The order of limits 1
The eight parts 2
Ultraviolet units 2
What each part hands to the next 2
Ultraviolet → infinite volume 2
Infinite volume → phases, reconstruction and symmetry 3
Reconstruction → identification 3
Symmetry → identification 3
Into the assembly 3
The work items of the limit theory 3
What does not follow the order of the parts 4
How reliable is this 4
This page describes how OpenAI's Lean proof of the construction of four-dimensional Yang–Mills theory is organised: its eight parts, what each part hands to the next, and the units that make up each part (detailed map).
It is read from the Lean statements and documentation and from the dependency graph. The mathematical framing follows the Yang–Mills group's assessment. Lean names link to the fixed source commit 3456e5e.
The construction uses Wilson lattice theories on tori of side \(L_p\), with lattice spacings \(a_N\) and couplings \(\beta_{N, p}\) satisfying \(\beta_{N, p} / \log (1 / a_N) \to 11 C_G / (24 \pi^2)\). The public theorem leaves the positive limiting constant unspecified; see the exact statement.
Correlations of renormalized local observables are first taken to the limit \(N \to \infty\) at fixed volume, along the whole sequence of spacings. They are then taken to the limit \(p \to \infty\), along a subsequence that may be extracted from any prescribed sequence of volumes. The public form of this statement is WilsonFamily.Converges.
The renormalization group is used only for the first limit. The second is obtained by compactness.
| part | name | units |
| I | Setting | I.1 |
| II | Data of the construction | II.1, II.2 |
| III | Ultraviolet construction | III.1–III.16 |
| IV | Infinite volume | IV.1 |
| V | Phases and reconstruction | V.1, V.2, V.3 |
| VI | Symmetry | VI.1, VI.2 |
| VII | Identification as Yang–Mills | VII.1, VII.2, VII.3, VII.4, VII.5, VII.6 |
| VIII | Assembly | VIII.1 |
Blocking, constants and Gaussian estimates: III.1, III.2, III.3, III.4.
Cluster expansion, stationary field, matching and dependence: III.5, III.6, III.7, III.8.
Coupling reset, nonlinear step and observable limits: III.9, III.10, III.11, III.12.
Alternative coordinated family, construction data and analytic tools: III.13, III.14, III.15, III.16.
The limit theory (Parts IV–VI) is written in the code around numbered work items; the units of those parts follow them.
Let \(G\) be a compact matrix group with simple nonabelian Lie algebra, and \(b > 0\) a parameter of the block-spin kernel. The ultraviolet construction provides a Wilson family with the following properties.
Specification. The family satisfies its normalization, covariance and budget conditions, and its couplings obey the asymptotic-freedom law above.
Chessboard bound, uniform in the volume. For each degree \(d\) and \(\varepsilon > 0\), every renormalized local observable at block scale \(m\) satisfies a chessboard bound of order \(a_m^{- d - \varepsilon}\), uniformly in the volume, once the lattice is fine enough. A chessboard bound is a bound on the root of a product of reflected copies.
Differences. The same holds for finite-difference quotients, so correlation kernels at \(r\)-separated points are bounded uniformly in the volume.
Fixed volume. At each fixed volume the correlation kernels converge along the whole sequence of spacings, to smooth limits.
Corrected fields. The same estimates hold for the corrected fields.
For the subsequential extraction, the infinite-volume part uses 1–3 and 5, together with the reflection-positivity inequality on cells, which holds for every \(\beta \geqslant 0\). Its conclusion: for every increasing sequence of volumes there is a subsequence along which all smeared correlations converge to kernels that are smooth away from coincident points. The fixed-volume convergence in item 4 then removes the spacing refinement; see IV.1. These two limits correspond to Theorems 4.1 and 5.1 of the group's assessment.
A caution about the code. About three quarters of the code filed under "full volume" (folder YM/FullVolume) belongs to the ultraviolet construction: it proves the volume-uniform bounds. The genuinely infinite-volume argument is short, about 7,000 lines.
The limit satisfies a Euclidean hierarchy:
translation homogeneity;
reflection positivity, including for smeared joint test functions;
smoothness on separated configurations, with controlled derivatives;
a compatible calculus of derivatives.
Reconstruction also needs a uniform bound on products of fields in thin time slabs.
For every phase, a completed Wightman theory: unique vacuum, a light-cone bound, heat-kernel decay, and a trivial weak commutant. The heat bound is polynomial; no mass gap follows.
A conserved stress tensor and its Ward identities. Both are compared, along sequences of scales tending to zero, with the corresponding objects for free Maxwell fields.
A single Wilson family serves every part of the public theorem, from extraction to the trace test; VIII.1 describes the assembly. The construction run exists for every allowed \(G\) and every \(b > 0\). III.13 describes an alternative coordinated family; its properties require a separate identification before they can be used for this family.
Parts IV–VI are written in the code as about seventy numbered work items. Each states one obligation, a "proof request" in the code's vocabulary, and proves it in its own file.
| family | items | part | content |
| D | 6 | IV | the limit passes to derivatives of the correlation kernels; the Euclidean hierarchy |
| B | 18 | V | time-slab operators, decomposition into phases, reflection positivity and smoothness for each phase |
| C | 29 | V | Osterwalder–Schrader → Wightman: continuation in time, boost covariance, gluing of tubes, locality |
| A | 17 | VI | so(4) brackets, integration to the rotation action, the averaged conserved stress tensor, boosts, comparison with free Maxwell fields |
These families interleave. D supplies the limit hierarchy; B begins with its OS spaces and slab operators. A uses those spaces and the free Maxwell comparison to obtain rotations and Ward identities. Those symmetry results enter both the phase decomposition in B and the boost and covariance arguments in C. The part numbers organize the subjects; they do not give a linear dependency order. See the strategy and the work-item introductions.
The same mechanism appears in the ultraviolet construction: families E–I cover the energy chain of one renormalization step, its stationary point and Hessian, label polynomials, and support bookkeeping.
See work-items.md for the complete list.
The Lean is not written in this order everywhere. Some definitions are needed before the part that explains their properties:
the reflection-positive cells and chessboard roots;
the record of the whole construction run;
the basic types of the renormalization group and of the labels.
Some sections are filed far from their mathematical home. Two examples:
the rotation-invariance estimate, in folder Next/ROTATIONS;
some renormalization-group pieces filed with the trace-anomaly code.
None of this involves new mathematics. A written account should simply state these definitions early.
Reachable declaration sets for seventeen named results were compared with OpenAI's compiled record. This comparison does not verify every edge or the mathematical interpretation of the interfaces (method).
The plain-language interfaces summarize statements and documentation. The unit index gives each outline's review status; its local limitations describe which arguments remain at statement level. The exact statement distinguishes the public theorem from additional construction conclusions.
Report a correction with the page, heading and source reference.