Momentum-Shell Renormalization Group

Real-space block spins are intuitive, but they carry a nagging ambiguity: what exactly counts as a "short-distance degree of freedom" when interactions spread across a lattice? Kenneth Wilson's answer was to move to Fourier space, where the question has a crisp answer — a short-distance degree of freedom is a field mode $\phi(\mathbf{q})$ with large wavevector. Define the theory with an ultraviolet cutoff $\Lambda$ (the smoothed inverse lattice spacing), and coarse-graining becomes something you can do exactly for the free theory and systematically for the interacting one: integrate out the modes in a thin shell $\Lambda/b<|\mathbf{q}|<\Lambda$ near the cutoff. The procedure is three clean steps. First coarse-grain: split the field into slow modes ( $|\mathbf{q}|<\Lambda/b$ ) and fast modes (in the shell), and integrate out the fast modes. Second rescale momenta, $\mathbf{q}'=b\,\mathbf{q}$ , to restore the cutoff to $\Lambda$ . Third rescale the field, $\phi'=\phi/\zeta$ , to bring the gradient term back to canonical form. The composition is a transformation on the space of couplings, and taking the shell infinitesimally thin ( $b=e^{\ell}$ , $\ell\to0$ ) turns the discrete map into a continuous flow. This article owns the foundations of that machinery: the exact Gaussian shell integration, the engineering dimensions that fall out of pure rescaling, and the tree-level result that the mass term is always relevant ( $y_r=2$ ) while the quartic coupling has $y_u=4-d$ — which changes sign at the upper critical dimension $d_c=4$ . That single sign change is the whole story of why mean-field theory works above four dimensions and fails below. The full nonperturbative payoff — the Wilson–Fisher fixed point and the $\epsilon=4-d$ expansion of the exponents — is developed in its own article; here we lay the scaffold on which it stands, and quote its leading result by name.

Key takeaways

Wilson's momentum-shell RG coarse-grains in Fourier space: integrate out field modes in a thin shell $\Lambda/b<|\mathbf{q}|<\Lambda$ , then rescale momenta and fields to restore the original cutoff.
Demanding the gradient term stay canonical fixes the field rescaling $\zeta=b^{(d+2-\eta)/2}$ (with $\eta=0$ at the Gaussian level); loop corrections to the $q^2$ coefficient define the anomalous dimension $\eta$ .
Pure rescaling gives the tree-level recursions $r'=b^2 r$ and $u'=b^{4-d}u$ , so $y_r=2$ (mass always relevant) and $y_u=4-d$ .
$y_u$ changes sign at $d=4$ : irrelevant above (Gaussian fixed point, mean-field exponents), relevant below — this is the upper critical dimension $d_c=4$ .
Taking the shell infinitesimal turns the discrete map into continuous beta-function flow $dg_i/d\ell=\beta_i(\{g\})$ ; the interacting Wilson–Fisher fixed point and $\epsilon$ -expansion build on this foundation.
Key idea: integrate out a thin momentum shell, then rescale momenta and fields to restore the theory; the resulting flow of couplings has fixed points and linearized eigenvalues that give universal critical exponents, and the tree-level power counting alone fixes $d_c=4$ .

The core insight

Integrate out a thin momentum shell near the cutoff, then rescale momenta and fields to restore the original theory: the resulting transformation of the couplings, taken in the infinitesimal-shell limit, is a continuous flow whose fixed points and linearized eigenvalues give universal critical exponents.

Where this lives today

The momentum-shell RG is the analytic workhorse of critical phenomena. Applied to the single-component $\phi^4$ theory it gives the $\epsilon$ -expansion of the 3D Ising exponents; extended to an $n$ -component order parameter it covers the XY ( $n=2$ ), Heisenberg ( $n=3$ ), and polymer ( $n\to0$ ) universality classes from one machine, with the Wilson–Fisher coupling scaling as $u^*\propto\epsilon/(n+8)$ . Its modern descendant is the functional (exact) renormalization group of Wetterich and Polchinski, which replaces the sharp shell with a smooth momentum regulator and flows the entire effective action rather than a few couplings — the tool of choice for strongly correlated electrons, the quark–gluon plasma, and even quantum gravity's asymptotic-safety program. The same tree-level power-counting that fixes $d_c=4$ here is what particle physicists call "dimensional analysis of operators," and it decides which interactions matter at low energy in every effective field theory, from the Standard Model to the theory of sound in a fluid. Wilson's thin shell turned out to be one of the most portable ideas in physics.

The physical picture

Picture the field's information as spread across momentum scales. A microscope with finite resolution simply cannot see modes near the cutoff; the RG asks what effective theory describes the modes it can see. Integrating the shell is exactly "averaging out" the finest unresolved wiggles — but done in Fourier space, where "fine" means "large $|\mathbf{q}|$ ." This is the same coarse-grain-and-rescale program as the renormalization group idea and its real-space block-spin cousin, now with an unambiguous notion of "short distance."

Figure

Momentum-shell coarse-graining: integrate out the fast modes in the shell Λ/b ＜ |q| ＜ Λ, keep the slow modes, then rescale q → bq

Rescaling is what makes the procedure a true flow rather than a one-shot blur. After removing the shell, the slow theory has a smaller cutoff $\Lambda/b$ ; magnifying lengths (equivalently, shrinking momenta back to $\Lambda$ ) lets us compare the new theory with the old on the same footing. A fixed point is the self-similar theory that looks identical at every magnification — scale invariance at criticality, expressed in momentum space. The field rescaling factor $\zeta$ is the subtle ingredient, and it is worth pinning down carefully because a common convention slip lives here. The Gaussian part of the action dictates its value: demanding that the coefficient of the $q^2$ gradient term stay fixed after coarse-graining and momentum rescaling forces a specific $\zeta$ . When loop corrections shift the propagator's momentum dependence, absorbing that shift into $\zeta$ introduces a small correction — the anomalous dimension $\eta$ — measuring how the field's scaling deviates from naive dimensional analysis.

Figure

Tree-level eigenvalues: the mass r is always relevant (y_r ＝ 2); the quartic u has y_u ＝ 4 − d, changing sign at d ＝ 4

The last figure is the punchline of the whole method's foundation. Two straight lines: the mass eigenvalue $y_r=2$ sits flat and positive (the mass is always relevant), while the quartic eigenvalue $y_u=4-d$ slopes downward and crosses zero exactly at $d=4$ . Above four dimensions the quartic coupling is irrelevant and mean-field theory is exact; below, it is relevant and drives the system to a new fixed point. That crossing point is the upper critical dimension. Relevant couplings (like the reduced temperature) grow under the flow and drive the system away from criticality; irrelevant ones decay, which is why universality holds — the number of relevant directions equals the number of parameters you must tune, typically two for the Ising class.

The core equations

We work with the Landau–Ginzburg–Wilson action for a scalar order-parameter field, defined with an ultraviolet momentum cutoff $\Lambda$ .

Z=\int\mathcal{D}\phi\;e^{-\mathcal{H}[\phi]},\qquad \mathcal{H}=\int_{|\mathbf{q}|<\Lambda}\frac{d^dq}{(2\pi)^d}\,\tfrac12\big(r+q^2\big)|\phi(\mathbf{q})|^2 + u\!\int(\phi^2)^2 \qquad (1)

where: $\phi(\mathbf{q})$ — Fourier mode of the order-parameter field; $\Lambda$ — ultraviolet cutoff (inverse microscopic length); $r$ — quadratic coupling, $\propto(T-T_c)$ at mean-field level; $q^2$ — gradient (stiffness) coefficient, normalized to one; $u$ — quartic self-coupling. The Gaussian part $\tfrac12(r+q^2)|\phi|^2$ is the free theory; $u$ is the interaction that makes criticality nontrivial. The first step splits the field by wavevector into slow and fast modes.

\phi(\mathbf{q})=\phi^{<}(\mathbf{q})\,\theta\!\big(\tfrac{\Lambda}{b}-|\mathbf{q}|\big)+\phi^{>}(\mathbf{q})\,\theta\!\big(|\mathbf{q}|-\tfrac{\Lambda}{b}\big)\theta\!\big(\Lambda-|\mathbf{q}|\big) \qquad (2)

where: $\phi^{<}$ — slow modes retained ( $|\mathbf{q}|<\Lambda/b$ ); $\phi^{>}$ — fast modes in the shell $\Lambda/b<|\mathbf{q}|<\Lambda$ ; $b>1$ — RG scale factor; $\theta$ — Heaviside step enforcing the shell boundaries. Coarse-graining is then the exact functional integral over the fast modes,

e^{-\mathcal{H}'[\phi^{<}]}=\int\mathcal{D}\phi^{>}\;e^{-\mathcal{H}[\phi^{<}+\phi^{>}]}, \qquad (3)

producing the effective action $\mathcal{H}'$ for the slow modes. Now the rescaling. Restore the cutoff with $\mathbf{q}'=b\,\mathbf{q}$ and rescale the field by a factor $\zeta$ chosen to keep the gradient term canonical.

\mathbf{q}'=b\,\mathbf{q},\qquad \phi'(\mathbf{q}')=\zeta^{-1}\phi^{<}(\mathbf{q}),\qquad \zeta=b^{(d+2-\eta)/2} \qquad (4)

where: $\mathbf{q}'$ — rescaled momentum restoring the cutoff to $\Lambda$ ; $\zeta$ — field rescaling factor; $\eta$ — anomalous dimension. At the Gaussian level $\eta=0$ and $\zeta=b^{(d+2)/2}$ ; the derivation below shows this follows from demanding gradient-term invariance. (A frequent slip: the momentum-space field-rescale factor is $\zeta=b^{(d+2-\eta)/2}$ , not $b^{(d-2+\eta)/2}$ — the latter is the real-space scaling dimension $\Delta_\phi=(d-2+\eta)/2$ of the field, a related but distinct object. Confusing the two flips the tree-level exponents.) Pure rescaling of each term of the action gives its tree-level recursion — the engineering dimensions.

r'=b^{2}\,r,\qquad u'=b^{4-d}\,u\qquad(\text{Gaussian / tree level}) \qquad (5)

where: $r'$ — rescaled quadratic coupling; $u'$ — rescaled quartic coupling. So $y_r=2$ (mass always relevant) and $y_u=4-d$ , which is positive for $d<4$ , zero at $d=4$ , negative for $d>4$ . Taking the shell infinitesimally thin turns the discrete map into continuous flow.

\frac{dg_i}{d\ell}=\beta_i(\{g\}),\qquad b=e^{\ell} \qquad (6)

where: $\ell=\ln b$ — logarithmic RG "time"; $\beta_i$ — the beta function of coupling $g_i$ . Fixed points ( $\beta_i=0$ ), the eigenvalues of the linearized flow, and the relevant/irrelevant classification then follow by ordinary-differential-equation analysis. The eigenvalues map to scaling exponents through

\lambda_i=b^{y_i},\qquad y_i=\frac{\ln\lambda_i}{\ln b}, \qquad (7)

interpreted by linearization about a fixed point and the operator classification, with the exponents themselves owned by critical exponents from the RG.

The core derivations

Derivation AThe Gaussian fixed point and the engineering dimensions

The free theory is where momentum-shell RG is exact, and everything universal about the tree level lives here. The goal is to show that integrating out the shell in the Gaussian theory is trivial, that the rescaling alone determines the field factor $\zeta$ and the recursions $r'=b^2 r$ , $u'=b^{4-d}u$ , and hence that the Gaussian fixed point sits at the origin with $y_t=2$ , $y_u=4-d$ . The strategy is pure power-counting: track how the measure, the momentum, and the fields each scale under $\mathbf{q}'=b\mathbf{q}$ , $\phi'=\zeta^{-1}\phi^<$ , and demand each term of the action return to canonical form. The one thing to get right is the direction of every power of $b$ — a single sign error here propagates into wrong exponents.

Step 1. Shell integration is trivial. With $u=0$ the action is a decoupled Gaussian, so $\int\mathcal{D}\phi^{>}e^{-\mathcal{H}_0[\phi^{>}]}$ is a field-independent constant; $\mathcal{H}'[\phi^{<}]$ is just $\mathcal{H}_0$ restricted to $|\mathbf{q}|<\Lambda/b$ . No new couplings are generated.

Step 2. Fix $\zeta$ from the gradient term. Under $\mathbf{q}'=b\mathbf{q}$ the measure gives $d^dq=b^{-d}d^dq'$ and $q^2=b^{-2}q'^2$ ; the field gives $\phi^<=\zeta\phi'$ . Then

\tfrac12\!\int d^dq\,q^2|\phi^<|^2=\tfrac12\,b^{-d-2}\zeta^2\!\int d^dq'\,q'^2|\phi'|^2.

Canonical form (coefficient back to $\tfrac12$ ) requires $b^{-d-2}\zeta^2=1$ , i.e. $\zeta=b^{(d+2)/2}$ (so $\eta=0$ at the Gaussian fixed point). Why this and not $b^{(d-2)/2}$ : the $b^{-d}$ from the measure and $b^{-2}$ from $q^2$ must be undone by $\zeta^2$ ; solving gives $\zeta^2=b^{d+2}$ . The exponent $(d-2)/2$ is the real-space scaling dimension of the field, obtained by a different bookkeeping.

Step 3. Recursion for $r$ . The mass term $\tfrac12 r\int d^dq|\phi|^2$ carries one measure and two fields:

r\,b^{-d}\zeta^2=r\,b^{-d}b^{d+2}=r\,b^{2}\ \Rightarrow\ r'=b^2 r,\quad y_r=2.

Step 4. Recursion for $u$ . The quartic $u\int(d^dq)^3(\phi)^4$ carries three measures and four fields:

u\,b^{-3d}\zeta^4=u\,b^{-3d}b^{2d+4}=u\,b^{4-d}\ \Rightarrow\ u'=b^{4-d}u,\quad y_u=4-d.

Step 5. Fixed point and eigenvalues. The only fixed point is $r^*=u^*=0$ (the Gaussian fixed point). The map is already diagonal, so the eigenvalues are $b^2$ and $b^{4-d}$ : $y_t=2$ gives $\nu=1/y_t=1/2$ (the mean-field value), and $u$ is irrelevant precisely for $d>4$ , marking the upper critical dimension $d_c=4$ . Verified: `verify_momentum-shell-rg.py` solves $b^{-d-2}\zeta^2=1$ for $\zeta=b^{(d+2)/2}$ , derives $r'=b^2 r$ and $u'=b^{4-d}u$ by power-counting, and confirms the sign change of $y_u$ at $d=4$ (checks 1–5).

Go deeper · Derivation B · 2 minThe Gaussian shell integral and the onset of interactions

Derivation BThe Gaussian shell integral and the onset of interactions

Having fixed the tree level, we set up — but do not fully execute — the interacting calculation, showing concretely where the Wilson–Fisher fixed point comes from while leaving its full construction to the dedicated article. The goal is to write down the Gaussian shell propagator, confirm the shell integral is finite and positive, and exhibit the structure of the one-loop corrections to $r$ and $u$ that generate a nontrivial fixed point for $d<4$ . The strategy: expand $e^{-\mathcal{H}_{\rm int}}$ in the quartic coupling and average over the fast modes with the Gaussian weight, whose propagator in the shell is $1/(r+q^2)$ . We stop at identifying the fixed-point condition; the loop integrals and the $\epsilon$ -expansion of the exponents are owned by the Wilson–Fisher article.

Step 1. The shell propagator. For fast modes in the shell, the Gaussian two-point function is $G^{>}(\mathbf{q})=1/(r+q^2)$ , and the one-loop corrections come from contracting fast-mode pairs in the interaction.

Step 2. The shell integral is finite. The relevant integral $\int_{\Lambda/b<|\mathbf{q}|<\Lambda}\frac{d^dq}{(2\pi)^d}\,\frac{1}{r+q^2}$ runs over a compact shell of a smooth positive integrand, so it is finite and positive; for a thin shell it is proportional to the shell width $d\ell=\ln b$ .

Step 3. Structure of the corrections. The tadpole (one contracted pair in $(\phi^2)^2$ ) shifts $r$ , and the bubble (two four-point contractions) shifts $u$ :

\frac{dr}{d\ell}=2r+\frac{Au}{1+r},\qquad \frac{du}{d\ell}=(4-d)\,u-\frac{Bu^2}{(1+r)^2},

with $A,B>0$ collecting combinatorial and angular (unit-sphere) factors. The tree-level pieces $2r$ and $(4-d)u$ are exactly the engineering dimensions from the previous derivation; the loop terms are the new physics.

Step 4. Where the new fixed point comes from. Setting $du/d\ell=0$ gives either $u^*=0$ (Gaussian) or the nonzero root

u^*=\frac{(4-d)(1+r^*)^2}{B}=\frac{\epsilon}{B}+O(\epsilon^2),\qquad \epsilon=4-d,

which exists only for $\epsilon>0$ ( $d<4$ ). This is the Wilson–Fisher fixed point; its full construction, stability analysis, and the $\epsilon$ -expansion of the exponents are developed there. We have only located the condition that produces it. Verified: `verify_momentum-shell-rg.py` confirms the Gaussian shell integral is finite and positive ( $d=3$ shell), and that $du/d\ell=0$ yields the nonzero root $u^*=\epsilon(1+r^*)^2/B$ (checks 6–7); the quoted $\nu=1/2+\epsilon/12$ is verified as the Wilson–Fisher result (check 8), attributed by name.

Limiting cases

The behaviour of the shell RG depends sharply on dimension and on the chosen scale factor. The exponents in the last column are owned by the critical-exponents and Wilson–Fisher articles; here they interpret the flow this method sets up.

Case	Result	Meaning
$d>4$ (above upper critical dimension)	$y_u=4-d<0$	The quartic coupling is irrelevant; the flow reaches the Gaussian fixed point and mean-field exponents ( $\nu=1/2$ , $\eta=0$ ) hold exactly.
$d=4$ (marginal)	$du/d\ell\propto -u^2$	$u$ is marginally irrelevant: it decays logarithmically, giving mean-field exponents with logarithmic corrections.
$d<4$ (Wilson–Fisher regime)	$u^*\sim\epsilon=4-d$	A new stable fixed point at $O(\epsilon)$ controls criticality; nontrivial exponents organized in powers of $\epsilon$ (owned by the WF article).
Infinitesimal shell $b=e^\ell,\ \ell\to0$	$dg_i/d\ell=\beta_i(\{g\})$	Discrete recursion becomes autonomous ODE flow — the natural, standard form of the momentum-shell RG.

Common trapThe sharp cutoff and the field-rescale convention both bite

Two cautions. First, a sharp momentum cutoff is only consistent at one-loop order; at higher orders it produces non-analytic, scheme-dependent terms, and the shell expansion does not commute naively with $b\to1$ for arbitrary observables — smooth cutoffs or dimensional regularization are used beyond one loop. Second, the field rescaling factor $\zeta$ is a choice dictated by keeping the gradient term canonical: $\zeta=b^{(d+2-\eta)/2}$ in momentum space, with $\eta$ defined by the loop shift of the $q^2$ coefficient — not an independent physical input. Writing $\zeta=b^{(d-2+\eta)/2}$ (the real-space scaling dimension) is a genuine and common error that flips the tree-level recursions; this article's base draft made exactly that slip in one place, corrected here.

Worked examples

Example 1 — The engineering dimensions, from power-counting alone. We compute the tree-level eigenvalues explicitly, the calculation that fixes the upper critical dimension.

Step 1. From the Gaussian gradient term, $\zeta=b^{(d+2)/2}$ (Step 2 of the first derivation). Take a concrete $d=3$ : $\zeta=b^{5/2}$ .

Step 2. Mass recursion: $r'=r\,b^{-d}\zeta^2=r\,b^{-3}b^{5}=r\,b^{2}$ , so $y_r=2$ , independent of $d$ .

Step 3. Quartic recursion: $u'=u\,b^{-3d}\zeta^4=u\,b^{-9}b^{10}=u\,b^{1}$ , so $y_u=1$ at $d=3$ . Since $y_u=4-d$ , this is positive — the quartic is relevant in three dimensions, so mean-field theory fails and a Wilson–Fisher fixed point exists.

Step 4. Repeat at $d=5$ : $\zeta=b^{7/2}$ , $u'=u\,b^{-15}b^{14}=u\,b^{-1}$ , so $y_u=-1<0$ — irrelevant, mean-field exact. The sign flip between $d=3$ and $d=5$ brackets the upper critical dimension at $d_c=4$ , where $y_u=0$ .

Example 2 — The Gaussian fixed point gives the mean-field correlation-length exponent. We check that the tree-level thermal eigenvalue reproduces the known mean-field $\nu$ .

Step 1. At the Gaussian fixed point the thermal eigenvalue is $y_t=y_r=2$ (the mass is the thermal scaling field, $r\propto T-T_c$ ).

Step 2. The correlation-length exponent is $\nu=1/y_t=1/2$ .

Step 3. Cross-check against mean-field theory: Landau theory gives $\xi\sim|t|^{-1/2}$ , i.e. $\nu=1/2$ . They agree — as they must, since above $d_c=4$ the Gaussian fixed point is the exact critical theory. $\nu=1/2$ (mean-field), exact for $d>4$ .

Step 4. For $d<4$ the flow leaves the Gaussian fixed point along the relevant $u$ direction and settles at the Wilson–Fisher fixed point, where $y_t=2-\epsilon/3+O(\epsilon^2)$ and hence $\nu=1/2+\epsilon/12+\cdots$ — the standard one-loop result, quoted here by name from the Wilson–Fisher article and verified numerically.

Origins · 1 minwilsonian renormalization and how this was discovered

wilsonian renormalization

Wilson's momentum-space formulation of the RG (1971) · Wilsonian renormalization is the viewpoint, introduced by Kenneth Wilson around 1971, that a physical theory is defined with a finite cutoff and that lowering the cutoff by integrating out high-momentum modes generates an effective theory with running couplings. In this picture renormalizability is not a formal requirement but a statement about which couplings are relevant near a fixed point, and the same machinery that tames ultraviolet divergences in quantum field theory computes critical exponents in statistical mechanics. The momentum-shell implementation — integrate a thin shell, rescale, repeat — is the concrete realization that turned this viewpoint into a calculational tool and earned Wilson the 1982 Nobel Prize.

History

1954

— Gell-Mann and Low introduce renormalization-group functional equations in quantum electrodynamics (Physical Review 95, 1300), the field-theory origin of the idea

1966

— Kadanoff's block-spin scaling supplies the real-space intuition Wilson would formalize

1971

— Wilson formulates the momentum-space renormalization group and the notion of integrating out modes shell by shell (Physical Review B 4)

1972

— Wilson and Fisher introduce the ε ＝ 4 − d expansion, computing critical exponents in 3.99 dimensions (Physical Review Letters 28, 240)

1982

— Wilson receives the Nobel Prize in Physics for the theory of critical phenomena in connection with phase transitions

Going further

References. Kardar, Statistical Physics of Fields (2007), Ch. 5 · Pathria & Beale, Statistical Mechanics (3rd ed.), Ch. 14 · Goldenfeld, Lectures on Phase Transitions and the RG (1992), Ch. 9 · Wilson & Kogut, Phys. Rep. 12 (1974) 75 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity (2nd ed.), Ch. 12.

Provenance. Equations and structure follow Kardar, Statistical Physics of Fields Ch. 5, and Goldenfeld Ch. 9. A verifier should recompute: (1) the engineering dimensions $y_r=2$ , $y_u=4-d$ from the rescaling power-counting; (2) the field rescaling $\zeta=b^{(d+2-\eta)/2}$ (momentum space) from gradient-term invariance; (3) the finiteness of the Gaussian shell integral; (4) the fixed-point condition $u^*=\epsilon(1+r^*)^2/B$ and the quoted one-loop $\nu=1/2+\epsilon/12$ (owned by the Wilson–Fisher article). Standard reference numbers: mean-field $\nu=1/2$ ; 3D Ising $\nu\approx0.630$ , $\eta\approx0.036$ , $\omega\approx0.8$ (conformal bootstrap / Monte Carlo).

Note on a corrected base error. The source draft's physical-picture prose wrote the field-rescale factor as $\zeta=b^{(d-2+\eta)/2}$ ; the correct momentum-space factor is $\zeta=b^{(d+2-\eta)/2}$ (the former is the real-space scaling dimension $\Delta_\phi$ ). The tree-level recursions $r'=b^2 r$ , $u'=b^{4-d}u$ follow only from the correct factor; this has been fixed throughout and flagged as a teaching point.

Builds on: The Renormalization Group Idea: Coarse-Graining and Rescaling · Block-Spin (Real-Space) Renormalization Group · RG Flow in Coupling Space. Leads to: Wilson-Fisher Fixed Point · Gaussian Fixed Point and Upper Critical Dimension · The Epsilon-Expansion · Critical Exponents from the RG.

Why work in momentum space instead of real space like the block-spin method?

Because "short-distance degree of freedom" is ambiguous on a lattice — interactions spread, and it is not obvious which spins to average — but in Fourier space it is crisp: a short-distance mode is one with large wavevector $|\mathbf{q}|$ . Integrating out a thin momentum shell near the cutoff is therefore unambiguous, and for the Gaussian theory it is exact. The trade-off is that you need a continuum field-theory description (the Landau–Ginzburg–Wilson action) rather than a lattice model, and you gain a systematic small parameter, $\epsilon=4-d$ , that the real-space method lacks. The two approaches give the same fixed points and exponents.

What exactly is the anomalous dimension

\eta

, and why is it zero at the Gaussian fixed point?

$\eta$ measures how the field's scaling deviates from naive dimensional analysis. At the Gaussian (free) fixed point the field rescaling $\zeta=b^{(d+2)/2}$ keeps the $q^2$ gradient term exactly canonical, so there is no deviation and $\eta=0$ . Once interactions are switched on, loop corrections modify the momentum dependence of the propagator — the coefficient of $q^2$ shifts — and to keep the gradient term canonical you must adjust $\zeta$ to $b^{(d+2-\eta)/2}$ . That required adjustment is $\eta$ . It is small: for the 3D Ising class $\eta\approx0.036$ , so the field scales almost, but not quite, by its engineering dimension.

Why does the quartic coupling being irrelevant above four dimensions mean mean-field theory is exact there?

Mean-field (Landau) theory ignores fluctuations, keeping only the Gaussian part of the action plus the uniform value of the quartic term. The RG makes precise when that is justified: if the quartic coupling $u$ is irrelevant ( $y_u=4-d<0$ , i.e. $d>4$ ), it flows to zero under coarse-graining, so at long distances the theory is Gaussian and mean-field exponents hold exactly. For $d<4$ , $u$ is relevant, grows under the flow, and the system is driven to the interacting Wilson–Fisher fixed point where fluctuations shift the exponents. This is the Ginzburg criterion made rigorous.

Where does the Wilson–Fisher fixed point come from, and why isn't it derived in full here?

It comes from the competition in the quartic flow $du/d\ell=(4-d)u-Bu^2/(1+r)^2$ : the tree-level term $(4-d)u$ pushes $u$ up (for $d<4$ ) while the one-loop bubble term $-Bu^2$ pushes it down, and they balance at $u^*\propto\epsilon=4-d$ . This article sets up that condition to show where the fixed point originates, but its full construction — the loop integrals, the stability analysis, and the $\epsilon$ -expansion of every exponent — is owned by the dedicated Wilson–Fisher fixed-point article to keep the scope clean. Here we establish only the Gaussian foundation and the tree-level power-counting the WF calculation is built on.

Why is

r^*

negative at the interacting fixed point?

The tadpole correction $+Au/(1+r)$ to $dr/d\ell$ is positive when $u^*>0$ , so for the flow to sit still ( $dr/d\ell=0$ ) the tree-level term $2r$ must be negative — forcing $r^*<0$ to leading order. Physically, the interaction shifts the critical temperature: the bare mass parameter $r$ at which the system is actually critical lies below its mean-field value of zero, because fluctuations help disorder the system, so you must start colder (more negative $r$ ) to reach criticality. The precise value is part of the Wilson–Fisher calculation.

Is the sharp momentum-shell cutoff a problem?

At one loop it is fine and physically transparent. Beyond one loop a sharp cutoff generates non-analytic, scheme-dependent contributions, and the shell procedure does not commute cleanly with the infinitesimal-shell limit for every observable. Practitioners therefore switch to smooth cutoffs (as in the functional RG) or to dimensional regularization for higher-order calculations. None of this affects the universal quantities — the exponents are scheme-independent — but it does affect intermediate, non-universal numbers, so the sharp cutoff is best regarded as a pedagogical and one-loop tool rather than a high-precision engine.