Past exam of the mathematics course of the University of Cambridge / 2001 / iii / Paper 72 (Phase Transitions and Collective Phenomena)

Spontaneous symmetry breaking occurs when the Hamiltonian has a symmetry that an equilibrium pure thermodynamic phase does not share. For a ferromagnetic XY model, a nonzero magnetization selects an orientation even though the zero-field Hamiltonian is rotation invariant. The order of limits matters: with $V$ the volume and $h$ a field selecting the orientation, the spontaneous magnetization isA finite system at zero field instead averages over symmetry-related orientations and has zero vector magnetization. Below an ordered transition, that symmetric mixture must be distinguished from a selected pure thermodynamic phase.

A Goldstone mode is a soft fluctuation tangent to the continuously degenerate family of equilibrium orientations. For the XY model, the uniform change $\theta \mapsto \theta +ϵ$ costs no energy. Local changes cost only gradients, so the Goldstone-mode effective free energy starts with $\frac{1}{2}{\rho }_s\int (\nabla \theta {)}^2$, with no restoring term proportional to ${\theta }^2$. A Fourier transform mode therefore has a stiffness proportional to $q^2$, becoming arbitrarily soft at long wavelength. The dimensionless phase stiffness here is $\bar{K}=\beta {\rho }_s$. These are the spin waves of the angle field. Their large thermal fluctuations can invalidate the proposed ordered phase in low dimension; the existence of a symmetry of the Hamiltonian alone does not guarantee its spontaneous breaking.

Fix the uniform angular mode, for example by fixing the mean phase in a large box, and introduce a microscopic ultraviolet cutoff $a$. The quadratic Goldstone-mode effective free energy becomes, under a Fourier transform,To evaluate the Gaussian functional integral, add a source $j$. Completing the square in every nonzero Fourier transform mode givesTaking two source derivatives yields the correlation functionThe cutoff and the treatment of the uniform mode regulate this expression. At separations much larger than $a$, its nonconstant part is the Green function satisfying $-\bar{K}{\nabla }^{2}G={\delta }^{(d)}(x)$. Rotation invariance and the radial Laplacian give $(r^{d-1}{G}^{\prime }{)}^{\prime }=0$ away from the origin. Integrating the Green function equation over a small sphere, whose unit-sphere area is $S_d=2{\pi }^{d/2}/\Gamma (d/2)$, fixes the flux:Thus for $d\ne 2$ the required long-distance correlation function isAt $d=2$, the finite separation-dependent limit isThe constant absorbs the divergent constant term in the $d\to 2$ limit. In low dimension the individual $G(r)$ has no regulator-independent infinite-volume value; its differences do. This is why the phase-difference variance, rather than an absolute phase variance, is the useful observable.

The dot product of the two magnetizations is ${\bar{m}}^2\cos [\theta (x)-\theta (0)]$. For the Gaussian spin wave field, the Gaussian phase averaging identity $⟨e^{iX}⟩=e^{-⟨X^2⟩/2}$ for a centered Gaussian variable givesKeeping this exponential is essential: expanding the cosine to quadratic order would fail when phase differences grow with separation. Equivalently, the phase-difference variance isFor $d>2$, the integral defining the absolute phase variance is finite at small momentum after fixing the uniform mode; its microscopic ultraviolet cutoff still matters. Its nonconstant Green function decays as $r^{2-d}$, so $D(r)$ tends to a finite constant $D_{\infty }$ andIn two dimensions, the logarithmic Green function gives $D(r)=(\pi \bar{K}{)}^{-1}\log (r/a)+O(1)$. ConsequentlyFor $1\le d<2$, $D(r)$ grows as $2r^{2-d}/[(2-d)S_d\bar{K}]$, up to bounded microscopic terms, and the correlation function again vanishes. In particular, in one dimension $S_1=2$ gives $D(r)=r/\bar{K}$ and exponential decay ${\bar{m}}^2{e}^{-r/(2\bar{K})}$. Therefore there is no conventional long-range order for $d\le 2$ at positive temperature in this short-range continuous-symmetry model. In two dimensions, algebraic decay can still support quasi-long-range order.

These infrared spin-wave correlations of the XY model explain the mechanism behind the Mermin-Wagner theorem. They are a test of the low-temperature spin wave approximation, not a claim that every $d>2$ XY model is ordered at every temperature. Large local fluctuations and defects destroy the ordered phase at sufficiently high temperature. Also, the assumed small angle is a local smooth-field approximation; phase differences at arbitrarily large separation need not remain small.

The physical angle takes values on a circle, so it is defined modulo $2\pi$. Around a closed curve enclosing a phase vortex,The integer winding number cannot change through a smooth single-valued deformation unless a defect crosses the curve. Thus a globally smooth Gaussian spin wave field omits whole topological sectors of the XY model.

Outside a vortex core of radius $a$, an isolated charge-$n$ phase vortex has $\theta =n\phi$ and $|\nabla \theta |=|n|/r$. Its dimensionless energy in a sample of size $L$ isThere are of order $(L/a{)}^2$ positions for its core, so the positional entropy contributes $2\log (L/a)$. The competition for unit charge is therefore $(\pi \bar{K}-2)\log (L/a)$, apart from core terms. Opposite charges have a logarithmic attraction and form a vortex Coulomb gas.

At low temperature, vortex-antivortex binding leaves no free charges on large scales. Spin waves still destroy true long-range order, but the correlation function decays algebraically with the renormalized phase stiffness. At the Berezinskii–Kosterlitz–Thouless transition, pairs unbind; free vortices screen the interaction and produce a finite correlation length and exponential decay above the transition. The energy-entropy estimate identifies the threshold stiffness $K_R=2/\pi$, where $K_R$ is the long-distance dimensionless stiffness, not necessarily the bare $\bar{K}$. It agrees with the universal stiffness jump and the limiting algebraic exponent $1/(2\pi K_R)=1/4$. The two-dimensional transition is vortex unbinding between quasi-long-range order and disorder, rather than the onset of nonzero spontaneous magnetization.

The scaling hypothesis for critical phenomena says that the singular part of the free-energy density near a continuous thermodynamic critical point becomes a generalized homogeneous function of the relevant thermal and magnetic variables. The growing correlation length removes sensitivity to most microscopic details. Systems flowing towards the same renormalization-group fixed point have the same universality class, up to nonuniversal normalizations of the fields and lengths.

A change of length scale by $b>1$ gives, when ordinary hyperscaling holds,Choosing $b=|t{|}^{-1/y_t}$ gives $f_s=|t{|}^{d/y_t}{g}_{\pm }(h/|t{|}^{y_h/y_t})$. More generally the specified homogeneous form uses exponents $2-\alpha$ and $\Delta$ without presupposing hyperscaling. Its thermodynamic derivatives determine the critical exponents. Analytic backgrounds are separated from this singular part.

Throughout the calculation, fractional powers of the thermal variable mean $|t|$, with different scaling functions above and below the transition. The spontaneous magnetization exponent uses the ordered branch $t<0$ with $h\to 0^{+}$ after the thermodynamic limit. Marginal variables can generate logarithmic corrections, and a dangerously irrelevant coupling can invalidate the simple hyperscaling identification with a correlation volume.

With the dimensionless magnetic field conjugate to the magnetization, $m=-\partial f_s/\partial h$. Set $y=h/|t{|}^{\Delta }$ and differentiate the homogeneous free-energy density:On the ordered branch, the one-sided value $g_{m,-}(0^{+})$ is nonzero, so the order-parameter critical exponent isTo reach the critical isotherm, take $|t|\to 0$ at fixed nonzero $h$. Suppose $g_m(y)\sim A\mathrm{sgn}(y)|y{|}^p$ at large $|y|$. Cancellation of the residual thermal factor requires $2-\alpha -\Delta -p\Delta =0$. Hence $m(0,h)\propto \mathrm{sgn}(h)|h{|}^{(2-\alpha -\Delta )/\Delta }$ andThe singular scaling function determines the nonanalytic critical-isotherm contribution; any analytic response background is subleading for the usual $\delta >1$.

A further derivative gives the magnetic susceptibility within a selected pure thermodynamic phase:Assuming a finite nonzero zero-field amplitude on the branch considered, comparison with $\chi (t,0)\asymp |t{|}^{-\gamma }$ givesTogether with part (a), this proves the Rushbrooke scaling relation $\alpha +2\beta +\gamma =2$ and the Widom scaling relation $\gamma =\beta (\delta -1)$. Below the transition, use the one-sided response of an ordered phase; differentiating through the cusp of the equilibrium free energy at $h=0$ would include phase switching rather than this critical response.

The hyperscaling relation expresses the idea that the singular dimensionless free energy in one correlation volume is of order one. A macroscopic volume $V$ contains of order $V/{\xi }^d$ such correlated regions, soSubstitution of the homogeneous correlation length givesThis has exactly the required homogeneous field variable. Comparing its thermal power with that of the singular free-energy density provesThe argument concerns powers, not equality of the two scaling functions or their amplitudes. It assumes the correlation length supplies the only singular length scale and that no dangerously irrelevant coupling changes the free energy per correlated region. For ordinary short-range quartic theory, hyperscaling holds below the upper critical dimension four, with the usual qualifications for marginal logarithms at four. Above four, mean-field critical exponents $\alpha =0$, $\nu =1/2$ do not satisfy $d\nu =2-\alpha$; the quartic coupling is dangerously irrelevant. The stated homogeneous form of $\xi$ alone therefore does not prove hyperscaling without the correlation-volume assumption.

First interpret the displayed critical correlation function as the connected correlation functionAbove the transition at zero field the subtraction is zero; in an ordered phase it is essential. For a dimensionless source coupled to the total magnetization $M=\int d^{d}xm(x)$, differentiating $Z$ twice gives $\chi =V^{-1}(⟨M^2⟩-⟨M{⟩}^2)$. Translation invariance then proves the correlation-function susceptibility sum ruleInsert the scaling form, use spherical coordinates and set $s=r/\xi$:For $\eta <2$, finite nonzero $g(0)$ and an integrable large-$s$ tail, the last integral tends to a finite nonzero constant. Microscopic separations add an analytic response background. Thus the singular magnetic susceptibility obeys ${\chi }_s\asymp {\xi }^{2-\eta }$, and at zero field $\xi \asymp |t{|}^{-\nu }$ gives the Fisher scaling relationA dimensional magnetic field would add a nonsingular inverse-temperature prefactor. Integrating the full unsubtracted ordered-phase correlation function instead would yield a volume-divergent term $V⟨m{⟩}^2$, which is not the intrinsic susceptibility appearing in this exponent identity.

Landau-Ginzburg theory describes a slowly varying order parameter by a local free-energy functional consistent with the microscopic symmetries. For a representative scalar system with spin-inversion symmetry, take the dimensionless functionalThe gradient term penalizes spatial variation, while the even local potential implements the zero-field symmetry $\varphi \mapsto -\varphi$. The positive quartic coefficient stabilizes a continuous transition. A negative quartic coefficient requires higher powers and can instead produce a first-order phase transition. A vector order parameter gives the corresponding rotationally invariant model with powers of $|\mathbf{\varphi }{|}^2$.

In the mean-field approximation, replace the field by a constant $M$ and minimize the local potential. Its equation of state and stable zero-field solutions areBelow the transition these two minima exhibit spontaneous symmetry breaking. Differentiating the equation of state gives $\chi =(r+3u{M}^2{)}^{-1}$: it is $1/r$ above the transition and $1/(2|r|)$ within either ordered phase. At $r=0$, $M=(h/u{)}^{1/3}$. The minimized zero-field free-energy density is zero above and $-r^2/(4u)$ below, apart from an analytic background; its second temperature derivative has a finite jump. Consequently

To include fluctuations, write $\varphi =M+\psi$. The quadratic part isA Gaussian functional integral therefore gives $⟨\psi (q)\psi (-q)⟩=(c{q}^2+r+3u{M}^2{)}^{-1}$, with the momentum-volume delta function understood. The Ornstein--Zernike correlation function has correlation length $\xi =\sqrt{c/(r+3u{M}^2)}$. Thus ${\nu }_{\mathrm{MF}}=1/2$, and at criticality the Gaussian covariance is proportional to $q^{-2}$, giving ${\eta }_{\mathrm{MF}}=0$. Continuous-symmetry ordered phases additionally have transverse Goldstone modes, for which the restoring mass vanishes.

The Gaussian integration contributes $\frac{1}{2}\int d^{d}q(2\pi {)}^{-d}\log (c{q}^2+r)$ to the disordered-phase free-energy density. Differentiating twice with respect to $r$ gives a critical contribution proportional to $\int d^{d}q(c{q}^2+r{)}^{-2}$; for $d<4$ its singular part grows as $r^{(d-4)/2}$ and at four dimensions it is logarithmic. Interacting fluctuations thus become important close to the critical point below four dimensions. The Ginzburg criterion tests this failure of the mean-field approximation; the renormalization group then explains modified exponents and universality. Mean-field theory finds the competing minima, while fluctuations determine whether its predicted critical behavior survives at long distance.

The Ginzburg criterion tests whether fluctuations on the scale of one correlation volume are small compared with the squared equilibrium order parameter. In the scalar quartic Landau-Ginzburg theory of part (a), the ordered mean-field value is $M^2=|r|/u$, and its longitudinal correlation length is proportional to $(c/|r|{)}^{1/2}$.

Average the fluctuation field over a region of size $\xi$. Its smoothing factor selects momenta of order ${\xi }^{-1}$ or smaller. The Gaussian correlation function then gives, up to a positive dimension-dependent constant,The last step follows directly by setting $q=\sqrt{|r|/c}k$. Dividing by $M^2$ proves the correlation-volume derivation of the scalar Ginzburg ratio:For $d<4$, the ratio diverges as the transition is approached, so there is a fluctuation-dominated critical region of a phase transition in which mean-field critical exponents cannot be trusted. With $r=r_{0}t$, an estimate of its width isin fixed microscopic units. Its numerical coefficient depends on the averaging convention and microscopic normalization.

For $d>4$, the ratio tends to zero and the long-distance mean-field approximation becomes self-consistent. Four is therefore the ordinary quartic upper critical dimension. At $d=4$ this test is marginal: its thermal power is zero, and the running quartic interaction produces logarithmic corrections. The marginal Ginzburg criterion alone cannot decide those logarithms. Short-distance contributions to an unaveraged variance renormalize the coefficients; they should not be mistaken for the long-wavelength fluctuations used in this criterion. Nor does the criterion by itself prove that a proposed ordered phase exists below its lower critical dimension.

The renormalization group organizes how a statistical system changes when observed at progressively larger length scales. Start with its partition function and a microscopic cutoff $\Lambda$. Split the Fourier transform field into modes below $\Lambda /b$ and modes in the shell $\Lambda /b<|q|<\Lambda$. Integrate the shell modes, rescale coordinates $x^{\prime }=x/b$ to restore the cutoff, and rescale the field to keep a chosen gradient normalization. This is a renormalization-group transformation. It preserves the partition function when the generated interactions and additive free energy terms are retained; truncating the resulting operator expansion is an approximation.

A renormalization-group fixed point has an invariant dimensionless interaction structure. The continuous flow of a nearby scaling field $v_i$ takes the linearized formA positive $y_i$ defines a relevant operator, a negative $y_i$ an irrelevant operator, and a zero $y_i$ a marginal operator requiring nonlinear analysis. The continuous-flow eigenvalues $y_i$ should not be confused with the discrete multipliers $b^{y_i}$. To reach a critical point, tune the relevant thermal variable and set the relevant symmetry-breaking source to zero. The resulting flow stays on the critical surface.

Microscopically different systems whose remaining couplings approach the same fixed point share a universality class. Irrelevant perturbations lose their influence on long-distance exponents. Relevant directions explain why varying the temperature or field moves the system away from criticality. The thermal scaling field obeys $t^{\prime }=b^{y_t}t$, while the rescaled correlation length is ${\xi }^{\prime }=\xi /b$. Taking $b=|t{|}^{-1/y_t}$ therefore gives $\nu =1/y_t$.

If the field has scaling dimension $x_{\varphi }=(d-2+\eta )/2$, the source term $\int h\varphi$ has eigenvalue $y_h=d-x_{\varphi }=(d+2-\eta )/2$. The singular free-energy density then obeyswhen the hyperscaling relation is valid. This derives the scaling hypothesis for critical phenomena and the exponent relations of Question 2 from the fixed-point picture.

At the Gaussian fixed point, keeping $\int (\nabla \varphi {)}^2$ invariant gives $x_{\varphi }=(d-2)/2$. The quadratic coupling has $y_r=2$ and the quartic coupling has $y_u=4-d$. Thus quartic interactions are relevant below four dimensions, marginal at four, and irrelevant above four. Below four, in dimensions where an ordinary continuous short-range scalar transition exists, an interacting Wilson-Fisher fixed point governs that transition. Above four, the Gaussian fixed point gives mean-field powers, but the quartic term remains necessary to stabilize the ordered state: it is a dangerously irrelevant coupling, explaining the failure of naive hyperscaling. The renormalization group connects scale invariance, universality, relevant tuning parameters and fluctuation corrections in one framework.

Write the finite-chain spin vector as $\sigma$ and the exchange matrix as $\mathsf{J}$. For $J>0$ and $\kappa >0$ this matrix is positive definite, as can be seen from the positive Fourier symbol in part (b), or from the explicit positive open-chain inverse. The linear-source multivariate Gaussian integral, obtained by completing the square, givesIndeed the exponent is $-(m-\mathsf{J}\sigma {)}^T{\mathsf{J}}^{-1}(m-\mathsf{J}\sigma )+{\sigma }^T\mathsf{J}\sigma$. Insert this Hubbard–Stratonovich transformation into the spin partition function. The field term remains $h{\sum }_i{\sigma }_i$, and each independent spin sum isConsequentlyThe interaction convention counts the full ordered pair sum, with no factor of one-half, so the auxiliary coupling is $2m_i{\sigma }_i$. Diagonal terms $J_{ii}{\sigma }_i^2$ only contribute a constant to the spin Hamiltonian; retaining them makes the given positive exchange matrix and normalization convenient. The auxiliary field is not itself the physical spin magnetization: the conditional spin expectation is $\tanh (2m_i+h)$, so $⟨{\sigma }_i⟩=⟨\tanh (2m_i+h)⟩$ in the auxiliary-field integral.

Use the infinite translation-invariant chain first. Set $\rho =e^{-\kappa }$ and Fourier-transform the exchange matrix:For positive $J,\kappa$ this is strictly positive. Its reciprocal is a linear polynomial in $\cos q$, proving the nearest-neighbour form of the exponentially coupled Ising inverse kernel:Expanding the squared difference givesSince $(\cosh \kappa -1)/\sinh \kappa =\tanh (\kappa /2)$, substitution into part (a) yieldsThis is the exact bulk identity. A finite periodic version uses the periodized coupling $J_{ij}=J{\sum }_{\ell \in \mathbb{Z}}{e}^{-\kappa |i-j+\ell N|}$; its discrete Fourier symbols are the same and the squared differences wrap around the ring. Simply truncating the exponential at open ends gives a different inverse at those ends.

For an open chain of $N\ge 2$ sites with the literal finite matrix $J{\rho }^{|i-j|}$, the open-chain endpoint corrections for an exponential Ising kernel are explicit:The inverse has endpoint diagonal entries $1/[J(1-{\rho }^2)]$, interior diagonal entries $(1+{\rho }^2)/[J(1-{\rho }^2)]$ and adjacent entries $-\rho /[J(1-{\rho }^2)]$. Multiplication by the original geometric matrix verifies every entry. These boundary terms affect boundary amplitudes, but not the bulk free-energy density in the thermodynamic limit at fixed $\kappa >0$.

For an auxiliary field varying slowly over lattice distances, measure $x$ in lattice-spacing units and replace the squared difference by $(dm/dx{)}^2$. The long-wavelength action becomesThis is the Euclidean path integral of a one-coordinate quantum particle, with the classical spatial coordinate serving as imaginary time and $\hslash =1$. To see the correspondence at the operator level, time-slice it by $ϵ$ and use the symmetric kernelActing on a smooth test function, Gaussian integration gives $K_{ϵ}\psi =\psi +ϵ{\psi }^{\prime \prime }/(2M_Q)-ϵU\psi +O({ϵ}^2)$. Thus the continuum auxiliary-field transfer operator for an Ising chain is $e^{-ϵ{\hat{H}}_Q}$, whereMultiplying kernels and integrating over intermediate fields gives $⟨m_f|e^{-L{\hat{H}}_Q}|m_i⟩$ for fixed endpoints, and $\mathrm{tr}{e}^{-L{\hat{H}}_Q}$ for periodic fields, up to the normalization already absorbed in $C$. This is the precise classical Ising chain as a quantum particle correspondence. It concerns an imaginary-time transition kernel; it is not the squared modulus of a real-time transition amplitude. The continuum approximation retains smooth long-distance modes, rather than asserting exact equality of the finite-spacing transfer operator with this differential operator.

At $h=0$, set $A=\tanh (\kappa /2)/J$. Expansion of the local potential givesThe potential is even and grows as $A{m}^2-2|m|$ at large $|m|$. It is a symmetric double well precisely whenThe nonzero stationary points satisfy $Am=\tanh (2m)$. For $m>0$, the ratio $\tanh (2m)/m$ decreases strictly from 2 to 0: writing $z=2m$, the sign follows from $\tanh z-z{\mathrm{sech}}^{2}z>0$, whose derivative is $2z{\mathrm{sech}}^{2}z\tanh z>0$. Hence there is one minimum on each side when $A<2$. At $A=2$ the potential instead has a single quartic minimum, and for $A>2$ a single quadratic minimum. Small $\kappa$ with fixed positive $J$ falls in the double-well regime, but the printed claim requires this coupling condition; $\kappa \ll 1$ alone does not imply it if $J$ scales to zero too quickly.

A weak magnetic field givesThe linear term is odd in $m$. For positive $h$, the positive well becomes deeper. If the zero-field minima are $\pm m_{\ast }$, their depth difference to first order is $U(m_{\ast },h)-U(-m_{\ast },h)=-2h\tanh (2m_{\ast })$. The magnetic field tilts the symmetric double well and favors the matching spin orientation. A large enough field can eliminate the metastable well.

Finally, two minima do not imply a true finite-temperature phase transition in this one-dimensional chain at fixed finite interaction range $1/\kappa$. A finite quantum barrier permits quantum tunnelling; the confining one-dimensional Schrödinger operator has a unique symmetric ground state at zero field and a nonzero first excitation gap. The quantum gap corresponds to a finite classical correlation length. A very small gap can mimic ordering on a finite sample, but genuine degeneracy requires a singular limit such as vanishing tunnelling. This distinction is central to interpreting the quantum mapping.