Here’s my latest Crucible column for Chemistry World. A techie one, but no harm in that. I also have a feature on nanobubbles in this (September) issue, and will try to stick that up, in extended form, on my website soon.
___________________________________________________________________
Bartosz Grzybowski of Northwestern University in Illinois, who has already established himself as one of the most inventive current practitioners of the chemical art, has unveiled a ‘chemo-informatic’ scheme called Chematica that can stake a reasonable claim to being paradigm-changing. He and his colleagues have spent years assembling the transformations linking chemical species into a vast network that codifies and organizing the known pathways through chemical space. Each node of the network is either a molecule or element, or a chemical reaction. Links connect reactants and products via the nexus of a known reaction. The full network contains around 7 million compound nodes and about the same number of reaction nodes. Grzybowski calls it a “collective chemical brain.”
I predict a mixed reaction from chemists. On the one hand the potential value of such a tool for discovering improved or entirely new synthetic pathways to drugs, materials and other useful products is tremendous, and has already been illustrated by Grzybowski’s team. On the other hand, Chematica seems to imply that chemistry is indeed, as the old jibe puts it, just cookery, and is something now better orchestrated by computer than by chemists.
I’ll come back to that. First let’s look at what Chematica is. Grzybowski first described the network in 2005 [1], when he was mostly concerned with its topological properties rather than with chemical insights. Like the Internet or some social networks, the chemical network has ‘scale-free’ connectivity, meaning that the distribution of nodes with different degrees of connectivity n is a power law: the number of nodes with n links is proportional to n(exp-α), where α is a constant. This means that a few very highly connected nodes are the hubs that bind the network together and provide shortcuts. The same structure is also found in the reaction network of compounds in metabolic pathways.
In a trio of new papers the researchers have now started to put the network to use. In the first, they perform an automated trawl for new one-pot reactions that can replace existing multi-step syntheses [2]. The advantages of single-step processes are obvious: no laborious separation and purification of products after each step, with consequent reductions in yield. Identifying potential one-pot processes linking molecular nodes that hitherto lacked a direct connection here means subjecting the relevant reactions to several filtering steps that check for compatibility – for example, checking that a water-solvated synthesis will not unintentionally hydrolyse functional groups. This filtering is painstaking in principle, but very quick once automated.
It is one thing to demonstrate that such one-pot syntheses are possible in principle, but Grzybowski and colleagues have ensured that at least some of those identified work in practice. Specifically, they looked for syntheses of quinoline-based molecules – common components of drugs and dyes – and thiophenes, which have useful electronic and optical properties. Many of the new pathways worked with high yields, in some cases demonstrably higher than those of alternative multi-step syntheses. Some false positives arise from errors in the literature used to build the network.
Another use of Chematica is to optimize existing syntheses – something previously reliant on manual or inexhaustive semi-automated searches. Looking for improved – basically, cheaper – routes to a given target is a matter of stepping progressively backwards from that molecule to preceding intermediates [3]. An algorithm can calculate the costs of all such steps in the network, working recursively backwards to a specified ‘depth’ (maximum number of synthetic steps) and finding the cheapest option. Applied to syntheses conducted by Grzybowski’s company ProChimia, Chematica offered potential savings of up to 45 percent if instituted for 51 of the company’s targets. The greatest the number of targets, the greater the savings because of the economies of shared ingredients and intermediates.
Finally, and perhaps most controversially, the researchers show how Chematica can be used to identify threats of chemical-weapons manufacture by terrorists [4]. The network can be searched for routes to harmful substances such as nerve agents using unregulated ingredients. Of course, it can also disclose such routes, but as with viral genomic data [5], open access to such data should be the best antidote to the risks they inherently pose.
Does all this, then, mean that synthetic organic chemists are about to be automated? The usual response is to insist that computers will never match human creativity. But that defence is looking increasingly under threat in, say, chess, maths and perhaps even music and visual art. In some ways chemical synthesis is as rule-bound as music if not chess, and thus ripe for an algorithmic approach. Perhaps at least some of the beauty rightly attributed to classic syntheses should be seen as illustrating human ingenuity in the face of tasks for which no better solution then existed. Synthetic schemes designed by humans surely won’t become obsolete any time soon – but there seems no harm in acknowledging that the time may come when the art and creativity of chemistry resides more solidly in our decisions of what to make, and why, than in how we make it.
References
1. M. Fialkowski, K. J. M. Bishop, V. A. Chubukov, C. J. Campbell & B. A. Grzybowski, Angew. Chem. Int. Ed. 44, 7263 (2005).
2. C. M. Gothard et al., Angew. Chem. Int. Ed. online publication 10.1002/anie.201202155 (2012).
3. M. Kowalik et al., Angew. Chem. Int. Ed. online publication 10.1002/anie.201202209 (2012).
4. P. E. Fuller, C. M. Gothard, N. A. Gothard, A. Wieckiewicz & B. A. Grzybowski, Angew. Chem. Int. Ed. online publication 10.1002/anie.201202210 (2012).
5. M. Imai et al., Nature 10.1038/nature10831 (2012).
Friday, August 31, 2012
Friday, August 24, 2012
Computers get emotional
Here’s more on Iamus, the computer programme that writes music you might actually want to listen to. It is published in Nature this week, although there are more references in this version. The CD of Iamus’ music is released in mid-September (not the 1st, as the Nature piece says), and is simply called Iamus (Melomics Records). I’m looking forward to that.
________________________________________________________________
If a computer can produce something that moves us, would this take artificial intelligence beyond an important threshold? That’s one of the questions seemingly raised by an algorithm called Iamus, developed by Francisco Vico and colleagues at the University of Malaga, which composes music from scratch.
There’s nothing new about computers making music. They have been used by composers since the early days of computer technology in the 1960s, most notably by the experimental Greek composer Iannis Xenakis. Neither is there anything mysterious about an algorithmic approach to composition: most music is highly rule-bound, even formulaic, and so lends itself to that. An early attempt at computer music in the 1980s, the programme called CHORAL devised by computer scientist Kemal Ebcioglu to harmonise chorales in the style of J. S. Bach [1], drew on well formulated principles of harmony and melody that were observed by Bach himself [2].
But it’s one thing to slavishly follow the rules, quite another to come up with original melodies and harmonic progressions that will captivate and even move the listener – especially if the composing is being performed without any human input. Computer scientists have been quite successful in making programmes that can learn from human examples. Early improvisational algorithms such as the jazz-inflected GenJam [3], devised by computer scientist John Biles, and GenBebop, a system created by cognitive scientists Lee Spector and Adam Alpern to produce solos in the style of Charlie Parker [4], gave indifferent results even by their creators’ admission. But Continuator, a programme devised by François Pachet at Sony’s Computer Science Laboratory in Paris, is much more convincing [5]. In a kind of ‘Turing test’ where Continuator alternated with an improvising human pianist and elaborated on his suggestions, expert listeners had trouble telling man and machine apart.
Contrary to common perception, however, improvising is fairly rule-bound too, so it’s not hard to see how human-derived musical material can be plausibly mutated and elaborated in an automated way. Coming up with the raw musical ideas is a much harder task for a computer. Before now, efforts to achieve this have been distinctly underwhelming, sounding like bland pastiche all too evidently shaped by clichéd harmonic progressions and melodic structures.
This is where Iamus’ creators claim to have something new. The algorithm, named after the legendary son of Apollo who could understand the language of birds, is inspired by Darwinian evolution. The computer generates very simple ‘musical genomes’ , rather like little motifs, which are then evolved, mutated and elaborated until they acquire genuine musical content and interest [6]. Genetic and evolutionary algorithms for making music are also not new (see, for example, ref. 7), but Iamus seems capable of dramatic invention: the music is far more than just a succession of transparent variations. Vico and colleagues, in collaboration with composer and pianist Gustavo Díaz-Jerez of the Conservatory of the Basque Country in San Sebastian, recently recorded some of Iamus’s scores with leading musicians, including two orchestral pieces played by the London Symphony Orchestra, for release in September. They broadcast a live performance of two of these pieces from the University of Malaga in July to commemorate the 100th anniversary of Alan Turing’s birth.
The recorded compositions are all in a modernist classical style – full of dissonance, but with hints of harmony and rich textures that might put a listener in mind of composers such as Gyorgy Ligeti and Krzysztof Penderecki. But the same approach can be used for other idioms too, and Vico and colleagues say that it might supply a cheap, convenient way of generating music for commercial purposes.
The willingness of professionals to perform the works marks out Iamus as unique. The LSO’s chairman Lennox Mackenzie was impressed with what it had achieved but felt that the music still fell short of that by good human composers. It struck him as “going nowhere” – a complaint often made of other modernist works – yet ultimately achieving an “epic” quality. Many of the other musicians were pleasantly surprised by the material, and found some of it genuinely expressive.
Which brings us to the initial question. If Iamus can simulate (and thus stimulate) emotionality, is it not just ‘thinking’ in the limited sense of the Turing test but apparently displaying human characteristics?
Here we should heed studies of music cognition which have shown that emotion in music is not some deeply mysterious process but has its own rules and regularities [8]. For example, certain musical structures, such as ‘false trails’ that create and then confound expectations, or judicious injections of dissonance followed by resolution, can elicit emotion quite reliably [9]. This should be no surprise to anyone whose emotions have been helplessly manipulated by formulaic film scores.
What’s more, the involvement of human performers is vital. Music lovers know very well that the same piece can be performed in a dry, unengaged manner or with heart-rending fervour. Good performers achieve expression with a wide range of ‘tricks’, such as subtle distortions of tempo, intonation and timbre [10].
Iamus’ work might therefore be considered to demonstrate the often neglected role of performer and listener in ‘making music’. The nineteenth-century Romantic tradition has fostered a deep-seated belief in the inherent genius of the composer, as though he or she has imbued the very notes with passion. It’s not to deny the undoubted sensitivity and skill of the greatest composers to say that a composition only becomes music in the mind of the listener, through the interaction of the composer’s and the performer’s choices with the wealth of learning and association that even allegedly ‘unskilled’ listeners possess.
This consideration ought also to diminish an inherent prejudice (evident in the critical responses to Iamus so far) against computer-composed music. Neuroscientists Stefan Koelsch and Nikolaus Steinbeis have shown that part of this prejudice is unconscious: the same piece of music may or may not activate parts of the brain associated with ascribing intention to others, depending on whether listeners have been told that the piece was composed by a human or by computer [11]. It’s possible that human performance of computer-made music might at least partly override this obstacle to emotional engagement. But we should also celebrate the way that Iamus, far from threatening our supposedly unique claim to creativity, can put the audience back in the picture as a participant in the creative act.
References
1. Ebcioğlu, K. Proc. 1984 Int. Computer Music Conf., 135-144, held in IRCAM, Paris. Computer Music Association, San Francisco, 1985.
2. Rohrmeier, M. & Cross, I. (2008), in Proc. 10th Int. Conf. on Music Percept. Cognit. (ICMPC 2008), Sapporo, Japan.
3. Biles, J. A., in Proc. 1994 Int. Computer Music Conf., 131-137. International Computer Music Association, San Francisco, 1994.
4. Spector, L. & Alpern, A., in Proc. Twelfth Natl Conf. Artificial Intelligence, AAAI-94, 3-8. AAAI Press/MIT Press, Menlo Park CA and Cambridge MA (1994).
5. Pachet, F., J. New Music Res. 32, 333-341 (2003).
6. Díaz-Jerez, G. Leonardo Music J. 21, 13-14 (2011).
7. MacCallum, R. M., Mauch, M., Burt, A. & Leroi, A. M. Proc. Natl Acad. Sci. USA 10.1073/pnas.1203182109 (2012).
8. Sloboda, J. A. The Musical Mind: The Cognitive Psychology of Music. Clarendon Press, Oxford, 1985.
9. Juslin, P. N. & Sloboda, J. A. (eds). Music and Emotion. Oxford University Press, Oxford, 2001.
10. Meyer, L. B. Emotion and Meaning in Music. University of Chicago Press, Chicago, 1956.
11. Steinbeis, N. & Koelsch, S. Cerebral Cortex 19, 619-623 (2009).
________________________________________________________________
If a computer can produce something that moves us, would this take artificial intelligence beyond an important threshold? That’s one of the questions seemingly raised by an algorithm called Iamus, developed by Francisco Vico and colleagues at the University of Malaga, which composes music from scratch.
There’s nothing new about computers making music. They have been used by composers since the early days of computer technology in the 1960s, most notably by the experimental Greek composer Iannis Xenakis. Neither is there anything mysterious about an algorithmic approach to composition: most music is highly rule-bound, even formulaic, and so lends itself to that. An early attempt at computer music in the 1980s, the programme called CHORAL devised by computer scientist Kemal Ebcioglu to harmonise chorales in the style of J. S. Bach [1], drew on well formulated principles of harmony and melody that were observed by Bach himself [2].
But it’s one thing to slavishly follow the rules, quite another to come up with original melodies and harmonic progressions that will captivate and even move the listener – especially if the composing is being performed without any human input. Computer scientists have been quite successful in making programmes that can learn from human examples. Early improvisational algorithms such as the jazz-inflected GenJam [3], devised by computer scientist John Biles, and GenBebop, a system created by cognitive scientists Lee Spector and Adam Alpern to produce solos in the style of Charlie Parker [4], gave indifferent results even by their creators’ admission. But Continuator, a programme devised by François Pachet at Sony’s Computer Science Laboratory in Paris, is much more convincing [5]. In a kind of ‘Turing test’ where Continuator alternated with an improvising human pianist and elaborated on his suggestions, expert listeners had trouble telling man and machine apart.
Contrary to common perception, however, improvising is fairly rule-bound too, so it’s not hard to see how human-derived musical material can be plausibly mutated and elaborated in an automated way. Coming up with the raw musical ideas is a much harder task for a computer. Before now, efforts to achieve this have been distinctly underwhelming, sounding like bland pastiche all too evidently shaped by clichéd harmonic progressions and melodic structures.
This is where Iamus’ creators claim to have something new. The algorithm, named after the legendary son of Apollo who could understand the language of birds, is inspired by Darwinian evolution. The computer generates very simple ‘musical genomes’ , rather like little motifs, which are then evolved, mutated and elaborated until they acquire genuine musical content and interest [6]. Genetic and evolutionary algorithms for making music are also not new (see, for example, ref. 7), but Iamus seems capable of dramatic invention: the music is far more than just a succession of transparent variations. Vico and colleagues, in collaboration with composer and pianist Gustavo Díaz-Jerez of the Conservatory of the Basque Country in San Sebastian, recently recorded some of Iamus’s scores with leading musicians, including two orchestral pieces played by the London Symphony Orchestra, for release in September. They broadcast a live performance of two of these pieces from the University of Malaga in July to commemorate the 100th anniversary of Alan Turing’s birth.
The recorded compositions are all in a modernist classical style – full of dissonance, but with hints of harmony and rich textures that might put a listener in mind of composers such as Gyorgy Ligeti and Krzysztof Penderecki. But the same approach can be used for other idioms too, and Vico and colleagues say that it might supply a cheap, convenient way of generating music for commercial purposes.
The willingness of professionals to perform the works marks out Iamus as unique. The LSO’s chairman Lennox Mackenzie was impressed with what it had achieved but felt that the music still fell short of that by good human composers. It struck him as “going nowhere” – a complaint often made of other modernist works – yet ultimately achieving an “epic” quality. Many of the other musicians were pleasantly surprised by the material, and found some of it genuinely expressive.
Which brings us to the initial question. If Iamus can simulate (and thus stimulate) emotionality, is it not just ‘thinking’ in the limited sense of the Turing test but apparently displaying human characteristics?
Here we should heed studies of music cognition which have shown that emotion in music is not some deeply mysterious process but has its own rules and regularities [8]. For example, certain musical structures, such as ‘false trails’ that create and then confound expectations, or judicious injections of dissonance followed by resolution, can elicit emotion quite reliably [9]. This should be no surprise to anyone whose emotions have been helplessly manipulated by formulaic film scores.
What’s more, the involvement of human performers is vital. Music lovers know very well that the same piece can be performed in a dry, unengaged manner or with heart-rending fervour. Good performers achieve expression with a wide range of ‘tricks’, such as subtle distortions of tempo, intonation and timbre [10].
Iamus’ work might therefore be considered to demonstrate the often neglected role of performer and listener in ‘making music’. The nineteenth-century Romantic tradition has fostered a deep-seated belief in the inherent genius of the composer, as though he or she has imbued the very notes with passion. It’s not to deny the undoubted sensitivity and skill of the greatest composers to say that a composition only becomes music in the mind of the listener, through the interaction of the composer’s and the performer’s choices with the wealth of learning and association that even allegedly ‘unskilled’ listeners possess.
This consideration ought also to diminish an inherent prejudice (evident in the critical responses to Iamus so far) against computer-composed music. Neuroscientists Stefan Koelsch and Nikolaus Steinbeis have shown that part of this prejudice is unconscious: the same piece of music may or may not activate parts of the brain associated with ascribing intention to others, depending on whether listeners have been told that the piece was composed by a human or by computer [11]. It’s possible that human performance of computer-made music might at least partly override this obstacle to emotional engagement. But we should also celebrate the way that Iamus, far from threatening our supposedly unique claim to creativity, can put the audience back in the picture as a participant in the creative act.
References
1. Ebcioğlu, K. Proc. 1984 Int. Computer Music Conf., 135-144, held in IRCAM, Paris. Computer Music Association, San Francisco, 1985.
2. Rohrmeier, M. & Cross, I. (2008), in Proc. 10th Int. Conf. on Music Percept. Cognit. (ICMPC 2008), Sapporo, Japan.
3. Biles, J. A., in Proc. 1994 Int. Computer Music Conf., 131-137. International Computer Music Association, San Francisco, 1994.
4. Spector, L. & Alpern, A., in Proc. Twelfth Natl Conf. Artificial Intelligence, AAAI-94, 3-8. AAAI Press/MIT Press, Menlo Park CA and Cambridge MA (1994).
5. Pachet, F., J. New Music Res. 32, 333-341 (2003).
6. Díaz-Jerez, G. Leonardo Music J. 21, 13-14 (2011).
7. MacCallum, R. M., Mauch, M., Burt, A. & Leroi, A. M. Proc. Natl Acad. Sci. USA 10.1073/pnas.1203182109 (2012).
8. Sloboda, J. A. The Musical Mind: The Cognitive Psychology of Music. Clarendon Press, Oxford, 1985.
9. Juslin, P. N. & Sloboda, J. A. (eds). Music and Emotion. Oxford University Press, Oxford, 2001.
10. Meyer, L. B. Emotion and Meaning in Music. University of Chicago Press, Chicago, 1956.
11. Steinbeis, N. & Koelsch, S. Cerebral Cortex 19, 619-623 (2009).
Why selfishness still doesn’t pay
Here’s my latest news story for Nature.
____________________________________________________________________
A recent finding that undermines conventional thinking on the evolution of cooperation doesn’t, after all, prevent altruistic behaviour from emerging.
For the past several decades, one of the central results of game theory has seemed to be that self-interest can drive social cooperation, because in the long term selfish behaviour hurts you as much as your competitors. Last May, two leading physicists, William Press of the University of Texas and Freeman Dyson of the Institute of Advanced Study in Princeton, argued otherwise. They showed how, in the classic ‘game’ from which cooperation seems to evolve, called the Prisoner’s Dilemma, it’s possible to be successfully selfish [1].
This apparently revolutionary idea has now been challenged by two evolutionary biologists at Michigan State University in East Lansing. In a preprint [2], Christoph Adami and Arend Hintze say that the strategy proposed by Press and Dyson is “evolutionarily unstable”. In a population of agents all seeking for the best way to play the Prisoner’s Dilemma, those using the new selfish strategy will eventually be bested by more generous players.
The Prisoner’s Dilemma is a simple ‘game’ that captures the fundamental problem faced by a population of organisms competing for limited resources: the temptation to cheat or freeload. You might do better acting together (cooperating) than alone, but the temptation is to let others put in the effort or face the risks while sharing yourself in the rewards.
In the Prisoner’s Dilemma as it was formulated by researchers in the 1950s, two prisoners accused of a crime are questioned. If one helps convict the other by testifying against him, he is offered a lighter sentence. But if both testify against the other, their sentences will be heavier than if both refuse to do so.
In a single ‘round’ of this game, it always makes sense to ‘defect’ – to shop the other guy. That way you’re better off whatever your opponent does. But if the game is played again and again – if you have repeated opportunities to cheat on the other player – you both do better to cooperate. This so-called iterated Prisoner’s Dilemma has been used to show how cooperation could arise in selfish populations: those genetically disposed to cooperate will be more successful than those predisposed to defect.
But what’s the best way to play the iterated game in a population of individuals using many different strategies? In the 1980s, political scientist Robert Axelrod tried to answer that question by staging computerized tournaments, inviting anyone to submit a strategy and then pitching them all against one another in many one-to-one bouts.
The winner was a very simple strategy called Tit-for-Tat (TfT), which merely copies its opponent’s behaviour from the last round. If the opponent defected in the last round, TfT does so in the current one. Against cooperators, TfT always cooperates; against defectors, it always defects. It is, in effect, ‘tough but fair’. The moral message seemed reassuring: it pays to be nice, but nastiness should be punished.
However, in further studies it became clear that TfT might not always dominate in evolutionary games where the most successful strategies are propagated from generation to generation. Slightly more forgiving strategies, which don’t get caught in cycles of mutual recrimination by a single mistaken defection, can do better in the long run. In fact, there is no single best way to play the game – it depends on your opponents. Nonetheless, the iterated Prisoner’s Dilemma seemed to explain how cooperation between unrelated individuals might evolve: why some animals hunt in packs and why we have altruistic instincts.
Press and Dyson seemed to shatter this cosy picture. They showed that there exists a class of strategies, which for technical reasons they call zero-determinant (ZD) strategies, in which one player can force the other to accept a less-than-equal share of the ‘payoff’ in the Prisoner’s Dilemma. In effect, the victim has to either grit his teeth and accept this unfair division, or punish the other player at a greater cost to himself. This turns the game into a different one, known as an Ultimatum game, in which one player is presented with the choice of either accepting an unequal distribution of the payoff or, if he refuses, both players losing out.
It turns out that TfT is just a special case of these ZD strategies in which the payoffs happen to be equal. Like a TfT player, a ZD player bases his next choice of cooperate/defect on what happened in the last round: it is said to be a ‘memory-one’ strategy. But instead of being rigidly deterministic – this previous outcome dictates that choice – it is probabilistic: the choice to cooperate or defect is made with a certain probability for each of the four possible outcomes of the last round. A judicious choice of these probabilities enables one player to control the payoff that the other receives.
According to William Poundstone, author of the 1992 book The Prisoner’s Dilemma, “The Press-Dyson finding directly challenges the two notions at the heart of the Prisoner’s Dilemma – that you can't fool evolution, and that the most successful strategies are fair strategies.” Nonetheless, Press says that “Freeman's and my paper has been warmly received by Prisoner’s Dilemma experts. More than one has expressed regret at not having discovered the ZD strategies previously.”
“The paper did indeed cause quite a stir, because the main result appeared to be completely new, despite intense research in this area for the last 30 years”, says Adami. It wasn’t totally new, however – in 1997 game theorists Karl Sigmund of the University of Vienna and Martin Nowak of Harvard University discovered strategies that similarly allow one player to fix the other’s payoff at a specified level [3]. But they admit that “we didn’t know about the vast and fascinating realm of zero-determinant strategies.” The work of Press and Dyson “opens a new facet in the study of trigger strategies and folk theorems for iterated games, and offers a highly stimulating approach for moral philosophers,” they say.
The ZD strategies are not as dispiriting as perhaps they sound, says Press, because they allow a new balance to be found if both players understand the principles. “Once both players understand ZD, then each has the power to set the other’s score, independent of the other’s actions. This allows them to make an enforceable treaty, not possible using simpler strategies.”
In other words, the ZD strategy forces players to reflect, to think ahead, to consider the opponent’s point of view, and not just try to get the highest possible score. It “allows for a whole range of careful, deliberative negotiations”, Press says. “This is a world in which diplomacy trumps conflict.”
But now Adami and Hintze say that this world might not exist – or not for long. They find that, in an evolutionary iterated Prisoner’s Dilemma game in which the prevalence of particular strategies depends on their success, ZD players are soon out-competed by others using more common strategies, and so they will evolve to become non-ZD players themselves. That’s because ZD players suffer from the same problem as habitual defectors: they do badly against their own kind.
There’s one exception: ZD players can persist if they can figure out whether they are playing another ZD player or not. Then they can exploit the advantages of ZD strategies against non-ZD players, but will switch to a more advantageous non-ZD strategy when faced with their own kind.
That in turn means, however, that non-ZD players could gain the upper hand by using strategies that look like ZD but are not, thus fooling ZD players into abandoning their extortionate strategy. This could lead to the same kind of ‘arms race’ seen in some kinds of biological mimicry, where a harmless species evolves to look like a harmful one, while the harmful one tries to evolve away from its imitator.
Sigmund and Nowak, with colleague Christian Hilbe, have also shown in work not yet published that the ZD strategy is evolutionarily unstable, but can pave the way for the emergence of cooperators from a more selfish community. “ZD strategies do not establish a strong foothold in the population”, says Sigmund.
Game theorist and economist Samuel Bowles at the Santa Fe Institute in New Mexico feels that these results demote the interest of the ZD strategies. “The question of their evolutionary stability is critical, and the paper makes their limitations clear. Because they are not evolutionarily stable, I’d call them merely a curiosity of little interest to evolutionary biology or any of the other biological sciences.”
Adami is not so sure that they won’t be fund in the wild. “We don’t usually have nearly enough information about animal decisions”, he says. “But in my experience, anything that is imaginable has probably evolved somewhere, sometime. To gather conclusive evidence about it is a whole different matter.”
Could they be found in human society too? “It’s not inconceivable”, says Adami, “but we have to keep in mind that humans very rarely make decisions based only on their and their opponent's last move. It is much more likely that this type of strategy is in use in automated trading programs, such as those involved in high-frequency trading of stocks and commodities. However, because these programs are usually secret, we wouldn't know about it.”
References
1. Press, W. H. & Dyson, F. J. Proc. Natl Acad. Sci. USA 10.1073/pnas.1206569109 (2012).
2. Adami, C. & Hintze, A. preprint http://www.arxiv.org/abs/1208.2666 (2012).
3. Nowak, M. A., Boerlijst, M. C. & Sigmund, K., Am. Math. Soc. Monthly 104, 303-307 (1997).
____________________________________________________________________
A recent finding that undermines conventional thinking on the evolution of cooperation doesn’t, after all, prevent altruistic behaviour from emerging.
For the past several decades, one of the central results of game theory has seemed to be that self-interest can drive social cooperation, because in the long term selfish behaviour hurts you as much as your competitors. Last May, two leading physicists, William Press of the University of Texas and Freeman Dyson of the Institute of Advanced Study in Princeton, argued otherwise. They showed how, in the classic ‘game’ from which cooperation seems to evolve, called the Prisoner’s Dilemma, it’s possible to be successfully selfish [1].
This apparently revolutionary idea has now been challenged by two evolutionary biologists at Michigan State University in East Lansing. In a preprint [2], Christoph Adami and Arend Hintze say that the strategy proposed by Press and Dyson is “evolutionarily unstable”. In a population of agents all seeking for the best way to play the Prisoner’s Dilemma, those using the new selfish strategy will eventually be bested by more generous players.
The Prisoner’s Dilemma is a simple ‘game’ that captures the fundamental problem faced by a population of organisms competing for limited resources: the temptation to cheat or freeload. You might do better acting together (cooperating) than alone, but the temptation is to let others put in the effort or face the risks while sharing yourself in the rewards.
In the Prisoner’s Dilemma as it was formulated by researchers in the 1950s, two prisoners accused of a crime are questioned. If one helps convict the other by testifying against him, he is offered a lighter sentence. But if both testify against the other, their sentences will be heavier than if both refuse to do so.
In a single ‘round’ of this game, it always makes sense to ‘defect’ – to shop the other guy. That way you’re better off whatever your opponent does. But if the game is played again and again – if you have repeated opportunities to cheat on the other player – you both do better to cooperate. This so-called iterated Prisoner’s Dilemma has been used to show how cooperation could arise in selfish populations: those genetically disposed to cooperate will be more successful than those predisposed to defect.
But what’s the best way to play the iterated game in a population of individuals using many different strategies? In the 1980s, political scientist Robert Axelrod tried to answer that question by staging computerized tournaments, inviting anyone to submit a strategy and then pitching them all against one another in many one-to-one bouts.
The winner was a very simple strategy called Tit-for-Tat (TfT), which merely copies its opponent’s behaviour from the last round. If the opponent defected in the last round, TfT does so in the current one. Against cooperators, TfT always cooperates; against defectors, it always defects. It is, in effect, ‘tough but fair’. The moral message seemed reassuring: it pays to be nice, but nastiness should be punished.
However, in further studies it became clear that TfT might not always dominate in evolutionary games where the most successful strategies are propagated from generation to generation. Slightly more forgiving strategies, which don’t get caught in cycles of mutual recrimination by a single mistaken defection, can do better in the long run. In fact, there is no single best way to play the game – it depends on your opponents. Nonetheless, the iterated Prisoner’s Dilemma seemed to explain how cooperation between unrelated individuals might evolve: why some animals hunt in packs and why we have altruistic instincts.
Press and Dyson seemed to shatter this cosy picture. They showed that there exists a class of strategies, which for technical reasons they call zero-determinant (ZD) strategies, in which one player can force the other to accept a less-than-equal share of the ‘payoff’ in the Prisoner’s Dilemma. In effect, the victim has to either grit his teeth and accept this unfair division, or punish the other player at a greater cost to himself. This turns the game into a different one, known as an Ultimatum game, in which one player is presented with the choice of either accepting an unequal distribution of the payoff or, if he refuses, both players losing out.
It turns out that TfT is just a special case of these ZD strategies in which the payoffs happen to be equal. Like a TfT player, a ZD player bases his next choice of cooperate/defect on what happened in the last round: it is said to be a ‘memory-one’ strategy. But instead of being rigidly deterministic – this previous outcome dictates that choice – it is probabilistic: the choice to cooperate or defect is made with a certain probability for each of the four possible outcomes of the last round. A judicious choice of these probabilities enables one player to control the payoff that the other receives.
According to William Poundstone, author of the 1992 book The Prisoner’s Dilemma, “The Press-Dyson finding directly challenges the two notions at the heart of the Prisoner’s Dilemma – that you can't fool evolution, and that the most successful strategies are fair strategies.” Nonetheless, Press says that “Freeman's and my paper has been warmly received by Prisoner’s Dilemma experts. More than one has expressed regret at not having discovered the ZD strategies previously.”
“The paper did indeed cause quite a stir, because the main result appeared to be completely new, despite intense research in this area for the last 30 years”, says Adami. It wasn’t totally new, however – in 1997 game theorists Karl Sigmund of the University of Vienna and Martin Nowak of Harvard University discovered strategies that similarly allow one player to fix the other’s payoff at a specified level [3]. But they admit that “we didn’t know about the vast and fascinating realm of zero-determinant strategies.” The work of Press and Dyson “opens a new facet in the study of trigger strategies and folk theorems for iterated games, and offers a highly stimulating approach for moral philosophers,” they say.
The ZD strategies are not as dispiriting as perhaps they sound, says Press, because they allow a new balance to be found if both players understand the principles. “Once both players understand ZD, then each has the power to set the other’s score, independent of the other’s actions. This allows them to make an enforceable treaty, not possible using simpler strategies.”
In other words, the ZD strategy forces players to reflect, to think ahead, to consider the opponent’s point of view, and not just try to get the highest possible score. It “allows for a whole range of careful, deliberative negotiations”, Press says. “This is a world in which diplomacy trumps conflict.”
But now Adami and Hintze say that this world might not exist – or not for long. They find that, in an evolutionary iterated Prisoner’s Dilemma game in which the prevalence of particular strategies depends on their success, ZD players are soon out-competed by others using more common strategies, and so they will evolve to become non-ZD players themselves. That’s because ZD players suffer from the same problem as habitual defectors: they do badly against their own kind.
There’s one exception: ZD players can persist if they can figure out whether they are playing another ZD player or not. Then they can exploit the advantages of ZD strategies against non-ZD players, but will switch to a more advantageous non-ZD strategy when faced with their own kind.
That in turn means, however, that non-ZD players could gain the upper hand by using strategies that look like ZD but are not, thus fooling ZD players into abandoning their extortionate strategy. This could lead to the same kind of ‘arms race’ seen in some kinds of biological mimicry, where a harmless species evolves to look like a harmful one, while the harmful one tries to evolve away from its imitator.
Sigmund and Nowak, with colleague Christian Hilbe, have also shown in work not yet published that the ZD strategy is evolutionarily unstable, but can pave the way for the emergence of cooperators from a more selfish community. “ZD strategies do not establish a strong foothold in the population”, says Sigmund.
Game theorist and economist Samuel Bowles at the Santa Fe Institute in New Mexico feels that these results demote the interest of the ZD strategies. “The question of their evolutionary stability is critical, and the paper makes their limitations clear. Because they are not evolutionarily stable, I’d call them merely a curiosity of little interest to evolutionary biology or any of the other biological sciences.”
Adami is not so sure that they won’t be fund in the wild. “We don’t usually have nearly enough information about animal decisions”, he says. “But in my experience, anything that is imaginable has probably evolved somewhere, sometime. To gather conclusive evidence about it is a whole different matter.”
Could they be found in human society too? “It’s not inconceivable”, says Adami, “but we have to keep in mind that humans very rarely make decisions based only on their and their opponent's last move. It is much more likely that this type of strategy is in use in automated trading programs, such as those involved in high-frequency trading of stocks and commodities. However, because these programs are usually secret, we wouldn't know about it.”
References
1. Press, W. H. & Dyson, F. J. Proc. Natl Acad. Sci. USA 10.1073/pnas.1206569109 (2012).
2. Adami, C. & Hintze, A. preprint http://www.arxiv.org/abs/1208.2666 (2012).
3. Nowak, M. A., Boerlijst, M. C. & Sigmund, K., Am. Math. Soc. Monthly 104, 303-307 (1997).
Thursday, August 23, 2012
Are we not reproducing enough?
Here is my Crucible column from the August issue of Chemistry World. This topic seems to be becoming a big deal, as witnessed for example by the creation of this initiative for replication of results from the PLoS journals – I note that their advisory board includes Brian Nosek, whose work I mention below.
________________________________________________________________
How much of the published literature should you believe? Not much, by some accounts. A 2005 paper by epidemiologist John Ioannidis of the University of Ioannina School of Medicine in Greece had the stark title “Why most published research findings are false” [1]. Ioannidis claimed that “for most study designs and settings, it is more likely for a research claim to be false than true”, and that often published claims simply reflected the prevailing bias of the field. Ioannidis suspected that some “established classics” in the literature wouldn’t stand up to close scrutiny.
His focus was on biomedical research, in particular clinical trials of drugs, where inferences have to be made from complex statistics, perhaps with small sample sizes. Here, not only might the effects being sought be rather marginal but there are also strong biases and prejudices introduced by financial pressures. Reports of drug trials certainly do have a bias towards positive outcomes, prompting valid calls for all drug trials to be registered before the study is undertaken so that negative findings can’t be quietly dropped.
These problems with pharmaceutical research are in themselves troubling for some chemists. But is this mostly an issue for Big Pharma, with its distorting profit motives and its reliance on statistics rather than more reductive, step-by-step experimentation? Probably not, Daniel Sarewitz of the Consortium for Science, Policy and Outcomes at Arizona State University claimed in Nature last May [2]. According to Sarewitz, systematic error due to bias, whether conscious or not, is “is likely to be prevalent in any field that seeks to predict the behaviour of complex systems – economics, ecology, environmental science, epidemiology and so on”. This figures: all these fields tend to depend on statistical inference of often marginal effects operating through mechanisms that may be poorly understood and perhaps nigh impossible to delineate.
But what about the subjects we like to think of as the “hard sciences” – like most of chemistry? Surely you can place more trust in spectra and rate constants and crystal structures than in scatter plots? Perhaps – but ‘trust’ is often what it is. Not many studies are ever repeated verbatim, and it’s generally acknowledged that crystallographic databases are probably full of errors, if only minor. The chance of experiments being replicated is probably proportional to the significance of the results. Maybe the greater good doesn’t suffer much from a literature full of flawed but uninteresting work – but that would offer scant support for science’s supposedly self-correcting nature.
And problems do crop up on close examination. Take, for example, the recent attempt by Darragh Crotty and colleagues at Trinity College Dublin to replicate the claims of Russian biochemist Anatoly Buchachenko and his coworkers, who since 2004 have been documenting (in good journals) the influence of a weak magnetic field on the rate of enzymatic production of ATP [3]. The Russians report that millitesla magnetic fields can more than double the reaction rate when the phosphorylating enzymes contain 25Mg (which is magnetic) rather than the other two stable isotopes 24Mg and 26Mg. Crotty and colleagues set out to test this because it bore on controversial claims of physiological effects from weak electromagnetic fields. They found no difference in reaction rate for all three magnesium isotopes [4]. So far the discrepancy remains puzzling.
If this is indeed a wider problem than is commonly recognized for all sciences, what to do? Sarewitz suggests reducing hype and strengthening ties between fundamental research and real-world testing. Ioannidis implores researchers to be honest with themselves about the ‘pre-study odds’ of their hypothesis being true. This purging of preconception and self-deception is what Francis Bacon called for in the seventeenth century when he argued that natural philosophers seeking truth must free themselves from ‘idols of the mind’. But as Ioannis recognizes, changing mindsets isn’t easy.
Another perspective is offered in a preprint by psychologist Brian Nosek of the University of Virginia and his colleagues [5]. They point out that professional success for scientists relies on publishing, but publication both favours positive results and prefers novelty over replication. What is needed is a way to rescue scientists’ ostensible aim – getting it right – from their short-term, pragmatic aim – getting it published. Among things that won’t work, the authors say, are journals devoted to replications and tougher peer review (which can already display stifling conservatism). Instead we need metrics for evaluating what is worth replicating, journal editorial policies that focus on soundness rather than ‘importance’, less focus on sheer publication productivity for job and tenure applicants, lower barriers to publication (so that it becomes less coveted in itself), and in particular, new ways of releasing results: open access to data, methods, tools and lab books. One can find problems with all of these, but the old ways of science publishing are looking increasingly archaic and flawed. What have we got to hide?
1. J. P. A. Ioannis, PLoS Med. 2, e124 (2005).
2. D. Sarewitz, Nature 485, 149 (2012).
3. A. L. Buchachenko & D. A. Kuznetsov, J. Am. Chem. Soc. 130, 12868-12869 (2008).
4. D. Crotty et al., Proc. Natl Acad. Sci. USA 109, 1437-1442 (2012).
5. B. A. Nosek, J. R. Spies & M. Motyl, preprint http://arxiv.org/abs/1205.4251.
________________________________________________________________
How much of the published literature should you believe? Not much, by some accounts. A 2005 paper by epidemiologist John Ioannidis of the University of Ioannina School of Medicine in Greece had the stark title “Why most published research findings are false” [1]. Ioannidis claimed that “for most study designs and settings, it is more likely for a research claim to be false than true”, and that often published claims simply reflected the prevailing bias of the field. Ioannidis suspected that some “established classics” in the literature wouldn’t stand up to close scrutiny.
His focus was on biomedical research, in particular clinical trials of drugs, where inferences have to be made from complex statistics, perhaps with small sample sizes. Here, not only might the effects being sought be rather marginal but there are also strong biases and prejudices introduced by financial pressures. Reports of drug trials certainly do have a bias towards positive outcomes, prompting valid calls for all drug trials to be registered before the study is undertaken so that negative findings can’t be quietly dropped.
These problems with pharmaceutical research are in themselves troubling for some chemists. But is this mostly an issue for Big Pharma, with its distorting profit motives and its reliance on statistics rather than more reductive, step-by-step experimentation? Probably not, Daniel Sarewitz of the Consortium for Science, Policy and Outcomes at Arizona State University claimed in Nature last May [2]. According to Sarewitz, systematic error due to bias, whether conscious or not, is “is likely to be prevalent in any field that seeks to predict the behaviour of complex systems – economics, ecology, environmental science, epidemiology and so on”. This figures: all these fields tend to depend on statistical inference of often marginal effects operating through mechanisms that may be poorly understood and perhaps nigh impossible to delineate.
But what about the subjects we like to think of as the “hard sciences” – like most of chemistry? Surely you can place more trust in spectra and rate constants and crystal structures than in scatter plots? Perhaps – but ‘trust’ is often what it is. Not many studies are ever repeated verbatim, and it’s generally acknowledged that crystallographic databases are probably full of errors, if only minor. The chance of experiments being replicated is probably proportional to the significance of the results. Maybe the greater good doesn’t suffer much from a literature full of flawed but uninteresting work – but that would offer scant support for science’s supposedly self-correcting nature.
And problems do crop up on close examination. Take, for example, the recent attempt by Darragh Crotty and colleagues at Trinity College Dublin to replicate the claims of Russian biochemist Anatoly Buchachenko and his coworkers, who since 2004 have been documenting (in good journals) the influence of a weak magnetic field on the rate of enzymatic production of ATP [3]. The Russians report that millitesla magnetic fields can more than double the reaction rate when the phosphorylating enzymes contain 25Mg (which is magnetic) rather than the other two stable isotopes 24Mg and 26Mg. Crotty and colleagues set out to test this because it bore on controversial claims of physiological effects from weak electromagnetic fields. They found no difference in reaction rate for all three magnesium isotopes [4]. So far the discrepancy remains puzzling.
If this is indeed a wider problem than is commonly recognized for all sciences, what to do? Sarewitz suggests reducing hype and strengthening ties between fundamental research and real-world testing. Ioannidis implores researchers to be honest with themselves about the ‘pre-study odds’ of their hypothesis being true. This purging of preconception and self-deception is what Francis Bacon called for in the seventeenth century when he argued that natural philosophers seeking truth must free themselves from ‘idols of the mind’. But as Ioannis recognizes, changing mindsets isn’t easy.
Another perspective is offered in a preprint by psychologist Brian Nosek of the University of Virginia and his colleagues [5]. They point out that professional success for scientists relies on publishing, but publication both favours positive results and prefers novelty over replication. What is needed is a way to rescue scientists’ ostensible aim – getting it right – from their short-term, pragmatic aim – getting it published. Among things that won’t work, the authors say, are journals devoted to replications and tougher peer review (which can already display stifling conservatism). Instead we need metrics for evaluating what is worth replicating, journal editorial policies that focus on soundness rather than ‘importance’, less focus on sheer publication productivity for job and tenure applicants, lower barriers to publication (so that it becomes less coveted in itself), and in particular, new ways of releasing results: open access to data, methods, tools and lab books. One can find problems with all of these, but the old ways of science publishing are looking increasingly archaic and flawed. What have we got to hide?
1. J. P. A. Ioannis, PLoS Med. 2, e124 (2005).
2. D. Sarewitz, Nature 485, 149 (2012).
3. A. L. Buchachenko & D. A. Kuznetsov, J. Am. Chem. Soc. 130, 12868-12869 (2008).
4. D. Crotty et al., Proc. Natl Acad. Sci. USA 109, 1437-1442 (2012).
5. B. A. Nosek, J. R. Spies & M. Motyl, preprint http://arxiv.org/abs/1205.4251.
Still talking about colour
Here’s another take on the recent paper on modelling of the evolution of colour terms – this time, published in Prospect.
___________________________________________
Languages are extremely diverse, but not arbitrary. Behind the bewildering diversity and the apparently contradictory ways in which different tongues elect to conceptualise the world, we can sometimes discern order and regularity. Many linguists have assumed that this reflects a hard-wired linguistic aptitude of the human brain. Some recent studies propose, however, that language ‘universals’ aren’t simply prescribed by genes but arise from the interaction between the biology of human perception and the bustle, exchange and negotiation of human culture.
Language has a perfectly logical job to do—to convey information— and yet is seemingly riddled with irrationality. Why all those irregular verbs, those random genders, those silent vowels and ambiguous homophones? You’d think languages would evolve towards some optimal model of clarity and concision, but instead they accumulate quirks that hinder learning, not only for foreigners but also in native speakers.
Traditionally, linguists have tended to explain the peculiarities of language through the history of the people who speak it. That’s often fascinating, but does not yield general principles about how languages have developed in the past—or how they will develop in future. As languages evolve and diverge, what guides their form?
Linguists have long suspected that language is like a game, in which individuals in a group or culture vie to impose their way of speaking. We adopt words and phrases we hear from others, and by using them, help them to propagate. Through face-to-face encounters, language evolves to reconcile our conflicting impulses as speakers or listeners. When speaking, we want to say our bit with minimal effort: we want language to be simple. As listeners, we want the speaker to make the meaning clear: we want language to be informative. In other words, speakers try to shift the effort onto listeners, and vice versa.
All this makes language what scientists call a complex system, meaning that it involves many agents interacting with each other via fairly well-defined rules. From these interactions there typically emerges an organised, global mode of behaviour that could not be deduced from local rules alone. Complex social systems have in recent years become widely studied by computer modelling: you define a population of agents, set the rules of engagement, and let the system run. Here the methods and concepts of the hard sciences—not so different to those used to model the behaviour of fundamental particles or molecules—are being imported into the traditionally empirical or narrative-dominated subjects of the social sciences. This approach has notched up successes in areas ranging from traffic flow to analysis of economic markets. No one pretends that a cultural artefact like language will ever be as tightly rule-bound or predictive as physics or chemistry, yet a complex-systems view might prove key to understanding how it evolves.
A significant success was recently claimed by and Italian group led by physicist Vittorio Loreto of the University of Rome La Sapienza. They looked at the paradigmatic example among linguists of how language segments and labels the objective world: the naming of colours.
As early anthropologists began to study non-Western languages in the nineteenth century, particularly those of pre-literate “savages”—they discovered that the familiar European colour terms of red, yellow, blue, green and so on are not as obvious and natural as they seem. Some indigenous people have far fewer colour terms. Many get by with perhaps three or four, so that for example “red” could refer to anything from green to orange, while blue, purple and black are all lumped together as types of black.
Inevitably, this was at first considered sheer backwardness. Researchers even concluded that such people were at an earlier stage of evolution, with a defective sense of colour vision that left them unable to tell the difference between, say, black and blue. Once they started testing natives using colour charts, however, they found them perfectly capable of distinguishing blue from black—they just saw no need to assign them different colour words. Uncomfortably for Western supremacists, we are in the same boat when it comes to blue, for Russians find it odd that an Englishman uses the same basic term for light blue (Russian goluboy) and dark blue (siniy).
Then in the 1860s the German philologist Lazarus Geiger proposed that the subdivision of colour always follows the same hierarchy. The simplest colour lexicons (such as the Dugerm Dani language of New Guinea) distinguish only black/dark and white/light. The next colour to be given a separate word is always centred on the red part of the visible spectrum. Then, according to Geiger, comes yellow, then green, then blue. Lazarus’s colour hierarchy was forgotten until restated in almost the same form in 1969 by US anthropologists Brent Berlin and Paul Kay, when it was hailed as one of the most significant discoveries in modern linguistics. Here was an apparently universal regularity underlying the way language is used to describe the world.
Berlin and Kay’s hypothesis has since fallen in and out of favour, and certainly there are exceptions to the scheme they proposed. But the fundamental colour hierarchy, at least in terms of the ordering black/white, red, yellow/green (either may come first) and blue, remains generally accepted. The problem is that no one could explain it.
Why, for example, do the blue of sky and sea, or the green of foliage, not register as distinct before the far less common red? It’s true that our visual system has evolved to be particularly sensitive to yellow (that’s why it appears so bright), probably because this enabled our pre-human ancestors to spot ripe fruit among foliage. But we have no trouble distinguishing purple, blue and green in the spectrum.
There are several schools of thought about how colours get named. “Nativists”, who include Berlin and Kay and Steven Pinker, argue that the concepts to which we attach words are innately determined by how we perceive the world. As Pinker has put it, “the way we see colours determines how we learn words for them, not vice versa”. In this view, often associated with Noam Chomsky, our perceptual apparatus has evolved to ensure that we make “sensible”—that is useful—choices of what to label with distinct words: we are hard-wired for particular forms of language. “Empiricists”, in contrast, argue that we don’t need this innate programming, but just the capacity to learn the conventional (but arbitrary) labels for things we can percieve.
In both cases, the categories themselves are deemed “obvious”: language just labels them. But the conclusions of Loreto and colleagues fit with a third possibility: the “culturist” view, which says that shared communication is needed to help organise category formation, so that categories and language co-evolve in an interaction between biological predisposition and culture. In other words, the starting point for colour terms is not some inevitably distinct block of the spectrum that we might decide to call ‘red’, ‘rouge’ and so on – but neither do we just divide up the spectrum any old how, because the human eye has different sensitivity to different parts of it. Given this, we have to arrive at some consensus not just on which label to use, but on what it labels.
The Italian team devised a computer model of language evolution in which new words arise through the game played by pairs of ’agents’, a speaker and a listener. The speaker uses words to refer to objects in a scene, and if she uses a word that is new to the listener (for a new colour, say), there’s a chance that the listener will figure out what the word refers to and adopt it. Alternatively, the listener might already have a word for that colour, but choose to replace it with the speaker’s word anyway. The language of this population of agents emerges and evolves from many such exchanges.
For colour, our visual physiology biases this process, picking out some parts of the spectrum as more worthy of a distinct colour term than others. The crucial factor is how well we can discriminate between very similar colours – we do that most poorly in the red, yellowish green and purple-violet. So we can’t distinguish two closely related reds as we can blues, say.
When the researchers included this bias in the colour-naming game, they found that colour terms emerged over time in their population of agents in much the same order proposed by Berlin and Key: first red, then violet, yellow, green, blue and orange. Violet doesn’t quite fit, but Loreto and colleagues think this is just an artefact of the way reddish hues crop up at both ends of the spectrum. Importantly, they don’t get the correct sequence unless they incorporate the colour sensitivity of actual human vision, but neither could the sequence be predicted from that alone, without the inter-agent negotiations that generate a consensus on colour words. You need both biology and culture to get it right.
The use of agent-based models to explore language evolution has been pioneered by Luc Steels of the Free University of Brussels, who is motivated by artificial intelligence: he wants to know how best to design robots so that they might develop a shared language. Steels and his coworkers have also favoured the acquisition of colour terms as their test case, and have previously argued in favour of the “cultural” picture that Loreto’s team now supports. The computer modelling of Steels’ group deserves much of the credit for starting to change the prevailing view of language acquisition, and the existence of near-universal patterns like Berlin and Kay’s colour hierarchy, from the influence of inherent, genetic factors to that of culture and environment.
Steels and his colleagues Joris Bleys and Joachim de Beule, for example, have presented an agent-based model of language negotiation, similar to that used by Loreto’s team, which purports to explain how a colour-language system can change from one based mostly on differences in brightness, using words like ‘dark’, ‘light’ and ‘shiny’, to one that makes distinctions of hue. (There are more ways to think about colour than Berlin and Kay’s rainbow-slicing.) The brightness system was used in Old English between around 600 and 1150, while Middle English (1150-1500) used hue-related words. A coeval switch was seen in other European languages, coinciding with the development of textile dyeing. This technology altered the constraints on what needed to be communicated: people now had to talk about a wider range of colours of similar brightness but different hue. Steels and colleagues showed that this sort of environmental pressure could tip the balance from a brightness-based colour terminology to a hue-based one. Again, it is one thing to tell that story, another to show that it really works in (a model of) the complex give and take of daily discourse. It increasingly seems, then, that language is determined not simply by ‘how we are’, but how it is used: by what we need to say.
___________________________________________
Languages are extremely diverse, but not arbitrary. Behind the bewildering diversity and the apparently contradictory ways in which different tongues elect to conceptualise the world, we can sometimes discern order and regularity. Many linguists have assumed that this reflects a hard-wired linguistic aptitude of the human brain. Some recent studies propose, however, that language ‘universals’ aren’t simply prescribed by genes but arise from the interaction between the biology of human perception and the bustle, exchange and negotiation of human culture.
Language has a perfectly logical job to do—to convey information— and yet is seemingly riddled with irrationality. Why all those irregular verbs, those random genders, those silent vowels and ambiguous homophones? You’d think languages would evolve towards some optimal model of clarity and concision, but instead they accumulate quirks that hinder learning, not only for foreigners but also in native speakers.
Traditionally, linguists have tended to explain the peculiarities of language through the history of the people who speak it. That’s often fascinating, but does not yield general principles about how languages have developed in the past—or how they will develop in future. As languages evolve and diverge, what guides their form?
Linguists have long suspected that language is like a game, in which individuals in a group or culture vie to impose their way of speaking. We adopt words and phrases we hear from others, and by using them, help them to propagate. Through face-to-face encounters, language evolves to reconcile our conflicting impulses as speakers or listeners. When speaking, we want to say our bit with minimal effort: we want language to be simple. As listeners, we want the speaker to make the meaning clear: we want language to be informative. In other words, speakers try to shift the effort onto listeners, and vice versa.
All this makes language what scientists call a complex system, meaning that it involves many agents interacting with each other via fairly well-defined rules. From these interactions there typically emerges an organised, global mode of behaviour that could not be deduced from local rules alone. Complex social systems have in recent years become widely studied by computer modelling: you define a population of agents, set the rules of engagement, and let the system run. Here the methods and concepts of the hard sciences—not so different to those used to model the behaviour of fundamental particles or molecules—are being imported into the traditionally empirical or narrative-dominated subjects of the social sciences. This approach has notched up successes in areas ranging from traffic flow to analysis of economic markets. No one pretends that a cultural artefact like language will ever be as tightly rule-bound or predictive as physics or chemistry, yet a complex-systems view might prove key to understanding how it evolves.
A significant success was recently claimed by and Italian group led by physicist Vittorio Loreto of the University of Rome La Sapienza. They looked at the paradigmatic example among linguists of how language segments and labels the objective world: the naming of colours.
As early anthropologists began to study non-Western languages in the nineteenth century, particularly those of pre-literate “savages”—they discovered that the familiar European colour terms of red, yellow, blue, green and so on are not as obvious and natural as they seem. Some indigenous people have far fewer colour terms. Many get by with perhaps three or four, so that for example “red” could refer to anything from green to orange, while blue, purple and black are all lumped together as types of black.
Inevitably, this was at first considered sheer backwardness. Researchers even concluded that such people were at an earlier stage of evolution, with a defective sense of colour vision that left them unable to tell the difference between, say, black and blue. Once they started testing natives using colour charts, however, they found them perfectly capable of distinguishing blue from black—they just saw no need to assign them different colour words. Uncomfortably for Western supremacists, we are in the same boat when it comes to blue, for Russians find it odd that an Englishman uses the same basic term for light blue (Russian goluboy) and dark blue (siniy).
Then in the 1860s the German philologist Lazarus Geiger proposed that the subdivision of colour always follows the same hierarchy. The simplest colour lexicons (such as the Dugerm Dani language of New Guinea) distinguish only black/dark and white/light. The next colour to be given a separate word is always centred on the red part of the visible spectrum. Then, according to Geiger, comes yellow, then green, then blue. Lazarus’s colour hierarchy was forgotten until restated in almost the same form in 1969 by US anthropologists Brent Berlin and Paul Kay, when it was hailed as one of the most significant discoveries in modern linguistics. Here was an apparently universal regularity underlying the way language is used to describe the world.
Berlin and Kay’s hypothesis has since fallen in and out of favour, and certainly there are exceptions to the scheme they proposed. But the fundamental colour hierarchy, at least in terms of the ordering black/white, red, yellow/green (either may come first) and blue, remains generally accepted. The problem is that no one could explain it.
Why, for example, do the blue of sky and sea, or the green of foliage, not register as distinct before the far less common red? It’s true that our visual system has evolved to be particularly sensitive to yellow (that’s why it appears so bright), probably because this enabled our pre-human ancestors to spot ripe fruit among foliage. But we have no trouble distinguishing purple, blue and green in the spectrum.
There are several schools of thought about how colours get named. “Nativists”, who include Berlin and Kay and Steven Pinker, argue that the concepts to which we attach words are innately determined by how we perceive the world. As Pinker has put it, “the way we see colours determines how we learn words for them, not vice versa”. In this view, often associated with Noam Chomsky, our perceptual apparatus has evolved to ensure that we make “sensible”—that is useful—choices of what to label with distinct words: we are hard-wired for particular forms of language. “Empiricists”, in contrast, argue that we don’t need this innate programming, but just the capacity to learn the conventional (but arbitrary) labels for things we can percieve.
In both cases, the categories themselves are deemed “obvious”: language just labels them. But the conclusions of Loreto and colleagues fit with a third possibility: the “culturist” view, which says that shared communication is needed to help organise category formation, so that categories and language co-evolve in an interaction between biological predisposition and culture. In other words, the starting point for colour terms is not some inevitably distinct block of the spectrum that we might decide to call ‘red’, ‘rouge’ and so on – but neither do we just divide up the spectrum any old how, because the human eye has different sensitivity to different parts of it. Given this, we have to arrive at some consensus not just on which label to use, but on what it labels.
The Italian team devised a computer model of language evolution in which new words arise through the game played by pairs of ’agents’, a speaker and a listener. The speaker uses words to refer to objects in a scene, and if she uses a word that is new to the listener (for a new colour, say), there’s a chance that the listener will figure out what the word refers to and adopt it. Alternatively, the listener might already have a word for that colour, but choose to replace it with the speaker’s word anyway. The language of this population of agents emerges and evolves from many such exchanges.
For colour, our visual physiology biases this process, picking out some parts of the spectrum as more worthy of a distinct colour term than others. The crucial factor is how well we can discriminate between very similar colours – we do that most poorly in the red, yellowish green and purple-violet. So we can’t distinguish two closely related reds as we can blues, say.
When the researchers included this bias in the colour-naming game, they found that colour terms emerged over time in their population of agents in much the same order proposed by Berlin and Key: first red, then violet, yellow, green, blue and orange. Violet doesn’t quite fit, but Loreto and colleagues think this is just an artefact of the way reddish hues crop up at both ends of the spectrum. Importantly, they don’t get the correct sequence unless they incorporate the colour sensitivity of actual human vision, but neither could the sequence be predicted from that alone, without the inter-agent negotiations that generate a consensus on colour words. You need both biology and culture to get it right.
The use of agent-based models to explore language evolution has been pioneered by Luc Steels of the Free University of Brussels, who is motivated by artificial intelligence: he wants to know how best to design robots so that they might develop a shared language. Steels and his coworkers have also favoured the acquisition of colour terms as their test case, and have previously argued in favour of the “cultural” picture that Loreto’s team now supports. The computer modelling of Steels’ group deserves much of the credit for starting to change the prevailing view of language acquisition, and the existence of near-universal patterns like Berlin and Kay’s colour hierarchy, from the influence of inherent, genetic factors to that of culture and environment.
Steels and his colleagues Joris Bleys and Joachim de Beule, for example, have presented an agent-based model of language negotiation, similar to that used by Loreto’s team, which purports to explain how a colour-language system can change from one based mostly on differences in brightness, using words like ‘dark’, ‘light’ and ‘shiny’, to one that makes distinctions of hue. (There are more ways to think about colour than Berlin and Kay’s rainbow-slicing.) The brightness system was used in Old English between around 600 and 1150, while Middle English (1150-1500) used hue-related words. A coeval switch was seen in other European languages, coinciding with the development of textile dyeing. This technology altered the constraints on what needed to be communicated: people now had to talk about a wider range of colours of similar brightness but different hue. Steels and colleagues showed that this sort of environmental pressure could tip the balance from a brightness-based colour terminology to a hue-based one. Again, it is one thing to tell that story, another to show that it really works in (a model of) the complex give and take of daily discourse. It increasingly seems, then, that language is determined not simply by ‘how we are’, but how it is used: by what we need to say.
Tuesday, August 14, 2012
Speaking from the attic
I discuss my new book Curiosity on the Guardian podcast here. No, I’m not at my most eloquent, but I hope you get the idea. Incidentally, water nerds (I know you’re out there) can see my talk on the chemistry of water at the UCLA “Fourth State of Water” symposium in March. This was an experiment in teleconferencing: I prepared a short intro movie using my cheap and cheerful Photo Booth, then recorded my Powerpoint talk with narration, and tuned in afterwards for a Q&A via Skype. All from my home study. And it seemed to kind of work – great to know that it’s possible to do this sort of thing now without a transatlantic flight.
Sunday, August 05, 2012
Falling down
I have mixed feelings about the Jonah Lehrer affair. No matter how I dislike the rivalries and jealousies of the writing world, it’s impossible for an old grafter, in these tough times of shrinking advances and dwindling opportunity for writers, not to feel a pang or two at the sight of a Wunderkind commanding whopping great speaker’s fees and being showered with adulatory reviews proclaiming him to be the future of science writing. But I’d like to think I was able not to be too begrudging, to feel that at least one of ‘us’ was making it big, and to recognize that this is just how things work, especially in the US. I’m sure that a fair bit of the venom that has come Jonah’s way in the light of the revelations of fabrication and self-plagiarism is fuelled by resentment at his youth and fame.
It is in any case all very sad. He may now be set up for life anyway, but I shudder to think how excruciating and embarrassing it must be to fall from such heights in such an ignominious way. And though it’s not what he meant by it, Lehrer’s remark in Imagine that “the young know less, which is why they often invent more” is not one that he’s going to be allowed to forget in a hurry.
He’ll recover, I expect, but it’s hard to see how he’ll ever quite shake off the stigma. And all for a few moments of there-but-for-the-grace-of-God hubris and deception. I don’t find it at all hard to understand the panic which compelled him to dig his hole deeper with outright lies about his fictitious sources on Dylan. It was foolish, of course, and unethical, but hardly a terrible sin.
Jonah was clearly far on the wrong side of grey territory in making up those quotes and pretending they were genuine. But there certainly is grey territory here, as there was in the case of Johann Hari using quotes from old interviews with his subjects as though they had been told directly to him. It’s not clear just how finickety must one be in making one’s sources plain – is it enough, for example, to say that your interviewee “has said…”, or do you need to mention to whom it was said? And there are no rules for how to ‘tidy up’ quotes. Do you just leave out repetitions and digressions? Or correct obviously unintended errors and grammatical slips? Or make a statement a bit more concise while preserving the meaning? I have certainly seen my own words recast, generally to good effect by making me sound much more articulate than I really am – I’ve no objection to that. I’ve also seen my meaning occasionally distorted once or twice, but evidently without that intention, and I’ve not been affronted by it. What Lehrer did obviously goes beyond such things, but I’ve a sense that, if you make a big blunder like this, the little slips and elisions are then hauled up as evidence against you too.
The charges of self-plagiarism are particularly ambiguous. I simply don’t know what the rules are here. It seems clear that one should never recycle articles for different publications unless there’s an explicit reason for that, and open acknowledgement of it. To accept a commission without admitting that you’ve written something similar or related before would be bad form. But what if you need to explain some particular concept or theory and feel that you couldn’t better the way you put it elsewhere? Is it okay to reuse a few phrases? A paragraph? I would think so, if they’re your own words anyway. There’s evidently a question of degree here – how much, how similar? I’ve not looked into exactly what Jonah has done in this regard, but self-plagiarism is a slippery concept. The comment by his publisher Houghton Mifflin Harcourt that “Jonah Lehrer fully acknowledges that Imagine draws upon work he has published in shorter form during the past several years and is sorry that was not made clear” sends a few shivers down the spine – is one really not meant to draw on one’s earlier, related journalistic work when preparing a book, or at least does one really have to specify in detail what's new and what you've said before?
And yet my response to this affair is coloured by another consideration. Some years ago, Jonah spoke to me while he was an editor for Seed magazine. He was preparing an article on people who were transforming science and how we think about it, or something, and for some reason I’d been selected as one of them. And you know, I have always remembered (at least, I think so, but Charles Fernyhough might argue about that) getting off the phone and thinking “Is it just me, or was there something sour in that bloke’s tone?” I’d sensed he was not persuaded that I quite deserved the accolade I was being given, that I had somehow disappointed him. Then he became famous, and I thought, ah OK, evidently here’s a very ambitious and competitive young man. So it goes. But now it seems there might have been just a tad too much of that.
It is in any case all very sad. He may now be set up for life anyway, but I shudder to think how excruciating and embarrassing it must be to fall from such heights in such an ignominious way. And though it’s not what he meant by it, Lehrer’s remark in Imagine that “the young know less, which is why they often invent more” is not one that he’s going to be allowed to forget in a hurry.
He’ll recover, I expect, but it’s hard to see how he’ll ever quite shake off the stigma. And all for a few moments of there-but-for-the-grace-of-God hubris and deception. I don’t find it at all hard to understand the panic which compelled him to dig his hole deeper with outright lies about his fictitious sources on Dylan. It was foolish, of course, and unethical, but hardly a terrible sin.
Jonah was clearly far on the wrong side of grey territory in making up those quotes and pretending they were genuine. But there certainly is grey territory here, as there was in the case of Johann Hari using quotes from old interviews with his subjects as though they had been told directly to him. It’s not clear just how finickety must one be in making one’s sources plain – is it enough, for example, to say that your interviewee “has said…”, or do you need to mention to whom it was said? And there are no rules for how to ‘tidy up’ quotes. Do you just leave out repetitions and digressions? Or correct obviously unintended errors and grammatical slips? Or make a statement a bit more concise while preserving the meaning? I have certainly seen my own words recast, generally to good effect by making me sound much more articulate than I really am – I’ve no objection to that. I’ve also seen my meaning occasionally distorted once or twice, but evidently without that intention, and I’ve not been affronted by it. What Lehrer did obviously goes beyond such things, but I’ve a sense that, if you make a big blunder like this, the little slips and elisions are then hauled up as evidence against you too.
The charges of self-plagiarism are particularly ambiguous. I simply don’t know what the rules are here. It seems clear that one should never recycle articles for different publications unless there’s an explicit reason for that, and open acknowledgement of it. To accept a commission without admitting that you’ve written something similar or related before would be bad form. But what if you need to explain some particular concept or theory and feel that you couldn’t better the way you put it elsewhere? Is it okay to reuse a few phrases? A paragraph? I would think so, if they’re your own words anyway. There’s evidently a question of degree here – how much, how similar? I’ve not looked into exactly what Jonah has done in this regard, but self-plagiarism is a slippery concept. The comment by his publisher Houghton Mifflin Harcourt that “Jonah Lehrer fully acknowledges that Imagine draws upon work he has published in shorter form during the past several years and is sorry that was not made clear” sends a few shivers down the spine – is one really not meant to draw on one’s earlier, related journalistic work when preparing a book, or at least does one really have to specify in detail what's new and what you've said before?
And yet my response to this affair is coloured by another consideration. Some years ago, Jonah spoke to me while he was an editor for Seed magazine. He was preparing an article on people who were transforming science and how we think about it, or something, and for some reason I’d been selected as one of them. And you know, I have always remembered (at least, I think so, but Charles Fernyhough might argue about that) getting off the phone and thinking “Is it just me, or was there something sour in that bloke’s tone?” I’d sensed he was not persuaded that I quite deserved the accolade I was being given, that I had somehow disappointed him. Then he became famous, and I thought, ah OK, evidently here’s a very ambitious and competitive young man. So it goes. But now it seems there might have been just a tad too much of that.
Saturday, August 04, 2012
Stop me if you've heard this before
My latest story for Chemistry World is strictly for chemistry nerds. But this one for Physics World – extended version below – has hopefully a little more general interest in its all-round boggleworthiness. Charles Bennett’s comments seem to imply that the key question is whether it is possible to find a way of experimentally distinguishing between this remarkable interpretation and the more prosaic one he suggests.
___________________________________________________________________
What you do today could affect what happened yesterday. This is the bizarre conclusion of a thought experiment in quantum physics described in a preprint by Yakir Aharonov of Tel-Aviv University in Israel and his colleagues.
It sounds impossible, indeed as though it is violating one of science’s most cherished principles: causality. But the researchers say that the rules of the quantum world conspire to decorously preserve causality by ‘hiding’ the influence of future choices until those choices have actually been made.
At the centre of the idea is the quantum phenomenon of nonlocality, in which two or more particles exist in inter-related (‘entangled’) states that remain undetermined until a measurement is made on one of them – whereupon the state of the other particle is instantly fixed too, no matter how far away it is. Albert Einstein first pointed out this instantaneous ‘action at a distance’ in 1935, when he argued that it meant quantum theory must be incomplete. But modern experiments have confirmed that this instantaneous action is real, and it now holds the key to practical quantum technologies such as quantum computing and cryptography.
Aharonov and his coworkers describe an experiment much like that proposed by Einstein, but on a large group of entangled particles rather than just two. They argue that under certain conditions, the experimenter’s choice of a measurement of the states of the particles can be demonstrated to affect the states they were in at an earlier time, when a very loose measurement was made. In effect, the earlier ‘weak’ measurements anticipate the choice made in the later ‘strong’ measurement.
The work builds on a way of thinking about entanglement proposed by Aharanov three decades ago. This entails looking at the correlations between particles in the four dimensions of spacetime rather than the three of space. “In three dimensions it looks like some miraculous influence between two distant particles”, says Aharonov’s coworker Avshalom Elitzur. “In spacetime as a whole, it is a continuous interaction extending between past and future events.” Quantum systems are generally described by a ‘state vector’: a set of quantum states propagating forward in time. But Aharanov’s view considers also a second state vector propagating from future to past – which is why it is called the ‘two state vector formalism’ (TSVF).
Aharonov and coworkers have now discovered a remarkable implication of the TSVF. It bears on the question posed by Einstein once the early quantum theorists began to appreciate how measurement not just reveals but may determine the state of quantum systems. If observation has this effect of fixing how the world is, said Einstein, then can we be so sure that the Moon is there when no one is looking?
“The ordinary physicist replies, ‘Go away, this is a philosophy not physics’”, says Elitzur. It is equivalent to asking what is the state of a particle between two measurements. “Of course you're not going to measure the particle, because then you will have the particle's state upon measurement rather than between measurements.” But Aharanov’s perspective shows that it is possible to get at the intermediate information – by making sufficiently ‘weak’ measurements on a whole bunch of entangled particles prepared in the same way, and then averaging the statistics. Elitzur explains that this amounts to saying “Give me sufficiently many particles during of this time interval and I'll tell you precisely what you want to know.”
The weak measurements tell you something about the probabilities of different states (spin value up or down, say) – albeit with a lot of error – without actually collapsing them into definite states, as a strong measurement does. The weak measurement does perturb the system, but not enough to fix an outcome for sure. “A weak measurement both changes the measured state and informs you about the resulting localized state”, says Elitzur. “But it does both jobs very loosely. Moreover, the change it inflicts on the system must be weaker than the information it gives you.”
As a result, Elitzur explains, “every single weak measurement in itself tells you nearly nothing. The measurements provide reliable outcomes only after you sum them all up. Then the errors cancel out and you can extract some information about the ensemble as a whole.”
In the researchers’ thought experiment, the conclusions of these weak measurements will agree with those of later strong measurements, in which the experimenter chooses freely which spin orientation to measure – even though, after the weak measurements, the particles’ states are still undetermined.
What this means within the TSVF, says Elitzur, is that “a particle between two measurements possesses the two states indicated by both of them, past and future!” This even seems to evade Heisenberg’s uncertainty principle, which forbids simultaneous precise knowledge of a particles position and momentum. “If you measured position first and momentum later, then the particle possesses both precise values, never mind Heisenberg”, says Elitzur. Heisenberg himself felt that his uncertainty principle undermined causality – he’d have been shocked to find this kind of backward causality actually seeming to undermine his own law.
But causality does emerge intact, after a fashion. For the catch is that the weak measurements in themselves appear to leave many options for what the particles states are. Only by adding subsequent information from the strong measurements can one reveal what the weak measurements were ‘really’ saying. This means that the weak measurements by themselves can’t show you what the later strong measurements will reveal. The information is there, but encrypted and only exposed in retrospect. So causality is preserved, even if it is not exactly causality as we normally know it.
Why there is this censorship is not clear, except from an almost metaphysical perspective. “Nature is known to be fussy about never appearing inconsistent”, says Elitzur. “So she's not going to appreciate overt backwards causality – people killing their grandfathers and so on.”
He says that some specialists in quantum optics have expressed interest in conducting the experiment, which he thinks should be no more difficult than previous studies of entanglement.
Charles Bennett of IBM’s research laboratories in Yorktown Heights, New York, a specialist on quantum information theory, is not convinced. For a start, he sees the TSVF as only one way of looking at the results. “People in quantum foundations are often so wedded to their own interpretation or formalism that they say it is the only reasonable one, when in fact quantum mechanics admits multiple interpretations, which except for a few outliers are entirely equivalent to one another. The differences are aesthetic and philosophical, not scientific.”
Bennett believes that the findings can be interpreted without any apparent ‘backwards causation’, so that the authors are erecting a straw man. “To make their straw man seem stronger, they use language that in my opinion obscures the crucial difference between communication and correlation. They say that the initial weak measurement outcomes anticipate the experimenter's future choice but that doing so causes no violation of causality because the anticipation is encrypted.” But he thinks this is a bit like an experiment in quantum cryptography in which the sender sends the receiver the decryption key before sending (or even deciding on) the message, and then claims that the key is somehow an ‘anticipation’ of the message. With this in mind, it is not clear whether even an experiment will resolve the issue, since it would come down to a matter of how to interpret the results.
Aharonov and colleagues suspect that their findings might even have implications for free will. “Our group remains somewhat divided on these philosophical questions”, says Elitzur. “I keep teasing Yakir that he will go down in history as the person who has abolished free choice. He on the other hand is confident that TSVF secures free will a place within physical formalism. His conclusion is somewhat Talmudic: Everything you're going to do is already known to God, but you still have the choice. On the other hand Yakir's God sharply differs from Einstein's in that she loves to play dice from morning to night.”
___________________________________________________________________
What you do today could affect what happened yesterday. This is the bizarre conclusion of a thought experiment in quantum physics described in a preprint by Yakir Aharonov of Tel-Aviv University in Israel and his colleagues.
It sounds impossible, indeed as though it is violating one of science’s most cherished principles: causality. But the researchers say that the rules of the quantum world conspire to decorously preserve causality by ‘hiding’ the influence of future choices until those choices have actually been made.
At the centre of the idea is the quantum phenomenon of nonlocality, in which two or more particles exist in inter-related (‘entangled’) states that remain undetermined until a measurement is made on one of them – whereupon the state of the other particle is instantly fixed too, no matter how far away it is. Albert Einstein first pointed out this instantaneous ‘action at a distance’ in 1935, when he argued that it meant quantum theory must be incomplete. But modern experiments have confirmed that this instantaneous action is real, and it now holds the key to practical quantum technologies such as quantum computing and cryptography.
Aharonov and his coworkers describe an experiment much like that proposed by Einstein, but on a large group of entangled particles rather than just two. They argue that under certain conditions, the experimenter’s choice of a measurement of the states of the particles can be demonstrated to affect the states they were in at an earlier time, when a very loose measurement was made. In effect, the earlier ‘weak’ measurements anticipate the choice made in the later ‘strong’ measurement.
The work builds on a way of thinking about entanglement proposed by Aharanov three decades ago. This entails looking at the correlations between particles in the four dimensions of spacetime rather than the three of space. “In three dimensions it looks like some miraculous influence between two distant particles”, says Aharonov’s coworker Avshalom Elitzur. “In spacetime as a whole, it is a continuous interaction extending between past and future events.” Quantum systems are generally described by a ‘state vector’: a set of quantum states propagating forward in time. But Aharanov’s view considers also a second state vector propagating from future to past – which is why it is called the ‘two state vector formalism’ (TSVF).
Aharonov and coworkers have now discovered a remarkable implication of the TSVF. It bears on the question posed by Einstein once the early quantum theorists began to appreciate how measurement not just reveals but may determine the state of quantum systems. If observation has this effect of fixing how the world is, said Einstein, then can we be so sure that the Moon is there when no one is looking?
“The ordinary physicist replies, ‘Go away, this is a philosophy not physics’”, says Elitzur. It is equivalent to asking what is the state of a particle between two measurements. “Of course you're not going to measure the particle, because then you will have the particle's state upon measurement rather than between measurements.” But Aharanov’s perspective shows that it is possible to get at the intermediate information – by making sufficiently ‘weak’ measurements on a whole bunch of entangled particles prepared in the same way, and then averaging the statistics. Elitzur explains that this amounts to saying “Give me sufficiently many particles during of this time interval and I'll tell you precisely what you want to know.”
The weak measurements tell you something about the probabilities of different states (spin value up or down, say) – albeit with a lot of error – without actually collapsing them into definite states, as a strong measurement does. The weak measurement does perturb the system, but not enough to fix an outcome for sure. “A weak measurement both changes the measured state and informs you about the resulting localized state”, says Elitzur. “But it does both jobs very loosely. Moreover, the change it inflicts on the system must be weaker than the information it gives you.”
As a result, Elitzur explains, “every single weak measurement in itself tells you nearly nothing. The measurements provide reliable outcomes only after you sum them all up. Then the errors cancel out and you can extract some information about the ensemble as a whole.”
In the researchers’ thought experiment, the conclusions of these weak measurements will agree with those of later strong measurements, in which the experimenter chooses freely which spin orientation to measure – even though, after the weak measurements, the particles’ states are still undetermined.
What this means within the TSVF, says Elitzur, is that “a particle between two measurements possesses the two states indicated by both of them, past and future!” This even seems to evade Heisenberg’s uncertainty principle, which forbids simultaneous precise knowledge of a particles position and momentum. “If you measured position first and momentum later, then the particle possesses both precise values, never mind Heisenberg”, says Elitzur. Heisenberg himself felt that his uncertainty principle undermined causality – he’d have been shocked to find this kind of backward causality actually seeming to undermine his own law.
But causality does emerge intact, after a fashion. For the catch is that the weak measurements in themselves appear to leave many options for what the particles states are. Only by adding subsequent information from the strong measurements can one reveal what the weak measurements were ‘really’ saying. This means that the weak measurements by themselves can’t show you what the later strong measurements will reveal. The information is there, but encrypted and only exposed in retrospect. So causality is preserved, even if it is not exactly causality as we normally know it.
Why there is this censorship is not clear, except from an almost metaphysical perspective. “Nature is known to be fussy about never appearing inconsistent”, says Elitzur. “So she's not going to appreciate overt backwards causality – people killing their grandfathers and so on.”
He says that some specialists in quantum optics have expressed interest in conducting the experiment, which he thinks should be no more difficult than previous studies of entanglement.
Charles Bennett of IBM’s research laboratories in Yorktown Heights, New York, a specialist on quantum information theory, is not convinced. For a start, he sees the TSVF as only one way of looking at the results. “People in quantum foundations are often so wedded to their own interpretation or formalism that they say it is the only reasonable one, when in fact quantum mechanics admits multiple interpretations, which except for a few outliers are entirely equivalent to one another. The differences are aesthetic and philosophical, not scientific.”
Bennett believes that the findings can be interpreted without any apparent ‘backwards causation’, so that the authors are erecting a straw man. “To make their straw man seem stronger, they use language that in my opinion obscures the crucial difference between communication and correlation. They say that the initial weak measurement outcomes anticipate the experimenter's future choice but that doing so causes no violation of causality because the anticipation is encrypted.” But he thinks this is a bit like an experiment in quantum cryptography in which the sender sends the receiver the decryption key before sending (or even deciding on) the message, and then claims that the key is somehow an ‘anticipation’ of the message. With this in mind, it is not clear whether even an experiment will resolve the issue, since it would come down to a matter of how to interpret the results.
Aharonov and colleagues suspect that their findings might even have implications for free will. “Our group remains somewhat divided on these philosophical questions”, says Elitzur. “I keep teasing Yakir that he will go down in history as the person who has abolished free choice. He on the other hand is confident that TSVF secures free will a place within physical formalism. His conclusion is somewhat Talmudic: Everything you're going to do is already known to God, but you still have the choice. On the other hand Yakir's God sharply differs from Einstein's in that she loves to play dice from morning to night.”