Critical Propagators and the Upper Critical Dimension

The previous page derived the continuum scalar field from the Ising model by an exact Hubbard–Stratonovich transformation. The Gaussian part of that field theory already knows a lot: it knows which momentum mode becomes soft, how the correlation length diverges, and why the two-point function near criticality has the universal form of a massive free scalar propagator.

This page asks the next question. Once the continuum action contains an interaction such as $\lambda \phi^4$, when is the Gaussian propagator trustworthy? The answer is not simply “when $\lambda$ is small.” Near a critical point the correlation length $\xi$ becomes large, so fluctuations are sampled over larger and larger regions. The true expansion parameter is a scale-dependent one,

$$g(\xi)\sim \lambda \xi^{4-D}.$$

Thus the Ising $\phi^4$ interaction becomes less important at long distances for $D>4$, marginal for $D=4$, and more important for $D<4$. The number

$$D_c=4$$

is the upper critical dimension. Above it, mean-field exponents are asymptotically correct. Below it, the Gaussian fixed point is unstable and the course is forced toward the Wilson–Fisher fixed point.

Required background. Hubbard–Stratonovich transformation and the continuum field supplies the lattice quadratic kernel, the auxiliary-field propagator, and the continuum $\phi^4$ action used below.

Critical modes of a lattice kernel

Propagator normalization

Throughout this page $D$ denotes the Euclidean spatial dimension of the statistical system. We use

$$\int_k \equiv \int {d^D k\over (2\pi)^D},$$

usually with an ultraviolet cutoff $\Lambda\sim a^{-1}$ inherited from the lattice spacing $a$.

The continuum Gaussian action is normalized as

$$S_0[\phi]= {1\over2}\int_k (k^2+m^2)\phi(k)\phi(-k),$$

so

$$G_0(k)=\langle \phi(k)\phi(-k)\rangle_0={1\over k^2+m^2}.$$

For the exact propagator we write

$$G(k)^{-1}=G_0(k)^{-1}+\Sigma(k) =k^2+m_0^2+\Sigma(k).$$

Some QFT books define the self-energy with the opposite sign. Nothing physical depends on this convention; what matters is the exact inverse two-point vertex $\Gamma^{(2)}(k)=G(k)^{-1}$.

For an interacting theory, $\Gamma^{(2)}(0)=\chi^{-1}$ is exactly the inverse susceptibility, not automatically the square of the exponential correlation-length mass. If

$$\Gamma^{(2)}(k)=\chi^{-1}+Z_2 k^2+O(k^4),$$

then the second-moment correlation length is $\xi_2^2=Z_2\chi$. After rescaling the field so that $Z_2=1$, it is convenient to write $m^2=\chi^{-1}=\xi_2^{-2}$. The exponential correlation length has the same critical scaling but need not agree with $\xi_2$ away from the scaling limit.

Start with a translation-invariant ferromagnetic interaction. In the spin language one may write schematically

$$H=-{1\over2}\sum_{x,y}J(x-y)\sigma_x\sigma_y,$$

with $J(r)$ short-ranged and positive near the origin. Its Fourier transform is

$$V(k)=\sum_r J(r)e^{ik\cdot r}.$$

The Hubbard–Stratonovich Gaussian propagator from the previous page has the form

$$G_{\mathrm{lat}}(k)\propto {1\over 1-\beta V(k)},$$

up to a smooth nonzero normalization factor. The important information is the zero of the inverse propagator. In the Gaussian or random-phase approximation, an instability occurs when some eigenvalue of the interaction kernel reaches

$$1-\beta_c^{\mathrm G} V(k_\star)=0.$$

For a ferromagnet, the largest value of $V(k)$ is at $k_\star=0$, so

$$\boxed{\beta_c^{\mathrm G} V(0)=1.}$$

If the maximum is instead at a nonzero wavevector $Q$, the soft field is not the uniform magnetization. The continuum expansion must be made around $Q$, which describes antiferromagnetic or modulated order. This is the same mathematical mechanism with a different ordering wavevector.

Near the ferromagnetic critical point, short-range interactions imply that $V(k)$ is analytic at $k=0$. Rotational symmetry at long distances gives

$$V(k)=V(0)-\rho k^2+O(k^4), \qquad \rho>0.$$

Define the reduced distance from the Gaussian critical point by

$$\tau=1-\beta V(0).$$

On the disordered side, $\tau>0$. Then

$$1-\beta V(k) = \tau+\beta\rho k^2+O(k^4).$$

After rescaling the field and momentum-independent constants, this becomes the Ornstein–Zernike propagator

$$\boxed{ G_0(k)\simeq {Z\over k^2+m^2}, \qquad m^2\propto \tau\propto T-T_c^{\mathrm G}. }$$

The parameter $m$ is the inverse correlation length:

$$\xi=m^{-1}.$$

Therefore the Gaussian theory predicts

$$\xi\sim |T-T_c^{\mathrm G}|^{-1/2}.$$

This is the mean-field value $\nu_{\mathrm{MF}}=1/2$.

The Gaussian instability occurs when the inverse propagator vanishes at the maximum of the lattice kernel. For a ferromagnet the soft mode is at $k=0$, and the analytic expansion $V(k)=V(0)-\rho k^2+\cdots$ gives $G_0(k)\simeq Z/(m^2+c k^2)$.

For the nearest-neighbor hypercubic Ising model,

$$V(k)=2J\sum_{\mu=1}^D\cos(k_\mu a).$$

At small $k$,

$$V(k)=2DJ-Ja^2k^2+O(k^4a^4).$$

Thus

$$1-\beta V(k) = (1-2D\beta J)+\beta J a^2 k^2+\cdots.$$

The Gaussian critical value is

$$\beta_c^{\mathrm{G}}J={1\over 2D},$$

and

$$G_0(k)\simeq {1\over (1-2D\beta J)+\beta Ja^2 k^2}.$$

The exact critical point is shifted by fluctuations in dimensions where the interaction is important. The Gaussian calculation is nevertheless the right first local approximation: it identifies the soft mode and the analytic structure around it. The superscript on $\beta_c^{\mathrm G}$ is essential; $1/(2D)$ is not the exact nearest-neighbor Ising critical coupling.

Position-space propagator and the correlation length

The continuum Gaussian propagator in position space is

$$G_0(x)=\int_k {e^{ik\cdot x}\over k^2+m^2}.$$

At criticality, $m=0$, dimensional analysis alone gives

$$G_0(x)\propto {1\over |x|^{D-2}},$$

for $D>2$. With the normalization above, the exact coefficient is

$$\boxed{ G_0(x) = {\Gamma\!\left({D\over2}-1\right)\over 4\pi^{D/2}} {1\over |x|^{D-2}}, \qquad m=0,\quad D>2. }$$

Away from criticality,

$$G_0(x) = {1\over (2\pi)^{D/2}} \left({m\over |x|}\right)^{D/2-1} K_{D/2-1}(m|x|),$$

where $K_\nu$ is a modified Bessel function. This formula has two important limits. For distances much shorter than $\xi$,

$$|x|\ll \xi, \qquad G_0(x)\sim {1\over |x|^{D-2}},$$

so the system looks critical. For distances much longer than $\xi$,

$$|x|\gg \xi, \qquad G_0(x)\sim {e^{-|x|/\xi}\over |x|^{(D-1)/2}},$$

so correlations are exponentially suppressed.

The susceptibility is the zero-momentum two-point function,

$$\chi=G_0(k=0)={1\over m^2}.$$

Since $m^2\propto T-T_c$, Gaussian theory gives

$$\chi\sim |T-T_c|^{-1}.$$

Thus $\gamma_{\mathrm{MF}}=1$. At this level the anomalous dimension is also zero, because at criticality

$$G_0(x)\sim {1\over |x|^{D-2}} = {1\over |x|^{D-2+\eta}} \quad\Longrightarrow\quad \eta_{\mathrm{MF}}=0.$$

The central issue is whether interactions preserve these exponents.

Dyson equation and critical mass tuning

The continuum action near the Ising critical point is

$$S[\phi] = \int d^D x\, \left[ {1\over2}(\partial\phi)^2 +{1\over2}m_0^2\phi^2 +{\lambda\over4!}\phi^4 +\cdots \right].$$

The dots include higher even powers and higher derivatives. For the present page the leading interaction is the quartic term. Perturbatively, the exact two-point function is organized by one-particle-irreducible self-energy insertions:

$$G(k)={1\over k^2+m_0^2+\Sigma(k)}.$$

The physical susceptibility is not fixed by the bare parameter $m_0$. At zero momentum the exact inverse propagator gives

$$\boxed{ \chi^{-1}\equiv \Gamma^{(2)}(0)=G(0)^{-1} =m_0^2+\Sigma(0). }$$

The critical point is the value of the microscopic parameters for which $\chi$ diverges,

$$\chi^{-1}=0.$$

Thus criticality is a tuning condition. The bare mass must cancel the fluctuation correction:

$$m_{0,c}^2+\Sigma_c(0)=0.$$

Near the critical point it is often cleaner to subtract the critical value:

$$\chi^{-1} = m_0^2-m_{0,c}^2 + \left[\Sigma(0;m)-\Sigma_c(0)\right].$$

This formula is the statistical-mechanics version of mass renormalization. The lattice cutoff makes all quantities finite, but the separation between the physical susceptibility scale and the microscopic bare mass is still essential. When the coefficient of $k^2$ has been normalized to one, we may abbreviate $\chi^{-1}$ as $m^2$ and identify it with $\xi_2^{-2}$, as in the convention note.

Self-energy insertions shift the inverse propagator. The critical point is defined by $\Gamma^{(2)}(0)=0$, not by the naive vanishing of the bare mass. With $G^{-1}=G_0^{-1}+\Sigma$, expanding $G$ produces alternating signs. In $\phi^4$ theory the one-loop tadpole shifts $\Gamma^{(2)}(0)$; momentum-dependent corrections begin at two loops.

Using a propagator with the renormalized scaling mass $m=\xi_2^{-1}$—equivalently, reorganizing perturbation theory around the physical quadratic term—the one-loop tadpole gives

$$\boxed{ \Sigma_1={\lambda\over2}\int^\Lambda {d^D q\over (2\pi)^D} {1\over q^2+m^2}. }$$

The factor $1/2$ follows from the conventional $\lambda\phi^4/4!$ normalization. Other normalizations of the quartic term move this numerical factor but not the scaling.

Write

$$I_D(m)=\int^\Lambda {d^D q\over (2\pi)^D}{1\over q^2+m^2} = {S_{D-1}\over(2\pi)^D} \int_0^\Lambda dq\,{q^{D-1}\over q^2+m^2},$$

where

$$S_{D-1}={2\pi^{D/2}\over \Gamma(D/2)}$$

is the area of the unit $(D-1)$-sphere. The ultraviolet behavior is

$$I_D(0)\sim {S_{D-1}\over(2\pi)^D} {\Lambda^{D-2}\over D-2}, \qquad D>2.$$

This cutoff-dependent constant shifts the critical temperature. It is not universal. The universal question is how the remaining long-distance part behaves as $m\to0$.

For $2<D<4$,

$$I_D(m)-I_D(0)\sim -C_D m^{D-2}, \qquad C_D>0.$$

For $D=4$,

$$I_4(m)-I_4(0) \sim -{m^2\over 8\pi^2}\log{\Lambda\over m} +\text{analytic terms}.$$

For $D>4$, the leading small-$m$ correction is analytic in $m^2$ once the cutoff is kept fixed:

$$I_D(m)-I_D(0)\sim -A_D \Lambda^{D-4}m^2+\cdots.$$

For noninteger $D>4$, a subleading nonanalytic term proportional to $m^{D-2}$ can also occur; it does not overturn the leading comparison with $m^2$. At even dimensions further logarithms appear in subleading orders.

The borderline behavior at $D=4$ is the first warning that four dimensions are special. The next section gives a cleaner and more general explanation.

Power counting at the Gaussian fixed point

At criticality the Gaussian action is

$$S_0={1\over2}\int d^D x\,(\partial\phi)^2.$$

Under the coordinate dilation

$$x'=b x,$$

the derivative scales as

$$\partial\mapsto b^{-1}\partial,$$

and the measure scales as

$$d^D x\mapsto b^D d^D x.$$

the transformed field that leaves $S_0$ invariant is

$$\phi'(x')=b^{-(D-2)/2}\phi(x).$$

Equivalently, the engineering dimension of the scalar field is

$$\boxed{ [\phi]={D-2\over2}. }$$

Now consider the quartic interaction

$$S_{\mathrm{int}}={\lambda\over4!}\int d^D x\,\phi^4.$$

The operator $\phi^4$ has dimension

$$[\phi^4]=2(D-2).$$

Since the action is dimensionless,

$$[\lambda]+D-2(D-2)=0.$$

Therefore

$$\boxed{ [\lambda]=4-D. }$$

This single equation encodes the upper critical dimension:

$$\begin{array}{ccl} D>4 &:& \lambda \text{ has negative dimension, so it is irrelevant at long distance},\\ D=4 &:& \lambda \text{ is classically marginal},\\ D<4 &:& \lambda \text{ has positive dimension, so it is relevant at long distance}. \end{array}$$

At a length scale $R$, the associated dimensionless coupling is

$$\boxed{ g(R)\sim \lambda R^{4-D}. }$$

Thus $g(R)\to0$ as $R\to\infty$ for $D>4$. This is the precise sense in which mean-field theory becomes exact at the longest distances above four dimensions. For $D<4$, $g(R)$ grows with $R$, so the Gaussian approximation eventually breaks down no matter how small the microscopic coupling was.

At $D=4$, power counting alone cannot decide the fate of $\lambda$. Quantum or statistical fluctuations generate logarithms, and the next page will turn those logarithms into a renormalization-group flow.

Position-space Ginzburg test

There is a useful way to see the same result without assigning dimensions abstractly. At criticality, the Gaussian propagator scales as

$$G_0(R)\sim {1\over R^{D-2}}.$$

The composite operator $\phi^2$ is the continuum representative of the local energy-density perturbation. After subtracting its expectation value, its Gaussian connected two-point function behaves as

$$C_0(R) \equiv \langle \phi^2(x_1)\phi^2(x_2)\rangle_{0,c} =2G_0(R)^2 \sim {1\over R^{2D-4}}, \qquad R=|x_1-x_2|.$$

Now perturb by one insertion of the quartic interaction. The first correction scales as

$$\delta C(R) \sim \lambda \int d^D z\, G_0(x_1-z)^2G_0(x_2-z)^2.$$

For the scaling estimate, the integration region of size $R$ contributes

$$\int d^D z\sim R^D,$$

and each propagator is of order $R^{-(D-2)}$. There are four propagators, so

$$\delta C(R) \sim \lambda {R^D\over R^{4D-8}} = {\lambda\over R^{3D-8}}.$$

The ratio of the first correction to the Gaussian answer is therefore

$$\boxed{ {\delta C(R)\over C_0(R)} \sim \lambda R^{4-D}. }$$

This is the same dimensionless coupling $g(R)$ found by power counting.

The integral also contains short-distance pieces from $z$ approaching either endpoint. Those pieces renormalize the composite operator and contribute local terms. The estimate above isolates the nonlocal dependence on the separation $R$, which is the part relevant to the infrared criterion.

At the Gaussian critical point $G_0(R)\sim R^{-(D-2)}$. A single $\phi^4$ insertion gives a relative correction $\delta C/C_0\sim \lambda R^{4-D}$. The correction decreases for $D>4$, grows for $D<4$, and is marginal at $D=4$.

The same estimate is often called the Ginzburg criterion. Evaluate the dimensionless coupling at the correlation length:

$$g(\xi)\sim \lambda \xi^{4-D}.$$

As $T\to T_c$, $\xi\to\infty$. Therefore:

• for $D>4$, $g(\xi)\to0$, so the Gaussian or mean-field approximation becomes asymptotically reliable;
• for $D=4$, $g(\xi)$ is marginal and receives logarithmic corrections;
• for $D<4$, $g(\xi)\to\infty$, so critical fluctuations invalidate mean-field exponents.

This does not mean the theory is uncontrolled below four dimensions. It means the correct fixed point is not the Gaussian one. The Wilson–Fisher fixed point, constructed perturbatively in $D=4-\epsilon$, is precisely the controlled replacement.

The random-walk meaning of 4 equals 2 plus 2

There is a beautiful interpretation of the number four. The Gaussian propagator has the Schwinger representation

$${1\over k^2+m^2} = \int_0^\infty dL\,e^{-L(k^2+m^2)}.$$

Fourier transforming gives

$$G_0(x) = \int_0^\infty dL\, {1\over(4\pi L)^{D/2}} \exp\left(-{x^2\over4L}-m^2L\right).$$

The parameter $L$ behaves like the length, or proper time, of a Brownian path. At criticality $m=0$, the exponential factor

$$\exp\left(-{R^2\over4L}\right)$$

says that paths contributing to a displacement $R$ typically have

$$L\sim R^2.$$

Thus a free critical propagator is geometrically like a random walk of fractal dimension $2$.

The quartic interaction couples two such walks when they meet. The elementary intersection-dimension estimate says that two sets of fractal dimension $2$ have a positive-dimensional intersection when

$$2+2>D.$$

The borderline is

$$D=2+2=4.$$

This is the geometric version of the upper critical dimension. Above four dimensions, intersections of two long independent walks are sufficiently sparse that the local interaction becomes irrelevant; below four dimensions they proliferate and reshape the long-distance theory. At the borderline, the estimate is marginal and logarithms decide the result. This is an interpretation of the power counting, not a substitute for the renormalization-group argument.

This random-walk viewpoint will reappear later in the course when field-theoretic propagators are related to sums over paths, and eventually when random surfaces enter the discussion.

Mean field above four dimensions and its caveat

For $D>D_c$, the long-distance critical two-point function is Gaussian:

$$G(k)\sim {1\over k^2+m^2}.$$

The critical exponents take their mean-field values:

$$\nu={1\over2}, \qquad \eta=0, \qquad \gamma=1.$$

However, one should be careful with the phrase “the interaction is irrelevant.” The quartic coupling is irrelevant at the Gaussian fixed point in the renormalization-group sense, but it is still needed to stabilize the ordered phase. If $m^2<0$, the potential

$${1\over2}m^2\phi^2$$

is unbounded unless the quartic term is kept. Thus $\lambda$ is dangerously irrelevant above four dimensions: it does not change the leading long-distance two-point critical behavior, but it enters thermodynamic quantities such as the magnetization amplitude in the ordered phase.

This subtlety is one reason critical phenomena are more delicate than a first reading of power counting suggests. Power counting identifies which fixed point is stable; it does not automatically tell us which amplitudes and scaling relations remain ordinary. In particular, bulk mean-field exponents hold for $D>4$, but hyperscaling and naive finite-size scaling fail because they are sensitive to the dangerously irrelevant coupling.

Example: the one-loop mass shift

Let us make the mass-renormalization statement concrete. In the symmetric phase, use

$$S[\phi]= \int d^D x\, \left[ {1\over2}(\partial\phi)^2 +{1\over2}m_0^2\phi^2 +{\lambda\over4!}\phi^4 \right].$$

At one loop,

$$m^2=m_0^2+{\lambda\over2}I_D(m)+O(\lambda^2).$$

Using $m$ rather than $m_0$ in the internal line is a convenient self-consistent notation; replacing it by the zeroth-order mass changes the result only beyond the displayed order away from the critical infrared singularity.

The critical bare mass is obtained by setting $m=0$:

$$m_{0,c}^2=-{\lambda\over2}I_D(0)+O(\lambda^2).$$

Therefore

$$m^2 = (m_0^2-m_{0,c}^2) + {\lambda\over2}\left[I_D(m)-I_D(0)\right] +O(\lambda^2).$$

The first term is the experimentally or microscopically tunable distance from the critical point. In a thermal system it is proportional to $T-T_c$ after nonuniversal normalization. The second term is the fluctuation correction.

For $D>4$, the bracket is proportional to $m^2$ times a cutoff-dependent coefficient, so it can be absorbed into a finite renormalization of the slope relating $m^2$ to $T-T_c$. Mean-field scaling survives.

For $D=4$, the bracket contains

$$-m^2\log{\Lambda\over m},$$

so logarithmic corrections appear.

For $2<D<4$, the bracket contains a nonanalytic contribution

$$-m^{D-2}.$$

This term becomes more singular than $m^2$ as $m\to0$. At and below $D=2$, even $I_D(0)$ has an infrared divergence, so the subtraction must be reformulated. In either case ordinary perturbation theory around the Gaussian point has lost control in the infrared. Near four dimensions, the $\epsilon$ expansion repairs this by moving the expansion point from the Gaussian fixed point to a nearby interacting fixed point.

Summary

The lattice interaction kernel determines which mode becomes critical. For a ferromagnet,

$$\beta_c^{\mathrm G} V(0)=1, \qquad V(k)=V(0)-\rho k^2+\cdots,$$

and the Gaussian propagator near the critical point takes the universal form

$$G_0(k)\simeq {Z\over k^2+m^2}, \qquad m=\xi^{-1}, \qquad m^2\propto |T-T_c^{\mathrm G}|.$$

At criticality,

$$G_0(x)\sim {1\over |x|^{D-2}},$$

so Gaussian theory predicts $\nu=1/2$, $\eta=0$, and $\gamma=1$.

Interactions shift the critical point through the self-energy:

$$G(k)^{-1}=k^2+m_0^2+\Sigma(k), \qquad \chi^{-1}=G(0)^{-1}.$$

After normalizing the $k^2$ coefficient, $\chi^{-1}=m^2=\xi_2^{-2}$.

The one-loop tadpole gives

$$\Sigma_1={\lambda\over2}\int^\Lambda {d^D q\over(2\pi)^D}{1\over q^2+m^2},$$

so the bare mass must be tuned against fluctuation corrections to reach $m=0$.

The quartic coupling has engineering dimension

$$[\lambda]=4-D.$$

Equivalently, its dimensionless strength at length scale $R$ is

$$g(R)\sim \lambda R^{4-D}.$$

Thus

$$D_c=4$$

is the upper critical dimension of the Ising $\phi^4$ theory. Above four dimensions the Gaussian fixed point controls the critical exponents. Below four dimensions the quartic interaction grows at long distances, and the correct critical theory is interacting.

Common pitfalls

Calling the Gaussian instability exact. The condition $\beta_c^{\mathrm G} V(0)=1$ is not generally the exact critical temperature. It is the point at which the quadratic approximation becomes unstable; fluctuations shift it.

Equating every definition of mass. Exactly, $G(0)^{-1}=\chi^{-1}$. It equals $\xi_2^{-2}$ only after normalizing the $k^2$ coefficient, while the exponential correlation length comes from the nearest complex-momentum singularity. These definitions share a critical exponent but can differ by finite factors.

Discarding the tadpole as a divergent nuisance. In a lattice statistical model it is finite, but it still changes the relation between microscopic temperature and physical correlation length. Renormalization is already present conceptually before taking a continuum cutoff to infinity.

Mistaking power counting for the full RG. Power counting identifies $D=4$ as the marginal dimension. At exactly four dimensions logarithms decide the flow; below four dimensions the Wilson–Fisher fixed point replaces the Gaussian fixed point.

Dropping the quartic term above four dimensions. It is irrelevant for leading bulk critical exponents but still stabilizes the ordered phase and affects amplitudes, hyperscaling, and finite-size scaling. This is the standard dangerously irrelevant caveat.

Exercises

Exercise 1

Let

$$G_{\mathrm{lat}}(k)={A\over 1-\beta V(k)}$$

and suppose

$$V(k)=V(0)-\rho k^2+O(k^4), \qquad \rho>0.$$

Assuming $\beta_c^{\mathrm G}V(0)=1$, show that near $\beta_c^{\mathrm G}$ the propagator can be written in the form

$$G_{\mathrm{lat}}(k)\simeq {Z\over k^2+m^2}.$$

Find $m^2$ in terms of $\beta$, $V(0)$, and $\rho$ up to an overall normalization convention.

Solution

Using the small-$k$ expansion,

$$1-\beta V(k) = 1-\beta V(0)+\beta\rho k^2+O(k^4).$$

Define

$$\tau=1-\beta V(0).$$

Then

$$G_{\mathrm{lat}}(k)\simeq {A\over \tau+\beta\rho k^2}.$$

Factor out $\beta\rho$ from the denominator:

$$G_{\mathrm{lat}}(k) \simeq {A\over \beta\rho} {1\over k^2+\tau/(\beta\rho)}.$$

Thus

$$Z={A\over\beta\rho}, \qquad m^2={\tau\over\beta\rho} ={1-\beta V(0)\over \beta\rho}.$$

Since $\beta_c^{\mathrm G}V(0)=1$, near the Gaussian transition

$$m^2\propto \beta_c^{\mathrm G}-\beta$$

on the disordered side. In terms of temperature this is proportional to $T-T_c^{\mathrm G}$, up to a positive nonuniversal constant.

Exercise 2

Use dimensional analysis to show that

$$G_0(x)=\int {d^D k\over(2\pi)^D}{e^{ik\cdot x}\over k^2}$$

scales as $|x|^{2-D}$.

Solution

Let $r=|x|$ and rescale the integration variable by

$$q=kr.$$

Then

$$d^D k={d^D q\over r^D}, \qquad k^2={q^2\over r^2}, \qquad e^{ik\cdot x}=e^{iq\cdot \hat x}.$$

Therefore

$$G_0(x) = \int {d^D q\over(2\pi)^D} {r^2\over r^D} {e^{iq\cdot \hat x}\over q^2} = r^{2-D} \int {d^D q\over(2\pi)^D}{e^{iq\cdot \hat x}\over q^2}.$$

The remaining integral is a dimensionless angular constant, after regularization of short-distance and long-distance singularities. Thus

$$G_0(x)\propto {1\over r^{D-2}}.$$

For $D>2$ the standard normalization gives

$$G_0(x)= {\Gamma(D/2-1)\over4\pi^{D/2}} {1\over r^{D-2}}.$$

Exercise 3

For

$$I_D(m)=\int^\Lambda {d^D q\over(2\pi)^D}{1\over q^2+m^2},$$

show by power counting that $I_D(0)$ has a UV divergence proportional to $\Lambda^{D-2}$ for $D>2$. What happens at $D=2$?

Solution

Using spherical coordinates,

$$I_D(0) = {S_{D-1}\over(2\pi)^D} \int_0^\Lambda dq\,{q^{D-1}\over q^2} = {S_{D-1}\over(2\pi)^D} \int_0^\Lambda dq\,q^{D-3}.$$

For $D>2$,

$$\int_0^\Lambda dq\,q^{D-3} = {\Lambda^{D-2}\over D-2},$$

up to the infrared lower limit. Therefore

$$I_D(0)\sim {S_{D-1}\over(2\pi)^D} {\Lambda^{D-2}\over D-2}.$$

At $D=2$, the integral becomes

$$\int_0^\Lambda {dq\over q},$$

which is logarithmic. In the massless theory this logarithm is also infrared divergent. With nonzero $m$, the denominator $q^2+m^2$ cuts off the infrared region and produces a logarithm of $\Lambda/m$.

Exercise 4

Find the engineering dimensions of $\phi$ and $\lambda$ in

$$S=\int d^D x\, \left[ {1\over2}(\partial\phi)^2 +{\lambda\over4!}\phi^4 \right].$$

Use them to identify the upper critical dimension.

Solution

The action is dimensionless. Since $[d^D x]=-D$ and $[\partial]=1$, the kinetic term gives

$$-D+2+2[\phi]=0.$$

Hence

$$[\phi]={D-2\over2}.$$

The quartic interaction gives

$$-D+[\lambda]+4[\phi]=0.$$

Substituting the field dimension,

$$[\lambda]=D-4[\phi] = D-2(D-2) = 4-D.$$

The coupling is dimensionless when

$$D=4.$$

Therefore the upper critical dimension of the $\phi^4$ interaction is

$$D_c=4.$$

For $D>4$, $[\lambda]<0$ and the interaction is irrelevant at the Gaussian fixed point. For $D<4$, $[\lambda]>0$ and it is relevant.

Exercise 5

At the Gaussian critical point, let

$$G_0(R)\sim {1\over R^{D-2}}.$$

Estimate the first-order correction from $\lambda\phi^4$ to

$$C(R)=\langle \phi^2(x_1)\phi^2(x_2)\rangle, \qquad R=|x_1-x_2|.$$

Show that the relative correction scales as $\lambda R^{4-D}$.

Solution

The Gaussian correlator is

$$C_0(R)\sim G_0(R)^2 \sim {1\over R^{2D-4}}.$$

The first correction from one insertion of the interaction is

$$\delta C(R) \sim \lambda \int d^D z\, G_0(x_1-z)^2G_0(x_2-z)^2.$$

For a scaling estimate at separation $R$, the integration region has volume $R^D$, and each propagator contributes a factor $R^{-(D-2)}$. Since there are four propagators,

$$\delta C(R) \sim \lambda R^D \left(R^{-(D-2)}\right)^4 = \lambda R^{D-4D+8} = \lambda R^{-3D+8}.$$

Therefore

$${\delta C(R)\over C_0(R)} \sim {\lambda R^{-3D+8}\over R^{-2D+4}} = \lambda R^{4-D}.$$

The correction decreases with $R$ for $D>4$, grows for $D<4$, and is marginal by power counting for $D=4$.

Exercise 6

Use the Schwinger representation

$${1\over k^2+m^2}=\int_0^\infty dL\,e^{-L(k^2+m^2)}$$

to show that a critical Gaussian propagator describes paths with typical length $L\sim R^2$ for endpoint separation $R$.

Solution

Fourier transforming the Schwinger representation gives

$$G_0(x) = \int_0^\infty dL \int {d^D k\over(2\pi)^D} e^{ik\cdot x}e^{-Lk^2}e^{-Lm^2}.$$

The Gaussian momentum integral is

$$\int {d^D k\over(2\pi)^D} e^{ik\cdot x}e^{-Lk^2} = {1\over(4\pi L)^{D/2}} \exp\left(-{x^2\over4L}\right).$$

Thus

$$G_0(x) = \int_0^\infty dL\, {1\over(4\pi L)^{D/2}} \exp\left(-{R^2\over4L}-m^2L\right).$$

At criticality $m=0$. The exponential factor suppresses $L\ll R^2$, because then $R^2/(4L)$ is large. The power $L^{-D/2}$ suppresses very large $L$ in dimensions where the integral is convergent. The natural scaling variable is

$$u={L\over R^2}.$$

Therefore the dominant path length scales as

$$L\sim R^2.$$

This is the Brownian scaling relation: the path has fractal dimension $2$.

References

• Ginzburg, Vitaly L. “Some Remarks on Phase Transitions of the Second Kind and the Microscopic Theory of Ferroelectric Materials.” Soviet Physics—Solid State 2 (1960): 1824–1834.
• Wilson, Kenneth G., and Michael E. Fisher. “Critical Exponents in 3.99 Dimensions.” Physical Review Letters 28, no. 4 (1972): 240–243. https://doi.org/10.1103/PhysRevLett.28.240.
• Wilson, Kenneth G., and John Kogut. “The Renormalization Group and the $\epsilon$ Expansion.” Physics Reports 12, no. 2 (1974): 75–199. https://doi.org/10.1016/0370-1573(74)90023-4.

Further reading

• Cardy, John. Scaling and Renormalization in Statistical Physics. Cambridge Lecture Notes in Physics 5. Cambridge: Cambridge University Press, 1996. Chapters 2–4.
• Goldenfeld, Nigel. Lectures on Phase Transitions and the Renormalization Group. Frontiers in Physics 85. Reading, MA: Addison-Wesley, 1992. Chapters 4–6.
• Polyakov, Alexander M. Gauge Fields and Strings. Contemporary Concepts in Physics 3. Chur: Harwood Academic Publishers, 1987. Chapters 1–3 and 9.
• Zinn-Justin, Jean. Quantum Field Theory and Critical Phenomena. 5th ed. Oxford: Oxford University Press, 2021. Chapters 14–17.