# Lecture notes Advanced Statistical Mechanics AP3021G

Jos Thijssen, Kavli Institute for Nanoscience Faculty of Applied Sciences Delft University of Technology

September 23, 2008

Preface
This is a set of lecture notes which is intended as a support for students in my course ‘advanced statistical mechanics’. This is a typical graduate course on the subject, including some non-equilibrium thermodynamics and statistical physics. The course has over the years been based on different books, but the informed reader will recognise the structure of the book by Pathria (Statistical Mechanics, 1992) in the ﬁrst part (equilibrium phenomena) and from several chapters of the book by Bellac, Mortessange and Batrouni (Equilibrium and non-equilibrium statistical thermodynamics, 2004). Another important contribution is provided by the lecture notes by Hubert Knops for his statistical mechanics courses at Nijmegen. My lecture notes are therefore by no means original, but they intend to combine the parts of all the sources mentioned into a coherent and clear story. However, this story does by no means qualify as a textbook, as it is too sketchy and superﬁcial for that purpose. It is merely intended as a support for students following my lecture course. I hope it helps. It should be noted that these notes do not fully cover the material of my course. I usually make a selection of about 80 % of the material in these notes, and ﬁll the remaining time with additional topics, e.g. the exact solution of the Ising model in 2D or the epsilon-expansion. I intend to include these topics into the course, together with a discussion of polymers and membranes.

ii

Contents
1 The statistical basis of Thermodynamics 1.1 The macroscopic and the microscopic states . . . . . . 1.2 Contact between statistics and thermodynamics . . . . 1.3 Further contact between statistics and thermodynamics 1.4 The ideal gas . . . . . . . . . . . . . . . . . . . . . . 1.5 The entropy of mixing and the Gibbs paradox . . . . . 1.6 The “correct” enumeration of the microstates . . . . . 2 Elements of ensemble theory 2.1 Phase space of a classical system . . . . . 2.2 Liouville’s theorem and its consequences 2.3 The microcanonical ensemble . . . . . . 2.4 Examples . . . . . . . . . . . . . . . . . 1 1 1 2 3 4 5 6 6 7 8 9 10 10 10 12 14 14 15 16 17 19 20 23 23 24 24 25 28 29 30 30 31 32 34

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

3 The canonical ensemble 3.1 Equilibrium between a system and a heat reservoir . . . . . . . . . . . . . . . . 3.2 A system in the canonical ensemble . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Physical signiﬁcance of the various statistical quantities in the canonical ensemble 3.4 Alternative expressions for the partition function . . . . . . . . . . . . . . . . . 3.5 Classical systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6 Energy ﬂuctuations in the canonical ensemble: correspondence with the microcanonical ensemble . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.7 Two theorems – the ‘equipartition’ and the ‘virial’ . . . . . . . . . . . . . . . . . 3.8 A system of classical harmonic oscillators . . . . . . . . . . . . . . . . . . . . . 3.9 The statistics of paramagnetism . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.10 Thermodynamics of magnetic systems: negative temperature . . . . . . . . . . . 4 The grand canonical ensemble 4.1 Equilibrium between a system and a particle-energy reservoir 4.2 Formal derivation of the grand canonical ensemble . . . . . 4.3 Physical signiﬁcance of the various statistical quantities . . . 4.4 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

5 Formulation of Quantum Statistics 5.1 Statistics of the various ensembles . . . . . . . . . . . . . . . . . . . . . 5.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.1 Electron in a magnetic ﬁeld . . . . . . . . . . . . . . . . . . . . 5.2.2 Free particle in a box . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Systems composed of indistinguishable particles . . . . . . . . . . . . . . 5.4 The density matrix and the partition function of a system of free particles iii

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . .2 Bose-Einstein systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Methods for studying phase behaviour . . . . . . . . .2 Local equation of state . . . . . . . . . . . . .3 Phonons and the speciﬁc heat . . . . . . . . . . . . . . . . . . . . . 10. .1 Diffusion . . . . . . . . . . . . . . . . . . . . . . . . 7. . . . . . . . . . .1 Planck distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Irreversible processes: macroscopic theory 11. . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . 10. . . . . . . . . . . . . . . . . . 10.3. . . . . . . . . . . . . . . . . . . . . . . . .2 The virial expansion and the Van der Waals equation of state . . . . . . . . .1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . .6 Renormalisation theory . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . .4 Landau Ginzburg theory and Ginzburg criterion 10. . . . . . . . . . . .1 Thermodynamics of free quantum gases 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Examples: gaseous systems composed of molecules with internal motion . . . 10.6 Derivation of hydrodynamic equations . . . . . . . . . . . . . . . . .2 The low-energy spectrum of helium . . . . . .1 About phase transitions . . . . . . . . . . . .1. . . . . 11. . . . . . . . . . . . . . . . . . . . . . . . . 10 Introduction to phase transitions 10. . . 8 Statistical mechanics of interacting systems: the method of cluster expansions 8. . . . . . . . 10. . . . . . . .4 General analysis of linear transport . . . . . . . . . . . 7. . . . . . .3 Viscosity . . . . . . . . . . . . . . . . . . . . . . . . .10Systems with continuous symmetries . . . . . . . . . . . . . . . .2 Bose–Einstein condensation . . . . . . . . . . . . .iv 6 The theory of simple gases 6. . . . . . . 12 Fluctuations and transport phenomena 12. . . 9 The method of quantized ﬁelds 9. . . . . . . . . . . . . 7. . . . 12. .2. . . . . . . . . . . . . . 12. . . . .1 Degenerate Fermi gas . . . . . 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Examples of quantum statistics 7. . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 Universality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Thermal conductivity .3 Landau theory of phase transitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7. . . . . . . . . . . .1 Cluster expansion for a classical gas .2. 9.7 Scaling relations . . .3 Heat and particle diffusion . . . . . . . . . 10. . . . . . . . . 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10. . . . . .5 Exact solutions . . . . . . . . . . . . . . . . . . . . . . . . . 7. . 10. . . .9 Examples of renormalisation transformations . . . . . . 7. . . . . . . .1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Landau diamagnetism . . . .2 Pauli paramagnetism . . . . . . . . . . . . . . . .1 The superﬂuidity of helium . . . . . .5 Coupling of different currents . . . . . . 12. 8. . . . . . . . . . . . 11. . . . . . . . . . . . 37 37 39 42 42 43 43 44 45 46 46 50 51 53 53 59 63 63 66 68 68 70 75 77 79 81 89 89 90 91 98 98 98 100 102 105 106 109 109 110 112 112 . . . . . . . . . . .1 An ideal gas in other quantum-mechanical ensembles – occupation numbers . 7. . . . . . . . . . . . . . . . . . . . . . . . .3. . . . . . . . . .1 Motion of particles . . .3 Fermions . . .3. . .

. . . . . . . . .4 Derivation of the Navier–Stokes equations . . . . . . . . .v 12. . . . . . . .4 General analysis of linear transport . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Fokker Planck equation and restoration of equilibrium .3 Fluctuations – the Wiener-Kintchine theorem . 113 12. . . . . . . 13. . . . . . 117 13 Nonequilibrium statistical mechanics 13. . . . . . .2 The Boltzmann equation . . . . . . . . . . . . . . .3 Equilibrium – deviation from equilibrium . . . 13. . . . . . . . . . . . . . . . . 13. . . . . . . . . . . . . . . . . . . . . . . . . .1 Langevin theory of Brownian motion . . . . . . . . . 122 122 124 126 129 . . . . . . . . . . 115 12. . . . . . . . . . . . . . . . . . . . .

the quantity Ω(0) is sharply peaked around its maximum as a function of E1 . • A macrostate is deﬁned by values of the macroscopic parameters which can be controlled. quantities which scale linearly with V (or N ) are called extensive. This number may not be countable. while quantities which do not scale with V (or N ) are called intensive. to an amount E0 . V2 . 1 and 2 are in thermal contact. E1 . For an isolated classical system. N2 .1 The macroscopic and the microscopic states Notions of statistical mechanics: • Extensive/intensive quantities: N . E1 . the most likely value of E1 is equal to the average value of E1 . That is. V2 . Therefore. We ﬁnd the most likely value by putting ∂ ln Ω(0) (N1 . N2 . In that case. E2 ) ∂ E1 1 . E1 )Ω(N2 .2 Contact between statistics and thermodynamics Two systems. E2 ). We let both quantities go to inﬁnity. volume V and particle number N . The density n = N /V is an example of an intensive quantity. with E = E1 + E2 constant. E2 ) = Ω(N1 .1 The statistical basis of Thermodynamics This chapter reviews material that you should have seen before in one way or another. 1. but they can exchange energy. • The quantity Ω(N . a microstate is a set of positions and momenta which are consistent with the prescribed energy E . V1 . V1 . V1 . is given by Ω(0) (N1 . For a thermally and mechanically isolated system. Therefore it is kept very brief. however. the number of states within a ﬁxed energy band is ﬁnite (for a ﬁnite volume). the total system has a number of microstates which. for a particular partitioning of the total energy (E1 . T ) is the number of microstates which are consistent with a particular macrostate. V2 . Because the particle numbers are very large. V. • A microstate is a particular state of a system which is consistent with the macrostate of that system. In that case. V are respectively the number of particles and the volume of the system. E and V . but we shall see that this problem is only relevant in the classical description – in a proper quantum formulation. The total energy is ﬁxed. while keeping the ratio n = N /V constant. E2 ). these are N . their respective volumes and particle numbers are ﬁxed. 1.

S µ= ∂E ∂N . S = k ln Ω(N . G = A + PV = E − T S + PV = µ N .2 equal to 0. The following relations can be derived straightforwardly: P=− ∂E ∂V . V.V Here. In fact.S T= ∂E ∂S . i. pressure and temperature are the same in both. N .g. we can study two systems which are not only in thermal equilibrium (i. E ) = kB T ∂N is the same in both. Finally. then the quantity ζ =− µ ∂ ln Ω(N . V and E ) by using the relation S = kB ln Ω and the thermodynamic relation dE = T dS − PdV + µ dN . which can exchange thermal energy) but also in mechanical equilibrium (i. E ) which has a well-deﬁned meaning in statistical physics (as do N . . . you should know the remaining most important thermodynamic quantities: • Helmholtz free energy • Gibbs free energy • Enthalpy A = E − T S. . through a hole). in addition to the temperature. which can change their volumes V1 and V2 while keeping the sum V1 + V2 = V0 constant). We have β = 1/(kB T ) S is the entropy and T the temperature.β denotes a partial derivative with respect to γ at constant α and β . This leads to the condition for equilibrium: ∂ ln Ω(N1 .3 Further contact between statistics and thermodynamics Similar to the foregoing analysis. The quantity µ is known as the chemical potential. V. V1 . We have derived relations between these and the fundamental quantity Ω(N .e. E2 ) = . the quantity ∂ ln Ω(N . 1. /∂ γ )α . V. P and µ are thermodynamic quantities. ∂ E1 ∂ E2 The partial derivative of ln Ω with respect to energy is called β . E ) P η= = kB T ∂V is the same in both systems. V1 . V.e. H = E + PV = G + T S. E1 ) ∂ ln Ω(N1 . (∂ . E ). N . If the systems can exchange particles (e. We then ﬁnd that. V.e.

The particles have wavefunctions ψ (x. N . The number of points in such a grid is called Γ.V ∂S ∂T = N . 2m L2 x 8mL2 x For N particles. The surface of a sphere of radius r in 3N dimensions is given by 2π 3N /2 r3N −1 . the number of ways the particles can be distributed in that volume scales as V N .P ∂ (E + PV ) ∂T = N . the energy is the square of the distance of the appropriate point on the 3N dimensional grid. But it might occur that none of the grid points lies precisely on the sphere (in fact.E For a consistent derivation of the entropy. (3N /2 − 1)! Multiplying this by δ r gives the volume of a spherical shell of thickness δ r. V ) = kB T ∂V N = kB .3 • Speciﬁc heat at constant volume CV = T • Speciﬁc heat at constant pressure: CP = T ∂S ∂T = N . y. i.V ∂E ∂T . We use the fact that each grid point occupies a unit volume to obtain the number of grid points within the shell. we therefore consider the number of points in a spherical shell of radius 2mE /ℏ2 and thickness much smaller than the radius (but much larger than the grid constant).z .P ∂H ∂T . E . V. Therefore P ∂ ln Ω(N . that is. L L L ℏ2 π 2 2 h2 2 2 2 n + n + n n2 + n2 = y z y + nz . In order to . Ω ∝ V N. The number Ω(N . that is rather likely)! In order to obtain sensible physics.y + n j.e.x + n j.4 The ideal gas If N particles in a volume V do not interact. z) = with energy 2 L 3/2 sin ny π y nz π z nx π x sin sin . N . V N . we consider a particular example: a quantum mechanical system consisting of particles within a cubic volume V = L3 and with total energy E .P 1. we have the relation E= E= h2 8m j =1 ∑ N 2 2 n2 j. E ) is the number of points on the surface of a sphere with radius 2mE /ℏ2 in a grid with unit grid constant in 3N dimensions.

2 5 CP = NkB . A straightforward analysis. before mixing.V . According to quantum mechanics. using the expression for the entropy derived in the previous section. This is to be expected because. N and S constant). leads to the relations 3 E = NkB T .4 include only positive values for each of the n j. 3V These relations are speciﬁc examples of more general thermodynamic ones. z. were at the same pressure and temperature. TV γ −1 = const with γ = 5/3. V1 V2 Now consider the case where the two gases contain the same kind of molecules.5 The entropy of mixing and the Gibbs paradox If we mix two gases which. we ﬁnd for the ratio of the two speciﬁc heats 5 CP = . y.α . PV γ = const. α = x. 2 From the last two relations. two conﬁgurations obtained by interchanging the particles must be considered .e. 1. This expression can be inverted to yield the energy as a function of S. CV 3 It can be veriﬁed that the change in entropy during an isothermal change of a gas (i. 2 We have neglected additional terms containing the thickness of the shell – it can be shown that these are negligibly small in the thermodynamic limit. when the two original gases consisted of different types of molecules. we arrive at the following expression for the entropy: S(N . leads to a mixing entropy ∆S: ∆S = kB N1 ln V1 + V2 V1 + V2 + N2 ln . we must multiply this volume by a factor 2−3N . 2 / 3 3NkB 4π mV This equation. V. N ) = 3h2 N 2S exp −1 . the entropy has increased tremendously by the fact that both species now have a much larger volume at their disposal.e. E ) = NkB ln V h3 4π mE 3N 3/2 3 + NkB . together with T −1 = (∂ S/∂ E )N . during an adiabatic change (i. V and N : E (S. the entropy has changed. Using ﬁnally the fact that N is large. 2 3 CV = NkB . This difference is called the mixing entropy. V. The work done by such an adiabatic process is given by (dE )adiab = −PdV = − 2E dV. then it turns out that after the mixing. N and T constant) is S f − Si = NkB ln (V f /Vi ) . Furthermore.

. V. The fact that the entropy changes without heat transfer is allowed. E ) = NkB ln V N 3 + NkB 2 2π mkB T 5 + ln 3 h2 . . The equals-sign only holds for reversible processes.5 as being identical. This paradox is known as the Gibbs paradox. n1 !n2 ! . A more careful rederivation of the entropy. For a system with n1 particles in quantum state 1. changes the entropy. In deriving the Sackur–Tetrode formula. we have taken the ni to be either 0 or 1. the mixing should not inﬂuence the entropy. taking the indistinguishability of the particles into account. It is a result of neglecting the indistinguishability. This formula is derived by multiplying Ω (the number of states) by 1/N ! in order to account for the indistinguishability in case alle particles occupy different quantum states. leads to the famous Sackur-Tetrode formula: S(N . . nor does it involve any work. but does not correspond to any heat transfer. as the second law of thermodynamics states that ∆Q ≤ T ∆S. This expression for the entropy leads to ∆S = 0 for identical particles. In that case. so the above result for the mixing entropy cannot be correct. Note that the process in which the wall is removed.6 The “correct” enumeration of the microstates This section argues how and why the indistinguishability of the particles should be included in the derivation of the entropy. by N! . 1. n2 particles in state 2 etcetera. The expression above for ∆S remains valid for non-identical particles. it boils down to dividing the total number of states calculated for distinguishable particles.

t ). For example. p ˙j = − ∂qj q ˙j = 6 . which nevertheless occur in the deﬁnition of H . q ˙ j . In our case. In fact. q j) = ∑ p jq ˙ j − L. j ∂L ∂q ˙j Note that H is a function of the p j and q j . Remember the generalised coordinates can be any coordinates which parametrise the accessible coordinate space within perhaps some given constraints. we shall most often be dealing with a volume within which the particles must move. the q ˙ j must be formulated in terms of q j and p j by inversion of the expression giving the p j . but not of the q ˙ j . The Euler-Lagrange equations of motion can now be formulated in terms of the Hamiltonian: ∂H . The motion of the particles is determined by the Lagrangian. The equations of motion are the Euler-Lagrange equations: ∂L d ∂L = . ∂ pj ∂H . which is a function of the generalised coordinates q j and their derivatives with respect to time q ˙ j: L = L(q j .1 Phase space of a classical system The phase space is the space of possible values of the generalised coordinates and canonical momenta of the system. we have p j = mq ˙ j. dt ∂ q ˙j ∂qj The canonical momenta are deﬁned as pj = − and these can be used to construct the Hamiltonian: H (p j. for a particle in 3D. so the coordinates are simply the variables ri (i labels the particles) which are constrained to lie within V .2 Elements of ensemble theory 2. for which q ˙ j can very easily be written in terms of the p j .

p)ρ (q.7 These equations are completely equivalent to the Euler-Lagrange equations. At that time. given by ρ v. Suppose we stay at some point (q. 2. and from the fact that the shape of ω can be arbitrary. In fact. p. ∂t We now work out the second term: ∂q ˙j ∂ p ˙j ∂ρ ∂ρ ∂ρ +∑ q ˙j + p ˙j +ρ ∑ + ∂t ∂qj ∂ pj ∂qj ∂ pj j j = 0. p. ρ (q. From these considerations. p) in phase space. q ˙): it is the velocity of the points in phase space. In statistical mechanics. the latter are second order differential equations with respect to time. t ). or trajectories starting outside ω and moving in. Any physical quantity is deﬁned in terms of the dynamical variables p j and q j . we see that the following relation must hold: ∂ρ + div(ρ v) = 0. The latter is a point in phase space. the density function ρ . i.e. t ) d 3N p d 3N q where the denominator is necessary in the case where ρ is not normalised. Rather we want to know the probabilities to ﬁnd our system in the points of phase space. This probability is also called the density function ρ (q. so that we can evaluate expectation values of physical quantities. The ﬂux of points across the boundary is the only cause of change in the density inside ω . Liouville’s theorem is about the change in the course of time of the density function. ∂t The density in some volume ω changes as ω ∂ρ dω ∂t (d ω is shorthand This change can only be caused by trajectories starting within ω and moving out of it. ρv · n where σ is the boundary of ω . p) ((q. Using Gauss’ theorem. t ) d 3N p d 3N q . the change of the density is ∂ρ . σ ˆ dσ . we are interested in the probability to ﬁnd a system in a particular point (q. . where v is shorthand for the vector ( p The number of points leaving and entering the system per unit of time can be evaluated as for d 3N p d 3N q). The Hamiltonian equations clearly give a recipe for constructing the time evolution in phase space given its original conﬁguration. p) is shorthand for all the coordinates). which are here reformulated as twice as many ﬁrst-order differential equations. this can be written as a volume integral: σ div(ρ v) d ω .2 Liouville’s theorem and its consequences We are not interested in the individual dynamical trajectories of a system. The ﬂux of phase space points is ˙. as there are no ‘sources’ or ‘sinks’ (trajectories do not disappear or appear). p. from which a dynamical trajectory starts. The expectation value of such a quantity f can be written as f = f (q.

This is usually assumed. occurs equally likely. . and we can use this in order to ﬁnd the average f as described above.3 The microcanonical ensemble Any particular choice for the density function is called an ensemble. The microstates all occur with a probability given by the density function – therefore. For an isolated system. the last group of terms is seen to vanish. They are for classical mechanics what commutators are for quantum mechanics. there are some mathematical difﬁculties involved in using a delta-function – one may formulate this as a smooth function with a variable width which can be shrunk to zero. ∂t where the expression on the left hand side can also be written as d ρ (t )/dt . the resulting equations for the classical density function and the quantum density operator in Heisenberg representation can be compared: dρ ∂ρ = + {ρ . It is useful to emphasise the difference between ∂ ρ /∂ t and d ρ /dt . corresponds to the expression given at the end of section 2. This refers to the idea of having a large collection of independent systems. so the question arises how this condition can be satisﬁed. Obviously. every point in phase space compliant with that value. time-dependent ﬁeld) but being a function of H .1 2. } are called the Poisson brackets. whereas the latter is the change of the density function as seen by an observer moving along with the trajectory of the system. H } = 0. The former is the change of the density function at a ﬁxed point in phase space.8 Using the equations of motion. volume. H } (class.1. the energy is ﬁxed. This now is Liouville’s theorem. One possibility is to have a density function which is constant in time and in phase space so that both ∂ ρ /∂ t and [ρ . particle number etcetera). p). It implies that for some particular value of H . all having the same values for their controllable external parameters (energy.). H ] (quantum) . The density function ρ gives us the probability that we ﬁnd a system in a state (q. We now deﬁne equilibrium as the condition that ∂ ρ /∂ t be equal to 0. so we can write ρ = ρ (H ) = δ [H (q. This is however not physically acceptable. and we are left with dρ ∂ρ ∂ρ ∂ρ = +∑ q ˙j + p ˙j dt ∂t ∂qj ∂ pj j = ∂ρ + {ρ . dt ∂t ℏ The intimate relation between classical and quantum mechanics is helpful in formulating statistical mechanics. p) − E ]. dt ∂t dρ ∂ρ i = − [ρ . In fact. determining the average of a physical quantity over this ensemble of systems. This average is called the ensemble 1 Parthia formulates this a bit differently on page 34. This is the famous postulate of equal a priori probabilities. but (in general) different microstates. Another possibility is to have ρ not depending explicitly on time (this is mostly the case – it means that there is no external. but I am not very happy with his description. The last equation expresses the fact that such an observer does not see a change in the density. The brackets {. H ] vanish. as inﬁnite momenta are allowed in that case.

the relation between the number of states Γ and the occupied volume ω in phase space is given by Γ = ω /h3N . Obviously. ω = k/m. it appears that there is a fundamental volume h3N which is occupied by quantum state of the system. Then the volume of the ellipse can be found to be ∆Γ = 2π ∆ . Therefore.4. that number is always inﬁnite (irrespective of the number of particles). p→ √ q→ 2mE 2E /mω with. the integral over q yields V N .4 Examples First we consider the case of N non-interacting point particles. In order to ﬁnd the volume (in our 2D phase space this is a surface area). the phase space volume accessible to the system is proportional to the number of microstates accessible to the system. these packets occupy a volume ∼ h per dimension in phase space. 2. Do not confuse this ω (angular frequency) with that representing the volume in phase space. If we consider phase space as continuous (as should be done in classical mechanics). ω As the potential vanishes. For the microcanonical ensemble. This result is called the fundamental postulate of statistical mechanics.9 average. We call this shell ∆ω . The number of states in the energy shell ∆ω is given by: ∆Γ = ∆ω d 3N p d 3N q. Therefore. ∆Γ = π mE ) 3 (3N /2 − 1)! h E . as the states are given as wavefunctions and not as points (q. however. As an example. In view of the Heisenberg uncertainty relation. The entropy can then be calculated as above. we can scale the coordinates in order to map the ellipse onto a circle: p q . This is equal to the time average of the quantity f in the stationary limit. Γ is the volume of a thin shell in phase space where the energy lies between E − ∆/2 and E + ∆/2. The connection between the two can be made by considering wavepackets which are localised in both q-space and p-space. the points with constant energy are seen to lie on ellipses. In order to take the indistinguishability of the particles into account. p). and we ﬁnd for the number of states within the shell: N V 1 ∆E 3/2 ( 2 . we must divide this number by N !. we consider the harmonic oscillator: p2 k 2 H= + q 2m 2 From this equation. The integral over the momenta can be evaluated using the results for an N -dimensional shell used in section 1. In quantum mechanics we do not have this problem. resulting in the Sackur–Tetrode formula. as usual. that is an ideal classical gas. Our general formalism also allows to evaluate the entropy of a single particle.

we speak of a canonical. Therefore the temperature of the subsystem will be a control parameter. isolated system which we divide up into a very small one and the rest. Pr is the Boltzmann distribution function. As the subsystem is very small in comparison with the total system. the multiplicity of the state with energy Er is Pr = Ω(E − Er ). T with β = 1/kB T . Writing ρ = ∑ |r Pr r| r leads to the same expression as above S = −kB ∑ Pr ln Pr . T) ensemble. In a quantum mechanical formulation. Therefore. Pr = P(Er ) ∝ exp(−Er /kB T ) = e−β Er 3. We then have: use Pr = exp[S(E − Er )/kB ] = exp so that we obtain S(E ) − ∂ S(E ) Er /kB ∂E = exp S(E ) − Er /kB . just as the number of particles N and its volume V . The small subsystem can exchange energy with the rest of the system.1 Equilibrium between a system and a heat reservoir Suppose we have a large.2 A system in the canonical ensemble A more formal approach can be taken in the calculation of the canonical and other distributions that we shall meet hereafter. and this number is given as Ω(E − Er ). this entropy is formulated terms of the quantum density operator as S = −kB Tr ρ ln ρ . and with energies distributed according to the Boltzmann factor. Consider a state r of the small subsystem with energy Er . V. but its volume and number of particles is constant. We know that Ω of the heat bath is given as exp(S/kB ). or (N. r 10 . If we consider a set of systems which are all prepared with the same N .3 The canonical ensemble 3. V and T . How likely is it to ﬁnd the subsystem in this state? That depends on the number of states which are accessible to the rest of the system (this is called the heat bath). where E is the energy of the total system. its temperature will be determined by the latter. which is based on a very general deﬁnition of entropy.

e. We now have an additional constraint. that is. In the context of quantum mechanics. the distribution which maximises the entropy is the one in which each state has the same probability of occurrence. This then leads to 1 Pr = . parametrised by λ . The Pr are thus all equal. We deﬁne F = S−λ and now require F to be stationary: r =1 ∑ Pr M ∂F = −kB (1 + ln Pr ) − λ ∂ Pr This leads to a family of solutions Pr = exp − kB + λ kB . Now suppose that there is a number M of states with the prescribed energy. E = ∑ Pr Er r is given. the solution of the problem is similar to that of the classical case. the Hamiltonian is conserved and it can be identiﬁed with the energy. r r . Let us ﬁrst note that the (NV E ) ensemble is the most natural one to deﬁne in classical or quantum mechanics: the number of degrees of freedom is well-deﬁned (via the particle number N ) and the potential does not explicitly depend on time (the volume is ﬁxed. this is a bit tricky as we must calculate the variation of an operator. We now have two Lagrange multipliers. it turns out that expressions for the entropy that can be derived from more physical arguments are all compatible with this general expression. However. We must now adjust λ such as to satisfy the constraint that Pr be normalised. Then. This is done using a Lagrange multiplier λ . M We see that for the microcanonical ensemble. V or E be ﬁxed. we may relax any of these conditions and require the expectation value to assume a certain value rather than the stronger condition that the parameter may only assume a certain value. i. the walls do not move). Here Er is the energy of the state r. We must ﬁnd the distribution Pr which makes S stationary under the constraint that Pr is normalised. one for the energy (which we call kB β ) and one (again λ ) for the normalisation: F = S − λ ∑ Pr − kB β ∑ Pr Er . the density matrix will assume a form which maximises the entropy deﬁned this way.11 The basic postulate now is that. Furthermore. Instead of requiring that each of the parameters N . We shall work this out for the energy. This expression for the entropy is often used in information theory. given expectation values for external parameters. if we assume that the density operator can be written in the form ρ = ∑ |r Pr r| r with |r being eigenstates of the Hamiltonian.

∂β ∑ r .12 Following the same procedure as above. Zλ where σ can be identiﬁed with −β µ . ∑r e−β Er The denominator ensures proper normalisation. we obtain Pr = 1 exp(−β Er − σ N ). we see that we can write U as U= E = U = E =− ∂ ln e−β Er . If we relax the particle number but ﬁx its expectation value. we ﬁnd Pr = 1 exp(−β Er ). Let us analyse the canonical partition function a bit further. The transformation from S to ln Q which leads to this type of relation is known as the Legendere transformation. The expectation value of the energy can be determined as ∂ ln Q. µ is the chemical potential. we can write the entropy as S = −kB ∑ Pr ln Pr = −kB ∑ Pr [− ln Q − β Er ] = kB ln Q − kBβ r r ∂ ln Q. Returning to the derivation of the canonical distribution function. The Lagrange parameter β which we can identify with 1/kB T serves as the parameter which can be tuned in order to adjust the expectation value of the energy to the required value. E =− ∂β Using Pr = exp(−β Er )/Q. We have: The Boltzmann distribution is the distribution which minimises the Helmholtz free energy. ∂β where we have used ∑ Pr = 1 and ∑r Pr Er = E . − =− T T The quantity A = E − T S is called the (Helmholtz) free energy. we note that the function we have maximised can be written as S−λ −T E 3. Qλ Q – the partition function – is deﬁned in terms of the multiplier λ – it serves merely to normalise the probability distribution. Looking at the above equation. Now we write this expression (disregarding the constant λ ) as E −TS A .3 Physical signiﬁcance of the various statistical quantities in the canonical ensemble Let us ﬁrst calculate the energy: ∑r Er e−β Er . We see that this quantity was minimised as a function of Pr . in particular it ensures that the average value of 1 is equal to 1.

Furthermore. it follows that the entropy vanishes at zero temperature (‘third law of thermodynamics’). that dA = dU − T dS − SdT = −SdT − PdV + µ dN .V ∂U ∂T N .V ∂ (A/T ) ∂ (1/T ) . We have seen that the probability Pr with which a conﬁguration with energy Er occurs. r From this relation. r Q is called the partition function. we have that is. by the expectation value of −kB T ln(QN Pr ): S = kB ln QN − kB ln(QN Pr ) = −kB ln Pr = −kB ∑ Pr ln Pr .13 It is useful to deﬁne QN = ∑ e−β Er . By taking the temperature derivative of U we obtain the expression for the speciﬁc heat: CV = Moreover.V U ∂ (kB T ln QN ) = kB ln QN − .V = −T 2 ∂ 2A ∂T2 . we see that if we keep the volume and the particle number constant. From this. P=− ∂A ∂V . isothermal system. Let us relate the quantities we have seen so far to thermodynamics. ∂T T We now replace U . For a system at constant V . this relation has become the starting point for studying information theory. N . dA = −PdV. N and T . we have: U = A+TS = A−T from which we can infer that A = −kB T ln QN .V . N . which is the expectation value of Er . from ∂A ∂T N . .V = −T 2 ∂ ∂T A T = N . (see above). we have: S=− ∂A ∂T N .V From the ﬁrst relation. This is obviously the same entropy as was introduced in the previous section. Pr = QN The entropy can be calculated as S=− ∂A ∂T = N . we know for the Helmholtz free energy A = E − T S. T .T µ= ∂A ∂N . The Hemholtz free energy represents the work which can be done by a closed. is given by the Boltzmann factor: e−Er /(kB T ) . where entropy is a measure for the reliability of communication.V dA = −SdT − PdV + µ dN . the change in free energy is completely determined by the work done by the system. N .

. provided we divide by the ‘unit phase space volume’ h3N and by N ! in order to avoid over-counting of indistinguishable conﬁgurations obtained from each other by particle permutation. when we evaluate the sum over some energydependent quantity.5 Classical systems We now show how to evaluate expectation values for a system consisting of interacting point particles.14 3. but the integral over the momenta can be evaluated analytically! The reason is that the exponential can be written as a product and the integral factorises into 3N Gaussian integrals: e−β ∑i pi /(2m) d 3N p = 2 e−β p1 /(2m) d 3 p1 2 e−β p2 /(2m) d 3 p2 . For a system with a strictly discrete spectrum. ∞ −β E dE −∞ g(E )e 3. the energies Ei might occur with a multiplicity (degeneracy) gi . In that case. we have ∑ → ∑ gi . p) = ∑ p2 i + V (q1 . when evaluating the sums over states r. one over py and one over pz . 2 e−β pN /(2m) d 3 pN . ∑i gi e−β Ei In practice. Note that the pi are vectors in 3D. the energy is usually continuous or almost continuous. p) d 3N qd 3N p. the probability of having an energy Ei is given by Pi = gi e−β Ei . e−α x dx = α −∞ . The expression for QN looks quite complicated. Therefore we have QN = where 1 h3N N ! N e−β H (q. . q3N ). .4 Alternative expressions for the partition function It is very important to realise that. In that case P(E ) = g(E )e−β E . . V ≡ 0. 2 m i=1 For the ideal gas. . In the previous chapter it was argued that the sum over the available quantum states can be replaced by a sum over e volume ω in phase space. . r i In fact. This quantity is deﬁned as follows: Number of quantum states with energy between E and E + dE = g(E )dE . we should not confuse those states with their energies. 2 The integral over one particle then factorises into one over px . H (q. In that case. Now we use the Gaussian integral result: ∞ 2 π . gi is replaced by the density of states g(E ).

the energy can take on any possible value between the ground state and inﬁnity. we ﬁnd that this quantity is a very strongly increasing function of the energy. V = 0 and we can evaluate the remaining integral: it yields V N (V is the volume of the system – do not confuse it with the potential!). we therefore have: ∆E = U √ kB T 2CV ∼ 1/ N . From U= E = where.e. we see that the energy. as usual. we see that ∑r Er e−β Er ∑r e−β Er 2 e−β Er ∑r E r = −(∆E )2 . From statistics. we expect the physics of the system to be almost the same of that in the microcanonical ensemble (where the energy is actually constant). To show that the energy is indeed sharply peaked around U . we calculate the ﬂuctuation. turns out to be almost constant. Therefore. The result is that the probability distribution of the energy is very sharply peaked around its mean value U = E . . Therefore. U In the thermodynamic limit (N → ∞). β = 1/kB T . the partition function of the ideal gas is found as: N (2π mkB T )(3/2) 1 V QN = N! h3 3. ∑r e−β Er ∂U = ∂β The quantity ∂ U /∂ β is equal to ∑r Er e−β Er ∑r e−β Er 2 − ∂U ∂U = kB T 2CV . the number of microstates at energy E .15 in order to obtain: 1 (2π mkB T )(3/2) QN = N! h3 N e−β V d 3N q. This result is usually referred to as the ‘equivalence of ensembles’. whereas the Boltzmann function exp(−E /kB T ) strongly decreases with energy. i. which scales linearly with N . = kB T 2 ∂β ∂T Realising that CV is an extensive quantity. For an ideal gas. The actual probability with which a particular value of the energy occurs is proportional to P(E ) ∝ g(E )e−E /kT . which is allowed to vary at will.6 Energy ﬂuctuations in the canonical ensemble: correspondence with the micro-canonical ensemble In the canonical ensemble. In general. we see that the relative width becomes very small. Therefore. we know that the width of the distribution is given by (∆E )2 = E 2 − E 2 . The prefactor g(E ) is the density of states – it is proportional to the Ω(E ).

. R and P represent all positions and momenta. the coordinates si are simply rescaled in such a way that they occupy a volume of the same shape as the ri . . Obviously. V 1/3sN )/(kB T )]d 3N S. the derivative with respect to V can be evaluated: −2/3 ˜N ∂Q ˜N − V = NV N −1 Q ∂V 3kB T ∑ si · ∇iU exp[−U (V 1/3 s1 . In the book. Now. . . This is done as follows. Therefore we can write: exp[−U (r1 . but everything is rescaled to a unit volume. . . . T = i i exp − ∑i 2m + V (R) /(kB T ) d 3N Pd 3N R ∑i 2m i + V (R) /(kB T ) d 3N Pd 3N R exp − ∑i 2m p2 p2 p2 . . We obtain T = 3N p2 2m exp[− p2 /(2mkB T )]d p 3NkB T . = 2 exp[− p2 /(2mkB T )]d p This result is known as the equipartition theorem: it tells us that the kinetic energy for each degree of freedom is kB T /2. To evaluate the volume-dependence of this object. .16 3.T 1 ∂ QN . rn )/(kB T )]d 3N R = V N exp[−U (V 1/3 s1 . The second theorem gives us an expression for the pressure P (the derivation given here is somewhat different from that of the book). . i . rn )/(kB T )]d 3 R is to be considered (note that we call the potential function U – this is to avoid confusion with the volume V). and they can be evaluated using the same factorisation which led to the evaluation of the partition function of the ideal gas (the integrals over the coordinates R cancel). All sums over i run from 1 to N . we write for the coordinates ri of the particles: ri = V 1/3 si . . . . . Every conﬁguration in a volume V has a one-to-one correspondence to a conﬁguration of the si . V 1/3 sN )/(kB T )]d 3N S where the prefactor arises because of the change of integration variables.5 allows us to calculate the expectation value of the kinetic energy T . . We know that P=− (see above).7 Two theorems – the ‘equipartition’ and the ‘virial’ The analysis of the classical gas in section 3. this theorem is proven more generally. QN ∂ V First we realise that the integral over the momenta is volume-independent – therefore only the part ˜N = Q exp[−U (r1 . the contributions to the result from each momentum coordinate of each individual particle are identical. . that is. Now we replace A by −kB T ln QN : P = kB T ∂A ∂V N .

we can take any pair instead of 1 and 2. . . r3 . rN )] d 3 r3 . N. j. . The free energy now follows as A = −kB T ln QN = NkB T ln ℏω kB T . The formal deﬁnition of g(r) is g(r) = V 2 exp [−β U (r1 . QN = (β ℏω )N where we have assumed that the oscillators are distinguishable. The result for oscillator system is: 1 . which is well known for the ideal gas. d 3 rN . For large separation r. . dr where we have introduced the pair correlation function.8 A system of classical harmonic oscillators Now we consider a classical system with an Hamiltonian given by H=∑ p2 mω 2 2 i + q . QN Because the particles are identical. ∂r 3. rN ) = i. the partition function can be evaluated analytically as the Hamiltonian is a quadratic function of both the momenta and the coordinates. r2 . . The virial theorem can be reformulated in terms of the pair correlation function: PV 2π N = 1− NkB T 3V kB T ∞ g(r) 0 ∂ u(r) 3 r dr. . 2 i i=1 2m N For this system. In that case. . j>i ∑ N u ri − r j . which gives the probability of ﬁnding a particle pair at separation r = |r2 − r1 |. . g(r). We see that for U = 0 we have PV = NkB T . g(r) tends to 1. The calculation therefore proceeds analogous to that for the ideal gas where the Hamiltonian is a quadratic function of the momenta only. the rightmost term in the virial theorem can be rewritten as i=1 ∑ ri · N ∂U ∂ ri = N (N − 1) ∂ u(r) r 2 ∂r = N (N − 1) 2 ∞ r 0 du g(|r1 − r2 |)d 3 r1 d 3 r2 . . . .17 Collecting all the terms we obtain PV 1 = 1− kB T N 3NkB T i=1 ∑ ri · ∂ ri N ∂U . Very often interaction potentials are modelled as a sum over all particle pairs: U (r1 .

Next we consider a collection of quantum harmonic oscillators in the canonical ensemble. the energy is equally divided over the potential and the kinetic energies. and we obtain: QN = ∑e n −β ℏω (n+1/2) N = e−β ℏω /2 1 − e−β ℏω N From the partition function we obtain. similar to the classical case: A = N ln And.18 From this we ﬁnd µ= P=− S=− U= From the last equation. + β ℏω U =N 2 e −1 S = NkB e−β ℏω And. The fact that U = NkB T is in agreement with the equipartition theorem as the Hamiltonian has two independent quadratic terms (for q and p) instead of only one. It shows that for harmonic oscillators. . β ℏω − ln 1 − e−β ℏω −1 ℏω ℏω . hence QN = {ni } ∑ e−β ℏω ∑ (n +1/2) . from this. ﬁnally CV = CP = NkB(β ℏω )2 eβ ℏω eβ ℏω − 1 2 . the quantum harmonic oscillator does not obey equipartition: we see that only the ﬁrst term in the expression for the energy is in accordance with that theorem – the second term gives a positive deviation from the equipartition result. ∂ (A/T ) ∂ (1/T ) = NkB T . µ = A/N . .V CV = NkB = CP .T = NkB ln N . This partition function factorises in a way similar to the classical system. = 0.V kB T ℏω +1 . Interestingly. i i where ∑{ni } denotes a sum over all possible values of all numbers ni . N . P = 0. The states for oscillator number i are labeled by ni . ℏω + kB T ln 1 − e−β ℏω 2 . This is simpler to evaluate than the classical case. N .T ℏω kB T . we ﬁnd ∂A ∂N ∂A ∂V ∂A ∂T = kB T ln V.

For high temperatures.4 0. the Langevin function behaves as L(x) ∼ x/3 (see ﬁgure 3.1). Each moment interacts with a magnetic ﬁeld H.1: The Langevin function. that is. β µH where L(x) is the Langevin function. for small values of x. The dashed line is the graph of x/3. 3. The partition function can now be evaluated: Q1 = eβ µ H cos ϑ sin ϑ d ϑ d ϕ = 4π sinh(β µ H ) . It is shown in ﬁgure 3. 3kB T The magnetic susceptibility is deﬁned as χ= ∂M ∂H .9 The statistics of paramagnetism Consider a system consisting of a set of magnetic moments. where ϑ is the angle between the moment and the z-axis. In that case we can consider again a system of only one magnetic moment and construct the partition function for N moments by raising that for a single moment to the N -th power. Without loss of generality we can take H along the z-direction so that H = −µ H cos ϑ . Note the difference between the Hamiltonian H and the ﬁeld H.2 0 0 2 4 6 8 10 12 Figure 3. The interaction Hamiltonian is given by H = −µ · H. β µH We can also calculate the average value of the magnetic moment: µz = 2π π β µ H cos ϑ cos ϑ sin ϑ d ϑ d ϕ 0 0 µe 2π π β µ H cos ϑ sin ϑ d ϑ d ϕ 0 0 e = µ coth(β µ H ) − 1 = µ L(β µ H ). but the interaction between the moments is neglected.19 1 0.6 0.8 0.1. so we have µ2 M= H.

= N µB tanh(β ε ). When we increase the temperature. S=− ∂A ∂T H = NkB {ln [2 cosh(β ε )] − β ε tanh(β ε )} . H We see that U = −MH . in which case the behaviour of the system approaches classical behaviour. Therefore we immediately ﬁnd that the partition sum is given as: QN (β ) = e−β ε + eβ ε N = [2 cosh(β ε )]N .e. These graphs show several interesting features. the spins assume values either ℏ/2 or −ℏ/2 when they are measured along an arbitrary axis. . i. This relation is known as the Curie law of paramagnetism. For high temperatures the spins assume more or less random orientations. the situation of real quantum systems (with smaller values of l ) is discussed further. In the usual way we obtain the thermodynamic properties from the partition function: A = −NkB T ln[2 cosh(ε /kB T )]. χ= 3. The graph of the magnetisation is also easily explained now. U = A + T S = −N ε tanh(β ε ) M=− ∂A ∂H T where µB is the Bohr magneton: the coupling constant between the spin and the external ﬁeld. U . this shows that for low temperatures nearly all spins are in line with the ﬁeld. ε = µB H . as could be expected. Also. so that the entropy is low. The fact that the term in brackets can simply be raised to the N -th power is a result of the fact that the spins do not interact mutually. there are therefore two possible values of the energy for these two orientations – we call these energies ε and −ε . In the next few ﬁgures we show the temperature dependence of S. M and CH . If we apply a magnetic ﬁeld H. and the entropy will approach a constant. The entropy vanishes for small temperature as it should.10 Thermodynamics of magnetic systems: negative temperature The case of paramagnetic s = 1/2 spins is the easiest example of a quantum magnetic system. the energy per spin is about −ε which is in agreement with this picture. Finally we have CH = ∂U ∂T = NkB (β ε )2 / cosh2 (β ε ). In that case. more and more spins ﬂip over and the entropy and energy increase. This law is found in nature for systems with high values of the angular momentum quantum number l . There will be a particularly strong increase in the entropy near kB T = ε as in that region the thermal energy is sufﬁciently strong for ﬂipping the spins over. The speciﬁc heat shows a maximum near kB T = ε for the reason just explained. In the book.20 therefore has the form C T where C is the so-called Curie constant.

the temperature is negative.4 0. In that case.4 U Nε -0. a strong magnetic ﬁeld is suddenly reversed in order to bring the spins in a conﬁguration where the majority is antiparallel to the ﬁeld. In order to reach equilibrium. .2 -0. This is not in contradiction with the laws of thermodynamics. the entropy will decrease with energy. as the system is far from equilibrium. In this technique.6 -0. the temperature will return to positive values.2: Entropy versus temperature 0 -0. which is reached when all spins would be antiparallel to the ﬁeld.8 -1 0 1 2 3 4 5 6 kB T /ε Figure 3. and it therefore has to pass through absolute zero. The system is therefore extremely cold for some time.3: Energy versus temperature A striking feature of the energy graph is that it does not approach its maximum value. In fact.5 0. when the energy is positive.3 0.21 0.6 0.7 0.2 0. This can be used in an experimental technique called magnetic cooling. as the entropy decreases with energy and 1/T = ∂ S/∂ E .1 0 0 1 2 3 4 5 6 S kB N kB T /ε Figure 3.

5 0.4: Magnetisation versus temperature 0.4 0.22 1 0.3 C kB N 0.1 0 0 1 2 3 4 5 6 kB T /ε Figure 3.6 M N µB 0.2 0 0 1 2 3 4 5 6 kB T /ε Figure 3.5: Speciﬁc heat versus temperature .4 0.8 0.2 0.

but also particles with the remainder of the large system (the remainder is again called a bath). which can exchange not only energy.s ∝ eS(E −Es.s ∝ Ω(E − Es . isolated system in which we deﬁne a subsystem.N −Nr )/kB ∝ e−β Es +β µ Nr . 23 . N − Nr ). ∂N T we obtain Pr. N V Figure 4. Now we consider a state s of the subsystem consisting of Nr particles and an energy Es .1 Equilibrium between a system and a particle-energy reservoir We derive the grand canonical distribution function (density function) in a way analogous to that of the canonical ensemble. E.1: The grand canonical ensemble. we note that the probability of occurrence of this state is proportional to the number of possible states of the bath: Pr.4 The grand canonical ensemble 4. Just as in the derivation of the canonical ensemble. Writing Ω = exp(S/kB ) and realising that ∂S 1 = ∂E T ∂S µ =− . The system under consideration (dashed square) can exchange energy and particles with its surroundings. We consider again a large.

The required normalisation factor is N =0 ∑ eβ µ N ∑ Ese−β E s ∞ s ≡Z. leading to the distribution pr (N ) = e−β Er (N )+β µ N . we note that the grand partition function can be written as Z = N =0 ∑ eβ µ N QN (N . The quantity Z is called the grand canonical or grand partition function. N r N r N r Taking the derivative with respect to pr (N ) leads to −kB ln pr (N ) − kB − λ − kB β Er (N ) + kBβ µ N = 0. The denominator in the last expression is called the grand canonical partition function Z = ∑ ∑ e−β Er (N )+β µ N . We do this by requiring that the expectation values of E and N are given. ∑N ∑r e−β Er (N )+β µ N as found in the previous section. T ) is the canonical partition function.2 Formal derivation of the grand canonical ensemble Using the principle of maximum entropy.3 Physical signiﬁcance of the various statistical quantities We can relate the thermodynamic quantities using the grand canonical distribution function. First of all. T ) ∞ where QN (N . V. V. 4. This then leads to a Lagrange function F = S − λ ∑ ∑ pr (N ) − kBβ ∑ ∑ pr (N )Er (N ) − kB β µ ∑ N ∑ pr (N ). .24 We see that the probability distribution is that of the canonical ensemble multiplied by an extra factor exp(β µ N ) and summed over N . which is related to the Helmholtz free energy A as QN = e−A/kB T . N r 4. Hence we must maximise the entropy S = −kB ∑ ∑ pr (N ) ln pr (N ) N r under the condition that ∑ ∑ pr (N )Er (N ) = N r E =U is given and that ∑ N ∑ pr (N ) = N r N . we can again derive the probability for the grand canonical ensemble.

T ) = . Using the Euler relation from thermodynamics. which has the form √ 3N VN 2π mkB T QN (N . In this way we ﬁnd kB T ln Z = µ N − A = µ N − U + T S.T µ . U = ST − PV + µ N . and the result obtained turns out to be correct. V.4 Examples We ﬁrst calculate the grand canonical partition function of the ideal gas. The energy can be obtained as U = kB T 2 ∂ q(z.V Note that in the derivative with respect to T . The parameter z is called the fugacity. the fugacity z = exp(µ /kB T ) is kept constant (though it depends on T ). which can be ﬁxed by noting that for N = 0. Let us now calculate the average value of N using the density function: N= β µ N e−A/kB T ∑∞ N =0 Ne = kB T ∞ ∑N =0 eβ µ N e−A/kB T ∂ q(µ . V. N! h .T µ . this only leads to an additive constant in the relation between µ N − A and kB T ln Z . The relations with thermodynamic quantities can most easily be formulated as N= P= S= ∂ kB T ln Z ∂µ ∂ kB T ln Z ∂V ∂ kB T ln Z ∂T V. z.V 4. T ) ∂µ . However. the right hand side should vanish. V.T Instead of the chemical potential µ . often the parameter z = exp(β µ ) is used. so that we may replace the sum by the summand at its peak value. Note that we have been a bit sloppy in replacing the sum over N by its maximum value – we should have included a width here. We start from the canonical partition function. ∞ Just as in the case of the energy in the canonical ensemble. the summand eβ (µ N −A) will be very sharply peaked near the equilibrium value N of N . we ﬁnd kB T ln Z = PV ≡ kB T q.25 The grand canoncial partition function can thus be written as Z = N =0 ∑ eβ (µ N−A) . T ) ∂T . V.

N= S = −NkB ln z − kB ln [1 − zφ (T )] + 1 . From this expression for the grand canonical partition function. dT The ﬁrst two of these relations can be combined into the well-known equation of state PV = NkB T . consisting of atoms vibrating around their eqilibrium position. This leads to the following form of the partition function: QN (N . This has two important implications: ﬁrst of all. Writing Q1 (T ) ≡ φ (T ). V. and. 1 − zφ (T ) . Interestingly. such as a system consisting of indistinguishable harmonic oscillators. From the partition sum the thermodynamic quantities can again be derived: zφ (T ) . this relation does not depend on Λ. T ) = (Q1 (T ))N . secondly. they are distinguishable. 1 − zφ (T ) zkB T φ ′ (T ) . In a solid. the partition function of one such oscillator does not depend on the volume. so it holds for other.26 Now Z = with N =0 ∑ ∞ eβ µ N V N (2π mkB T )3N /2 ∞ ξ N ∑ = exp(ξ ) h3N N ! N =0 N ! 2π mkB T h2 3/2 ξ =Vz ≡ V zΛ−3 . zkB T Λ3 zV N= 3 Λ P= U = zV kB T 2 d Λ−3 dT S = −NkB ln z + zV kB T d Λ−3 + Λ−3 . uncoupled systems too. 1 − zφ (T ) A = NkBT ln z + kB T ln [1 − zφ (T )] . this leads straightfordly to Z = ∑ zN [φ (T )]N = N We see that zφ (T ) must be smaller than 1 in order for this sum to converge. The quantity Λ = h2 /(2π mkB T ) is called the de Broglie wavelength – it depends on T only. the thermodynamic quantities can easily be evaluated using the relations given in the previous section. 1 − zφ (T ) zkB T 2 φ ′ (T ) U= . the oscillators are localised.

The equilibrium is achieved for a gas density Ng 1 = 3 . we have φ (T ) = [2 sinh(ℏω /2kB T )]−3 . As a result. Vg Λ φ (T ) For low vapour density and high enough temperature. we therefore ﬁnd Ng 1 kB T . Vg √ with Λ the de Broglie wavelength h/ 2π mkB T . φ (T ) = ∑ e−β ℏω (n+1/2) = n For classical harmonic oscillators we have. h2 We see that two parameters enter this equation: the energy difference ε and the frequency ω . the atom is bound to the solid. and we need a certain amount of energy to remove it from there and move it to the gas. which you should forget as soon as possible). For 3D harmonic oscillators. we see that N ≈ 1 − 1/N zφ (T ) = 1+N for large N . Solid and vapour are in equilibrium when their chemical potentials are equal. From the second of these equations. For quantum harmonic oscillators. We then arrive at an expression for the vapour pressure 2π mkB T 3/2 [2 sinh(ℏω /2kB T )]3 e−β ε . which we describe a system composed of many independent oscillators. S/(NkB ) = ln φ (T ) + T φ ′ (T )/φ (T ). on the other hand. For the gas. (16) of Pathria. we must include a factor exp(β ε ) in the product φ (T )Λ3 . we have 1 e−β ℏω /2 = . We have however not taken into account the fact that the energy at the equilibrium point of the harmonic oscillator describing an atom is lower than the energy of a gas atom: after all. P = kB T = 3 Vg Λ φ (T ) which follows immediately from the ideal gas equation of state. we have zg = Ng Λ3 . as the grand partition function is independent of the volume (see Eq. We now use these results in order to analyse the solid-vapour euqilibrium. This renders the other relations a bit simpler: U /N = kB T 2 φ ′ (T )/φ (T ).27 Note that calculating the pressure for this system is nonsensical. P = kB T . φ = (β ℏω )−1 . These two parameters precisely determine shape and offset of the parabola’s deﬁning the energy felt by an atom in the solid. For the solid. we have zs = 1/φ (T ). β ℏ ω − 2 sinh(β ℏω /2 1−e A/N = −kB T ln φ (T ).

Now we shall consider statistical mechanics more strictly in the context of quantum mechanics. 28 .e. no explicit time dependence). Just as in the classical case. together with the probability for the system to be in any of those states. we note that in equilibrium ˙ must vanish. ρ i where pi is the normalised probability for the system to be in state |ψi (∑i pi ≡ 1). ρ ] . ℏ This is called the quantum Liouville theorem. but only the set of possible states which the system can be in. and inferred from that theorem that in equilibrium the density function depends on the Hamiltonian only: ρ (q. The analog of the density function now becomes the density operator. we shall leave the hats from operators unless confusion may arise. sometimes we needed to take quantum mechanics into account in order to have well-behaved partition functions. we have mainly considered classical statistical mechanics. ˆρ ˆ −ρ ˆH =H From now on. p)]. Remember however that the density functions we have considered so far were essentially classical: we have derived Liouville’s theorem from the classical (Hamilton) equations of motion. Of course. where ‘well-behaved’ means in this context that entropy and (free) energies scale linearly with N . In case we have a stationary Hamiltonian (i.5 Formulation of Quantum Statistics Up to now. This operator can be useful when we do not know the actual state of the system. The density operator is then ˆ = ∑ pi |ψi ψi | . We see that we have an equation quite analogous to Liouville’s theorem: ˙= ρ i [H . we see that iℏ ˆ ∂ρ = iℏ ∑ pi ∂t i ∂ |ψ ˆ |ψ =H ∂t ∂ |ψ | ˆ = ψ| H ∂t ∂ |ψi ∂t ψi | + |ψi ∂ ψi | ∂t = ˆ |ψi H ∑ pi i ψi | − |ψi ˆ ψi | H ˆ. we have ρ ρ = ρ (H ). From the time-dependent Schr¨ odinger equation iℏ and its Hermitian conjugate −iℏ which hold for any state |ψ . and also that integrals over phase space are dimensionless. p) = ρ [H (q.

. . it is evaluated as Tr A = ∑ φn |A| φn . for large particle numbers. . the physical wavefunctions have coordinates r1 . . spin degrees of freedom might be included. it does not matter which representation we choose because. Tr is the trace operator. the density operator of a quantum system is given as ρ = δ [H − EI ] where I is the unit operator. Then we can express ρ with respect to this basis: ρnm = φn |ρ | φm . Suppose we have an orthonormal basis set φn which forms a basis in our Hilbert space. even continuum) for a single particle. . such a state is entangled. and zero for energies above E . xN ). Also.1 Statistics of the various ensembles Just as in the classical case. E + ∆E ). but instead. and the trace boils down to adding the diagonal elements of the matrix representation of the operator being traced. In a ﬁnite-dimensional Hilbert space. In general. The density matrix then reads ρnm = an (t )a∗ m (t ). . A general state of the system is a linear combination of these basis states. n we have ρ = |ψ ψ | .. count the states in a narrow interval (E . . Then a complete set of states for a system consisting of N particles is ψn1 . Now suppose we have a complete set of basis states φn (x) (n might assume an inﬁnite number of values.. where xi is supposed to include all degrees of freedom of a single particle. . 5. also use the theta-function which is constant for all energies smaller than E .29 We recall here that for any operator G. In a many-particle system. . we do not rigorously implement a delta-function. n In a ﬁnite-dimensional Hilbert space. . instead of using the delta-function. In practice. . . φnN (xN ). . the expectation value is easily evaluated as G = Tr ρ G . the basis is ﬁnite. As mentioned in the ﬁrst chapter. We can. . rN . we therefore speak of the density matrix rather than an operator. Tr ρ Here. In case we have a state |ψ which can be expressed in this basis as |ψ = ∑ an (t ) |φn .nN (x1 . . xN ) = φn1 (x1 )φn2 (x2 ) . For any orthonormal basis |φn . This is the matrix representation.. the dominant contributions to the entropy come from energies very close to E in the latter representation. The wavefunction in general can therefore be written as ψ (x1 ..

= ρ= −β µB B β B µ σ β µ B − β µ B − B z B B 0 e +e Tr e e . we have a density operator ρ = Ce−β H . the density matrix becomes diagonal: ρnn = 1/Γ 0 for En < E . From this. ˆ If we express this operator with respect to an energy-basis. We consider only the interaction of the magnetic moment with the magnetic ﬁeld. The ﬁrst example is that of an electron in a magnetic ﬁeld. In the canonical ensemble. The grand canonical ensemble is formulated using the particle number operator n ˆ in addition to the Hamiltonian: ˆ ˆ . we calculate properties for some systems we have considered before in the classical context. an orthonormal basis of eigenfunctions of the Hamiltonian with eigenvalues En : ρmn = Ce−β En δnm . and not the orbital degrees of freedom (i.30 The entropy is given as S = kB ln Γ.e. for En ≥ E . we have H = −µB (σ · B). where Γ is the number of states with energy in a narrow band (E . E + ∆E ). Then we can use the fact that the exponential of a diagonal operator is again a diagonal operator with the exponentials of its eigenvalues on the diagonal: 1 e−β BµB σz eβ µB B 0 . perhaps in some potential). V.1 Electron in a magnetic ﬁeld In order to practice the quantum formulation a bit. T ) = ∑ eβ µ N ∑ e−β Es = ∑ eβ µ N QN (T ).2 Examples 5. We work in the representation in which σz is diagonal. the particle number operator commutes with the Hamiltonian.2. the motion of the electron. N s N 5. In a basis of eigenstates of the Hamiltonian. The calculation is most conveniently done in the canonical ensemble. ρ = Ce−β H +β µ n In most cases. The grand canonical partition function is then found again as 1/C = Z (µ . the normalisation is easily found as 1/C = Tr e−β H = ∑ e−β En = QN (T ). that is. ˆ n just as in the classical case. Considering only a single spin.

nz ). That the transition from the sum to the integral requires an extra factor L3 /(2π )3 can be seen as follows. First we choose as a basis the eigenfunctions which we denote as |k : k e−β H k′ = e−ℏ so that the partition function becomes QT = Tr e−β H = ∑ k e−H /(kB T ) k′ = ∑ e−β ℏ k k 2 k2 /(2m) 2 k2 /(2mk T ) B δ (k − k′). ny .31 Then we obtain for the average expectation value of σz : σz = Tr (ρσz ) = eβ µB B − e−β µB B = tanh(β µB B). governed by the Hamiltonian H =∑ i p2 i . eβ µB B + e−β µB B A comparison with sections 3. the volume occupied by a k-point is (2π /L)3 . d 3 k) divided by the volume per point. . L with k = 2π /L(nx . On the grid of k-values. It is instructive to derive the same partition function using the r-representation: r e−β H r′ = ′ 2 2 1 eik·(r −r) e−ℏ k /(2mkB T ) L3 ∑ k 1 ′ 2 2 ≈ d 3 k eik·(r −r) e−ℏ k /(2mkB T ) 3 (2π ) = m 2πβ ℏ2 3/2 exp − m r − r′ 2β ℏ2 2 .10 shows that these results are correct. 2m We must choose a basis of the Hilbert space in order to evaluate the trace.2 Free particle in a box We now consider a free particle in a box.e. 2m Inside the cubic box of size L in which the particles move.2. The eigenfunctions which are compliant with these boundary conditions are 1 3/2 i(kx x+ky y+kz z) ψ (r) = e . outside. The sum runs over the points in a certain volume. The corresponding energies are E (k) = ℏk2 . 5. This is then equal to that volume (i. the potential is zero. ≈ d 3 k e−β ℏ 2 k2 /2m L3 (2π )3 =V m 2πβ ℏ2 3/2 . we assume periodic boundary conditions.9 and 3.

equipartition is satisﬁed. Let us evaluate the expectation value of the energy. .3 Systems composed of indistinguishable particles To ﬁx the ideas. . n=1 N where Hn un = εn un . The coordinates of the particles are denoted by q = (q1 . 2m 2 that is. 5. the partition function can be evaluated as 3/2 m −β H 3 r e Q1 (T ) = r d r =V 2πβ ℏ2 which is obviously the same as the one found above. We have seen in the case of the quantum harmonic oscillator (see exercises and the next section) that the equipartition theorem does no longer < hold for kB T ∼ ℏω . Then. ). We could also use H =− ∂ ln Tr (e−β H ) ∂β which leads to the same result. T = Tr e−β (T +V ) The contributions from exp(−β V ) in numerator and denominator do no longer cancel. we start by considering the noninteracting case. q2 . Therefore. we must evaluate Tr Te−β (T +V ) . Because we have periodic boundary conditions. the Hamiltonian has the form H= n=1 ∑ Hn N where Hn only acts on the coordinates of particle n. this must not depend on r. as a result of the fact that T and V do not commute. where qn denotes the coordinate(s) of particle n. Using this. the equipartition theorem no longer holds in the quantum case. . as we have found. The eigenstates of the Hamiltonian then have the form ψE (q) = ∏ un (qn ). the expression r |ρ | r′ gives the probability that a particle suddenly moves from r to r′ as a result of a thermal ﬂuctuation. One might ask how general the equipartition theorem is. This is most easily evaluated in the krepresentation: H = Tr (He−β H ) 1 V = β H − Q1 (2π )3 ) Tr (e ℏ2 k2 −ℏ2 k2 /(2mkB T ) 3 3 e d k = kB T . On the other hand. . In order to check whether this theorem still holds for the kinetic energy only. The quantity r |ρ | r which occurs in these expressions (remember ρ = exp(−β H )) represents the probability density of ﬁnding the particle at position r.32 The Fourier integral will be discussed in the exercises. .

33 The total energy is given as E=
n=1

∑ εn .

N

Note that n denotes a particle, not a particular energy level. Now suppose that the particles are identical – in that case, the form of the Hamiltonians, and their spectra should be identical. Now suppose that we have N particles with energy E . As each of the particles occupues energies of the same spectrum, there might be more than one particle in the one state with energy εi . We must have: N = ∑ ni , E = ∑ εi .
i i

The states can then be written as

ψE (q) =

m= 1

∏ u1(qm ) ∏ u2(qm ) . . .
m= 1

n1

n2

Now, if the particles are identical, we know that a permutation of them leaves the total Hamiltonian invariant. If this is the case, the Hamiltonian commutes with the permutation operator: PH = HP. If an operator commutes with the Hamiltonian, it must be possible to construct the eigenstates of the Hamiltonian in such a way, that they are simultaneously eigenstates of that operator. You might recall from your mathematics course that any permutation can be written as a product of particle exchanges (a particle exchange means that we exchange a particle pair, i, j, say). Let us call Pi, j a particle exchange for the pair i, j. We obviously have Pi2 , j = 1. Then also the eigenvalues λ of Pi, j 2 should satisfy λ = 1. As we furthermore know that, since Pi, j is Hermitian, λ is real, we must have λ = ±1. We see that the particle wavefunctions are either symmetric under particle exchange (λ = 1) of antisymmetric (λ = −1). It turns out that for a particular kind of particles, we have either one or the other possiblity. Particles whose wavefunction is symmetric with respect to exchange, are called bosons; those which have antisymmetric wavefunctions are called fermions. The fact that any permutation operator can be written as a product of exchanges, leads to the conclusion that always PψE = ±ψE .

This notion directly leads to the conclusion that the microstates are invariant with respect to any permutation of the particles. Therefore, the numbers ni , which tell us how many particles can be found in state i, deﬁne a microstate uniquely, and additional correction factors for proper counting should not be included in sums over the states. Even if the particles interact, we can use the same representation (although the ui are no longer eigenstates of single-particle Hamiltonians). The reason why interaction does not matter is that the products of single particle states form a basis of the Hilbert space for many particles, whether they interact or not. Finally we note that, for fermions, we can construct wavefunctions constructed from singleparticle states ui as follows: ui (q1 ) ui (q2 ) u j (q1 ) u j (q2 ) 1 ψ (q) = √ . . . . N! . . ul (q1 ) ul (q2 ) . . . ui (qN ) . . . u j (qN ) . .. . . . . . . ul (qN )

34 where the vertical bars |. . .| denote a determinant. This wavefunction is called a Slater determinant. The prefactor follows automatically from the normalisation condition (the ui are considered to be normalised). In the case of Bosons, we have a similar expression, but the minus-signs in evaluating the determinant all turn into plus-signs. Another way of writing the wavefunction is

ψ (q1 , . . . , qN ) = ∑ δ [P] Pui (q1 )u j (q2 ) · · · ul (qN ),
P

where ∑P denotes a sum over all possible permutations, δ = 1 for bosons and −1 for fermions and [P] is the sign of the permutation. The sign of the permutation is determined as the number of exchange operations it is composed of. Note that the permutation operator acts on the arguments qn of the wavefunctions only, not on the labels i, j, . . .. Note that this state is not normalised in the case of bosons: ψ |ψ = n1 !n2 ! · · · where n1 etcetera are the occupation numbers of the different states. To see that this factor occurs indeed in the normalisation, look at a system consisting of two bosons, both in the same state u: √ 1 ψ = √ [u(q1 )u(q2 ) + u(q2 )u(q1 )] = 2u(q1 )u(q2 ). N! We see that the norm of this state is 2! = 2. For fermions we do not have this problem, as no two particles can be in the same state.

5.4

The density matrix and the partition function of a system of free particles

We know the wavefunctions already for a single particle (see section 5.3): uk (q) = 1 ik·r e . L3/2

With the above deﬁnition of a many-particle basis function, we must therefore evaluate
′ k1 , . . . , kN |exp(−β H )| k′ 1 , . . . kN ′ where the states |k′ 1 , . . . kN are (anti)symmetrised states. In order to evaluate this equation we note the following:

• On the left hand side and on the right hand side, we have actually a sum over all permutations. • For a particular permutation for different ki on the left hand side, the operator exp(−β H ) in the middle forces the states on the right hand side to correspond in a one-to-one fashion to those on the left hand side. √ • The normalising prefactors 1/ N ! on the left and right hand side yield a factor 1/N !. Combining all these considerations, we see that
′ −ℏ k1 , . . . , kN |exp(−β H )| k′ 1 , . . . kN = e
2 (k2 +···+k2 )/(2mk T ) B N 1

∏ δ (ki − k′i)
i=1

N

35 where the normalisation factor n1 !n2 ! . . . of the Slater determinant has already be divided out in order to work with normalised states. This factor amounts to 1 if all ki are different. When taking the trace, we must sum over a state k1 , . . . , kN . Note that this sum must include a restriction such that for the set of permutations of the quantum numbers k1 , . . . , kN only one member is chosen, otherwise we are overcounting. But we can relax any such restriction provided we divide by the number of possible permutations N! . n1 ! · n2 ! · · · Therefore we have as the ﬁnal result QN = 1 L3N N ! (2π )3
N

exp −ℏ2 k2 /(2mkB T ) d 3 k

=

1 N!

V Λ3

N

.

We see that we have obtained the correct expression of the partition function. Note that we have been sloppy in this calculation: if all k would be different, the result would be correct. But by integrating over all possile combinations k1 , . . . , kN , we have included the cases where two k’s are the same. If there is such overlap, the 3N -dimensional integral no longer factorises into N three-dimensional integrals. It turns out that the correction turns out to be very small when V 1/3 ≫ Λ: then for the vast majority of conﬁgurations, no two k’s overlap. Obviously, we could have evaluated this partition function using the coordinate (r) basis rather than the k-basis. The book uses this representation, but the calculation is more difﬁcult. We shall work out here the example of a two-particle system. We ﬁrst evaluate r1 , r2 e−β H r1 , r2 = 1 L6 2 (2π )6 {1 ± cos[(k1 − k2 ) · (r1 − r2 )]} e−β ℏ
2 (k2 +k2 )/(2m) 1 2

d 3 k1 d 3 k2 .

In this expression, the ﬁrst term is the one where the particles are both on the left and right hand side of the matrix element in the same state (both either k1 or k2 ), and the second term corresponds to different states on the left and right hand side of the matrix element – the + sign is for bosons, the − for fermions. Evaluating the integrals we obtain r1 , r2 e−β H r1 , r2 = 1 1 ± exp(−2π (r12 /Λ)2 ) . 2Λ6
2

Taking the trace means that we must integrate over r1 and r2 : Q2 = 1 2λ 6 1 + exp(−2π (r12 /Λ)2 ) d 3 r1 d 3 r2 = 1 2 V Λ3 1± 1 23/2 Λ3 V .

We see that the result reduces to the correct partition function for the case where Λ ≪ V 1/3 = L. The probability density for the particles to be at r1 and r2 can be considered for the boson- end fermi case. We deﬁne the statistical potential vs (r) as e−β vs (r12 ) = ρ (r1 , r2 ). The statistical potential is shown in ﬁgure 5.1. We see that the potential has the correct behaviour in the sense that it prevents fermi particles to come close, and favours bose particles to come close. It is instructive to repeat the calculation for the two-particle partition function in terms of the k basis. For bosons, we have a possible state for each combination k1 , k2 . In this case, for k1 = k2 the wavefunction is: 1 ψ = √ (ψk1 (1)ψk2 (2) + ψk1 (2)ψk2 (1)) . 2

we have a possible state only when k1 = k2 : 1 ψ = √ (ψk1 (1)ψk2 (2) − ψk1 (2)ψk2 (1)) . 2 If we calculate the partition function. we must integrate over k1 and k2 . In general. but we see that we must single out the contributions for k1 = k2 . this results in Tr e −β H 1 = 2 V (2π )3 2 2π mkB T ℏ2 3 1 V ± 2 (2π )3 π mkB T ℏ2 1 2 3/2 = 2 V (2π )3 2π mkB T ℏ2 3 1± 1 Λ3 .1: Statistical two-particle potential for bose (solid line) and fermi (dashed line) systems.5 0 -0.8 1 1. k)d 3 k. for an operator A which is diagonal in phase space we have 2 V 1 V 1 A(k1 . Tr A = 2 (2π )3 2 (2π )3 For the trace.2 0. √ Note that the two expressions differ by a factor 2.36 1 0.2 β vs (r) r /Λ Figure 5.5 -1 0 0. 23/2 V . k2 )d 3 k1 d 3 k2 ± A(k. the wavefunction is: ψ = ψk1 (1)ψk1 (2). For fermions.6 0.4 0. For k1 = k2 .

n2 . we sum over all possible conﬁgurations for distinguishable particles and then divide by N ! Therefore. we must have ∑ ni = N . The two are equivalent only if each state contains at most one particle. . If we have N particles. if we have a set of single-particle quantum states |i . A quantum state of a collection of identical particles is a fully antisymmetric (for fermions) or symmetric (bosons) many-particle state. each set (n1 . n1 ! · n2 ! · · · In any sum over all the eigenstates. The canonical partition function can be evaluated as Q(N . From this it follows that. n2 . . In Maxwell-Boltzmann counting. . . thereby rendering a full analytic solution feasible. .) = n1 ! · n2 ! · · · instead of the correct factor g = 1 which is taken into account in Bose-Einstein counting. .)e−β ∑ ε n ′ i i i 37 .6 The theory of simple gases In the last chapter we have laid the foundations for quantum statistical mechanics. which occurs at high enough temperature and large enouh volume. n2 . we can evaluate the energy of the system to be E = ∑ ni εi . .) into account is 1 g(n1 .) should be counted only once. In this chapter we shall work out further the case of non-interacting particles for which the partition functions usually factorise. Such a state can be constructed from single-particle states by a Slater determinant (in the case of fermions) or a symmetrised linear combination of products of single-particle states (bosons). 6. . T ) = {ni } ∑ g(n1 . n2 . i The number of single-particle products in such a state is N! . the many-particle state is speciﬁed by the numbers ni of particles in each such state. Let us summarise the most important results here. the effective weight with which we take the conﬁguration (n1 . .1 An ideal gas in other quantum-mechanical ensembles – occupation numbers Quantum states for ideal gases are fully characterised by specifying how many particles there are in each available state. . . i If the single-particle states are eigenstates of the single particle Hamiltonian with energies εi . .

. that in order for the grand canonical partition function to be well-deﬁned. Each of the factors is a geometric series which can be evaluated analytically: n=0 ∑ eβ (µ −ε )n = 1 − eβ (µ −ε ) ∞ 1 Note. T ) = n1 =0 ∑ eβ (µ −ε )n ∑ eβ (µ −ε )n 1 1 2 ∞ ∞ 2 n2 =0 ··· . where ε0 is the ground state energy. i ∂ β ( µ − εi ) +1 e . the factors are identiﬁed as the power series expansions of the exponential function: n=0 ∑ n! eβ (µ −ε )n = exp ∞ 1 eβ (µ −ε ) . For Bose-Einstein statistics. we obtain ni = ∂ 1 ln 1 + eβ (µ −εi ) = β (ε −µ ) . . . For FD statistics. n=0 ∑ eβ (µ −ε )n = 1 + eβ (µ −ε ). where the restriction ∑i ni = N does not come into play: Z (µ . we obtain: β ( µ −ε1 )n1 · · · ∞ ∑ni =0 ni eβ (µ −εi )ni · · · ∑∞ n =0 e . In order to proceed. the weight factor g = 1. with 1/(n1 ! · n2 ! · · · ). . we look at the grand canonical partition function. however.)eβ ∑ (µ −ε )n . i i i The nice property of this partition function is that it factorises into a product of sums over ni . . In the case of BE statistics: Z (µ . Note that for Bose-Einstein (BE) and Fermi-Dirac (FD) statistics. ∂ β ( µ − εi ) 1 − eβ (µ −εi ) e i −1 β ( µ −εi )ni ∑∞ ni =0 ni e β ( µ −εi )ni ∑∞ ni =0 e This is the famous Bose–Einstein distribution function. For Fermi-Dirac statistics. it is necessary that µ < ε0 . except for the i-th factor. . Because of this last restriction.) with ∑i ni = N . it is not easy to evaluate this partition function. T ) = {ni } ∑ g(n1 . which yields: ni = This can be evaluated as ni = 1 ∂ 1 ln = β (ε − µ ) . n2 . and that for Maxwell-Boltzman statistics g = 1/(n1 ! · n2 ! · · · ). It is also possible to evaluate the average occupations of the levels. ni = 1 β ( µ −ε1 )n1 · · · ∞ ∑∞ ∑ni =0 eβ (µ −εi )ni · · · n1 =0 e All factors in the numerator and the denominator are identical. 1 For Maxwell-Boltzmann counting. the situation is even simpler: each of the ni only assumes the values 0 or 1. n2 .38 where ∑′ {ni } denotes a sum over all conﬁgurations (n1 .

i The factor gi is the multiplicity (degeneracy) of the state i. V ) = N! where the single-molecule partition function has the form: Q(1. To obtain this expression. V )]N Q(N . In the usual way. Finally. The centre of mass coordinates of the molecules yield the free. These can include electronic or nuclear spin. so that Boltzmann counting is justiﬁed.5 1 FD 0. We do not have to include the counting factor 1/n1 ! since n1 ≤ 1 in the regime considered.39 3 2. T .1 we show the different distribution functions. it is necessary to split up the degrees of freedom: we consider the centre of mass coordinates separately from the internal degrees of freedom. which is justiﬁed in the gas phase when their mutual separations are on average very large. T .5 2 MB BE ni 1. molecular partition function j(T ): j(T ) = ∑ gi e−εi /(kB T ) . In ﬁgure 6. . and vibrational or rotational motions of the nuclei.1: Bose–Einstein. for Maxwell-Boltzmann counting. We neglect the interaction between different molecules.5 0 -2 -1 0 1 2 3 (ε − µ )/kBT Figure 6. we have. we may factorise the partition function into partition functions of the individual molecules: 1 [Q(1. whereas the internal degrees of freedom generate the internal. 6. not really surprisingly: ni = ∂ ln exp eβ (µ −εi ) ∂ β ( µ − εi ) = eβ (µ −εi ) . Fermi–Dirac and Maxwell–Boltzmann distribution functions. ideal gas partition function. T . We furthermore suppose that the thermal wavelength is much smaller than the system size. V ) = V 2π mkB T h2 3/2 j(T ).2 Examples: gaseous systems composed of molecules with internal motion Consider a gas consisting of molecules with internal degrees of freedom.

which is the value of the total angular momentum. ℏω . the corresponding frequency is high: in fact the distance between the vibrational levels. ∂T µ = µideal − kB T ln j. We ﬁrst consider the vibrations along the axis connecting the nuclei. 2 where ∂ ln j(T ). Calling the energy levels εJ . 2 dT Using PV = NkB T . If on the other hand. The spin does not inﬂuence the energy (if we neglect the (hyper)ﬁne structure) and only the chemical potential and entropy are affected by the spin. ∂V V Note that j(T ) does not contribute to the pressure. = NkB + 2 dT N . We have j = 2S + 1. From quantum mechanics. as en example. Let us. ∂T Also the speciﬁc heat at constant volume can be evaluated as Eint = NkB T 2 3 dEint CV = NkB + . we have j = ∑(2J + 1)e−β εJ . ∂ A NkB T = . we know that the ﬁne structure energy correction is determined by the value of the quantum number J . Other quantities which can be evaluated are the chemical potential and the entropy: S = Sideal + NkB ln j + T ∂ ln j . as does not depend on the volume (which is natural. J Diatomic molecules consist of two atoms. which means that these vibrations can only be seen for temperatures of that order.40 For these systems.P where we have again used the fact that the internal degrees of freedom do not depend on V . We see that j always inﬂuences the values of these two quantities. We then have additional degrees of freedom: two related to rotations and one to vibrations. The energy can be evaluated as P=− 3 E = NkB T + Eint . whereas those of the energy and of the speciﬁc heat are determined only by the temperature-dependence of j. we obtain for the speciﬁc heat at constant pressure: CP = ∂ (E + PV ) ∂T dEint 5 . is of the order of 103 K. It turned out that for T ≫ ℏω the system satisﬁes equipartition. As the atomic bonds are relatively stiff. the electron has orbital angular momentum in addition to its spin. leading to . consider a monatomic gas for which the electrons or the nuclei have spin S. then there is ﬁne structure splitting. since j(T ) includes only internal degrees of freedom). We have already evaluated the partition function of the harmonic oscillator.

we see that for T < θr . 2IkB This temperature is in general much lower than the vibrational temperature. If we disregard ﬁne structure. Again we can deﬁne the temperature where rotations become important. θv . for T ≫ θr . θv . the sum can be replaced by an integration: jr (high T ) ≈ (2l + 1) exp [−θr l (l + 1)/T ] dl = T . only the ﬁrst few terms will contribute signiﬁcantly to the partition function: jr (low T ) ≈ 1 + 3e−2θr /T + 5e−6θr /T + · · · . The full behaviour is given by eβ ℏω CV = NkB (β ℏω ) β ℏω . θv . ﬁrst the rotational freedom will affect the speciﬁc heat. for θr ≪ kB T ≪ θv . we have 3 CV = NkB . 2 .41 a constant speciﬁc heat. the vibrational degrees of freedom will become noticeable. (e − 1)2 The temperature ℏω /kB is often denoted as θv : it is the temperature where the vibrations become noticeable in the speciﬁc heat. The partition function for the rotations is ∞ jr = ∑ (2l + 1) exp [−θr l (l + 1)/T ] . • For low temperatures. and that the speciﬁc heat for small temperatures decays to zero. θr This results in a contribution NkB to the speciﬁc heat. This is θr = ℏ2 . l =0 For low temperatures. In summary we can say that • At all temperatures. • For even higher temperatures. for T ≪ θr . 2 7 CV = NkB . ﬁne structure effects may become noticeable in the speciﬁc heat. we see the effect of spin degeneracy in the entropy and the chemical potential. The quantum mechanical energies for a rigid rotator are given as El = ℏ2 l (l + 1) 2I where I is the moment of inertia perpendicular to the molecular axis. • For high enough temperatures. 2 5 CV = NkB . For high temperatures. A diatomic molecule can be considered as a ‘rigid’ rotator if we neglect the coupling between vibrations and rotations.

For Bosons (minus-sign) the pressure becomes smaller. which indicates that the particles seem to repel each other as a result of the Pauli principle. 2 +β µ where the + sign is for fermions. Bauer). strongly negative µ corresponds to high temperatures. 7. after expanding to second order in exp(β µ ): P 1 = 3 3/2 kB T λ π d 3 x e−x eβ µ ∓ e2β µ 2 ≪ 1. this leads to nλ 3 = eβ µ ∓ 1 23/2 e2β µ + · · · . this d 3 x e−2x 2 = 1 1 eβ µ ∓ 5/2 e2β µ . Therefore. we restrict ourselves here to a brief review of the major applications. where the + sign corresponds to Fermi. we have the expression 1 1 P = ln Z = ± 3 3/2 kB T V λ π d 3 x ln 1 ± e−x 2 +β µ . we can write 1 Nλ 3 = 3/2 V π d3x 1 ex2 −β µ ±1 . For the pressure. This equation shows that when we keep the density ﬁxed. In the classical limit. and the − sign for bosons. the second term gives the quantum correction. and the − sign to Bose statistics. 2π mkB T In the classical limit for which µ is strongly negative. Furthermore λ=√ h . indicating an effective attraction. Examples have been covered extensively in the statistical physics course of the third year (G. 3 λ 2 If we now substitute eβ µ by the expansion obtained above. e−x yields.7 Examples of quantum statistics Quantum statistics involves either Bose–Einstein or Fermi–Dirac counting in the evaluation of physical quantities. For Fermions we have the plus-sign. 42 .1 Thermodynamics of free quantum gases As we have seen in the exercises. we see that P = nkB T 1 ± 2−5/2 nλ 3 The ﬁrst term is the classical results.

it follows that the photons are created and annihilated freely.1 Planck distribution Take an empty box in contact with a reservoir of temperature T . the factor of 2 arises from the fact that there are two transverse modes (only transverse modes are allowed by Maxwell’s equations). and we have divided the volume of the spherical shell with thickness dk in k-space by the volume (2π /L)3 of each k-point. The photons carry energy ℏω . If we take a cube of size L × L × L. the k vectors are 2π /L(nx . 8 2 7. ny . . The reservoir can interact with the box by emitting electromagnetic (EM) ﬁeld waves into it. In quantum language we say that photons can travel into the box.V =V ∂P ∂T µ . λ 2 T 3/2 eβ µ Combining this with the classical relation eβ µ = nλ 3 gives S = NkB which. Therefore. so that there number cannot be controlled by a chemical potential. c3 π 2 In this expression. Which k vectors are accessible is determined by the shape of the box.2 Bose-Einstein systems 7.43 Finally we can derive the entropy: S=− ∂ kB T ln Z ∂T µ . the number of modes available at frequency ω is given by ω 2dω . and used ω = ck. 2 h2 /(2π mkB T ) and E /N = 3kB T /2 can be written in the form: S = NkB 3 E 4π m V 3 ln + ln + ln + 5/2 . Now we use the fact that the photons are bosons (they are spin-1 particles) and we have for the energy radiated at frequencies between ω and ω + d ω : N (k)dk = 2 · 4π k2 dkL3 /(2π )3 = V u(ω )d ω = n(ω )ℏω d ω = ℏ ω 3dω . Now we copy the lowest order term in the expansion found above for the pressure P: S =V m ∂ (kB T )5/2 ∂T 2π ℏ2 V 5 µ = 3 eβ µ kB − .V . so µ = 0. The result is 1 S = SClass ± √ λ 3 NkB .2. The quantum corrections can be evaluated analogous to the case of the pressure. Creating a new photon in particular does not involve any cost except for its energy. multiplied by the energy of the mode and the occupancy. π 2 c3 eℏω /kB T − 1 where the BE distribution is recognised. From the quantum theory of the elactromagnetic ﬁeld. nz ) with Ni integer. with ω = c|k|. 2 N N 2 3h2 This is again the Sackur–Tetrode formula for the entropy of an ideal gas. for λ = 5 + ln(λ 3 /n) .

Let us. the energy corresponding to the longest wavelength). for such an acceptable µ .61. the analysis fails. This fraction corresponds to nnormal λ 3 = 2. In fact. evaluate the number of particles as a function of µ : N =∑ k 1 eβ (ℏ2 k2 /(2m)−µ ) − 1 ≈ V (2π )3 1 eβ (ℏ2 k2 /(2m)−µ ) − 1 4π k2 dk. this transition is not justiﬁed if ε0 − µ really approaches the value 0. Note that the integral depends on β µ only. and the only way out is by questioning the transition from the sum over k to an integral. which now is macroscopically occupied (that is. . What happens is that the gas splits up into two parts.e.5 0 βµ Figure 7. The occupation of the ground state is V nG = 1 eβ (εG −µ ) − 1 . This imposes a temperature-dependent maximum on the particle density – beyond this value. a ﬁnite fraction of the particles is in the ground state). We therefore know that nG = n − nnormal particles per unit volume will be in the ground state.5 0 -3 -2. In function 7.5 -1 -0.2 Bose–Einstein condensation The BE distribution is well deﬁned only for values of µ below the ground state energy – otherwise. the rest of the particles occupied the ground state. Reparametrising ℏ2 k2 /(2mkB T ). In that case we must split off the term corresponding to the ground state energy. For β µ → 0 this function approaches the value 2. where ε0 is the ground state energy (i.61.2. The normal part ﬁlls the energy levels according to the BE distribution in the usual way.5 1 0. If the total particle density is greater than prescribed by this limit.1.44 3 2. we plot the function g(β µ ) as a function of β µ . we obtain for the particle density 4 nλ 3 = √ π ∞ 0 x2 ex2 −β µ − 1 dx ≡ g(β µ ). 7.5 2 g(β µ ) 1.5 -2 -1. the occupancy becomes negative.1: The Bose–Einstein function g.

+ ∑ ℏω /(k T ) i B 2 −1 i e The ﬁrst term on the right hand side is the energy-offset. µ = 0 and the pressure remains constant. Now we keep the temperature constant.5 1 1. µ will vary and assume negative values.8 0.2.2 0 0 0. We now can deduce that εG − µ ≈ 1/(β V nG ) i.4 0. and the excitations can be described in terms of a collection of independent harmonic oscillators (see the classical mechanics course). and let the density vary. For densities lower than the critical density. the larger the system. any system can be described in terms of harmonic interactions.e. As we have seen in section 7. .3 Phonons and the speciﬁc heat Phonons are lattice vibrations. The critical density corresponds to 1/2. the pressure is given by P=− kB T 4 √ λ3 π dx x2 ln(1 − e−x 2 +β µ ). i.5 2 2. as the denominator in the formulae for the density is now replaced by the logarithm. Figure 7.5 λ 3 /n Figure 7.6 P 0. 7.1. The energy for such a system can easily be found: we simply add up the expectation value of the energies of the oscillators at frequencies ωi and at temperature T : U (T ) = Φ0 + ∑ i ℏωi ℏωi .2: Pressure versus λ 3 /n. a system of particles connected by harmonic springs. even when µ is close to ε .. the second is the zero-point energy of the harmonic oscillator. and the rightmost term is the average energy due to the energy quanta nℏωi .61 on the horizontal axis. the closer the chemical potential will be to the ground state.45 1 0.e. and they can be understood by realising that the system of interacting atomic nuclei can be approximated by a harmonic system. For densities higher than the critical value. The x = 0 term no longer contributes.2 shows the pressure as a function of the inverse density. Close to the conﬁguration of minimum potential energy.

The chemical potential at T = 0 is called the Fermi energy. The parameter θD is called the Debye temperature. at which the oscillators would vibrate.1 Degenerate Fermi gas The name ‘degenerate Fermi gas’ is used for a dense system consisting of noninteracting Fermions. For low temperatures. the square shape gets rounded of. but positive T . ∂T (e − 1)2 where x = ℏωE /(kB T ) = θE /(kB T ). Dense means that nλ 3 = eβ µ is larger than 1. the distribution function has a square shape. 7. Using the ω 2 distribution. where x0 = ℏωD /(kB T ) = θD /(kB T ). This leads directly to x2 ex ∂U CV (T ) = = 3NkB x . The parameter θE is called the Einstein temperature. In the ground state.46 We can evaluate the sums by transforming them into integrals. all the one-particle levels are ﬁlled for energies smaller than µ . The Debye function D(x) is deﬁned as 3 x0 x4 ex D(x0 ) = 3 dx. the Debye result for the speciﬁc heat is CV (T ) = C(T /θD )3 . however. the Einstein result does not match the experimental results very well. For particles . This distribution is however cut off as there cannot be more modes than particles: ωD 0 VCω 2 d ω = N . ωE . For high temperatures. the chemical potential µ is positive. At higher temperatures. The value of the proportionality constant depends on the sound speeds for transverse and longitudinal waves. as shown in ﬁgure 7. but Einstein avoided this in 1907 by requiring that there was approximately only one frequency. In that case. x0 0 so that we ﬁnd again CV → 3NkB as it should be for high temperature. εF . which is occupied for T = 0. For T = 0. and for small.3 Fermions 7.3. it approaches the classical result 3NkB (why?). The speciﬁc heat is found as CV (T ) = 3NkB D(x0 ).3. we can perform a Taylor expansion of the integrand in the expression for D which yields 3 x0 2 D(x0 ) ≈ 3 x dx = 1. x0 0 (ex − 1)2 For low temperatures. Peter Debye took the actual distribution of modes which we have already encountered above for the photons.

2 0 0 0. We conclude that 2 ℏ2 (3nπ 2 )2/3 ℏ2 kF εF = = . To show that the chemical potential for a degenerate electron gas deviates only to order (kB T /εF )2 is not easy. If we do so. These excited electrons come from a band of width ≈ kB T below the Fermi energy. In a particular kind of stars. we can easily evaluate the speciﬁc heat of the degenerate electron gas. The number of k-points which lie in the range k.4 1. D(ε ) = N εF 2 .2 0.6 0.4 0. It is convenient to count the number of states at a speciﬁc energy – this is called the density of states. What happens is that some electrons with ε < εF are excited to ε > εF .3: The Fermi distribution function for T = 0 and small. The radius of this sphere is as usual related to the particle number: N =2 L3 4π 3 k .4 0. Degenerate Fermi gases are quite familiar: in a metal.6 0.2 1 0. k + ∆k is given as D(ε )∆ε = 2 therefore D(ε )∆ε = which can also be written as The total energy is now given by E = d ε D(ε )ε f (ε .47 1. (2π )3 3/2 √ ε ∆ε . T ). 3 −3/2 √ ε.8 T= 0 n 0. 2m 2m From ﬁgure 7. moving in a cubic box with side L. much larger than room temperature. the particles ﬁll up a sphere in k-space. since E = ℏ2 k2 /(2m).2 E Figure 7. We shall take this for granted here.8 kT T> 0 1 1. and it takes care of the double spin-degeneracy. the electrons have a Fermi temeperature of 107 K.3 it is clear that for positive T . We speak of a degenerate Fermi gas when kB T ≪ εF . (2π )3 3 F The factor 2 is for the particular case of electrons. and they occupy states in a band kB T above the Fermi energy. but positive T . the chemical potential will remain more or less constant. V 2π 2 2m ℏ2 L3 4π k2 ∆k. the valence electrons have a fermi energy corresponding to about 50000 K. the so-called ‘white dwarfs’.

except in a band kB T around εF . the electronic contribution dominates. well above the maximum contribution of the electrons. we obtain: CV = 1 ∂ E = ∂T kB T 2 ∞ 0 d ε D(ε ) f (ε . The lower boundary εF /(kB T ) of the integral is large but negative – we replace it by −∞. T ). This allows us to take D(εF ) out of the integral. T ) for the Fermi-Dirac distribution function and we will stick to this convention from now on. which starts by expressing N as an integral over the function ∆(ε ) which is an integral of the density of states over the energy: ∆(ε ) = Then we can write. T ) d ε = − ∞ 0 ∆(ε ) f (ε . in particular the fact that µ was replaced by εF . Taking the temperature derivative of this expectation value yields the speciﬁc heat. d ε D(ε ) e(ε −εF )/(2kB T ) + e−(ε −εF)/(2kB T ) ε − εF 2 Note that the integrand is small everywhere. we have 2 CV = kB T D(εF ) ∞ −εF /(kB T ) x2 ex/2 + e−x/2 2 dx. If we compare this with phonons. 3 (ex/2 + e−x/2 )2 we have π2 2 k T D(εF ). Using ∞ ∞ π2 x2 dx = . This growth stops only at the Fermi temperature. 3 B Substituting the value D(εF ) = 3/2N /(kB T ) we obtain: CV = CV = T π2 NkB 2 TF where the Fermi temperature TF is deﬁned by εF = kB TF . grows as T 3 . dε . as we have seen.48 We have used f (ε . For very high temperatures (higher than TF ). a few things have been wiped under the carpet. the phonon contribution saturates at 3NkB . Changing to the integration variable x = (ε − εF )/(kB T ). If we take the derivative with respect to T and assume that µ is approximately temperature-independent (this is not quite correct. T ). see below). which. using partial integration: ∞ ∞ 0 D(ε ′ )d ε ′ . We now present a correct calculation. We see that the speciﬁc heats grows linearly with T . whereas for temperatures well above the Debye. we see that for low temperature. Using the fact that N= we can write E = N εF + d ε D(ε )(ε − εF ) f (ε . lies fairly high. where the speciﬁc heat due to the phonons. Although this calculation yields the correct result. T ) dε . N= 0 D(ε ) f (ε . the speciﬁc heat saturates at 3NkB /2.

cV = ε ∂T ∂T ∂T We have ∂f ε − µ + T µ′ 1 . the rightmost term is much smaller than the ﬁrst two.49 The partial integration has the advantage that the energy derivative of f is nonzero only in a small interval around µ so that we can expand D around µ : N= ∆(µ ) + (ε − µ )D(µ ) + 1 1 (ε − µ )2 ′ D (µ ) 2 kB T e(ε −µ )/(2kB T ) + e−(ε −µ )/(2kB T ) 2 dε . no reference to the explicit form of D(ε ) has been made. dT 3 6 For low temperatures. kB T + D(µ )kB 3 3 Using the relation obtained above from the vanishing temperature-derivative of the particle number N then yields: 2 cV = D(µ )kB T π 2 /3. where two arise from the ﬁrst integral. we obtain two terms: π2 (kB T )2 . 6 where in the ﬁrst term we could integrate directly as the integrand is proportional to the energyderivative of f . and in the second integral we have used the same result as was used in the simpliﬁed derivation. and the third one from the second integral: 2 π2 2 2 π T . . This implies that its derivative with respect to the temperature should vanish: N = ∆( µ ) + D′ ( µ ) π2 2 π2 dN = D(µ )µ ′ + D′ (µ ) kB T + D′′ (µ )µ ′ (kB T )2 . . if we want to calculate the speciﬁc heat at constant density. the number of particles should be ﬁxed. along the lines of our simple derivation above: ∂f ∂f ∂f d ε = µ D(ε ) d ε + D(ε )(ε − µ ) dε . . . 3 B We now perform a similar analysis for the speciﬁc heat. (ε − µ ) 2 kB T e(ε −µ )/(2kB T ) + e−(ε −µ )/(2kB T ) 2 dε Carefully analysing these integrals gives three dominant terms. We have used the prime ′ to indicate a derivative with respect to T . cV = µ µ ′ D(µ ) + µ D′ (µ ) Substituting the explicit expression for the density of states yields the result obtained above. Again selecting only the even integrands. = ∂ε kB T 2 e(ε −µ )/(2kB T ) + e−(ε −µ )/(2kBT ) 2 D( µ ) µ ′ + D′ ( µ ) Substituting this into the two integrals appearing in the expression for cV . . we obtain: cV = µ D(µ ) + (ε − µ )D′ (µ ) + . 1 ε − µ + T µ′ dε + kB T 2 e(ε −µ )/(2kB T ) + e−(ε −µ )/(2kB T ) 2 ε − µ + T µ′ 1 D(µ ) + (ε − µ )D′ (µ ) + . so that we have π2 2 k T = 0. Note that in this derivation. The main observation now is to realize that.

f± are the Dirac functions for the appropriate spin state. all relevant physical quantities can be determined. We do this by calculating the total number of particles and then require that this is constant: N = N+ + N− = ∑ [ f+ (i) − f− (i)] i where the sum is over the orbital states i. Then. the only thing which changes is that the energies are shifted over ±µ ∗ B. Now we assume that the ﬁeld B is very small (i. The sum over the orbital states can again be replaced by an integral over the energy: M= µ∗ 2 d ε [D(ε ) f (εi − µ ∗ B) − D(ε ) f (εi + µ ∗ B)] . we obtain: M = µ ∗2B d ε D(ε ) f ′ (ε ).3. 2 Substituting this back into the integral expression for N . ∞ 0 Substituting the same Taylor expansion for the distribution functions as above. First we analyse how the chemical potential changes with the magnetic ﬁeld. A sum over the orbital states can however be replaced by an integral over the energy if we insert the density of states: 1 d ε [D(ε ) f (ε − µ ∗ B) + D(ε ) f (ε + µ ∗ B)] .e. If the magnetic ﬁeld is switched on. Note that µ ∗ is the magnetic moment – it should not be confused with chemical potential which is µ without the asterisk. N= 2 The integral is over the orbital energies.and down spins. The magnetisation is given as the difference between the number of spin-up and -down electrons: M = µ ∗ (N+ − N− ) = µ ∗ ∑ [ f+ (i) − f− (i)] i where the sum is over the orbital states i – f± are the Dirac functions for the appropriate spin state. that is.9 and 10). Here we shall consider the full quantum description for spin-1/2 fermions – you may think of electrons in a solid. . We are interested in how the system reacts to an applied ﬁeld. We see that to ﬁrst order in B the density does not change if we keep the chemical potential constant. and the magnetic ﬁeld only enters in the Fermi distribution functions. The factor 1/2 in front of the integral takes into account that D(ε ) includes up. we have N= d ε D(ε ) f (ε ) + O (B2 ) + . The important issue now is that this density of states does not depend on the spin degrees of freedom (this is a direct consequence of the fact that B = 0).50 7. Once we know this function. Then we can expand the distribution functions about B = 0: f (ε ± µ ∗ B) ≈ f (ε ) ± µ ∗ B f ′ (ε ) + (µ ∗ B)2 ′′ f (ε ). hence we conclude that µ varies with B only to second order. all the properties of the electrons are determined by the density of states D(ε ) (see the previous section). . µ ∗ B smaller than kB T ). .2 Pauli paramagnetism In chapter 3 we have already considered paramagnetism (section 3. we want to calculate the magnetisation as a function of the ﬁeld strength B. where the sign depends on the spin. . Suppose we have no magnetic ﬁeld.

If we apply a magnetic ﬁeld in the zdirection. For the free electron gas. dB We see that measuring the susceptibility at low temperatures tells us what the density of states near the Fermi level is. we obtain M ≈ µ ∗ 2 BD(εF ). the Fermi function decays rapidly from 1 to 0 near the Fermi energy. For evaluating particle numbers and magnetisations. but also as a result of their orbit. This is most easily done by ﬁrst evaluating the grand partition function: Z = ∏ 1 + eβ (µ −εi ) . the susceptibility can be expanded in powers of kB T . i We then have. The spectrum is given by p2 eℏB ( j + 1/2) + z .and y degrees of freedom.3 Landau diamagnetism Electrons moving in a solid have a magnetic moment not only as a result of their spin. we need to know the density of states.51 For small T .3. pz ) = m 2m This problem has been treated in the exercise class of your quantum course (believe it or not). . The magnetic susceptibility χ tells us how the magnetisation varies with the ﬁeld: χ= dM = µ ∗ 2 D(εF ). in other words the multiplicity of these levels. h We can then evaluate the number of particles and the magnetic moment. ε ( j. In addition. with z = exp(β µ ): ln Z = ∑ ln 1 + ze−β εi . This holds in particular for the x and y degrees of freedom. i The index i denotes the states. as we can simply perform an integral over pz when summing over all states. hence f ′ (ε ) ≈ −δ (ε − εF ). the particles will have quantized ﬁeld levels associated with the x. they have an energy associated with their motion in the z-direction. the density of states was found in the previous section – we ﬁnd for the susceptibility in this case 3 µ ∗2 χ= . 2 εF For higher temperature. 1− 2 εF 12 εF 7. the result is χ≈ π 2 kB T 3 µ ∗2 + ··· . which for our particular problem are deﬁned by pz and j: ln Z = ∞ p2 eℏB Lz eB d pz ∑ Lx Lz ln 1 + z exp −β ( j + 1/2) + z 2π ℏ h m 2m j =0 . This is called the orbital magnetic moment. It can be argued that the multiplicity for the energy levels associated with the orbital motion in the xy plane is given by D( j ) = Lx Ly eB . Using this.

(2π mkB T )1/2 2 h 2 sinh [eℏB/(2mkB T )] Now the desired quantities can easily be evaluated. where z ≪ 1. The result we have obtained is similar to that of 3. . Expanding the argument of the logarithm. With λ = h/(2π mkB T ). we have zV x ∂Z N=z = 3 ∂z λ sinh x and 1 ∂ zV x cosh x 1 . x = β Beh/(4π m) and µeff = eh/(4π m). except for a minus sign. This means that the magnetisation is now opposite to the ﬁeld – this effect is called diamagnetism.9.52 This partition function can be evaluated in the classical limit. we get ln Z = zVeB h2 e−β pz /(2m) d pz ∑ e−β eℏB( j+1/2)/m = 2 ∞ j =0 zVeB 1 . where L is the Langevin function L(x) = coth x − 1/x. M= ln Z = 3 µeff − β ∂B λ sinh x sinh2 x We can write M = −N µeff L(x).

. σ r 12 − σ r 6 .8 Statistical mechanics of interacting systems: the method of cluster expansions Up to this point we have considered ideal gases only. . These were sometimes derived from interacting systems. . we have seen that the partition function always factorises into an integral over the momenta. . however. ∞ 0 for r < a. i.. . In this chapter.1 Cluster expansion for a classical gas In the analysis of the classical gas.e. which are written as V (r1 . rN ) = ∑ u(|ri − r j |). qN ) ∑ 2m N p2 d 3N p d 3N q. In general. Relevant interactions are those which are described in terms of pair-interactions. . . i< j Important examples of pair interactions are the hard sphere interaction: u(r) = and the Lennard-Jones interaction uLJ (r) = 4ε This interaction is shown in ﬁgure 8. crystal lattices) which could be transformed to a system of independent oscillators. we consider a method for evaluating the correct equation of state. The interaction then plays a relevant role. for r ≥ a. involving the kinetic energy.1. we cannot transform away the interaction like we have done in these harmonic systems. i=1 53 . 8. such as systems with harmonic interactions (e.g. which for a classical ideal gas reads: PV = α NkB T with α = 3/2 for a noninteracting system in 3 dimensions and α = 3 for a system of uncoupled harmonic oscillators. and an integral over the orbital coordinates: QN (T ) = 1 h3N N ! exp −β i + V (q1 . .

If we write out the product occurring in the integral for Z .k. i< j We shall use the notation f (ri j ) ≡ fi j from now on. the factors exp[−β u(r)] do not decay to zero – they will tend to 1 for large separation since the interaction then vanishes. V ) = exp −β ∑ u(ri j ) d 3N r. The integral over the momenta can be performed analytically (it is a product of elementary Gaussian integrals): 1 QN (T ) = ZN (T . we immediately see that the conﬁguration integral can be written as Z (N . Figure 8. then between triplets.5 3 r /σ Figure 8. V.54 2 1.5 -1 0 0. So. we obtain ∏ [1 + f i j ] = 1 + ∑ f i j + ∑ i< j i< j ′ fi j fkl + i. Therefore we introduce the Mayer functions f .1: The Lennard-Jones potential. .5 1 VLJ /ε 0.5 0 -0.5 2 2. T ) = d 3N r e−β ∑i< j u(ri j ) = d 3N r ∏ e−β u(r i< j ij) = d 3N r ∏ [1 + f (ri j )] . and larger and larger clusters.2 shows the Mayer function for the Lennard-Jones potential. But if we want to neglect contributions beyond a certain cluster size. Using the Mayer function.l i< j<k ∑ fi j f jk fki + · · · . we must have some expressions which vanish rapidly beyond some interaction range. how does it work? The idea behind the cluster expansion is to include ﬁrst only the interactions between particle pairs. which indeed decay to 0 for large r. V ) N !λ 3N with ZN (T . i< j The cluster expansion is a way to systematically evaluate the so-called conﬁguration integral ZN .5 1 1. Clearly. deﬁned as f (r) = exp[−β u(r)] − 1. j.

The sum with the prime ∑′ is over all possible pairs i. . Note that the actual expansion consists of a sum over all possible conﬁgurations. 1 2 3 1 2 3 f 25 4 5 6 4 f25 f 16 5 6 1 2 3 1 2 3 f25 f26 4 5 6 4 5 6 Figure 8. The second order term contains .3: Some cluster conﬁgurations.5 3 f ij (r) VLJ (r) r Figure 8. j. j. In ﬁgure 8. the other two are the second order terms. l which are distinct. a sum over all possible triplets. Note however that we still include conﬁgurations of the form i.5 1 0. and a sum over all ‘distinct pairs of distinct particle pairs’. . k.5 0 -0. .5 2 2.5 1 1. . The top left cluster corresponds to the lowest order term.2: The Lennard-Jones interaction potential and the Mayer function. It is clear that the ﬁrst order term contains N (N − 1)/2 pairs. j and k. This means that we have a sum over all distinct pairs in the ﬁrst order term.55 2 1.5 -1 0 0.3 we indicate the possibilities.

and for the lower right part we have N (N − 1)(N − 2)/6.1 we list the ﬁrst ﬁve clusters. we do not have to impose additional conditions on the integration over r12 .3. . . We call n j the number of vertices in a cluster of type j. . V 2V 2 V2 Generally. . In order to avoid double counting of these pairs (corresponding to interchanging pair 1 and pair 2) we must include one more factor 1/2. which leads to the required result. This number is obtained as follows. In table 8. Obviously. d 3 rN .56 N (N − 1)(N − 2)(N − 3)/8 terms of the form corresponding to the upper right part of ﬁgure 8. The number of ways in which . which yields an additional factor V . Higher terms in the expansion are considered similarly. the expansion is built up as follows. As the interaction is short-ranged. For the ﬁrst term we have N (N − 1)/2 possibilities. the integral over all coordinates except r1 and r2 can be performed directly. and the volume is large. We rewrite the integral over r1 and r2 as one over r1 and r12 = r2 − r1 . Inserting the ﬁrst two terms of the product expansion in the expression for the conﬁguration integral and integrating over the coordinates. . d 3 rN = V N −1 f (r12 ) d 3 r12 . and we obtain f (r12 ) d 3 r1 d 3 r2 . after some calculation that the expansion for the conﬁguration integral can be written as Z = VN 1+ N (N − 1)(N − 2)(N − 3) 2 N (N − 1)(N − 2) N (N − 1) b2 + b2 + b3 + . and for each of these possiblities we have (N − 2)(N − 3)/2 for the second pair. Obviously. . two connected points have j = 2 etcetera. 2+ 3! We then see. For the terms corresponding to the conﬁguration in the lower left part of ﬁgure 8. We label the different types of clusters by the index j. . b3 deﬁned above contains the contributions from diagrams 3a and 3b. Now we deﬁne 1 d 3 r f (r) b2 = 2 and 1 b3 = 2b2 d 3 r1 d 3 r2 f (r1 ) f (r2 ) f (r12 ). We work out the term f (r12 ) d 3 r1 d 3 r2 . The integral over r1 can be performed over r1 . .3. d 3 rN = V N −2 f (r12 ) d 3 r1 d 3 r2 . we obtain Z = V N + N (N − 1)V N −1 V N −2 N (N − 1)(N − 2) 2 N (N − 1)(N − 2)(N − 3) d 3 r1 d 3 r2 f (r1 ) f (r2 )+ 8 N (N − 1)(N − 2) d 3 r1 d 3 r2 f (r1 ) f (r2 )+ V N −2 d 3 r1 d 3 r2 f (r1 ) f (r2 ) f (r12 )+ · · · 6 1 2! d 3 r f (r) + V N −2 The prefactors (the powers of V ) arise from the integrations over the particles not present in the clusters and from the integration over one of the coordinates of each independent cluster itself. noting that d 3 r1 d 3 r2 f (r1 ) f (r2 ) = 4b2 2. so we have f (r12 ) d 3 r1 d 3 r2 . we have N (N − 1)(N − 2)/2 possibilities. A single point (vertex) has label j = 1.

the cluster integral b j is deﬁned as b j (T ) = 1 n j !V vertex permutations ∑ d 3 r1 · · · d 3 rn j ∏ fi j . j Graph b j (T ) 1 1 2 d 3 r f (r) 3a 1 2 d 3 r1 d 3 r2 f (r1 ) f (r2 ) 3b 1 3! d 3 r1 d 3 r2 f (r1 ) f (r2 ) f (r12 ) 4 1 8 d 3 r1 d 3 r2 d 3 r3 f (r1 ) f (r2 ) f (r3 ) f (|r1 + r2 − r3 |) we can divide N particles into m1 clusters of type 1. V. j Now let’s go back to the expansion containing only the terms to order V N −2 . so that we have A = −kB T ln 2π mkB T h2 3N /2 1 Z (N . N! . their labelling and the corresponding cluster integrals. For each type of cluster. i. T ) . etcetera is given by the combinatorial factor N! ∏ j m j !(n j !)m j Obviously.57 Table 8.1: Different types of clusters. N = ∑ j m j n j . If we calculate the free energy. we ﬁrst have a contribution arising from the momenta. m2 clusters of type 2.

. This is a speciﬁc example of a feature which occurs at all levels: the diagrams remaining in the expansion do not contain any lines by which the diagram can be cut into two disjoint pieces. . 1 This expansion should not be read as an expansion for small arguments. The remaing diagrams are called star diagrams or irreducible diagrams. + 2 V V 2 2 V2 N (N − 1) b2 N (N − 1)(N − 2) = A0 − kB T b2 − 2 (2N 3 − 5N 2 + 3N ) + b3 + . . 2 V V V2 If we now use the fact that N − 1 etcetera can be replaced by N if N is large. • The second order term b3 − 2b2 2 corresponds to the triangle term 1 3! d 3 r1 d 3 r2 f (r1 ) f (r2 ) f (r12 ). we have A = A0 − kB T b2 N (N − 1)(N − 2)(N − 3) N 2 (N − 1)2 N (N − 1)(N − 2) N (N − 1) b2 + 2 − b3 + . For a general proof of this fact. Using P=− we directly obtain 2 P = kB T n 1 − b2 n − 2(b3 − 2b2 2 )n + . . The aim is merely to group terms with a particular power of n = N /V . we obtain A = −kB T ln 2π mkB T h2 3N /2 VN − N! kB T ln 1 + N (N − 1)(N − 2)(N − 3) 2 N (N − 1)(N − 2) N (N − 1) b2 + b2 + b3 . ∂A ∂V = kB T n 1 + a2 n + a3 n2 + . but we shall refrain from this step for simplicity. . V 2V 2 V2 Performing a Taylor expansion for the logarithm1 .58 Including the remaining terms. we can write this expansion in terms of the density n = N /V : A = A0 − kB T N nb2 + n2 (b3 − 2b2 2) + . . . . A better way to perform the expansion is to move to the grand canonical ensemble. . you should consult the book of Mayer and Mayer. . . . but as a formal expansion in terms which could perhaps be relatively large. . . the equation of state can be determined. Now that we have the partition function. This cancellation must happen at all orders of the expansion in order to guarantee that the free energy scales linearly with N (it is an extensive variable). Two important remarks are to be made at this stage: • The coefﬁcient of b2 2 is proportional to N as a result of a cancellation of two terms proportional to N 2 .

The ﬁrst theory which explained the phenomenon of phase transitions was the Van der Waals theory from 1873. On average. The ideal gas equation of state reads PV = NkB T . In practice. some fraction of this volume is excluded as the strong repulsive interaction for short particle separation prevents them from coming too close. After measuring the pressure accurately as a function of the density. are called virial coefﬁcients. the Van der Waals theory is based on a particular equation of state. the excluded volume is 1 (0 + 8V0 + 2 · 8V0 + · · · + (N − 1) · 8V0 ) ≈ 4NV0 . This suggests that we replace the volume by V − Nb. However. where V0 is the volume occupied by the ‘hard core’ of each particle. the volume-dependent part of the entropic term for the ideal gas S = kB N ln V. In fact. Also. only V − 2 · 8V0 is available etcetera. To guess the value of the parameter b. the parameters occurring in the potential can be ﬁtted in order to match the virial coefﬁcients for the model potential to the experimental results. In fact. For the third particle. The fact that the particles have less space to move in. the form of which can heuristically be motivated as follows. Now the volume occurring in this equation is the total volume of the system.2 The virial expansion and the Van der Waals equation of state In chapter 11 we shall concern ourselves with phase transitions. The ﬁrst particle does not experience the presence of other particles. which contains several parameters. in practice.59 The coefﬁcients a2 = −b2 = − 1 2 f (r)d 3 r 1 3 f1 f2 f12 dr1 dr2 a3 = −2b3 + 4b2 2=− etcetera. N We see that b ≈ 4V0 . the excluded volume is 8V0 . 8. we note that for a spherical hard core of volume V0 . The second particle however has only a volume V − 8V0 at its disposal. directly affects the entropic contribution to the free energy. in magnets phase transitions occur – there they are associated with a change in the magnetisation. These are transitions which take place when the control parameters are changed and which cause the system to move from a phase with certain values for thermodynamic properties to a phase where these properties are signiﬁcantly different. is now replaced by S = kB N ln(V − Nb). Now let’s put N particles in the volume V . an educated guess for the form of the potential is made. It can be argued that b ≈ 4V0 . . The virial coefﬁcients contain information concerning the potential. An example is the LennardJones potential which contains two parameters: ε and σ . Common example of phase transitions are the liquid-gas transition and the solid-liquid transition.

This number is in ﬁrst approximation proportional to the density. we have µ= Using the thermodynamic relation E − T S + PV . only one temperature can be found. When are we at the leftmost branch of the curve. the pressure is shown as a function of the volume per particle for several temperatures. The value of the parameter a can be guessed as follows. and correcting for the double counting of pairs. Using E = T S − PV + µ N . This term will for each particle be proportional to the number of particles within the range of the potential. Now suppose that we ﬁx the temperature at some value below Tc and lower the pressure. So if we add up this contributions for all particles.4. we note that the internal energy is reduced by the attractive part of the potential. In ﬁgure 8. V where E0 is the internal energy of the ideal gas and a is determined by the details of the attractive potential. and when do we cross over to the right hand? The answer to this question is given by a thermodynamical argument. For T below a treshold value Tc . but the other two do. The result is n u(r)d 3 r. It turns out that the middle value does not correspond to a thermodynamically stable phase. V from which the equation of state follows as P=− so P+a N V N ∂A = −a ∂V V 2 2 + NkBT 1 . N . V − Nb (V − Nb) = NkB T . N dE = T dS − PdV + µ dN . three values for the density are possible. We see that for large temperature. Take a particular particle and calculate its interaction with the particles nearby. we see that 1 u(r)d 3 r. a=− 2 All in all. we ﬁnd for the volume-dependent part of free energy A=− aN 2 − NkB T ln(V − Nb). we obtain dµ = V dP − SdT . For all particles we have N E = E0 − N a. for a particular value for the pressure.60 In addition to this effect.

Now let’s work out the ﬁrst virial coefﬁcient a2 for low temperatures: a2 = 1 2 (e−β u(r) − 1)d 3 r ≈ 4V0 + 1 2kB T u(r)d 3 r = b − a . . N If the phase on the left branch coexists with the phase on the right branch at a given pressure and temperature.2 000000000000 111111111111 111111 000000 000000000000 111111111111 000000 111111 000000 111111 T<Tc 0. the chemical potentials of these two phases should be equal. we see that a2 = b − a3 = b2 etcetera.4 must be zero.5 T>Tc T=Tc Pressure 0. we have dµ = 1 V dP.4: Isotherms for the Van der Waals gas. . which means that the hatched area in ﬁgure 8.3 0.3 0. Starting from P= NkB T N −a V − Nb V 2 = a NkB T 1+ b− V kB T N V + b2 N V 2 + b3 N V 3 + .2 0. .7 V /N Figure 8.1 0 0 0.4 0.61 0. As we keep the temperature constant. kB T a kB T We see that this matches precisely the expansion of the Van der Waals equation of state! This unfortunately does not hold for the higher orders... Now we compare the Van der Waals equation of state to the cluster expansion. Comparing this with the virial expansion for the pressure.4 0.6 0.5 0. . The Van der Waals expansion approximates the effect of the hard core by the series 1 + 4V0 n + 16V02 n2 + 64V03 n3 + .6 0.1 0. This means that we must have: d µ = V dP.

. . The Van der Waals equation takes part of all higher virial coefﬁcients into account.62 whereas the correct series from the cluster expansion yields 1 + 4V0 n + 10V02 n2 + 18.4V03 n3 + . . That is the reason why it can predict the phase transition behaviour rather well.

cannot lower its energy by interactions with the pipe. This implies that the momentum of the helium becomes p (before the exchange. Nevertheless. which normally have integer spin (spin=0). and that this state does not experience any friction with walls. This means that these atoms could form a Bose-Einstein condensate at low temperatures. A way to see this is to consider a ﬂow with speed v of helium thourgh a pipe of mass M . It is assumed that a ﬁnite fraction of particles occupies the same quantum state.17 K. a kind of BE condensation can also take place in the case of interacting systems. The phenomenon of superﬂuidity has been brought in connection with a BE condensate a long time ago. so that it moves without friction. and so far. Helium-4 consists of atoms. You may compare this with the description of electrons as independent particles in the solid: we consider the particles moving in a potential which is represents the average interaction potential of each particle with its counterparts. This means that helium which ﬂows through a pipe. Now suppose that the liquid and the pipe exchange momentum p. 2M . If we place ourselves in the rest frame of the liquid. The energy of the pipe with momentum P is given as E= P2 . Strictly speaking however. it becomes a ‘superﬂuid’. using v = P/M and letting M → ∞: ε (p) = v · p + 63 p2 ≈ v · p. which can either be determined by experiment (scattering experiments) or theoretically.1 The superﬂuidity of helium If liquid helium is cooled down to a temperature of 2.9 The method of quantized ﬁelds 9. the fact that helium forms a liquid at low temperatures. tells us that the interactions between the atoms become important. a Bose-Einstein (BE) condensate has been studied in the case of noninteracting particles. This curious phenomenon can be understood using techniques of statistical quantum ﬁeld theory. 2M 2M From this we have. We can now setup the energy balance between the helium and the pipe: ε (p) = P2 (P − p)2 − . the helium was at rest) and that that of the pipe is decreased by an amount p. 2M For the helium we assume a relation between energy and momentum given as ∆ + ε (p). we see the pipe moving with a speed v in the opposite direction. provided a description in terms of noninteracting particles is appropriate for the system at hand.

the critical velocity can be determined. only a ﬁnite fraction of the liquid is in the ground state. It is seen from ﬁgure 9. If the energy-dispersion curve is measured. It is in this context important to realise that if helium is superﬂuid as a result of BE condensation.1: Dispersion relation for helium.1 which is tangent to the energy-momentum curve. and the mass ﬂow is given by the quantum mechanical expression of the ﬂux: j(r) = ℏ ℏ [Ψ∗ (r)∇Ψ(r) − Ψ(r)∇Ψ∗ (r)] = a2 ∇γ (r). For such low velocities. much higher than the critical velocity which was measured directly in experiments. We can express this by splitting the total density in a normal and a superﬂuid fraction: ρ = ρs + ρn . we should have pv > ε ( p). it turns out to be 60 m/s. The fastest speed at which no energy transfer is possible corresponds to the straight line in ﬁgure 9. and the rest is in a normal state.1. Then. energy transfer is not possible which means that helium is superﬂuid. 2im m . It turns out that ε ( p) has a shape depicted in ﬁgure 9. Now suppose that the pipe would cause excitations of the helium. in order for the container to have available such an amount of energy.64 45 40 35 30 25 ε 20 15 10 5 0 0 1 2 3 4 5 6 p Figure 9. The superﬂuid fraction of the liquid consists of particles which are all in the same state Ψ(r) which can be written in the form: Ψ(r) = a(r)eiγ (r) . The superﬂuid number density is given by ρs (r) = a2 . The difference can be explained by considering rotational motion in the superﬂuid.1 that this is possible only in case the speed is large enough.

Therefore. we must assume that ∇ × v = 0 holds in a region of the plane without holes in it. For such a ﬂow we have ∇ × us = 0. just as the force is a gradient of a potential. The expectation that superﬂuid helium cannot rotate was checked by putting helium inside a rotating cylinder (‘rotating bucket experiment’). In order to prove the latter relation. m From this. we should have at each point r: v = ω × r. The function γ (r) plays the role of the potential. we have ∇γ · d l = 2π n. In experiments. we see that for the velocity we have us = ℏ ∇γ (r). where ω is the angular velocity.65 Given the fact that the ﬂow is density × velocity. causing the above relation to be modiﬁed to ρn ω 2 r 2 . a ﬂow derived from a potential as sketched above). in the case where the ﬂuid would rotate in a cylinder. (In the analogue of classical mechanics this equation expresses the fact that the work done along a path only depends on the start and end point of that path). For a normal ﬂuid. m We see that the ﬂow can be considered as a gradient of a function. but that the ‘core’ of this ‘vortex motion’ is not superﬂuid. . we read off two striking features: (i) the rotational velocity of a vortex is quantized and (ii) this velocity depends directly on Planck’s constant. Therefore we have u · dl = n integer. we may satisfy the condition ∇×v = 0 without the usual relation following from it: us · d l = 0. ∇ × v = 2ω . z(r) = ρ 2g the experiment showed that also the superﬂuid fraction participated in the rotation. We can circumvent this condition by assuming that the superﬂuid rotates around an axis. The explanation for this fact comes from the notion that the function γ (r) is not a usual potential: it is a phase and hence deﬁned modulo 2π . the meniscus assumes the shape z(r) = ω 2 r2 2g which can easily be checked by minimising the total energy of the ﬂuid as a function of z(r). When we follow a path surrounding the vortex core. although only the normal fraction was supposed to contribute to the rotation. nh . This last condition seems to prohibit circular ﬂows. We see that the second equation is certainly incompatible with a potential ﬂow ﬁeld (ie. For example.

cannot be justiﬁed within the approach in which we analyse the low energy states. in comparison to terms where two subscripts are zero. q = 0. we have N0 in the superﬂuid phase.k′ . we analyze the system in the case where the part of the helium which is in the superﬂuid phase much larger than the part which is the normal phase. an operator term containing b† k+q bk′ −q bk′ bk with k + q. We then are left with (∑′ k means that k = 0 is excluded): 2 2 2 ˆ = u0 N0 + ∑′ ℏ k b† bk + u0 N0 ∑′ 4b† bk + b† b† + bk b−k = H k −k k 2m k k k 2V 2V 2 u0 [N0 + 2N0 (N − N0 )] u0 N0 ′ ℏ2 k2 † ′ † † + b bk + ∑k 2b† k bk + bk b−k + bk b−k . • k = k′ = 0. for all q. k = 0. † ′ ′ In that case. the ﬁrst term is ﬁxed and can be neglected. we have Using the fact that N0 0 0 2 2 2 ˆ = u0 N + ∑′ ℏ k 2b† bk + u0 N0 ∑′ 2b† bk + b† b† + bk b−k . The correct treatment of the problem . • k′ = q = 0.q k+q k −q k 2m In order to make the Hamiltonian tractable. In fact. H k k −k k k k 2m 2V 2V For a ﬁxed total number of particles. The fact that half of the term 4bk b† k in the potential energy has been split off and approximated.2 The low-energy spectrum of helium In this section we want to derive the low-energy spectrum of helium from the microscopic Hamiltonian: 2 2 ˆ = ∑ ℏ k b† bk + 1 ∑ vq b† b† ′ bk′ bk . H k 2V k. and we consider the case where N − N0 ≪ N .66 9. for which vq = u0 . 2V 2V ∑k 2m k 2 + 2N (N − N )2 ≈ N 2 . Secondly. k − q. k and k being nonzero. and which scales as (N − N0 )). k′ = 0. • q = k′ = −k = 0. for N particles. • q = k′ . • k = q = 0. is negligible (it scales with (N − N0 )2 . • q = −k. This means that we must single out all possibilities where at least two b’s have subscript zero. we restrict ourselves to a delta-function potential. k′ = 0. k = 0. This is possible in seven different ways: • q = k = k′ = 0.

is: † bk = ηk cosh θk − η− k sinh θk . and a quadratic form can be put in diagonal form by a linear transformation. . The proper choice. we note that the Hamiltonian is essentially a quadratic expression of the b-operators. we see that the second term in the square root dominates. and are generally denoted as ‘quasi-particles’. Here. b−k = η−k cosh θk − ηk It is straightforward to check that the ηk satisfy the proper boson commutation relations: † = 1. In this case we must however take care that the linear transformations are chosen such that the operators still satisfy tractable commutation relations. To solve this problem. V 2m +2 For small k. = −∑k ′ ℏ2 k2 u0 N0 † † + cosh θk sinh θk ηk η− k + η− k ηk + 2m V u0 N0 ′ cosh2 θk + sinh2 θk V ∑k † † ηk η− k + η− k ηk . and we see that the energy is linear in k. † sinh θk . we are left with   2 2 2 2 2 2 2 2 2 2 u0 N0 ℏ k 2u0 N0 ℏ k ℏ k ℏ k u0 N0  ′ ′ † +2 + − − . we can immediately infer the energy eigenvalues with respect to the ground state: E (k) = ℏ2 k2 2m 2 Substituting this into the expression for the diagonal part. a there is only a single vibrating particle.67 should include the chemical potential. the theory does not see the difference between the energy quanta of this system and ‘particles’ of energy ℏω which can be created and destroyed by the ladder operators a± . H = ∑k + ∑k  ηk ηk 2m V 2m 2m V 2m V u0 N0 ℏ2 k2 . The simples example of a quasi-particle is the energy quantum of a harmonic oscillator. The question is now what the eigenvalues of this Hamiltonian are. A correct treatment of this is presented in the book of Fetter and Walecka. However. we obtain an off-diagonal contribution: Ho. From this. ηk .d. ηk If we substitute the new expression for the bk in terms of the ηk into the Hamiltonian. We see that these off-diagonal terms vanish when tanh 2θk = u0 N0 /V ℏ2 k 2 2m + u0 N0 /V . ﬁrst made by Bogoliubov. Note that the elementary excitations are linear combinations of the single-k modes which describe single-particle excitations: the elementary excitations describe excitation quanta. which precisely cancels this term.

ij i What will be the behaviour of the model? To answer this questions. In two dimensions. there may be an external magnetic ﬁeld present which favours all spins to be either + or −. and an instructive system for studying phase transitions more generally is the system in which the degrees of freedom reside on lattice sites. In the ﬁrst half of the twentieth century it has become clear that the reason for this nonanalytic behaviour lies in the fact that macroscopic objects (such as a glass of water) consist of large (almost inﬁnite) numbers of molecules: a function depending on a paremeter. and a sudden emergence of order in the case of freezing. we see such abrupt. We may consider the spins as magnetic moments which obviously interact – usually. the positive magnetisation will be favoured. These fascinating phenomena have for long time been poorly understood. It seems paradoxal that although we believe the world around us to be governed by smooth functions and differential equations leading to analytical solutions. it is necessary to always identify a parameter which characterises the degree and/or type of order present in the system. we start by examining the ground state. the density is the proper order parameter. In the case of freezing. If the magnetic ﬁeld h is zero. The model describing such a system is the famous Ising model. but the square lattice is quite popular): on each lattic point. these two phases have the same energy and the system will choose either one or the other. For a N × N 2 lattice. In order to analyse phase transitions. the system is formulated on a square lattice (in fact. This parameter is called the order parameter. the opposite sign is the stable phase. These considerations lead to the following Hamiltonian: H = −K ∑ si s j − h ∑ si . different choices for the lattice can be made. A phase transition is always characterised by a sudden change in the degree or the type of order in the system. There exist however other types of phase transitions than freezing and boiling. may become nonanalytic if the parameter becomes inﬁnite. The ﬁrst theory of phase transitions which gave excellent results was the Van der Waals theory. These transitions are associated with an abrupt change of the density in the case of boiling.1 About phase transitions The most common examples of phase transitions are the freezing and boiling of ordinary water.10 Introduction to phase transitions 10. the interaction is limited to nearest neighbours. it might be the structure factor. non-analytic behaviour. Furthermore. Now let us consider 68 . a spin-up or -down can be placed. whereas for negative ﬁeld. The ﬁrst interaction favours all spins to be equal: all spins + or all spins −. We have discussed this at the end of chapter 8 of the notes and it is recommended that you go back and study this theory once again. which we call + and −. which is analytic for every ﬁnite value of that parameter. For a positive ﬁeld. In the case of the boiling of water. which will be the phase at absolute zero. and can assume only two different values. there are thus 2N possible conﬁgurations.

This phenomenon is called symmetry breaking. ξ (T ) ∝ |T − Tc |−ν T < Tc . sometimes spontaneous symmetry breaking. It turns that then in 2D and on a square lattice.44 . but this turns out not to be the case. As an example. When the system passes the transition temperature.5 Figure 10. as it is not imposed by changing the model itself – it is a well-known phenomenon which is relevant in many areas of physics and astronomy. the symmetry between up. The Ising transition occurs in two dimensions at K /(kB T ) = 0. Which of the two will be the low-temperture phase is not known beforehand. the h = 0 case at nonzero temperatures. This case is interesting because the Hamiltonian yields the same value for any particular conﬁguration as for the one in which all spins are reversed.5 -1 0 0. the interactions between the spins become irrelevant. m ∼ |T − Tc |1/8 .1. The magnetisation assumes a value zero above a ﬁnite temperature. deﬁned as the average value of the spins. β → 0. m(T ) ∝ (−T + Tc )β . Physical quantities usually vary as broken algebraic power functions of the system parameters. will be zero. This behaviour is only one example of many similar ones. There exist other critical exponents for other physical quantities and/or parameters: χm = ∂m ∂h ch (T ) ∝ |T − Tc | T −α ∝ |T − Tc |−γ . The magnetisation is shown in ﬁgure 10.5 2 2.1: The magnetisation of the Ising model as a function of temperature for zero ﬁeld. The parameter β = 1/8 (do not confuse this β with 1/(kB T )!) is called critical exponent. This means that the average magnetisation. The behaviour close to the transition point is interesting. consider the variation of the magnetisation when the transition temperature is approached from below. .69 1 0. and their values will be completely random. Suppose the system starts at high temperature and is then cooled down.. At very high temperatures.and down is broken. One now might think that the magnetisation will decay with increasing temperature to reach zero at T → ∞. the phase in which all spins have either the value +1 or −1 must be chosen.5 1 k BT/K 1. . Once the system chooses one of the two values.5 magnetisation 0 -0.

continuous or second order. the system is in some state ρ . that is. nor for Ising models including farther than nearest neighbour interactions. no such solutions exist. This is one of the most important results obtained in theoretical physics of the 20th century. the spin is actually ﬂipped. For three dimensions. If this number is smaller than exp(−β ∆E ). these methods boil down to the following: in a computer. the Ising model in two dimensions on a square lattice was solved for zero ﬁeld analytically by Lars Onsager in 1944. The change in Pρ is due to the combined effect of the system leaving the state ρ at the next step to enter a state σ . (10. Also for the Ising model on a triangular lattice. although some progress has been made in recent years. Suppose this cost is ∆E . we know the values of the exponents from the exact solution: α = 0. The order refers to the derivative of the (free) energy which jumps or is continuous. If. γ = 7/4. The ﬁrst type of event results in a decrease of Pρ and the second one in a gain. Consider the probability Pρ (t ) that at some instance of time t . All in all we have Pρ (t + 1) − Pρ (t ) = ∑ −T (ρ → σ )Pρ + T (σ → ρ )Pσ σ . and which are called ﬁrst order transitions. If we consider however the magnetisation as a function of the external ﬁeld for ﬁxed temperature (which is taken below the phase transition temperature). the ﬂip is performed. ν = 1.1) For the case of the two-dimensional Ising model on a square lattice. moreover. for nonzero ﬁeld. in which the magnetisation is a continuous function of the parameter which is changed (the temperature in that case). but of a different kind than the one described above. If ∆E < 0. no analytic solution is possible. Many results concerning spin models can be obtained using Monte Carlo techniques which are performed on a computer. a spin is chosen at random.2 Methods for studying phase behaviour The results given in the previous section for the behaviour of the Ising model can be derived in various ways. We shall from now on focus on the second kind of phase transition. β = 1/8. except for inﬁnite temperature. The Monte Carlo algorithm leads to conﬁgurations occuring with a probability proportional to exp(−β E ) as required in the canonical ensemble. Tc ) = h1/δ . 10. as opposed to transitions in which the order parameter jumps discontinuously. δ = 15. However. else the spin is not ﬂipped. if there is an energy gain by ﬂipping the spin. on the other hand.2) For nonzero magnetic ﬁeld. then the spin is ﬂipped with probability P = exp(−β ∆E ). Then the energy cost or gain associated with ﬂipping that spin is calculated. This can be seen as follows. (10.70 and. A random number between 0 and 1 is chosen . an analytic solution was obtained by Houtappel in 1950. First of all. and entering the state ρ from any different state σ . Performing a spin ﬂip with this probability is done as follows. there is an energy cost associated with the spin ﬂip. This is also a phase transition. then we see a sudden ﬂip of the magnetisation which jumps from positive to negative or vice versa. we have an exponent for the behaviour of the magnetisation with varying small magnetic ﬁeld at the transition temperature: m(h. the spin ﬂip is carried out. In a nutshell. which is called critical. The phase transition no longer occurs for nonzero ﬁeld. the magnetisation will always be along the ﬁeld.

A method for obtaining analytic results is the mean ﬁeld approximation. 4 neighbours in 2D and 8 in 3D. j i K K − qNm2 − (h + Kqm) ∑ si − K ∑ (si − m)(s j − m) ≈ − qNm2 − (h + Kqm) ∑ si . but in practice. interacting with a ‘ﬁeld’ which incorporates the contribution from the average magnetisation. see ﬁgure 10. This can easily be done as the partition function has been reduced to that of uncoupled spins. 2 2 i i i. we have Pσ = exp −β (Eσ − Eρ ) = exp (−β ∆E ) . In the last expression. For the (hyper)cubic lattice in D dimensions.2). we have T (ρ → σ ) Pσ = . which we call q. the coordination number is 2D which gives indeed 2 neighbours in 1 dimension. We can rewrite the Hamiltonian in the form: H =K ∑ (si − m + m)(s j − m + m) − h ∑ si = i. If the probability distribution Pρ becomes stationary. 2 2 /2 ∏∑ i si eβ (qmK+h)si . On the Ising lattice. We shall now sketch the mean ﬁeld approximation for the Ising model. eβ (qmK+h) + e−β (qmK+h) . T (σ → ρ ) Pρ For the Boltzmann distribution. the coordination number. which does not depend on i for a homogeneous Hamiltonian (we assume the Hamiltonian satisﬁes periodic boundary conditions so that a site on the edge of the lattice couples to the corresponding site on the opposite edge. we have neglected the quadratic contribution of ﬂuctuations of the magnetisation around its equilibrium value. each site has a number of neighbours. This approximation occurs in many different forms. We want to ﬁnd the average magnetisation m = si . but always boils down to replacing the interactions between a particular particle and its neighbours by the average value of the interactions between that particle and all its neighbours. j (the number N represents the number of lattice sites). It is however important that apart from the ﬁnite size of the system. where si is the average value of the spin.71 where T (ρ → σ ) is the probability to go to state σ provided the system was in a state ρ . We can evaluate this from our mean-ﬁeld Hamiltonian: m = si = eβ (qmK+h) − e−β (qmK+h) = tanh β (qmK + h) . Pρ The Monte Carlo method is ﬂexible in the sense that in principle any dimension and many types of interaction can be treated in this way. no systematic approximation is introduced. Note that this expression for the free energy still contains the unknown average magnetisation m. Now we can evaluate the free energy by factorising the partition function: F = −kB ln Z = −kB T ln e−β KqNm This expression can be evaluated as F= K qNm2 − kB T N ln {2 cosh [(Kqm + h)/kB T ]} . the results are subject to statistical errors and will be not inﬁnitely accurate.

25 in two dimensions. We see that the number of such points depends on the value of qJ . which is 0. it turns out that the m = 0 corresponds to a higher value of the free energy than the nonzero values. and still the point at m = 0. which give an equal value of the free energy. which now lies at β K = 1/q.5 1 qJ= 2 q J=1 q J= 0. Considering the free energy. For qJ > 1.5 0 0.2: Periodic Boundary conditions in the Ising model. Obviously.5 Figure 10. there are three such points. the system will choose one of these two. . the two should be the same. we show the leftand right hand side as a function of m for h = 0 and for different values of the parameter β qK ≡ qJ .3: The self-consistency relation for the Ising model for different values of the parameter qJ = β qK . We see that this is a self-consistency relation for m: on the right hand side. there is only one intersection point. The self-consistent values for the magnetisation correspond to acceptable values. which we have evaluated as the result on the left hand side.5 1 0. The phase diagram is therefore the same as the one described in the previous section.72 Figure 10. m is the average magnetisation which we imposed before evaluating the very same quantity.3. In ﬁgure 10.5 0 -0. at m = 0. 1. two of which have opposite nonzero values of m. For qJ < 1. Which of these three points will the system choose. Therefore.5 -1 -1 -0. except for the location of the critical point.

For 3D. kB T 3 kB T 1/3 . T Differentiating the self-consistency relation at arbitrary T with respect to h and then putting h → 0 yields Tc 1 χ= χ+ 1 − m2 . This means that the values of the neighbours of a + spin differs from those of the neighbours of a − spin. The reason behind this is that in the mean ﬁeld approximation. It seems that the mean ﬁeld approximation can cause large errors in the exponents. irrespective of the value of si . to be compared with δ = 15 for the 2D Ising model on a square lattice. First we analyse the magnetisation as a function of temperature. ordinary critical points have indeed β = 1/2. It turns out that for dimensions greater than four.73 We can now calculate the critical exponents for the Ising model. Mean ﬁeld theory is very successful in systems with long range interactions. The procedure is similar to that followed in calculating β : we simply must replace qJm by m + h/kB T m = (m + from which it follows that m= 3h kB T 1 h 3 h ) − (m + ) . . 3 from which we obtain. we neglect correlations. . Now we study how m varies with h at the critical point. m = 0 and we see that χ ∼ |T − Tc |. kB [T − Tc (1 − m2 )] . Tc Thus we see that the exponent β is found to be 1/2. We can expand the tanh function on the right hand side: 1 m = qJm − (qJ )3 m3 + . hence γ = 1. we have however replaced these values by m. T T from which we have χ= For T > Tc . Now let us calculate the other critical exponents. The magnetic susceptibility is deﬁned as χ= ∂m ∂h . to be compared with γ = 7/4 for the 2D Ising model on a square lattice. Above. the exponent β is 0. the mean ﬁeld approximation becomes better with increasing dimension. which is already quite a bit closer to the mean ﬁeld value. quite different from the exact result β = 1/8 for 2D (see above). 1 − m2 .324. In general. This means that we must set qJ = 1 and then include the magnetic ﬁeld into the expression for m and see how the latter varies with h. writing qJ = Tc /T : √ m = ± 3 1− T . In higher dimensions (or when a spin has very many neighbours) these correlations become less important and the mean ﬁeld result becomes more and more reliable. giving δ = 3.

In fact. bx rx rx . i and j are not necessarily neighbours. For short distances. the calculation of the exponent proceeds a bit differently. The behaviour of correlations is generally seen from the correlation function This function is deﬁned as 1 g(|ri − r j |) = si s j = ∑ si s j exp(−β H ). ξ : g(r) ∼ exp(−r/ξ ) − m2 . we have seen above that the correlation length diverges near the transition point according to a scaling law with the critical exponent ν = 1. If the correlation function decays with a given exponent ξ . we have an algebraic decay of the correlation function. The speciﬁc heat is calculated as Ch = 1 kB (Tc − T ) ∂E ∂T = 0 for T > Tc . rx where x is another example of a critical exponent. At the critical point. it does not change its shape under a recaling r → br: g(br) ∼ 1 1 1 ∼ . The shape of the correlation function is nearly always exponential.74 For T < Tc . The distance ξ over which the correlations decay increases when the critical point is increased. h where we have used the fact that m = 0 for T > Tc . the correlation function changes from exponential to algebraic: g(r) ∼ 1 . χ= which again leads to γ = 1. which is the same as in the 2D Ising model on the square lattice. which is called the correlation length. the correlation function becomes more or less isotropic. 2 Tc 2 This tells us that the critical exponent α = 0. we have Ch = 3 KqN 3 = kB N . we consider the exponent of the speciﬁc heat. If however. First we note that the energy for h = 0 is given by ∂βF 1 E= = − KqNm2 ∂β 2 where we have used the self-consistency relation to replace the tanh occurring in the second term by m. For longer distances however. We have m ≈ putting this into the above equation leads to 3/Tc (Tc − T )1/2 . Z {i} Note that now. The term m2 is the value which we expect for long distances: the average values of the spins si and s j are not correlated and equal to the single-site average m. and the absolute values are justiﬁed. a rescaling of the physical space will change the correlation length accordingly. Finally. with a typical length scale. using the absolute value on the left hand side is not justiﬁed as the expression on the right hand side is anisotropic. For T < Tc .

An Ising model at the critical point is an example of a fractal structure. here we ﬁx a priori the magnetisation to its predeﬁned value.3 Landau theory of phase transitions Suppose we ﬁx the magnetisation of our Ising model to the value m. The parameter q furthermore can be set to zero by a suitable redeﬁnition of the zero of energy.75 This behaviour is a manifestation of the fact that the system is scale invariant. but we must restrict ourselves to those conﬁgurations for which the total magnetisation sums up to Nm. with s > 0. we see that F = −hm − kB T ln ∑ exp J ∑ si s j . . ′ {si } i. . in order for the free energy to be acceptable. We then have F = rm2 + sm4 . . Therefore. where b = β h and the primed sum denotes the restricted sum over the conﬁgurations with magnetisation m. j Note that the second term should be even in m as it does not have any preference for up. the picture looks the same. we prefer using the reduced temperature t as a parameter: t= T − Tc . . Tc Close to the critical point t is small. and we would not notice the anisotropy of the lattice at short scales. The Taylor coefﬁcents q. which as usual does not affect the physics of the problem. 10. we can no longer sum over all possible conﬁgurations with si = ±. In ﬁgure 10. we know that the minima of F as a function of m correspond to equilibrium. .or down directions. We see that for r > 0 there is only one minimum of the free energy at m = 0. When we evaluate the partition function. we should have s > 0. F = −hm + q + rm2 + sm4 + . if we now relax the value m. instead of temperature. and we may expand the free energy in m. with Z= {si } ∑ ′ exp J ∑ si s j + bm i. Fractals are examples of scale-invariant objects. In fact. A system at ﬁxed temperature will occupy. . s. From the expression for the free energy: F = −kB T ln Z . We furthermore study ﬁrst the case where h = 0. If we do so. Note the difference with the mean ﬁeld approximation: there we started with calling m the average magnetisation which was to be calculated. Cutting off the expansion beyond the fourth order term. j . at its equilibrium. If we would look at an Ising model from a distance from nearby.4 we show F (m) for several values of q. Note that r and s are functions of the temperature t . states which correspond to a minimum of the free energy. depend on temperature. otherwise the free energy would be lowest at large m which is clearly invalid close to the critical point. Close to the critical point. the free energy as calculated from this partition function is the free energy evaluated for the particular value of the magnetisation we started with. the magnetisation is small. This . r.

where r1 . as in the mean ﬁeld approach. we would expect this magnetisation to vary linearly with h. Usually. The assumption is that the Taylor coefﬁcients are regular functions of t .4: The free energy in the Landau expansion for various r. For example. we can again derive the critical exponents. just as in the mean ﬁeld theory. s = s0 + s1t . minimum turns into a maximum for r < 0 and is accompanied by two minima at ±ms . Also for the other critical exponents. Another exponent that we can ﬁnd is δ . r and s behave as r(t ) = r1 t . . so that δ = 3. for small t : m ∝ t 1/2 . We assume that r1 is nonzero. but here we ﬁnd a different behaviour. Taking the expression found above: F= K qNm2 − kB T N ln {2 cosh [(Kqm + h)/kB T ]} . 2 −r 2s . From the Landau theory. the mean ﬁeld values are found. At the critical point. This is not surprising. Substituting the . Around t = 0. s0 and s1 are some constants which do not depend on t . This exponent tells us how the magnetisation varies with h at the critical point for small magnetic ﬁeld. hence β = 1/2 the same as in the mean ﬁeld approximation. as it can be shown that the mean ﬁeld free energy can be expanded in a series which is equal to the Landau expansion. r = 0. the exponent β is found by analysing how m goes to zero when t approaches 0. where ms is the value of the spontaneous magnetisation. The minimum of the Landau expansion for the free energy is found at m = ± expansions for r and s we obtain. so F varies with m as F = −hm + sm4 . We know that criticality corresponds to t = 0 and r(t ) = 0. We see that the critical exponents from the Landau theory and for the mean ﬁeld approach are the same.76 8 7 6 5 4 F 3 2 1 0 -1 -3 -2 r >0 r =0 r<0 m s -1 0 1 ms 2 3 m Figure 10. We see that m ∝ h1/3 .

A full calculation of the free energy starting from this Hamiltonian is difﬁcult because of the presence of the m4 term – the way to proceed is by a diagramatic expansion as done in Wilson’s renormalisation theory. As we are interested in phenomena close to the critical point. and within each cell. deﬁned as g(r) = m(r0 )m(r0 + r) − m(r0 ) 2 . the energy (Hamiltonian) may be described by a Landau expansion as in the previous section. The possibility to vary h and m with position enables us to evaluate the correlation function. We take a magnetic ﬁeld which only couples to the spin located at r = 0. u = Kq/3(Kq/kB T )3 . even in the presence of an external ﬁeld. after replacing the difference by a gradient and integrating over the volume: H= k [∇m(r)]2 + h(r)m(r) + rm2 (r) + sm4 (r) d 3 r. this contribution can be cast in the form α m(r′ ) − m(r) + β m(r′ ) − m(r) + . writing up such a Hamiltonian for each isolated cell. . Therefore. . the term on the right hand side does not depend on r0 . moreover. For a homogeneous system. The correlation function with the Landau-Ginzburg Hamiltonian is found as follows.77 and expanding this in terms of m. we can assume the critical behaviour found in the previous section to be valid and use our new Hamiltonian only to evaluate the correlation function. with q = −kB ln 2. However. we ﬁnd: F = q + rm2 + sm4 + . This is important because the integral over the correlation function is precisely the term which was neglected in the mean ﬁeld theory. where r and r′ are the coordinates of neighbouring cells. In the Landau theory it is assumed that the free energy is determined by the average value of the magnetisation. r = Kq(1 − Kq/kB T ).4 Landau Ginzburg theory and Ginzburg criterion In order to obtain more insight into the approximations made in formulating the Landau theory. However. The idea behind this formulation is as follows. means that we neglect the couplings between neighbouring cells. the ﬂuctuations of the magnetisation must be considered as well. Note that the magnetic ﬁeld h varies with position. We divide the volume of the system up into cells which are very small in comparison to the system volume. We therefore obtain a consistency criterion fo this theory. 10. we have. The fact that these ﬂuctuations were neglected is an approximation similar to the mean ﬁeld approximation. . we now formulate a mesoscopic form of this theory. Then the average spin in such a cell is a continuous variable which we shall call m(r): r is the location of the cell. This explains why the mean ﬁeld exponents were recovered in the Landau theory. In reality. it will not depend on the sign of that difference. Keeping only the lowest order term in this expansion. but still large enough to contain many spins (in the case of a gas/liquid. which is beyond the scope of this course. we are dealing with small values of m(r). and m assumes values between −1 and +1. these subvolume must contain many particles). 2 4 This form of the Hamiltonian reproduces the results of the previous section is m(r) does not vary with r. . This means that we have a term . This coupling will depend on the difference between the magnetisation in these cells.

Now we evaluate the magnetisation at some point r (not necessarily at the origin): ∏r dm(r′ ) exp −β H ({m(r′ )}) d 3 r + hm(0) m(r) . For h = 0 we recover the Landau result m(r) = m0 = 0 for T > Tc (r > 0) and m(r) = m0 = For small h. we only have to evaluate its derivative in order to ﬁnd the correlation function. Substituting this in our differential equation we obtain r ∇2 g(r) − g(r) = k r 2 ∇ g(r) + 2 g(r) = k The solution for these equations is g(r) = where    k r −k 2r −r for T < Tc (r < 0). This is the solution m(r) which minimises H ({m(r)}) d 3 r + hm(0). we may identify ϕ (r) with g(r). Varying m(r) by δ m(r) we have 2rm(r)δ m(r) + 4sm3 δ m(r) + 2k∇m(r)∇ [δ m(r)] + hδ m(r)δ (r) d 3 r = 0. 2k 1 −r/ξ e . 2s 1 δ (r) for T > Tc and 2k 1 δ (r) for T < Tc . 2 Remember r is a parameter of the Hamiltonian – it does not denote |r|. 8kπ r for T > Tc and for T < Tc . ξ= . and require that δ m(r) vanishes at the boundary of the system in order to ﬁnd the equation h −k∇2 m(r) + rm(r) + 2sm3 (r) + δ (r) = 0. We apply Green’s theorem to the term with the gradients. m(r) = ∏r dm(r′ ) exp {−β [ H ({m(r′ )}) d 3 r + hm(0)]} The correlation function is then given by β g(r) = dm(r) = β [ m(r)m(0) − m(r) m(0) ] dh so if we ﬁnd m(r) as a function of h. and because g(r) = dm/dh.78 hm(r = 0) in the Hamiltonian. We ﬁnd m(r) by requiring that only the contribution which maximises the Boltzmann should be counted in evaluating the equilibrium value. we can write m(r) = m0 + hϕ (r).

but sketch the ideas and apply them to the simple case of the one-dimensional Ising model. . ∼ r m2 0 We see that the ﬂuctuations can safely be neglected for d > 4.si+1 = exp Jsi si+1 + (si + si+1 ) = si |T |si+1 2 where we have used Dirac notation in the last expression. N We use periodic boundary conditions. so that sN +1 ≡ s1 . still yielding η = 0. |sN −1 sN −1 |T |sN sN |T |s1 {si =±1} .5 Exact solutions Another way for studying phase transitions is by exact solutions. 10. which leads to ξ 2−d r−(2−d )/2 = r(d −4)/2 . The partition function of this model can be written as Z= {si =±1} ∑ exp J ∑ si si+1 + B ∑ si i=1 i=1 N N = {si =±1} i=1 ∑ ∏ exp (Jsisi+1 + Bsi) . For smaller dimensions. For dimension other than 3. m2 0 For T approaching Tc from below. Exact solutions are generally quite difﬁcult to obtain. Quite a few spin models on regular lattices have been solved exactly. the above result can be generalised to g(r) ∼ 1 rd −2+η . The required result enables us now to estimate the term which was neglected in Landau (or mean ﬁeld) approximation: the spatial ﬂuctuation of the magnetisation. we see that the critical exponent ν = 1/2. At the critical point. Now we deﬁne the transfer matrix as follows B Tsi . |r| which gives an exponent η = 0. we can expect corrections to the classical exponents. The solution is well known: 1 g(r) ∼ . and we shall refrain from treating them in detail here. we have m0 = −r/s. The relative importance of this ﬂuctuation can be estimated as follows: ξ d a g(r)d r ξ 2 d a m0 d r ∼ ξ 1 d a |r|d −2 d r ξ d m2 0 ∼ ξ 2−d . we must use r = 0 in the differential equations for g: h ∇2 g(r) = δ (r) k which is recognised as the electrostatic problem of ﬁnding the potential of a point charge. We can now rephrase the partition function as follows Z = ∑ s1 |T |s2 s2 |T |s3 s3 | .79 Since r is proportional to T − Tc (see previous section). .

i=1 N For large N . and the second largest is λ2 etcetera. Then Z = ∑ λiN .80 which. |si si si |T .e. we would replace the transfermatrix by its largest eigenvalue. Z ≈ λ1 We now can calculate the average value of some spin si somewhere in the Ising chain. . can immediately be written in the form Z = Tr T N . using the fact that ∑s |s s| is the unit operator. However. If we follow the same argument as we used in the calculation of the partition function. This eigenvalue is λ2 . λ2 This correlation length becomes inﬁnite when λ1 = λ2 .the second term cancels against the ﬁrst and the result is zero. The largest (in absolute value) eigenvalue is called λ1 . We assume j > i and j − i ≪ N . This the term in which we replace the part between i and j (the term T j−i ) by the second largest eigenvalue of T. Slightly more difﬁcult is the evaluation of the correlation function gi j = si s j − si s j . symmetric matrix. It is easy to see that this is given by si = {si =±1} ∑ s1 |T |s2 s2 |T |s3 s3 | . |sN −1 sN −1 |T |sN sN |T |s1 . where φ1 is the – normalised – eigenvector corresponding to the largest eigenvalue λ1 of T . Therefore. i. in that case. Above we have indicated that the critical point is characterised by a divergence of the correlation length. we are left with gi j = λ2 λ1 j −i | φ1 |s|φ2 |2 − ( φ1 |s|φ1 )2 . . and therefore we have N . Then s1 |T i−1 |si si si |T j−i |s j s j s j |T N − j+1 |s1 − gi j = N λ1 Using again the fact that an expression like this is dominated by the eigenvector with the largest eigenvalue of T . We see that g decays exponentially with correlation length ξ = ln λ1 . . The main contribution to the transfer matrix comes from the second largest term in of the ﬁrst part. Therefore its eigenvalues are real. si = φ1 |φ1 φ1 |s|φ1 φ1 |φ1 2 . . It is easy to see that T is a real. the contribution from the largest eigenvalue will dominate the sum. when the largest eigenvalue becomes degenerate. we obtain φ1 |si |φ1 . so we identify the critical point with the point where the eigenvalues of T become .

5.6 Renormalisation theory Close to a critical phase transition. as argued above in section 11. a model which has only a ﬁnite number of degrees of freedom in the direction perpendicular to the transfer-matrix direction. never exhibits critical behaviour. The largest droplets are of a size of the order of the correlation length. . Important in this respect is the Frobenius theorem. but there we see droplets with spin opposite to the direction of the overall magnetisation. The foregoing description suggests that a critical model is scale invariant. 1964). . within droplets. For the Ising model above the critical temperature. Schulz. The two dimensional Ising model can be solved exactly using the transfer matrix method (Onsager 1944.5: Kadanoff’s droplet picture for the Ising model.81 _ + _ + _ + _ + + _ + + _ + _ + _ _ + + _ _ Figure 10. if this becomes inﬁnitely large. We shall discuss the general ideas behind this theory by considering an Ising model with nearest neighbour coupling J and next . degenerate. This does not mean that there is no structure characterised by length scales smaller than the correlation length. Mattis and Lieb. the largest of which are of a size comparable to the correlation length. The situation is different in the two-dimensional Ising model. Baxter (1982) has written a book about exactly solved models in statistical mechanics. Kadanoff has characterised the critical phase by a droplet model. the onedimensional Ising model does not have a critical phase transition. this droplet picture incorporates large regions (droplets) of one spin direction. . This theorem tells us that the largest eigenvalue of a symmetric matrix with all positive elements is nondegenerate. which contain smaller droplets of the opposite spin. In general.2 in connection with the divergence of the correlation length and the shape of the correlation function (power law). the correlation length diverges. within droplets. and Frobenius theorem no longer applies. as shown in ﬁgure 10. and these droplets contain in turn smaller droplets with the ﬁrst spin direction etcetera. Thus we have droplets. 10. The scale invariance of a model at its critical point is the notion which lies at the basis of the renormalisation theory which is described in this section. Therefore. Working out the transfer matrix for the 1D model and its eigenvalues is left as an exercise. For T < Tc the picture is similar. . Then the transfer matrix becomes inﬁnite.

the actual value is chosen at random with probabilities 1/2 for both values. j si s j where i. Then we can assign an energy to this conﬁguration using the following rule: exp(−β H ′ {tk }) = {si } ∑ e−β H ({s })W (tk . K space. as in ﬁgure 10. The coarse spins are denoted as t instead of s. The spins t can assume the values ±1. s4 ) (k ) (k ) (k ) (k ) where {tk } is a conﬁguration of plaquette-spins tk . s3 . In a renormalisation transformation we try to formulate the model on a coarser scale. The t -spins are located at the centers of a subset of the plaquettes of the lattice. These values are determined by the values of the spins at the corners of the plaquette according to the following rules: • If s1 + s2 + s3 + s4 > 0 then t = 1. s3 . We study the model in the J. Now suppose we ﬁx the values of t on each plaquette. whereas i. Speciﬁcally. and W are the probabilities to have a spin t given the values of the four corner spins – these probabilities follow directly from the rules given above: • W (t = 1. k denotes the plaquettes. j denotes nearest neighbour pairs. we want to formulate the partition function in terms of new spins which are deﬁned on a lattice with a larger length scale. si denotes the spins surrounding plaquette k. . just as in the ordinary Ising model. s4 ) = 1 if s1 + s2 + s3 + s4 > 0. s2 . • If s1 + s2 + s3 + s4 < 0 then t = −1.6: Example of a coarsening procedure as is performed in a renormalisation transformation nearest neighbour coupling K (including a factor 1/(kB T )): H = −kB T J ∑ si s j + K i. j ∑ i.82 s1 t s3 s2 s’ 1 t’ s’ 2 s4 s’ 3 s’ 4 Figure 10. j are the next-nearest neighbour pairs. • If s1 + s2 + s3 + s4 = 0 then t = −1 or 1. s2 .6. s1 . s1 i (k ) .

s2 . which we take +1 to be speciﬁc. This transformation is the renormalisation transformation. s3 . (k ) (k ) (k ) {tk } {si } We can now move the sum over the conﬁgurations {tk } to the right as the term e−β H ({si }) does not depend on the ti : ′ (k ) (k ) (k ) (k ) ∑ e−β H ({tk }) = ∑ e−β H ({si }) ∑ W (tk . most of the lattice is single-colored (either red or blue). We say that we have integrated out all degrees of freedom at length scale of the lattice constant a and are left with a new Hamiltonian of the same form as the original one. It is important to realise that what we have done is a clever scale transformation. the plaquette spins will be +1. If we calculate the tk . Thus. . s2 . but with different values of the interaction constants. Now we show what we can do with this new interaction. and new values of the coupling constants. {tk } ∑ e−β H ({t }) = ∑ e−β H ({s }) = Z . and the small deviating color spots have dissapeared. s1 . a point J. as the distance between two neighbouring t -spins is twice that between two si spins. We have now merely deﬁned a new interaction for the tk . In other words. Therefore it looks like a system at T = 0. Let us however assume that we can approximate H ′ – up to an additive constant – reasonably well by a form similar to H . Now let us consider the renormalisation transformation in the J. the coupling constants J and K are mapped onto new ones. s4 ). We shall come back to this point later. s4 ) = 1 if s1 + s2 + s3 + s4 < 0. Under a renormalisation transformation. s1 ′ k i (k ) . s1 . s1 . For all other conﬁgurations. How will these points transform? We consider a few special cases. s2 . with small spots of minority color. s2 . the weight factors W are zero. Then we repeat this procedure over and over. This means that in going from the si to the tk . • W (t . K ′ . s2 . Then only few spins si might deviate from the majority value. hence the new coupling constants J ′ and K ′ describing the t -spins will be larger than J and K respectively. We calculate the partition function for the tk : {tk } ∑ e−β H ({t }) = ∑ ∑ e−β H ({s })W (tk . s4 equals 1 (the W have been designed this way). even when an isolated spin si = −1 is found at the corner of the plaquette. s3 . the lattice looks either red or blue. In the low temperature phase. {tk } {si } {tk } Now we note that the sum over {tk } of the weights W for any ﬁxed conﬁguration of the four spins (k ) (k ) (k ) (k ) s1 . but with a lattice constant 2a. K plane. • Consider a point where J and K are large (low-temperature case). s4 ). K is mapped onto a new point J ′ . Therefore. If we look at the lattice from a distance so that we no longer distinguish the ﬁne detail. the renormalisation transformation causes the temperature to go down. in the low temperature. s3 . ′ k i {si } We see that the new spins tk form a model with a partition function which is exactly the same as the original one which was deﬁned in terms of the si . s3 . s3 . Another way to understand this is to imagine that we color the + spins red and the − spins blue.83 • W (t = −1. the tk are more likely to assume the majority value than the si . J ′ and K ′ . s4 ) = 1/2 for t = ±1 if s1 + s2 + s3 + s4 = 0. A problem is that the form of H ′ might differ from H .

the small couplings J and K transform into even smaller values J ′ and K ′ . • At high temperatures. the edges have become broader and broader and ﬁnally these regions where the t spins are randomly +1 or −1 cover the whole lattice.7. K plane. If we coarsen however. Obviously. This means that the edge becomes fuzzier. Again. We therefore have opposite ﬂows towards two ﬁxed points. when we look from a distance at a lattice at the lattice consisting of red and blue patches. the effective couplings J and K are weak. so we have a system at inﬁnite temperature. By extending the arguments above for describing the ﬂow in the J. a positive coupling J will be generated in the renormalisation transformation. the two attraction basins must be separated by a line.from the high temperature phase. the rescaling has made the patches smaller. Then there is a third ﬁxed point on the separatrix.84 K J Figure 10. Collecting all this information we have the picture shown in ﬁgure 10. in other words. and in the end they have the same size as the (new) lattice constant.7: Renormalisation ﬂow. in other words. The system will consist of patches of a single color – these patches have a size larger than the lattice constant a because the spins still feel the ferromagnetic interaction. near the edges of such a patch. Points on that line must ﬂow along that line. K plane looks like: there is a ﬂow towards the origin J = K = 0 and there is a ﬂow towards inﬁnity J = K = ∞. and the reverse is also true: from the point J > 0. or. K = 0 we move to a point where both are positive. Two of these (at zero and inﬁnity) are so-called trivial ﬁxed points where the physics of the model is easy to understand and relatively trivial. the new spins t on the coarser lattice will assume random values as on the edge there are about as many spins + as there are −. Now we can imagine what the ﬂow diagram generated by the renormalisation transformation in the J. We see that there are three ﬁxed points. we can see that for J = 0. After a few renormalisation steps. because it separates the low. The entire separatrix is interpreted as a line consisting of critical points. called separatrix. The average spin value is 0 – there is no net magnetisation. We see that the t seem to live in a system at higher temperature than the system of the s spins. we see in the end a mixture of red and blue. K > 0. . Both ﬁxed points have an attraction basin which is the region of points which ﬂow eventually towards the ﬁxed points.

There is a clear relation between N and N ′ : N ′ = N /2d as can be seen directly from the renormalisation procedure. One of these. we must realise that the latter scales with l . If we call the coordinates of the ﬁxed point J ∗ . λ and µ . Therefore. as shown above: e−β N f (J. J ′ ) = ξ (K . corresponds to the points on the separatrix and will be smaller than 1. we relax the rescaling constant 2 to assume any value larger than 1. However.85 So far. This constant also depends on J and K . K ′ ) + c where c = kB TC/N . we ﬁnd two real eigenvalues. K ∗ .K) = Z (J. we consider the partition function in more detail. so ξ (K ′ . ′ ′ ′ ′ ′ ′ The constant C is the difference between the original and the renormalised Hamiltonian. so the correlation length does not change under the renormalisation transformation. Now we can write Z as the partition function of the Ising model either in terms of the ‘original’ spins si . To proceed. and call it l . we linearise the renormalisation transformation. essentially invariant. Points on the separatrix ﬂow to that point under the renormalisation transformation. To keep the discussion general. K ). K ) = l −d f (J ′ . if we measure the correlation length in units of the grid constant.K )−C = e−β N f (J . This means that the corresponding eigenvector lies along the . The free energy is related to the partition function by F = −kB T ln Z . The renormalisation transformation leaves the physics at length scales beyond the scales over which we have integrated out the degrees of freedom. The result we can infer for the free energy per spin is: f (J. ∆J ′ and ∆K ′ can be written as ∆J ′ ∆K ′ =A ∆J ∆K . We can analyse the transformation of the correlation in a similar fashion. or in terms of the coarsened spins tk . K ∗ + ∆K ′ ). where A is a 2 × 2 matrix whose elements do not depend on ∆J and ∆K . Now let us analyse the behaviour of the transformation close to the (nontrivial) ﬁxed point. If we diagonalise the matrix A. we have only discussed the mechanics of the renormalisation transformation but have not derived any useful result from it. To ﬁrst order in ∆J and ∆K .K )−C . we have RT (J ∗ + ∆J. the free energy per spin f = F /N (for N spins) satisﬁes: e−β N f (J. To describe the behaviour near the ﬁxed point. J )/l . say λ .K) = {tk } ∑ e−H (J . whereas points in a direction perpendicular to that line ﬂow away from the ﬁxed point. K ∗ + ∆K ) −→ (J ∗ + ∆J ′ .

Therefore. J and K can be reparametrised in terms of s and t . Therefore. 1). Note that any point on the separatrix moves under the renormalisation transformation eventually to the ﬁxed point. B). we must have y < 0 and z > 0. Other critical exponents follow from a similar analysis of the free energy. Choosing now n such that l nzt = 1. the behaviour at long length scales of these points is dominated by the properties of this ﬁxed point. Now there exists an important theorem by Wegner which says that the eigenvalues λ and µ must have the following dependence on l : λ = l y. there is an outward ﬂow. We therefore can identify t= T − Tc . The indices y and z are called scaling indices. as varying the temperature corresponds to moving on a straight line through the origin. Note that close to the critical point. corresponding to a positive scaling index (eigenvalue > 1) and a negative scaling index (eigenvalue < 1). Repeated application of the transformation leads to the rule: s(n) = l ny s. µ . We have seen that near a ﬁxed point. µ = lz.86 sepatrix at the ﬁxed point. t and h. The other. From the renormalisation ﬂow plot. will be larger than 1 and the corresponding eigenvector lies in the outﬂowing direction. as the magnetic ﬁeld breaks the symmetry and destroys the critical . For t → 0. K . Scaling ﬁelds with positive indices are called relevant. we have ξ (s. For the correlation length we therefore have: ξ (s. l nzt ). we have three scaling ﬁelds. and using the fact that y/z < 0 (see above). it is useful to expand the parameter space to include the magnetic ﬁeld B (including the factor 1/kB T ): f = f (J. The corresponding eigenvectors are called Wegner scaling ﬁelds. which we denote as s. we can infer thet t must be related to the temperature. To analyse these exponents. Tc and we see that the critical exponent ν is given by 1/z. 1) ≡ A t 1/z . We call these scaling ﬁelds s (corresponding to λ ) and t (corresponding to µ ). those with negative indeces are irrelevant. Note that the corresponding scaling ﬁeld must be relevant. t ) = l n ξ (l ny s. t ) = t −1/z ξ (0. and those with index 0 are called marginal. t (n) = l nz t . we have ξ (s. Since λ < 1 and µ > 1. t ) = t −1/z ξ (t −y/z s.

l nv h). If we are at a critical point and change either the temperature or the ﬁeld. as these determine the critical properties. t ′ . ht −v/z ). The direction corresponding to h is perpendicular to the plane of the renormalisation ﬂow plot in ﬁgure 10. We now neglect c(s. and therefore these points would be critical. However. We call v the scaling index. and we are interested in the part of the free energy which contains singularities. so that The exponent β describes the behaviour of the magnetisation when the critical temperature is approached from below. • s corresponds to a particular combination (e. h′ ) + c(s. If it were irrelevant. h=0 The contribution to the magnetisation arising from the singular part of the free energy is therefore m ∝ t d /z t −v/z . Ch = ∂ t 2 h=0 Restricting attention to the singular part of f . . t and h. Therefore we have fsing (s. From this expression we can derive the critical exponents in terms of the scaling indices.87 behaviour. We call the scaling ﬁelds s. a judicious change of both J and K keeps the system at the critical point. h). h) = l −d f (s′ . we have fsing (s. We have seen before that the free energy per spin transforms as f (s. t . Recall that the magnetisation is given by m= α = 2 − d /z. ∂f ∂h . The interpretation of these ﬁelds is as follows: • t corresponds to the temperature. we calculate the exponent α for the speciﬁc heat per particle. there would be points corresponding to nonzero ﬁeld which would ﬂow to the ﬁxed point. the ratio) of J and K .7. as we know. we see that Ch ∝ t d /z−2 . t . This is found from the free energy per particle as ∂2 f . Again we take l nz t = 1. so the exponent β is found as β = (d − v)/z. is not the case. we destroy the criticality. ±1. h) as this is a regular function. associated with h. First. h) = l −nd fsing (l ny s. l nz t . which. h) = t d /z f (st −y/z .g. t . After n renormalisation transformations. t . • h corresponds to the magnetic ﬁeld. We may ﬁrst note that again st −y/z will approach 0 (see the discussion concerning the correlation length). t .

t |h|−z/v . the exponent δ describes how the magnetisation varies with the applied magnetic ﬁeld. we must realise that the algebraic decay sets in beyond a distance of the order of the lattice constant a. Now the exponent η can be derived by evaluating the contribution from the algebraic component of g(r) in the integral. close to the critical point. rd −2+η Now we note that the susceptibilty χ is related to the correlation function by the relation χ= This relation is derived as follows. this decay persists up to the distance ξ . Given the fact that the magnetisation is given as ∑ s(r = 0) exp [H0 ({s(r)}) + h ∑r s(r)] m= r ∑r exp [H0 ({s(r)}) + h ∑r s(r)] and using χ = dm/dh. as we anticipated above. r where use has been made of the translation invariance of the system. d −v A different kind of analysis enables us to ﬁnd the critical exponent associated with the correlation function.88 For the susceptibility exponent γ . we choose n such that l nv h = ±1. Then. dh ∂h Finally. Therefore we have ξ (t ) 1 χ ∝ t (d −2v)/z = rd −1 dr. d − r 2+η a . where the ﬁrst term derives from the numerator. The decay is deﬁned in terms of a critical exponent η : yielding γ = −(d − 2v)/z. so that m = |h|d /v−1 . we ﬁnd χ= and we have dm ∂ 2 f = 2 ∝ t d /z t −2v/z . We recognise the right hand side of this equation as g(r)d d r. In order to ﬁnd this exponent. and that. t . δ= g(r) ∼ 1 . g(r)d d r. and the second from the denominator in the expression for m. h) = |h|d /v f (s|h|−y/v . In order to do this properly. where the correlation length ξ is large but ﬁnite. decays algebraically with distance at the critical point. g(r)d d r ≈ ∑ s(0)s(r) − N s(0) 2 . v . but vanishes beyond. we see that χ= ∑r s(r = 0) exp [H0 ({s(r)}) + h ∑r s(r)] ∑r s(r = 0) ∑r s(r) exp [H0 ({s(r)}) + h ∑r s(r)] −N ∑r exp [H0 ({s(r)}) + h ∑r s(r)] ∑r exp [H0 ({s(r)}) + h ∑r s(r)] 2 . which. fsing (s. ±1).

• β = (d − v)/z. 10. 10. ν = 1. ν = 1/2. • α = 2 − d /z. These new terms then have their own parameters. so dν = 2 − α .8 Universality Suppose that we add new interactions or ﬁelds to the Hamiltonian. but there are six critical exponents. If the new ﬁelds are relevant. The important notion here is that adding irrelevant ﬁelds leaves the critical exponents unchanged. If the terms are irrelevant they have absolutely no effect on the critical exponents.7 Scaling relations In the previous section. this means that the ﬁxed point lies on a line of critical points. It is clear that the critical exponents of the Ising model (α = 0. Therefore. There are only two such ﬁelds.89 where the angular degrees of freedom have been separated out from the integral. β = 1/2. the critical exponents may vary. δ = 3. so the total parameter space of the model acquires extra dimensions. Obviously the effect of the new terms strongly depends on whether they are relevant. • γ = −(d − 2ν )/z. the indices corresponding to the relevant scaling ﬁelds. and on this line. β = 1/8. then their presence will move the system away from the critical point under the renormalisation transformation. ‘New’ here means that they cannot be expressed in terms of the interactions and ﬁelds already present in the Hamiltonian. δ = 15. η = d + 2 − 2v. From this. and from ξ ∝ t −ν = t −1/z . as the integrand depends only on r.e. ν = 1/z. η = d + 2 − 2v. irrelevant or marginal. η = 0) satisfy these relations for d = 4. the exponents must be related among each other. so α + 2β + γ = 2. . γ = −(d − 2ν )/z. we have seen that the critical behaviour is governed by the positive scaling indices z and v. η = 1/4. have the same critical behaviour. δ = v/(d − v) so β (δ − 1) = γ . α = 2 − d /z. γ = −(d − 2ν )/z. All the models which differ from the original one by irrelevant terms. γ = 7/4 satisfy these scaling relations for d = 2 whereas the mean-ﬁeld or Landau exponents (α = 0. β = (d − v)/z. If they are marginal. we have • ν = 1/z. we see that so that t (d −2v)/z ∝ ξ 2−η ∝ t (η −2)/z . Inspecting the expressions for these exponents. so γ = (2 − η )ν . i. The class of models which is governed by one and the same ﬁxed point is called the universility class of the system.

8.8. given that the two + values of the plaquette spins is e−H (+. The reason we choose this lattice is that the transformation is simpler than for the square lattice. since both plaquettes have four possible conﬁgurations given their majority spin. Therefore. e ′ which leads to the following explicit form of J (J ). Two neighbouring coarse plaquettes are arranged as shown on the right hand side of ﬁgure 10. we consider the Ising model in two dimensions on a triangular lattice. and take the magnetic ﬁeld equal to zero. For opposite spins we have e−β H (+.90 t t’ Figure 10. J′ = e8J + 3e4J + 2e2J + 3 + 6e−2J + e−4J 1 ln .+) e2J = −H (+.+) = e8J + 3e4J + 2e2J + 3 + 6e−2J + e−4J . i.−) = 2e4J + 2e2J + 4 + 6e−2J + 2e−4J . The plaquette spins are chosen according to the majority rule: the plaquette spin is the same as the majority of spin values at the three corners of the plaquette.−) .9 Examples of renormalisation transformations As an example of a renormalisation transformation. H includes the factor 1/kB T = β .8: The renormalisation transformation for the triangular lattice in 2 dimensions. j where J ′ obviously includes the factor −β and where the renormalised coupling constant J ′ is given by the requirement ′ e−H (+. We take for the interaction only nearest neighbour couplings into account. 10. If both plaquette spins are positive. calling the plaquette spins ti we have a Hamiltonian of the form Hrenormalised ({ti }) = ∑ J ′tit j . 2 2e4J + 2e2J + 4 + 6e−2J + 2e−4J . The total number of weights on the right hand side adds up to 16. The transformation is shown in ﬁgure 10. then the sum of the Boltzmann weights corresponding to all possible conﬁgurations.

It is straightforward to include the magnetic ﬁeld into this renormalisation procedure. Moreover.365. i.8750).0.79.544 – this is the value of the eigenvalue λ . There is a ﬁxed point at the value J ≈ 0. 10. which describes the roughening of a crystal surface with increasing temperature.91 1 0. to be compared with the exact value of 1.9.10). It then follows directly that the scaling index z = 0. the degrees of freedom have a 2π periodicity: they can be viewed as the angles of unit vectors which in turn can be considered as ‘planar spins’ (see ﬁgure 10.02 (exact value 1. The Hamiltonian of the model is given by H = −K ∑ cos(θi − θ j ). This model has physical realisations in superﬂuid helium ﬁlms and arrays of superconducting islands separated by Josephson junctions.5 0 0 0. The model is formulated on a 2D lattice. All in all these values are encouraging.5 Figure 10.9: Renormalised coupling constant J ′ as a function of the original coupling constant J of the Ising model.274. Moreover. this procedure can be extended straightforwardly to encompass larger and larger clusters (Niemeyer and Van Leeuwen) and the values obtained in this way are within less then a percent of the exact ones. j This Hamiltonian favours the spins to be aligned (θi = θ j ).10 Systems with continuous symmetries Another example of a system for which a renormalisation transformation can be carried out is the XY model. The behaviour at low temperatures of this models is dominated by two types of excitations. These are excitations of the form θ (r) = A exp(ik · r) . This leads to a scaling index for the magnetisation of 2. The slope of the plot of the function J ′ (J ) turns out to be 1. In ﬁgure 10. the model can be mapped onto a roughening model. The exact solution for the Ising model on a triangular lattice gives a value 0. the renormalised coupling constant is plotted as a function of J . The ﬁrst type of excitations are spin waves.

In that case we can approximate the cosine occurring in the Hamiltonian by a quadratic expression: H ≈ −K ∑ i. θ (ri ) = ∑ θ k Therefore. j 1 1 − (θi − θ j )2 . We set ˜ (k) = 1 ∑ θ (ri )eik·ri θ L i where we assume that we are dealing with an L × L square lattice. Keeping the second term leads to the so-called Gaussian model. ˜ (k)e−ik·ri . . we obtain 2 ˆ )] ∑ [θ (ri ) − θ (ri + ax i 2 =∑ i ˆ) ˜ (k) e−ik·ri − e−ik·(ri +ax ∑θ k 2π aL (nx . as the Boltzmann factor has a Gaussian form. we obtain ˜ (k)θ ˜ (−k). ˆ )]2 = ∑[2 − 2 cos(akx )]θ ∑ [θ (ri ) − θ (ri + ax i k .0 .10: The XY model with planar spins on a square lattice. The Gaussian model can be solved exactly by Fourier transforming the variables θi . with wavelengths 2π /|k| much larger than the lattice spacing a. ny ). Then.92 Figure 10. 2 The ﬁrst terms in the square brackets add up to a constant which does not inﬂuence the model. since k assumes the values k = we have ∑ eik·r i i = N δk. Using the fact that.

Surprisingly. Jk2 The Gaussian integral gives precisely the same result as the corresponding Gaussian integral occurring in the denominator. Jk2 . so after dividing this factor out. For the numerator we can work out a factor. ˜ (k)θ k2 θ 2∑ k Deﬁning J = β K . k θ (k)θ 2 This is a product of Gaussian integrals which can be evaluated analytically to yield Z=∏ k π . We see that the expressions in the numerator and denominator factorize.93 Including the terms which are neighbours along the y direction. the model is critical for all temperatures! This can be seen by working out the correlation function cos[θ (r) − θ (0)] = ei[θ (r)−θ (0)] (the sin term disappears as a result of antisymmetry). the result is exp − 2 µ1 . Jk2 The integral has a lower bound corresponding to 2π /(La) and an upper bound of 2π /a. Finally. Jk2 For the free energy we then ﬁnd F = −kB T ln Z = −kB T ln 2π 3 d k. As the action (Hamiltonian) is quadratic in θk . we can evaluate this expectation value. We write θk = X + iY and µk = µ1 + iµ2 . and taking |k| small. Furthermore. We refrain from working it out but emphasise that this free energy is an analytical function of temperature (which is hidden in the parameter J ) so there is no phase transition. to obtain ei(µk θk +µ−k θ−k e−Jk 2 θ k θ −k d θk = e2i(µ1 X −µ2Y ) e−Jk 2 (X 2 +Y 2 ) dX dY . We see that the two integrals factorise. For the integral over X we have e2iµ1 X e−Jk 2X2 dX = exp −Jk2 X − i µ1 Jk2 2 exp − 2 µ1 . we realise that θ (−k) = θ ∗ (k) and similar for µk . the partition function can be written as Z= ˜k exp ∏θ −∞ k ∞ J 2˜ ˜ (−k) . we combine the integrals over k and −k into a single integral over k. which now runs over half the reciprocal space only. we have H =− K ˜ (−k). e i(θr −θ0 ) = e i N ∑k (eik·r −1)θk = ∏k d θk eiµk θk e− 2 ∑k k J 2 |θ |2 k 2 ∏k d θk e− 2 ∑k k J 2 |θ k| with µk = eik·r − 1.

The difference between neighbouring spin angles will therefore be 2π /(2π r/a) = a/r. and we obtain Etotal ∼ π K ln(R/a). the Hamiltonian favours spins to be aligned. We see that this correlation is critical (power law). This is an example of the Mermin-Wagner theorem (1966) which states that systems with continuous symmetry cannot exhibit long range order. ∑ 2 2J k k 2J k k2 Let us call the result of the sum f (r): f (r) = ∑ k 2 − 2 cos(k · r) . The energy stored in the spins around the circle is therefore E (r) ≈ 2π rK a a2 r 2 ≈ Kπ a . However. we must integrate this for radii up to the order of the system size R. π so that we ﬁnd for the correlation function g(r) = g(r) ∝ e− ln(r)/(2π J) = 1 r1/(2π J) . in the continuum limit: ∂2 ∂2 + ∂ x2 ∂ y2 f (r) = ∑ 2 cos(k · r) = 2N δr. r For the total energy. Jk2 Carrying out the product over k over the entire reciprocal space. The solution of this equation is 1 f (r) = ln r.94 For the integral over Y . we obtain exp − multiplying both terms gives exp − |µk |2 . The number of spins on that ring will be of order 2π r/a where r is the lattice constant. k2 Then. we obtain g(r) = exp − |µk |2 2 − 2 cos(k · r) 1 1 = exp − ∑ .0 . the XY model can exhibit excitation of a vortex character. In fact. k This is recognised as the Poisson equation for a line charge (or a point charge in two dimensions). In addition to spin-waves. Jk2 2 µ2 . the spin waves can be formed at such low energy cost that they will destroy any attempt to build up a real ‘long range order’. Consider a circle of radius a around the center of the vortex. and realising that the product of exponentials can be written as the exponential of the sum. .11. in which the expectation value for spins very far apart approaches a constant. First we study a single vortex as in ﬁgure 10. The energy of such a vortex can be calculated as follows.

e j = ±1. which we associate with a phase transition. We see that for low temperatures. two vortices have a ﬁnite energy (as opposed to a single vortex) and the entropy helps these pairs to proliferate at low temperatures. corresponding to positive and negative charges. We therefore expect a sudden. It can be rigourously shown that the vortex system behaves as a system of charges in two dimensions. The vortex system can be described by a so-called Coulomb gas in two dimensions: this is a gas consisting of charges which ﬂoat on the lattice. A vortex pair of opposite sign has a total energy of Epair = −π Jei e j ln ri − r j where ei . the dipoles will ‘melt’ and free charges will occur: the insulator becomes a conductor. whereas for high temperatures. . To see what the situation is like in the two-vortex case. The entropy associated with a single vortex is proportional to the logarithm of the number of ways we can place the vortex on the lattice: S ∼ kB ln(R/a)2 .95 Figure 10. it costs free energy to build up an isolated vortex. Therefore. spontaneous proliferation of vortices when the temperature exceeds a particular value. The vortices have two possible winding directions. we ﬁrst note that a vortex centered at r0 is described by a solution of the equation: ∇ × ∇ [θ (r)] = ±2πδ (r − r0 ). free energy is gained when a vortex is formed.11: An isolated vortex in the XY model. Beyond a transition temperature of the order of Tc = π K /(2kB ). the system will for charge dipoles and therefore be an insulator. The picture we have developed so far of the behaviour of this model is that at low temperatures. The free energy of a single vortex can therefore be estimated to be F = E − T S = (π K − 2kB T ) ln(R/a).

but with a prefactor given by the dielectric constant. dl dz = (2 − π J )z. the correlation function is simply g(r) = exp(−π J˜ln r) = r−π J . The renormalisation transformation can also be obtained using a self-consistency requirement for the linearly screened. the renormalisation equations have the form dJ = −4π 2 J 2 z2 . as anticipated above on the basis of a simple energy-entropy balance argument. The Hamiltonian is −π J ∑ ei e jV (|ri − r j |) + µ ∑ e2 i. Suppose we start at some values for J and z.e. i. We see that we end up at z = 0. the renormalisation trajectory brings us to a point with equivalent critical behaviour. Kosterlitz and Thouless have performed a renormalisation procedure to analyse the Coulomb gas model. Deﬁning z = exp(µ ). In the low-temperature phase. This is a disordered phase. In this phase. the potential with still be logarithmic. ˜ This is therefore a critical phase. i. The transition temperature lies at J˜ = 2/π . For high temperatures. This potential is deﬁned as the free energy associated with two inﬁnitesimal charges placed at 0 and r. dl The ﬂow diagram is shown schematically in ﬁgure 10. a phase with very low vortex density and a renormalised coupling constant J˜. when free charges can exist. the system becoms a conductor and the potential acquires an exponential form. j i The second term is added to have the freedom of changing the chemical potential of the charges. If the system is an insulator (T < Tc ). the behaviour of the system corresponds to that of a system with an inﬁnite concentration of vortices and high temperature.12: Renormalisation ﬂow for the Coulomb gas. .96 z π/2 1/J Figure 10.12. or effective potential. The transision point lies at J = 2/π . Beyond the critical temperature.

As we have seen in chapter 10. we have found for the superﬂuid velocity: ℏ us = ∇γ (r). At the phase transition. The coupling constant is ℏ2 ρ . we can have vortices associated with a rotation of the phase around a centre where the superﬂuid density ρ (r) = a2 (r) vanishes. On page 63. is in fact the superﬂuid density. This means that the phase can exhibit vortices. Similar phenomena have been observed for coupled arrays of Josephson junctions (Herre van der Zant) and for surface roughening transitions. the superﬂuid density is roughly constant. The phases γ (r) have the same property as the degrees of freedom in the XY model. and they carry a particular density of (quasi)-particles which are all in the same quantum state.97 What consequences does this have for physics? We shall describe the vortex physics of superﬂuid helium ﬁlms. These ﬁlms are essentially two-dimensional. This implies that the measured superﬂuid density jumps from ρcrit = 2kB T m π ℏ2 to zero. so we can evaluate the kinetic energy for the superﬂuid fraction as H= mρ 2 2 u2 s (r)d r = ℏ2 ρ 2m [∇γ (r)]2 d 2 r. m Outside the vortex centres. which is characterised as ψ (r) = a(r)eiγ (r) . J˜ jumps from the value 2/π to zero. J= m kB T What is measured in experiment is the renormalised coupling constant which. as we infer from the last equation. . This has been conﬁrmed experimentally by Bishop and Reppy in 1978. in the sense that they are periodic with period 2π .

or other quantities. this may no longer hold. as can be seen by expression (11.2 Local equation of state When we want to consider the ﬂow of a quantity Ai in the sense of the previous section. If Ai is a conserved quantity (for example particle number. However. by exchange of energy. dt b=a The quantities Φi represent ﬂuxes: their dimension is the dimension of Ai per unit of time. Note that a necessary condition for deﬁning Ai is that the subcell must be much larger than the microscopic length scale (atomic correlation length) and much smaller than the distance over which thermodynamic quantities vary – we then are in the hydrodynamic limit. The conservation law for quantity Ai can be formulated as dAi (t ) = − ∑ Φi (a → b) + Φi (sources → a). We focus on a particular quantity Ai within subcell a. t ) = V (a) ρi (t ) d 3 r (11. in terms of ρi . and identify a subvolume within this system of size (much) larger than the correlation length. For example. If we consider a system in equilibrium. Now consider such subcells inside some larger system. ji and σi : ∂ ρi + ∇ · ji = σi . particles may move from one cell to another [process (i)] or the cell may be connected to a source or drain of particles [process (ii)]. (ii) a loss or increase due to some source or sink for the quantity Ai inside the cell. We have seen in the beginning of 98 . we ﬁnd that the thermodynamic properties within this system are equivalent to that of the entire system. The systems will however tend to equalise these properties. using the divergence theorem.1) for any cell a much larger than the correlation length.1 Introduction Up to this point. ∂t 11.11 Irreversible processes: macroscopic theory 11. Then we may also deﬁne a local ﬂux ji and sourec term σ such that the conservation law above may be formulated. we have been exclusively concerned with equilibrium statistical mechanics. energy) then a change of Ai (a) may take place through two mechanisms: (i) a loss or increase due to Ai ﬂowing to or from neighbouring cells b. Now suppose that Ai can be deﬁned in terms of a density ρi : Ai (a. This exchange can be formulated in terms of ﬂuxes of the quantity under consideration. if we are not yet in equilibrium. momentum. and neighbouring subsystems will have different thermodynamic properties.1). we must take this Ai to be extensive.

It can be shown that the entropy does not change when we impart an equal velocity to all particles in our system. there is a conjugate. We ﬁnd the derivative with respect to the momentum component pi as follows: ∂ S(E . ∂ Ai As S is extensive. particle number. as are the extensive quantities Ai . p) ∂ S(E . we see that there exist pairs of conjugate variables. 0 ). What may ﬂow. but which is not subject to a conservation law as in the previous section. ∂N T P ∂S = . ∂E T ∂S µ =− . which is the quantity which ﬂows from one cell to the other. it is convenient to take the entropy as the statistical potential: this potential is deﬁned in terms of extensive quantities and the corresponding intensive quantities can be determined as derivatives of the entropy: 1 ∂S = . is the momentum of the particles in our subvolume. we see that γi (r.99 this course. The energy measured with respect to the system box. t ) d 3 r = S. which will change in the two cells as a result of energy transport. energy and total momentum p. Intensive partners of extensive variables are found by taking the derivative of . we have S(λ Ai ) = λ S(Ai ). γi : γi = ∂S . In more general terms. t ) From this. p) = S(E − p2 /(2m). and its conjugate variable temperature. and the other extensive. Taking the derivative with respect to λ and then setting λ ≡ 1. however. 0 ) pi ui =− =− . we call the conjugate variable of an extensive variable Ai . As the distinction between extensive quantities and their intensive. intensive variable. conjugate partners is quite explicit in the study of nonequilibrium and transport. t )γi (r. Therefore we have: S(E . t ) = Summarizing. As an example we may consider the energy. δS . will however change by an amount p2 /2m. this is ∑ i V ρi (r. ∂ pi ∂E m T where ui is the mean velocity. that with each extensive variable. ∂V T The last relation is irrelevant for our purposes as we do not consider the volume as a ﬂowing quantity. so we must consider the entropy for ﬁxed volume. δ ρi (r. one of which is intensive. ∂ Ai ∑ i For the local quantities. Note that p is an extensive quantity. we have S = ∑ Ai i ∂S = γi Ai .

β for cartesion directions. αβ αβ We need a so-called constitutive equation to relate ε to the temperature.E (r. t ) = −κ ∇ (T ) . we shall always use α . t ) can be related to the gradient of the afﬁnity γ along the cartesian volume: then the current jα i direction α : αβ i.β In this chapter. b)∂β γ j (r. We then obtain ∂ T (r.100 the entropy S with respect to the extensive variable. t ) − jα (r. Note that we do not include particle transport. and together with the expression for the current just derived. The energy density is called ε . j This can be cast into a local form by taking the volumes a and b very small and dividing by the i (r. or afﬁnity. t ) = ∑ Li j (a. where κ is the thermal conductivity. t ) − κ ∇2 [T (r. The conservation law leads to an interesting result. 11. We are usually interested in small deviations from equilibrium. We see that in this case the tensor LEE is diagonal: LEE = κ T 2 δαβ . which we assume to be independent of temperature. t )] = 0.eq i jα (r. c ∂t . The heat current can be directly derived using the results of the last section: 1 αβ jα . We assume that energy is the only ﬂowing quantity – this type of transport is called thermal conduction. which is the difference of the intensive variable in neighbouring cells and the current Φi (a → b): Φi (a → b) = ∑ Li j [γ j (b) − γ j (a)] . where c is the speciﬁc heat (per unit volume). An obvious example is temperature difference which drives the transport of energy. it enters in the conservation equation to give 1 ∂ε + κ T 2∇ · ∇ ∂t T = 0. and in that case we may postulate a linear relation between the driving force. The variation of the intensive variable drives the transport of the extensive variable. j. This is ε = cT . t ) = ∑ LEE ∂β .3 Heat and particle diffusion We can apply the general analysis of the previous sections to the particular examples of particle and heat diffusion. T αβ In an isotropic medium the relation between temperature and heat current is given by the familiar relation jE (r. the heat diffusion is driven by the gradient of the temperature. t ). As already mentioned.

N (r. . t ) = − ∑ LNN ∂β β αβ µ (r. t ) = − Hence we ﬁnd D= LNN 1 ∇n(r. This is driven by µ /T according to the previous section. This solution shows that sharp features (delta functions) decay to smooth shapes in the course of time.e. no energy transport). which involves another relation between µ and n. The current is related to the gradient of this driving force by the linear response relation: jα . This comparison necessitates an additional step. t ). and its deﬁnition is κT = − From the Gibbs-Duhem relation 1 V ∂V . t ) = exp −r2 /(4Dt ) . we may compare this with the familiar expression jN (r. The solution with intial condition T (r.101 This is the famous heat. For particle diffusion in a system with uniform temperature (i. we can focus on the particle transport. This relation is 1 ∂µ = . t = 0) = δ 3 (r) is given as 1 T (r. ∂P Nd µ + SdT − V dP = 0. this compressibility indicates how difﬁcult it is to compress a material. In fact. (4π Dt )3/2 with D = κ /c. T which proves the relation used above. we obtain for the current jN (r. T κT n2 LNN ∂µ = 2 . t ) . ∂ nLNN n κT T The particle conservation equation then leads to a diffusion equation similar to that for the heat transport obtained above. T so that we obtain 1 V =− κT N =n T =n T ∂P ∂µ T ∂µ ∂n = n2 T ∂µ ∂n . it is seen that ∂P ∂µ ∂P ∂ (V /N ) ∂P ∂n = n. ∂ n T κT n2 where κT is the isothermal compressibility. T Just as in the previous section. or diffusion equation. t ) = −D∇n. All in all. known as Fick’s law.

which gives us the change of energy. . isotropic ﬂuid at rest and in equilibrium. . AN ) (Ai − A S(A1 . . aN ) = ∞ γi j −∞ da1 · · · daN exp − ∑i j 2 ai a j where γi j = − ˜ 1. depends on extensive quantities Ai . . for example a box surrounding the centre. A ˜ 2. A ˜N ) 1 ∂ 2 S(A . A ˜ 2. . . as we are in equilibrium. and therefore with heat generation. . Now we imagine a box that we move at uniform velocity along the ﬂuid. . momentum. A2 . If we now imagine again a box in the system. and combining this with the fundamental postulate if statistical mechanics. We have distinguished reversible and non-reversible currents. entropy. Therefore.4 General analysis of linear transport Suppose we have a homogeneous. etcetera. but these currents will no longer be reversible. The fact that the ﬁrst-order term is not included is due to the fact that S was expanded around its maximum – hence. Obviously. A ˜ N ) + 1 ∑ ∂ S(A1 . ˜ 1. which tends to maximise its entropy. the following for the probability of having a state where the quantities ai = Ai − A expression: γi j ai a j exp − ∑i j 2 . Suppose we have an isolated system. or dissipative. entropy will move in. where Ω is the number of states accessible to the system. the enetropy is maximal.and outside. momentum. . but as an exchange does not change the total entropy. The problem of non-reversible. However. The latter are associated with an increase of entropy. . which says that each of these states is equally probable. there will be an exchange of entropy.102 11. The corresponding variation in the entropy can be expanded in a Taylor series: 2 ˜ ˜ ˜ ˜ i)(A j − A ˜ j) + . vanishes. we shall conﬁne ourselves to the case where the system is close enough to equilibrium to justify dropping the higher order terms in the expansion. . We know however that the net ﬂux. the ﬁrst derivatives are all zero. We call A Fluctuations correspond to deviations of the Ai from their equilibrium values. AN ) = S(A 2 i. Recalling that S = kB ln Ω. P(a1 . . . non-dissipative currents. we know that the entropy is not conserved. our imaginary box does not change the physics of the problem. . Now imagine another process in which we place many particles at the centre of a box. In particular. . The particles will diffuse from the centre to ﬁll the box homogeneously (if we are in the gas phase). if we divide the system in two parts. . ∑ kB i. we have again currents of energy. Obviously. we have ˜ i are nonzero. in addition in each box there will be an increase or decrease which is not cancelled by an opposite change in the other box. The fact that the entropy strives to its maximum is the driving force which causes ﬂuctuations to dampen out. the ﬂuid now moves. The analysis proceeds as follows. entropy and so on. j ∂ Ai ∂ A j From now on. . Fluctuations of a system from the equilibrium state can be related to the transport properties of that system. The entropy ˜ i the values for which the entropy assumes its maximum. hence energy and momentum are transported through the walls of our imaginary box. . and within the box the ﬂuid is still in equilibrium. currents is of particular interest from now on. . . j ∂ Ai ∂ A j . All these currents are therefore reversible. Also. We shall now analyse the transport problem in a general way. . and that it will increase in the entire box. . .

In a linear response Ansatz. for i = j we ﬁnd along the same lines: ai ∂S ∂aj = 0. we have the following relation between the γi and the time-derivative of ai : a ˙i = ∑ Li j γ j . Now it is however time to object: above. we used a different picture. and this ﬂow was related to the spatial variation of the afﬁnity [in electrical terms: with the electric ﬁeld = spatial variation of the potential (=afﬁnity)]. and the way to reach this maximum is by changing the values of the ai . in which a current was really associated with the ﬂow of a quantity. The rate of change of ai is often called a current. This is true at the maximum. we ﬁnd ∂S = γi j a j . P ∂ ai kB ∂ ai Taking the average on the left and right hand side with respect to the distribution P we obtain ai ∂S ∂ ai = da1 · · · daN ai ∂S P = kB ∂ ai da1 · · · daN ai ∂P ∂ ai Integrating by parts the integral on the right hand side leads to ai ∂S ∂ ai = −kB . You may object that the ﬁrst derivative of S with respect to the ai is zero.103 Note that the fact that S has a maximum implies that the matrix γi j has positive eigenvalues. We know that the entropy strives towards becoming maximal. This change only stops when ∂ S/∂ ai = γi is zero. From this we can derive a simple expression for correlation functions of the ai . and the afﬁnities are called generalised forces. but near that maximum. The entropy changes in the course of time as ∂S dS = ∑a = a ˙i ˙i γi . ∂ ai ∑ j Using this we rewrite the result above as ∑ γik k ak a j = kB δi j . and hence had a direction. How can we translate the above analysis to the problem of currents? . Note that 1 ∂P 1 ∂S = . j This equation relates the rate of change of ai to the afﬁnities γ j . Moreover. dt ∂ ai ∑ i i where we have used the deﬁnition of the afﬁnity γi (see above).

We see that the linear transport coefﬁcients Li j are found as the time correlation functions of the ﬂuctuations: −1 ai (t + τ )a j (t ) − ai (t )a j (t ) . In that case. ai (t + τ )a j (t ) = ai (t − τ )a j (t ) we see that the transport coefﬁcients must be symmetric: Li j = L ji . If its α component increases. This relation is used a lot by chemists to construct phenomenological systems of equations which describe exchange of thermal energy. t )r d 3 r. We can analyse further the relation between Li j and the correlator. This nontrivial property follows from what is called microscopic reversibility as it reﬂects symmetry properties of microscopic correlation functions which follow from the underlying time reversal symmetry of the microscopic dynamics. Li j = kB τ From this. Therefore. If we take t much larger than the correlation time. but their ﬁrst moments: mi = V ρi (r. Suppose we had in the above derivation not multiplied ai (t + τ ) with a j (t ) but with a j (0). Now we approximate ﬁnite differences by time derivatives: Li j ≈ 1 1 ai (t )a ˙ j (0) = kB kB t 0 a ˙i (t ′ )a ˙ j (0) dt ′ . increases. Now suppose that this moment changes in time. using time-reversibility of the correlation function. we arrive at the result: Li j = −1 kB τ ai (t + τ )a j (0) − ai (t )a j (0) = −1 kB τ ai (t )a j (−τ ) − ai (t )a j (0) where we have used time translation symmetry. This can only happen when there is a net current.104 The key is to not consider the ﬂowing quantities Ai themselves as variables. Hence we see ˙ i . Onsager received the Nobel prize for this formulation of non-equilibrium transport. we see that Li j ≈ 1 kB ∞ 0 a ˙i (t ′ )a ˙ j (0) dt ′ . momentum etcetera. the present analysis directly carries over to currents and their driving forces. . that ji ∝ m Now we consider the expectation value ai (t + τ )a j (t ) = ai (t )a j (t ) + τ a ˙i (t )a j (t ) = ai (t )a j (t ) + τ ∑ Lim m ∂S a j (t ) ∂ am = ai (t )a j (t ) − τ Li j kB . particle species. the slope of A in the increasing direction rα .

However. and light particles scattering off these centres. This implies that LNN ∇ −µ T + LEN ∇ 1 T = 0. The formalism enables us to directly formulate the currents using the transport coefﬁcients: −µ 1 + LEN ∇ . We now turn to a problem in which there is only a single current. If an electric potential can be considered to be more or less constant over the subvolume.5 Coupling of different currents In section 11.3.3. we consider ﬂow of energy and current at the same time. . under the restriction that the coupling constants are symmetric according to microscopic reversibility. or of neutrons off heavy atoms. the total momentum of the light particles changes at the collisions and is absorbed by the scatterers. we have studied linear transport of a single quantity: either the energy or the particle density. t ) = T 2L −1 NN LEE LNN − L2 NE ∇T . Examples of such systems are the scattering of electrons off impurities in a solid. which now is driven by the chemical potential and by en electric ﬁeld. Substituting this into the equation for the energy current. In this section we study the simultaneous occurence of such currents. T T −µ 1 jN = LNN ∇ + LNE ∇ . For the particles. which according to the general theory may be coupled. As a ﬁrst example.105 11. Suppose we have a system consisting of ﬁxed scattering centres. we obtain jE (r. the fact they are allowed to move alters the thermal conductivity 1 κT = 2 LEE LNN − L2 NE T LNN with respect to that found in section 11. which are not included in the currents under study. The scattering is considered to be elastic. n is the number density N /V and s is the entropy density S/V . T T jE = LEE ∇ Microscopic reversibility (the ‘Onsager relation’) tells us that LNE = LEN . ∂n where e is the charge of the particles. We see that. From this we infer that µ = −T ∂s = µ0 + eΦ. the energy levels will be shifted by that potential. even when there is no net ﬂow of particles. Thermal conductivity is the process which takes place when there is no particle current. The ﬂowing quantity is the charge. this effect is indistinguishable from an shift of the chemical potential which in the grand canonical ensemble is a kind of ‘zero point’ energy which is assigned to every particle. so the energy of the light particles does not change. and the current is the familiar electric current. Then we analyse the relation between drift and diffusion within the context of the general formulation.

t ) e2 LNN e e2 LNN ∇Φ(r. at these collisions. ∂n ∂µ T σel /e2 . In a stationary system. ∂n kB T ∂µ = . the density will be uniform. t ) = 0. Note that this is a striking result: the way in which particles can diffuse through a medium determines their behaviour under a driving force completely. consisting of structureless particles (i. We conﬁne ourselves to the simplest case of isotropic ﬂuids. we have ∂µ LNN . t )u(r.e. The archetypical material is liquid argon. t ) = T We see that conductivity σel = e2 LNN /T . conservation of momentum and energy (in addition to the trivial particle conservation: the particles do not undergo chemical reactions). we derive hydrodynamic equations. equal to that found in section 11.3. no electric or magnetic dipoles or charges). ∂t where ρ is the mass density. ∂t . the particles will collide and exchange energy and momentum. These equations describe the ﬂow of a ﬂuid. We see that the current is composed of a diffusive part. and transport phenomena on the other – they go under the name of ﬂuctuation dissipation theorems.6 Derivation of hydrodynamic equations In this section. we have seen that D= so that we have D= For an ideal gas. which is an intrinsic quantity. This can even be put on a more general level: the transport coefﬁcient is related to dissipation (the current generates heat through the resistivity of the material). t ) + E jel (r. ne2 This is an example of an Einstein relation which between the diffusion constant D and a transport coefﬁcient σ . Mass conservation is expressed by the relation ∂ ρ (r. t ) = −e ∇n(r. In terms of u. In section 11. There exists theorems which establish general relations between equilibrium ﬂuctuations on one hand. which leads to 11. We have. whereas the diffusion tells us how much the particle positions ﬂuctuate in equilibrium. and a drift part which is caused by the electric ﬁeld.3. jel (r. t )] = 0. t ) − T T T ∂n T where E is the electric ﬁeld. This then leads to Ohm’s law: e2 LNN E = σel E. During the ﬂow. The current represents the average motion of the particles.106 The current which we calculated in section 11. however. t ) + ∇ · j(r. t ) + ∇ [ρ (r.3 can directly seen to be modiﬁed: LNN ∂ µ0 (r. the mass ﬂow is j = ρ u. and the ﬁrst term vanishes. so we have ∂ ρ (r. We call the average velocity in a small subvolume u. ∂n n D = kB T σel 1 . t ) = − LNN ∇µ0 (r.

P is a tensor containing the viscous forces which change the momentum. Now consider a small volume V within the ﬂuid.e. momentum and energy. which. i. Therefore we have S ρ uα ∑ uβ daβ = β V ∂β ρ uα uβ d 3 r = F. the temperature does not vary in space. which are called viscosities. We ﬁrst make the assumption that our ﬂuid is isothermal.107 Working out the gradient. Therefore. ∂t β β β The quantity ρ uα uβ is the ﬂow along the Cartesian direction β of the α -component of the momentum. The effect of momentum ﬂow across the boundary is called the pressure. Usually. that is. the ﬁrst of these is replaced by the linear combination 1 1 ∂α uβ + ∂β uα − δαβ (∇ · u). we obtain ρ ∂ uα − uα [ρ ∇ · u + (u · ∇)ρ ]+ ∑ ∂β Pαβ = 0. ∂t Now we consider the momentum ﬂow. The conservation of the α -component of the momentum is expressed by ∂ ρ uα + ∑ ∂β Pαβ = 0. Furthermore we assume that the main contribution to the momentum is due to its own afﬁnity. we see that ∂ρ + ρ ∇ · u + (u · ∇)ρ = 0. ∂t β After some manipulation. that is. where S is the surface bounding V . the second and third term of this equation can be rewritten to arrive at ρ ∂ uα + ρ ∑(uβ ∂β )uα − ∑ ∂β (ρ uα uβ ) + ∑ ∂β Pαβ = 0. T and uα . we should construct Pαβ such that it be isotropic and symmetric in α and β . in equilibrium. by µ . This leads to the two possibilities δαβ (∇ · u) and ∂α uβ + ∂β uα . The ﬂow of momentum across the boundaries of this small volume determines the rate of change of the momentum inside this volume. Finally. has the form ρ uα uβ = Pδαβ where P is the scalar pressure for the isotropic ﬂuid. d a is a outward normal vector to the surface and we have used the divergence theorem to get the second experession. 2 3 . this must be driven by the afﬁnity of the particle number. But the rate of change of the total momentum is the net force acting on the volume. Based on the general theory. to uα . ρ uα uβ is called the pressure tensor. We see that the term occurring in our momentum conservation equation is simply the force. ∂t β Here. The momentum density is ρ u. Both have their own transport coefﬁcient. Writing out the ﬁrst term and using the mass conservation law. Now we are left with the viscous tensor Pαβ .

using a similar analysis.108 Now we can write down the conservation equations for the momentum: ∂u η 1 η 1 + (u · ∇)u + ∇P = ∇2 u + + ζ ∇ (∇ · u) . we can put P equal to ρ kB T so that this turns into a closed equation. The ﬂow of entropy and energy can be expressed in a separate equation. . In good approximation. ∂t ρ ρ ρ 3 This is the Navier-Stokes equation for the momentum ﬂow.

The meaning of these quantities is related to what happens between two collision events: the average time between two collisions is the free ﬂight time. from studying the ﬂuctuations in an equilibrium system. but if we wait long enough. it will take some time before equilibrium is restored. The main message of this chapter is that these three types of motion are strongly related. Initially.e. systems in which the particles move freely most of the time and experience collisions with their counterparts every now and then. which are related according to l = |v| τ . where |v| is the average absolute velocity. in a small subvolume of the liquid. τ . This is called the drift.1 Motion of particles Particles move under the inﬂuence of their counterparts. we have restricted ourselves to equilibrium phenomena. a local equilibrium will be realised. Therefore. In this chapter we shall concentrate on ﬂuctuations from the equilibrium phase and on transport phenomena.12 Fluctuations and transport phenomena Up to this moment. and which is the result of thermal ﬂuctuations. these velocities become equal. and the average distance a particle travels during that time is the mean free path. We now focus on dilute systems. In this context it is useful to speak of the mean free path. • A drag induced by the interaction of a particular particle with the other particles. The motion of a particle can be split into three contributions: • A motion caused by an external force acting on the particles. where u is the average velocity (also called the wind velocity) of the subvolume. and of the free ﬂight time. 109 . This means for example that if we stir a liquid. If a system is moved out of equilibrium. the particles will be distributed according to the distribution: P(v) = exp −m(v − u)2 /(2kB T )) . • A motion on top of the drift. This motion occurs also in equilibrium. l . 12. The latter are intrinsically outside of the domain of equilibrium phenomena: transport does – on average – not occur in equilibrium. i. Two neighbouring subvolumes will have different average velocities. which are clearly the non-equilibrium properties of the system. we can deduce the transport properties.

From this we infer directly that τ= t 1 t = . We shall derive the diffusion equation for one dimension. t )] ≈ α a2 ∂ 2 ρ (x. which is caused by thermal ﬂuctuations kicking the particles in arbitrary directions. their density. The average number of encounters is then given by the volume of this tube times the particle density – see ﬁgure 12. based on a Maxwell velocity distribution. t )dx of ﬁnding the particle at time t in the interval dx located at x. but the average relative velocity between the particles. First we focus on diffusion. The probability density satisﬁes the following equation: ρ (x. We want to evaluate the probability ρ (x. If we carry out the average for relative velocities. and we should correct for this. t + h) − ρ (x.1: The volume a particle ‘sees’ when travelling through the system. ∂ x2 . |vrel | is not the average velocity. = N |vrel | t σscat n |vrel | σscat n In this formula. A particle is placed on the x-axis and performs a step in some random direction at regular time intervals. and their average velocity. The free ﬂight time can be calculated if we know the scattering cross section of the particles in the system.1 In agreement with what has been said in the previous section.1. To that end. t ) + ρ (x − a. t ) = α [ρ (x + a.1. is called α . and the particle will remain at its position with probability 1 − 2α .110 σ Figure 12. We take the step size equal to a and the time step equal to h. The probability that at each time step a jump to the left or right is made. 2 |v| σscat n Diffusion 12. we set up a Master equation similar to that introduced in connection with the Monte Carlo method. A particle ‘sees’ other particles within a tube of cross section equal to the scattering cross section σscat and of length |v| t . t ) − 2ρ (x. we can distinguish two mechnisms for transport of particles through a gas or liquid: diffusion. we ﬁnd √ |vrel | = 2 |v| and we have τ=√ 1 . t ) . and drift caused by an external force.

111 where we have taken the small-a limit in the last expression. we have D=α l2 . we may approximate the left hand side by the time derivative in order to obtain: ∂ ρ (x. t ) − ρ (x + a. It is obvious that the width should increase in time as the position of the particle should become more and more uncertain in the course of time. t ). in 3D this becomes 6Dt . we ﬁnd for the ﬂux: This relation is known as Fick’s law of diffusion. t ) = 0. τ We can also calculate the ﬂux. with a width which grows proportional to t . which can be seen to take on the value a2 . we have 1 −r 2 /(4Dt ) ρ (r. t ) = √ e−x /(4Dt ) . ρ (x. which seems to be the natural choice for this case. t ) =D ∂t ∂ x2 where D is the diffusion constant. h h ∂x ∂x which can be generalised in 3D to J = −D∇ρ (r. ∂t The solution to the diffusion equation with initial condition that there is a particle at the origin at t = 0. which is the net number of particles moving from one position to its neighbouring position. t ) + ∇ (D(r)∇) ρ (r. t ) a2 ∂ ρ ∂ρ ≈ −α = −D . is 1 2 ρ (x. as an extra factor a would have to be included in both the left. The ﬂow to the right is given by D= α aρ (x. J = αa . t ) and that to the left by α aρ (x + a. Keeping the same convention in the derivation of the diffusion equation does not alter the result. (4π Dt )3/2 We can calculate the average square displacement of a particle in one dimension (the average displacement is obviously zero because of symmetry): (∆x)2 = 2Dt . t ).and right hand side. If the diffusion constant depends on position. Therefore. h By interpreting a as the mean free path and h as the free ﬂight time. 4π Dt √ The shape of this distribution is Gaussian at all times. t ) ∂ 2 ρ (x. and taking α = 1. the diffusion equation reads ∂ ρ (r. The factors a in front of the ρ ’s in these equation come from the fact that ρ is in fact deﬁned as the number of particles per unit length. In 3D. For t = 0 this reduces to a delta-function. t ) = e . Taking also h small.

1. v dz . v where ε (z) is the average energy of a particle at height z. we obtain Pxz = which directly leads to dux (z) 1 . we ﬁnd for the viscosity 1 η = nm v ℓ. To ﬁx the ideas. which is realised by putting the system in between two planes perpendicular to the z-axis. which are kept at different temperatures.2 Thermal conductivity In order to study the transport of thermal energy. jE = − n v ℓ 3 dz Now we write ε = cT where c is the speciﬁc heat per molecule. the particles have an energy which is the average energy for the height z − ℓvz /v. We want to study the transport of momentum px across that plane. px = mux z − ℓ v Inserting this into the above expression for Pxz . coming either from above or from below the plane. after some manipulation: d ε (z) 1 . 3 m ux (z) − vz dux (z) n(v)vz d 3 v. the thermal conductivity is given by κT = 12. jE = n(v)vz ε (z) − ℓ v dz Now we may substitute the Maxwell distribution for n(v) to obtain. jE = − n v ℓc 3 dz and therefore. The average height at which these particles last collided was vz h=ℓ . Then we ﬁnd for the energy ﬂux dT (z) 1 . Similar to the previous section. Now let us consider a plane at height z. we note that the average momentum of the particles moving across. but with a temperature gradient in the z-axis. is vz .3 ncl v . Pxz = − nm v ℓ 3 dz Therefore. Performing a Taylor expansion for ε around z gives vz d ε (z) 3 d v. we consider a system like in the previous subsection. v After this last collision. T1 and T2 .112 12.1. Particles will cross this plane. This is given by Pxz = px vz n(v)d 3 v. 3 Viscosity Now we consider the transport of momentum tangential to a plane across that plane. imagine a constant z plane. The ﬂux of energy through the plane at height z is therefore given by vz 3 jE = n(v)vz ε z − ℓ d v.

t ) + F · ∇v f (r. v. This means that we must replace the total collision section by the differential one. t ) + v · ∇r f (r. t + ∆t ) = m f (r. Now we need a more reﬁned expression. t + ∆t ) = f (r + v∆t . which does not look too unfamiliar. n|v|σ . in which the particles collide every now and then. 1 . The Boltzmann equation describes the motion of a collection of particles in a hydrodynamic cell: that is. v. using F = ∇rV (r) allows for a solution f (r. v. More speciﬁcally. or a collision may change the velocity v to some other velocity. In that case we have v · ∇r f (r. However. v(t + ∆t ). keeps the volume element d 3 r d 3 v constant. t ). The interactions between the particles must also be taken into account. v) = exp −β mv2 + V (r) 2 . First we consider the question how f changes with time if we disregard the collisions. as the particles move independently of each other. t )d 3 r d 3 v = number of particles within d 3 r and velocity within d 3 v. The change of position and speed has a direct effect on f . However. which depends on the difference between the in. t ) + which. we consider the particles as moving freely for some average time τ and then colliding with each other. If we consider the dilute limit as in the previous section. There are two issues which we have to consider: ﬁrst the particles move in space due to the fact that they have a velocity and they change their velocity as a result of some (external) force which acts on them.and outgoing velocity. This quantity is called the distribution function f (r. the volume elements d 3 r and d 3 v may also change in time. a cell which contains many molecules but which is much smaller than the scale over which hydrodynamic quantities vary appreciably. m ∂ F · ∇v f (r. v. t ) + f (r. For the frequency of occurrence of a collision we have derived τ= where σ is the total collision cross section. t ): a collision at r may have a particle with velocity v as an end product. and not explicitly on time. This is done in the dilute limit.113 12. Therefore we have f (r(t + ∆t ). This is a more formal and consistent formulation of the ideas treated in the previous section. v(t ) + F ∆t .2 The Boltzmann equation We now turn to a central topic in nonequilibrium statistical mechanics: the Boltzmann equation. v. which includes the in. f only depends on r and v. by Liouville’s theorem. The collisions will result in a loss and an increase of the distribution function f (r. v. each particle is subject to a Hamiltonian evolution which. t ): f (r. t ) = 0. m ∂t If we are in equilibrium.and outgoing angle. The central quantity is the number of particles in such a cell of size d 3 r and velocity within the volume d 3 v in velocity-space. v. v. v. Do not confuse this force with the interactions between the particles: we save those for the collisions which are not taken into account til further notice.

t )|v − v2 |P(v. v2 . v2 ) vanishes unless the velocities satisfy momentum and energy conservation. v1 . v2 . v. Hence. v2 . • P should respect the general conservation laws for momentum and energy. we also may write ′ ′ ′ P(v1 . v2 ). v2 ) = P(−v1 . v. v2 . the loss term due to the collisions can be written as I− (v) = ∆t f (r. 2 ′ ′ v2 1 + v2 = v 1 + v 2 . t ) due to collisions occurring in a small time interval ∆t . v2 . v1 .114 if we travel along with one of the incoming particles. if the particles all have the same mass m: ′ v1 + v2 = v′ 1 + v2 . v′ 1 . v2 . As the probability distribution is in general symmetric under reversal of all velocities (space inversion symmetry). v2 . A similar analysis as for the loss term results in a gain term I+ = ∆t 3 3 3 ′ f (r. The probability that two such particles meet in this time interval is given by f (r. t ) f (r. Therefore. v2 . and using momentum conservation. −v2 . We ﬁrst analyse the loss of the distribution f (r. t ) due to the collisions results from collisions of particles with any velocities v1 and v2 which results in one of the outgoing particles having end velocity v. v′ 1 . v′ 2 ) d v1 d v2 d v2 . v2 . v2 ) = P(v1 . v′ 1 . this particles is a target at rest and we see the other particle hitting it. v′ 1 . 2 2 ′ Therefore the transition probability density P(v1 . we may reformulate energy conservation as ′ |v1 − v2 | = v′ 1 − v2 . v′ 1 . 2 2 2 2(v2 1 + v2 ) = (v1 + v2 ) + (v1 − v2 ) . v2 ) d v2 d v1 d v2 . t )|v − v2 |∆t . v2 ) which gives us the probability density that given two particles which collide with incoming velocities v1 and v2 . v. ϕ ) are the polar angles of the difference between in. v2 . This loss is caused by collisions with particles with a velocity v2 at r. and Using ′ which also holds for v′ 1 and v2 . t ) f (r. −v2 ). The gain term to f r. t )|v1 − v2 |P(v1 . More generally. v. t ) ′ 3 3 ′ 3 ′ f (r. This probability density must satisfy several requirements: • Time reversal symmetry: ′ ′ ′ P(v1 . the result is outgoing particles with velocities v′ 1 and v′ 2 .and outgoing velocity of the incoming particle. Ω = (ϑ . −v1 . . we introduce a transition amplitude ′ P(v1 . v.

115 To be able to combine this expression conveniently with I− we use time reversal symmetry to swap the arguments of the collision distribution P and rename the integration variables to obtain I+ = ∆t
′ ′ ′ 3 ′ 3 ′ 3 f (r, v′ 1 , t ) f (r, v2 , t )|v − v2 |P(v, v2 ; v1 , v2 ) d v1 d v2 d v2 .

Combining the loss and gain term, we arrive at the ﬁnal form of the Boltzmann equation, including collisions: ∂ f + v · ∇r f = Icoll ( f ), ∂t with Icoll =
′ ′ ′ 3 3 ′ 3 ′ |v − v2 |P(v, v2 ; v′ 1 , v2 ) f (r, v1 , t ) f (r, v2 , t ) − f (r, v, t ) f (r, v2 , t ) d v2 d v1 d v2 .

It is important to realise that in this derivation, we have assumed that the probability for two particles with velocity v1 and v2 , is given by the product of the single particle distribution functions: Prob(r, v1 , v2 , t ) = f (r, v1 , t ) f (r, v2 , t ). This implies a neglect of velocity correlations: a probability depending on v1 and v2 which cannot be written as the above product is excluded. This implicit assumption goes by the name molecular chaos. Note furthermore that the collision term does not affect the number density – it only inﬂuences the velocity distribution. We end this section by writing up the Boltzmann transport equation, which includes the collision term derived above: F ∂ ∂ ∂ f (r, v, t )+ v · f (r, v, t )+ · f (r, v, t ) = ∂t ∂r m ∂v

′ ′ ′ 3 |v − v2 | P(v, v2 ; v′ 1 , v2 ) f (r, v1 , t ) f (r, v2 , t ) − f (r, v, t ) f (r, v2 , t ) d v

12.3 Equilibrium – deviation from equilibrium
For equilibrium we now that, in the absence of the collision term, the local distribution function is the Boltzmann distribution. If we add the collision term, it should not affect the equilibrium distribution. That this is indeed the case follows from the fact that
′ f (r, v′ 1 , t ) f (r, v2 , t ) − f (r, v, t ) f (r, v2 , t ) = 2

n2 (r) exp −m v′ 1 + v′ 2 /(2kB T ) − exp −m v2 + v2 2 /(2kB T )

2

= 0,

where the last equality follows from energy conservation at the collision. If we deviate from equilibrium, the collisions should drive us back to equilibrium. This process is expressed in terms of a new quantity H , which is commonly refered to as the Boltzmann function.1 The Boltzmann function is deﬁned as H (t ) = f (r, v, t ) ln f (r, v, t )d 3 r d 3 v.

It is clear that this quantity is related to the entropy by H = −kB S.

1 Usually

people view H as a character of the latin alfabet. Boltzmann however used the Greek capital form of eta (η ).

116 The quantity H is a function of time only, so we may calculate dH /dt . We can now evaluate dH = dt

∂f (1 + ln f ) d 3 r d 3 v. ∂t

From now on we shall use the obvious abbreviations: f (r, v1 , t ) ≡ f1 ; and so on. Furthermore
′ ′ ′ ˜ |v1 − v2 | P(v1 , v2 → v′ 1 , v2 ) ≡ P(1, 2 → 1 , 2 ).

f (r, v′ 1 , t ) ≡ f1′

We then have, after substituting Boltzmann’s equation in the equation for dH /dt : dH =− dt ˙ · ∇r f1 [1 + ln f1 ] d 3 r d 3 p1 + r ˜ (1, 2 → 1′ , 2′ ) ( f1′ f2′ − f1 f2 ) (1 + ln f1 ) d 3 rDv; P

3 ′ where Dv stands for d 3 v1 d 3 v2 d 3 v′ 1 d v2 . Integrating the ﬁrst term and assuming that f vanishes if we are far away (outside the volume), we keep only the second integral. By using symmetry under exchanging coordinates 1 and 2, we obtain

dH = dt

˜ (1, 2, → 1′ , 2′ ) ( f1′ f2′ − f1 f2 ) (1 + ln f2 ) d 3 r Dv. P

We can add the two last expressions for the time derivative to obtain dH = dt ˜ (1, 2, → 1′ , 2′ ) ( f1′ f2′ − f1 f2 ) 1 + 1 ln f1 f2 d 3 r Dv. P 2

As a ﬁnal step, we use the time reversal symmetry property according to which the swap 1, 2 ↔ 1′ , 2′ should not change the integral. This leads to another expression for the time derivative of the Boltzmann function: dH = dt ˜ (1, 2, → 1′ , 2′ ) ( f1 f2 − f1′ f2′ ) 1 + 1 ln f1′ f2′ d 3 r Dv. P 2

Adding this new form to the old one leads to dH = dt ˜ (1, 2, → 1′ , 2′ ) ( f1′ f2′ − f1 f2 ) 1 ln f1 f2 d 3 r Dv ≤ 0. P 4 f1′ f2′

The inequality on the right hand side follows from the fact that (y − x) ln(x/y) is always less than or equal to zero. We see that H monotonically decreases. Furthermore, H is a positive number. Therefore, H will decrease in time until it has reached its minimum value. This value corresponds to the equilibrium state. So what does this state look like? It is reached when f1 f2 = f1′ f2′ , or, ln f1 + ln f2 = ln f1′ + ln f2′ .

117 In a homogeneous system, the equilibrium distribution does not depend on r, and the last condition for equilibrium must be satisﬁed for the f ’s being functions of the momentum coordinate. As we know that momentum and energy are conserved during collisions, this requirement can be satisﬁed for any function f of the form p2 ln f (p) = A + b · p + C , 2m in other words p2 f (p) = exp A + b · p + C . 2m This is the general form of the Maxwell distribution of a gas with a nonzero total momentum.

12.4 Derivation of the Navier–Stokes equations
In this section we present a derivation of the Navier–Stokes equations from an approximate Boltzmann equation through a Chapman–Enskog procedure. If the particles would simply ﬂow according to their initial velocity, without interaction, equilibrium would never be reached: the role of the collisions is to establish local equilibrium, that is, a distribution which is in equilibrium in a small cell with ﬁxed volume, constant temperature, density and average velocity u. We know this equilibrium distribution; it was derived in the previous section: f eq (r, v) = f (r) exp −m(v − u)2 /(2kB T ) , (12.1)

which holds for cells small enough to justify a constant potential. We have neglected external forces which would change the velocities for simplicity – they can be included straightforwardly. Once the liquid is in (local) equilibrium, the collisions will not push it away from equilibrium. It can be shown that the collisions have the effect of increasing the entropy – hence they generate heat. Before we continue, we note that the mass must always be conserved, whether there are collisions or not. The mass density is found as

ρ (r, t ) =

m f (r, v, t ) d 3 v.

(12.2)

Its evolution can be calculated by integrating the Boltzmann equation, multiplied by the single particle mass m, over the velocity:

∂ ρ (r, t ) + ∂t

mv · ∇r f (r, v, t ) d 3 v =

m

df dt

d 3 v.
collisions

(12.3)

The second term of this equation can be written as ∇ · j(r, t ) where j denotes the mass ﬂux, or momentum density, of the ﬂuid: j(r, t ) = vm f (r, v, t )d 3 v = ρ u, (12.4)

where u is the average local velocity. The collisions change the velocity distribution, but not the mass density of the particles – hence the right hand side of (12.3) vanishes and we obtain the familiar continuity equation: ∂ ρ (r, t ) + ∇ · j(r, t ) = 0. (12.5) ∂t Another interesting equation describes the conservation of momentum. We would like to know how j(r, t ) changes with time. This is again evaluated straightforwardly by multiplying the Boltzmann

118 equation by v and integrate over the velocity. Using the indices α and β for the Cartesian coordinates, we obtain ∂ jα df + mvα ∑ vβ ∂β f (r, v, t )d 3 v = mvα d 3 v, (12.6) ∂t dt collisions β where ∂β denotes a derivative with respect to the coordinate rβ . For the right hand side, a similar statement can be made as for the equivalent term in the mass equation: although individual particles involved in a collision change their momenta, the total momentum is conserved at the collisions. After thus putting the right hand side to zero, we write (12.6) in short hand notation as

∂ jα + ∂β Pαβ = 0, ∂t
where we have introduced the momentum ﬂow tensor Pαβ = mvα vβ f (r, v, t )d 3 v,

(12.7)

(12.8)

and where we have used the Einstein summation convention in which repeated indices (in this case β ) are summed over. The derivative with respect to rβ is in our notation denoted by ∂β . Assuming that we are in equilibrium, we can evaluate the momentum tensor by substituting for f (r, v, t ) the form (12.1): Pαβ =
eq

mvα vβ n(r) exp −m(v − u)2 /(2kB T ) d 3 v = ρ (r) kB T δαβ + uα uβ .

(12.9)

This result can be derived by separately considering α = β and α = β , and working out the appropriate Gaussian integrals. Noting that ρ kB T equals the pressure P,1 we arrive at the following two equations:

∂ ρ (r, t ) + ∇ · j(r, t ) = 0 (mass conservation); ∂t ∂ (ρ u) + ∇r · (PI + ρ uu) = 0 (momentum conservation). ∂t
Using the ﬁrst equation, we can rewrite the second as

(12.10a) (12.10b)

∂ u(r, t ) 1 + [u(r, t ) · ∇r ] u(r, t ) = − ∇r P(r, t ). ∂t ρ (r, t )

(12.11)

The equations (12.10a) and (12.10b) or (12.11) are the Euler equations for a ﬂuid in equilibrium. When the ﬂuid is not everywhere in local equilibrium, the collisions will drive the system towards equilibrium – hence their effect can no longer be neglected. As mentioned above, the additional currents which arise on top of the equilibrium ones increase the entropy and are therefore called dissipative. Hence these terms describe the viscous effects in the ﬂuid. We now split the distribution function into an equilibrium and a nonequilibrium part: f (r, v, t ) = f eq (r, v) + f noneq (r, v, t ). The equilibrium term satisﬁes (12.1).
1 Here,

(12.12)

we consider the ﬂuid as an ideal gas; a realistic equation of state may be used instead.

Before noneq has an important property: its we proceed to work out (12.17) For the calculation of the nonequilibrium stress tensor.16) Realizing that this expression represents the change in the average kinetic energy due to the collisions. To ﬁnd the lowest-order contribution to a systematic expansion of the density. t )d 3 v. t ) − f eq (r. we see that the collision effects enter explicitly in this momentum ﬂux. In this second equation. If we consider the evolution of this ﬂux using the Boltzmann equation. which we have calculated above assuming equilibrium.18c) . Note that if we integrate this equation over the velocity. (12. v. the right hand side vanishes as the collisions do not affect the mass density.14): Pαβ noneq = mvα vβ nnoneq d 3 v = −τ mvα vβ ∂ f eq 3 d v+ ∂t mvα vβ v · ∇r f eq d 3 v . m(vα − uα )(vβ − uβ ) f eq (r. This is calculated from the density f (r. This can be seen by writing out this trace: noneq = ∑ Pαα α v2 f noneq (r.119 How can we represent the effect of the collision term? There is an approach due to Maxwell. It can be shown that this is an expansion in the parameter ℓ/L. t ) dt =− f noneq f (r. v) f noneq (r. The momentum ﬂux is deﬁned in (12. or are dominated by a single. v. and the nonequilibrium part will now be calculated using (12.15). In that case: d f (r. v)d 3 v = ρ 1 u ˙α = − ∑ uβ ∂β uα − (∂α P). v) =− . we note that the tensor Pαβ trace vanishes.15) further. The mass ﬂux can still be written as ρ u. t ) and it can therefore be split into an equilibrium and nonequilibrium part.8). relaxation time τ . v. v. we immediately see that it vanishes as the (instantaneous) collisions leave the total energy invariant: TrP noneq = 0. (12. v.18b) (12. the momentum ﬂux Pαβ occurs. the collisions do not change the mass conservation equation.14) + v · ∇r f eq = − ∂t τ This is an explicit equation for the nonequilibrium term. m αβ (12. (12. which can easily be seen to hold for the equilibrium distribution: m f eq (r.18a) (12. which should always be valid. The equilibrium part was calculated in Eq. where ℓ is the mean free path.15) where we have again used the notation ∂α for a derivative with respect to the α -component of r. (12. v)d 3 v = ρ (r). τ τ (12. Eq.13) collisions As mentioned above. which is based on the assumption that all relaxation processes have the same. as we shall see. the collisions leave the total momentum unchanged. (12. Moreover. and L is the typical length scale over which the hydrodynamic quantities vary. t ) . ρ β kB T δ = Pδαβ .9). The ﬂux j occurring in the mass conservation equation also occurs in the momentum conservation equation. The equation for the ﬂux will however acquire a contribution from the nonequilibrium part of the distribution function. we replace n on the left hand side of the Boltzmann equation by its equilibrium version: ∂ f eq (r. (12. we use the following equations.

(12. (12.15) can be written. we obtain. this can be written as eq ∂t Pαβ = ∂t Pδαβ + ρ uα uβ = ˙ δαβ − ∑ ∂γ (ρ uγ )uα uβ + ρ uα uγ (∂γ uβ ) + ρ uβ uγ (∂γ uα ) − uβ ∂α P − uα ∂β P. This tensor must therefore be 2 P noneq = −Pτ ∂α uβ + ∂β uα − δαβ ∂γ uγ . This can only happen when the trace occurring in the last equation cancels the trace of the remaining terms in the expression for P noneq . using Eqs.21)]: P(∂β uα + ∂α uβ ) + δαβ The terms ˙ + ∑ uγ (∂γ P) + P∂γ uγ P γ . (12.21) Adding the two terms of Eq. (12. v) d 3 v = ∂γ uα uβ uγ + uα Pδβ γ + uβ Pδβ γ + uγ Pδαβ . When we write this term out.19) and (12. (12.120 where in the last equation it is understood that the velocities are those evaluted for the equilibrium eq distribution: this equation is the Euler equation. (12. We ﬁrst work out the ﬁrst term in the square brackets on the right hand side in (12. again with wα = vα − uα : ∂ ∂t mw2 f d 3 v + ∑ uγ ∂γ γ mw2 f d 3 v = mw2 ∂f + uγ ∂γ n ∂t ∑ γ d3v = 1 τ mw2 f noneq d 3 v. many terms occuring in the last equations cancel – the ones that remain are [(12.25) mwα wβ f noneq d 3 v.18b) and the equilibrium distribution. (12.20) The second term can now be worked out and yields ∑ γ uα uβ uγ ∂γ ρ + ρ uβ uγ (∂γ uα ) + ρ uα uγ (∂γ uβ ) + ρ uα uβ (∂γ uγ )+ ∂γ Puγ δαβ + ∂γ (Puα )δβ γ + ∂γ (Puβ )δαγ .10a) and (12. using the quantity wα = vα − uα . τ Now we use the fact that TrP noneq vanishes.23) can be calculated using (12.22) ˙ + ∑ uγ (∂γ P) P γ (12.19) P γ The second term in the square brackets of (12.24) 1 (12.15).26) 3 This is the trace of the tensor .15). After some manipulation.9). (12.11) which can also be written as ∂t Pαβ (we use ∂t to denote a partial derivative with respect to time).10b). (12.18b)]: (uα + wα )(uβ + wβ )(uγ + wγ )∂γ neq (r.9) and (12. (12. in the form [see also (12.

as 1 1 ∂u + u · ∇u = ∇P + ν ∇2 u + ν ∇(∇ · u). which may be ρ = constant for an incompressible ﬂuid. We need therefore an additional equation. Note that the case where ρ = const also implies ∇ · u = 0 from the continuity equation. or P ∝ ρ for the isothermal case.121 Using this. u and P. ∂t ρ 3 (12.27) The mass conservation equation and the momentum conservation equation together are insufﬁcient to give us the four unknown ﬁeld: ρ . we can formulate the momentum conservation equation. with ν = τ kB T /m. which in turn causes the last term in the last equation to become negligible. .

Consider a solution containing polymers or ions which are much heavier than the solvent molecules. As the kinetic energy is on average divided equally over the degrees of freedom. The motion of ﬂuid particles exhibits strong time correlations and therefore the effects of their collisions should show time correlation effects. whereas in reality it executes a Brownian motion. they will change their momenta only after many collisions with the solvent molecules and the picture which emerges is that of the heavy particles forming a system with a much longer time scale than the solvent molecules. Time correlations affect the form of the friction term which. (13. gravitation. In the following we shall restrict ourselves to a particle in one dimension. In order to avoid complications we shall proceed with the simpler form (13.1 Langevin theory of Brownian motion In this section we consider the Langevin equation. Equation (13.1).13 Nonequilibrium statistical mechanics 13. We then arrive at the following equation: m dv (t ) = −γ v(t ) + F (t ) + R(t ) dt 122 (13. Moreover.3) .2) This form of the equation must also hold for lighter particles. in Eq. (13. the analysis for more particles in two or three dimensions is similar. This process can be carried out analytically through a projection procedure but here we shall sketch the method in a heuristic way.1).1) (t ) = −γ v(t ) + F(t ) dt where γ is the friction coefﬁcient and F the external or systematic force. How can we model the effect of the solvent particles without taking into account their degrees of freedom explicitly? When a heavy particle is moving through the solvent. the ions or polymers will move much more slowly than the solvent molecules. etc. because of their large mass. which describes Brownina motion on a microscopic level. walls.1) has the unrealistic effect that if the external forces are absent the heavy particle comes to rest. To make the model more realistic we must include the rapid variations in the force due to the frequent collisions with solvent particles on top of the coarse-grained friction force. This difference in time scale can be employed to eliminate the details of the degrees of freedom of the solvent particles and represent their effect by forces that can be treated in a simple way. due to the other heavy particles. This suggests the following equation of motion for the heavy particle: dv (13. the collisions with the solvent particles will on average have the effect of a friction force proportional and opposite to the velocity of the heavy particle. Therefore. has been taken dependent on the instantaneous velocity but which in a more careful treatment should include contributions from the velocity at previous times through a memory kernel: m m dv (t ) = − dt t −∞ dt ′ γ (t − t ′ )v(t ′ ) + F(t ). it will encounter more solvent particles in the front than in the back.

9) ∆t For the continuum case ∆t → 0 (13.123 where R(t ) is a ‘random force’.6) (13.9) converges to the δ -distribution function R(t )R(t + τ ) = qδ (τ ). so we arrive at q Rn Rm = δnm . the expectation value of the random force should vanish: R(t ) = 0. dRm exp − 21q ∑m l =n Rl ∆t Rn Rm 2 dRn dRn+1 .3) we ﬁnd [v(t )]2 = v2 0 exp(−2γ t /m) + q (1 − e−2γ t /m).8) which yields the value 0 for n = m. . but they are neglected once more and the force is subject to the following conditions. In the discretised time case. Using (13. Again. (13.13) . For n = m we ﬁnd the value q/∆t . (13. dRm exp − 21q ∑m l =n Rl ∆t (13. the correlation function for the Rn reads Rn Rm = 2 dRn dRn+1 . in accordance with the previous assumptions.4) • The values of R are taken to be uncorrelated: R(t )R(t + τ ) = 0 for τ > 0. we may assume that the random force is constant over each time step: at step n.3) with F (t ) ≡ 0. the time correlations present in the ﬂuid should show up in this force.10) We now return to the continuum form of the Langevin equation (13. the value of the random force is Rn .7) with q a constant to be determined. All these above assumptions can then be summarised in the following prescription for the probability for a set of random forces to occur between t0 and t1 : P[Ri (t )]t0 <t <t1 ∼ exp − 1 2q t1 t0 dt R2 i (t ) (13. (13. (13. The expectation value of v2 is determined in a similar way.12) which is to be expected for a particle subject to a friction force proportional and opposite to the velocity. • The values of R are distributed according to a Gaussian: P[R(t )] = (2P R2 )−1/2 exp(−R2 /2 R2 ). 2γ m (13.11) Because the expectation value of R vanishes we obtain v(t ) = v(0) exp(−γ t /m) (13. • As the average effect of the collisions is already absorbed in the friction. .10) and (13. For this case. This can be solved analytically and the result is v(t ) = v(0) exp(−γ t /m) + 1 m t 0 exp [−(t − τ )γ /m] R(τ )d τ . . .5) Now let us discretise time. (13.

the random force is no longer uncorrelated – it is constructed with correlations in accordance with the ﬂuctuation-dissipation theorem: R(0)R(t ) = v2 γ (t ). is F v=− . We now establish an equivalent of the diffusion equation which includes the effect of a driving force. Such correlation effects are often neglected and the systematic force is simply added to the friction and the Langevin term. In that case.16) so this equation deﬁnes the value of q necessary to obtain a system with temperature T . The results presented here are easily generalised to more than one dimension.18) 13. (13. the same will hold for the velocity – the width is given by (13.14).14) According to (13.11). in which the random force is neglected.17) The absence of a long time tail in this correlation function reﬂects the oversimpliﬁcations in the construction of the Langevin equation. and the friction counteracts this and in the end there is a balance between that friction and the force.124 which for large t reduces to [v(∞)]2 = q . 2γ m (13. this equation is again no longer valid if external forces are included. In section 12. This is precisely the Maxwell distribution if we write q = 2kB T γ . However.2). .15) exp[−mv(t )2 γ /q] P[v(t )] = Pq for large t . (13. using the Fokker-Planck equation. v depends linearly on the random forces R(t ) and as the latter are distributed according to a Gaussian. This force may be derived from a stationary potential. the force tries to increase the velocity along the force direction. in particular the absence of correlations in the random force and the fact that the frictional force does not depend on the ‘history’ of the system.11): v(0)v(t ) = v(0)2 e−γ t /m . A further reﬁnement is the inclusion of memory kernels in the forces. The velocity autocorrelation function can also be obtained from (13. The stationary solution for the velocity equation.4 we shall discuss Langevin types of equations in a more formal way.2 Fokker Planck equation and restoration of equilibrium In the previous section we have formulated an equation for a single particle which diffuses and which feels a force F which tries to establish some distribution which differs from the homogeneous one. (13. γ that is. (13. However. (13. The ﬂux due to diffusion is jdiff = −D∇ρ (r. including a force acting between the heavy particles causes problems if this force exhibits correlations with the random force. and Eq. t ). so we have γ m 1/2 (13.16) is no longer valid in that case. similar to the approach in Eq.

as is clear from the fact that for ∇ρ0 (r) F= kB T ρ0 (r) the Fokker–Planck distribution will yield a stationary distribution ρ0 (r) which is realised for long times. t ) . or drift ﬂux is given by jdrift = ρ (r. This turns out to be the case. Using Gauss’. t ) F(r. we can rewrite the right hand side as a volume integral: V ∂ ρ (r. t ) = ρ0 (r). As this must hold for any volume within the system. γ A well-known example of the last relation is Ohm’s law. t )d 3 r = − ∂t V ∇ · jtotal d 3 r. in a closed system with some external potential. t ). If we now substitute for ρ0 (r) the Boltzmann factor: ρ0 (r) = exp [−V (r)/(kB T )] . . and checking that the right hand side vanishes for ρ (r. we have kB T D ∂ ρ (r. ∂t Now we can substitute the expressions above for the diffusive and the drift ﬂux: ∂ F ρ (r. t ). We can check whether this equation makes sense by investigating whether. t ) = ∇ · D [∇ − kBT F] ρ (r. The diffusion equation can be derived from the requirement that the ﬂux through the surface A of a volume V equals the change in density inside that volume: d dt V ρ (r. where the ﬂux (current) is linearly related to the force (which is proportional to the applied electric ﬁeld). the density will be distributed according to the Boltzmann distribution. t ) = ∇ · D∇ − ρ (r. t ) = 0. we obtain the continuity equation: ∂ ρ (r. or divergence theorem. This can be checked by putting the left hand side of the Fokker–Planck equation to zero. ∂t This equation is called the Fokker–Planck equation. t )d 3 r = − A jtotal · d A where A is a unit vector perpendicular to a surface element dA.125 and the ﬂux due to the driving force. t ) + ∇ · jtotal (r. ∂t γ Recalling the relation γ= from the previous section.

we see that indeed the Boltzmann distribution is the correct stationary distribution. The quantity K (s) satisﬁes the property K (s) = K (−s). hence d ρ (r) eEx ρ (r) = dx kB T e2 Dρ (r) . which determines the charge density ρ (r): ρ (r) = ρ0 e−eV (r)/(kB T ) . The expression for the electric drift current: Jel = gives the same result provided that 1 D = . This relation is known as the Einstein relation. There is a stationary situation when the diffusion current cancels the drift current. this result tells us again that the relation γ = kB T /D must be valid: if the proportionality factor would be different. Using the expression for the drift current. There we have assumed that this random force had no time correlation. Consider a quantity A (you may think of a more realistic random force in the Langevin equation) which has an average value of 0 but which ﬂuctuates in time. we have encountered a ﬂuctuating quantity: the random force. 0 This is a time averaged quantity.126 and realise that for F(r) = −∇V (r). the Boltzmann distribution would not be found as the stationary distribution. In fact.3 Fluctuations – the Wiener-Kintchine theorem In our discussion of the Langevin equation. kB T γ in accordance with what was found above. For an electric ﬁeld along the x-direction: Jdrift = eD d ρ (r) . kB T e2 Ex ρ (r) γ from which we ﬁnd for the conductivity σ= This is essentially the famous Drude formula. If the boundary conditions are such that there is a source and drain (the contacts). where the net current should be zero indeed. In this section we shall study ﬂuctuating quantities which are correlated in time. dx The electric ﬁeld is minus the gradient of the potential V (r). It should be noted that the above derivation is performed in the context of a stationary equilibrium state. The autocorrelation function is deﬁned as K (s) = A(t )A(t + s) = lim 1 T →∞ T T A(t )A(t + s)dt . we can calculate transport coefﬁcients. . 13. First we analyse electric transport which is due to the acceleration of charges by an electric ﬁeld. then a net current survives.

The spectral density of the variable A is deﬁned as ˜ (ω )A ˜ ∗ (ω ) . We expect furthermore that. is short-ranged so that t1 ≈ t2 and taking the limit of large t . but refrain from imposing a non-coherence for the random force. in the case where A is a component of an electric ﬁeld. As we have seen. For example. It turns out that the autocorrelation function is intimately related to the so-called spectral density. since A(t ) = 0. The equation of motion mv ˙ = −γ v + R(t ) with R the random force still holds. and that it will attain its maximum for s = 0 (then the integrand in the average is always positive). A is such that A2 is related to some energy. we obtain. Using the fact that R(t1 )R(t2 ) = KF (t2 − t1 ). This relation is known as the Wiener-Kintchine theorem. A 0 T where the limit T → ∞ is implicitly assumed – a convention we shall adhere to from now on. which has a ﬁnite width. Usually. then the width of K (ω ) will be 1/τ (reminiscent of the Heisenberg uncertainty relation in quantum mechanics). A2 is the contribution of that component to the energy stored in the oscillating ﬁeld. Let us now come back to the Langevin equation. If the width of K (s) is called τ . S(ω ) = A We can evaluate the right hand side as follows: S(ω ) = 1 T T 0 0 T A(t )eiω t A(t ′ )e−iω t dtdt ′ ′ = 1 T A(t )A(t + s) eiω t e−iω (t +s) dtds where the shift of the integration variable was made possible because the quantity in brackets is known to be K (s). the solution of this equation can be written as mv(t ) = e−γ t /m t 0 eγ t /m R(t ′ )dt ′ ′ where we have assumed that the initial velocity was zero. 2 . using the variables T= t1 + t2 and τ = t2 − t1 . The latter is deﬁned as T 1 ˜ (ω ) = √ A(t )eiω t dt . We now evaluate the average kinetic energy: m 2 v (t ) = 2 1 −2γ t /m e 2m t 0 eγ t1 /m R(t1 )dt1 t 0 eγ t2 /m R(t2 )dt2 where the expectation value is over the different possible realisations of the random force. deﬁned in terms of the Fourier transform of A in the time domain. the autocorrelation correlation function vanishes for large values of s. If we take the limit for T → ∞. Note the absence of the factor 1/T in front of the Fourier transform of K . we ﬁnd S(ω ) = T /2 −T /2 K (s)e−iω s ds ≡ K (ω ).127 as can readily be checked by inspecting its deﬁnition.

. v(0)v(t ) = 2m where the equipartition theorem has been used to rewrite v2 (0) in terms of kB T . this friction exponent directly determines the value of the transport coefﬁcients. analyse the frequency dependence. for t → ∞ reduces to Ekin = As before. We have evaluated this function in section 12. it does not depend on ω for frequencies (much) smaller than m/γ . we may put the kinetic energy equal to kB T /2 to obtain γ= 1 ˜ KF (0). The Fourier transform of this gives us 1 kB T . i. which is a non-equilibrium property. we have derived the result γ= 1 ˜ Kv (0). kB T The parameter γ determines the transport properties. by the autocorrelation function. the result obtained there was kB T −γ |t |/m e . such as the conductance.e. Now we shall focus on this last example. To this end. In the previous section. and instead of the ω = 0 case.128 Ekin = e−2γ t /m 2m t 0 e2γ t 0 t e2γ T /m dT K (τ )d τ = 1 ˜ K (0). we can evaluate the transport coefﬁcient. Therefore. which is the inverse relaxation time of the electrons. Kv (ω ) = γ 1 + (ω m/γ )2 According to the Wiener-Kintchine theorem.3. In particular. This is a striking result: the autocorrelation function is a property of the equilibrium system – from this. 2γ P ˜ KF (0) 1 − e−2γ t /m 2γ which. the conductance is found as σ= ne2 ne2 kB T = . This means that we have white noise up to this limit. is ﬂat. we have j = nev = ne2 E γ where E is a component of the electric ﬁeld. ˜ v (0) γ K This means that we can evaluate the transport coefﬁcient σ from the autocorrelation function for the velocities. such as the electric conductivity. kB T As mentioned before. this is related to the power spectrum: |v(ω )|2 = kB T 1 . induced by the ﬂuctuations. γ 1 + (ω m/γ )2 This formula implies that the spectrum of the current power. we replace the expectation value of v2 (t ) on which the above derivation was based.

Suppose we have an isoalted system. . which tends to maximise its entropy. . . 2 In equilibrium. we must have white noise analogously to the case discussed above. The fact that the entropy strives to its maximum is the driving force which causes ﬂuctuations to dampen out. The electric energy stored in the inductance is given by the expression L Eel = I 2 . x ˜N ) + ˜1 . .129 The Wiener Kintchine theorem has an important application in electric circuits. are distributed in space. ﬂuctuations of a system from the equilibrium state can be related to the transport properties of that system.4 General analysis of linear transport Quite generally. . 13. . Consider for example an inductance in some electric circuit. j ∂ xi ∂ x j The fact that the ﬁrst-order term is not included is due to the fact that S was expanded around its maximum – hence. x ˜2 . We call these numbers xi and we call x ˜i the values for which the entropy assumes its maximum. I (ω ) d ω = 2 2 This result is known as Nyquist’s theorem. . I (ω ) d ω = 2 2 therefore. . The entropy depends on how the energy and the particle density. x ˜N ) 1 ∂ 2 S(x (xi − x ˜i )(x j − x ˜ j) + . xN ) = S(x ˜1 . This means that I acts as a generalised coordinate of the Hamiltonian. where the ‘other terms’ do not depend on I . The corresponding variation in the entropy can be expanded in a Taylor series: S(x1 . . . . . The analysis proceeds as follows. Fluctuations correspond to deviations of the xi from their equilibrium values. . We can store the information concerning these distributions in a set of numbers – you may think of the Fourier coefﬁcients of the energy and/or denisty distribution. the total power should yield L 2 kB T . . the ﬁrst derivatives are all zero. ∑ 2 i. and the power containes within a frequency window d ω should be kB T L 2 dω . . and it should therefore satisfy the equipartition theorem LI 2 kB T = . 2 2 The power stored in a mode with frequency ω is L 2 I (ω ) = J (ω ). for frequencies (much) smaller than those corresponding to the relaxation time of the circuit. 2 It is generally assumed that the total energy of the circuit can be written as Etotal = Eel (I ) + other terms. . and perhaps some other quantities. x ˜2 .

The entropy changes in the course of time as dS ∂S ˙i = δa . which says that each of these states is equally probable. Moreover. You may object that the ﬁrst derivative of S with respect to the ai is zero. . x ˜2 . we can identify δ x we shall consider an example). but near that maximum. . ∑ kB i. . ˙i as some kind of current (below As ai represents a ﬂuctuation in some quantity. From this we can derive a simple expression for correlation functions of the ai . we ﬁnd ∂S = γi j a j . ∂ ai ∑ j Using this we rewrite the result above as ∑ γik ij ak a j = kB δi j . for i = j we ﬁnd along the same lines: ai ∂S ∂aj = 0. γi j ∞ −∞ da1 · · · daN exp − ∑i j 2 ai a j γ γi j = ∂ 2 S(x 1 ˜1 . . we have for the probability of having a state where the quantities ai = xi − x ˜i the following expression: P(a1 . Note that 1 ∂P 1 ∂S = . P ∂ ai kB ∂ ai Taking the average on the left and right hand side with respect to the distribution P we obtain ai ∂S ∂ ai == da1 · · · daN ai ∂S P = kB ∂ ai da1 · · · daN ai ∂P ∂ ai Integrating by parts the integral on the right hand side leads to ai ∂S ∂ ai = −kB . where Ω is the number of states accessible to the system. x ˜N ) . . . . j ∂ xi ∂ x j Note that the fact that S has a maximum implies that the matrix γi j has positive eigenvalues. aN ) = where ij ai a j exp − ∑i j 2 . dt ∑ ∂ ai i . and combining this with the fundamental postulate if statistical mechanics. This is true at the maximum. .130 Recalling that S = kB ln Ω.

Suppose we had in the above derivation not multiplied ai (t + τ ) with a j (t ) but with a j (0). We can analyse further the relation between Li j and the correlator. We see that the linear transport coefﬁcients Li j are found as the correlation functions of the ﬂuctuations: 1 ai (t + τ )a j (t ) − ai (t )a j (t ) . Now we assume a linear relationship between the forces Xi and their resulting deviations ai : Ji = a ˙i = ∑ Li j j ∂S = Li j X j . Now we approximate ﬁnite differences by time derivatives: Li j ≈ − 1 1 ai (t )a ˙ j (0) = − kB kB ∞ 0 t 0 a ˙i (t ′ )a ˙ j (0) dt ′ .131 which suggests that we can interpret ∂ S/∂ ai = Xi as the driving force corresponding to that current. In that case. We use the thermodynamic relation T dS = −µ dN + dE − 1 T ∑ Xiai . ∂aj ∑ j Now we consider the expectation value ai (t + τ )a j (t ) = ai (t )a j (t ) + τ δ a ˙i (t )δ x j (t ) = ai (t )a j (t ) + τ ∑ Lim m ∂S a j (t ) ∂ am = ai (t )a j (t ) + τ Li j kB . If we take t much larger than the correlation time. This nontrivial property follows from what is called microscopic reversibility as it reﬂects symmetry properties of microscopic correlation functions which follow from the undrlying time reversal symmetry of the microscopic dynamics. we see that Li j ≈ − 1 kB a ˙i (t ′ )a ˙ j (0) dt ′ . Li j = kB τ From this. such as diffusion? First we must identify the dependence of the entropy on the density. Recall that a ˙i is the current ji . i . How does this relate to a speciﬁc process. The parameters ai are here the coordinates of the particles in the system. and we see that we have found an expression for the linear transport coefﬁcient Li j in terms of a time correlation function of the currents: Li j ≈ − 1 kB ∞ 0 Ji (t ′ )J j (0) dt ′ . we arrive at the result: Li j = 1 kB τ ai (t + τ )a j (0) − ai (t )a j (0) = 1 kB τ ai (t )a j (−τ ) − ai (t )a j (0) where we have used time translation symmetry. we see that the transport coefﬁcients must be symmetric: Li j = L ji .

we can derive relations for spatial ﬂuxes through some trick as we shall now show. The resulting ﬂuxes are not necessarily the ﬂuxes which are of interest to us. It is clear that for the entropy density s we have ds ∂ s = + ∇ · js . With the tilde ˜ we have indicated that the transport coefﬁcient may deviate from the coefﬁcient of the Onsager relation. The latter is given as vi (t )v j (0) = v2 (0) e−t /τ δi j . In particular. t )vi (r. t ) = ρ (r. ∇− Xi T as forces for the appropriate ﬂuxes. However. the ﬁrst two terms are always present. ∂t We assume that similar conservation laws hold for the other quantities occuring in the problem. ∇ T T . we have for the driving force for the particle ﬂux: 1 − ∇µ . but simply time derivatives. and that of particle transport is the gradient of −µ /T . For the particle density we have a similar formula. dt ∂t where js is the entropy ﬂux. together with the conservation of particle density gives ∂ρ + ∇ · j = 0. because the current is not a time derivative. after replacing all time derivatives by the appropriate ﬂuxes. they are not spatial ﬂuxes.132 In this expression. T The linear relationship between the current and this force can be cast in the form Ji = − 1 T ˜ij . which. dt T T T i This is of the same form as found above provided that we interpret the quantities −∇ 1 µ . ∑L ∂xj j ∂µ where i and j assume the values 1. 2 and 3. We immediately see that the driving force for heat transport is the gradient of 1/T . From the thermodynamic relation above. which tells us how much entropy ﬂows through a wall per unit area. We use Ji (r. we have Xi ∂ ai ∂s µ ∂ρ 1 ∂E =− + −∑ . t ) and the fact that any time correlation function of the currents is dominated by the velocity time correlation function. and the last term is used to store additional contributions resulting from external forces. ∂t T ∂t T ∂t i T ∂t Then we see. Provided we do not apply a temperature gradient to the system. using the conservation laws that µ ds 1 Xi j+∇ = −∇ jE − ∑ Ji .

µ (ρ ) = µ (ρ0 ) + kB T ln so that we obtain Ji = D hence J = D ∇ρ . ρ0 ∂ρ . kB T ∂ xi In order to arrive at the diffusion equation. we must realize that. for low densities. . ρ kB T ∂ xi D ∂µ . Adding an electric term ρ . Integrating this over time gives ˜ij = 1 L kB So we see that vi = − so that Ji = − ∞ 0 vi (t )v j (0) dt = 1 D τδi j = δi j . the diffusion coefﬁcient. kB kB 1 D ∂µ . ∂ xi 1 − V (x)qd ρ (x) T to the expression for the entropy. The Onsager formulation provides an abstract framework from which the various forces and currents can quickly be derived.133 where τ = D. gives us the force on particles with charge q in a potential: J= D ∂µ q + E kB ∂ xi T nicely in line with our previous results.