Posts Tagged ‘free energy’

Computing Free Energies

October 3, 2007

In my last post, I discussed phase transitions, and how computing the free energy for a model would let you work out the phase diagram. Today, I want to discuss in more detail some methods for computing free energies.

The most popular tool physicists use for computing free energies is “mean-field theory.” There seems to be at least one “mean-field theory” for every model in physics. When I was a graduate student, I became very unhappy with the derivations for mean-field theory, not because there were not any, but because there were too many! Every different book or paper had a different derivation, but I didn’t particularly like any of them, because none of them told you how to correct mean-field theory. That seemed strange because mean-field theory is known to only give approximate answers. It seemed to me that a proper derivation of mean-field theory would let you systematically correct the errors.

One paper really made me think hard about the problem; the famous 1977 “TAP” spin glass paper by Thouless, Anderson, and Palmer. They presented a mean-field free energy for the Sherrington-Kirkpatrick (SK) model of spin glasses by “fait accompli,” which added a weird “Onsager reaction term” to the ordinary free energy. This shocked me; maybe they were smart enough to write down free energies by fait accompli, but I needed some reliable mechanical method.

Since the Onsager reaction term had an extra power of 1/T compared to the ordinary energy term in the mean field theory, and the ordinary energy term had an extra power of 1/T compared to the entropy term, it looked to me like perhaps the TAP free energy could be derived from a high-temperature expansion. It would have to be a strange high-temperature expansion though, because it would need to be valid in the low-temperature phase!

Together with Antoine Georges, I worked out that the “high-temperature” expansion (it might better be thought of as a “weak interaction expansion”) could in fact be valid in a low-temperature phase, if one computed the free energy at fixed non-zero magnetization. This turned out to be the key idea; once we had it, it was just a matter of introducing Lagrange multipliers and doing some work to compute the details.

expansion001.jpg

It turned out that ordinary mean-field theory is just the first couple terms in a Taylor expansion. Computing more terms lets you systematically correct mean field theory, and thus compute the critical temperature of the Ising model, or any other quantities of interest, to better and better precision. The picture above is a figure from the paper, representing the expansion in a diagrammatic way.

We found out, after doing our computations but before submitting the paper, that in 1982 Plefka had already derived the TAP free energy for the SK model from that Taylor expansion, but for whatever reason, he had not gone beyond the Onsager correction term or noted that this was a technique that was much more general than the SK model for spin glasses, so nobody else had followed up using this approach.

If you want to learn more about this method for computing free energies, please read my paper (with Antoine Georges) “How to Expand Around Mean-Field Theory Using High Temperature Expansions,” or my paper “An Idiosyncratic Journey Beyond Mean Field Theory.”

This approach has some advantages and disadvantages compared with the belief propagation approach (and related Bethe free energy) which is much more popular in the electrical engineering and computer science communities. One advantage is that the free energy in the high-temperature expansion approach is just a function of simple one-node “beliefs” (the magnetizations), so it is computationally simpler to deal with than the Bethe free energy and belief propagation. Another advantage is that you can make systematic corrections; belief propagation can also be corrected with generalized belief propagation, but the procedure is less automatic. Disadvantages include the fact that the free energy is only exact for tree-like graphs if you add up an infinite number of terms, and the theory has not yet been formulated in an nice way for “hard” (infinite energy) constraints.

If you’re interested in quantum systems like e.g. the Hubbard model, the expansion approach has the advantage that it can also be applied to them; see my paper with Georges, or the lectures by Georges on his related “Dynamical Mean Field Theory,” or this recent paper by Plefka, who has returned to the subject more than 20 years after his original paper.

Also, if you’re interested in learning more about spin glasses or other disordered systems, or about other variational derivations for mean-field theory, please see this post.

Advertisement

Tags:, , , , ,
Posted in Inference, Physics, Science, Statistical Physics | 2 Comments »

Phase Transitions and Free Energies

October 2, 2007

Much of condensed matter and statistical physics is concerned with the explanation of phase transitions between different forms of matter. A familiar example is water, which has a transition from the solid phase of ice to the liquid phase, and then from the liquid phase to the gaseous phase of steam, as the temperature is increased.

In fact, many different materials have all sorts of exotic phases, as you vary the temperature, pressure, or composition of the material. Constructing “phase diagrams” which show what the different phases of a material should be as a function of the varying parameters is one of the main preoccupations of physicists.

How can these phase diagrams be constructed? Condensed matter physicists tend to follow the following algorithm. First, from arguments about the microscopic physics, construct a simple model of the local interactions of the molecules making up the material. Secondly, choose some method to approximately compute the “free energy” for that simple model. Finally, find the minima of the approximate free energy as a function of the adjustable parameters like the temperature. The phase diagram can be constructed by determining which phase has the lower free energy at each point in parameter space.

If the results disagree with experiment, your model is too simple or your approximation for the free energy is not good enough, so you need to improve one or the other or both; otherwise write up your paper and submit it for publication.

To illustrate, let’s consider magnetism. Although it is less familiar than the phase transition undergone by water, magnets also have a phase transition from a magnetized “frozen” phase at low temperature to an unmagnetized “paramagnetic” phase at high temperature.

The simplest model of magnetism is the Ising model. I’ve discussed this model before; to remind you, I’ll repeat the definition of the ferromagnetic Ising model: “In this model, there are spins at each node of a lattice, that can point ‘up’ or ‘down.’ Spins like to have their neighbors point in the same direction. To compute the energy of a configuration of spins, we look at all pairs of neighboring spins, and add an energy of -1 if the two spins point in the same direction, and an energy of +1 if the two spins point in opposite directions. Boltzmann’s law tells us that each configuration should have a probability proportional to the exp(-Energy[configuration] / T), where T is the temperature.”

Of course, as defined the Ising model is a mathematical object that can and has been studied mathematically independent of any relationship to physical magnets. Alternatively, the Ising model can be simulated on a computer.

Simulations (see this applet to experiment for yourself) show that at low temperatures (and if the dimensionality of the lattice is at least 2), the lattice of spins will over time tend to align so that they point together up more than down, or they all point down more than up. The natural symmetry between up and down is “broken.” At low temperatures one will find “domains” of spins pointing in the “wrong” direction, but these domains only last temporarily.

At high temperatures, on the other hand, each spin will typically fluctuate between pointing up and down, although again domains of like-pointing spins will form. The typical time for a spin to switch from pointing up to pointing down will increase as the temperature decreases, until it diverges towards infinity (as the size of the lattice approaches infinity) as one approaches the critical temperature from above.

Intuitively, the reason for this behavior is that at low temperature, the configurations where all the spins point in the same direction have a much lower energy, and thus a much higher probability, than other configurations. At high temperatures, all the configurations start having similar probabilities, and there are many more configurations that have equal numbers of up and down spins compared to the number of aligned configurations, so one typically sees the more numerous configurations.

This balance between energetic considerations (which make the spins align) and entropic considerations (which make the spins favor the more numerous unaligned configurations) is captured by the “free energy” F, which is given by the equation F=U-TS, where U is the average energy, S is the entropy, and T is the temperature. At low temperatures, the energy dominates the free energy, while at high temperatures, the entropy dominates.

All this intuition may be helpful, but like I said, the Ising model is a mathematical object, and we should be able to find approximation methods which let us precisely calculate the critical temperature. It would also be nice to be able to precisely calculate other interesting quantities, like the magnetization as a function of the temperature, or the susceptibility (which is the derivative of the magnetization with respect to an applied field) as a function of the temperature, or the specific heat (which is the derivative of the average energy with respect to the temperature) as a function of the temperature. All these quantities can be measured for real magnets, so if we compute them mathematically, we can judge how well the Ising model explains the magnets.

This post is getting a bit long, so I’ll wait until my next post to discuss in more detail some useful methods I have worked on for systematically and precisely computing free energies, and the other related quantities which can be derived from free energies.

 

Tags:, ,
Posted in Physics, Science, Statistical Physics | 1 Comment »

Generalized Belief Propagation

August 5, 2007

In 2002, I gave a lecture at the Mathematical Sciences Research Institute on the work I did, together with Bill Freeman and Yair Weiss on Generalized Belief Propagation, and the correspondence between free energy approximations and message passing algorithms. The lecture is available as a streaming video, together with a pdf for the slides, here.

It’s worth mentioning that there are many other interesting research lectures available in MSRI’s video archive, and that the more recent ones are of higher production quality.

Here is our most recent and comprehensive paper on this subject, published in the July 2005 issue of IEEE Transactions on Information Theory, which gives many additional details compared to the lecture: MERL TR2004-040.

If that paper is too difficult, you should probably start with this earlier paper, which was more tutorial in nature: MERL TR2001-22.

If you’re looking for generalized belief propagation software, your best bet is this package written by Yair’s student Talya Meltzer.

P.S.: I realized I haven’t told those of you who don’t know anything about it what generalized belief propagation is. Well, one answer is to that is look at the above material! But here’s a little background text that I’ve copied from my research statement to explain why you might be interested:

Most of my current research involves the application of statistical methods to “inference” problems. Some important fields which are dominated by the issue of inference are computer vision, speech recognition, natural language processing, error-control coding and digital communications. Essentially, any time you are receiving a noisy signal, and need to infer what is really out there, you are dealing with an inference problem.

A productive way to deal with an inference problem is to formalize it as a problem of computing probabilities in a “graphical model.” Graphical models, which are referred to in various guises as “Markov random fields,” “Bayesian networks,” or “factor graphs,” provide a statistical framework to encapsulate our knowledge of a system and to infer from incomplete information.

Physicists who use the techniques of statistical mechanics to study the behavior of disordered magnetic spin systems are actually studying a mathematically equivalent problem to the inference problem studied by computer scientists or electrical engineers, but with different terminology, goals, and perspectives. My own research has focused on the surprising relationships between methods that are used in these communities, and on powerful new techniques and algorithms, such as Generalized Belief Propagation, that can be understood using those relationships.

I’ll tell you more in future posts; I promise.

Tags:, , , ,
Posted in AI, Algorithms, Computer Science, Education, Inference, Probability, Science, Statistical Physics, Video | 3 Comments »