble to recover the structure of a 500 kD molecule. Finally, by recovering the structure of the test
faint object to high resolution from scattering snapshots of random molecule from a collection of simulated noisy diffraction patterns of
orientation, and demonstrate its power by recovering the structure unknown orientation, it is shown that the orientational accuracy we
of a single biological molecule to 1.8 Å. This serves as a proof of achieve is sufficient for structure determination to high resolution.
Department of Physics, University of Wisconsin Milwaukee, 1900 E. Kenwood Blvd, Milwaukee, Wisconsin 53211, USA. *e-mail: ourmazd@uwm.edu.

Conceptual framework Latent space of Manifest space of
A loose but illuminating analogy can be used to illustrate our orientations measured intensities
(three-dimensional) (p-dimensional)
approach. Consider the eroded fragments of an ancient Greek tp
vase recovered in a dig. The vase can be reconstructed from
Th
the correlations between the fragments. The most likely shape is Mapping
r e e-
dim e n s i o n a l m
obtained when the eroded pieces are maximally correlated with φ
each other. For best results, the correlations considered should not
be limited to the shapes of neighbouring fragments, but include p-dimensional space t2
the elaborate patterns spanning all of the fragments. In other
words, correlations in the entire data set must be considered θ t1
simultaneously. This is the basis of our approach: we exploit
the correlations in the entire ensemble of diffracted photons Figure 2 | Latent and manifest spaces. The relationship between the
to reconstruct the 3D diffraction volume. An iterative phasing three-dimensional latent (or hidden) space of orientations and the
algorithm is then used to recover the molecular structure. p-dimensional manifest (or accessible) space of measured intensities. The
A compact nomenclature is needed to describe our approach molecular orientation has only three degrees of freedom. As the molecule
in more detail. Although the algorithm does not rely on any rotates, the tip of the vector representing its diffraction pattern is confined
particular data representation, consider the ensemble of scattered to a 3D manifold in the pD manifest space. This manifold is a nonlinear
photons as a collection of diffraction patterns, each emanating mapping of the space of orientations. The mapping function can be
from a random orientation of the object. (In reality, the diffraction determined by well-known manifold-embedding techniques, and used to
patterns stem from randomly oriented members of a set of relate a particular diffraction pattern to a specific orientation of
objects assumed to be identical.) The nomenclature consists of the molecule.
representing each snapshot by a vector, the components of which
are the measured intensity values at the pixels of the snapshot Determining the orientations of diffraction patterns
(see Supplementary Information, Fig. S1). The diffracted photon We now demonstrate the ability to recover the orientations of
ensemble is then a matrix consisting of the ensemble of diffraction simulated diffraction patterns of the protein chignolin at signal-to-
pattern vectors. As described in the Methods section, individual noise levels corresponding to the data shown by the red circles in
pixels in each diffraction pattern span the interval needed to ensure Fig. 1, using no information other than the dimensionality of the
optimum information capture—‘oversampling’ in the sense of orientational space. Consider first the case where the molecule can
refs 3,8–11 . (We call a gauge satisfying the sampling requirement assume any orientation about one axis. In this case, the appropriate
an ‘appropriate sampling gauge.’) sampling angle, the natural length scale for the orientational
A molecule in a specific orientation gives rise to a vector in the accuracy needed for 1.8 Å resolution is 3.2◦ . Figure 3 shows a plot
so-called ‘manifest space’ of measured pixel intensities (Fig. 2). As of the determined versus actual orientations for a collection of
the molecular orientation is changed in the hidden or ‘latent space’ 3,000 simulated diffraction patterns (1) with no noise (infinite
of orientations, the vector representing the diffraction pattern signal) and (2) at the signal-to-noise corresponding to the data
traces out a path in the p-dimensional manifest space of measured shown by the red circles in Fig. 1. The noise-free case produces a
intensities. Because the molecule resides in 3D space, it has only root-mean-square (r.m.s.) orientation error of 1.4◦ . (The minimum
three orientational degrees of freedom. Thus, the tip of the vector error due to discretization of orientational space into 3.2◦ bins is
in the p-dimensional intensity space is confined to a 3D manifold. 0.9◦ .) When the signal is reduced to that of the data shown by the
To translate a particular position on the manifold in the manifest red circles in Fig. 1, the r.m.s. orientation error amounts to 3.8◦ . The
intensity space to a specific orientation, that is, a specific point in same accuracy was achieved when 72,000 diffraction patterns were
the 3D latent space of orientations, we must determine the mapping analysed. Diffraction patterns were therefore oriented to within 1.2
between the manifold and the latent space of orientations. This can appropriate sampling angles. As shown below, this is ample for
be done by embedding a 3D manifold in the manifest space so as structure determination to high resolution.
to include all vector tips to within noise, subject to the constraints Next, consider the case where the molecule can assume any
imposed by the geometry of the latent space. Once this function is orientation in 3D space. The possible orientations are now
known, the position of each vector in the manifest intensity space represented by points on the surface of the unit 4-sphere22 .
can be directly related to a point in the latent space of orientations, With appropriate sampling for 1.8 Å resolution, ∼105 distinct
that is, to a specific molecular orientation. orientations must be recognized, requiring ∼106 diffraction
A number of manifold-embedding techniques are available15,16 . patterns. This exceeds our current desktop computational
We use generative topographic mapping (GTM), a Bayesian capabilities. We have therefore limited our simulations to random
nonlinear factor-analytical approach originally developed for data orientations over 30◦ × 30◦ × 30◦ patches of the surface of the
projection and visualization17–20 and neural network (see, for unit 4-sphere. For a set of 103 diffraction patterns with a signal
example, ref. 21 and references therein) applications. This approach level of the data shown by the red circles in Fig. 1, the r.m.s.
determines the maximum likelihood manifold in the manifest error in orientation determination is 5.2◦ . (For further details, see
space of experimental intensity measurements by fitting the Supplementary Information, Fig. S7.) When the molecule is free to
correlations in the diffracted photon ensemble, subject to the assume any orientation in 3D, the appropriate sampling angle for
constraints imposed by the geometry of the latent space. Through 1.8 Å resolution is 5◦ . The orientational accuracy achieved is thus
its discrete treatment of the latent and manifest spaces, GTM 1.04 times the appropriate sampling angle.
enables natural classification of similar patterns into orientational
classes, and thus noise reduction through averaging. We note, Recovering the structure of the molecule
however, that averaging is carried out after the orientation As noted earlier, a full 3D orientational recovery is beyond
of each diffraction pattern ‘snapshot’ has been determined. our current computing resources. In principle, however, the 3D
In other words, GTM functions at the actual experimental molecular structure can be deduced from diffraction patterns
signal-to-noise level without the need for prior classification and obtained when the molecule is free to assume any orientation
averaging. This is a key attribute. Further details are provided in about a single axis. In practice, the curvature of the Ewald sphere
Supplementary Information. means that only part of the diffraction volume is covered by

(ii)
Determined angles (π) 2
(i)
0 1 2
Correct angles (π)
Figure 3 | Orientation recovery. Plot of determined versus actual Figure 4 | Structure recovery. Isosurfaces of electron density of the protein
orientations (modulo 2π) for 3,000 simulated diffraction patterns. (i) No chignolin, recovered from 72,000 diffraction patterns of unknown
noise (infinite signal). (ii) MPC of 4×10−2 per pixel at 1.8 Å resolution, with orientation at a MPC of 4 × 10−2 per pixel (see text). The molecular model
shot noise. The red dots represent actual results, the blue lines best linear is represented by the stick figure, with C bonds shown in yellow, N in blue
fits. The y-intercepts represent unimportant rigid rotations of the molecule. and O in red. The 1, 2 and 3σ electron density contours are shown in blue,
pink and red, respectively, with σ denoting the r.m.s. deviation from the
the ensemble of diffraction patterns, leaving regions devoid of mean electron density.
diffraction data. These gaps can be eliminated by allowing the
molecule to assume any orientation about each of two orthogonal to the present code, a 100-node computing cluster (each node a
axes in turn. We therefore used the following procedure to 2.33 GHz Intel Core 2 Duo with 4 GB random-access memory),
recover the 3D structure of the test molecule chignolin. (1) With and no symmetry in the molecule (S = 1), it should be possible to
the beam along the positive z direction and the molecule free recover the structure of a 500 kD molecule to 3 Å, a 1 MD molecule
successively to assume any orientation about the x and the y axis, to 4 Å and a 2 MD molecule to 5 Å, respectively. This range includes
a total of 72,000 diffraction patterns were simulated for random important macromolecules, nanoparticles and colloids. Clearly,
orientations of the molecule. (2) The orientations of the molecule the computational cost for recovering the structure of symmetric
were determined from the diffraction patterns. At the signal level particles is lower. For example, reconstructing the diffraction
corresponding to the data shown by the red circles in Fig. 1, the volume for the satellite tobacco necrosis virus (Protein Data Bank
r.m.s. orientational error was 1.2 times the appropriate sampling designation: 2buk) to 1 nm is no more difficult than doing so for
angle for 1.8 Å resolution. (3) The diffraction patterns belonging chignolin to 1.8 Å, the example used in this article.
to the same orientational classes, each spanning the appropriate The approach we have outlined accurately determines the
sampling angle were averaged. (4) The data were combined to orientations of diffraction patterns at very low signal levels, thus
produce a diffraction volume on a regular Cartesian grid of points filling a critical gap in the proposed single-particle experiments.
in reciprocal space. (5) An iterative phasing algorithm9 with ‘charge These proposals have assumed a single conformational state for the
flipping’ of low electron densities23 and ‘phase shifting’ of weak molecules exposed to X-ray pulses. No means have been suggested
reflections24 was used to recover the electron density shown in to deal with cases where this assumption is not valid. In our
Fig. 4. It is clear that the orientational accuracy achieved is ample approach, if the beam of molecules consists of a number of distinct
for high-resolution structure recovery at very low signal levels. conformations (or a number of different molecular types), each
should produce a different manifold in the manifest intensity space.
Implications and outlook It should be possible to fit a manifold to each type separately to
It is important to estimate the range of particle sizes amenable to our determine the structure of a number of molecular (sub)types, or
approach. The lower limit is set by the number of photons scattered include structural variations as an extra latent variable and study
1/3
to large angle. This varies as Natoms (ref. 12), where Natoms is the structural variations within a given species. This might mitigate
number of (non-hydrogen) atoms in the molecule, in effect, the the need for conformational and/or chemical homogeneity, and
molecular weight. A 500 kD molecule scatters ∼4 × 10−2 photons potentially enable the study of reactions.
per pixel at 1.8 Å resolution, with heavier molecules producing Our algorithm exploits the entire diffracted photon ensemble.
larger signals. The MPC of 10−2 per pixel, the smallest signal So far, we have been able to determine particle orientations at
level at which we can at present recover orientation, corresponds MPCs as low as 10−2 per pixel using a total of 105 scattered
to a molecular weight of 500 kD × (10−2 /4 × 10−2 )3 ∼ 8 kD. This photons. A 500 kD particle in an XFEL beam can scatter ∼109
represents our current lower limit for the molecular weight, with photons to high angle in a few minutes12 . It may therefore be
the upper limit unbounded by intensity considerations. possible to trade the per-pulse dose against the total number of
Available computational resources set the upper bound. This diffraction patterns recorded, enabling the former to be reduced.
stems from the increasingly tight orientational accuracy needed for If the per-pulse dose can indeed be reduced to below the single-
larger objects. To image a particle with diameter D to resolution molecule damage threshold, the data collection window increases
d, the number of independent orientations to be recognized is from the 20 fs (refs 25,26) anticipated at present to the 100 ps–10 ns
given by Nnodes = (8π2 /S)(D/d)3 ≈ (8π2 a3 Natoms /Sd 3 ), where S range, depending on the molecular size and rotational energy.
denotes the number of asymmetric units in the particle, and a Lower per-pulse doses might also bring single-particle structure
the interatomic spacing. We have characterized the computational determination within range of non-FEL X-ray sources, albeit at
requirements of the elementary steps in our approach, and reduced resolution. Two questions then arise. (1) What is the lowest
conducted a feasibility study of the resources needed for large practical per-pulse dose needed for structure recovery? (2) Is this
molecules and nanoparticles. Assuming appropriate modifications dose below the ‘acceptable’ damage threshold of a single molecule?

A practical application of our approach also requires the ability 6. Chapman, H. N. et al. Femtosecond diffractive imaging with a soft-X-ray
to maintain sensitivity in the presence of background scattering, free-electron laser. Nature Phys. 2, 839–843 (2006).
which is instrument dependent. These issues highlight important 7. Chapman, H. N. et al. Femtosecond time-delay X-ray holography. Nature 448,
676–679 (2007).
directions for future work. 8. Gerchberg, R. W. & Saxton, W. O. A practical algorithm for the determination
Finally, our approach reconstructs an object from sections of of the phase from image and diffraction plane pictures. Optik 35,
any shape and dimension with no orientational information. Other 237–246 (1972).
potential applications therefore include, but are not limited to, 9. Fienup, J. R. Reconstruction of an object from the modulus of its Fourier
transform. Opt. Lett. 3, 27–29 (1978).
the reconstruction of faint, radiation-sensitive objects by ultralow-
10. Sayre, D. in Imaging Processes Coherence in Physics Vol. 112
dose electron microscopy, diffraction imaging of nanoparticles and (eds Schlenker, M. et al.) 229–235 (Lecture Notes in Physics, Springer, 1980).
colloids and rapid tomography of faint macroscopic objects. 11. Miao, J., Charalambous, P., Kirz, J. & Sayre, D. Extending the methodology of
X-ray crystallography to allow imaging of micrometer-sized non-crystalline
Methods specimens. Nature 400, 342–344 (1999).
Unless otherwise stated, the pixel size is the ‘appropriate
√ sampling pixel,’ which 12. Shneerson, V. L., Ourmazd, A. & Saldin, D. K. Crystallography without
‘oversamples’ the diffracted amplitude by a factor of s 2, with s being the crystals. I. The common-line method for assembling a three-dimensional
dimensionality of the space considered (1, 2 or 3). The angle √ subtended by diffraction volume from single-particle scattering. Acta Crystallogr. A 64,
such a pixel is the ‘appropriate sampling angle’ 1θ = (d/D s 2), where d is the 303–315 (2008).
resolution, and D the diameter of the particle. For our test molecule, the 1D and 13. Bortel, G. & Faigel, G. Classification of continuous diffraction patterns: A
3D appropriate sampling angles for 1.8 Å resolution are 3.2◦ and 5◦ , respectively. numerical study. J. Struct. Biol. 158, 10–18 (2007).
With the molecule free to assume any orientation about one axis, the MPC is 14. Honda, S., Yamasaki, K., Sawada, Y. & Morii, H. 10 residue folded peptide
4 × 10−2 per pixel at 1.8 Å. The 1D sampling pixel size was used in the iterative designed by segment statistics. Structure 12, 1507–1518 (2004).
phasing algorithm. 15. Tenenbaum, J. B., de Silva, V. & Langford, J. C. A global geometric
At high angle, in the so-called Wilson statistics regime, the signal averaged framework for nonlinear dimensionality reduction. Science 290,
over a diffraction shell at a given distance from the origin is essentially independent 2319–2323 (2000).
of the detailed structure of the molecule. We scale the entire intensity distribution 16. Donoho, D. L. & Grimes, C. Hessian eigenmaps: Locally linear embedding
such that its value at 1.8 Å resolution is 4 × 10−2 photons per pixel (with s = 1). techniques for high-dimensional data. Proc. Natl Acad. Sci. 100,
Note that the photon count of 1.4 × 10−2 at ‘large angle’ in ref. 12 corresponds to 5591–5596 (2003).
4 × 10−2 at 1.8 Å. 17. Bishop, C. M. Neural Networks for Pattern Recognition
No thermal broadening was included in the calculations, because the (Oxford Univ. Press, 1995).
appropriate value for single molecules is not known. Using the value typical of 18. Bishop, C. M. in Learning in Graphical Models (ed. Jordan, M. I.) 371–403
biological crystals reduces the calculated intensity at 1.8 Å by less than a factor (MIT Press, 1999).
of 2, well within the accuracy of diffracted intensity estimates12 . (40 × 40) pixel 19. Svensen, J. F. M. The Generative Topographic Mapping. Thesis,
diffraction patterns of the protein chignolin in random orientations were simulated Aston Univ. (1998).
out to a scattering angle corresponding to 1.8 Å resolution using an incident 20. Bishop, C. M. & Tipping, M. E. A hierarchical latent variable model for data
photon wavelength of 1 Å. Shot noise was incorporated as Poisson statistics. The visualization. IEEE Trans. Pattern Recognition Machine Intelligence 20,
incident photon intensity was successively reduced so as to produce down to 281–293 (1998).
10−2 scattered photons per pixel at 1.8 Å. The innermost central pixels were not 21. Howard, R. E., Jackel, L. D. & Graf, H. P. Electronic neural networks. J. Am.
used for analysis, because, despite their higher photon counts, they contain little Telephone Telegraph Co. 67, 58–64 (1988).
orientational information. Up to ∼103 pixels from each diffraction pattern were 22. Kuipers, J. B. Quaternions and Rotation Sequences
provided to the algorithm with no information other than the dimensionality of the (Princeton Univ. Press, 1999).
orientational space. The pixels stemmed from rectangular strips at the perimeter, 23. Oszlányi, G. & Sütő, A. Ab initio structure solution by charge flipping. Acta
annuli excluding the innermost central pixels or pixels with the highest variance Crystallogr. A 60, 134–141 (2004).
across all diffraction patterns. Typically, the 103 pixels used by the program 24. Oszlányi, G. & Sütő, A. Ab initio structure solution by charge flipping. II. Use
contained a total of ∼100 photons. of weak reflections. Acta Crystallogr. A 61, 147–152 (2005).
25. Neutze, R., Huldt, G., Hajdu, J. & van der Spoel, D. Potential impact of an
Received 16 June 2008; accepted 9 October 2008; X-ray free electron laser on structural biology. Rad. Phys. Chem 71,
published online 9 November 2008 905–916 (2004).
26. Jurek, Z., Faigel, G. & Tegze, M. Dynamics in a cluster under the influence of
intense femtosecond hard X-ray pulses. Eur. Phys. J. D 29, 217–229 (2004).
References
1. Solem, J. C. & Baldwin, G. C. Microholography of living organisms. Science
218, 229–235 (1982). Acknowledgements
2. Neutze, R., Wouts, R., van der Spoel, D., Weckert, E. & Hajdu, J. Potential for We acknowledge valuable discussions with M. Schmidt and P. Schwander. We are
biomolecular imaging with femtosecond X-ray pulses. Nature 406, grateful to V. Elser for stimulating us to think about general methods for determining
752–757 (2000). orientations, and to D. Starodub for the suggestion to consider the application of our
3. Miao, J., Hodgson, K. O. & Sayre, D. Extending the methodology of X-ray approach to multicrystalline materials.
crystallography to allow imaging of micrometer-sized non-crystalline
specimens. Proc. Natl Acad. Sci. 98, 6641–6645 (2001). Additional information
4. Huldt, G., Szőke, A. & Hajdu, J. Diffraction imaging of single particles and Supplementary Information accompanies this paper on www.nature.com/naturephysics.
biomolecules. J. Struct. Biol. 144, 219–227 (2003). Reprints and permissions information is available online at http://npg.nature.com/
5. Gaffney, K. J. & Chapman, H. N. Imaging atomic structure and dynamics with reprintsandpermissions. Correspondence and requests for materials should be addressed
ultrafast X-ray scattering. Science 316, 1444–1448 (2007). to A.O.

ARTICLES
Multiphase transformation and Ostwald’s rule of

stages during crystallization of a metal phosphate
Sung-Yoon Chung1,2 *, Young-Min Kim3 , Jin-Gyu Kim3 and Youn-Joong Kim3
Although the classical picture of crystallization depicts a simple and immediate transformation from an amorphous to
a crystalline phase, it has been argued that, in selected systems, intermediate metastable phases exist before a stable
state is finally reached. However, most experimental observations have been limited to colloids and proteins, for which the
crystallization kinetics are fairly slow and the size is comparatively large. Here, we demonstrate for the first time in an inorganic
compound at an atomic scale that an amorphous phase transforms into a stable crystalline state via intermediate crystalline
phases, thus directly proving Ostwald’s rule of stages. Through in situ high-resolution electron microscopy in real time at a
high temperature, we show the presence of metastable transient phases at an atomic scale during the crystallization of an
olivine-type metal phosphate. These results suggest a new description for the kinetic pathway of crystallization in complex
inorganic systems.
T
he final shapes and sizes of crystals, factors that govern Taking amorphous LiFePO4 as a multi-component model
their resulting physical properties1 , are critically influenced compound in this study, we revealed the existence of intermediate
by the kinetic pathway of the phase transformation during metastable crystalline states during crystallization. For direct
crystallization from amorphous phases and melts. Accordingly, atomic-level observations, we used in situ high-resolution electron
a number of studies have been conducted on transition kinetics microscopy (HREM) at a high temperature. Recent progress
and thermodynamics in a variety of systems2–4 . A traditional in HREM has led to enhanced resolution for rapid and clear
model for the early stage of crystallization can be described by imaging even at elevated temperatures. Thus, it enables observation
two distinct processes2,3 : nucleation for the formation of stable of structural variations in real time in a variety of nanoscale
atomic (or molecular) clusters and subsequent growth of the nuclei. materials18–22 . HREM image processing23–25 based on two-
In contrast to this conventional description for crystallization, dimensional electron crystallography was also used to determine
a system in an unstable state does not necessarily transform the atomic arrays of intermediates phases, showing the different
directly into the most thermodynamically stable state, thus implying crystallographic characteristics between each phase.
the existence of metastable intermediate states, as empirically A series of in situ HREM images was acquired during the
described by Ostwald in 1897 and dubbed the rule of stages5 . crystallization of amorphous LiFePO4 at 450 ◦ C. Figure 1 shows the
In his conjecture, Ostwald stated that an unstable system could initial series of images between 40 and 44 min of heating. As shown
transform into another transient state, the formation of which is for the earlier stage in another series of images (Supplementary
attained by the smallest loss of free energy before finally reaching a Information, Fig. S1), the nuclei, which can be distinguished
stable state. Therefore, the recognition of such intermediate states through their level of contrast in the images, initially form in
would provide new insight into transformation kinetics during the matrix and subsequently grow as nanocrystals over time.
crystallization and be the first crucial step towards ultimate control Coalescence of nanocrystals can also be seen in Fig. 1b. Figure 1b–d
of overall crystallization behaviour. includes fast Fourier transforms (FFTs) corresponding to the
Since the notable work by ten Wolde and Frenkel showing the nanocrystal denoted by the arrow in Fig. 1a. Although the crystal
presence of an intermediate dense fluid during the crystallization structure of the nanocrystal does not change, the appearance of
of globular proteins6 , many simulations and predictions have been diffraction spots with a high-order index in the FFT of Fig. 1d
offered in attempts to shed light on possible intermediate states together with the clearer development of the lattice fringe indicates
during the early stage of crystallization, largely in organic proteins7 , that the crystallinity of the nanocrystal improves with heating time.
colloids8–10 and Lennard-Jones fluids11,12 . With the exception However, lattice defects, which seem to be stacking faults, are also
of several experimental attempts in such selective systems13–17 , observed in the crystal, as clearly shown in the Fourier-filtered
however, no atomic-scale evolutions during crystallization in image for the region outlined by a green rectangle in Fig. 1d.
inorganic compounds—which would provide direct evidence for The thermodynamically stable crystalline form of LiFePO4 is
the Ostwald rule—have been observed. Compared with organic known to be an olivine structure with cation ordering (Li at
proteins and colloidal crystals with comparatively large units, most the M 1 octahedral site and Fe at the M 2 octahedral site)26–29 in
inorganic compounds have ångström-scale lattice parameters in the Pnma space group (no. 62 (ref. 30)). Here, the diffraction
the unit cell. Furthermore, their crystallization usually occurs at pattern shown in the FFTs of Fig. 1 cannot be derived from
high temperatures. Owing to such limitations, a direct experimental the ordered olivine structure. Therefore, this result demonstrates
observation in real time of the transformation kinetics is difficult in that the nanocrystal shown in Fig. 1 is in a metastable transient
most inorganic systems. crystalline state, thus implying further transitions to the stable
1 Department of Materials Science and Engineering, Inha University, Incheon 402-751, Korea, 2 Nalphates LLC, Wilmington, Delaware 19801, USA, 3 Korea
Basic Science Institute, Daejeon 305-333, Korea. *e-mail: nalphates@gmail.com.

a 40 min 0 s b a
42 min 0 s 44 min 0 s
(450 °C)
(450 °C)
5 nm
c 42 min 30 s d 43 min 30 s
5 nm
b
44 min 30 s
Figure 1 | First series of in situ HREM images and corresponding FFTs

during the crystallization of LiFePO4 at 450 ◦ C. a–d, Images taken at 30 s c
intervals. FFTs of a nanocrystal that nucleated from the matrix (denoted by
45 min 0 s
an arrow in a) show the same pattern, indicating an identical crystal
structure. The presence of lattice defects, which appear to be stacking
faults, can be distinguished in the Fourier-filtered image in d for the region
outlined by the green square.
phase. Figure 2 reveals multiple phase transformations and the

existence of resultant intermediate crystalline states of LiFePO4
before an ordered stable olivine structure is finally achieved during
crystallization. The nanocrystal shown in Fig. 2a (red), which has
a crystallographic structure identical to that shown in Fig. 1d,
sequentially transforms into other phases with a different crystal
d
structure over the course of a few minutes, as represented in 48 min 0 s
different colours (yellow in Fig. 2b and blue in Fig. 2c). The change
in diffraction patterns in each FFT and the different high-resolution
lattice fringes of each structure also confirm the crystal-to-crystal
transformations. The final phase (green) shown in Fig. 2d was
identified as ordered olivine-type LiFePO4 in the [122] projection
–
based on a simulated electron diffraction pattern, which is shown be 210
identical to the FFT, along with the reflection indices. The magnified –
201
HREM image (green) and the simulated [122]-projection image 000
(black and white) are confirmed to be in good agreement with –
011
each other.
During real-time HREM, we observed two distinct transition
behaviours when the metastable intermediate crystalline phases Atomic displacement
transformed into the next phase. First, it was found that the phases
shown in Fig. 2a,b make a fairly rapid transition, showing no
Surface nucleation and growth
nucleation inside the nanocrystal for the next new crystalline state.
Such a martensitic-type massive transformation can be achieved by
fast atomic displacement at an ångström scale if the phases before Figure 2 | Second series of in situ HREM images and corresponding FFTs
and after the transformation do not differ considerably from each of a LiFePO4 nanocrystal at 450 ◦ C. a–d, Both the HREM images and the
other. As mentioned in a recent review31 , the precise determination related FFTs demonstrate that the crystal structures differ from each other.
of the atomic structure in nanoscale materials remains challenging. For clear discrimination, the images and the FFTs are presented in different
Although there is currently no robust method to fully identify the colours (red, yellow and blue for the intermediate crystalline states). The
intermediate states shown in Fig. 2, we were able to demonstrate final state (green) shown in d was identified as an ordered olivine structure
their crystallographic relationship based on the HREM images. in the [122] projection. The inset in d shows a magnified image of the
HREM images, which are constructed by transmitted and diffracted region outlined by the square. A simulated HREM image in black and white
electron beams, afford crystallographic phase information along is also superimposed to confirm the ordered olivine structure. The image
with the amplitude information of the diffracted electrons25,32 . simulation was carried out under conditions of t = 40 Å (specimen
Therefore, the resulting two-dimensional atomic potentials (or thickness) and 1f = −60 Å (defocus length). An electron diffraction
atomic arrangement) can be determined accurately by electron pattern simulated under a dynamical condition is also presented with the
crystallography via HREM image processing23–25 . spot indices in d.

Figure 3 shows the projected potential contours reconstructed, a

44 min 30 s
using CRISP24,25,33–35 , from the HREM images in Fig. 2b,c. Each
inset also shows the black-and-white potential information
extracted from the images. On the basis of these image-processing
results, the plausible atomic arrays of Fe and P for each intermediate
crystalline phase can be estimated, although it is difficult to
determine the exact positions of Li and O owing to their low
atomic numbers. Red and yellow spheres superimposed on the
1 nm
contours denote Fe and P atoms, respectively. From the potential
contour maps, we can easily see that the linear connectivity between
Fe and P in Fig. 3a has been changed to a zigzag geometry in High
Fig. 3b, implying the ångström-scale displacement of atoms during
the transformation. Given the atomic potentials, corresponding
plane groups (or two-dimensional space groups) and the resultant
symmetry elements shown in Fig. 3 can be identified for these a 3.5 Å
Low
metastable phases: a p1m1 plane group (no. 3) for Fig. 3a and a b
p1g 1 plane group (no. 4) for Fig. 3b. Symbols and diagrams showing
each symmetry element in the unit cell for all 17 plane groups30 are b 45 min 0 s
summarized in Supplementary Information, Table S1 and Fig. S2.
The change of the atomic configuration into a p1g 1 plane group,
which contains glide planes instead of mirror planes parallel to
the a axis (Supplementary Information, Fig. S2), directly indicates
that the phases shown in Fig. 3 differ crystallographically from
each other. The plane groups and the parameters of each unit cell
outlined by a white rectangle in the contour maps for the phases are
1 nm
summarized in the table in Fig. 3.
In addition to the displacive transformation shown in Fig. 2a–c,
another distinguishable transition behaviour was observed during
the in situ analysis. This behaviour is characterized by nucleation
for a new crystalline phase and its subsequent growth, as generally
observed in phase-transition phenomena. Figure 4 shows a series
of real-time HREM images taken at intervals of 30 s for a heating
time between 45 min (Fig. 2c) and 48 min (Fig. 2d). In contrast to
a 3.1 Å
the transitions shown in Fig. 2b,c, the formation of new crystalline
states inside the nanocrystal can be observed, as denoted by the b
arrows in Fig. 4c. However, compared with the lattice image shown
in Figs 2d and 4f, the new crystalline phases nucleated in the Lattice parameters (Å) Plane groups
nanocrystal shown in Fig. 4c do not seem to be the ordered Phases (γ = 90°)
olivine structure that should finally form. This again confirms a b
the occurrence of multiple phase transformation and the resulting
presence of intermediate crystalline states. (a) Yellow 9.83 3.46 p1m1
Another noteworthy feature of Fig. 4 is that the new crystalline
(b) Blue 10.28 3.07 p1g1
phases seem to nucleate at the surface of the nanocrystal, as
indicated by the arrows in Fig. 4c. More detailed lattice fringes
between the surface and the interior bulk are compared in Fig. 4e. Figure 3 | Atomic potential contour maps calculated by crystallographic
A magnified lattice image for the surface region, outlined by a image processing. a,b, The HREM images (left) shown in Fig. 2b,c were
green square, and its FFT directly show that the surface region processed to obtain the potential information. Each inset shows the
has transformed into a stable olivine phase, consistent with potential information extracted from the images. Plausible atomic arrays
the simulated HREM image and the FFT shown in Fig. 2d. In for Fe (red spheres) and P (yellow spheres) are indicated on each map,
contrast, the bulk region, outlined by a blue square, remains in suggesting possible connectivity between the oxygen (grey spheres)
an intermediate crystalline state, as can be seen in the magnified octahedra and tetrahedra on the left. Crystallographic information for the
HREM image in Fig. 4e. Subsequent growth of the stable phase intermediate phases is listed in the table.
into the interior with time is observable in the magnified HREM
image in Fig. 4f, finally resulting in an ordered olivine structure in result verifies that the transition behaviours observed in this study
the [122] projection for the entire crystallographic configuration of are not unique to a specific crystallite, but instead prevail in most
the nanocrystal. crystals during the crystallization process. Although the FFTs of the
Multiple transformations such as this could also be observed in initial intermediate phases (red) in Figs 2a and 5a show different
other regions of the amorphous sample. Figure 5 shows another patterns from each other (see the comparison in Supplementary
series of in situ HREM images and their FFTs, further ensuring Information, Fig. S3a), the rapid transition to the other states based
the presence of the intermediate transient crystalline states. In on the atomic displacement was consistently similar between the
particular, when the FFTs in Fig. 2a–c are compared with those two series.
shown in Fig. 5a–c, it can be recognized that each intermediate The final intermediate crystalline state (blue, Fig. 5c) was also
phase in Fig. 5 is compatible with the metastable phases in Fig. 2. confirmed to transform into the stable olivine structure (green,
(For a direct and straightforward comparison, see Supplementary Fig. 5d) through surface nucleation and growth, the behaviour of
Information, Fig. S3. The a and b axes of the FFTs presented in which is analogous to that shown in Fig. 4. A detailed series of
both Figs 2a–c and 5a–c are aligned in the same directions.) This HREM images is presented in Supplementary Information, Fig. S4.

a b
45 min 0 s 45 min 30 s
(450 °C)
5 nm
c d
46 min 0 s 46 min 30 s
e f
47 min 0 s 47 min 30 s
1 nm
Figure 4 | Third series of in situ HREM images of a LiFePO4 nanocrystal at 450 ◦ C. a–f, Images taken at 30 s intervals. In contrast to Fig. 2a–c, this series
of images shows the transformation with the observable nucleation and subsequent growth of new phases inside the nanocrystal, as indicated by the
arrows in c and e. The magnified HREM images shown in e demonstrate that the interior bulk (blue square) remains in a metastable transient state,
compared with the stable surface region (green square). The enlargement shown in f and its FFT confirm that the nanocrystal finally transforms by growth
of a phase with an olivine structure.
A simulated electron diffraction pattern of LiFePO4 in the [010] the antisite defects in LiFePO4 with temperature37 , it is anticipated
zone direction is compared with the FFT in Fig. 5d. Although the that the final stable crystals shown in Figs 2d and 5d would have
relative intensities of the reflection spots are different, the good considerable cation intermixing between the M 1 and M 2 sites due
agreement between the FFT and the simulated pattern reveals to the much lower heating temperature and insufficient time to
that the nanocrystal (green) is of the olivine structure in the complete the ordering.
[010] projection. The present findings, which represent the first direct evidence at
Because the ionic radii of Li+ (0.76 Å) and Fe2+ (0.78 Å) in an atomic level of multiple transitions via intermediate metastable
a coordination octahedron of anions are similar to each other36 , crystalline phases in an inorganic compound, suggest a plausible
it is not likely that the two different cations are distributed in a pathway for the crystallization of LiFePO4 , as schematically
completely ordered manner onto octahedral interstitial sites of the represented in Fig. 6. The overall driving force for the crystallization
intermediate phases. In addition, on the basis of a recent study of to the finally stable ordered olivine structure is denoted by 1G. The

a b
39 min 30 s 42 min 0 s
5 nm
c d
44 min 30 s 48 min 30 s
200
+ 101
+
+
+
Figure 5 | Extra series of in situ HREM images of another LiFePO4 nanocrystal at 450 ◦ C. a–d, HREM images and corresponding FFTs presented in
different colours to discriminate the different crystal structures consistently, as shown in Fig. 2. The FFT of the final state (green) shown in d represents a
pattern that is identical to the simulated electron diffraction pattern of LiFePO4 in the [010] projection, confirming that the final state is of the olivine
structure. The simulation was carried under a dynamical condition. Plus signs in the simulated pattern denote nearly invisible spots with low intensity.
between nanocrystals, suggesting local variations, as compared in

ΔG2* ΔG4*
ΔG1* ΔG3* Supplementary Information, Fig. S3a.
Amorphous The observations in this study are based on HREM, which is
usually carried out in comparatively thin regions of a sample for
A better electron transparency under electron irradiation. Thus, in
Free energy
B reality, not all of the nanocrystals are predicted to show precisely

ΔG identical transition rates and activation energies from a statistical
C
standpoint. In addition, quantitative energy variations have not
been provided by which to judge how stable each intermediate
phase becomes after transformation in terms of the relative energy
Stable crystal stability. In this respect, statistical investigations accompanied
by quantitative simulations can be suggested to make further
Structure (or reaction time) progress in understanding the multiple phase transformation
during crystallization.
Figure 6 | Schematic diagram showing the transition pathway during the
crystallization of LiFePO4 . 1G indicates the total driving force for Methods
crystallization from an amorphous to a finally stable olivine-type crystal. A, Sample preparation. An amorphous LiFePO4 powder sample was prepared
using high-purity lithium carbonate (Li2 CO3 , 99.99%, Aldrich), iron(ii)
B and C represent metastable intermediate states with a local energy
oxalate dihydrate (Fe(ii)C2 O4 ·2H2 O, 99.99%, Aldrich) and ammonium
minimum, and each 1G∗ denotes the activation energy that should be dihydrogenphosphate (NH4 H2 PO4 , 99.999%, Aldrich). A small amount (2 mol%)
overcome for the transition to the next phase. of potassium carbonate (K2 CO3 , 99.995%, Aldrich) was also added as a flux for
ready formation of an amorphous phase at a low temperature. A stoichiometric
structure having the lowest energy barrier to transition crystallizes powder mixture of the three starting materials with the additive was ball-milled in
acetone for 24 h with zirconia milling media. After drying, the slurry was calcined
first, before transforming to another structure with the next lower at 350 ◦ C for 5 h in a stream of high-purity Ar (99.999%) at 400 s.c.c.m. The
energy state. Consistent with the Ostwald rule, the activation calcined powder sample was confirmed to be amorphous by transmission electron
energy 1G∗1 for the first transition to A is expected to be the microscopy (TEM) before in situ analysis.
lowest among the other activation energies, showing a sequence of
In situ HREM and crystallographic image processing. A transmission electron
1G∗1 < 1G∗2 < 1G∗3 < 1G∗4 . When the energy barrier is sufficiently microscope (JEM-ARM 1300S, JEOL Ltd) was used along with a hot-stage heating
low, it can be easily overcome by thermal fluctuation, leading to a holder (Gatan) for atomic-scale in situ observation at a high temperature. Images
fast transition via atomic displacement, as demonstrated in Figs 2 were recorded using a CCD (charge-coupled device) camera (2,048 × 2,048 pixels,
and 5. However, as a large activation is required, for example, SPUS1000 HV-IF, Gatan Inc.) mounted in the post-column high-voltage Gatan
in the case of 1G∗4 , a displacive transformation is no longer imaging filter. After heating at a rate of 10 ◦ C min−1 to 450 ◦ C, the samples were
stabilized for 20 min without electron irradiation to prevent drifting during the
likely to occur. Therefore, as shown in Fig. 4, the formation of observation. A two-dimensional Difference Filter (HREM Research Inc.) was used
energetically more stable nuclei that have overcome the energy to eliminate the background noise of the HREM images taken during the in situ
barrier along with their subsequent growth are favourable for experiment and thereby to obtain clearer FFTs along with accurate crystallographic
further transition. It is also noted that the first intermediate information. For image processing based on the electron crystallography with
crystalline phase from an amorphous solid is readily achieved by the lattice fringe images and their FFTs, CRISP (Calidris Inc.) was used to determine
the projected atomic potentials and estimate the plausible positions of Fe and P.
smallest reduction of energy along with the lowest energy barrier. During the data analysis, an optimal plane group for each image was determined,
Consequently, the initial transient phase corresponding to A in and thus the values of the difference of the amplitude of symmetry-related
Fig. 6 may not always be of the same crystallographic structure reflections (Rsym ) and the mean phase error (ϕRes ) are as low as possible.

Received 12 May 2008; accepted 24 October 2008; 22. Sutter, P. W. & Sutter, E. A. Dispensing and surface-induced crystallization of
published online 23 November 2008 zeptolitre liquid metal-alloy drops. Nature Mater. 6, 363–366 (2007).
23. Hovmöller, S., Sjögren, A., Farrants, G., Sundberg, M. & Marinder, B.-O.
Accurate atomic positions from electron microscopy. Nature 311,
References 238–241 (1984).
1. Burda, C, Chen, X., Narayanan, R. & El-Sayed, M. A. Chemistry and properties 24. Weirich, T. E., Ramlau, R., Simon, A., Hovmöller, S. & Zou, X. A crystal
of nanocrystals of different shape. Chem. Rev. 105, 1025–1102 (2005). structure determined with 0.02 Å accuracy by electron microscopy. Nature
2. Oxtoby, D. W. Homogeneous nucleation—theory and experiment. J. Phys. 382, 144–146 (1996).
Condens. Matter 4, 7627–7650 (1992). 25. Zou, X. & Hovmöller, S. Electron crystallography: Imaging and single-crystal
3. Oxtoby, D. W. Nucleation of first-order phase transitions. Acc. Chem. Res. 31, diffraction from powders. Acta Crystallogr. A 64, 149–160 (2008).
91–97 (1998). 26. Santoro, R. P. & Newmann, R. E. Antiferromagnetism in LiFePO4 . Acta
4. Mullin, J. W. Crystallization 4th edn 181–288 (Elsevier, 2004). Crystallogr. 22, 344–347 (1967).
5. Ostwald, W. Studien über die Bildung und Umwandlung fester Körper. 27. Papike, J. J. & Cameron, M. Crystal chemistry of silicate minerals of
Z. Phys. Chem. 22, 289–330 (1897). geophysical interest. Rev. Geophys. Space Phys. 14, 37–80 (1976).
6. Ten Wolde, P. R. & Frenkel, D. Enhancement of protein crystal nucleation by 28. Chung, S.-Y., Bloking, J. T. & Chiang, Y.-M. Electronically conductive
critical density fluctuations. Science 277, 1975–1978 (1997). phospho-olivines as lithium storage electrodes. Nature Mater. 1,
7. Vekilov, P. G. Dense liquid precursor for the nucleation of ordered solid 123–128 (2002).
phases from solution. Cryst. Growth Des. 4, 671–685 (2004). 29. Tang, P. & Holzwarth, N. A. W. Electronic structure of FePO4 , LiFePO4 , and
8. Auer, S. & Frenkel, D. Prediction of absolute crystal-nucleation rate in related materials. Phys. Rev. B 68, 165107 (2003).
hard-sphere colloids. Nature 409, 1020–1023 (2001). 30. International Tables for Crystallography 5th edn, Vol. A: Space-Group
9. Anderson, V. J. & Lekkerkerker, H. N. W. Insight into phase transition kinetics Symmetry (ed. Hahn, T.) (Springer, 2005).
from colloid science. Nature 416, 811–815 (2002). 31. Billinge, S. J. L. & Levin, I. The problem with determining atomic structure at
10. Sanz, E., Valeriani, C., Frenkel, D. & Dijkstra, M. Evidence for the nanoscale. Science 316, 561–565 (2007).
out-of-equilibrium crystal nucleation in suspensions of oppositely charged 32. Massa, W. Crystal Structure Determination 2nd edn (Springer, 2004).
colloids. Phys. Rev. Lett. 99, 055501 (2007). 33. Hovmöller, S. CRISP: crystallographic image processing on a personal
11. Anwar, J. & Boateng, P. K. Computer simulation of crystallization from computer. Ultramicroscopy 41, 121–135 (1992).
solution. J. Am. Chem. Soc. 120, 9600–9640 (1998). 34. Gramm, F. et al. Complex zeolite structure solved by combining powder
12. Lutsko, J. F. & Nicolas, G. Theoretical evidence for a dense fluid precursor to diffraction and electron microscopy. Nature 444, 79–81 (2006).
crystallization. Phys. Rev. Lett. 96, 046102 (2006). 35. Baerlocher, C. et al. Structure of the polycrystalline zeolite catalyst IM-5 solved
13. Yau, S.-T. & Vekilov, P. G. Quasi-planar nucleus structure in apoferritin by enhanced charge flipping. Science 315, 1113–1116 (2007).
crystallization. Nature 406, 494–497 (2000). 36. Shannon, R. D. Revised effective ionic radii and systematic studies of
14. Gasser, U., Weeks, E. R., Schofield, A., Pusey, P. N. & Weitz, D. A. Real-space interatomic distances in halides and chalcogenides. Acta Crystallogr. A 32,
imaging of nucleation and growth in colloidal crystallization. Science 292, 751–767 (1976).
258–262 (2001). 37. Chung, S.-Y., Choi, S.-Y., Yamamoto, T. & Ikuhara, Y. Atomic-scale
15. Gliko, O. et al. A metastable prerequisite for the growth of lumazine synthase visualization of antisite defects in LiFePO4 . Phys. Rev. Lett. 100, 125502 (2008).
crystals. J. Am. Chem. Soc. 127, 3433–3438 (2005).
16. Boerrigter, S. X. M. et al. In situ observations of epitaxial polymorphic Acknowledgements
nucleation of the model steroid methyl analogue 17 norethindrone. J. Phys. This work was supported by the Korea Research Foundation, grant no. 2007-331-D00196.
Chem. B 106, 4725–4731 (2002).
17. Stoica, C. et al. Epitaxial 2D nucleation of metastable polymorphs: A 2D Author contributions
version of Ostwald’s rule of stages. Cryst. Growth Des. 5, 975–981 (2005). S.-Y.C. conceived, designed and carried out the experiments, analysed the data,
18. Ross, F. M., Tersoff, J. & Tromp, R. M. Coarsening of self-assembled Ge interpreted and discussed the results and wrote the paper. Y.-M.K., J.-G.K. and Y.-J.K.
quantum dots on Si(001). Phys. Rev. Lett. 80, 984–987 (1998). provided the TEM facilities and contributed the acquisition of TEM data.
19. Hansen, P. L. et al. Atomic-resolved imaging of dynamic shape changes in
supported copper nanocrystals. Science 295, 2053–2055 (2002). Additional information
20. Helveg, S. et al. Atomic-scale imaging of carbon nanofibre growth. Nature 427, Supplementary Information accompanies this paper on www.nature.com/naturephysics.
426–429 (2004). Reprints and permissions information is available online at http://npg.nature.com/
21. Howe, J. M. & Saka, H. In situ transmission electron microscopy studies of the reprintsandpermissions. Correspondence and requests for materials should be addressed
solid–liquid interfaces. MRS Bull. 29, 951–957 (2004). to S.-Y.C.

ARTICLES
Navigability of complex networks

Marián Boguñá1 *, Dmitri Krioukov2 and K. C. Claffy2
Routing information through networks is a universal phenomenon in both natural and man-made complex systems. When each
node has full knowledge of the global network connectivity, finding short communication paths is merely a matter of distributed
computation. However, in many real networks, nodes communicate efficiently even without such global intelligence. Here, we
show that the peculiar structural characteristics of many complex networks support efficient communication without global
knowledge. We also describe a general mechanism that explains this connection between network structure and function.
This mechanism relies on the presence of a metric space hidden behind an observable network. Our findings suggest that real
networks in nature have underlying metric spaces that remain undiscovered. Their discovery should have practical applications
in a wide range of areas where networks are used to model complex systems.
N
etworks are ubiquitous in all domains of science and large networks. Our analysis reveals that, remarkably, real networks
technology, and permeate many aspects of daily human satisfy the topological conditions that maximize their navigability
life1–4 , especially on the rise of the information technology within this framework. Therefore, hidden metric spaces offer
society5,6 . Our growing dependence on them has inspired a burst explanations of two open problems in complex networks science:
of activity in the new field of network science, keeping researchers the communication efficiency networks so often exhibit and their
motivated to solve the difficult challenges that networks offer. unique structural characteristics.
Among these, the relation between network structure and function
is perhaps the most important and fundamental. Transport is one Node similarity and hidden metric spaces
of the most common functions of networked systems. Examples Our work is inspired by the seminal work of sociologist Stanley
can be found in many domains: transport of energy in metabolic Milgram on the small-world problem. The small-world paradigm
networks, of mass in food webs, of people in transportation refers to the existence of short chains of acquaintances among
systems, of information in cell signalling processes or of bytes across individuals in societies14 . At Milgram’s time, direct proof of such
the Internet. a paradigm was impossible owing to the lack of large databases
In many of these examples, routing—or signalling of informa- of social contacts, so Milgram conceived an experiment to analyse
tion propagation paths through a complex network maze—has the small-world phenomenon in human social networks. Randomly
a determinant role in the transport properties of the system, in chosen individuals in the United States were asked to route a letter
particular in systems such as the Internet or airport networks that to an unknown recipient using only friends or acquaintances that,
have transport as their primary function. The observed efficiency of according to their judgement, seemed most likely to know the
this routing process in real networks poses an intriguing question: intended recipient. The outcome of the experiment revealed that,
how is this efficiency achieved? When each element of the system without any global network knowledge, letters reached the target
has a full view of the global network topology, finding short routes recipient using, on average, 5.2 intermediate people, demonstrating
to target destinations is a well-understood computational process. that social acquaintance networks were indeed small worlds.
However, in many networks observed in nature, including those in The small-world property can be easily induced by adding a
society and biology (signalling pathways, neural networks and so small number of random connections to a ‘large world’ network15 .
on), nodes efficiently find intended communication targets even More striking is the fact that social networks are navigable without
though they do not possess any global view of the system. For global information. Indeed, the only information that people used
example, neural networks would not function so well if they could to make their routing decisions in Milgram’s experiment was a set
not route specific signals to appropriate organs or muscles in the of descriptive attributes of the destined recipient, such as place of
body, although no neurone has a full view of global inter-neurone living and occupation. People then determined who among their
connectivity in the brain. contacts was ‘socially closest’ to the target. The success of the
Here, we identify a general mechanism that explains routing experiment indicates that social distances among individuals—even
conductivity, or navigability of real networks based on the concept though they may be difficult to define mathematically—have a role
of similarity between nodes7–12 . Specifically, intrinsic characteristics in shaping the network architecture and that, at the same time,
of nodes define a measure of similarity between them, which we these distances can be used to navigate the network. However, it
abstract as a hidden distance. Taken together, hidden distances is not clear how this coupling between the structure and function
define a hidden metric space for a given network. Our recent of the network leads to efficiency of the search process, or what the
work shows that these spaces explain the observed structural minimum structural requirements are to facilitate such efficiency16 .
peculiarities of several real networks, in particular social and Here, we show how network navigability depends on the
technological ones13 . Here, we show that this underlying metric structural parameters characterizing the two most prominent
structure can be used to guide the routing process, leading to and common properties of real complex networks: (1) scale-
efficient communication without global information in arbitrarily free (power-law) node degree distributions characterizing the
1 Departament de Física Fonamental, Universitat de Barcelona, Martí i Franquès 1, 08028 Barcelona, Spain, 2 Cooperative Association for Internet Data
Analysis (CAIDA), University of California, San Diego (UCSD), 9500 Gilman Drive, La Jolla, California 92093, USA. *e-mail: marian.boguna@ub.edu.

heterogeneity in the number of connections that different nodes E Observable network topology
have and (2) clustering, a measure of the number of triangles in D
C A
the network topology. We assume the existence of a hidden metric F
B
space, an underlying geometric frame that contains all nodes of
the network, shapes its topology and guides routing decisions,
as illustrated in Fig. 1. Nodes are connected in the observable Hidden metric space
topology, but a full view of their global connectivity is not available
at any node. Nodes are also positioned in the hidden metric space
and identified by their coordinates in it. Distances between nodes
in this space abstract their similarity7–12 . These distances influence
both the observable topology and routing function: (1) the smaller
the distance between two nodes in the hidden space, that is, the
more similar the two nodes, the more likely they are connected
in the observable topology; (2) nodes also use hidden distances to
select, as the next hop, the neighbour closest to the destination in Figure 1 | How hidden metric spaces influence the structure and function
the hidden space. Kleinberg introduced the term greedy routing of complex networks. The smaller the distance between two nodes in the
to describe this forwarding process16 . Greedy routing and its hidden metric space, the more likely they are connected in the observable
modifications have been studied extensively in recent computer network topology. If node A is close to node B, and B is close to C, then A
science literature17–29 (see also Kleinberg’s review30 and references and C are necessarily close because of the triangle inequality in the metric
therein). However, most of these works do not study greedy routing space. Therefore, triangle ABC exists in the network topology with high
on scale-free topologies, which are known as the common signature probability, which explains the strong clustering observed in real complex
of many large-scale self-evolving complex networks1–3 . networks. The hidden space also guides the greedy-routing process: if node
We use the class of network models developed in recent A wants to reach node F, it checks the hidden distances between F and its
work13 . They generate networks with topologies similar to two neighbours B and C. Distance CF (green dashed line) is smaller than BF
those of real networks—small-world, scale-free and with strong (red dashed line); therefore, A forwards information to C. Node C then
clustering—and, simultaneously, with hidden metric spaces carries out similar calculations and selects its neighbour D as the next hop
lying underneath. The simplest model in this class uses a on the path to F. Node D is directly connected to F. The result is path
one-dimensional circle as the underlying metric space, in which A → C → D → F shown by green edges in the observable topology.
nodes are uniformly distributed. The model first assigns to each
node its expected degree k, drawn from a power-law degree metric strength: the more powerful is the influence of the network’s
distribution P(k) ∼ k −γ , with γ > 2, and then connects each pair of underlying metric space on the observable topology, the more
nodes with connection probability r(d;k,k 0 ) that depends both on strongly it is clustered.
the distance d between the two nodes in the circle and their assigned Although our toy model is not designed to exactly match any
degrees k and k 0 , specific real network, it generates graphs that are surprisingly
similar to some real networks, such as the Internet at the
r(d;k,k 0 ) ≡ r(d/dc ) = (1 + d/dc )−α , autonomous system level or the USA airport network. See
Supplementary Information for details.
where α > 1 and dc ∼ kk 0 , (1)
Navigability of modelled networks
which means that the probability of link connection between two We use the model to generate scale-free networks with different
nodes in the network decreases with the hidden distance between values of power-law degree distribution exponent γ and clustering
them (as ∼d −α ) and increases with their degrees (as ∼(kk 0 )α ). strength α, covering the observed values in most documented
These two properties have a clear interpretation. The connection complex networks1–3 . We then simulate greedy routing for a
cost increases with hidden distance, thus discouraging long-range large sample of paths on all generated networks, and compare
links. However, in making connections, rich (well-connected, high- the following two navigability parameters: (1) the average hop
degree) nodes care less about distances (connection costs) than poor length τ from source to destination of successful greedy-routing
nodes. Furthermore, the characteristic distance scale dc provides a paths and (2) the success ratio ps , defined as the percentage of
coupling between node degrees and hidden distances, and ensures successful paths. Unsuccessful paths are paths that get stuck at nodes
the following three topological characteristics that we commonly without neighbours closer to the destination in the hidden space
see in real networks. First, pairs of richly connected, high-degree than themselves. These nodes usually have small degrees. See the
nodes—hubs—are connected with high probability regardless of Methods section for simulation details.
the hidden distance between them because their characteristic Figure 2 shows the impact of the network’s degree distribution
distance dc is so large that any actual distance d between them and clustering on the average length τ of greedy-routing paths.
will be short in comparison: regardless of d, connection probability We observe a straightforward dependency: paths are shorter for
r in equation (1) is close to 1 if dc is large. Second, pairs of smaller exponents γ and stronger clustering (larger α values). The
low-degree nodes will not be connected unless the hidden distance dependency of the success ratio (the fraction of successful paths)
d between them is short enough to compare to the small value of ps on the two topology parameters γ and α is more intertwined.
their characteristic distance dc . Third, following similar arguments, Figure 3 shows that the effect of one parameter, γ , on the success
pairs composed of hubs and low-degree nodes are connected only ratio depends on the other parameter, the level of clustering. If
if they are located at moderate hidden distances. clustering is weak (low α), the percentage of successful paths decays
The parameter α in equation (1) determines the importance of with network size N regardless of the value of γ (Fig. 3a). However,
hidden distances for node connections. The larger α, the more with strong clustering (large α), the percentage of successful paths
preferred are connections between nodes close in the hidden space. increases with N and attains a maximum for large networks if
Consequently, the triangle inequality in the metric space leads to γ . 2.6, whereas it degrades for large networks if γ > 2.6 (Fig. 3b).
stronger clustering in the network, see Fig. 1. Clustering has a clear Figure 3c shows this effect for networks of the same size (N = 105 )
interpretation in our approach as a reflection of the network’s with different γ and α. The value of γ = 2.6 ± 0.1 maximizes the

a 8 b 15
α = 1.5 γ = 2.5 α = 1.5

7 α = 2.5 α = 2.5
α = 3.5 α = 3.5
10
6 α = 4.5 α = 4.5
τ
τ 5
5
4
γ = 2.2
3 0
103 104 105 2.0 2.2 2.4 2.6 2.8 3.0
N γ
Figure 2 | Average length of greedy-routing paths. a, The average hop length of successful paths, τ , as a function of the network size N for different values
of γ and α. Results for values of γ > 2.5 look similar but with longer paths and are omitted for clarity. In all cases, the path length grows polylogarithmically
with the network size: the observed values of τ are fitted well by τ (N) = A[logN]ν (solid lines), where A and ν are some constants. b, τ as a function of γ
and α for networks of fixed size N ≈ 105 . The effect of the two parameters on average path length is straightforward: paths are shorter for smaller
exponents γ and stronger clustering (larger α values).
a 0.2 b 0.7
0.6
γ = 2.1
γ = 2.2
0.1 0.5
ps
ps
γ = 2.6
γ = 2.3
γ = 2.7
γ = 2.4
γ = 2.8
γ = 2.5 0.4
γ = 2.9
γ = 3.0 α = 1.1 α = 5.0
0 0.3
102 103 104 105 102 103 104 105
N N
c d 3.0
Non-navigable region
0.6
2.5 Web of trust

0.4 α = 1.1
α = 1.5
ps
Metabolic
α = 2.0
α = 3.0 Internet
0.2
α = 5.0 2.0
Airports
Navigable region
0
2.0 2.2 2.4 2.6 2.8 3.0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
γ C
Figure 3 | Success probability of greedy routing. a,b, Success probability ps as a function of network size N for different values of γ with weak (a) and
strong (b) clustering. c, ps as a function of γ and α for networks of fixed size N ≈ 105 . d, Parameter α mapped to clustering coefficient C (ref. 15) by
computing C for each network with given γ and α. For each value of C, there is a critical value of γ = γc (C) such that the success ratio in networks with this
C and γ > γc (C) decreases with the network size (ps (N) −→ 0 as N → ∞), whereas ps (N) reaches a constant value for large N in networks with γ < γc (C).
The solid line in the plot shows these critical values γc (C), separating the low-γ , high-C navigable region, in which greedy routing remains efficient in the
large-graph limit, from the high-γ , low-C non-navigable region, where the efficiency of greedy routing degrades for large networks. The plot labels
measured values of γ and C for several real complex networks. Internet is the global Internet topology of autonomous systems as seen by the Border
Gateway Protocol39 ; Web of trust is the Pretty Good Privacy social network of mutual trust relationships40 ; Metabolic is the network of metabolic
reactions of E. coli41 ; and Airports is the network of the public air transportation system42 .
number of successful paths once clustering is above a threshold, exceeding γ ≈ 2.6—maximizing the percentage of successful paths.
α ≥ 1.5. These observations mean that for a fixed clustering We explain these findings in the next section, but we note
strength, there is a critical value of the exponent γ (Fig. 3d) below here that qualitatively, this navigable parameter region contains
which networks remain navigable as their size increases, but above most complex networks observed in reality1–3 , as confirmed in
which their navigability deteriorates with their size. Fig. 3d, where we juxtapose a few paradigmatic examples of
In summary, strong clustering improves both navigability communication, social, biological and transportation networks
metrics. We also find a delicate trade-off between values of versus the identified navigable region of clustering and degree
γ close to 2 minimizing path lengths, and higher values—not distribution exponent. Interestingly, power grids, which propagate

a
Toksook Bay
Bethel
Anchorage
Paris
Ibiza
Valencia
Detroit
b 3.0 c
¬4.0
Paris 2
2.5 ¬4.5
Detroit
¬5.0
Logarithm of airport degree
Logarithm of airport degree

2.0
¬5.5
Anchorage Valencia ¬6.0
1.5
Bethel
1 ¬6.5
1.0 ¬7.0
Ibiza
¬7.5
0.5 Toksook bay
¬8.0
¬8.5
9,000 8,000 7,000 6,000 5,000 4,000 3,000 2,000 1,000 0 4,032 3,000 2,000 1,000 69
Distance to Ibiza (km) Distance to destination (km)
Figure 4 | Greedy routing in the airport network. a, The structure of the single greedy-routing path from Toksook Bay to Ibiza (image from NASA). At each
intermediate airport, the next hop is the airport closest to Ibiza geographically. Sizes of symbols representing the airports are proportional to the logarithm
of their degrees. b, The changing distance to Ibiza versus the degree of the visited airports. c, The structure of greedy-routing paths between a collection of
airports in the USA (ref. 43). We include an airport pair in the collection if the distance between the airports is between 3,900 and 4,100 kilometres. The
number of airport pairs in this collection is 7,620. We use colour to indicate how often paths in the collection go through an airport of a given degree
located at a given geographical distance from the destination: blue/red indicates exponentially less/more visits to those airports, or more specifically, the
colour is the logarithm of the normalized density of visited airports.
electricity rather than route information, are neither scale-free travelling longer and longer distances per hop towards hubs,
nor clustered15,31 . thus ‘zooming out’ from the starting point. The second phase
corresponds to a fine-grained search, ‘zooming in’ onto the
Air travel by greedy routing as an explanation destination. The turning point between the two phases appears
We illustrate the greedy-routing function, and the structure of naturally: once we are in a hub near the destination, the probability
networks conductive to such routing, with an example of passenger that it is connected to a bigger hub closer to the destination sharply
air travel. Suppose we want to travel from Toksook Bay, Alaska, decreases, but at this point we do not need hubs anyway, and greedy
to Ibiza, Spain, by the public air transportation network. Nodes routing directs us to smaller airports at shorter distances next to
in this network are airports, and two airports are connected if the destination.
there is at least one flight between them. We travel according This zoom-out/zoom-in mechanism works efficiently only if the
to the greedy-routing strategy using geography as the underlying coupling between the airport network topology and the underlying
metric space. At each airport we choose the next-hop airport geography satisfies the following two conditions: the sufficient
geographically closest to the destination. Under these settings, our hubs condition and the sufficient clustering condition. The first
journey goes first to Bethel, then to Anchorage, to Detroit, over the condition ensures that a network has enough hub airports (high-
Atlantic to Paris, then to Valencia and finally to Ibiza, see Fig. 4. degree nodes) to provide an increasing sequence during the zoom-
The sequence and sizes of airport hops reveal the structure of our out phase. This condition is fulfilled by the real airport network and
greedy-routing path. The path proceeds from a small airport to by other scale-free networks with small values of degree distribution
a local hub at a small distance, from there to a larger hub at a exponent γ , because the smaller the γ , the larger the proportion of
larger distance, and so on until we reach Paris. At that point, when hubs in the network.
the distance to the destination becomes sufficiently small, greedy However, the presence of many hubs does not ensure that greedy
routing leads us closer to our final destination by choosing not routing will use them. Unlike humans, who can use their knowledge
another hub, but a less-connected neighbouring airport. of airport size to selectively travel via hub airports, greedy routing
We observe that the navigation process has two, somewhat uses only one constraint at each hop: minimize distance to the
symmetric phases. The first phase is a coarse-grained search, destination. Therefore, the network topology must satisfy the

1.0 1.0 ¬12 ¬11 ¬10 ¬9 ¬8 ¬7 ¬6 ¬5 ¬4 ¬3

α = 1.1, γ = 2.2 Colour: logarithm of the normalized
0.8 0.8 density of visited points
Pup(k,d)
Pup(k,d)
Logarithm of visited node’s degree

0.6 0.6 γ = 2.2, α = 5.0 γ = 2.2, α = 1.1
0.4 0.4
0.2 α = 5.0, γ =2.2 0.2
0 0
100 101 102 103 100 101 102 103
k k
1.0 1.0
α = 1.1, γ = 2.5
Source Target
0.8 0.8
Pup(k,d)
0.6 Pup(k,d) 0.6

Distance to destination
0.4 0.4
γ = 2.5, α = 5.0 γ = 2.5, α = 1.1
Logarithm of visited node’s degree

0.2 0.2
α = 5.0, γ =2.5
0 0
100 101 102 103 100 101 102 103
k k
1.0 1.0
α = 1.1, γ = 3.0
α = 5.0, γ =3.0
0.8 0.8
d = 100
Pup(k,d)
Pup(k,d)
0.6 0.6
d = 1,000
0.4 0.4 d = 10,000
d = 100,000
0.2 0.2
Distance to destination
0 0
100 101 102 103 100 101 102 103
k k Logarithm of visited node’s degree γ = 3.0, α = 5.0 γ = 3.0, α = 1.1
Figure 5 | Probability that greedy routing travels to higher-degree nodes.

The probability Pup (k,d) that the greedy-routing next hop after a node of
degree k located at distance d from a destination has higher degree k0 ≥ k
and is closer to the destination. The distance legend in the bottom right plot
applies to all of the plots. The results are for the large-graph limit N → ∞.
second condition, which ensures that Bethel is larger than Toksook

Bay, Anchorage larger than Bethel and so on. More generally, this
condition is that the next greedy hop from a remote low-degree Distance to destination
node probably has a higher degree, so that greedy paths typically
head first towards the highly connected network core. But the Figure 6 | The structure of greedy-routing paths. We visualize the results
network metric strength is exactly the required property: preference of our simulation of greedy routing in modelled networks with different
for connections between nodes nearby in the hidden space means values of γ and α observed in real complex networks. The hidden distance
that low-degree nodes are less likely to have connectivity to distant between the starting point and the destination is always approximately 104 ,
low-degree nodes; only high-degree nodes can have long-range and the network size N and number of attempted paths is always 105 for
connection that greedy routing will effectively select. The stronger each (γ ,α) combination, but the number of successful paths and path hop
this coupling between the metric space and topology (the higher lengths vary, see Figs 2, 3. All paths start and end at low-degree nodes
α in equation (1)), the stronger the clustering in the network. located, respectively, in the bottom left and bottom right corners of the
diagrams (see top left plot). For each (γ ,α), we show a single typical path
To illustrate, imagine an airport network without sufficient in black and, as in Fig. 4, use colour to indicate how often paths included a
clustering, one where the airport closest to our destination (Ibiza) node of a given degree located at a given distance from the destination. The
among all airports connected to our current node (Toksook Bay, simulations confirm that only when γ is small and α is large does the
Alaska) is not Bethel, which is bigger than Toksook Bay, but average path structure follow the zoom-out/zoom-in pattern that
Nightmute, Alaska, a nearby airport of comparable size to Toksook characterizes successful greedy routing in real networks, for example, in the
Bay. As greedy routing first leads us to Nightmute, then to another airport network in Fig. 4.
small nearby airport, and then to another, we can no longer get
to Ibiza in few hops. Worse, travelling via these numerous small calculate (in Supplementary Information) the probability that the
airports, we could reach one with no connecting flights heading next hop from a node of degree k located at hidden distance d
closer to Ibiza. Our greedy routing would be stuck at this airport from the destination has a larger degree k 0 > k, in which case the
with an unsuccessful path. path moves towards the high-degree network core, see Fig. 5. In
These factors explain why the most navigable topologies the most navigable case, with small degree-distribution exponent
correspond to scale-free networks with small exponents of the and strong clustering, the probability of increasing the node degree
degree distribution, that is, a large number of hubs, and with strong along the path is high at low-degree nodes, and sharply decreases to
clustering, that is, strong coupling between the hidden geometry zero after reaching a node of a critical degree value, which increases
and the observed topology. with distance d. This observation implies that greedy-routing paths
first propagate up to higher-degree nodes in the network core and
The structure of greedy-routing paths then exit the core towards low-degree destinations in the periphery.
We observe the discussed zoom-out/zoom-in mechanism in In contrast, with low clustering, paths are less likely to find higher-
analytical calculations and numerical simulations. Specifically, we degree nodes regardless of the distance to the destination. This

path structure violates the zoom-out/zoom-in pattern required for where d(θ,θ 0 ) is the geodesic distance between the two nodes on the circle and hki
efficient navigation. is the average degree. It can be shown that the average degree of nodes with hidden
Figure 6 shows the structure of greedy-routing paths in variable κ, k̄(κ), is proportional to κ (ref. 37). This proportionality guarantees
that the shape of the node degree distribution P(k) in generated networks is
simulations, further confirming our analysis. We again see that for γ −1
approximately the same as the shape of ρ(κ). The choice of ρ(κ) = (γ −1)κ0 κ −γ ,
small degree-distribution exponents and strong clustering (upper κ > κ0 ≡ (γ − 2)hki/(γ − 1), γ > 2, generates random networks with a power-law
left and middle left), the routing process quickly finds a way to degree distribution of the form P(k) ∼ k −γ , where γ is a model parameter
the high-degree core, makes a few hops there and then descends that regulates the heterogeneity of the degree distribution in the network. This
parameter abstracts the heterogeneity of node degrees in real networks, where
to a low-degree destination. In the other, non-navigable cases, the
degree distributions may not perfectly follow power laws, or may exhibit various
process can almost never get to the core of high-degree nodes. forms of high-degree cutoffs38,39 . The specific effects are less important than the
Instead, it wanders in the low-degree periphery, increasing the overall measure of heterogeneity. We note that instead of a circle in our model we
probability of getting lost at low-degree nodes. could use any isotropic space of any dimension13 .
Discussion Numerical simulations. Our model has three independent parameters: exponent
γ of power-law degree distributions, clustering strength α and average degree hki.
Our main motivation for this work comes from long-standing We fix hki to 6, which is roughly equal to the average degree of some real networks
scalability problems with the Internet routing architecture32 . To of interest39,40 , and vary γ ∈ [2.1,3] and α ∈ [1.1,5], covering their observed ranges
route information packets to a given destination, Internet routers in documented complex networks1–3 . For each (γ ,α) pair, we produce networks of
must communicate to maintain a coherent view of the global different sizes N ∈ [103 ,105 ] generating, for each (γ ,α,N ), a number of different
network instances—from 40 for large N to 4,000 for small N . In each network
Internet topology. The constantly increasing size and dynamics instance G, we randomly select 106 source–destination pairs (a, b) and execute
of the Internet thus leads to immense and quickly growing the greedy-routing process for them starting at a and selecting, at each hop h, the
communication and information processing overhead, a major next hop as the h’s neighbour in G closest to b in the circle. If for a given (a, b),
bottleneck in routing scalability33 causing concerns among Internet this process visits the same node twice, then the corresponding path leads to a loop
experts that the existing Internet routing architecture may not and is unsuccessful. We then average the measured values of path hop lengths τ
and percentage of successful paths ps across all pairs (a, b) and networks G for the
sustain even another decade32 . Discovery of the Internet’s hidden same (γ ,α,N ). Note that we are not concerned with the absolute values of the
metric space would remove this bottleneck, eliminating the need success ratio ps . Instead we use it as a measure of navigability to compare networks
for the inherently unscalable communication of topology changes. with different (γ ,α,N ). For this purpose, we could use the success ratio of any
Instead routers would be able to just forward packets greedily to the (improved) modification of standard greedy routing.
destination based on hidden distances.
Shortest path versus shortest time. All results derived in the present paper
In a similar manner, reconstruction of hidden metric spaces are about finding short paths across a network topology. The total physical
underlying other real networks may prove practically useful. For time from source to destination is implicitly assumed to be proportional to the
example, in social or communication networks (for example, the number of hops. In real transportation systems, such as the Internet or the airport
Web, overlay or online social networks) hidden spaces would yield network, the finite capacity of nodes implies that the end-to-end path latency
efficient strategies for searching specific individuals or content may be longer when intermediate nodes are congested. Although our results
most cleanly apply to uncongested systems, there are obvious modifications,
based only on local knowledge. The metric spaces hidden under such as choosing the second- or third-nearest rather than the nearest neighbour,
some biological networks (such as neural, gene regulatory networks, that could still find nearly shortest paths while reducing and balancing load on
signalling or even protein folding34,35 pathways) can become a the system.
powerful tool in studying the structure of information or signal
flows in these networks, enabling investigation of such processes Received 12 March 2008; accepted 26 September 2008;
without detailed global knowledge of the network structure published online 16 November 2008
or organization. References
The natural question we thus face is how to proceed towards 1. Albert, R. & Barabási, A.-L. Statistical mechanics of complex networks. Rev.
discovery of the explicit structure of hidden metric spaces Mod. Phys. 74, 47–97 (2002).
underlying real networks. We do not expect spaces underlying 2. Newman, M. E. J. The structure and function of complex networks. SIAM Rev.
different networks to be exactly the same. For example, the 45, 167–256 (2003).
3. Dorogovtsev, S. N. & Mendes, J. F. F. Evolution of Networks: From Biological
similarity spaces of Web pages9 and Wikipedia editors11 probably Nets to the Internet and WWW (Oxford Univ. Press, 2003).
differ. However, the main contribution of this work establishes 4. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M. & Hwang, D.-U. Complex
the general mechanisms behind navigability of scale-free, strongly networks: Structure and dynamics. Phys. Rep. 424, 175–308 (2006).
clustered topologies that characterize many different real networks. 5. Castells, M. The Rise of the Network Society (Blackwell, 1996).
The next step is to find the common properties of hidden spaces 6. Pastor-Satorras, R. & Vespignani, A. Evolution and Structure of the Internet. A
Statistical Physics Approach (Cambridge Univ. Press, 2004).
that render them congruent with these mechanisms. Specifically, 7. Watts, D. J., Dodds, P. S. & Newman, M. E. J. Identity and search in social
we are interested in what geometries of hidden spaces lead to networks. Science 296, 1302–1305 (2002).
such congruency36 . 8. Girvan, M. & Newman, M. E. J. Community structure in social and biological
In general, we believe that the present and future work on networks. Proc. Natl Acad. Sci. USA 99, 7821–7826 (2002).
hidden metric spaces and network navigability will deepen our 9. Menczer, F. Growing and navigating the small world Web by local content.
Proc. Natl Acad. Sci. USA 99, 14014–14019 (2002).
understanding of the fundamental laws describing relationships 10. Leicht, E. A., Holme, P. & Newman, M. E. J. Vertex similarity in networks.
between structure and function of complex networks. Phys. Rev. E 73, 026120 (2006).
11. Crandall, D., Cosley, D., Huttenlocher, D., Kleinberg, J. & Suri, S. Proc. ACM
Methods Int. Conf. on Knowledge Discovery and Data Mining (ACM, 2008).
A model with the circle as a hidden metric space. In our model, we place all 12. Clauset, A., Moore, C. & Newman, M. E. J. Hierarchical structure and the
nodes on a circle by assigning them a random variable θ, that is, their polar angle, prediction of missing links in networks. Nature 453, 98–101 (2008).
distributed uniformly in [0,2π). The circle radius R grows linearly with the total 13. Ángeles Serrano, M., Krioukov, D. & Boguñá, M. Self-similarity of complex
number of nodes N , 2πR = N , to keep the average density of nodes on the circle networks and hidden metric spaces. Phys. Rev. Lett. 100, 078701 (2008).
fixed to 1. We next assign to each node its expected degree κ drawn from some 14. Travers, J. & Milgram, S. An experimental study of the small world problem.
distribution ρ(κ). The connection probability between two nodes with hidden Sociometry 32, 425–443 (1969).
15. Watts, D. J. & Strogatz, S. H. Collective dynamics of small-world networks.
coordinates (θ ,κ) and (θ 0 ,κ 0 ) takes the form
Nature 393, 440–442 (1998).
16. Kleinberg, J. Navigation in a small world. Nature 406, 845 (2000).
d(θ ,θ 0 ) −α

(α − 1)
r(θ ,κ;θ 0 ,κ 0 ) = 1 + , µ= , 17. Kleinberg, J. STOC ’00: Proc. Thirty-Second Annual ACM Symp. on Theory of
µκκ 0 2hki Computing 163–170 (ACM, 2000).

18. Kleinberg, J. M. in Small-World Phenomena and the Dynamics of Information 33. Krioukov, D., claffy, kc, Fall, K. & Brady, A. On compact routing for the
(eds Dietterich, T. G., Becker, S. & Ghahramani, Z.) 431–438 Internet. Comput. Commun. Rev. 37, 41–52 (2007).
(MIT Press, 2001). 34. Ravasz, E., Gnanakaran, S. & Toroczkai, Z. Network structure of protein
19. Manku, G. S., Naor, M. & Wieder, U. Proc. 36th ACM Symp. on Theory of folding pathways. Preprint at <http://arxiv.org/abs/0705.0912> (2007).
Computing (STOC) 54–63 (ACM, 2004). 35. Toroczkai, Z., Kozma, B., Bassler, K. E., Hengartner, N. W. & Korniss, G.
20. Martel, C. 23rd ACM Symp. Principles of Distributed Computing (PODC) Gradient networks. J. Phys. A: Math. Theor. 41, 155103 (2008).
179–188 (ACM, 2004). 36. Krioukov, D., Papadopoulos, F., Boguñá, M. & Vahdat, A. Efficient navigation
21. Nguyen, V. & Martel, C. SODA ’05: Proc. Sixteenth Annual ACM-SIAM Symp. in scale-free networks embedded in hyperbolic metric spaces. Preprint at
on Discrete Algorithms 311–320 (Society for Industrial and Applied <http://arxiv.org/abs/0805.1266> (2008).
Mathematics, 2005). 37. Boguñá, M. & Pastor-Satorras, R. Class of correlated random networks with
22. Nguyen, V. & Martel, C. Proc. Fourth Workshop on Analytic Algorithmics and hidden variables. Phys. Rev. E 68, 036112 (2003).
Combinatorics 213–227 (SIAM, 2008). 38. Giot, M. et al. A protein interaction map of Drosophila melanogaster. Science
23. Simsek, Ö. & Jensen, D. in Proc. Nineteenth Int. Joint Conf. in Artificial 302, 1727–1736 (2003).
Intelligence (eds Kaelbling, L. P. & Saffiotti, A.) 304–310 (Professional Book 39. Mahadevan, P. et al. The Internet AS-level topology: Three data sources and
Center, 2005). one definitive metric. Comput. Commun. Rev. 36, 17–26 (2006).
24. Lebhar, E. & Schabanel, N. in Proc. 31st Int. Colloquium on Automata, 40. Boguñá, M., Pastor-Satorras, R., Díaz-Guilera, A. & Arenas, A. Models of social
Languages and Programming Vol. 3142 (eds Díaz, J., Karhumäki, J., Lepistö, A. networks based on social distance attachment. Phys. Rev. E 70, 056122 (2004).
& Sannella, D.) 894–905 (Lecture Notes in Computer Science, Springer, 2004). 41. Jeong, H., Tombor, B., Albert, R., Oltvai, Z. N. & Barabási, A.-L. The
25. Fraigniaud, P. in Proc of the 13th Annual European Symposium Vol. 3669 large-scale organization of metabolic networks. Nature 407, 651–654 (2000).
(eds Brodal, G. S. & Leonardi, S.) 791–802 (Lecture Notes in Computer 42. Barrat, A., Barthélemy, M., Pastor-Satorras, R. & Vespignani, A. The
Science, Springer, 2005). architecture of complex weighted networks. Proc. Natl Acad. Sci. USA 101,
26. Fraigniaud, P., Gavoille, C. & Paul, C. Eclecticism shrinks even small worlds. 3747–3752 (2004).
Distrib. Comput. 18, 279–291 (2006). 43. <http://www.transtats.bts.gov/>
27. Fraigniaud, P., Gavoille, C., Kosowski, A., Lebhar, E. & Lotker, Z. in Proc.
Nineteenth Annual ACM Symp. on Parallel Algorithms and Architectures
(eds Gibbons, P. B. & Scheideler, C.) 1–7 (ACM, 2007). Acknowledgements
28. Fraigniaud, P. & Gavoille, C. in Proc. Twentieth Annual Symposium on We thank M. Á. Serrano for useful comments and discussions. This work was
Parallelism in Algorithms and Architectures (eds auf der Heide, F. M. & supported in part by DGES grant FIS2007-66485-C02-02, Generalitat de Catalunya
Shavit, N.) 62–69 (ACM, 2008) SPAA. grant No. SGR00889, the Ramón y Cajal program of the Spanish Ministry of Science,
29. Chaintreau, A., Fraigniaud, P. & Lebhar, E. in Automata, Languages and by NSF CNS-0434996 and CNS-0722070, by DHS N66001-08-C-2029 and by
Programming Vol. 5125 (eds Aceto, L. et al.) 133–144 (Lecture Notes in Cisco Systems.
Computer Science, Springer, 2008).
30. Kleinberg, J. Complex networks and decentralized search algorithms. Proc. Int.
Congr. Math. (ICM) 3, 1019–1044 (2006). Additional information
31. Amaral, L. A. N., Scala, A., Barthélemy, M. & Stanley, H. E. Classes of Supplementary Information accompanies this paper on www.nature.com/naturephysics.
small-world networks. Proc. Natl Acad. Sci. 97, 11149–11152 (2000). Reprints and permissions information is available online at http://npg.nature.com/
32. Meyer, D., Zhang, L. & Fall, K. (eds). RFC4984 (The Internet Architecture reprintsandpermissions. Correspondence and requests for materials should be addressed
Board, 2007). to M.B.

Phsctanu 1 Re

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Phsctanu 1 Re

Uploaded by

Copyright:

Available Formats

editorial

2009 and all that

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 1

© 2009 Macmillan Publishers Limited. All rights reserved

2009 and all that

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 1

© 2009 Macmillan Publishers Limited. All rights reserved

2 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 3

© 2009 Macmillan Publishers Limited. All rights reserved

Know the unknown

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 5

© 2009 Macmillan Publishers Limited. All rights reserved

The autist aesthete

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 7

© 2009 Macmillan Publishers Limited. All rights reserved

Lucky for some

Modern Art (MoMA) is

© Jasper Johns / Licenced by VaGa, new york, ny

8 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

Lucky for some

Modern Art (MoMA) is

© Jasper Johns / Licenced by VaGa, new york, ny

8 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

© istockphoto / Felix Moeckel

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 9

© 2009 Macmillan Publishers Limited. All rights reserved

© istockphoto / AmAndA Rohde

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 11

© 2009 Macmillan Publishers Limited. All rights reserved

Electric spin orchestra

12 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

Electric spin orchestra

12 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

How do your crystals grow?

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 13

© 2009 Macmillan Publishers Limited. All rights reserved

How do your crystals grow?

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 13

© 2009 Macmillan Publishers Limited. All rights reserved

14 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

Vortices wiggled and dragged

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 15

© 2009 Macmillan Publishers Limited. All rights reserved

fundamental research on atomic physics,

16 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

fundamental research on atomic physics,

16 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 17

© 2009 Macmillan Publishers Limited. All rights reserved

18 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

© 2009 Macmillan Publishers Limited. All rights reserved

18 nature physics | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics

Measurement-based quantum computation

NATURE PHYSICS | VOL 5 | JANUARY 2009 | www.nature.com/naturephysics 19

Device Area (µm2 ) RN (k) EC (K) EJ /EC