2026-10-01 · chapter Yang–Mills and the mass gapPart 6 Constructive field theorytriviality

Triviality of φ⁴ in four dimensions — the limit made on the lattice collapses to a Gaussian

A Euclidean field theory is a probability measure, and its content appears in the correlation functions $S_n(x_1,\dots,x_n)=\langle\Phi(x_1)\cdots\Phi(x_n)\rangle$ (6-01). For a Gaussian measure, the correlations at an even number of points are determined just by dividing the two-point functions into pairs (1-08).

The sum is over all the ways $\pi$ of grouping the $2n$ points in twos. This Wick rule means there is no interaction, and for that reason Gaussian fields have been called trivial (ADC §1.1). The first quantity measuring the deviation is the four-point Ursell function. Known

In models with the Lee–Yang property (the zeros of the partition function in the magnetic field lie on the imaginary axis; the Lee–Yang theorem of 2-08), Gaussianity follows if $U_4$ vanishes in the limit (Newman 1975, ADC §6.3). Known Example: for a single spin taking $\pm1$ with equal probability, $\langle\sigma^4\rangle-3\langle\sigma^2\rangle^2=-2$, and for a sum of $N$ independent ones the ratio $U_4/\mathrm{Var}^2$ collapses like $-2/N$ (1-08). Spins near the critical point are strongly correlated, so this reason cannot be used, and three-dimensional $\phi^4$ gives a non-Gaussian measure (JW §6.2, 6-03).

The object is the model with a real number $\phi_x$ at each point of the lattice $\mathbb Z^d$, with neighbours linked by a ferromagnetic coupling $J\ge0$ (ADC formula (1.12); the lattice version of formula (1) of 6-03).

As $b=2\lambda\to\infty$, the distribution at a point concentrates on $\pm1$ and it becomes the Ising model (ADC formula (1.20)). Taking the scale at which we look to be $L$ in lattice units (the lattice spacing in continuum units is $a=1/L$), we follow the smeared sum divided by the standard deviation $\sqrt{\Sigma_L}$ of the sum over the box $\Lambda_L$ ($\Lambda_L=[-L,L]^d$, ADC formulas (1.14)(1.15)).

The volume is first taken to infinity and then $L\to\infty$, and $(J,\lambda,b)$ may be moved along the way. To get smooth correlations at scale $L$, one approaches the critical point to make the correlation length at least of order $L$. If the correlation length stays finite in lattice units, the limit is white noise (ADC footnote 1, 2-08). For the Ising model ($b=2\lambda\to\infty$), $J$ is written $\beta J$, $\beta$ is called the inverse temperature, and $\beta_c$ the critical value at which the correlation length diverges (2-08). $S_2\to0$ at large distance is the side $\beta\le\beta_c$ (ADC formula (1.10)), and this chapter is restricted to that side. Known

A generalised Gaussian process is a distribution-valued random field for which every finite collection of $T_f(\Phi)$, one for each test function $f$ (the limit as $L\to\infty$ of the $T_{f,L}$ of formula (4)), follows a multivariate Gaussian distribution (Gaussian measures in 1-09).

Triviality in five or more dimensions was shown by two inequalities (Aizenman 1981, 1982, Fröhlich 1982): the tree-diagram bound, which bounds the four-point deviation by "the diagram in which four legs meet at one point $u$", and the infrared bound, by which the two-point function does not exceed the free-field form even at the critical point (Ising model, ADC formulas (1.22)(1.23)).

If the four points are at mutual distances of order $L$, then $S_4\sim S_2^2\sim L^{-2(d-2)}$; on the right-hand side the sum over $u$ gives $L^d$ and the four factors $L^{-4(d-2)}$.

For $d\gt4$ this goes to 0, and apart from details this is the backbone of the theorem in five or more dimensions (ADC §1.3). For $d=4$ it is $L^0$, and the primary source writes 'clearly inconclusive for d = 4'. The problem statement also writes that in five or more dimensions scalar fields with a quartic self-interaction do not exist, and that the proof applies to one-component fields (JW §6.2). Known The same exponent also appears in the dimension of the coupling $[\lambda]=4-d$ in formula (7) of 6-01; for $d\gt4$ it points to the coupling ceasing to act at long distances. Physics

The origin of the tree-diagram bound is the random current representation. Taylor-expanding the weight $e^{\beta J\sigma_x\sigma_y}$ edge by edge, viewing the degree $n(x,y)$ as an integer on the edge (a current), and summing over spins, only the configurations survive in which the set $\partial n$ of points with odd total degree (sources) coincides with the set $A$ of points whose correlation is taken (ADC formula (3.5)).

Configurations without sources are collections of loops, and the two-point correlation is "the effect of inserting one path joining $x$ and $y$". An identity that exchanges sources (the switching lemma) gives the following (ADC formula (3.11)).

$\mathbf P^{xy,zt}$ is the law of two independent currents $n_1,n_2$ with sources $\{x,y\}$ and $\{z,t\}$ respectively (the weights of formula (7) turned into probabilities). $\mathcal C(x)$ is the set of points connected to $x$ by following edges where the sum $n_1+n_2$ is positive, and likewise $\mathcal C(z)$. $\mathcal C(x)\cap\mathcal C(z)\ne\varnothing$ means that $x$ and $z$ are connected in $n_1+n_2$. The deviation from Wick's rule is exactly the probability of this meeting. Known

The two clusters are not independent. So one adds currents without sources and replaces them by two independent clusters. Doing so can only increase the probability of intersecting (ADC formulas (3.13)(3.14)). Bounding that probability by the expected number of intersection points and using the switching lemma once more gives the first inequality of formula (5) (the tree-diagram bound) (ADC §3.2). Known

Intuition comes from picturing the two clusters as random walks joining $x$ and $y$, and $z$ and $t$ (the primary source also uses this analogy as a guide). In four dimensions, consider two trajectories starting from two points a distance $L$ apart and walking until they leave a region of radius about $L$ (about $L^2$ steps). The probability of intersecting goes to 0 like $1/\log L$, yet the expected number of intersection points does not become smaller than a positive constant independent of $L$. Intersections are rare, but when they intersect the number of intersection points grows like $\log L$ (ADC §3.2). Known

The rough guide is that the trajectory of a random walk is two-dimensional (about $L^2$ points within radius $L$), and the borderline for two two-dimensional objects to meet is $2+2=4$. More precisely, the content of this guide is that the expected number of intersection points behaves like the $L^{4-d}$ of formula (6) (the $E$ of Table 2).

ComputedThe points P and Table 2 were computed on this machine. The pictures of random walks are samples from random numbers with fixed seeds (the first picture was chosen to be an intersecting example). Most of the places that look overlapping in the projection do not coincide in $d$ dimensions.

ComputedTable 2 Two simple random walks started from $\pm(L/2)e_1$ and run until they leave the ball of radius $2L$: the mean number of common lattice points $E$, the intersection probability $P$, and the mean given intersection $E/P$ (6000 pairs each; the standard error of $P$ is 0.003–0.0044; $L=8,32$ are in Figure 1).

$E$ is proportional to $L$ in three dimensions, decreases like $1/L$ in five, and is nearly constant in four, just as the $L^{4-d}$ of formula (6). In four dimensions $P$ decreases slowly and $E/P$ increases: the shape "rare, but when they meet they meet many times at once". With finite $L$ this does not confirm the rate $1/\log L$. Computed

The key to improving the tree-diagram bound in four dimensions is the bubble diagram $B_L$ cut off at distance $L$ (ADC formula (1.25)): the diagram that joins two two-point functions and goes once round. For the free-field two-point function $G$ (the inverse of the lattice Laplacian) it is proportional to the expected number of times two random walks from the same point visit the same point, and since $G(x)\approx1/(4\pi^2|x|^2)$ far away in four dimensions, it grows logarithmically. Known

ComputedTable 1 The lattice bubble of the free field $\sum_{x\in[-L,L]^d}G(x)^2$ (infinite lattice, mass 0, $G(x)=\int_0^\infty\prod_ie^{-2t}I_{x_i}(2t)\,\mathrm dt$ integrated numerically. $I_n$ is the modified Bessel function, and $e^{-2t}I_n(2t)$ is the probability that a one-dimensional continuous-time random walk jumping left and right each at rate 1 is at $n$ at time $t$). The increment is the difference when $L$ doubles; formula (9) gives $\ln2/(8\pi^2)=0.00878$.

The four-dimensional increment approaches 0.00878; in three dimensions the increment doubles each time ($\propto L$), and in five it halves and converges. Computed Theorem 1.3 (the improved tree-diagram bound) gives, for the four-dimensional nearest-neighbour ferromagnetic Ising model with $\beta\le\beta_c$, $L\le\xi(\beta)$, and four points mutually farther apart than $L$, with some $c,C\gt0$, the following (ADC formula (1.24)).

If the two-point function nearly satisfies the infrared bound (the case expected in four dimensions), $B_L$ is of order $\log L$ and $|U_4|/S_4=O((\log L)^{-c})$; if it is much smaller, formula (5) alone suffices. Beyond four points, it closes by the characteristic function of the quantity in formula (4) being close to Gaussian with error $(\log L)^{-c}$ (ADC Proposition 1.4). The centre of the proof is a multi-scale analysis showing that, given intersection, the number of intersection points grows at least like $(\log L)^c$ (ADC §4, §6). Known

Theorem 1.2 does not depend on the choice of counterterms. The result rests not on computing counterterms but on dimension-dependent relations among the Schwinger functions (the $S_n$ of §01) that can appear in any limit (ADC §1.1). It includes "hard" $\phi^4$ and the Ising model (the Griffiths–Simon construction, which makes $\phi^4$ from averages of blocks of spins). On the other hand, the object is restricted to limits of nearest-neighbour, ferromagnetic, one-component models with a lattice cutoff, and it is not a theorem about every method of construction. Nor does it give the exact size of the logarithmic correction or the covariance of the limit (ADC §1.2). Known

KnownTable 3 The main results on triviality (classification from ADC §1.1–1.2. §1.1 also counts some of the small-$\lambda$ results below among the partial results for $d=4$. Journals at the end).

Since then, a one-page correction (2024) and an extension to effective dimension four and above (Panis 2026) have appeared (bibliographic data only).

The reason on the physics side is that the quartic coupling in four dimensions is not asymptotically free (JW §6.2). What follows is a one-loop picture, not a theorem. Physics If you have not read 4-04, you may accept just the right-hand end of formula (11) and read on. Write the interaction as $\frac{u}{4!}\phi^4$ (for formula (3), $u=24\lambda$). At one loop the four-point coupling shifts by one inserted bubble; there are three ways of inserting it (the three pairs of formula (2)), each with symmetry factor $\tfrac12$, and the bubble is the same $\frac{1}{8\pi^2}\ln$ as in formula (9) (the sum over positions $\sum G^2$ and the integral over momenta are equal by Parseval's identity; 1-11). With $\Lambda$ the cutoff and $\mu$ the scale we look at,

(the middle is the integral of the flow equation $\mu\,du/d\mu=3u^2/16\pi^2$). However large the bare coupling $u(\Lambda)$ is made, raising the cutoff sends $u(\mu)$ to 0. Conversely, keeping $u(\mu)=1$ and extending upwards, the denominator becomes 0 at $\ln(\Lambda/\mu)=16\pi^2/3\approx52.6$ ($\Lambda/\mu\approx10^{22.9}$) (the Landau pole). For Yang–Mills the sign is reversed (formula (2) of 6-07), and fixing $g^2(\mu)$ and raising $\Lambda$ only sends the bare coupling $g^2(\Lambda)$ to 0. PhysicsComputed

We look at formula (11) in numbers. The two left columns of Table 4 are $\phi^4$ and the right end is Yang–Mills; Figure 2 draws the same quantities as functions of the scale.

ComputedThe numbers in this figure were computed on this machine (formula (11) and formula (2) of 6-07; they agree with Table 4).

ComputedPhysicsTable 4 $u(\mu)$ of formula (11), and for $SU(3)$ the bare coupling $g^2(\Lambda)$ keeping $g^2(\mu)=1$ ($b_0=11\cdot3/48\pi^2$).

The rate of collapse is $1/\ln\Lambda$, the same type of slowness as the $(\log L)^{-c}$ of formula (10). The one-loop formula cannot be trusted where the coupling is large, and is not a proof of triviality. In 2000 the problem statement wrote 'all insights from random walks, perturbation theory, and renormalization analysis point toward triviality' (JW §6.2). The answer as a theorem is §02–§07. Known

Detour: the side that becomes free in the infrared

The problem statement asks that 'a non-trivial quantum Yang–Mills theory exists on R⁴', and requires that the correlation functions agree at short distances with the predictions of asymptotic freedom and perturbative renormalisation (JW §4, 0-03). Known In the prediction of perturbation theory, at short distances the coupling goes to 0 logarithmically, like $g^2(\ell)\approx1/(2b_0\ln(\ell_0/\ell))$, but is not identically 0 (4-06; an example is the right end of Table 4). Physics The meaning of the requirement can be organised into three points (this chapter's reading).

1. Non-triviality is a property to be shown separately. It does not follow from ultraviolet estimates or from the existence of a limit (level 5 of 6-07). Four-dimensional $\phi^4$ is an actual example that, within the family of nearest-neighbour, ferromagnetic, one-component lattice models, becomes Gaussian however the parameters are moved (if $S_2\to0$ at large distance).
2. The theorem for $\phi^4$ does not show that Yang–Mills is trivial. The problem statement contrasts the two by whether they are asymptotically free, and calls Yang–Mills the 'natural candidate' (JW §6.2). The Gaussian limit of the $SU(2)$ Yang–Mills–Higgs model comes from a way of taking the limit that sends the gauge coupling to 0 in step with the lattice spacing, and non-Gaussian limits are said to be open (Chatterjee 2026).
3. The tools of the proof cannot be carried over as they are. The random current representation is a structure of Ising-type models.

The present state is as in the problem statement's 'at present we do not know any non-trivial relativistic field theory that satisfies the Wightman (or any other reasonable) axioms in four dimensions' (JW §6), and the survey's 'there is yet no construction of a continuum limit of a lattice gauge theory in any dimension higher than two where Wilson loop variables have been shown to have nontrivial behavior' (Chatterjee 2018 §6). Non-triviality of four-dimensional Yang–Mills is unproved, and the whole of where it stops is seen in 6-10. Known

What this chapter can say

Sources and reproduction

Next chapter to read: 6-09 Stochastic quantisation and regularity structures (subcritical and critical: the face, on the side of equations, of four dimensions being the borderline).

← 6-07 Four dimensionsContents6-09 Stochastic quantisation and regularity structures →