The Renormalization Group

Every reference cited below sets \(\hbar=c=1\). The dictionary is the one given once in Remark 100.4; three entries are used so heavily here that they are repeated. A scalar field maps to its reduced form Equation (108.2); a coupling of a dimension-\(D\) operator maps to the hatted number of Equation (108.5), which is what those texts write with no hat; and the “mass dimension” of an operator is the integer \(D\) of Equation (108.3), which in this book is a power of metres. To go the other way, restore an \(\hbar c\) for every unit by which the operator's dimension exceeds four. No derivation in this chapter is carried out with \(\hbar\) or \(c\) set to one; this remark exists so that a reader can check the chapter against its sources, which is a different thing.

Theorem 108.7 uses no perturbation theory and no regulator. It uses one physical assumption — that the theory contains no mass with which \(Q\) could be compared — and one piece of dimensional analysis. Everything specific to a theory sits in the one function \(\beta\), and computing \(\beta\) is what perturbation theory is for. Two consequences are worth stating at once, because they are the skeleton of the whole chapter.

First, a zero of \(\beta\) is a scale-invariant theory: by Equation (108.9) the coupling then does not move, and by Equation (108.10) the integral diverges as \(\bar{\alpha}\) approaches the zero, so the zero is reached only as \(\ln Q\to\pm\infty\). Such a point is a fixed point (Section 108.4.3).

Second, if \(\beta\) has no zero between \(\alpha_{\kappa}\) and some value, Equation (108.10) converts a statement about a dimensionless coupling into a statement about an energy. Where the integral converges at the upper limit the theory acquires a finite scale out of nothing: this is dimensional transmutation, and it is why \(\Lambda_{\mathrm{QCD}}\) exists (Definition 102.38) and why the one-loop Landau pole of Proposition 100.64 has a location at all.

Theorem 108.7 and Theorem 108.11 are the same statement addressed to different objects. Equation (108.9) governs a measured effective coupling and contains only \(\beta\); Equation (108.13) governs Green functions, which depend on the arbitrary \(Z_{\phi}\), and pays for that generality with the extra term \(n\gamma_{\phi}\). In a scheme where the renormalized coupling is defined as the effective coupling at \(Q=\bar{\mu}\) the two beta functions coincide, which is how Proposition 100.63 was obtained in Quantum Electrodynamics and Renormalization without writing Equation (108.13) at all.

Minimal subtraction is not more physical than any other scheme; it is more convenient, for three reasons that are worth separating from any claim of naturalness. It is mass-independent, so the beta function is a polynomial in the couplings and multi-loop calculations are tractable. It respects gauge invariance, because dimensional regularization does (Section 108.3.1), so Ward–Takahashi identities hold order by order without extra counterterms. And the modification of [Bardeen:1978] — removing \(-\gamma_{\mathrm{E}}+\ln4\pi\) along with the pole — removes a combination that would otherwise appear at every order in every process, so that residual coefficients are of order one rather than of order ten. That third point is bookkeeping and is the reason the modified scheme, not the minimal one, is the default for every number quoted in Quantum Chromodynamics.

Dimensional regularization was introduced independently by 't Hooft and Veltman [tHooft:1972] and by Bollini and Giambiagi [Bollini:1972], and it is the only regulator in common use that preserves gauge invariance and Lorentz invariance simultaneously. The reason is narrow and worth naming: it never introduces a preferred momentum into the integrand, so the shift \(\ell\mapsto\ell+k\) of a loop momentum remains legitimate, and every Ward–Takahashi identity of Quantum Electrodynamics and Renormalization — each of which is proved by exactly such a shift — survives order by order. A momentum cutoff does introduce a preferred momentum, the shift is then illegal by a surface term, and the surface term is precisely a photon mass counterterm forbidden by gauge invariance. This is not an aesthetic preference. It is what made 't Hooft's proof that the spontaneously broken gauge theory of Electroweak Unification and the Higgs Boson is renormalizable possible at all, and the same point is recorded there in the discussion preceding Remark 104.50.

Two honest qualifications.

The first is technical and specific to chiral theories: there is no definition of \(\gamma^{5}\) away from four dimensions that both anticommutes with every \(\gamma^{\mu}\) and keeps \(\tr\left(\gamma^{5}\gamma^{\mu}\gamma^{\nu}\gamma^{\rho} \gamma^{\sigma}\right)\) non-zero, because the two requirements are contradictory once the number of dimensions is not four. Every consistent prescription therefore breaks one of them and needs compensating finite counterterms; the point is taken up where it bites, in Remark 104.50. That the difficulty is unavoidable rather than a failure of ingenuity is the same fact that makes the Adler–Bell–Jackiw anomaly of Theorem 106.87 real: a symmetry of the classical Lagrangian that no regulator can preserve is not a symmetry of the quantum theory.

The second is a matter of standing. The physical world of this book has three space dimensions and one of time, and every instantiated result below is at \(d=4\). The continuation in \(d\) is a device for defining an integral, exactly as the continuation of a factorial to the gamma function is a device for defining a product; the limit \(\varepsilon\to0\) is taken before any number is quoted. The same caution applies to the \(\varepsilon\) expansion of Section 108.4.4, where the continuation is in the number of space dimensions and the physical case is always the observed three.

It was regarded so for thirty years, and the reversal of that judgement is the single most consequential thing the renormalization group did to particle physics. Corollary 108.27 says that a theory with \(D_{i}>4\) cannot be extended to arbitrarily high energy with a finite number of parameters. It does not say the theory is wrong, and it does not say it is unpredictive: at energies below the scale \(E_{\star}\) set by Equation (108.5), the operator's contribution is suppressed by \((E/E_{\star})^{D-4}\), so a calculation to a stated accuracy needs only finitely many operators. Fermi's theory predicted the muon lifetime to a few per cent for forty years without knowing that the \(W\) existed, and that was not luck. The systematic version of the argument is Section 108.6.

Theorem 108.26 controls the overall divergence of a diagram; a graph whose \(D\) is negative may still contain a divergent subgraph, and at two loops the electron self-energy already contains two vertex subgraphs that overlap (Definition 100.51). That the recursive subtraction terminates, that the forest formula solves it in closed form, and that the counterterms so generated are exactly those of the local Lagrangian, is the theorem of Bogoliubov and Parasiuk [Bogoliubov:1957], Hepp [Hepp:1966] and Zimmermann [Zimmermann:1969], stated as Theorem 100.52. It is what licenses every statement in this chapter about “the” beta function of a renormalizable theory: without it, a beta function would be an order-by-order accident rather than a property.

Theorem 108.32 derives every exponent relation from Equation (108.32), and derives nothing at all about the two numbers \(y_{t}\) and \(y_{h}\). Worse, the assumption that the blocked Hamiltonian has the same form as the original is false as stated: blocking a nearest-neighbour Ising model generates couplings between next-nearest neighbours, four-spin couplings, and infinitely many others. Kadanoff's argument therefore explains why the exponents satisfy relations and why the same relations hold for a fluid and a magnet, but it cannot compute a single exponent and it rests on a premise known to be wrong.

Wilson's contribution was to accept the proliferation of couplings and to work in the infinite-dimensional space of Hamiltonians, where Equation (108.32) is not an assumption but the linearization of a flow about a fixed point, and where \(y_{t}\) and \(y_{h}\) are its eigenvalues, computable. That is Section 108.4.2 and Section 108.4.3. The three-dimensional critical exponents themselves, and the experimental verdict, are Section 108.4.4; the phenomenology they explain is collected in Section 123.5.

Equation (108.38) integrates degrees of freedom out. Information is discarded and cannot be recovered: \(\mathcal{R}_{b}\) has no inverse, and the “group” of the name is in fact a semigroup, \(b\geq1\) only. The name is historical, from [Stueckelberg:1953] where the transformations really did form a group because nothing was being discarded — only relabelled. The distinction is not pedantic: it is the reason the flow has a direction, which is the subject of Section 108.7.2.

The Ginzburg criterion [Ginzburg:1960] compares the fluctuation of the order parameter within a correlation volume to its mean, and yields a computable window of reduced temperature inside which mean-field theory fails for \(d<4\); that estimate belongs to Section 123.4.4. Corollary 108.37 reaches the same conclusion from the SI dimension of one coupling and no estimate at all. The two are the same statement: a coupling with a positive length dimension is one whose dimensionless strength grows as the probe scale grows, and the correlation length near \(T_{c}\) is the probe scale that diverges.

\(l=\ln b\) increases towards longer distances, so \(\dd/\dd l=-\dd/\dd\ln\bar{\mu}\). Translating Equation (108.42) accordingly, and using the dimensionless \(\hat{g}\) of Equation (106.34) in place of \(u\), the field-theoretic beta function of a single real scalar (\(N=1\), \(\varepsilon=0\)) is

positive, so a scalar self-coupling grows with energy exactly as the electromagnetic coupling does. That is the sign that will produce Section 108.5.3. A reader comparing with the sources should note that many of them write \(\dd/\dd l\) and many others \(\dd/\dd\ln\bar{\mu}\), with no warning, so an apparent sign disagreement between two texts is usually this.

Equation (108.38) can be written as an exact differential equation for the cutoff dependence of the effective Hamiltonian rather than as a perturbative expansion. Polchinski's form [Polchinski:1984] follows the flow of the Wilsonian action with a smooth cutoff and is the version usually used to prove things, including renormalizability; Wetterich's form [Wetterich:1993] flows the effective action of Definition 106.41 instead, with an added infrared regulator, and is the version usually used to compute in strongly coupled regimes, since truncating it does not require a small coupling. Both are exact statements about a functional and both require truncation to be evaluated; neither is free of that approximation, and the systematic error of a truncation is the honest weak point of the method.

That the Wilsonian construction computes rather than merely interprets was established by Wilson's numerical renormalization group, which solved the Kondo problem — a magnetic impurity in a metal, whose perturbation series diverges logarithmically at low temperature — by discretizing the conduction band logarithmically and diagonalizing shell by shell [Wilson:1975]. It produced the impurity susceptibility and the Wilson ratio to a few per cent, in a regime where no expansion in the coupling exists at all.

Three facts that had no explanation before Wilson follow at once, and Fisher's review sets them out in the order in which they were understood [Fisher:1974]. That only two exponents are independent is Theorem 108.43 with \(n_{r}=2\). That a fluid and a uniaxial magnet share exponents is Corollary 108.44, the two Hamiltonians lying on the same critical surface because they have the same space dimension and the same order-parameter symmetry and differ only in irrelevant couplings. And that the exponents are not simple rational numbers is because an eigenvalue of a linearized flow on an infinite-dimensional space has no reason to be: the mean-field values \(\tfrac{1}{2}\), \(1\), \(3\) are rational precisely because they come from naive dimensional counting at a Gaussian fixed point, and the true values do not.

Theorem 108.47 is an expansion in a parameter that Nature does not supply. Its purpose is to give a controlled approximation scheme in which the fixed point is close to the Gaussian one and therefore calculable; the physical case is \(\varepsilon=1\), where the expansion parameter is not small and the series is in fact asymptotic rather than convergent. Setting \(\varepsilon=1\) in Equation (108.52) gives, for the Ising class \(N=1\), \(y_{t}=5/3\) and \(\nu=0.600\) against the true \(0.6300\), and for the \(XY\) class \(N=2\), \(y_{t}=8/5\) and \(\nu=0.625\) against the true \(0.6717\): the leading term captures roughly two thirds of the departure from the mean-field \(\tfrac{1}{2}\) and no more. This is honest arithmetic about a one-loop truncation and not a defect of the framework; resummation of the series to high order, or methods that do not use the continuation at all, do far better, as Table 108.1 shows.

Three candidates are on the table and none has been eliminated. The measurement may carry an unmodelled systematic, most plausibly from the finite size of the sample cell, which cuts off the divergence of \(\xi\) at a few centimetres and enters the fit as a correction whose form must be assumed. The theory may be misestimating its own truncation error, the resummations involved being asymptotic. Or the comparison may be incomplete: corrections to scaling enter the measured specific heat with an exponent \(\omega\approx0.79\) and an amplitude that is not universal, and the two analyses do not treat them identically. The honest summary is that a two-standard-deviation disagreement between the sharpest measurement and the sharpest calculation in critical phenomena has stood for over twenty years without being resolved or explained, and that it is recorded here rather than smoothed over.

It claims that the exponents and certain amplitude ratios depend only on the space dimension, the number of order-parameter components and the range of the interaction. It does not claim that \(T_{c}\) is universal — that is Equation (108.43)'s non-universal \(\Lambda^{2}\) term, which depends on every microscopic detail — nor that amplitudes themselves are. And it fails where its premises do: for long-range interactions, for which the gradient term of Equation (108.39) is not the leading one; at a transition governed by a different fixed point, such as the Berezinskii–Kosterlitz–Thouless transition, which belongs to Section 123.6.2, where the singularity is essential rather than a power and the exponent language does not apply at all; and above the upper critical dimension, where hyperscaling Equation (108.35) breaks because the dangerous irrelevant variable \(u\) cannot be set to zero in the free energy.

Among renormalizable theories in the observed \(3+1\) dimensions — that is, among theories whose vertices all have \(D_{i}=4\) in the sense of Corollary 108.27 — a negative one-loop beta function requires a non-abelian gauge field. Scalar self-couplings contribute with the sign of Equation (108.48), Yukawa couplings and abelian gauge couplings with the sign of Equation (108.22), and only the gluon self-interaction contributes the other way, by Lemma 102.31. Adding matter can only reduce \(b_{0}\), never increase it, which is Corollary 108.53. The theorem to this effect was proved by Coleman and Gross in the same year as [Gross:1973] [Politzer:1973]; the treatise's bibliography does not carry that paper, and the statement is recorded here without a citation rather than attached to a source that does not contain it.

Equation (108.55) can be read backwards. The pair \((a(\kappa),\kappa)\) — a pure number and an energy — can be traded for the single energy at which the right-hand side would vanish, which is Equation (102.45):

and \(\Lambda_{\mathrm{QCD}}\) does not depend on which \(E_{0}\) was used, as Definition 102.38 verifies by differentiation. A theory whose Lagrangian contains no dimensionful parameter at all has acquired one — Phenomenon 102.39 — and it is of the observed size, a few hundred \(\mathrm{MeV}\), corresponding to a hadronic length of about \(10^{-15}\,\mathrm{m}\).

What this costs is worth stating plainly, because it is easy to present dimensional transmutation as though a scale had been created from nothing. It has not. The number of free parameters is unchanged: one dimensionless coupling quoted at one energy has been re-expressed as one energy. What is genuinely explained is why that energy is far below any scale appearing in the Lagrangian, since the exponential in Equation (102.45) converts a modest coupling into an enormous ratio, and hence why \(99\,\mathrm{\%}\) of the proton's mass has no counterpart among the quark masses (Remark 102.40).

The one-loop curve of Table 102.2 misses the \(\tau\)-mass point by about ten per cent, and a reader is entitled to ask whether that is a failure of the theory or of the truncation. Table 108.2 answers it without any new input: the next term in a series whose coefficient is fixed by Equation (108.58) and is not adjustable moves the prediction by almost exactly the discrepancy, in the right direction, at the scale where the expansion parameter is largest and nowhere else. The correction at \(m_{Z}c^{2}\) is nil by construction and at \(1\,\mathrm{TeV}\) is one part in a hundred. That pattern — a truncation error that grows as the coupling grows and vanishes where it is small — is what an asymptotic series behaving properly looks like, and it is a stronger statement than agreement at any single point.

The same apparatus applied to Equation (108.22) runs \(\alpha\) the other way: from \(\alpha^{-1}=137.035999177(21)\) at zero momentum transfer [Mohr:2025] to \(\alpha^{-1}\approx128.94\) at \(E=m_{Z}c^{2}\) [Navas:2024], a rise of about six per cent in the coupling, of which roughly half is leptonic and computable from Equation (100.83) and roughly half hadronic and extracted from measured annihilation cross-sections through a dispersion relation. The derivation is Phenomenon 100.62. Two couplings of the same theory therefore run in opposite directions, by the same mechanism and with the same machinery, and both runs are measured. If the renormalization group were bookkeeping, that could not happen. The measurement at zero momentum transfer belongs to Experiment: The Electron Anomalous Magnetic Moment; the scaling violations that measure \(\alpha_{s}\) by a slope rather than a rate belong to Section 102.5.2 and to the experiment chapter Experiment: Deep Inelastic Scattering.

The four objections stated in Remark 100.65 for the electromagnetic case apply verbatim here and are not repeated; the essential one is that Equation (108.60) extrapolates a one-loop formula over forty decades in energy, into a regime where the expansion parameter it is built from is of order one. The location of the pole is not a prediction. Two things are worth adding.

First, the hypercharge pole sits at about \(9\times 10^{41}\,\mathrm{GeV}\), which is some \(7\times10^{22}\) times the Planck energy \(\sqrt{\hbar c^{5}/G}=1.22\times 10^{19}\,\mathrm{GeV}\); the QED pole of Equation (100.85) sits far higher still. In neither case is the question reachable, and any statement about it is a statement about an extrapolation and not about an experiment.

Second, and this is the point that makes the discussion worth having at all: the existence of a scale at which the description fails, rather than its value, is a structural fact. A theory with a positive \(\beta_{0}\) carries an ultraviolet expiry date in exactly the way Fermi's theory carried one through Equation (101.9) — and in that case the date was \(292.8\,\mathrm{GeV}\), the theory did fail there, and what replaced it was found.

The perturbative pole is not the strongest available statement. For the scalar theory of Equation (108.39) there are rigorous non-perturbative results, and they say something sharper.

Above four dimensions, the continuum limit of the lattice \(\phi^{4}\) theory — and of the Ising model, which the cited work treats in the same framework and for the same reason — is a free field theory: the renormalized quartic coupling vanishes when the cutoff is removed [Aizenman:1981]. This is the rigorous counterpart of Corollary 108.37: above the upper critical dimension the quartic coupling is irrelevant, so it flows to zero, and nothing survives the limit but the Gaussian fixed point. Exactly at four dimensions the coupling is marginal and the question is delicate; the answer, obtained only in 2021, is that the scaling limits of the critical four-dimensional Ising and \(\lambda\phi^{4}\) models are Gaussian as well, marginally so [Aizenman:2021]. There is therefore no non-trivial continuum \(\phi^{4}\) theory in the observed number of dimensions.

For abelian gauge theory the corresponding statement is not a theorem but a lattice result: numerical study of lattice quantum electrodynamics with charged matter finds the continuum limit to exist only at zero renormalized coupling [Gockeler:1998], consistent with triviality and not proving it. The strength of that conclusion depends on the lattice action used — compact and non-compact \(\U(1)\) have different phase structures at strong coupling — so what is recorded here is the conclusion the cited work draws and not a general statement about lattice \(\U(1)\).

Corollary 108.62 is an ultraviolet statement about a coupling that grows. Section 104.8.1 reaches the same place from the opposite direction, and it does so at a far lower energy. The quartic coupling's beta function Equation (104.72) is dominated by the top Yukawa and is negative at the top mass, \(\dd\hat{\lambda}/\dd\ln\mu =-0.0206\) by Proposition 104.66; integrating the coupled system sends \(\hat{\lambda}\) through zero at about \(10^{10}\)–\(10^{11}\,\mathrm{GeV}\) [Degrassi:2012] [Buttazzo:2013], where the potential turns over.

The growth that Equation (108.48) describes for a self-coupled scalar therefore never happens in the Standard Model, because the fermion loop wins first. Asking where it would have arrived requires care with the normalization, because Equation (108.48) is written for a single real field with the interaction \(-\left(g/4!\right)\phi^{4}\), while \(\hat{\lambda}=0.126\) is the quartic coupling of a complex doublet in the normalization of Equation (104.27), where the interaction is \(-\lambda\left(\Phi^{\dagger}\Phi\right)^{2}\). The beta function belonging to the second is the leading term of Equation (104.72), \(16\pi^{2}\dd\hat{\lambda}/\dd\ln\mu =24\hat{\lambda}^{2}\), and not the \(3\hat{g}^{2}\) of the first: the two normalizations differ by a factor of eight in the coefficient, and the eight sits in an exponent. Dropping every other term of Equation (104.72) and integrating that one from the top mass,

with \(m_{t}c^{2}=172.57\,\mathrm{GeV}\). Feeding \(\hat{\lambda}\) into the single-real-scalar coefficient instead would give \(417.8\) e-folds and \(10^{184}\,\mathrm{GeV}\), roughly a hundred and sixty decades higher: that is what a normalization mismatch costs in an exponentiated quantity, and it is why the normalization is stated rather than assumed. Either number is far beyond reach and neither is a prediction. The honest composite statement is that the scalar sector's extrapolation fails at \(10^{10}\,\mathrm{GeV}\) for a reason having nothing to do with triviality, and would fail eventually for triviality if it survived; Remark 104.67 states how fragile the first of those numbers is, and none of it is evidence for any particular completion.

Theorem 108.64 presupposes a mass-dependent scheme. In minimal subtraction the beta function contains no mass at all (Theorem 108.18), so a heavy quark keeps contributing its full \(-\tfrac{2}{3}\) to \(b_{0}\) at every scale, including scales far below its own mass, and \(\alpha_{s}\) computed that way at \(E=1\,\mathrm{GeV}\) would carry the top quark. That is not a defect of the scheme but the price of its convenience, and the remedy is a construction rather than a correction:

• Below the threshold one uses a different theory, the effective theory with \(n_{f}-1\) flavours, in which the heavy quark does not appear as a field at all.

• At a scale near the threshold, conventionally \(\bar{\mu}c=m_{Q}c^{2}\), one matches: the two theories are required to give the same value for a set of low-energy observables, which fixes the couplings of the light theory in terms of those of the heavy one. At one loop the matching is continuity, \(\alpha_{s}^{(n_{f}-1)}=\alpha_{s}^{(n_{f})}\); the difference first appears at order \(\alpha_{s}^{2}\).

• Above and below, one runs with the appropriate \(b_{0}\).

The observable consequence is the kink already used in the derivation of Table 108.2: the slope \(\dd(1/\alpha_{s})/\dd\ln E\) changes from \(1.2202\) to \(1.3263\) on crossing the \(b\) threshold downwards, a discontinuity of about \(0.106\) in a quantity that is measured. Decoupling in \(\overline{\mathrm{MS}}\) is therefore not automatic; it is imposed, and the imposing is matching.

Two failures matter physically, and both are visible in the electroweak theory. Neither contradicts Theorem 108.64: each violates one of its hypotheses, and saying which is the useful part.

A mass that is not an independent parameter. The theorem takes \(M\to\infty\) with the couplings of the theory fixed. In the Standard Model a fermion's mass is \(m_{f}c^{2}=\hat{y}_{f}E_{v}/\sqrt{2}\) by Section 104.4.3, so making the top heavy means making \(\hat{y}_{t}\) large, and the limit is not the one the theorem contemplates. The consequence is measured: the correction \(\Delta\rho\) of Equation (104.52) grows as \(m_{t}^{2}\) rather than falling, so the top does not decouple from the electroweak fit — which is exactly why that fit located the top mass before the top was produced. A non-decoupling effect is not a pathology; here it was the measurement.

A relevant operator. Equation (108.6) says that an operator of dimension \(D<4\) has a dimensionless strength growing as \((E_{\star}/E)^{4-D}\) when the probe energy falls. The scalar mass term is such an operator, \(D=2\), and a heavy field of mass \(M\) coupled to \(\Phi\) shifts \(m_{h}^{2}c^{4}\) by an amount of order \((Mc^{2})^{2}\) times a loop factor: the heavy state's effect on the one dimensionful parameter of the light theory grows with its mass instead of decaying. The same statement applies to the vacuum energy, \(D=0\), more strongly still. What follows from this and what does not is set out in Section 104.8.2, and the honest summary is repeated here because it is a renormalization-group statement: it is an argument about the form a more fundamental theory would have to take, it is not an observation, and none of the programmes proposed to address it has observational support of the kind this book requires.

Chiral symmetry breaking is sometimes given as a third example. It is better described as the failure of a different assumption: the light theory below \(\Lambda_{\mathrm{QCD}}\) is not the Standard Model Lagrangian with heavy fields removed but a theory of different degrees of freedom altogether — pions rather than quarks — and matching onto it is Section 108.6.2, not Theorem 108.64.

An effective theory has two expansion parameters, not one. The first is \(E/E_{\star}\), the operator expansion Equation (108.64). The second is the loop expansion, and for a theory whose interactions are generated by a single scale the natural loop factor is \(\left(E/4\pi E_{\star}\right)^{2}\) — the \(4\pi\) being the \(1/(4\pi)^{2}\) that Lemma 100.28 attaches to every four-dimensional loop. The two expansions are therefore not independent: one loop costs as much as two extra dimensions of operator, up to \(16\pi^{2}\), and a consistent calculation at a given order in \(E/E_{\star}\) must include both. That correspondence is what makes a power counting formula — an integer attached to each diagram counting its total suppression — possible at all.

Wilson's operator product expansion [Wilson:1969] asserts that the product of two local operators at short separation may be written

with the whole singularity as \(x\to0\) carried by the c-number coefficients \(C^{i}_{ab}\) and none by the operators \(O_{i}\). The degree of the singularity is fixed by dimensions: by Equation (108.3), \(C^{i}_{ab}\sim\abs{x}^{D_{i}-D_{a}-D_{b}}\) up to logarithms, so operators of higher dimension are suppressed by more powers of the separation. This is Equation (108.64) with the two scales exchanged: an effective field theory expands in the ratio of a low energy to a high one, and Equation (108.67) expands in the ratio of a short distance to a long one, but the organizing integer is the same \(D\) and the factorization of long from short is the same factorization. In an asymptotically free theory the coefficients are computable in perturbation theory while the matrix elements of the \(O_{i}\) are not, which is exactly the division of labour of Theorem 102.53, and the logarithms the coefficients carry are governed by the anomalous dimensions Equation (102.62) of the operators — the same \(\gamma\) that Definition 108.12 introduced.

Equation (108.70) says: if the neutrino masses come from the unique dimension-five operator, and if its coefficient is of order one, then the scale at which the Standard Model is superseded is about \(10^{15}\,\mathrm{GeV}\). Both conditionals matter. The first is a parametrization of ignorance, not a mechanism: Equation (108.68) records that lepton number is violated somewhere above the electroweak scale and says nothing about what does the violating. The second is unconstrained, and \(E_{\Lambda}\) scales linearly with \(\hat{c}_{5}\): a coefficient of \(10^{-10}\) would place the scale at \(10^{5}\,\mathrm{GeV}\), and the data cannot distinguish. Nor is the operator the only possibility — the neutrinos may be Dirac particles with Yukawa couplings smaller than the electron's by seven orders of magnitude, the ratio \(0.0495\,\mathrm{eV}/0.511\,\mathrm{MeV}=9.7\times 10^{-8}\), which is equally unexplained and equally consistent. The alternatives, and the honest statement that no direct evidence selects among them, belong to Flavour Physics and Neutrinos and to What We Observe but Do Not Understand.

What is not conditional is the structural point, and it is considerable: the smallness of the neutrino masses is, within this framework, not a puzzle but a measurement. An irrelevant operator is suppressed by the scale of the physics that generated it, so a very small mass is the signature of a very large scale. This is the only place in the Standard Model where an observed number points above the electroweak scale, and the pointing is the content.

At \(D=6\) the list is no longer short. The first systematic construction [Buchmuller:1986], corrected and completed into what is now the standard basis [Grzadkowski:2010], contains \(59\) independent baryon-number-conserving operators for a single fermion generation, up to hermitian conjugation; flavour indices multiply that by a large factor for three generations. Their coefficients are constrained by global fits to electroweak precision data, to Higgs couplings and to high-energy tails of distributions at the LHC, and the constraints are usually quoted as lower bounds on \(E_{\Lambda}/\sqrt{\hat{c}}\) of a few \(\mathrm{TeV}\): at \(E=1\,\mathrm{TeV}\) the suppression Equation (108.64) is \((E/E_{\Lambda})^{2}\), which is \(0.11\) for \(E_{\Lambda}=3\,\mathrm{TeV}\) and \(0.01\) for \(10\,\mathrm{TeV}\), so a measurement at the per-cent level probes the ten-\(\mathrm{TeV}\) range and no further.

A small subset of the dimension-six operators violates baryon number. They take the form of four-quark-and-lepton operators such as \((\hat{Q}\hat{Q})(\hat{Q}\hat{L})\), they conserve \(B-L\), and they mediate proton decay.

Three statements, in decreasing order of security.

Secure. Baryon number is conserved to an extraordinary accuracy, and any physics that violates it does so through operators whose coefficient-to-scale ratio satisfies Equation (108.73). This is a bound on \(\hat{c}_{6}/E_{\Lambda}^{2}\) and on nothing else. A theory with \(\hat{c}_{6}=10^{-12}\) and \(E_{\Lambda}=3\times 10^{13}\,\mathrm{GeV}\) is equally consistent with Equation (108.72), and there is no measurement that separates them.

Weaker. The near-coincidence of the neutrino scale Equation (108.70), at \(10^{15}\,\mathrm{GeV}\), with the proton bound Equation (108.73), at \(3\times 10^{16}\,\mathrm{GeV}\), is suggestive of a common origin. It is one order of magnitude apart in quantities whose coefficients are unknown by many orders of magnitude, so it is suggestive and not more.

Not established at all. The programme that historically motivated the proton-decay searches — grand unification, in which the three gauge couplings are supposed to meet at a single scale — has no observational support. Its central prediction is proton decay, and the prediction has been tested for forty years with a null result that has excluded the simplest versions. By editorial rule 1 it is named here to record that fact and for no other reason. The renormalization-group statement that survives is the one that does not depend on it: an unobserved decay is a bound on a coefficient over a squared scale, and Equation (108.73) is that bound.

Corollary 108.81 and Phenomenon 102.39 are one statement in two languages. In the language of the running coupling, a massless Lagrangian generates the scale \(\Lambda_{\mathrm{QCD}}\) of Equation (102.45) by dimensional transmutation, and every hadron mass is a pure number times \(\Lambda_{\mathrm{QCD}}/c^{2}\). In the language of the trace anomaly, the same fact says that the classical scale invariance of the massless theory is destroyed by quantization with a coefficient proportional to \(\beta\), and the proton's mass is what the destruction produces. Neither is more fundamental; the second is the operator statement of the first, and both are consequences of the single fact that \(\beta\neq0\). Quantum Chromodynamics notes the identification in passing at Remark 102.40; this is where it is made precise.

At a zero of \(\beta\) in a massless theory the trace vanishes and the theory is scale invariant. In every case relevant here it is more: Lorentz invariance, locality and scale invariance together promote the symmetry to the full conformal group, which in \(3+1\) dimensions has fifteen parameters against the eleven of Poincaré plus dilatations, the four extra being the special conformal transformations. The promotion is a theorem under technical hypotheses and not an identity — there are known counterexamples in other dimensions — so it is recorded here as what it is.

The consequences are not decorative. Conformal invariance constrains two- and three-point functions completely, up to constants, and it is those constraints, combined with unitarity and the associativity of the operator product expansion Equation (108.67), that the conformal bootstrap turns into numerical bounds. That is how the three-dimensional Ising exponents in the last column of Table 108.1 were obtained [ElShowk:2014]: not by computing a beta function at all, but by asking which sets of dimensions are consistent, and finding that consistency alone determines them to six figures. The critical point of a fluid or a uniaxial magnet is a conformal field theory, and the exponents measured in Section 123.5 are its operator dimensions.

Coarse graining loses information — Remark 108.35 — and in every case that has been examined the loss shows up as a reduction in the number of degrees of freedom describing the theory. A ferromagnet above its Curie point is described by the full lattice of spins; at long wavelength it is described by one order parameter. Quantum chromodynamics at short distance has quarks and gluons, six flavours and eight colours; below \(\Lambda_{\mathrm{QCD}}\) it has pions, and below the pion mass it has nothing at all in the strong sector. A heavy field integrated out by Theorem 108.64 is gone. The direction is always the same, and never the reverse.

If the flow is a gradient flow of a monotone function, then the space of possible ultraviolet completions of a given low-energy theory is constrained: a candidate completion must have a larger \(a\) than the theory it completes, which is a necessary condition that can be checked. That is the reason the theorem is of interest here rather than only in mathematical physics.

Three qualifications must be attached, and the third is the one usually left out.

First, Theorem 108.85 is a statement about flows between conformal fixed points. The Standard Model is not such a flow. Its infrared is not a conformal theory: it contains a massless photon and nearly massless neutrinos alongside a gapped strong sector, and its ultraviolet is not a fixed point at all — by Corollary 108.62 the scalar and hypercharge couplings grow and the description fails. The theorem does not apply to the theory this book is about.

Second, the quantity \(a\) is not measurable. It is a coefficient in the response of a theory to a background curvature, and no experiment has measured one.

Third, and consequently, no classification of the possible ultraviolet completions of the observed \(3+1\) Standard Model exists, and none is in prospect. What exists is a theorem about a class of theories to which the Standard Model does not belong, together with the empirical observation of Remark 108.84 that in every examined case the flow has run one way. That is worth recording and it is not worth more than that. It is recorded here rather than in What We Observe but Do Not Understand, which lists the measured discrepancies and the one missing existence proof and does not carry the absence of a classification of ultraviolet completions as an entry — correctly, since nothing is measured that such a classification would explain.

The measured content is small, sharp and repeatedly confirmed. Two dimensionless couplings of the Standard Model have been watched to change with the energy at which they are measured, in opposite directions and by amounts predicted with no free coefficient (Phenomena 100.62 and 108.56). A critical exponent has been measured to four decimal places against a calculation of comparable precision, and the two do not quite agree (Phenomenon 108.49). The universality of critical behaviour across chemically unrelated systems is explained, not merely described (Corollary 108.44). And an observed number — the neutrino mass — points to a scale fifteen orders of magnitude above the electroweak one, conditionally but for a stated reason (Phenomenon 108.73).

The unmeasured content is larger and should not be confused with the above. The location of any Landau pole is an extrapolation over dozens of decades. The triviality of the scalar sector is a theorem about \(\phi^{4}\) in isolation, not about the Standard Model. The cutoff of the Standard Model is unmeasured, and every statement about what lies above it is a statement about a bound on a coefficient divided by a power of a scale. The renormalization group tells one how to be predictive without knowing what lies above; it does not tell one what lies above, and the two must not be traded for one another.