InteractionBased Computing in Physics
Latest version View entry history
 164 Downloads
Keywords
Lyapunov Exponent Cellular Automaton Energy Landscape Extended System Couple Differential EquationGlossary
 Correlation

The correlation between two variables is the difference between the joint probability that the two variables take some values and the product of the two probabilities (which is the joint probability of two uncorrelated variables), summed over all possible values. In an extended system, it is expected that the correlation among parts diminishes with their distance, typically in an exponential manner.
 Critical phenomenon

A condition for which an extended system is correlated over extremely long distances.
 Extended system

A system composed by many parts connected by a network of interactions that may be regular (lattice) or irregular (graph).
 Graph, lattice, tree

A graph is set of nodes connected by links, oriented or not. If the graph is translationally invariant (it looks the same when changing nodes), it is called a (regular) lattice. A disordered lattice is a lattice with a fraction of removed links or nodes. An ordered set of nodes connected by links is called a path. A closed path not passing on the same links is a loop. A cluster is a set of connected nodes. A graph can be composed by one cluster (a connected graph) or more than one (a disconnected graph). A tree is a connected graph without loops.
 Mean field

An approximate technique for computing the value of the observables of an extended system, neglecting correlations among parts. If necessary, the dynamics is first approximated by a stochastic process. In its simpler version, the probability of a state of the system is approximated by the product of the probability of each component, neglecting correlations. Since the state of two components that depend on a common “ancestor” (that interact with a common node) is in general not uncorrelated, and this situation corresponds to an interaction graph with loops, the simplest meanfield approximation consists in replacing the graph or the lattice of interactions with a tree.
 Monte Carlo

A method for producing stochastic trajectories in the state space designed in such a way that the timeaveraged probability distribution is the desired one.
 Nonlinear system

A system composed by parts whose combined effects are different from the sum of the effects of each part.
 Percolation

The appearance of a “giant component” (a cluster that contains essentially all nodes or links) in a graph or a lattice, after adding or removing nodes or links. Below the percolation threshold, the graph is partitioned into disconnected clusters, none of which contains a substantial fraction of nodes or links, in the limit of an infinite number of nodes and links.
 Probability distribution

The probability of finding a system in a given state, for all the possible states.
 State of a system

A complete characterization of a system at a given time, assigning or measuring the positions, velocities, and other dynamical variables of all the elements (sometimes called a configuration). For completely discrete systems (cellular automata) of finite size, the state of the system is just a set of integer numbers, and therefore the state space is numerable.
 State space

The set of all possible states of a system.
 Trajectory

A sequence of states of a system, labeled with the time, i.e., a path in the state space.
Definition
Theoretical physics investigation is based on building models of reality. Due to our limited cognitive capabilities, In order to understand a phenomenon, we need to represent it using a limited amount of symbols. Clearly, it is generally impossible to reproduce all aspects of a physical system with a simple model. However, even using simple building blocks, one can obtain systems whose behavior is quite complex. Therefore, most of the job of theoretical physics is that of identifying and investigating simple models that reproduce the main aspects of complex phenomena. Although some progress can be done analytically, most of models are also investigated by computer simulations.
Computers have changed the way a physical model is studied. Computers may be used to calculate the properties of a very complicated model representing a real system or to investigate experimentally what are the essential ingredients of a complex phenomenon. In order to carry out these explorations, several basic models have been developed, which are now used as building blocks for performing simulations and designing algorithms in many fields, from chemistry to engineering, from natural sciences to psychology. Rather than being derived from some fundamental law of physics, these blocks constitute artificial worlds still to be completely explored.
In this entry, we shall first present a pathway from Newton’s laws to cellular automata and agentbased simulations, showing (some) computational approaches in classical physics. Then, we shall present some examples of artificial worlds.
Introduction: Physics and Computers
Some 60 years ago, shortly after the end of the Second World War, computers become available to scientists. Actually, computers had already been used during the last years of the war for performing computations about a specific field of physics: the atomic bomb (Harlow and Metropolis 1983; Metropolis et al. 1980).
Up to then, the only available tool, except experiments, was paper and pencil. Starting with Newton and Leibnitz, humans discovered that continuous mathematics (i.e., differential and integral calculus) allowed to derive many consequences of a given hypothesis just by the manipulation of symbols. It seemed natural to express all quantities (e.g., time, space, mass) as continuous variables. Notice however that the idea of a continuous number is not at all “natural”: one has to learn how to deal with it, while (small) integer numbers can be used and manipulated (added, subtracted) by illiterate humans and also by many animals. A point which is worth to be stressed is that any computation refers to a model of certain aspects of reality considered most important, while others are assumed to be not important.
Most of the investigations in physics only using paper and pencils are limited to almostlinear systems, or systems whose effective number of variables is quite small. On the other hand, most of the naturally occurring phenomena can be successfully modeled only using nonlinear elements . Therefore, a great deal of the precomputer physics is essentially linear physics, although astronomers (like other scientists) used to integrate numerically, by hand, the nonlinear equations of gravitation, in order to compute the trajectories of planets. This computation, however, was so cumbersome that no playing with trajectories was possible.
How can a computer solve physical problems? While analog computers have been used for integrating differential equations, the much more flexible digital computers are deterministic discrete systems. The way of working of a (serial) computer is that of a very fast automaton that manipulates data following a program.
In order to use computers as fast calculators, scientists ported and adapted existing numerical algorithms and developed new ones. This implied the use of techniques able to approximate the computations of continuous mathematics using computer algebra. However, numbers in computers are not exactly the same as human numbers; in particular, they have finite (and varying) precision.
This intrinsic precision limit has deep consequences in the simulations of nonlinear systems, in particular of chaotic ones. Indeed, chaos was numerically discovered by Lorenz (Oestreicher 2007) after the observation that a simple approximation, a number that was retyped with fewer decimals, caused a macroscopic change in the trajectory under study.
With all their limits, computers can be fruitfully used just to speed up computations that could eventually be performed by humans. However, since the increase in velocity is of several order of magnitude, it becomes possible to include more and more details into the basic model of the phenomenon under investigation, well beyond what would be possible with an army of “human computers.” The idea of exploiting the brute power of fast computers has originated a fruitful line of investigation in numerical physics especially in the field of chemistry, biological molecules, and the structure of matter. The power of computers has allowed, for instance, to include quantum mechanical effects in the computation of the structure of biomolecules (Car and Parrinello 1985), and although these techniques may be targeted as “brute force,” the algorithms developed are actually quite sophisticated.
However, a completely different usage of computers is possible: instead of exploiting them for performing computations of models that already proved to approximate the reality, one can use computers as experimental apparatuses to investigate the patterns of theoretical models , which are generally nonlinear. This is what Lorenz did after having found the first example of chaos in computational physics. He started simplifying his equations in order to enucleate the minimal ingredients of what would be called the butterfly effect .
Much earlier than Lorenz, Fermi, Pasta, and Ulam (and the programmer Tsingou (Daxois et al. 2005)) used one on the very first available computers to investigate the basis of statistical mechanics: how energy distributes among the oscillation modes of a chain of nonlinear oscillators (Fermi et al. 1955).
Also in this case the model is simplified at its maximum, in order to put into evidence what are the fundamental ingredients of the observed pattern and also to use all the available power of computers to increase the precision, the duration, and the size of the simulation.
This simplification is even more fruitful in the study of systems with many degrees of freedom that we may denote generically as extended systems . We humans are not prepared to manipulate more than a few symbols at once. So, unless there is a way of grouping together many parts (using averages, like, for instance, when considering the pressure of a gas as an average over extremely many particle collisions), we are in difficulties in understanding such systems. They may nevertheless be studied performing experiments on computers. Again, the idea is that of simplifying at most the original model, in order to isolate the fundamental ingredients of the observed behavior. It is therefore natural to explore systems whose physics is different from the usual one. These artificial worlds are preferably formulated in discrete terms, more suitable to be implemented in computers (see section “Artificial Worlds”).
This line of investigation is of growing interest today: since modern computers can easily simulate thousands or millions of elementary agents, it may be possible to design artificial worlds in which artificial people behave similarly to real humans. The rules of these worlds are not obtained from the basic laws of the real one, since no computer can at present simulate the behavior of all the elements of even a small portion of matter. These rules are designed so to behave similarly to the system under investigation and to be easily implemented in digital terms. There are two main motivations (or hopes): to be able to understand real complex dynamics by studying simplified models and to be so lucky to discover that a finely tuned model is able to reproduce (or forecast) the behavior of its real counterpart.
This is so promising that many scientists are performing experiments on these artificial worlds, in order to extract their principal characteristics, to be subsequently analyzed using paper and pencil!
In the following, we shall try to elucidate some aspects of the interplay between computer and physics. In section “From Trajectories to Statistics and Back,” we shall illustrate possible logic pathways (in classical mechanics) leading from Newton’s equations to research fields that use computers as investigative tool, like agentbased investigations of human societies. In section “Artificial Worlds,” we shall try to present succinctly some example of artificial systems that are still active research topics in theoretical physics.
From Trajectories to Statistics and Back
The outline of this section is the following. Physics is often denoted the most fundamental science, and one may think that given powerful enough computers, one should be able to reconstruct any experimental situation simply implementing the fundamental laws of physics. I would like to show that any investigation is based on models, requiring approximations and additional assumptions, and that any change of scale implies a change of model. However, an important suggestion from physics is that similar models can be used to interpret different situations, and therefore, the associated computational techniques can be reused in different contexts. We shall follow this line from Newton’s equations to agentbased modeling.
Let us assume that a working model of the reality can be built using a set of dynamical equations, for instance, those of classical mechanics. We shall consider the model of a system formed by many particles , like a solid or a fluid. The state of the resulting system can be represented as a point in a highdimensional space, since it is given by all coordinates and velocities of all particles. The evolution of the system is a trajectory in such space. Clearly, the visualization and the investigation of such a problem is challenging, even using powerful computers.
Moreover, even if we were able to compute many trajectories (in order to have an idea of fluctuations), this would not imply that we have understood the problem. Let us consider, for instance, the meteorology: one is interested in the probability of rain or in the expected wind velocity, not in forecasting the trajectories of all molecules of air. Similarly, in psychology, one is interested in the expected behavior of an individual, not in computing the activity of all neurons in his brain.
Since physics is the oldest discipline that has been quantified into equations, it may be illuminating to follow some of the paths followed by researchers to reduce the complexity of a highdimensional problem to something more manageable or at least simpler to be simulated on a computer.
In particular, we shall see that many approaches consist in projecting the original space onto a limited number of dimensions, corresponding to the observables that vary in a slow and smooth way and assuming that the rest of the dynamics is eventually approximated by “noise. ” The noise can be so small, compared to the macroscopic observables that it can be neglected. In such cases, one has a deterministic, lowdimensional dynamical system, for instance, the usual models for rigid bodies, planets, etc. In the opposite case, the resulting system is stochastic, and one is interested in computing the average values of observables, over the probability distribution of the projected system.
Newton’s Laws
The success of Newton in describing the motion of one body, subjected to a static field force (say, gravitational motion of one planet, oscillation of a body attached to a spring, motion of the pendulum, etc.), clearly proved the validity of his approach and also the validity of using simple models for dealing with natural phenomena. Indeed, the representation of a body as a point mass, the idea of massless springs and strings, and the concept of force fields are all mathematical idealizations of the reality.
The natural generalization of this approach is carried out in the eighteenth century by Lagrange, Hamilton, and many others. It has brought to the mathematizing of the mechanics and the derivation of rational mechanics. The resulting standard (or historical) way of modeling physical systems is that of using differential equations, i.e., a continuous description of time, space, and other dynamical quantities.
From an abstract point of view, one is faced with two different options: either concentrate on systems described by a few equations (lowdimensional systems) or try to describe systems formed by many components.
Low Dimensionality
Historically, the most important problem of Newton’s times was that of three bodies interacting via the gravitational attraction (the Sun, the Earth, and the Moon). By approximating planets with point masses, one gets a small number of coupled differential equations. This reduction of dimensionality is an example of a scale separation: the variables that describe the motion of the planets vary slowly and smoothly in time. Other variables, for instance, those that describe the oscillations of a molecule on the surface of a planet can be approximated by a noise term so small that can be safely neglected. This approximation can also be seen as a “meanfield” approach , for which one assumes that variables behave not too differently from their average. Using these techniques, one can develop models of many systems that result in the same mathematical scheme: a few coupled equations. The resulting equations may clearly have a structure quite different from that resulting from Newtonian dynamics (technically, Hamiltonian systems).
However, the reduction of the number of variables does not guarantee the simplicity of the resulting model. The problem of three gravitational bodies cannot be split into smaller pieces, and the computation of an accurate trajectory requires a computer. In general, a nonlinear system in a space with three or more dimensions is chaotic. This implies that it may react to a small perturbation of the parameters or of the initial conditions with large variations of its trajectory. This sensibility to variation implies the impossibility of predicting its behavior for long times, unless one is content with a probabilistic description.
The reduction from a highdimensional system to a small number of equation can in general be considered the result of a projection operation, similar to studying the dynamics of couple of dancers only looking at his shadow. The problem is of course that of finding the right projection, for instance, that able to show that the couple is formed by two distinct bodies that only occasionally separate.
High Dimensionality
In many cases, the projection operation results in a system still composed by many parts. For instance, models of nonequilibrium fluids neglect to consider the movement of the individual molecules, but still one has to deal with the values of the pressure, density, and velocity in all points. In these cases, one is left with a highdimensional problem. Assuming that the noise of the projected dimensions can be neglected, one can either write down a large number of coupled equation (e.g., in modeling the vibration of a crystal) or use a continuous approach and describe the system by means of partial differential equations (e.g., the model of a fluid).
Linear Systems
In general, high and lowdimensional approaches can be systematically developed (with paper and pencil) only in the linear approximation. Let us illustrate this point for the case of coupled differential equation: if the system is linear, one can write the equations using matrices and vectors. One can in principle find a (linear) transformation of variables that make the system diagonal, i.e., that reduces the problem to a set of uncoupled equations. At this point, one is left with (many) onedimensional independent problems. Clearly, there may be mathematical difficulties, but the path is clear. A similar approach (for instance, using Fourier transforms) can be used also for dealing with partial differential equations.
The variables that result from such operations are called normal modes , because they behave independently one from the other (i.e., they correspond to orthogonal or normal directions in the state space). For instance, the linear model of a vibrating string (with fixed ends) predicts that any pattern can be described as a superposition of “modes,” which are the standing oscillations with zero, one, two, etc. nodes (the harmonics).
However, linear systems behave in a somewhat strange way, from the point of view of thermal physics. Let us consider, for instance, the system composed by two uncoupled oscillators. It is clear that if we excite one oscillator with any amount of energy, it will remain confined to that subsystem. With normal modes, the effect is the same: any amount of energy communicated to a normal mode remains confined to that mode, if the system is completely linear. In other words, the system never forgets its initial conditions.
On the contrary, the longtime behavior of normal systems does not depend strongly on the initial conditions. One example is the timbre or “sound color” of an object. It is given by the simultaneous oscillations on many frequencies, but in general an object emits its characteristic sound regardless of how exactly is excited. This would not be true for linear systems.
Since the distribution of energy to all available modes is one of the assumptions of equilibrium statistical mechanics, which allows us to understand the usual behavior of matter, we arrived at an unpleasant situation: linear systems, which are so “easy” to be studied, cannot be used to ground statistical physics on mechanics.
Molecular Dynamics
Nowadays, we have computers at our disposal, and therefore we can simulate systems composed by many parts with complex interactions. On can do this simply discretizing Newton’s equations of motion so that a digital computer can approximately integrate the set of coupled differential equations. This technique is sometimes called molecular dynamics .
One is generally interested in computing macroscopic quantities. These are defined as averages of some function of the microscopic variables (positions, velocities, accelerations, etc.) of the system. A measurement on a system implies an average, over a finite interval of time and over a large number of elementary components (say, atoms, molecules, etc.) of some quantity that depends on the microscopic state of that portion of the body.
Chaos and Probability
It was realized by Poincaré (with paper and pencil) and Lorenz (with one of the very first computers) that also very few (three) coupled differential equations with nonlinear interactions may give origin to complex (chaotic) behavior. In a chaotic system, a small uncertainty amplifies exponentially in time, making forecasting difficult. However, chaos may also be simple: the equations describing the trajectory of dice are almost surely chaotic, but in this case the chaos is so strong that the tiniest perturbation or uncertainty in the initial conditions will cause in a very small amount of time a complete variation of the trajectory. Our experience says that the process is well approximated by a probabilistic description. Therefore, chaos is one possible way of introducing probability in dynamics.
Chaotic behavior may be obtained in simpler models, called maps, that evolve in discrete time steps. As May (1976) showed, a simple map with a quadratic nonlinearity (logistic map) may be chaotic. One can also model a system using coupled maps instead of a system of coupled differential equations (Kaneko 1985). And indeed, when a continuous system is simulated on a computer, it is always represented as an array of coupled maps.
Computers are also used to simulate stochastic systems, which at first may seem quite strange: how can a deterministic system, like a computer, generate random data? Although there is the possibility of using unpredictable events (like network data), amplifying thermal noise or even use quantummechanical events for generating true random numbers, in most of cases the pseudorandom number generators just use a chaotic map, properly initialized (for instance, using a true random number).
Discretization
Many extended systems (for instance, fluids) are described by means of partial differential equations , which are however quite difficult to be handled by computers. When possible, it is more efficient to reduce the original system to a (small) number of ordinary differential equations that are then converted to discretetime maps.
There is a therefore a progression of discretization from partial differential equations, coupled differential equations, coupled map lattices: from systems that are continuous in space, in time, and in the dynamical variables to systems that are discrete in time and space and continuous only in the dynamical variables. The further logic step is that of studying completely discrete systems, called cellular automata .
Cellular automata show a wide variety of different phenomenology. They can be considered mathematical tools or used to model reality. In many cases, the resulting phenomenological models follow probabilistic rules, but it is also possible to use cellular automata as building blocks. For instance, it is possible to simulate the behavior of a hydrodynamic system by means of a completely discrete model, called cellular automata lattice gas (Frisch et al. 1986; Hardy et al. 1973).
Statistics
The investigation of chaotic extended systems proceeds generally using a statistical approach. The idea is the following: any system contains a certain degree of nonlinearity that couples otherwise independent normal modes. Therefore, (one hopes that) the initial condition is not too important for the asymptotic regime. If moreover one assumes that the motion is so chaotic that any trajectory spans the available space in a characteristic way (again, not depending on the initial conditions), we can use statistics to derive the characteristic probability distribution : the probability of finding the system in a given portion of the available space is proportional to the time that the system spends in that region. See also the paragraph on equilibrium.
Random Walks
Another approach is that of focusing on a small part of a system, for instance, a single particle. The rest of the system is approximated by noise. This method was applied, for instance, by Einstein in the development of the simplest theory of Brownian motion, the random walk (Haw 2005). In random walks, each step of the target particle is independent on previous steps, due to collisions with the rest of particles. Collisions, moreover, are supposed to be uncorrelated. A more sophisticated approximation consists in keeping some aspects of motion, for instance, the influence of inertia or of external forces, still approximating the rest of the world by noise (which may contain a certain degree of correlation). This is known as the Langevin approach, which includes the random walk as the simplest case. Langevin equations are stochastic differential equations.
The essence of this method relies in the assumption that the behavior of the various parts of the systems is uncorrelated. This assumption is vital also for other types of approximations that will be illustrated in the following. Notice that in the statistical mechanics approach, this assumption is not required.
In the Langevin formulation, by averaging over many independent realizations of the process (which in general is not the same of averaging over many particles that simultaneously move, due, for instance, to excluded volumes), one obtains the evolution equation of the probability of finding a particle in a given portion of space. This is the Kolmogorov integraldifferential equation that in many case can be simplified, giving a differential (FokkerPlank) equation. The diffusion equation is just the simplest case (Gardiner 1994; van Kampen 1992).
It is worth noticing that a similar formulation may be developed for quantum systems: the Feynman pathintegral approach is essentially a Langevin formulation, and the Schroedinger equation is the corresponding FokkerPlank equation.
Random walks and stochastic differential equations find many applications in economics, mainly in stock market simulations. In these cases, one is not interested in the average behavior of the market, but rather in computing nonlinear quantities over trajectories (time series of good values, fluctuations, etc.).
TimeSeries Data
In practice, a model is never derived ab initio, by projecting the dynamics of all the microscopic components onto a limited number of dimensions, but is constructed heuristically from observations of the behavior of a real system.
It is therefore crucial to investigate how observations are made, i.e., the analysis of a series of time measurements. In particular, a good exercise is that of simulating a dynamical or stochastic system, analyze the resulting timeseries data of a given observable, and see if one is able to reconstruct from it the relationships or the equations ruling the time evolution.
Let us consider the experimental study of a chaotic, lowdimensional system. The measurements on this system give a time series of values that we assume discrete (which is actually the case considering experimental errors). Therefore, the output of our experiment is a series of symbols or numbers, a time series. Let us assume that the system is stationary, i.e., that the sequence is statistically homogeneous in time. If the system is not extremely chaotic, symbols in the sequence are correlated, and one can derive the probability of observing single symbols, couples of symbols, triples of symbols, etc. There is a hierarchy in these probabilities, since the knowledge of the distribution of triples allows the computation of the distribution of couples, and so on.
It can be shown that the knowledge of the probability distribution of the infinite sequence is equivalent to the complete knowledge of the dynamics. However, this would correspond to performing an infinite number of experiments, for all possible initial conditions.
The usual investigation scheme assumes that correlations vanish beyond a certain distance, which is equivalent to assume that the probability of observing sequences longer than that distance factorize. Therefore, one tries to model the evolution of the system by a probabilistic dynamics of symbols as shown in section “Probabilistic Cellular Automata.” Timeseries data analysis can therefore be considered as the main experimental motivation in developing probabilistic discrete models. This can be done heuristically comparing results with observations a posteriori or trying to extract the rules directly from data, like in the Markov approach.
Markov Approximation
The Markov approach, either continuous or discrete, also assumes that the memory of the system vanishes after a certain time interval, i.e., that the correlations in time series decay exponentially. In discrete terms, one tries to describe the process under study as an automaton, with given transition probabilities. The main problem is, given a sequence of symbols, what is the simplest automaton (hidden Markov chains (Rabiner 1989)) that can generate that sequence with maximum “predictability,” i.e., with transition probabilities that are nearest to zero or one? Again, it is possible to derive a completely deterministic automaton, but in general it has a number of states equivalent to the length of the time series, so it is not generalizable and has no predictability (see also section “Probabilistic Cellular Automata”). On the contrary, an automaton with a very small number of nodes will have typically intermediate transition probabilities, so predictability is again low (essentially equivalent to random extraction). Therefore, the good model is the result of an optimization problem that can be studied using, for instance, Monte Carlo techniques.
Mean Field
Finally, from the probabilities one can compute averages of observables, fluctuations, and other quantities called moments of the distribution. Actually, the knowledge of all moments is equivalent to the knowledge of the whole distribution. Therefore, another approach is that of relating moments at different times or different locations, truncating the recurrences at a certain level. The roughest approximation is that of truncating the relations at the level of averages, i.e., the meanfield approach. It appears so natural that it is often used without realizing the implications of the approximations. For instances, chemical equations are essentially meanfield approximations of a complex phenomenon.
Boltzmann Equation
Another similar approach is that of dividing an extended system into zones and assume that the behavior of the system in each zone is well described by a probability distribution. By disregarding correlations with other zones, one obtains the Boltzmann equation, with which many transport phenomena may be studied well beyond the elementary kinetic theory. The Boltzmann equation can also be obtained from the truncation of a hierarchy of equations (BBGKY hierarchy) relating multiparticle probability distributions. Therefore, the Boltzmann equation is similar in spirit to a meanfield analysis.
Equilibrium
One of the biggest successes of the stochastic approach is equilibrium statistical mechanics . The main ingredient of this approach is that of minimum information, which, in other words, corresponds to the assumption: what is not known is not harmful. By supposing that at equilibrium the probability distribution of the systems maximizes the information entropy (corresponding to a minimum of information on the system), one is capable of deriving the probability distribution itself and therefore the expected values of observables (ensemble averages, see section “Ising Model”). In this way, using an explicit model, one is capable to compute the value of the parameters that appear in thermodynamics. If it were possible to show that the maximum entropy state is actually the state originated by the dynamics of a mechanical (or quantum) system, one could ground thermodynamics on mechanics. This is a longinvestigated subject, dating back to Boltzmann, which is however not yet clarified. The main drawback in the derivations is about ergodicity. Roughly speaking, a system is called ergodic if the infinitetime average of an observable over a trajectory coincides with its average over a snapshot of infinitely many replicas. For a system with fixed energy and no other conserved quantities, a sufficient condition is that a generic trajectory passes near all points of the accessible phase space. However, most systems whose behavior is “experimentally” well approximated by statistical mechanics are not ergodic. Moreover, another ingredient, the capability of quickly forgetting the information about initial conditions appears to be required; otherwise trajectories are strongly correlated, and averages over different trajectories cannot be “mixed” together. This capability is strongly connected to the chaoticity or unpredictability of extended systems, but unfortunately these ingredients make analytic approaches quite hard.
An alternative approach, due to Jaynes (1957), is much more pragmatic. In essence, it says: let design a model with the ingredients that one thinks are important, and assume that what is not in the model does not affect its statistical properties. Compute the distribution that maximizes the entropy with the given constraints. Then, compare the results (averages of observables) with experiments (possibly, numerical ones). If they agree, one has captured the essence of the problem; otherwise one has to include some other ingredients and repeat the procedure. Clearly, this approach is much more general than the “dynamical” one, not considering trajectory or making assumptions about the energy, which is simply seen as a constraint. But physicists would be much more satisfied by a microscopic derivation of statistical physics.
In spite of this lack of strong basis, the statistical mechanics approach is quite powerful, especially for systems that can be reduced to the case of almost independent elements. In this situation, the probability distribution of the system (the partition function ) factorizes, and many computations may be performed by hand. Notice however that this behavior is in strong contrast to that of truly linear systems: the “almost” attribute indicates that actually the elements interact and therefore share the same “temperature.”
Monte Carlo
The Monte Carlo technique was invented for computing, with the aid of a computer, thermal averages of observables of physical systems at equilibrium. Since then, this term is often used to denote the technique of computing the average values of observables of a stochastic system by computing the timeaverage values over artificial trajectories.
In equilibrium statistical physics, one is faced by the problem of computing averages of observables over the probability distribution of the system, and since the phase space is very high dimensional, this is in general not an easy task: one cannot simply draw random configurations, because in general they are so different from those typical of the given value of the temperature, that their statistical weight is marginal. And one does not want to revert to the original, stillmorehighly dimensional dynamical system, which typically requires powerful computers just to be followed for tiny time intervals.
First of all, one can divide (separate) the model into almost independent subsystems that, due to the small residual interactions (the “almost” independency), are at the same temperature. In the typical example of a gas, the velocity components appear into the formula of energy as additive terms, i.e., they do not interact with themselves or with other variables. Therefore, they can be studied separately giving the Maxwell distribution of velocities. The positions of molecules, however, are linked by the potential energy (except in the case of an ideal gas), and so the hard part of the computation is that of generating configurations. Secondly, statistical mechanics guarantees that the asymptotic probability distribution does not depend on the details of dynamics. Therefore, one is free to look for the simplest dynamics still compatible with constraints. The Monte Carlo computation is just a set of recipes for generating such trajectories. In many problems, this approach allows to reduce the (computational) complexity of the problem of several orders of magnitude, allowing to generate artificial trajectories that span statistically significant configuration with small computational effort. In parallel with the generation of the trajectory, one can compute the value of several observables and perform statistical analysis on them, in particular the computation of time averages and fluctuations.
By extension, the same terms Monte Carlo is used for the technique of generating sequences of states (trajectories) given the transition probabilities, and computing averages of observables on trajectories, instead of over the probability distribution.
Stochastic Optimization
One of the most interesting applications of Monte Carlo simulations concerns stochastic optimization via simulated annealing . The idea is that of exploiting an analogy between the status of a system (and its energy) and the coding of a particular procedure with corresponding cost function. The goal is that of finding the best solution, i.e., the global minimum of the energy given the constraints. Easy systems have a smooth energy landscape, shaped like a funnel, so that usual techniques like that of always choosing the displacements that locally lowers the energy (gradient descent) are successful. However, when the energy landscape is corrugated, there are many local minima where algorithms like gradient descent tend to get trapped. Methods from statistical mechanics (Monte Carlo), on the contrary, are targeted to generating trajectories that quickly explore the relevant parts of the state space, i.e., those that correspond to the largest contributions to the probability distribution that depends on the temperature, an external or control parameter. If the temperature is high, the probability distribution is broad and the generated trajectory does not “see” the minima of energy that are below the temperature, i.e., it can jump over and off the local minima.
By lowering the temperature, the probability distribution of system obeying statistical mechanics concentrates around minima of energy, and the Monte Carlo trajectory does the same. The energy (or cost) function of not extremely complex problems is shaped in such a way that the global optimum belongs to a broad valley, so that this lowering of the temperature increases the probability of finding it.
Therefore, a sufficiently slow annealing should furnish the desired global minimum. Moreover, it is possible to convert constraints into energy terms, which is quite convenient since or many problems, it is difficult to generate configurations that satisfy the constraints. Let us think, for instance, to the problem of generating a school timetable, keeping into consideration that lessons should not last more than three consecutive hours, that a teacher or students cannot stay in two classes at the same time, and that a teacher is not available on Monday, another prefers the first hours, etc. It is difficult to generate generic configurations that obey all constraints, while it is easy to formulate a Monte Carlo algorithm that generates arbitrary configurations, weighting them with a factor that depends on how many constraints are violated. At high temperature, constraints do not forbid the exploration of the state space and therefore to try creative solutions. At low temperature, constraints become important. At zero temperature, the configurations with lower energy are those that satisfy all constraints, if possible, or at least the largest part of them.
In recent years, physics have dealt with extremely complex problems (e.g., spin glasses (Dotsenko 1994; Mezard et al. 1987)), in which the energy landscape is extremely rough. Special techniques based on nonmonotonous “walks” on temperature have been developed (simulated tempering (Marinari and Parisi 1992)).
Critical Phenomena
One of the most investigated topics of statistical mechanics concerns phase transitions . This is a fascinating subject: in the vicinity of a continuous phase transitions, correlation lengths diverge, and the system behave collectively, in a way which is largely independent of the details of the model. This universal behavior allows the use of extremely simplified models that therefore can be massively simulated.
The philosophy of statistical mechanics may be exported to nonequilibrium systems: systems with absorbing states (that correspond to infinitely negative energy), driven systems (live living ones), strongly frustrated systems (that actually never reach equilibrium), etc. In general, one defines these systems in terms of transition probabilities, not in terms of energy. Therefore, one cannot invoke a maximum entropy principle, and the results are less general.
However, many systems exhibit behavior reminiscent of equilibrium systems, and the same language can be used: phase transitions, correlations, susceptibilities, etc. These characteristics, common to many different models, are sometimes referred as emergent features .
One of the most famous problems in this field is percolation: the formations of giant clusters in systems described by a local aggregation dynamics, for instance, adding links to a set of nodes. This basic model has been used to describe an incredibly large range of phenomena (Stauffer and Aharony 1994).
Equilibrium and nonequilibrium phase transitions occur for a welldefined value of a control parameter. However, in nature one often observes phenomena whose characteristics resemble that of a system near a phase transition, critical dynamics , without any finetuned parameter. For such system, the term selforganized criticality has been coined (Bak et al. 1987), and they are the subject of active researches.
Networks
A recent extension of statistical physics is the theory of networks. Networks in physics are often regular, like the lattice of a crystal, or only slightly randomized. Characteristics of these networks are the fixed (or slightly dispersed around the mean) number of connections per node, the high probability of having connected neighbors (number of “triangles”), and the large time needed to cross the network. The opposite of a regular network is a random graph, which, for the same number of connections, exhibits low number of triangles and short crossing time. The statistical and dynamical properties of systems whose connection are regular or random are generally quite different.
Watts and Strogatz (1998) argued that social networks are never completely regular. They showed that the simple random rewiring of a small number of links in a regular network may induce the small world effect : local properties, like the number of triangles, are not affected, but largedistance ones, like the crossing time, quickly became similar to that of random graphs. Also the statistical and dynamical properties of models defined over a rewired network are generally similar to those correlated to random graphs.
After this finding, many social networks were studied, and they revealed a yet different structure: instead of having a welldefined connectivity, many of them present a few highly connected “hubs” and a lot of poorly connected “leafs.” The distribution of connectivity is often shaped as a power law (or similar (Newman 2005)), without a welldefined mean (scalefree networks (Albert and Barabasi 2002)). Many of phenomenological models are presently reexamined in order to investigate their behavior over such networks. Moreover, scalefree networks cannot be laid down; they need to be grown following a procedure, similar in this to fractals. It is natural therefore to include such a procedure in the model, so that they not only evolve over the networks but also evolve the network (Boccaletti et al. 2006).
Agents
Many of the described tools are used in the socalled agentbased modeling. The idea is that of exploiting the powerful capabilities of present computers to simulate directly a large number of agents that interact among them. Traditional investigations of complex systems, like crowds, flocks, traffic, and urban models, have been performed using homogeneous representation: partial differential equations (i.e., mean field), Markov equations, cellular automata, etc. In such an approach, it is supposed that each agent type is present in many identical copies, and therefore they are simulated as macrovariables (cellular automata ) or aggregated like independent random walkers in the diffusion equation. But live elements (cells, organisms) do not behave in such a way: they are often individually unique, carry information about their own past history, and so on. With computers, we are now in the position of simulating large assemblies of individuals, possibly geographically located, like, for instance, humans in an urban simulation.
One of the advantages of such approach is that of offering the possibility of measuring quantities that are inaccessible to field researchers and also playing with different scenarios. The disadvantages are the proliferation of parameters that are often beyond experimental confirmation.
Artificial Worlds
A somewhat alternative approach to that of traditional computational physics is that of studying an artificial model, built with little or no direct connection with reality, trying to include only those aspects that are considered relevant. The goal is to be able to find the simplest system still able to exhibit the relevant features of the phenomena under investigation. The resulting models, though not directly applicable to the interpretation of experiments, may serve as interpretative tools in many different situations. For instance, the Ising model was developed in the context of the modeling of magnetic systems but has been applied to opinion formation, social simulations, etc.
Ising Model
The Ising (better: LenzIsing) model is probably one of the most known models in statistical physics. Its history (Niss 2005) is particularly illuminating in this context, even if it took place well before the advent of computers in physics. It is also a model for which the Monte Carlo and simulated annealing techniques are readily applied.
Let us first illustrate schematically the model. I shall present the traditional version, with the terminology that arises from the physics of magnetic systems. However, it is an interesting exercise to reformulate it in the context, for instance, of opinion formation. Let us simply replace “spin up/down” with “opinion A/B,” “magnetization” with “average opinion,” “coupling” with “exchange of ideas,” “external magnetic field” with “propaganda,” and so on.
The quantity E(x) can be thought as a landscape, with lowenergy configurations corresponding to valley and highenergy ones to peaks. The distribution P(x) can be interpreted as the density of a gas, each “particle” corresponding to a possible realization (a replica) of the system. This gas concentrates in the valleys for low temperatures and diffuses if the temperature is increased. The temperature is related to the average level of the gas.
With coupling turned on (J > 0), it may happen that the local field is strong enough to “resist” H, i.e., a compact patch of spins oriented against H may be stable, even if the energy could be lowered by flipping all them, because the flip of a single spin would rise the energy (actually, this flip may happen but is statistically reabsorbed in short times). The fate of the patch is governed by boundaries. A spin on a boundary of a patch feels a weaker local field, since some of its neighbors are oriented in the opposite. Straight boundaries in two or more dimensions separate spins that “know” the phase they belong to: since most of their neighbors are in that phase, the spins on the edges may flip more freely. Stripes that span the whole lattice are rather stable objects and may resist an opposite external field since spins that occasionally flip are surrounded by spins belonging to the opposite phase and therefore feel a strong local field that pushes them towards the phase opposed to the external field.
In one dimension with finiterange coupling, a single spin flip is able to create a “stripe” (perpendicularly to the lattice dimension) and therefore can destabilize the ordered phase. This is the main reason for the absence of phase transitions in one dimension, unless the coupling extends on very large distances or some coupling is infinite (see the part on directed percolation, section “Probabilistic Cellular Automata”).
This model was proposed in the early 1920s by Lenz to Ising for his Ph.D. dissertation as a simple model of a ferromagnet. Ising studied it in one dimension, found that it shows no phase transition, and concluded (erroneously) that the same happened in higher dimensions. Most of the contemporaries rejected the model since it was not based on Heisenberg’s quantum mechanical model of ferromagnetic interactions. It was only in the forties that it started gaining popularity as a model of cooperative phenomena, a prototype of orderdisorder transitions. Finally, in 1940, Onsager (1944) provided the exact solution of the twodimensional LenzIsing model in zero external field. It was the first (and for many years the only) model exhibiting a nontrivial secondorder transition whose behavior could be exactly computed.
The divergence of the correlation length indicates that there is no characteristic scale (ξ), and therefore fluctuations of all sizes appear. In this case, the details of the interactions are not so important, so that many different models behave in the same way, for what concerns, for instance, the critical exponents. Therefore, models can be grouped into universality classes, whose details are essentially given by “robust” characteristics like the dimensionality of space and of the order parameter, the symmetries, etc.
The powerlaw behavior of the correlation function also indicates that if we perform a rescaling of the system, it would appear the same or, conversely, that one is unable to estimate the distance of a pattern by comparing the “typical size” of particulars. This scale invariance is typical of many natural phenomena, from clouds (whose height and size are hard to be estimated), to trees and other plant elements, lungs, brain, etc.
Many examples of power laws and collective behavior can be found in natural sciences (Sornette 2006). Differently from what happens in the LenzIsing model, in these cases there is no parameter (like the temperature) that has to be finetuned, so one speaks of selforganized criticality (Bak et al. 1987).
Since the LenzIsing model is so simple, exhibits a critical phase, and can be exactly solved (in some cases), it has become the playground for a variety of modifications and applications to various fields. Clearly, most of the modifications do not allow analytical treatment and have to be investigated numerically. The Monte Carlo method allows to add a temporal dimension to a statistical model (Kawasaki 1972), i.e., to transform stochastic integrals into averages over fictitious trajectories. Needless to say, the LenzIsing model is the standard test for every Monte Carlo beginner, and most of the techniques for accelerating the convergence of averages have been developed with this model in mind (Swendsen and Wang 1987).
Near a secondorder phase transition , a physical system exhibits critical slowing down , i.e., it reacts to external perturbations with an extremely slow dynamics, with a convergence time that increases with the system size. One can extend the definition of the correlation function including the time dimension: in the critical phase also the temporal correlation length diverges (as a power law). This happens also for the LenzIsing model using the Monte Carlo dynamics, unless very special techniques are used (Swendsen and Wang 1987). Therefore, the dynamical version of the LenzIsing model can be used also to investigate relaxational dynamics and how this is influenced by the characteristics of the energy landscape. In particular, if the coupling J _{ ij } changes sign randomly for every couple of sites (or the field H has random sign for each site), the energy landscape becomes extremely rugged. When spins flip in order to align to the local field, they may invert the field felt by neighboring ones. This frustration effect is believed to be the basic mechanism of the large viscosity and memory effects of glassy substances (Dotsenko 1994; Mezard et al. 1987).
The rough energy landscape of glassy systems is also challenging for optimization methods, like simulated annealing (Kirkpatrick et al. 1983) and its “improved” cousin, simulated tempering (Marinari and Parisi 1992). Again, the LenzIsing model is the natural playground for these algorithms.
The dynamical LenzIsing model can be formulated such that each spin is updated in parallel (Barkema and MacFarland 1994) (with the precaution of dividing cells into sublattices, in order to keep the neighborhood of each cell fixed during updates). In this way, it is essentially a probabilistic cellular automata, as illustrated in section “Probabilistic Cellular Automata.”
Cellular Automata
In the same period in which traditional computation was developed, in the early 1950s, John Von Neumann was interested in the logic basis of life and in particular in selfreproduction, and since the analysis of a selfreproduction automata following the rules of real physics was too difficult, he designed a playground (a cellular automaton) with just enough “physical rules” in order to make its analysis possible (von Neumann and Burks, 1966). It was however just a theoretical exercise; the automaton was so huge that up to now it has not yet completely implemented (Von Neumann universal constructor 2008).
The idea of cellular automata is quite simple: take a lattice (or a graph) and put on each cell an automaton (all automata are equal). Each automaton exhibits its “state” (which is one out of a small number) and is programmed to react (change state) according to the state of neighbors and its present one (the evolution rule). All automata update their state synchronously.
Cellular automata share many similarities with the parallel version of the LenzIsing model. Differently from that, their dynamics is not derived from an energy, but is defined in terms of the transition rules. These rules may be deterministic or probabilistic. In the first case (illustrated in this section), cellular automata are fully discrete, extended dynamical systems. Probabilistic cellular automata are illustrated in section “Probabilistic Cellular Automata.” The temporal evolution of deterministic cellular automata can be computed exactly (regardless of any approximation) on a standard computer.
Let us illustrate the simplest case, elementary cellular automata, in Wolfram’s jargon (1983). The lattice here is onedimensional, so to identify an automaton it is sufficient to give one coordinate, say i, with i = 1, …, N. The state of the automaton on cell i at time t is represented by a single variable, x _{ i }(t), that can take only two values, “dead/live” or “inactive/active” or 0/1. The time is also discrete, so t = 1, 2, …
Since x _{ i } = 0,1, there are only eight possible combinations of the triple {x _{ i−1 }(t), x _{ i }(t), x _{ i+1 }(t)}, from {0,0,0} to {1,1,1}. For each of them, f(x _{ i−1 }(t), x _{ i }(t), x _{ i+1 }(t)) is either zero or one, so the function f can be simply coded as a vector of eight bits, each position labeled by a different configuration of inputs. Therefore, there are only 2^{8} = 256 different elementary cellular automata that have been studied carefully (see, for instance, Wolfram (1983)).
Complex CA have large transients, during which interesting structures may emerge. They finally relax into class1 automata. It has been conjectured that they are able of computation, i.e., that one can “design” a universal computer using these CA as building blocks, as has been proved to be possible with the Game of Life. Another hypothesis, again confirmed by the Game of Life, is that these automata are “near the edge” of selforganizing complexity. One can slightly “randomize” the Game of Life, allowing sometimes an exception to the rule. Let us introduce a parameter p that measures this randomness, with the assumption that p = 0 is the true “life.” Well, it was shown (Nordfalk and Alstrøm 1996) that the resulting model exhibits a secondorder phase transition for a value of p very near zero.
Deterministic cellular automata have been investigated as prototypes of discrete dynamical systems, in particular for what concerns the definition of chaos. Visually, one is tempted to use this word also to denote the irregular behavior of “class3” rules. However, the usual definition of chaos involves the sensitivity to an infinitesimally small perturbation: following the time dynamics of two initially close configurations, one can observe an amplification of their distance. If the initial distance (δ _{0}) is infinitesimal, then the distance grows exponentially for some time (δ(t) ≃ δ _{0} exp.(λt)), after which it tends to saturate, since the trajectories are generally bounded inside an attractor or due to the dimensions of the accessible space. The exponent λ depends on the initial configuration, and if this behavior is observed for different portions of the trajectory, it fluctuates: a trajectory spends some time in regions of high chaoticity, after which may pass through “quiet” zones. If one “renormalizes” periodically this distance, considering one system as the “master” and the other as a measuring device, one can accumulate good statistics and define a Lyapunov exponent λ that gives indications about the chaoticity of the trajectory, through a limiting procedure.
The accuracy of computation poses some problems. Since in computer numbers are always approximate, one cannot follow just “one” trajectory. The small approximations accumulate exponentially, and the computer time series actually jumps among neighboring trajectories. Since the Lyapunov exponent is generally not so sensible to a change of precision in computation, one can assume that the chaotic regions are rather compact and uniform, so that in general one associates a Lyapunov exponent to a system, not to an individual trajectory. Nevertheless, this definition cannot apply to completely discrete systems like cellular automata.
In any case, chaoticity is related to unpredictability. As first observed by Lorenz, and following the definition of Lyapunov exponent, the precision of an observation over a chaotic system is related to the average time for which predictions are possible. Like in weather forecasts, in order to increase the time span of a prediction, one has to increase the precision of the initial measurement. In extended system, this also implies to extend the measurements over a larger area. One can also consider a “synchronization” approach. Take two replicas of a system and let them evolve starting from different initial configurations. With a frequency and a strength that depends on a parameter q, one of these replicas is “pushed” towards the other one, so as to reduce their distance. Suppose that q = 0 is the case of no push and q = 1 is the case of extremal push, for which the two systems synchronize in a very short time. There should be a critical value q _{ c } that separates these two behaviors (actually, the scenario may be more complex, with many phases (Bagnoli and Cecconi 2001)). In the vicinity of q _{ c }, the distance between the two replicas is small, and the distance δ grows exponentially. The critical value q _{ c } is such that the exponential growth compensates the shrinking factor and is therefore related to the Lyapunov exponent λ.
Finitesize cellular automata always follow periodic trajectories. Let us consider, for instance, Boolean automata of N cells. The number of possible different states is 2 N and due to determinism, once that a state has been visited twice the automata has entered a limit cycle (or a fixed point). One may have limit cycles with large basins of transient configurations (configurations that do not belong to the cycle). Many scenarios are possible. The set of different configurations may be divided in many basins, of small size (small transient) and small period, like in class1 and class2 automata. Or one may have large basins, with long transients that lead to short cycles, like in class4 automata. Finally, one may have one or very few large basins, with long cycles that include most of the configurations belonging to the basin (small transients). This is the case of class3 automata. For them, the typical period of a limit cycle grows exponentially, like the total number of configurations, with the system size, so that for moderately large system it is almost impossible to observe a whole cycle in a finite time. Another common characteristic of class3 automata is that the configurations quickly decorrelate (in the sense of the correlation function) along a trajectory. If one takes into consideration as starting points two configurations that are the same but for a local difference, one observes that this difference amplifies and diffuses in class3 automata, shrinks or remains limited in class 1 and class 2, and has an erratic transient behavior in class 4, followed by the fate of class 1 and class 2. Therefore, if one considers the possibility of not knowing exactly the initial configuration of an automaton, unpredictability grows with time also for such discrete systems. Actually, also (partially) continuous systems like coupled maps may exhibit this kind of behavior (Bagnoli and Cecconi 2001; Cecconi et al. 1998; Crutchfield and Kaneko 1988; Politi et al. 1993). Along this line, it is possible to define an equivalent of the Lyapunov exponents for CA (Bagnoli et al. 1992). The synchronization procedure can be applied also to cellular automata, and it correlates well with the Lyapunov exponents (Bagnoli and Rechtman 1999).
An “industrial” application of cellular automata is their use for modeling gases. The hydrodynamical equations, like the NavierStokes ones, simply reflect the conservation of mass, momentum, and energy (i.e., rotational, translational, and time invariance) for the microscopic collision rules among particles. Since the modeling of a gas via molecular dynamics is rather cumbersome, some years ago it was proposed (Frisch et al. 1986; Hardy et al. 1973) to simplify drastically the microscopic dynamics using particles that may travel only along certain directions with some discrete velocities and jumping in discrete time only among nodes of a lattice, indeed, a cellular automaton. It has been shown that their macroscopic dynamics is described by usual hydrodynamics laws (with some odd features related to the underlining lattice and finiteness of velocities) (Rothman and Zaleski 2004; WolfGladrow 2004).
The hope was that these Lattice Gas Cellular Automata (LGCA) could be simulated so efficiently in hardware to make possible the investigation of turbulence, or, in other words, that they could constitute the Ising model of hydrodynamics. While they are indeed useful to investigate certain properties of gases (for instance, chemical reactions (Lawniczak et al. 1991), or the relationship between chaoticity and equilibrium (Bagnoli and Rechtman 2009)), they resulted too noisy and too viscous to be useful for the investigation of turbulence. Viscosity is related to the transport of momentum in a direction perpendicular to the momentum itself. If the collision rule does not “spread” quickly the particles, the viscosity is large. In LGCA there are many limitations to collisions, so that in order to lower viscosity one has to consider averages over large patches, thus lowering the efficiency of the method.
However, LGCA inspired a very interesting approximation. Let us consider a large assembly of replicas of the same system, each one starting from a different initial configuration, all compatible with the same macroscopic initial conditions. The macroscopic behavior after a certain time would be the average over the status of all these replicas. If one assumes a form of local equilibrium, i.e., applies the meanfield approximation for a given site, one may try to obtain the dynamics of the average distribution of particles, which in principle is the same of “exchanging” particles that happen to stay on the same node among replicas.
It is possible to express the dynamics of the average distribution in a simple form: it is the Lattice Boltzmann Equation (LBE) (Chopard et al. 2002; Succi 2001; WolfGladrow 2004). The method retains many properties of LGCA like the possibility of considering irregular and varying boundaries and may be simulated in a very efficient way with parallel machines (Succi 2001). Differently from LGCA, there are numerical stability problems to be overcome.
Probabilistic Cellular Automata
In deterministic automata, given a local configuration, the future state of a cell is univocally determined. However, let us consider the case of measuring experimentally some pattern and trying to analyze it in terms of cellular automata. In timeseries analysis, it is common to perform averages over spatial patches and temporal intervals and to discretize the resulting value. For instance, this is the natural result of using a camera to record the temporal evolution of an extended system, for instance the turbulent and laminar regions of a fluid. The resulting pattern symbolically represents the dynamics of the original system, and if it is possible to extract a “rule” out of this pattern, it would be extremely interesting for the construction of a model. In general, however, one observes that sometimes a local configuration is followed by a symbol, and sometimes the same local configuration is followed by another one. One should conclude that the neighborhood (the local configuration) does not univocally determine the following symbol.
One can extend the “range” of the rule, adding more neighbors farther in space and time (Rabiner 1989). By doing so, the “conflicts” generally reduce but at the price of increasing the complexity of the rule. At the extremum, one could have an automaton with infinite “memory” in time and space that perfectly reproduces the observed patterns but with almost nonpredictive power, since it is extremely unlucky that the same huge local configuration is encountered again.
So, one may prefer to limit the neighborhood to some finite extension and accept that the rule sometimes “outputs” a symbol and sometimes another one. One defines a local transition probability τ(x _{ i }(t + 1)X _{ i }(t)) of obtaining a certain symbol x _{ i } at time t + 1 given a local configuration X _{ i } at time t. Deterministic cellular automata correspond to the case τ = 0,1. The parallel version of the LenzIsing model can be reinterpreted as a probabilistic cellular automaton .
Starting from the local transition probabilities, one can build up the transition probability T(xy) of obtaining a configuration x given a configuration y. T(xy) is given by the product of the local transition probabilities τ (Bagnoli 2000). One can read the configurations x and y as indexes, so that T can be considered as a matrix. The normalization of probability corresponds to the constraint Σ_{ x } T ( xy) = 1 ∀ y.
For the parallel LenzIsing model, the elements of the matrix T are given by the product of local transition rules of the Monte Carlo dynamics. They depend on the choice of the algorithm but essentially have the form of exponentials of the difference in energy, divided by the temperature. Although a definitive proof is still missing, it is plausible that matrices with all elements different from zero correspond to some equilibrium model, whose transition rules can be derived from an energy function (Georges and le Doussal 1989).
Since probabilistic cellular automata are defined in terms of the transition probabilities, one is free to investigate models that go beyond equilibrium. For instance, if some transition probability takes the value zero or one, in the language of equilibrium system, this would correspond to some coupling (like the coupling factor J of the LenzIsing model) that becomes infinite. This case is not so uncommon in modeling. The inverse of a transition probability corresponds to the average waiting time for the transition to occur in a continuoustime model (one may think to chemical reactions). Some transitions may have a waiting time so long with respect to the observation interval, to be practically irreversible. Therefore, probabilistic cellular automata (alongside other approaches like for instance annihilating random walks) allow the exploration of outofequilibrium phenomena .
The parallel version of a directed percolation process can be mapped onto probabilistic cellular automata. The simplest case, in one spatial dimension and with just two neighbors, is called the DomanyKinzel model (Domany and Kinzel 1984). It is even more general than the usual directed percolation, allowing “nonlinear” interactions among sites in the neighborhood (e.g., two wet sites may have less probability of percolating than one alone).
These processes are interesting because there is an absorbing state (Bagnoli et al. 2001; Hinrichsen 1997), which is the dry state for the wetting phenomenon and the healthy state for the spreading of epidemics. Once the system has entered this absorbing state, it cannot exit, since the spontaneous appearing of a wet site or of an ill individual is forbidden. For finite systems, the theory of Markov chains says that the system will “encounter,” sooner or later, this state that therefore corresponds to the unique asymptotic state. For infinitely extended systems, a phase transition can occur, for which wet sites percolate for all “times,” the epidemics become endemic, and so on.
Again these are examples of critical phenomena, with exponents different from the equilibrium case.
AgentBased Simulations
Cellular automata can be useful in modeling phenomena that can be described in lattice terms. However, many phenomena require moving particles. Examples may be chemical reactions, ecological simulations, and social models. When particles are required to obey hydrodynamics constraints, i.e., to collide conserving mass, momentum, and energy, one can use lattice gas cellular automata or the lattice Boltzmann equation. However, in general one is interested in modeling just a macroscopic scale, assuming that what happens at lower level is just “noise”. In according with the degree of complexity (or “intelligence”) assigned to particles, one can develop models based on the concept of walkers that move more or less randomly. From the simulation point of view, walkers are not very different from the graphs succinctly described above. In this case, the identifier i is just a label that allows to access walker’s data, among which there are the coordinates of the walker (that may be continuous), its status, and so on.
In order to let walkers interact, one is interested in finding efficiently all walkers that are nearer than a given distance from the one under investigation. This is the same problem one is faced with when developing codes for molecular dynamics simulations (Rapaport 2004): scanning all walkers in order to compute their mutual distance is a procedure that grows quadratically with the number of walkers. One “trick: is that of dividing the space in cells, i.e., to define an associated lattice. Each cell contains a list of all walkers that are located inside it. In this way, one can directly access all walkers that are in the same or neighboring cells of the one under investigation.
Moreover, one can exploit the presence of the lattice to implement on it a “cellular automaton: that may interact with walkers, in order to simulate the evolution of “fields”. Just for example, the simulation of a herd of herbivores that move according to the exhaustion of the grass and the parallel growing of the vegetation can be modeled associating the “grass” to a cellular automaton and the herbivores to walkers. This simulation scheme is quite flexible, allowing to implement random and deterministic displacements of moving objects or agents , continuous or discrete evolution of “cellular objects,” and their interactions. Many simulation tools and games are based on this scheme (Repast – recursive porus agent simulation toolkit 2008; Wilensky 1999), and they generally allow the contemporary visualization of a graphical representation of the state of the system. They are valuable didactic tool and may be used to “experiment” with these artificial worlds. As often the case, the flexibility is paid in terms of efficiency and speed of simulation.
Future Directions
Complex systems, like, for instance, human societies, cannot be simulated starting from physical laws. This is sometimes considered a weakness of the whole idea of studying such systems from a quantitative point of view. We have tried to show that actually even the “hardest” discipline, physics, always deals with models that have finally to be simulated on computers, making various assumptions and approximations. Theoretical physics is accustomed for a long time to extreme simplifications of models, hoping to enucleate the fundamental ingredients of a complex behavior. This approach has proved to be quite rewarding for our understanding of nature.
In recent years, physics have been seen studying many fields not traditionally associated to physics: molecular biology, ecology, evolution theory, neurosciences, psychology, sociology, linguistics, and so on. Actually, the word “physics” may refer either to the classical subjects of study (mainly atomic and subatomic phenomena, structure of matter, cosmological, and astronomical topics) or to the “spirit” of the investigation that may apply to almost any discipline. This spirit is essentially that of building simplified quantitative models, composed by many elements, and study them with theoretical instruments (most of the times, applying some form of meanfield treatment) and with computer simulations.
This approach has been fruitful in chemistry and molecular biology, and nowadays many physical journals have sections devoted to multidisciplinary studies. The interesting thing is that not only physicists have brought some mathematics into fields that are traditionally more qualitative (which often corresponds to linear modeling, plus noise), but have also discovered many interesting questions to be investigated and new models to be studied. One example is given by the current investigations about the structure of social networks that were “discovered” by physicists in the nontraditional field of social studies.
Another contribution of physicists to this “new way” of performing investigations is the use of networked computers. For a long time, physicists have used computers in performing computations, storing data and diffusing information using the Internet. Actually, the concept of what is now the World Wide Web was born at CERN, as a method for sharing information among laboratories (Cailliau 1995). The highenergy physics experiments require a lot of simulations and data processing, and physicists (among others) developed protocols to distribute this load on a grid of networked computers. Nowadays, a European project aims to “open” grid computing to other sciences (European Grid Infrastructure).

Theory of evolution, especially for what concerns evolutionary medicine

Social epidemiology, coevolution of diseases and human populations, interplay between sociology and epidemics

Molecular biology and drug design, again driven by medical applications

Psychology and neural sciences: it is expected that the “black box” of traditional psychology and psychiatry will be replaced by explicit models based on brain studies

Industrial and material design

Earth sciences, especially meteorology, volcanology, and seismology

Archaeology: simulation of ancient societies, reconstruction of historical and prehistorical climates
Nowadays, the term cellular automata has enlarged its meaning, including any system whose elements do not move (in opposition to agentbased modeling). Therefore, we now have cellular automata on nonregular lattices, nonhomogeneous, with probabilistic dynamics (see section “Probabilistic Cellular Automata”), etc. (El Yacouby et al. 2006). They are therefore considered more as a “philosophy” of modeling rather than a single tool. In some sense, cellular automata (and agentbased) modeling is opposed to the spirit of describing a phenomena using differential equations (or partial differential equations). One of the reasons is that the language of automata and agents is simpler and requires less training than that of differential equations. Another reason is that at the very end, any reasonable problem has to be investigated using computers, and while the implementation using discrete elements is straightforward (even if careful planning may speed up dramatically the simulation), the computation of partially differential equations is an art in itself.
However, the final success of this approach is related to the availability of highquality experimental data that allow to discriminate among the almost infinite number of models that can be built.
Bibliography
Primary Literature
 Albert R, Barabasi AL (2002) Statistical mechanics of complex networks. Rev Mod Phys 74:47–97ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Bagnoli F, Rechtman R, Ruffo S (1991) Some facts of life. Physica A 171:249–264ADSMathSciNetCrossRefGoogle Scholar
 Bagnoli F (2000) Cellular automata. In: Bagnoli F, Ruffo S (eds) Dynamical modeling in biotechnologies. World Scientific, Singapore, p 1CrossRefGoogle Scholar
 Bagnoli F, Cecconi F (2001) Synchronization of nonchaotic dynamical systems. Phys Lett A 282(1–2):9–17ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Bagnoli F, Rechtman R (1999) Synchronization and maximum Lyapunov exponents of cellular automata. Phys Rev E 59(2):R1307–R1310ADSCrossRefzbMATHGoogle Scholar
 Bagnoli F, Rechtman R (2009) Thermodynamic entropy and chaos in a discrete hydrodynamical system Phys Rev E 79:041115Google Scholar
 Bagnoli F, Rechtman R, Ruffo S (1992) Damage spreading and Lyapunov exponents in cellular automata. Phys Lett A 172:34ADSCrossRefzbMATHGoogle Scholar
 Bagnoli F, Boccara N, Rechtman R (2001) Nature of phase transitions in a probabilistic cellular automaton with two absorbing states. Phys Rev E 63(4):046116ADSCrossRefGoogle Scholar
 Bak P, Tang C, Weisenfeld K (1987) Selforganizing criticality: an explanation of 1/f noise. Phys Rev A 38:364–374ADSCrossRefGoogle Scholar
 Barkema GT, MacFarland T (1994) Parallel simulation of the ising model. Phys Rev E 50(2):1623–1628ADSCrossRefGoogle Scholar
 Berlekamp E, Conway J, Guy R (1982) What is life? Games in particular, vol 2. Academic, London. Chap. 25Google Scholar
 Binney J, Dowrick N, Fisher A, Newman MEJ (1993) The theory of critical phenomena. Oxford Science/Clarendon Press, OxfordzbMATHGoogle Scholar
 Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang DU (2006) Complex networks: structure and dynamics. Phys Rep 424(4–5):175–308ADSMathSciNetCrossRefGoogle Scholar
 Broadbent S, Hammersley J (1957) Percolation processes I. Crystals and mazes. Proc Camb Philos Soc 53:629–641ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Cailliau R (1995) A short history of the web. http://www.netvalley.com/archives/mirrors/robert_cailliau_speech.htm. Accessed 10 Apr 2017
 Car R, Parrinello M (1985) Unified approach for molecular dynamics and densityfunctional theory. Phys Rev Lett 55(22):2471–2474ADSCrossRefGoogle Scholar
 Cecconi F, Livi R, Politi A (1998) Fuzzy transition region in a onedimensional coupledstablemap lattice. Phys Rev E 57(3):2703–2712ADSCrossRefGoogle Scholar
 Chopard B, Luthi P, Masselot A, Dupuis A (2002) Cellular automata and lattice Boltzmann techniques: an approach to model and simulate complex systems. Adv Complex Syst 5(2):103–246MathSciNetCrossRefzbMATHGoogle Scholar
 Crutchfield J, Kaneko K (1988) Are attractors relevant to turbulence? Phys Rev Lett 60(26):2715–2718ADSMathSciNetCrossRefGoogle Scholar
 Daxois T, Peyrard M, Ruffo S (2005) The FermiPastaUlam ‘numerical experiment’: history and pedagogical perspectives. Eur J Phys 26:S3–S11MathSciNetCrossRefGoogle Scholar
 Domany E, Kinzel W (1984) Equivalence of cellular automata to Ising models and directed percolation. Phys Rev Lett 53(4):311–314ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Dotsenko V (1994) An introduction to the theory of spin glasses and neural networks. World Scientific, SingaporezbMATHGoogle Scholar
 El Yacouby S, Chopard B, Bandini S (eds) (2006) Cellular automata, Lecture notes in computer science, vol 4173. Springer, BerlinGoogle Scholar
 European Grid Infrastructure. https://www.egi.eu/. Accessed 10 Apr 2017
 Fermi E, Pasta J, Ulam S (1955) Los alamos report la1940. In: Segré E (ed) Collected papers of Enrico Fermi. University of Chicago Press, ChicagoGoogle Scholar
 Frisch U, Hasslacher B, Pomeau Y (1986) Latticegas automata for the navierstokes equation. Phys Rev Lett 56(14):1505–1508ADSCrossRefGoogle Scholar
 Gardiner CW (1994) Handbook of stochastic methods for physics, chemistry, and the natural sciences, Springer series in synergetics, vol 13. Springer, BerlinzbMATHGoogle Scholar
 Georges A, le Doussal P (1989) From equilibrium spin models to probabilistic cellular automata. J Stat Phys 54(3–4):1011–1064ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Hardy J, Pomeau Y, de Pazzis O (1973) Time evolution of a twodimensional classical lattice system. Phys Rev Lett 31(5):276–279ADSCrossRefGoogle Scholar
 Harlow H, Metropolis N (1983) Computing & computers – weapons simulation leads to the computer era. Los Alamos Sci 4(7):132Google Scholar
 Haw M (2005) Einstein’s random walk. Phys World 18:19–22CrossRefGoogle Scholar
 Hinrichsen H (1997) Stochastic lattice models with several absorbing states. Phys Rev E 55(1):219–226ADSCrossRefGoogle Scholar
 Jaynes E (1957) Information theory and statistical mechanics. Phys Rev 106(4):620–630ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Kaneko K (1985) Spatiotemporal intermittency in coupled map lattices. Progr Theor Phys 74(5):1033–1044ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Kawasaki K (1972) Kinetics of Ising model. In: Domb CM, Green MS (eds) Phase transitions and critical phenomena, vol 2. Academic, New York, p 443Google Scholar
 Kirkpatrick S, Gelatt CG Jr, Vecchi MP (1983) Optimization by simulated annealing. Science 220:671–680ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Lawniczak A, Dab D, Kapral R, Boon JP (1991) Reactive lattice gas automata. Phys D 47(1–2):132–158MathSciNetCrossRefzbMATHGoogle Scholar
 Marinari E, Parisi G (1992) Simulated tempering: a new Monte Carlo scheme. Europhys Lett 19:451–458ADSCrossRefGoogle Scholar
 May R (1976) Simple mathematical models with very complicated dynamics. Nature 261:459–467ADSCrossRefGoogle Scholar
 Metropolis N, Hewlett J, Rota GC (eds) (1980) A history of computing in the twentieth century. Academic, New YorkzbMATHGoogle Scholar
 Mezard M, Parisi G, Virasoro MA (1987) Spin glass theory and beyond. World scientific lecture notes in physics, vol 9. World Scientific, SingaporezbMATHGoogle Scholar
 Newman ME (2005) Power laws, Pareto distributions and Zipf’s law. Contemp Phys 46:323–351ADSCrossRefGoogle Scholar
 Niss M (2005) History of the LenzIsing model 1920–1950: from ferromagnetic to cooperative phenomena. Arch Hist Exact Sci 59(3):267–318MathSciNetCrossRefzbMATHGoogle Scholar
 Nordfalk J, Alstrøm P (1996) Phase transitions near the “game of life”. Phys Rev E 54(2):R1025–R1028ADSCrossRefGoogle Scholar
 Onsager L (1944) Crystal statistics. I. A twodimensional model with an orderdisorder transition. Phys Rev 65:117–149ADSMathSciNetCrossRefzbMATHGoogle Scholar
 Oestreicher C (2007) A history of chaos theory. Dialogues Clin Neurosci 9(3):279–289. Available online https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3202497/
 Politi A, Livi R, Oppo GL, Kapral R (1993) Unpredictable behaviour of stable systems. Europhys Lett 22(8):571–576ADSCrossRefGoogle Scholar
 Rabiner L (1989) A tutorial on hidden Markov models and selected applications in speech recognition. Proc IEEE 77(2):257–286CrossRefGoogle Scholar
 Rapaport DC (2004) The art of molecular dynamics simulation. Cambridge University Press, CambridgeCrossRefzbMATHGoogle Scholar
 Repast – recursive porus agent simulation toolkit (2008) http://repast.sourceforge.net/. Accessed 10 Apr 2017
 Rothman DH, Zaleski S (2004) Latticegas cellular automata. Monographs and texts in statistical physics. Collection AleaSaclay, ParisGoogle Scholar
 Sornette D (2006) Critical phenomena in natural sciences, Springer series in synergetics. Springer, BerlinzbMATHGoogle Scholar
 Stauffer D, Aharony A (1994) Introduction to percolation theory. Taylor Francis, LondonzbMATHGoogle Scholar
 Succi S (2001) The lattice Boltzmann equation for fluid dynamics and beyond. Numerical mathematics and scientific computation. Oxford University Press, OxfordzbMATHGoogle Scholar
 Swendsen R, Wang JS (1987) Nonuniversal critical dynamics in Monte Carlo simulations. Phys Rev Lett 58(2):86–88ADSCrossRefGoogle Scholar
 van Kampen NG (1992) Stochastic processes in physics and chemistry. NorthHolland, AmsterdamzbMATHGoogle Scholar
 von Neumann J, Burks AW (1966) Theory of selfreproducing automata. University of Illinois Press, Urbana/LondonGoogle Scholar
 Von Neumann universal constructor (2008) http://en.wikipedia.org/wiki/Von_Neumann_Universal_Constructor. Accessed 10 Apr 2017
 Watts D, Strogatz SH (1998) Collective dynamics of ‘smallworld’ networks. Nature 393:440–441ADSCrossRefGoogle Scholar
 Wilensky U (1999) Netlogo. Center for connected learning and computerbased modeling, Northwestern University, Evanston. http://ccl.northwestern.edu/netlogo/. Accessed 10 Apr 2017
 WolfGladrow D (2004) Latticegas cellular automata and lattice Boltzmann models: an introduction, Lecture notes in mathematics, vol 1725. Springer, BerlinzbMATHGoogle Scholar
 Wolfram S (1983) Statistical mechanics of cellular automata. Rev Mod Phys 55:601–644ADSMathSciNetCrossRefzbMATHGoogle Scholar
Books and Reviews
 Boccara N (2004) Modeling complex systems. In: Graduate texts in contemporary physics. Springer, BerlinGoogle Scholar
 Bungartz HJ, Mundani RP, Frank AC (2005) Bubbles, jaws, moose tests, and more: the wonderful world of numerical simulation, Springer VideoMATH. Springer, Berlin. (DVD)zbMATHGoogle Scholar
 Chopard B, Droz M (2005) Cellular automata modeling of physical systems. In: Collection AleaSaclay: monographs and texts in statistical physics. Cambridge University Press, CambridgeGoogle Scholar
 Deisboeck S, Kresh JY (2006) Complex systems science in biomedicine. In: Deisboeck S, Kresh JY (eds) Topics in biomedical engineering. Springer, New YorkGoogle Scholar
 Gould H, Tobochnik J, Christian W (2007) An introduction to computer simulation methods: applications to physical systems. AddisonWesley, New YorkGoogle Scholar
 Landau RH (2005) A first course in scientific computing: symbolic, graphic, and numeric modeling using maple, java, Mathematica, and Fortran90. Princeton University Press, PrincetonzbMATHGoogle Scholar
 Open Source Physics. http://www.opensourcephysics.org/. Accessed 10 Apr 2017
 Resnick M (1994) Turtles, termites, and traffic jams. Explorations in massively parallel microworlds. In: Complex adaptive systems. MIT Press, CambridgeGoogle Scholar
 Shalizi C Cosma’s home page. http://www.cscs.umich.edu/~crshalizi/. Accessed 10 Apr 2017