You are on page 1of 3
The Physical Basis of Computability R. B. Laughlin Department of Physics, Stanford University, Stanford, California 94305∗ (Dated: 1 Jan 02) Simulations work in practice because they exploit higher-level organizing principles in nature. Good code writing requires faithfulness to these principles and the discipline not to exceed their limits of validity. An important exception is the use of simulation to search for new kinds of emergence. [Published as Computing in Science and Engineering 4, 27 (2002).] PACS numbers: 02.60.Cb, 07.05.Tp, 82.20.Wt, 83.85.Pt, 85.40.Bh, 87.15.Aa Once, as a graduate student attending an international ently fuzzy disciplines, this is quite easy to demonstrate conference, I made the mistake of admitting during my in physics. Consider, for example, the celebrated spin talk that I could not calculate the optical absorption glass problem - a set of N half-integral quantum spins spectrum of a certain structural defect in an insulator. interacting by random Heisenberg exchanges. Because I meant that I could not calculate accurately, of course, this system’s configuration space has dimension 2N , a but that did not matter. Before I had time to think or straight solution of the quantum mechanics requires the qualify my statement an ambitious young assistant pro- diagonalization of a 2N × 2N matrix, something we know fessor in the back leaped to his feet and yelled, “Maybe how to do algorithmically. However even for the case of you can’t calculate this spectrum, but I can!” His words N = 200 this matrix has 2400 = 2.6 × 10120 elements, ring in my ear to this day; one tends not to forget such a number vastly larger than all the atoms in the visi- things. I got through my presentation somehow, retired, ble universe. The computational task would obviously and then later went about finding out how this guy had exceed the memory capacity of any conventional digital managed to do a computation I had found impossible. computer that could ever be built and is therefore fun- It turned out to be a wild goose chase. He could no damentally impossible. This example is not merely aca- more do an honest calculation of that spectrum than I demic. Spin glasses exhibit an array of important ther- could. He had simply redefined “calculation” to mean a modynamic behaviors - the glass transition, remanance, postdiction of a complicated model with lots of param- huge low-temperature specific heats - that to not occur in eters fit to the data. He had also hidden these weak- ordered antiferromagnets. Most of the fundamental lim- nesses in a large, proprietary, poorly-documented com- its to physical simulations we care about - fully-developed puter program, so they could not be discovered without turbulence, reaction chemistry, cuprate superconductiv- considerable work. But I was motivated, had time on ity, life, and so forth - are like this. They proceed from my hands, and was eventually able to get to the truth— the configurational nature of matter, particularly quan- and truth it was. To this day no one has ever calcu- tum matter, to the conclusion that the size of the com- lated these spectra correctly from first principles. Their putation must increase exponentially with the number shape is dominated by Franck-Condon broadening (corre- of degrees of freedom and rapidly become qualitatively lated electronic and nuclear motion) and complex multi- larger than the resources of any imaginable computer. electron shake-off effects1 . This calculation, like many The importance and universality of this effect is beauti- others I have encountered in my career, is just too hard. fully articulated by P. W. Anderson in his famous paper It is understandably difficult for any of us to admit “More is Different”3 . that a calculation is just too hard. Large segments of A formally correct but physically problematic response our society now view the understanding of natural phe- to this argument is the quantum computer4 . One way to nomena, in the sense of Bohr and Einstein, to be a quaint “compute” the motion of the spins in the previous exam- anachronism rendered obsolete by computers. I strongly ple is simply to initialize them physically and measure disagree with this view, and can defend this position with their behavior as a function of time using conventional sobering accounts of its failure, but it is part of our cul- spectroscopy. This could be done without employing all ture at the moment and something with which we have the atoms in the universe and would constitute one ex- to live. From this perspective a computer code or compu- ample of simulation by quantum computer - namely time tational strategy unable to produce some essential result evolution of the system itself. The quantum computer is just outdated technology, something to be supplanted evades the problem of thermodynamic size by being it- shortly by either a market challenge or the next upgrade. self fundamentally different from conventional comput- None of us can afford to be uncompetitive, so none of us ers. Indeed the unique ability of quantum computers is anxious to admit weakness, even when it is true and to solve such problems is what makes them so interest- even when the reason for the weakness is fundamental. ing. However, the very property that allows quantum The fundamental limits to simulating physical phe- computers to solve such problems also makes them more nomena by computer are real2 . In contrast to the sit- noise-prone and unpredictable. This effect is the micro- uation in business, economics, or law, which are inher- scopic origin of the second law of thermodynamics and 2 is fundamental. Whether it can be overcome technically by the owner. We are only human, and there are just is an important and deep question, but one that can be too many incentives and opportunities for misrepresent- answered only by experiment. I am aware of no such ing the truth. These are sometimes subtle and difficult to technology on the immediate horizon. detect. For example, one strategy is to write a complex In light of the configurational difficulty and the stupen- code based on bogus principles, adjust its parameters to dous variety and unpredictability of the natural world it fit a handful of experiments, and then use comparison is remarkable that anything can be simulated reliably. with one of these experiments as a “test”. Another is to One is moved to ask why simulation is possible at all, “test” the code in some extreme limit where it cannot and what distinguishes a system that can be simulated fail and does not matter. Another is to define the differ- from one that cannot. ence between what the code actually produces and what The physical world is simple because it is regulated it should have produced as acceptable error. Another by higher-level physical law5 . The laws in question are is to benchmark the code against another code that is collective in nature and emergent - meaning that they wrong in the same way. Another is to dismiss as wrong are exact only in the thermodynamic limit, encoded only the experiment with which the code should have agreed. indirectly by the underlying laws of quantum mechanics, Another is to declare the code valid only in regimes that and in a deep sense independent of them. The laws of hy- are experimentally inaccessible. I have had so much un- drodynamics are a good example of this, as are the laws pleasant personal experience with these and other failure of crystalline elasticity, the laws of plasticity, the laws modes that I have now come to automatically mistrust of superfluidity, the laws of magnetism, and the laws of physical simulations not based on principles I fully under- thermodynamics. When hydrodynamics is working, say, stand and that are rigorously tested against experiment in the collision or two very large nuclei, a hydrodynamic in their entire range of validity. simulation can predict the scattering experiment’s essen- An important exception to this rule is the class of tial features without taking into account the equations simulation that aims to identify new emergent princi- of motion of individual nuclei6 . When the laws of plastic ples. There are legendary successes of this approach in flow are working, say, in the motion of a glacier, we can physics: Fermi, Pasta, and Ulam discovered the soliton predict where the ice will flow without knowing where effect in simulations of lattice thermization in the late each atom goes7 . It is the existence of these higher-level 1940s8 . Alder and Wainwright discovered the Kosterlitz- laws - the simple mathematical relationships among mea- Thouless transition9 , long-time tails, and atomic-scale sured quantities created and enforced by emergent phys- hydrodynamics10 in molecular dynamics simulations of ical phenomena - that makes meaningful computer sim- fluids. Mitchell Feigenbaum discovered scaling on the ulation possible. period-doubling path to chaos in turbulence studies11 . Most of us have an intuitive understanding of emer- Edward Lorentz discovered the strange attractor and the gent laws and respect them even though we rarely talk principle of chaos while trying to model the weather12 . about them. In praising a good code we often say that The vast majority of commercial applications - e.g. ur- it “captures the physics.” We mean by this that the au- ban growth13 , war strategy14 , crop forcasts15 , and drug thor skillfully exploited one or more legitimate collective design searches16 are in this category. In these and simi- organizational principles in writing the code, disciplined lar applications the underlying physical basis is not fully himself to stay within the limits of validity of those prin- understood and simulation aims only for “some” quanti- ciples, and did not just make things up. The latter is tative analysis on the grounds that it is better than none. centrally important because computer simulations, like However, for such a simulation to be correct, as opposed the equations of mathematical physics on which they are to merely useful, it must predict one or more contentful based, are symbolic representations of physical law and experimental facts. It is not generally possible to start are meaningful only insofar as they are faithful to that from the wrong equations and get the right result. This law. can only happen if the answer is insensitive to details, The reliance of physical simulations on organizational and therefore reliable, because of some as-yet-unknown principles for their validity means that they must be higher-level emergent law. The identification of such laws judged by higher standards than those we use for other by means of simulation is still in its infancy, and I be- software. Even if they work well and produce breathtak- lieve it to be one of the great outstanding challenges for ing graphics, simulations can be, and often are, wrong. computer science. It is, however, potentially vulnerable Sometimes this does not matter, as in the recreational to abuse, especially since the tests are hard to quantify F-22 simulator my teenage son flies, and sometimes it and administer. does. When I first heard about the plan to design air- I believe that the long-term health of computing as planes with electronic wind tunnels, for example, a happy a branch of physical science is invested in the opposite vision flashed into my mind of sending the people respon- approach—the repudiation of market-based science and sible for this decision up on the first test flight. a demand of total fidelity with established physical law. When the validity of a simulation matters one must be Mark Twain captured the problem well when he said that profoundly uncomfortable with its central physical prin- truth is always stranger than fiction because fiction is ciples being proprietary or relying on “tests” designed forced to stick to possibilities, while truth is not. Real 3 science always begins with careful observations of nature cause strong winter storms can now be predicted a few and thoughtful consideration of facts that “ought not to days in advance. Computer modeling of the sun has led be true” but nonetheless are. Studying only a simulated to the discovery of the neutrino flux deficit18 and the world based on what one thinks is true rather than what likely detection neutrino oscillations19 . Modern quanti- actually is true automatically precludes rethinking the tative hydrodynamic simulations have vastly decreased facts, and therefore automatically precludes making a the cost of airplane design20 . However, there is a serious fundamental discovery. Klaus von Klitzing’s discovery danger of this power being misused, either by accident of the quantum hall effect is a beautiful case in point17 . or through deliberate deception. All of us trained in sci- No one before von Klitzing had ever bothered to mea- ence and concerned about the the integrity of physical sure the hall conductance accurately because it was not law must be committed to preventing this. As caretakers supposed to be quantized. of a tradition that is perhaps the greatest contribution It has become clear to many of us that the central to humanity western civilization has ever made, we are task of physics in our time is the identification and enu- obligated to remind our fellow citizens and the younger meration of as-yet undiscovered higher organizing princi- generations that real science deals with truth, and is as ples of nature. This is a vast frontier, and an incredibly different from falsehood as night is from day. Nothing exciting one, especially in light of developments in the that has happened, or will happen, in government or the life sciences. It is my great hope that simulation will economy can change this. After all of us are memories play the important role in this drama that its nature and and the last grant proposal, program review, and IPO traditions suggest it should. Computers are enormously have passed into history there will still be truth, reason, powerful tools, and they can do great good when used experimental discipline, and the majesty of physical law. properly and wisely. Many lives are spared each year be- ∗ 11 R. B. Laughlin: http://large.stanford.edu M. J. Feigenbaum, “Quantitative Universality for a Class 1 G. D. Mahan, Many-Particle Physics, Second Edition of Nonlinear Transformations,” J. Stat. Phys. 19, no. 1, (Plenum, 1990). 25-52 (1978). 2 12 N. Orsekes, K. Shrader-Frechette, and K. Belitz, “Verifi- E. N. Lorentz, “Deterministic Nonperiodic Flow,” J. At- cation, Validation, and Confirmation of Numerical Models mos. Sci. 20, 130-141 (1963). 13 in the Earth Sciences,” Science, 263, no. 5147, 641-646 R. Liggett, “Automated Facilities Layout: Past, Present, (1994). and Future,” Automation in Construction 9, no. 2, 197-215 3 P.W. Anderson, “More is Different,” Science 177, no. 4047, (2000). 14 393-396 (1972). S. I. Irwin, “Commanders Want Realistic Sim- 4 M. A. Nielson and I. I. Chuang, Quantum Computation ulations,” National Defense Magazine, April 2001 and Quantum Information (Cambridge U. Press, Cam- (www.nationaldefensemagazine.org). 15 bridge, 2000). Y. W. Jame and H. W. Cutforth, “Crop Growth Models 5 R.B. Laughlin and D. Pines, “The Theory of Everything,” for Decision Support Systems,” Canadian J. Plant Science Proc. Nat’l Acad. Sci., 97, no. 1, 28-31 (2000). 46, no. 1, 9-19 (January 1996). 6 16 U. Ornik, F. W. Pottag, and R. M. Weiner, “High-Energy R. S. Dewitte, A. V. Ischenko, and E. I. Shakhnovich, “De Reactions and Relativistic Hydrodynamics in Three Di- Novo Method Based on Simple, Fast, Accurate Free Energy mensions,” Phys. Rev. Lett. 63, no. 24, 2641-2644 (1989). Estimates 2: Case Studies in Molecular Design,” J. Am. 7 G. A. Olyphant, “Computer Simulation of Rock-Glacier Chemical Sciences, 119, no. 20, 4608-4617 (1997). 17 Development under Viscous and Pseudo-Plastic Flow, K. von Klitzing, G. Dorda, and M. Pepper, “New Method Bull. Geo. Soc. Am. 94, 499-505 (1983). for High-Accuracy Determination of the Fine-Structure 8 E. Fermi, J. R. Pasta, and S. M. Ulam, “Studies of Nonlin- Constant Based on Quantized Resistance,” Phys. Rev. ear Problems,” in Sets, Numbers, and Universes, ed. by S. Lett. 45, no. 6, 494-497 (1980). 18 Ulam, W. Beyer, J. Mycielski, and Gian-Carlo Rota (MIT J. N. Bahcall, “Solar Neutrinos: Where We Are, Where Press, Cambridge, 1974), pp. 491-501. We Are Going,” Astrophys. J. 467, 475-484 (1996). 9 19 B. J. Alder and T. E. Wainwright, “Phase Transition in Y. Fukuda et al., “Evidence for Oscillation of Atmospheric Elastic Disks,” Phys. Rev. 127, no. 2, 359-361 (1962). Neutrinos,” Phys. Rev. Lett. 81, no. 8, 1562-1567 (1998). 10 20 B. J. Alder and T. E. Wainwright, “Decay of the Velocity R. Agarwal, “Computational Fluid Dynamics of Whole- Autocorrelation function,” Phys. Rev. A 1, no. 1, 18-21 Body Aircraft,” Ann. Rev. Fluid Mech. 31, 125-169 (1999). (1970).