Home·The theorem·The proof in one reading·Architecture·Units·Manuscript·Glossary·FAQ·Work items·Data·Method·Provenance·Changes

OpenAI's construction of four-dimensional quantum Yang–Mills theory: an overview of the proof

Contents

The theorem 2

The strategy 5

Part I. The lattice theory 7

Part II. Fields, insertions and polymer activities 8

Part III. The ultraviolet construction 9

The block-spin step and the background field 9

Gaussian integration and the cluster expansion 11

Large fields and one nonlinear step 12

The marginal coupling and the schedule 13

Renormalized observables, uniformly in the volume 15

Convergence at fixed volume and the construction run 16

Part IV. Infinite volume 17

Part V. Phases and relativistic reconstruction 19

Part VI. Ward identities, rotations and the stress tensor 21

Part VII. Identification as Yang–Mills theory 23

Free Maxwell fields at short distance 23

Running coupling 24

The trace law and nonfreeness 24

Operator product expansion and contact extensions 25

Part VIII. Assembly 26

Relation to earlier work, and where the difficulty lies 27

Steps whose informal account is incomplete 28

Reading guide 29

Level 2 of the five levels: the proof in one reading. Written by an AI model from the unit outlines, then checked against the outlines and the Lean and edited by a second model. It will be revised as the unit outlines are completed.

This is a substantial overview, about 10,000 words. The contents links locate individual topics; the exact statement fixes the scope of every claim.

Contents

The theorem

Let \(G \subset U (m)\) be a closed subgroup whose Lie algebra \(\mathfrak{g}\) is simple and nonabelian, for instance \(SU (N)\) with \(N \geqslant 2\); \(G\) need not be connected. The main theorem is YangMills.jointExistence; its precise statement is in statement.md. It asserts that there is one family of Wilson lattice gauge theories for \(G\) on four-dimensional tori, consisting of:

The family has the following property. For every strictly increasing sequence \(p (k)\) of volume indices there are a further subsequence \(p (q (k))\) and a continuum hierarchy satisfying (1)–(5) below.

(1) Lattice origin. Let \(\text{Conf}_n \subset (\mathbb{R}^4)^n\) be the open set of configurations of distinct points. For every \(n\), every list of tokens \(v\) and every \(f \in C_c^{\infty} (\text{Conf}_n)\),

\(\displaystyle \lim_{k \to \infty} \hspace{0.17em} \lim_{N \to \infty} \hspace{0.17em} a_N^{4 n} \sum_y f (\hat{y}) \hspace{0.17em} \left\langle \prod_{i = 1}^n I_{N, p (q (k))} (v_i, y_i) \right\rangle_{N, p (q (k))} = \int f (x) \hspace{0.17em} S^{\text{mix}} (\pi v ; x) \hspace{0.17em} dx.\)

Here the sum runs over \(n\)-tuples of sites, \(\hat{y}\) are their physical positions, and \(\langle \cdot \rangle_{N, p}\) is the Wilson measure.

(2) Phases. On \(\text{Conf}_n\), \(S^{\text{mix}} = \int_X S^{\xi} \hspace{0.17em} d \nu (\xi)\). Here \(X\) is a compact metric space, \(\nu\) is a probability measure of full support, and \(\xi \mapsto S^{\xi} (a ; x)\) is continuous. All correlations \(S^s\), \(s \in X \sqcup \{ \text{mix} \}\), are smooth and permutation symmetric on \(\text{Conf}_n\).

(3) Reconstruction. A selected phase \(\xi_{\ast}\) carries a Wightman theory:

Its Wightman functions are boundary values, on product test functions, of the continuation of \(S^{\xi_{\ast}}\) that is holomorphic in the ordered complex times.

(4) Operator product expansion. One system of coefficients, with finitely many local fields of each degree, satisfies finite-order OPE bounds in the mixed state and in every phase.

(5) Nonfreeness. Let \(e_4\) be the time direction. A field \(\Theta\) of degree 4 with vanishing vacuum expectation satisfies

\(\displaystyle \lim_{r \downarrow 0} \hspace{0.17em} r^8 \log^2 (1 / r) \hspace{0.17em} G_{\Theta} (r) = \frac{3 \dim_{\mathbb{R}} \mathfrak{g}}{\pi^4}, \qquad G_{\Theta} (r) = S^{\xi_{\ast}} \left( \Theta (- \tfrac{r}{2} e_4) \hspace{0.17em} \Theta (\tfrac{r}{2} e_4) \right) .\)

Moreover, \(G_{\Theta}\) agrees on no interval \((0, \delta)\) with a finite Wick profile, that is, a finite sum of products of free mass-shell kernels \(\int E^q h (\mathbf{p}) (2 E)^{- 1} e^{- rE} \hspace{0.17em} d^3 \mathbf{p}\).

The lattice data are fixed before the volume sequence. The refinement and all continuum objects may depend on that sequence.

The stronger statement. The theorem is a projection of a longer statement, YangMills.constructed. That statement holds for every value \(b > 0\) of the penalty parameter of the block-spin kernel; the public theorem takes \(b = 1\). The construction supplies the following facts; the slope value is established in the proof of regularity, rather than stated as a separate conjunct of YangMills.constructed:

This overview describes the proof of the stronger statement.

What is not claimed.

Normalizations. On \(\mathfrak{g}\), put \(\langle X, Y \rangle_{\mathfrak{g}} = - \tfrac{1}{m} \text{Re} \hspace{0.17em} \text{tr} (XY)\), and define \(C_{\mathfrak{g}} > 0\) by \(\text{tr} (\text{ad} X \hspace{0.17em} \text{ad} Y) = - C_{\mathfrak{g}} \langle X, Y \rangle_{\mathfrak{g}}\). Two facts fix the coupling convention:

Hence \(\beta = 1 / g_0^2\) for the continuum action \(\frac{1}{2 g_0^2} \int \sum_{\mu < \nu} |F_{\mu \nu} |^2_{\mathfrak{g}}\). The constructed couplings satisfy, uniformly in \(p\),

\(\displaystyle \frac{\beta_{N, p}}{\log (1 / a_N)} \longrightarrow \kappa_G = \frac{11 C_{\mathfrak{g}}}{24 \pi^2} = 2 b_0, \qquad b_0 = \frac{11 C_{\mathfrak{g}}}{48 \pi^2} .\)

For \(SU (N)\) in the defining representation, \(C_{\mathfrak{g}} = 2 N^2\) and \(\kappa_G = 11 N^2 / (12 \pi^2)\). In that case \(\beta\) is the usual Wilson coupling \(2 N / g^2\), and the law above is the one-loop running \(1 / g^2 \simeq (11 N / 24 \pi^2) \log (1 / a)\) of Gross–Wilczek and Politzer.

The strategy

Order of limits. One lattice family is fixed first. Correlations of renormalized insertions are taken to the limit \(a_N \to 0\) at fixed volume, along the whole sequence of spacings. The volume then tends to infinity along a subsequence of the prescribed one. The renormalization group (RG) is used only for the first limit. Its levels run from the Wilson lattice (level \(N\)) to the torus of side \(L_p\) at unit spacing (level \(0\)). The bare coupling is tuned so that the effective coupling at level \(0\) is one fixed number \(\bar{\beta}\), the same for all \(N\) and \(p\). No RG step is taken above the unit scale; infinite volume is reached by compactness, from local bounds that are uniform in the volume.

Ultraviolet construction (Parts I–III). At each level the density is written as two pieces:

One block-spin step integrates the fine links against an exactly normalized, gauge-covariant probability kernel:

The same steps transport inserted observables. This second layer produces what later parts use:

Euclidean limit (Part IV).

Symmetry (Part VI).

Phases and reconstruction (Part V).

Identification (Part VII).

The logical order is not the order of the parts. Part VI uses the OS spaces of the mixed state from Part V and the free comparison from Part VII. Part V uses the \(O (4)\) action from Part VI. Part VIII assembles everything for one family and one limit table.

Part I. The lattice theory

Unit: I.1 — Lattice, symmetries, gauge groups and the Wilson measure.

On \(\Lambda_{\ell} = (\mathbb{Z}/ \ell \mathbb{Z})^4\), a configuration assigns \(U (y, \mu) \in G\) to each positively oriented link. With \(h\) the normalized Haar measure, the Wilson measure is

\(\displaystyle d \mu_{\ell, \beta} (U) = Z^{- 1} \exp \left[ - \beta \sum_y \sum_{\mu < \nu} \left( 1 - \tfrac{1}{m} \text{Re} \hspace{0.17em} \text{tr} \hspace{0.17em} U_{y ; \mu \nu} \right) \right] \prod_{(y, \mu)} dh (U (y, \mu)) .\)

The energy lies in \([0, 12 \ell^4]\), so this is a probability measure for every real \(\beta\). It is invariant under gauge transformations, translations and the signed permutations of the axes.

Lattice schedule.

Group facts. On \(G\) the unit proves:

These are qualitative statements about single group elements; quantitative control comes in Part II.

Role of simplicity. Simplicity enters through one fact: every \(\text{Ad}\)-invariant symmetric bilinear form on \(\mathfrak{g}\) is a multiple of \(\langle \cdot, \cdot \rangle_{\mathfrak{g}}\). Together with lattice symmetry, this leaves a single marginal coupling. The same fact is used in the one-loop computation and in the uniqueness of the stress tensor.

Reflection positivity. The Wilson measure is reflection positive for every \(\beta \geqslant 0\), for reflections through planes of sites and through planes between sites (the argument of Osterwalder and Seiler). This is proved in Part IV for the original lattice measure, not for an effective action, and it is used only through the chessboard inequality.

Part II. Fields, insertions and polymer activities

Units: II.1 — Field labels and their lattice representatives; II.2 — Polymer activities, norms, the Wilson family, the construction record, reflection-positive cells.

Labels. In radial gauge, a connection near a point is described by its Taylor coefficients, and the coefficient of order \(j\) has weight \(j + 2\). \(V_d\) is the space of \(\text{Ad} (G)\)-invariant complex polynomials of weight at most \(d\) in these coefficients. Equivalently, it consists of the gauge-invariant polynomials of weight at most \(d\) in the curvature and its covariant derivatives. It is finite-dimensional, and \(1 \in V_d \subset V_{d + 1}\).

Lattice representatives. The basic lattice variables are "slots": plaquette curvatures transported to a base point along walks.

Two properties follow:

Renormalized insertions. The insertion of \(v \in V_d\) at a site \(o\) is

\(\displaystyle O_{N, p} (v, o) = a_N^{- d} \hspace{0.17em} \Phi^{(d)}_{S_{1 / \theta_N} \tau^{- 1} v} (o) .\)

The RG defines \(\tau\). The running label coordinates of an insertion obey a transfer relation from level \(N\) down to level \(m\), and \(\tau\) is their value at a reference level \(r (d)\) fixed in advance. So each insertion is normalized to have running coordinates exactly \(v\) at level \(r (d)\). All labels of weight at most \(d\), including the unit, may mix; no triangular structure is imposed. The large-field part of the inserted density goes into reserved forced polymers, so the construction never divides by the vacuum density.

Equation relations. Haar invariance gives the lattice Schwinger–Dyson identity \(\langle \langle A, \nabla_e S_W \rangle \hspace{0.17em} S \rangle_{\beta} = 0\) for \(S\) independent of the link \(e\). Products of such forces with slot polynomials span the linearized free-Maxwell equation ideal, up to terms of higher weight. They give lattice observables whose correlations with distant spectators vanish exactly. Two facts are established later, through the comparison with free theory: these relations become null in the limit, and they exhaust \(N_d\).

Polymers, budgets and marks.

Part III. The ultraviolet construction

The block-spin step and the background field

Units: III.1 — The block-spin step and the symmetric gauge fixing; III.6 — The stationary branch: background field of the effective action; III.3 — Free lattice kernels and covariances.

At level \(m\), relative to product Haar measure, the density is

\(\displaystyle \rho_m (U) = \exp \left( v| \Lambda_m | - \beta \hspace{0.17em} \mathcal{K}(U) -\mathcal{L}(U) -\mathcal{E}(U) \right) \hspace{0.17em} \mathcal{G} (U) .\)

The kernel. One step is \(\rho_m (V) = \int k_{\beta} (V, U) \hspace{0.17em} \rho_{m + 1} (U) \hspace{0.17em} dU\). Given the fine configuration \(U\), a finite tag \(t\) is drawn with probability \(\lambda (U, t)\). Each coarse link is then drawn independently, with Haar density

\(\displaystyle Z_{\beta}^{- 1} \exp \left( - \beta \kappa \left( 1 - \tfrac{1}{m} \text{Re} \hspace{0.17em} \text{tr} \hspace{0.17em} V_b Q_b (U, t)^{- 1} \right) \right),\)

where \(\kappa > 0\) is the penalty. The predictor \(Q_b\) is built in two steps:

  1. On a patch of radius \((J + 3) L\) around each coarse site, a gauge frame is chosen.

  2. Fine transports of length \(L\), read in these frames, are combined by a nonlinear (minimizing) average with a \(J\)-fold convolution weight, with a random fallback.

Properties of the kernel:

Background field. For the integration, the links of a spanning forest are set to \(1\), and each remaining link is \(\exp (a)\) times a fixed lift of the coarse field \(x\). The background is the critical branch \(a = \rho (x)\) of the classical Taylor model of the action (quadratic, cubic and quartic terms) plus the blocking penalty, at fixed \(x\). The steps are:

Free kernels. The free covariances are written in Fourier space through the actual averaging filters, which are entire and multiplicative under composition of blockings. Aliased momenta carry extra powers of \(1 / L\).

Gaussian integration and the cluster expansion

Units: III.4 — Gaussian integration: change of Gaussian sector, determinants, window law; III.5 — Cluster expansion of the polymer gas; III.8 — Connected activities and backward dependence cones.

Change of variables. Active links are written in exponential coordinates around the background. Inactive links stay in Haar measure.

Cluster expansion. The gas is \(\Xi (w) = \sum_s \prod_{i \in s} w_i\), summed over families of polymers with pairwise disjoint supports and real weights. With \(B = \lceil (\log \hat{\beta})^2 \rceil\), the smallness condition is

\(\displaystyle \sum_{i : \hspace{0.17em} x \in X_i} |w_i | \hspace{0.17em} e^{(a + \delta) \hspace{0.17em} \text{cov}_B (X_i)} \leqslant \gamma \text{for every site } x, \qquad 4 (2 B + 1)^4 \gamma / \delta \leqslant q < 1.\)

Under this condition:

So the vacuum logarithm is extensive, but a local change costs only its size. A second, graph-free argument based on a deletion recursion gives bounds of the same type.

Connected activities and dependence cones.

Large fields and one nonlinear step

Unit: III.10 — One full nonlinear renormalization step: bounds on irrelevant terms.

At reference coupling \(\hat{\beta}\) the small-field threshold is

\(\displaystyle t = \hat{\beta}^{- 1 / 2} P (\hat{\beta}), \qquad P (\hat{\beta}) = \exp ((\log \log \hat{\beta})^2),\)

well above the Gaussian fluctuation scale \(\hat{\beta}^{- 1 / 2}\).

Seeds. A seed is one of:

Energy lower bound. For every store \(A\) of seeds,

\(\displaystyle \kappa \hspace{0.17em} t^2 \hspace{0.17em} \sharp A \leqslant \text{quadratic energy of the step integrand, after subtracting the stationary predictor} .\)

The energy is a sum over five channels, one of which is the blocking penalty. The proof assigns charges to fine cells with bounded multiplicity, and controls the linearization defects on cells without seeds.

Why this suffices. Since \(\beta t^2 \simeq P^2\), each seed gains a factor \(e^{- cP^2}\). \(P^2\) grows faster than any power of \(\log \hat{\beta}\), so this gain beats:

The decomposition into large-field sectors is exact: large-field sectors are integrated, not discarded.

One step. The step takes inputs at level \(m + 1\): regular norms at most \(\hat{\beta}^{- 11 / 10}\), forced norms at most \(e^{- P^{1 / 5}}\), and the coupling in the band \([\hat{\beta} / 2, 2 \hat{\beta}]\). For \(L\) large, and then \(\hat{\beta}\) large, it produces:

The marginal and vacuum terms are extracted, and the level-\(m\) bounds restored, by the reset below.

The marginal coupling and the schedule

Units: III.7 — Matching the Wilson energy to fourth order; III.9 — The marginal flow: the running coupling and its reset; III.2 — Choosing the constants: the order of choice and the RG schedule.

Matching and the one-loop coefficient.

The value of \(\kappa_G\) comes from an explicit momentum integral. The comparison symbol is

\(\displaystyle h (k) = 4 C_{\mathfrak{g}} \left( \frac{k_1^2}{|k|^6} - \frac{4 k_1^2 k_0^2}{|k|^8} - \frac{1}{|k|^4} \right) .\)

Its angular average is \(4 C_{\mathfrak{g}} (\tfrac{1}{4} - \tfrac{1}{6} - 1) |k|^{- 4} = - \tfrac{11}{3} C_{\mathfrak{g}} |k|^{- 4}\), and \(\int_{1 / L < |k| \leqslant 1} |k|^{- 4} \hspace{0.17em} d^4 k = 2 \pi^2 \log L\). Hence \((2 \pi)^{- 4} \int_{1 / L < |k| \leqslant 1} h = - \kappa_G \log L\). Comparing the step's Hessian with \(h\) up to \(O (|k|^{- 3})\) gives the bound on \(\alpha_{N, m}\).

The reset. After the step, the blocked density is rewritten exactly, at every configuration including large fields:

The correction \(\delta\) is the isotropic Hessian coefficient of the regular activities:

Altogether, with \(\omega_j\) a drift allowance attached to regularity index \(j\),

\(\displaystyle | \beta_{N, p, m + 1} - \beta_{N, p, m} - \kappa_G \log L_m | \leqslant \omega_{J_{m + 1}} .\)

Endpoint. For real \(|b| < r_{\ast}\) there is a trajectory with level-\(0\) coupling \(\bar{\beta} + b\). Its bare coupling is unique within the construction, and the trajectory continues analytically to complex \(|b| < r_{\ast}\), with \(r_{\ast}\) independent of \(N\) and \(p\). The family uses \(b = 0\), so \(\beta_{N, p, 0} = \bar{\beta}\) for all \(N\) and \(p\). An intermediate-value shooting argument also produces such trajectories; the analytic version is the one used for convergence at fixed volume below.

Running. Put \(s_m = \log (1 / a_m)\) and \(\hat{\beta}_m = \bar{\beta} + \kappa_G s_m\). Induction from level \(0\) gives \(| \beta_{N, p, m} - \hat{\beta}_m | \leqslant \sum_{k \leqslant m} \omega_{J_k}\), and the schedule makes this \(o (s_m)\):

\(\displaystyle \beta_{N, p, m} = \bar{\beta} + \kappa_G \log (1 / a_m) + o (\log (1 / a_m)) \qquad \text{uniformly in } N \geqslant m \text{ and } p.\)

At \(m = N\) this is the slope law with \(b_{\ast} = \kappa_G\). Since \(\bar{\beta}\) is chosen large, every level stays in the band \([\hat{\beta}_m / 2, 2 \hat{\beta}_m]\). The statement is leading-logarithmic only: there is no two-loop term.

Why the block ratios grow. The one-loop error per step is bounded but does not decrease with \(L\). For the accumulated error to be \(o (s_m)\), the allowance per unit of logarithmic scale must tend to zero. The schedule therefore imposes two conditions:

The ratios are constant while the regularity index is fixed, \(L_m = \ell_{J_{m + 1}}\). The regularity index \(J\) is nondecreasing, rises by at most one per level, tends to infinity toward the fine end, and stays constant on long windows of levels. One sequence \((L_m)\) serves every cutoff \(N\).

Order of choices. The constants are chosen in a fixed order:

  1. the kernel with its penalty, and for every index \(j\) its regularity profile (derivative order, spatial power, support exponent, widths, path counts, drift allowance), together with the one-loop error constant \(D\);

  2. all block ratios \(\ell_j\), so that the allowances and the logarithms of the combinatorial costs are at most \(j^{- 2} \log \ell_j\);

  3. the smallness exponents;

  4. the entry couplings and the volume conditions;

  5. \(\bar{\beta}\);

  6. the assignment of regularity profiles to levels.

Each degree \(d\) has a reference level \(r (d)\) with regularity index at least \(d + 3\). Beyond \(r (d)\) the derivative order exceeds \(d\) and the spatial power is at least \(d\). All choices precede any observable or extraction order, so every degree is served by the same family.

Renormalized observables, uniformly in the volume

Units: III.11 — Comparison with the free Gaussian measure: interacting rank and determinant budgets; III.12 — Bounds uniform in the volume and convergence as the spacing tends to zero; III.15 — Cost accounting for observables and the remainder comparison; III.16 — Analytic toolkit.

The chessboard root. The quantity propagated for observables is the chessboard root. For a function \(F\) of the links in one cell of a tiling of the torus by cells of side \(a_m\),

\(\displaystyle \lVert F \rVert_T = | \hspace{0.17em} \mathbb{E}_{\beta_{N, p}} \left[ \prod_{\operatorname{cells}} \text{reflected copies of } F \right] |^{1 / \sharp \text{cells}} .\)

The main bound reads as follows. For every degree \(d\), radius \(R\) and \(\varepsilon > 0\) there is \(A\) such that, for every \(m \geqslant r (d)\), there is \(N_0\) independent of \(p\) with

\(\displaystyle \lVert O_{N, p} (v, o) \rVert_{T_m (o)} \leqslant A \hspace{0.17em} a_m^{- d - \varepsilon} \qquad (N \geqslant N_0, \text{all } p, \lVert v \rVert \leqslant R) .\)

It is proved by running the insertion through the same RG steps; \(\varepsilon\) absorbs losses smaller than any power. Reflection positivity enters only afterwards, through the chessboard inequality. That inequality makes the root a seminorm and bounds correlations of separated observables by products of roots.

No loss of \(\log \beta\) per site. A root over \(c\) copies tolerates a fixed cost per copy, but not a factor \(\log \beta\) per site. The determinant budget avoids this by a gauge count on a terminal torus with \(c = L^4\) sites:

Field content. Free Gaussian detectors give uniform rank detection: \(\lVert \pi x \rVert \leqslant Z \hspace{0.17em} \lVert \text{probe} (\pi x) \rVert\), where \(\pi\) is the projection onto quotient labels. The limit field spaces are isomorphic, degree by degree, to those of free Maxwell theory, compatibly with conjugation and reflections.

Difference quotients.

These bounds are propagated through the RG; they do not come from differentiating a sup norm.

A rank-free trace bound. Let \(X \geqslant 0\) be a matrix and let the insertions \(b_a\) be separated by gaps of more than \(2 r\) factors of \(X\). Moving \(m \leqslant r\) factors from one gap to another changes \(\text{tr} \prod_a b_a X^{Y_a}\) by at most \((\text{cost} / r) \hspace{0.17em} m\). The cost is a product of trace norms with no dimension factor. The proof is a Hölder inequality for complex exponents, obtained from Hadamard's three-lines theorem, followed by a Cauchy estimate. By quadrature it extends to continuous positive kernels. Part IV uses it for equicontinuity.

Bookkeeping. The numerical conditions of all steps can be met in one fixed order: constants before \(L\), then \(L\), then the later constants, then \(\beta\) large. This holds uniformly across regularity profiles.

Convergence at fixed volume and the construction run

Units: III.12; III.13 — Extension to all labels: one common Wilson family; III.14 — The construction run: flat representations and the ultraviolet rate.

The obstacle. At fixed volume, two cutoffs \(N\) and \(N'\) have different bare couplings, each fixed by the condition \(\beta_{N, p, 0} = \bar{\beta}\). Continuity in a fixed bare coupling therefore compares nothing.

Fixed-level convergence and tails. The construction has a one-parameter family of trajectories:

Combined with the difference bounds, this gives at each fixed volume:

The family used by the public theorem. VIII.1 assembles the construction for one family, using the bounds and convergence results described in III.9–III.12. III.13 describes an alternative coordinated family with root bounds, an analytic shooting family and Maxwell rank comparisons. Its existence statements do not identify that family with the one used by YangMills.constructed; properties from the two constructions should not be combined without such an identification.

The construction run (III.14). III.14 packages the construction run. It also proves quantitative convergence of the order-zero quadratic part of the effective action (its position-space Hessian kernel) at each fixed level \(i\), at rate \((a_N / a_i)^c\).

Part IV. Infinite volume

Unit: IV.1 — Infinite volume: extraction of a subsequence and the Euclidean axioms of the limit.

Local bounds.

Equicontinuity. Let \(\text{Gen}_n\) be the set of configurations whose points have pairwise distinct coordinates on every axis.

  1. Integrating out the links along one axis gives a positive transfer kernel \(K\). A correlation becomes \(\text{Tr} \prod_k B_k K^{Y_k} / \text{Tr} \hspace{0.17em} K^M\), with slab kernels \(B_k\) and gaps \(Y_k\).

  2. Moving one point shifts units between two gaps. The trace is entire in the size of the shift.

  3. The rank-free Hölder bound controls the trace on a disc, and a Cauchy estimate gives a Lipschitz bound.

  4. The trace norms of the slab kernels are themselves chessboard roots of \(2 q\) alternating copies, so the bound is uniform in the volume.

The result is \(|C (x) - C (y) | \leqslant 16 nA^n |x - y| / R\) on compact subsets of \(\text{Gen}_n\) whose coordinate gaps exceed \(R\). No derivative bound is used.

Extraction. Arzelà–Ascoli is applied twice, over countably many degrees, basis labels and compact sets: first to extract a spacing subsequence that works for all volumes, then a volume subsequence. This gives limit kernels on \(\text{Gen}_n\). The complement \(\text{Conf}_n \setminus \text{Gen}_n\) is a null set, and the kernels are locally bounded on \(\text{Conf}_n\), so dominated convergence gives the smeared limits.

Derivatives and smoothness.

\(\displaystyle \partial_{x_i^a} S (\ldots, v_i, \ldots ; x) = S (\ldots, \partial_a v_i, \ldots ; x) \qquad \text{in } \mathcal{D}' (\text{Conf}_n) .\)

Iteration and a Fourier argument give \(S \in C^{\infty} (\text{Conf}_n)\), with derivatives polynomially bounded near collisions.

What the limit inherits. The limit inherits:

It does not inherit continuous rotations, clustering or uniqueness.

Removing the spacing refinement. The extraction refines both the spacing and the volume sequences. At each admissible fixed volume, the smeared limit along the whole spacing sequence exists (Part III) and agrees with the subsequential one. This removes the spacing refinement from the final statement.

Part V. Phases and relativistic reconstruction

Units: V.1 — Decomposition of the limit into phases; V.2 — Osterwalder–Schrader reconstruction of a Wightman theory; V.3 — Reconstruction for every phase and the reconstructed theory.

The OS space of the mixture. Positive-time field words, modulo the null space of \(\langle \theta w, w' \rangle_S\) and completed, form a Hilbert space. It carries:

The joint spectrum lies in a half-space. No positive spectral gap is asserted. The zero-energy space may be large. Its projection \(\Pi\) is the strong limit of \(e^{- tH}\) as \(t \to \infty\).

Thin slabs. A word \(w\) placed in a thin time slab defines an operator \(B\) on the OS space. To bound it, consider the train \(v_{n + 1} = B^{\ast} Bv_n\), which contains \(n\) reflected copies of \(w\) and an arbitrary exterior word.

So the exterior word is paid once and disappears under the \(n\)-th root.

Phases.

The compressions commute:

  1. Place four clusters on the rays \(n (ue_4 + ve_1)\), with \((u, v) = (0, 0), (1, 2), (2, 1), (3, 3)\).

  2. Exchanging the time axis with a spatial axis reverses the order of the middle pair.

  3. \(O (4)\) covariance and permutation symmetry of \(S\) then give \(\langle \Omega, M_x M_y M_w M_z \Omega \rangle = \langle \Omega, M_x M_w M_y M_z \Omega \rangle\).

  4. By density, \([M_y, M_w] = 0\).

The resulting phase decomposition:

Reconstruction. For each phase, slab operators bounded by \(C \hspace{0.17em} p^{- A}\) (with \(p\) the slab half-width) define

\(\displaystyle F_n (z ; \mathbf{x}) = \left\langle \Omega, \hspace{0.17em} B (v_1) \hspace{0.17em} e^{i (z_1 - z_2) H} U (\mathbf{x}_1 -\mathbf{x}_2) \hspace{0.17em} B (v_2) \cdots \Omega \right\rangle .\)

This function is holomorphic for \(\text{Im} \hspace{0.17em} z_1 < \ldots < \text{Im} \hspace{0.17em} z_n\), and it equals the Euclidean correlation at imaginary times. The Wightman theory is obtained in five steps:

  1. Boundary values. Bounds of order \(k^{- m}\), when all height gaps exceed \(k\), give tempered boundary values \(W_n\). They converge against every joint Schwartz test as the vector of heights tends to zero through all ordered vectors.

  2. Positivity and fields. This mode of convergence allows two independently regularized vectors to be compared. So \(W\) is a limit of Gram matrices of OS vectors: it is positive, and the Wightman space embeds isometrically into the OS space. Fields act on finite words modulo the null space by prepending a letter. This gives the invariant domain, the formal adjoints and the cyclic vacuum.

  3. Covariance and spectrum. The rotation Ward identities of Part VI, in planes containing the time axis, continue to infinitesimal boost identities. These integrate to Lorentz covariance without any spectral assumption. The spectral measure of a Wightman vector equals that of its OS image, so it is supported in \(p^0 \geqslant 0\); covariance then moves the support into the forward cone.

  4. Locality. Only adjacent transpositions are needed. For a spacelike adjacent pair, both orderings have holomorphic germs, constructed with complex boosts. The germs agree to all orders on the Euclidean slice, because the Euclidean kernel is symmetric and smooth away from coincidences.

  5. Faithfulness. A label whose Wightman functions vanish has a vanishing tube function, hence a vanishing Euclidean kernel.

Every phase. The low-energy bound uses only locality and the spectrum condition. Take \(A\) a finite sum of field words with compactly supported tests.

\(\displaystyle \mu_{A \Omega_{\xi}} (0 < E \leqslant u) \leqslant M_A u^2 (0 < u \leqslant 1), \qquad 0 \leqslant \langle A \Omega_{\xi}, (e^{- sH_{\xi}} - P_{0, \xi}) A \Omega_{\xi} \rangle \leqslant M_A s^{- 2} (s \geqslant 1) .\)

The constant \(M_A\) is independent of \(\xi\), by compactness of \(X\).

Vacuum uniqueness then extends from almost every phase to every phase:

The comparison map from the Wightman space to the OS space is surjective, proved independently by pushing time-ordered vectors apart. This transfers vacuum uniqueness to the OS space. Each phase theory is irreducible. Its Wightman distributions are continuous in \(\xi\), and their \(\nu\)-average is the boundary value of the continuation of \(S^{\text{mix}}\). The bounds allow spectrum down to zero and give only polynomial decay \(s^{- 2}\) in Euclidean time, not exponential clustering.

Part VI. Ward identities, rotations and the stress tensor

Units: VI.1 — Ward identities, the rotation action and the conserved stress tensor; VI.2 — The free Maxwell stress tensor and the local Ward charge.

The lattice has only the hypercubic symmetries. The proof obtains the generators of rotations from stress fluxes, and the group action by comparison with free Maxwell theory. It uses the OS spaces of the mixed state in the coordinate directions and the free comparison of Part VII. It uses no decay of correlations.

Free Maxwell theory. Consider \(\dim \mathfrak{g}\) free Maxwell fields, with field spaces the \(G\)-invariant polynomials in on-shell curvature jets.

Conserved tensor. There are scaling sequences \(r_k \downarrow 0\) with frames \(U_k\), isomorphisms for large \(k\). Along them the rescaled interacting divergence converges to the free one.

Shell identities. Lattice integration by parts in the link variables, with ultraviolet control of the error terms, gives shell identities. Take \(\chi\) equal to \(1\) near \(x_0\) and vanishing near the spectators:

\(\displaystyle - \sum_a \int \partial_a \chi (y) \hspace{0.17em} S \left( J_{ab} (y) \hspace{0.17em} B (x_0) \cdots \right) \hspace{0.17em} dy = S ((p_b B) (x_0) \cdots) .\)

With weight \(K (y - x_0)\), for a rotation generator \(K\), the same construction gives maps \(Y_K\). Here \(p_b\) and \(Y_K\) are linear maps on labels, not yet identified with derivatives and rotations. Moving the centre of the shell gives \(\partial p = p \hspace{0.17em} \partial\) and \(\partial_b Y_K + \sum_c K_{cb} \hspace{0.17em} p_c = Y_K \partial_b\).

Flux at infinity.

Normalization. In free theory, the finitely many relations satisfied by \(p\) and \(\partial\) at a given degree have a one-dimensional solution space. Injectivity is an open condition, so the same holds along the scaling sequence. Hence \(p_b = c \hspace{0.17em} \partial_b\), with \(c\) independent of the degree and \(c \ne 0\). The tensor \(T = J / c\) satisfies

\(\displaystyle - \sum_a \int \partial_a \chi (y) \hspace{0.17em} S^s \left( T_{ab} (y) \hspace{0.17em} B (x_0) \cdots \right) \hspace{0.17em} dy = \frac{\partial}{\partial x_0^b} \hspace{0.17em} S^s (B (x_0) \cdots)\)

in the mixed state. By the operator identities of Part V, it holds in every phase. Since \(\partial \chi\) vanishes near every insertion, no contact terms arise.

All separated configurations. Positivity of Gram functions, together with the Ward values already obtained, gives analytic-vector bounds \(\lVert f^{(m)} \rVert \leqslant C (\rho / 2)^{- m} m!\). The angular Ward functional is therefore real-analytic along walks that move one point at a time. Starting from configurations on a line, it vanishes on all configurations whose points have pairwise distinct first coordinates, and by continuity on all of \(\text{Conf}_n\).

Integration to \(O (4)\).

Part VII. Identification as Yang–Mills theory

Free Maxwell fields at short distance

Units: VII.1 — Free Maxwell comparison theory; VI.2 — Interacting stress comparison; VII.6 — Scaling variants and fixed-field estimates. The comparison is assembled in VIII.1.

The reference theory. It consists of \(\dim \mathfrak{g}\) free Maxwell fields built from \(\Gamma (x) = 1 / (4 \pi^2 |x|^2)\).

The comparison. There are scales \(r_k \downarrow 0\) and linear frames \(U_k\) on the field spaces with the following property. The correlations of the mixed state, rescaled by \(r_k\) and transformed by \(U_k\), converge to the free Maxwell correlations:

The stress tensor converges to the Maxwell stress tensor. An upper bound on \(\dim F_d\) makes the frames isomorphisms for large \(k\), so \(F_d \cong M_d\).

The frames are necessary: the claim is not that a fixed normalization of each composite field has a limit along every sequence of radii. The schedule contains a real parameter that moves the scale at which observables are read, and the estimates are uniform in it. Tuning this parameter, every sequence of radii tending to zero has a subsequence along which the comparison holds.

Running coupling

Unit: VII.5 — Asymptotic freedom of the running coupling.

With \(g_{N, p} (m) = \beta_{N, p, m}^{- 1 / 2}\), Part III gives

\(\displaystyle g_{N, p} (m) \hspace{0.17em} \sqrt{\log (1 / a_m)} \longrightarrow \frac{1}{\sqrt{2 b_0}} \qquad (m \to \infty),\)

uniformly in \(N \geqslant m\) and \(p\). This is the leading-logarithmic running of the coupling of the construction's effective actions. No two-loop term, no \(\Lambda\) parameter and no comparison with another renormalization scheme is claimed. For \(SU (N)\), the conventionally normalized coupling \(g_{\text{std}} = \sqrt{2 N} \hspace{0.17em} g\) satisfies \(g_{\text{std}} \sqrt{\log (1 / a_m)} \to (11 N / 24 \pi^2)^{- 1 / 2}\).

The trace law and nonfreeness

Units: VII.2 — The trace of the stress tensor and the trace anomaly; VII.3 — The theory is not free: Gram positivity and exclusion of free-field profiles.

Leading form of the trace. The classical Maxwell and Yang–Mills stress tensors are traceless. The quantum trace comes from the scale dependence of the coefficient of curvature squared.

Along the scaling sequences, with \(\beta_k\) the running coupling at the RG level that reads scale \(r_k\), and \(O_{F^2}\) the normalized curvature-squared label,

\(\displaystyle \beta_k \hspace{0.17em} r_k^4 \hspace{0.17em} U_k^{- 1} \Theta \longrightarrow 4 d_2 \hspace{0.17em} O_{F^2}, \qquad d_2 = - \frac{\kappa_G}{4}, \qquad \frac{\log (1 / r_k)}{\beta_k} \longrightarrow \frac{1}{\kappa_G} .\)

In perturbative terms, this is the familiar \(\Theta \simeq \frac{\beta (g)}{2 g^3} F^2\), with \(\beta (g) / (2 g^3) = - b_0 / 2 = d_2\) at one loop. The trace is a curvature-squared field suppressed by \(g^2 (r) \simeq 1 / (\kappa_G \log (1 / r))\).

The two-point law. Here \(\Theta = \sum_{\mu} T_{\mu \mu} - S^{\xi_{\ast}} \left( \sum_{\mu} T_{\mu \mu} \right) \boldsymbol{1}\), in the selected phase, which has a unique vacuum.

  1. The free two-point function of \(O_{F^2}\) is \(3 \dim \mathfrak{g}/ (\pi^4 |x|^8)\).

  2. So at lattice-fitted configurations along the scaling sequence, \(r_k^8 \log^2 (1 / r_k) \langle \Theta \Theta \rangle \to (4 d_2 / \kappa_G)^2 \cdot 3 \dim \mathfrak{g}/ \pi^4 = 3 \dim \mathfrak{g}/ \pi^4\).

  3. Rotation invariance moves the pair onto the time axis.

  4. The spectral representation in the phase gives \(G_{\Theta} (r) = \int e^{- rE} \hspace{0.17em} d \mu (E)\) with a positive measure. So \(G_{\Theta}\) is nonincreasing, which sandwiches \(G_{\Theta} (r)\) between its values at fitted points.

  5. If the limit failed along some sequence of radii, the flexible scaling above would give a subsequence along which it holds, a contradiction.

Hence the limit exists over all radii:

\(\displaystyle \lim_{r \downarrow 0} \hspace{0.17em} r^8 \log^2 (1 / r) \hspace{0.17em} G_{\Theta} (r) = \frac{3 \dim_{\mathbb{R}} \mathfrak{g}}{\pi^4} .\)

The amplitude does not depend on the value of the one-loop coefficient. \(\kappa_G\) enters once through the trace and once through the running, and \((4 d_2 / \kappa_G)^2 = 1\) identically. So the test detects the inverse-logarithmic suppression and the free normalization of \(F^2\), not the value of \(b_0\); that value is established by the running law of Part III.

Exclusion of free profiles. Consider a finite Wick profile \(W (r) = \sum_{\alpha} c_{\alpha} \prod_{\beta} K_{\alpha \beta} (r)\). Each factor \(K (r) = \int E^q \hspace{0.17em} h (\mathbf{p}) (2 E)^{- 1} e^{- rE} \hspace{0.17em} d^3 \mathbf{p}\) has \(E = \sqrt{\textsf{m}^2 + |\mathbf{p}|^2}\) and \(h\) homogeneous.

This excludes exactly the finite class of profiles in the statement, and it concerns only \(G_{\Theta}\) on the time axis.

The stronger statement adds further tests:

Operator product expansion and contact extensions

Unit: VII.4 — Operator product expansion and contact extensions.

The lattice statement. The product of insertions in one cell equals a single insertion of the same total degree, plus a remainder with small chessboard root. The proof is an induction over the RG levels between the cluster scale and the observation scale: the part of degree at most \(D\) is kept, and the rest is paid into the remainder.

The continuum coefficients. They are read off rather than given by a formula.

Consider an \((r, M, \eta)\)-cluster: points within \(Mr\) of \(o\) and at mutual distance at least \(\eta r\), carrying fields of total degree \(w\). Take spectators at distance at least \(c\) from \(o\) and from each other. Then, for all sufficiently small \(r\),

\(\displaystyle |c_{j, \kappa} (a ; x, o) | \leqslant r^{\hspace{0.17em} j - w - \varepsilon}, \qquad | S^s (a, x ; b, y) - \sum_{j \leqslant D} \sum_{\kappa} c_{j, \kappa} (a ; x, o) \hspace{0.17em} S^s (O_{j, \kappa}, o ; b, y) | \leqslant r^{\hspace{0.17em} D + 1 - w - \varepsilon} .\)

Uniformity and limits of the statement.

Contact extensions. These use Taylor subtractions over forests of collapsing clusters on dyadic shells. The OPE remainder makes the shell sums converge. For every state and every list of labels this gives a distribution of finite order on \((\mathbb{R}^4)^n\) that agrees with the correlation on \(\text{Conf}_n\). Not established: uniqueness, reflection positivity on overlapping supports, contact Ward identities, real-time products, and convergence of the lattice correlations on the diagonals.

Part VIII. Assembly

Unit: VIII.1 — Assembly of the main theorem.

The construction run depends only on \(G\) and the penalty \(b\). It is chosen before any extraction order or observable. For each extraction order (a spacing sequence and a volume sequence), the assembly composes further refinements:

To these it adds the reconstructions of all phases, the OPE prescription, the nested and contact results and the Maxwell comparison. Everything is transported to a common refinement, so all objects refer to one table.

The public theorem is the case \(b = 1\), with the full spacing sequence and the given volume sequence; its \(q\) is the volume part of the refinement.

No new estimate enters, but identifying all objects with one table is a real constraint. Separate existence results for a lattice limit, a Wightman theory, an OPE and a trace law would not show that they concern the same correlation functions. The formal construction supplies the hypotheses under which the limit theory is first proved. The outlines have not explained every supplier in full; the assembly unit records the limits of its source reading.

Relation to earlier work, and where the difficulty lies

Classical sources.

Departures from Balaban's scheme.

Beyond the ultraviolet analysis, the proof does the following:

Where the difficulty lies.

  1. Gauge degeneracy and large fields must be controlled with a bounded cost per site and per seed. This is done by the gauge count of the determinant budget, the energy lower bound for seeds, and the exactness of the kernel.

  2. Effective-action estimates must become estimates for observables: for every composite field, its derivatives and its fixed-volume limit, inside one family.

  3. The passage from a lattice mixture to relativistic phases has no continuous symmetry and no clustering to start from. It is handled by Ward identities with free comparison, the commutative zero-energy algebra, and spectral bounds based on locality.

The order of choice of constants and the transport of estimates through the construction account for a substantial part of the formal development. Its length alone does not measure mathematical novelty.

Steps whose informal account is incomplete

The theorem is proved in Lean, and the conditional statements of the limit theory are discharged within its proof. The unit outlines, however, describe the following steps only at the level of statements, or with readings marked tentative. A written proof will have to supply them.

Reading guide

The exact statement separates the public theorem from the stronger construction. The architecture describes what each part hands to the next. For the stated inputs and remaining gaps in this account, follow the relevant unit's “What it uses” and “What remains unexplained” sections. The assembly unit describes how the construction is put together.

Question Units to read
What are the fields, and how are their normalizations fixed? II.1, III.12, IV.1, VII.1
How does the ultraviolet analysis relate to Balaban's? III.1, III.3, III.4, III.6, III.5, III.10
Where does the one-loop coefficient enter? III.7, III.9, VII.5, VII.2
Why does one family serve every observable and every limit? III.2, II.1, III.12, VIII.1; III.13 treats an alternative family
How are bounds made uniform in the volume? III.11, III.12, III.16
How is infinite volume reached without an infrared RG? III.16, IV.1
Why are there phases, and why does every phase reconstruct? V.1, V.2, V.3; VII.4 for the train bounds
How are rotations and the stress tensor recovered? VI.2, VI.1, with VII.1 and VII.6 for the free comparison
What makes the theory nonfree? VII.2, VII.3, with VII.6 for arbitrary short-distance scales
What does the OPE say, and what is known at coincident points? VII.4, then VIII.1 for the assembly