Physics 212: Statistical mechanics II, Fall 2006 Lecture XII

The main result of the last lecture was a calculation of the averaged magnetization in mean-eld theory in Fourier space when the spin at the origin is xed up, m(k) C . 1 J(1 R2 k 2 ) (1)

Its Fourier transform is m(r), which we argued above to be equal to G(r), the correlation function between two spins separated by r. First rewrite it as G(k) C CR2 CR2 = 2 = 2 KR2 k 2 + 1 K k + tR2 k + 2 (2)
1 2

with = Rt1/2 , and t = (T Tc )/T . The result = Rt1/2 denes the critical exponent = for the growth of the correlation length near criticality.

The Fourier transform described above, which in general dimensionality can be written in terms of Bessel functions, gives that the correlation function on long length scales falls o as (now forgetting about units; additional prefactors of C and the lattice spacing R must be added to get the units right) G(r) er/ /r(d1)/2 (3) starting from G(r) CR2 dd k exp(ikx) = CR2 2d k 2 + 2 dd k exp(ikx/) . k2 + 1 (4)

Actually G(r) has two dierent asymptotic limits in MFT: for large r G(r)
er/ r(d1)/2 1 rd2

if r if r

(5)

(Huang describes this Fourier transform in one problem but has a wrong power in the rst limit.) Note that the rst limit in d = 1 matches our result for the correlation function of the onedimensional Ising chain, in the limit of large separation. With this information about correlations, we are able to justify our earlier claim about the validity of mean-eld theory above four dimensions. The approach will be to estimate the neglected part of the Hamiltonian and compare it to the part we found in the mean-eld calculation. This procedure is carried out for lattices in the book of Cardy; instead of working with the lattice mean eld theory, we now will introduce a phenomenological approach to phase transitions in the continuum known as Landau theory. The continuum approach to phase transitions is a rare example where phenomenology can give exactly correct answers even for certain numerical quantities, such as critical exponents. For now, think of this phenomenology as motivated by a desire to describe continuum systems such as the liquid-gas transition in water, which was claimed in Lecture I to share some properties, including critical exponents, with the (lattice) Ising model. 1

We start o by following Landau and conjecturing that the free energy density near a phase transition can be expanded in powers of the order parameter describing the transition. For the Ising transition, w choose the magnetization to be the order parameter: the requirement is that the symmetry that is broken at the transition should transform the order parameter from one ordered state to another. In terms of the dimensionless magnetization, we can write this expansion as F (m) = m H 1 + r0 m2 + s0 m3 + u0 m4 + . . . . kT 2 (6)

Remark: We will later on want to consider spatially varying congurations and also uctuations: let me quickly introduce a dierent way of looking at the Landau free energy. We assume that near the critical point, the coarse-grained spin m(r) varies only slowly for the lowest-energy congurations. Hence we guess that the Gibbs free energy in zero magnetic eld takes the phenomenological form, 1 1 (m(r), H(r)) = | m(r)|2 + r0 m(r)2 + u0 m(r)4 + . . . . 2 2 (7)

Here the coecients are unknown functions of T , and m is rescaled so that the rst term has coecient 1/2 (see next paragraph). The purpose of this is that the partition function can be written in terms of an integral over functions m(r): Z(T, H) C (Dm(r))e
dd r (m,H)

(8)

This is similar to the improved derivation of MFT, where we rst carried out the spin sum to obtain a free energy as a function of the uniform variational parameter m; the Landau free energy (1937) is essentially a phenomenological guess for a general form of the free energy.
1 Choosing the coecient of the gradient term to be 2 is just a convention that eectively denes the scale of m: if we started with a dierent coecent, then we could rescale m to set it equal to 1 . 2 Note that here m is not even bounded between 1 and 1, although we will assume u0 positive so that congurations with m large in magnitude are quite unlikely. This is known as a soft-spin free energy since there is no longer a hard constraint on the values of m. The Landau theory of phase transitions gives a powerful way of deriving such results as the Ginzburg criterion found below. It is currently believed that the phenomenological form, with a suitably dened functional integral, has exactly the same critical exponents and other universal properties as the original lattice Ising model. This is perhaps not too surprising if one remembers that the liquid-gas transition in the continuum has the same universal properties as well. End remark

Looking at the terms in this free energy, we can interpret them quite easily, at least with zero magnetic eld and symmetry between m, so s0 = 0. Changing the sign of r0 while keeping u0 positive will change between having one or two minima of the energy in the uniform case. We will show now that this is analogous, under a simple assumption about the temperature dependence of r0 to our solution of heuristic mean-eld theory. (Landau mean-eld theory is obtained by approximating the functional integral in the remark above by its saddle point, ignoring uctuations; then we do not even need to worry about the functional integral.) The mean value m in the ordered state is then r0 m= . (9) 4u0 2

Similarly, in this saddle-point approximation, it is even easier than before to calculate mean-eld critical exponents like = 1 and = 3, where the parameter driving the transition is r0 = a0 t for some constant a0 and t the reduced temperature t= T Tc . Tc (10)

There are several assumptions that go into the Landau (continuum) version of mean-eld theory: there is an order parameter that transforms under the broken symmetry; there is a Taylor series expansion of the free energy in terms of the order parameter, with coecients that are analytic functions of temperature; and the saddle point of the integral is a good approximation to its behavior. The Ginzburg criterion for validity of mean-eld theory is most easily derived on the ordered side of the transition. We want to compare the neglected uctuations near the critical point with the mean m. A detailed calculation shows that the correlation length exponent just on the disordered side, which we derived earlier this lecture, is exactly the same as the correlation length exponent on the ordered side (recall that the correlation function on the ordered side is dened as the correlation of the uctuation part). We want to estimate the relative magnitude of the double-uctuation term, m(r)m(r ) G(r r ) = . (11) m2 m2 At a separation r r of order , this becomes (using the Coulomb law derived previously) 2d 2d 2d = = . (r0 /4u0 ) (r0 /4u0 ) (a0 t/4u0 ) The mean-eld prediction is that (since = 1/2) the correlation length grows as t1/2 , (13) (12)

so the uctuations scale with t as t(d2)/2 /t = td4/2 . This blows up near the transition for d < 4, but goes to zero as d > 4, suggesting that close to the transition mean eld theory becomes a worse approximation in low dimensions, but a better approximation in high dimensions. The model obtained from neglecting the cubic, quartic, and higher terms in the Landau free energy is known as the Gaussian model. We will have more to say about this model later. One main result of the previous two lecures was a certain form for the correlation functions near a critical point. In fact such scaling forms were conjectured historically by Widom, Kadano, and others before the modern RG approach to phase transitions. The point of this lecture is to show that a great deal of information can be derived from simple scaling assumptions. An example of such a scaling assumption is that correlations near the transition are controlled by the diverging length scale , as we showed for mean-eld theory last time. This part will have fewer equations than the previous sections on phase transitions: its goal is to put the specic calculations done before in the broader context of the scaling theory of phase transitions. One motivation for the scaling laws is that they seem to hold exactly for almost all of the transitions encountered in physical systems, even though many of the laws have still not been proven rigorously. (To be precise, most of the laws have been rigorously proven as inequalities, 3

although they are believed to hold as equalities.) Assuming the scaling laws hold, only two of the critical exponents at a classical phase transition need to be known in order to deduce all the others. Let us make a list of critical exponents and in the process review the critical exponents discussed so far for mean-eld theory. First we give a list of six critical exponents and their denitions for the Ising model, and then for a liquid-gas transition near the critical point. For brevity we dene the reduced temperature T Tc t (14) Tc and reduced magnetic eld H h . (15) kTc 1. The critical exponent describes the singularity in the specic heat at a second-order transition. Recall that the specic heat is proportional to the second derivative of the free energy with respect to temperature: F S S= , NC = , (16) T T and that at a generic second-order point the rst derivative is continuous and the second derivative is singular. Here N is the number of spins in the system in order to make the specic heat an intensive quantity. We might expect dierent singularities on the ordered and disordered sides of the transition, A|t| if t > 0 C . (17) A |t| if t < 0 It turns out that = (this is predicted by the RG, and conrmed by exact results) so we will just call this exponent . The amplitudes A and A are nonuniversal (but it happens that their ratio A/A is universal; such universal amplitude ratios are an example of universal critical quantities other than critical exponents!) We didnt calculate the mean-eld theory specic heat because it is actually somewhat pathological; it is discontinuous but not otherwise singular, so = 0. 2. The critical exponent describes the spontaneous magnetization (that is, H = 0) on the ordered side of the transition: M (t) , t < 0. (18) For mean-eld theory we calculated the value = 1 . 2 3. The critical exponent was not discussed so far: it gives the divergence of the susceptibility near the transition. The susceptibility is related to the second derivative of the free energy with respect to magnetic eld (the rst derivative is the magnetization), so should in general have a singularity for the same reason that the specic heat does. This singularity is described as = M H |H=0 |t| . (19)

As for the specic heat, we might have expected dierent values for on the ordered and disordered sides of the transition, but theory and exact results suggest + = so we will just write . Meaneld theory predicts = 1. 4. The critical exponent describes the magnetization in a eld H = 0 at the critical temperature (T = Tc ): M h1/ . (20) 4

For mean-eld theory we found = 3. 5. The divergence of the correlation length near the transition is given by the critical exponent . The correlation function goes as G(r) = (0 m)(r m) er/ (polynomial factors), and |t| . (22) Once again, we might worry that would take dierent values on the dierent sides of the transition, but that does not appear to be the case. Note that and the following exponent both are dened in terms of the correlation functions, while the rst four exponents were dened just in terms of 1 the free energy and other translation-invariant quantities. For mean-eld theory we found = 2 . 6. At criticality (H = 0, T = Tc ), the correlation falls o as a power-law rather than as an exponential, because the correlation length has gone to innity. The coecient of this power-law is dened using the critical exponent : G(r) For mean-eld theory = 0. (7. For completeness we mention a critical exponent that comes up in anisotropic critical phenomena, which we will probably not deal with in this course. You can imagine a problem where, for example, the diverging correlation length scale in the x direction might scale dierently from the diverging correlation length scale in the y direction; then one denes the anisotropy critical exponent z through x (y )z . (24) This is chiey important for dynamical problems where the time scale near criticality is a power of the length scale, and time can be thought of as an extra dimension. For isotropic critical phenomena z = 1.) Now we quickly give the denitions of the rst four exponents in the uid case, which are natural analogies to those in the Ising magnet. The specic heat exponent in the spin case corresponds to the specic heat at constant volume: CV |t| at = c . The exponent gives the density dierence near the critical point, which also determines the shape of the coexistence curve in a system of xed volume: L G (t) . The critical exponent is related to the isothermal compressibility T |t| . Finally, the critical exponent gives the density dierence in response to a change in pressure at the critical temperature: L G = (p pc )1/ . (25) 1 rd2+ at criticality. (23) (21)

The six main critical exponents turn out to be governed by just two independent quantities. This will turn out to be a very simple prediction of the renormalization-group theory of the Ising model. Before this modern understanding, however, it is important to note that by intuitive/physical arguments, people already understood scaling relations such as = 2 d, 5 (26)

 = (2 ), + 2 + = 2, + (1 + ) = 2.

(27) (28) (29)

Now we introduce the scaling forms for free energy and correlation functions in order to understand where the above relations come from. Let us assume that the free energy per site near a phase transition can be separated into a smooth part fn and a part fs whose derivatives contain the interesting singularities: f (t, h, . . .) = fn (t, h, . . .) + fs (t, h, . . .). (30) Now focus on the singular part fs and assume that the only variables controlling it are t and h. At h = 0, we know that the singular part should give rise in its second derivative to the specic heat singularity, or fs (t, 0) |t|2 . (31) It will be convenient to introduce two scales t0 and h0 that stay nite at the transition in order to make quantities dimensionless: then this becomes fs (t, 0) = |t/t0 |2 . (32)

Now the main assumption of the scaling theory is that tuning away from the critical point with some magnetic eld h is related to tuning away from the critical point with temperature, so that rather than having fs (t, h) be an arbitrary function of t and h, its dependence is restricted to fs (t, h) = |t/t0 |2 (|h/h0 ||t/t0 | ) (33)

for some single-variable function (known as the scaling function) and some exponent . By our denitions (0) = 1. Our current understanding is that dierent problems in the same universality class dier only in the scales h0 and t0 : the scaling function , not just the individual critical exponents that it predicts, is universal. We can relate to the other critical exponents as follows: the spontaneous magnetization at h = 0 is, assuming (0) to be a nite number, f (t)2 = 2 . h The susceptibility is given by the second derivative of fs with respect to h, or 2f (t)22 = + 2 2. h2 (35) (34)

Just from the above simple assumption about the form of fs , we can now derive one of the identities for critical exponents: combining 2 + in order to eliminate the unknown quantity , we nd 2 + = 2 + 2 + = 2, (36) one of the scaling relations mentioned above. A slightly more involved trick can be used to express in terms of and to obtain another exponent identity. We will return to these identities once the RG theory has been developed to explain the scaling form used above.

Lastly, consider a scaling form for the correlation function. A rst guess, which turns out not to work except for simple cases like mean-eld theory, is that near criticality the only important length scale is the correlation length , so that G(r) = r f (r/), (37)

where is some exponent and f some unknown scaling function. For somewhat subtle reasons, this formula is not suciently general: it essentially describes a case where all the uctuations are on the length scale , where what we want in general is a description of uctuations on all length scales between lattice spacing a and the correlation length . Hence consider the more general form G(r) = r f (r/, a/). (38)

The value = (d 2) is suggested by our mean-eld calculation. This prefactor will be justied in more detail later when we discuss dimensional analysis; for now focus on the function f . The assumption corresponding to the scaling assumption in the free energy is that f (r/, a/) (a/) g(r/) for a. (39)

To see why this is a powerful assumption, recall (or convince yourself) that the susceptibility is related to the correlation function through G(r)dd r. (40)

Then the above scaling form predicts = 2 or = (2 ), another identity for critical exponents. For simple cases with = 0, it is acceptable to use the simpler form of f with only one argument, but for most transitions (like the 3D Ising model) this does not give the correct picture of correlations. The next lectures will explain why = 0 is an example of an anomalous dimension, i.e., a critical exponent not calculable simply from dimensional analysis, and how this insight led to the modern theory of critical phenomena.