Wilson's Effective Action and the Exact RG Equation

The block-spin and momentum-shell renormalization group both work the same way: integrate out a shell of fast modes, then rescale, and watch a handful of couplings flow. Wilson's deeper realization was that this procedure does not really track a few numbers — it tracks the entire action functional. Define the Wilsonian effective action $S_\Lambda[\phi]$ by integrating out every field mode with momentum above a running cutoff $\Lambda$ while holding the slow field fixed. Lower $\Lambda$ by an infinitesimal amount and $S_\Lambda$ changes in a way that leaves the partition function and all long-distance observables exactly invariant. That change is governed by an exact functional differential equation — the exact renormalization group, or ERG. Renormalization stops being an approximation scheme and becomes an exact statement about a trajectory in the infinite-dimensional space of all possible actions. Two formulations dominate. The Polchinski equation flows the Wilsonian action itself, with a smooth cutoff function inserted into the propagator; it is the instrument behind the cleanest modern proofs that renormalizable theories have only finitely many relevant couplings. The Wetterich equation instead flows the effective average action $\Gamma_k$ — the generator of one-particle-irreducible vertices with only modes faster than an infrared scale $k$ integrated out. Its right-hand side is a single trace with the structure of one loop, yet the equation is exact, because the full field-dependent inverse propagator appears inside the trace rather than the bare one. This deceptive simplicity has made it the standard tool of the functional RG. Everything from earlier in this chapter reappears here as a property of a continuous flow on a function space: fixed points are scale-invariant theories, relevant and irrelevant operators are the eigen-directions of the linearized flow, and the Gaussian and Wilson-Fisher fixed points, the upper critical dimension, and the ε-expansion all emerge as limits or truncations of the same functional equation. Because the exact equations cannot be solved in closed form, one truncates — the derivative expansion (leading term: the local potential approximation) or the vertex expansion — and these non-perturbative truncations already deliver accurate three-dimensional critical exponents. The ERG is thus both the conceptual foundation of Wilson's renormalization group and a quantitatively competitive calculational tool.

Key takeaways

Wilson's effective action $S_\Lambda[\phi]$ encodes the physics of all modes below a running cutoff $\Lambda$ after faster modes are integrated out; lowering $\Lambda$ generates an exact, continuous flow in the infinite-dimensional space of couplings.
Demanding that the partition function be independent of the cutoff yields the Polchinski equation, an exact functional differential equation for the Wilsonian action with a smooth regulator.
An equivalent, computationally powerful form is the Wetterich equation for the effective average action $\Gamma_k$ : a one-loop-exact flow $\partial_t\Gamma_k=\tfrac12\,\mathrm{Tr}[(\Gamma_k^{(2)}+R_k)^{-1}\partial_t R_k]$ .
These exact-RG equations reproduce perturbation theory, the ε-expansion, and the Wilson-Fisher fixed point, but are genuinely non-perturbative once truncated (derivative or vertex expansion).
Fixed points of the flow are scale-invariant theories; the linearized flow around them gives critical exponents through its eigenvalue spectrum — the continuous version of momentum-shell RG.
Key idea: integrating out fast modes continuously turns the renormalization group into an exact functional flow, and the change of the effective average action under an infinitesimal cutoff variation is a one-loop-exact trace, $\partial_t\Gamma_k=\tfrac12\,\mathrm{Tr}\,[(\Gamma_k^{(2)}+R_k)^{-1}\,\partial_t R_k]$ .

The core insight

Integrating out fast modes continuously turns the renormalization group into an exact functional flow: the change of the effective average action under an infinitesimal cutoff variation is given by a one-loop-exact trace, $\partial_t\Gamma_k=\tfrac12\,\mathrm{Tr}\,[(\Gamma_k^{(2)}+R_k)^{-1}\,\partial_t R_k]$ .

Where this lives today

The exact renormalization group is the standard non-perturbative tool wherever perturbation theory fails and lattice methods are expensive. Solving the Wetterich equation in the derivative expansion gives accurate critical exponents for the three-dimensional O( $n$ ) universality classes — for the 3D Ising model $\nu\approx0.63$ and $\eta\approx0.04$ , competitive with conformal-bootstrap and Monte-Carlo results, all from truncating a single functional equation. The Polchinski equation, meanwhile, underlies some of the cleanest proofs that renormalizable field theories contain only finitely many relevant couplings, realizing Wilson's picture of renormalization as rigorous mathematics. Beyond critical phenomena, functional RG flows of the effective average action probe strongly coupled quantum field theory: chiral symmetry breaking and the phase diagram of the quark-meson model in QCD, where standard perturbation theory breaks down, and — most speculatively — the asymptotic-safety program in quantum gravity, which applies the Wetterich equation to the gravitational effective action in search of a non-trivial ultraviolet fixed point. In every case the same idea does the work: integrate fluctuations out continuously from the ultraviolet to the infrared, and track the entire action along the way rather than a handful of couplings.

The physical picture

Imagine the partition function written with a knob $\Lambda$ that declares "only modes slower than $\Lambda$ remain as integration variables; everything faster is already baked into the action." Turning the knob down forces the information in the just-integrated shell into the couplings of the action. The exact-RG equation is nothing more than the bookkeeping that keeps the total probability — the partition function itself — fixed as the knob turns. Because the shell is infinitesimal, the bookkeeping closes into a differential equation rather than a discrete map, and the discrete momentum-shell recursion becomes a continuous trajectory through the space of all actions. The Wetterich construction adds a scale-dependent "mass term" $R_k$ that gives the slow modes ( $|p|<k$ ) a large mass and freezes them, while the fast modes ( $|p|>k$ ) stay light and get integrated. Lowering $k$ thaws more modes. The flow term $\partial_t R_k$ is sharply peaked at $|p|\approx k$ , so at each instant the flow integrates out just the shell near $k$ — the smooth analogue of one momentum shell. Because $R_k$ acts as an infrared regulator, the effective average action interpolates between two clean limits: $\Gamma_{k\to\Lambda}$ equals the bare microscopic action $S$ , and $\Gamma_{k\to0}$ equals the full effective action $\Gamma$ , generator of all 1PI vertices. A single trajectory carries you from microphysics to macrophysics. Fixed points appear once you rescale to dimensionless variables and the action stops changing. Around such a point, small deformations grow or shrink at rates that are exactly the scaling dimensions of operators — the relevant directions are the few knobs an experimentalist must tune to sit at criticality, the irrelevant ones flow away and produce universality. The whole content of the earlier momentum-shell renormalization group survives, now as the linearization of a functional flow. The one genuinely new feature is that the flow is exact: no perturbative expansion was made in writing the equation, only in solving it.

Figure

The FRG interpolates: Γ_k slides from the bare action S at the UV to the full effective action Γ at the IR

Figure

Punchline — infinitely many couplings flow in theory space, yet a finite LPA truncation lands on the Wilson-Fisher fixed point

The core equations

The starting point is the definition of the Wilsonian effective action itself — the object whose flow everything else describes. Integrating out the fast modes while holding the slow field fixed produces an action for the slow field alone that reproduces the same low-momentum correlators.

e^{-S_\Lambda[\phi_<]}=\int\mathcal{D}\phi_>\;e^{-S_{\Lambda_0}[\phi_<+\phi_>]} \qquad (1)

where: $S_\Lambda$ — Wilsonian effective action at scale $\Lambda$ ; $\phi_<$ — slow modes, $|p|<\Lambda$ ; $\phi_>$ — fast modes, $\Lambda<|p|<\Lambda_0$ , integrated out; $S_{\Lambda_0}$ — bare action at the UV cutoff $\Lambda_0$ . The functional integral over $\phi_>$ is what packs the fast-mode physics into new, generically infinitely many, couplings of the slow-field action. Demanding cutoff-independence of the partition function and differentiating (1) infinitesimally gives the Polchinski equation — an exact functional flow with the structure of a diffusion equation in field space. Note the two pieces: a bilinear "tree" term and a second-derivative "loop" term.

\Lambda\frac{\partial S_\Lambda}{\partial\Lambda}=\frac12\int_p \Lambda\frac{\partial C(p)}{\partial\Lambda}\left[\frac{\delta S_\Lambda}{\delta\phi_p}\frac{\delta S_\Lambda}{\delta\phi_{-p}}-\frac{\delta^2 S_\Lambda}{\delta\phi_p\,\delta\phi_{-p}}\right] \qquad (2)

where: $C(p)=K(p/\Lambda)/p^2$ — regularized propagator carrying the smooth cutoff $K$ ; $\int_p=\int\mathrm{d}^dp/(2\pi)^d$ ; $\delta/\delta\phi_p$ — functional derivative in the mode of momentum $p$ . The bilinear term rearranges couplings at tree level; the second-derivative term is the genuine fluctuation (loop) contribution. The equation is closed and exact — no perturbative approximation has entered. The most-used form flows the effective average action $\Gamma_k$ rather than the Wilsonian action. Its right-hand side is a single momentum-and-index trace with a one-loop appearance that is nonetheless exact.

\partial_t\Gamma_k[\phi]=\frac12\,\mathrm{Tr}\!\left[\big(\Gamma_k^{(2)}[\phi]+R_k\big)^{-1}\,\partial_t R_k\right],\qquad t=\ln(k/\Lambda) \qquad (3)

where: $\Gamma_k$ — effective average action at IR scale $k$ ; $\Gamma_k^{(2)}=\delta^2\Gamma_k/\delta\phi\,\delta\phi$ — full field-dependent inverse propagator; $R_k(p)$ — infrared regulator suppressing modes with $|p|<k$ ; $\mathrm{Tr}$ — trace over momenta and internal indices; $t$ — RG "time," decreasing as $k\to0$ . The single trace looks like one loop but is exact: because $\Gamma_k^{(2)}$ is the full, field-dependent propagator, iterating the equation generates every loop order. The flow needs boundary conditions, and they are exactly the two physical endpoints of the RG.

\Gamma_{k\to\Lambda}[\phi]=S[\phi],\qquad \Gamma_{k\to0}[\phi]=\Gamma[\phi] \qquad (4)

where: $S$ — microscopic (bare) action at the UV cutoff; $\Gamma$ — full thermodynamic/quantum effective action, generator of all 1PI vertices. The regulator freezes everything at the UV so $\Gamma_k$ starts at $S$ , and vanishes at the IR so $\Gamma_k$ ends at $\Gamma$ . For the regulator to do its job it must satisfy three limits — mass the slow modes, free the fast modes, and freeze everything at the UV.

\lim_{p^2/k^2\to0}R_k(p)>0,\qquad \lim_{k^2/p^2\to0}R_k(p)=0,\qquad \lim_{k\to\Lambda}R_k(p)\to\infty \qquad (5)

where: $R_k(p)$ — cutoff function. A convenient choice is the Litim regulator $R_k(p)=(k^2-p^2)\,\theta(k^2-p^2)$ : at $p=0$ it gives a mass $k^2$ , and at $p=k$ it vanishes — exactly the "freeze slow, free fast" behavior. Projecting the Wetterich equation onto a constant field gives the flow of the effective potential, the workhorse of the local potential approximation (LPA).

\partial_t U_k(\phi)=\frac12\int\frac{\mathrm{d}^d q}{(2\pi)^d}\,\frac{\partial_t R_k(q)}{q^2+R_k(q)+U_k''(\phi)} \qquad (6)

where: $U_k(\phi)$ — effective potential at scale $k$ ; $U_k''=\partial^2U_k/\partial\phi^2$ — field-dependent mass; $q^2+R_k$ — regulated inverse propagator in LPA (wavefunction renormalization set to one). This nonlinear partial differential equation for $U_k(\phi)$ already contains the Wilson-Fisher fixed point.

Derivation AFrom mode elimination to the Polchinski flow

This derivation obtains the exact Wilsonian flow by demanding that integrating out an infinitesimal shell of fast modes leaves the partition function untouched, so that the change forced on the action takes the form of a functional diffusion equation. The strategy is to write the theory with a smooth cutoff inserted into the Gaussian propagator, impose $\Lambda\,\mathrm{d}Z/\mathrm{d}\Lambda=0$ , and recognize that the cutoff dependence sits entirely in the propagator so that differentiating it produces a heat-kernel operator acting on $e^{-S_\Lambda}$ . The two terms of the Polchinski equation then fall out as the tree-level and loop-level pieces of that operator. Keep in mind that no expansion in the coupling is made anywhere — the result is exact.

Step 1. Write $Z=\int\mathcal{D}\phi\,\exp[-\tfrac12\int_p\phi_p C^{-1}(p)\phi_{-p}-S_\Lambda[\phi]]$ , separating the cutoff propagator $C(p)=K(p/\Lambda)/p^2$ from the interaction $S_\Lambda$ . Why: the cutoff lives entirely in the Gaussian weight through $C^{-1}$ , so all $\Lambda$ -dependence can be tracked there.

Step 2. Impose cutoff independence, $\Lambda\,\mathrm{d}Z/\mathrm{d}\Lambda=0$ : the physics below $\Lambda$ must not depend on where the cutoff sits. Why: $\Lambda$ is an arbitrary bookkeeping scale, and demanding $Z$ be invariant is exactly what forces $S_\Lambda$ to compensate the change in $C^{-1}$ .

Step 3. Recognize the functional heat-kernel identity $\Lambda\partial_\Lambda e^{-S_\Lambda}=\tfrac12\int_p \Lambda\partial_\Lambda C(p)\,\dfrac{\delta^2}{\delta\phi_p\delta\phi_{-p}}\,e^{-S_\Lambda}$ . Why: shifting $C^{-1}$ in a Gaussian is equivalent to acting with a second functional derivative — the same structure as diffusion, where lowering $\Lambda$ "diffuses" the action in field space.

Step 4. Expand the second derivative using $\delta^2 e^{-S}/\delta\phi^2=[(\delta_\phi S)^2-\delta_\phi^2 S]\,e^{-S}$ , which converts the identity into the boxed Polchinski equation with a bilinear term and a loop term. Why: differentiating the exponential twice produces exactly the "tree" $(\delta S)^2$ piece and the "loop" $\delta^2 S$ piece — the two contributions of equation (2). *Verified: perturbative expansion of $S_\Lambda$ reproduces the one-loop beta functions; the bilinear-minus-loop structure and its match to Polchinski (1984) checked in verify_aq-sm-wilson-effective-action-exact-rg.py.*

Go deeper · Derivation B · 2 minWilson-Fisher fixed point in the local potential approximation

Derivation BWilson-Fisher fixed point in the local potential approximation

This derivation solves the dimensionless LPA flow near four dimensions to recover the non-trivial Wilson-Fisher fixed point and the leading correction to the correlation-length exponent, demonstrating that the functional equation reproduces the ε-expansion of the companion article on the Wilson-Fisher fixed point. The plan is to truncate the effective potential to a quartic polynomial, extract the beta functions of the dimensionless mass and quartic couplings including the canonical (engineering) terms in $d=4-\varepsilon$ , locate the interacting fixed point where the quartic beta function vanishes at nonzero coupling, and linearize about it to read the relevant eigenvalue that fixes $\nu$ . The key physical input is that the quartic coupling's engineering dimension changes sign at $d=4$ , which is what makes the interacting fixed point appear below four dimensions.

Step 1. Truncate the dimensionless potential to $u_k(\varphi)=\tfrac12 m^2(t)\varphi^2+\tfrac{1}{4!}\lambda(t)\varphi^4$ , keeping the mass and quartic couplings. Why: the leading singular behavior near the Gaussian point is captured by these two couplings; higher powers are irrelevant at one loop.

Step 2. Read the beta functions from the LPA flow: $\partial_t\lambda=-\varepsilon\lambda+c'\lambda^2+\cdots$ , where $-\varepsilon\lambda$ is the canonical (engineering-dimension) term in $d=4-\varepsilon$ and $c'>0$ is a positive regulator-dependent threshold constant carrying the O( $n$ ) combinatorial factor. Why: the engineering dimension $[u]=4-d=\varepsilon$ gives the linear term; the loop integral gives the $\lambda^2$ term whose coefficient counts the one-loop diagrams.

Step 3. Solve $\partial_t\lambda=0$ : besides the Gaussian solution $\lambda=0$ , there is a Wilson-Fisher solution $\lambda_\ast=\varepsilon/c'>0$ , which exists precisely for $\varepsilon>0$ , i.e. $d<4$ . Why: a positive fixed-point coupling requires $\varepsilon>0$ ; above four dimensions only the Gaussian point survives, below it the interacting point appears.

Step 4. Linearize the flow about $\lambda_\ast$ to obtain the relevant eigenvalue, giving $\nu^{-1}=2-\tfrac13\varepsilon+\mathcal O(\varepsilon^2)$ and hence $\nu=\tfrac12+\tfrac{1}{12}\varepsilon+\mathcal O(\varepsilon^2)$ , with $\eta=0$ at this order (LPA has no wavefunction renormalization). Why: the eigenvalue of the linearized flow at the fixed point is the thermal RG eigenvalue $y_t=1/\nu$ , and inverting it delivers the exponent — identical to the one-loop ε-expansion result. *Verified: $\nu=\tfrac12+\varepsilon/12$ and $\nu^{-1}=2-\varepsilon/3$ for $n=1$ , the fixed-point $\lambda_\ast=\varepsilon/c'>0$ , and the O( $n$ ) one-loop factor $n+8$ all confirmed in verify_aq-sm-wilson-effective-action-exact-rg.py (checks 1–6).*

Limiting cases

The exact-RG contains the familiar renormalization-group results as special limits and truncations.

Case	Result	Meaning
Perturbative / one-loop	$\partial_t\Gamma_k\big\vert_{\Gamma_k^{(2)}\to S^{(2)}}$	Replacing the full propagator by the bare one and iterating reproduces ordinary perturbation theory and the standard beta functions order by order.
Local potential approximation	$\Gamma_k=\int\mathrm{d}^dx\,[\tfrac12(\partial\phi)^2+U_k(\phi)]$	The leading derivative expansion: only $U_k$ flows, giving a PDE that already yields the Wilson-Fisher fixed point.
ε-expansion	$d=4-\varepsilon,\;\lambda_\ast\sim\mathcal O(\varepsilon)$	Expanding the LPA fixed-point equation near the upper critical dimension recovers the perturbative Wilson-Fisher exponents to leading order in $\varepsilon$ .
Sharp cutoff	$R_k\to\infty\,\theta(k-\vert p\vert)$	Recovers the original Wilson-Kogut and Wegner-Houghton flow; smooth regulators avoid the resulting non-analyticities.

Common trapThe Wetterich equation looks like one loop but is exact

The most common error is treating $\partial_t\Gamma_k=\tfrac12\,\mathrm{Tr}[(\Gamma_k^{(2)}+R_k)^{-1}\partial_t R_k]$ as a literal one-loop formula. The full field-dependent inverse propagator $\Gamma_k^{(2)}+R_k$ appears, not the bare one, so iterating the equation generates all loop orders; the field dependence closes an infinite hierarchy that a truncation must cut. A second trap: $\Gamma_k$ is not the Legendre transform of the Wilsonian $S_\Lambda$ . They are different objects — one the generator of 1PI vertices, the other the coarse-grained action — that flow to the same physics but obey different equations. Do not identify them term by term.

Worked examples

Example 1 — The Wetterich right-hand side has the right dimension. Verify by dimensional analysis that the trace on the right of equation (3) carries the same dimension as the flow of the effective potential density $\partial_t U_k$ , working in $d$ spatial dimensions with $\hbar=k_B=1$ .

Step 1. Catalogue the dimensions. The inverse propagator has $[\Gamma_k^{(2)}]=[R_k]=\text{mass}^2$ , so the regulated propagator $(\Gamma_k^{(2)}+R_k)^{-1}$ has dimension $\text{mass}^{-2}$ . The flow of the regulator has $[\partial_t R_k]=\text{mass}^2$ . The trace is a momentum integral, $\mathrm{Tr}\to\int\mathrm{d}^dq/(2\pi)^d$ , contributing $\text{mass}^{d}$ .

Step 2. Multiply: the right-hand side scales as $\text{mass}^{d}\times\text{mass}^{-2}\times\text{mass}^{2}=\mathbf{\text{mass}^{d}}$ .

Step 3. Compare to the left. The effective potential is an energy density, $[U_k]=\text{mass}^{d}$ , so $[\partial_t U_k]=\text{mass}^{d}$ as well. The two sides match. The Wetterich RHS has dimension $\text{mass}^{d}$ , equal to $[\partial_t U_k]$ — the equation is dimensionally consistent.

Example 2 — LPA reproduces the one-loop $(n+8)$ and the exponent $\nu$ . Show that the local potential approximation for the O( $n$ ) model gives the same one-loop quartic beta-function coefficient and the same $\nu$ as the ε-expansion of the Wilson-Fisher fixed point.

Step 1. The one-loop renormalization of the O( $n$ )-symmetric quartic vertex sums three channels of the invariant tensor $T_{ijkl}=\delta_{ij}\delta_{kl}+\delta_{ik}\delta_{jl}+\delta_{il}\delta_{jk}$ . Contracting two such vertices over the internal index and projecting back onto $T$ yields the combinatorial factor $n+8$ : for $n=1$ this is $9$ , for $n=2$ it is $10$ , for $n=3$ it is $11$ .

Step 2. With that factor the LPA quartic beta function is $\partial_t\lambda=-\varepsilon\lambda+a(n+8)\lambda^2$ ( $a>0$ ), whose non-trivial zero is $\lambda_\ast=\varepsilon/[a(n+8)]\propto\varepsilon$ . Linearizing gives, for $n=1$ , $\nu=\tfrac12+\tfrac{1}{12}\varepsilon$ .

Step 3. Evaluate at $\varepsilon=1$ ( $d=3$ ): the leading LPA value is $\nu=\tfrac{7}{12}\approx\mathbf{0.583}$ . Higher orders of the derivative expansion climb toward the accepted three-dimensional Ising value $\nu\approx0.63$ with a small $\eta\approx0.04$ . The LPA reproduces the Wilson-Fisher fixed point exactly at one loop ( $\nu=\tfrac12+\varepsilon/12$ , coefficient $n+8$ ); the full derivative expansion then gives $\nu\approx0.63$ , $\eta\approx0.04$ for the 3D Ising class.

Origins · 1 minjoseph polchinski and how this was discovered

joseph polchinski

1954–2018, American theoretical physicist · Best known for D-branes in string theory, Polchinski also wrote the 1984 paper that put Wilson's renormalization group on an exact functional footing and made proofs of renormalizability transparent.

Joseph Polchinski (1954–2018) was an American theoretical physicist at the University of California, Santa Barbara, celebrated above all for his 1995 discovery that D-branes carry Ramond-Ramond charge, a cornerstone of the second superstring revolution, and for his two-volume textbook String Theory. A decade earlier, in a 1984 paper in Nuclear Physics B, he recast Wilson's renormalization group as an exact functional differential equation for the Wilsonian action with a smooth cutoff, turning the proof of perturbative renormalizability into a clean statement about the flow of relevant couplings. His later co-authorship of the black-hole "firewall" paradox reshaped debates on quantum gravity. He died in 2018 after an illness he described with characteristic candor in a memoir.

History

1971

— Wilson introduces the momentum-shell renormalization group and the Wilsonian effective action

1973

— Wegner and Houghton write a sharp-cutoff exact RG equation for the action

1974

— Wilson and Kogut's Physics Reports review systematizes the continuous renormalization group

1984

— Polchinski derives the exact RG equation for the Wilsonian action and proves renormalizability cleanly

1993

— Wetterich formulates the exact flow of the effective average action, the modern functional RG

2002

— Berges, Tetradis and Wetterich review non-perturbative FRG methods and 3D critical exponents

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., Ch. 14 · Kardar, Statistical Physics of Fields, Ch. 5 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, Ch. 12 · Wilson & Kogut, Phys. Rep. 12 (1974) 75 · Polchinski, Nucl. Phys. B231 (1984) 269 · Wetterich, Phys. Lett. B301 (1993) 90 · Berges, Tetradis & Wetterich, Phys. Rep. 363 (2002) 223.

Provenance. Definitions and structure follow Wilson & Kogut (1974), Polchinski (Nucl. Phys. B231, 1984) for the Polchinski equation, and Wetterich (Phys. Lett. B301, 1993) for the effective-average-action flow, with pedagogical treatment in Kardar Ch. 5, Sethna Ch. 12, and Pathria & Beale Ch. 14. Recomputed here from scratch: the bilinear-minus-loop structure of the Polchinski equation; the LPA potential flow and its leading ε-expansion $\nu=\tfrac12+\tfrac1{12}\varepsilon$ ; the O( $n$ ) one-loop combinatorial factor $n+8$ ; the dimensional consistency of the Wetterich right-hand side; and the regulator limits. The 3D Ising values $\nu\approx0.63$ , $\eta\approx0.04$ are quoted from the FRG/derivative-expansion literature and stated as higher-order results beyond the one-loop LPA.

Builds on: Momentum-Shell Renormalization Group · RG Flow in Coupling Space · Fixed Points of the RG · The Renormalization Group Idea: Coarse-Graining and Rescaling. Leads to: Wilson-Fisher Fixed Point · Critical Exponents from the RG · Linearization About a Fixed Point and Eigenvalue Spectrum · Relevant, Irrelevant, and Marginal Operators.

Why do we need an exact RG if the ε-expansion already gives exponents?

The ε-expansion is a perturbative series in $\varepsilon=4-d$ , asymptotic and divergent, that must be extrapolated (and resummed) to reach $d=3$ . The exact RG is non-perturbative: once you truncate it — say to the local potential approximation or a higher derivative expansion — you get results directly in three dimensions without any small parameter. It also handles problems the ε-expansion cannot touch, such as strongly coupled gauge theories, non-renormalizable theories like gravity, and full scaling functions rather than just exponents.

In what sense is the equation "exact" if I still have to truncate it to solve it?

The equation itself is derived without any approximation — it is an identity following from the definition of $\Gamma_k$ and the regulated Legendre transform. The approximation enters only when you solve it, by projecting onto a finite set of operators (derivative or vertex expansion). This is a genuine advantage: the truncation is systematic and controlled, and one can test convergence by adding terms and check regulator dependence to estimate the error, rather than expanding in a coupling that may not be small.

How does the exact RG relate to the discrete momentum-shell RG?

It is the continuous, exact version. Instead of integrating a finite shell $[\Lambda/b,\Lambda]$ and then rescaling, the exact RG integrates infinitesimal shells and tracks the entire action functional as the cutoff slides. The discrete coupling-space flow of the momentum-shell renormalization group is a finite-dimensional projection of the functional exact-RG trajectory: keep only a few couplings and you recover the familiar recursion relations.

What is the regulator R_k, and does the final answer depend on my choice of it?

$R_k(p)$ is an infrared cutoff that gives modes with $|p|<k$ a mass of order $k^2$ and leaves faster modes untouched; $\partial_t R_k$ is peaked at $|p|\approx k$ , so the flow integrates one shell at a time. Exact results are regulator-independent — any admissible $R_k$ gives the same physics — but truncated calculations carry a residual scheme dependence. That dependence is not a bug: comparing exponents across regulators (or optimizing the regulator) is a standard way to estimate the truncation error.

Why is there a factor of one-half and a supertrace in the Wetterich equation?

The one-half is the usual symmetry factor of a closed loop, inherited from the Gaussian integration over the regulator term $\tfrac12\int\phi R_k\phi$ . The "trace" (a supertrace for theories with fermions or several fields) runs over momenta and all internal indices — the loop momentum is integrated and the matrix structure of $(\Gamma_k^{(2)}+R_k)^{-1}$ is traced. For a single scalar it reduces to a momentum integral times a scalar propagator, as in the LPA flow of the potential.

Does the Wilson-Fisher fixed point look different in this functional language?

Its content is the same but its form is richer. After rescaling to dimensionless variables it is a full fixed-point function $u_\ast(\varphi)$ — a scale-invariant effective potential — not just a fixed value of one coupling. Linearizing the flow around it gives an eigenvalue spectrum whose single relevant direction fixes $\nu$ and whose subleading directions give corrections-to-scaling exponents. The one-loop truncation collapses this to the familiar $\nu=\tfrac12+\varepsilon/12$ , but the functional picture makes clear that infinitely many couplings are flowing and being organized by the fixed point.