Disclaimer: this is unverified, and we are simply attempting to map out these results and their implications carefully.
Metadata from the proof (due to Levent Alpöge)
| Result | Role | Statement |
|---|
At the level of declarations, the proof constitutes a directed graph on 128,479 vertices and 573,122 edges; we note here that an edge from $u$ to $v$ implies that the statement or proof of $u$ refers to $v$. The density of this graph is $3.5 \times 10^{-5}$, and its in-degree distribution is heavily tailed, in the sense that 58% of declarations are cited at most once, about 0.4% are cited at least a hundred times, and the most cited is used 6,641 times. Since Lean forbids circular dependencies, the graph is acyclic apart from 27 small cycles introduced by the name matching that we used to construct it; the shortest paths from riemannHypothesis reach depth 56, while the longest dependency chain has 139 steps.
Contracting each module group to a vertex gives a connected graph on 771 vertices and 3,862 edges, with mean degree 10, diameter 9, mean distance 3.1 and clustering coefficient 0.52. This gives a small-world pattern of a few hubs joining many dense local clusters. Then, the degree assortativity of the graph, $$r \;=\; \frac{\frac{1}{M}\sum_{e} j_e k_e \;-\; \Big[\frac{1}{M}\sum_{e} \tfrac{1}{2}(j_e + k_e)\Big]^2}{\frac{1}{M}\sum_{e} \tfrac{1}{2}\big(j_e^2 + k_e^2\big) \;-\; \Big[\frac{1}{M}\sum_{e} \tfrac{1}{2}(j_e + k_e)\Big]^2},$$ where the sums run over the $M$ edges and $j_e, k_e$ are the degrees of the two endpoints of the edge $e$, is $r = -0.13$, so such hubs attach primarily to small groups. Five groups hold 65% of the declarations, and Development.AbelianVarietyTheory.Proofs alone has normalised betweenness 0.48.
We call a pillar one of the six results in the upper layers of the tree on which the remainder of the proof primarily relies. These include the analytic reduction (rh_of_every, critLine), the Galois representations of Row0b, the theta lifts of TLiftK, the local Langlands correspondence (recTExists), the Selmer bound VWin, and the bridge from VWin to nonvanishing. A declaration is reached by $p$ if it can be arrived at from $p$ by following references, that is, if it appears in the statement or proof of $p$, or in the statement or proof of a declaration that does so, and so on down to the foundations. For a pillar $p$ let $D(p)$ be the set of declarations reached by $p$, and let $$E(p) \;=\; D(p) \setminus \bigcup_{q \neq p} D(q)$$ be its exclusive support. That is, $E(p)$ is the component of the formalization that serves $p$ and is exclusive to $p$. The six pillars that we identified together reach 128,464 of the 128,479 declarations under riemannHypothesis.
Three findings emerge from this pillar-based characterization. First, the pillars of local Langlands and of the theta lifts are not independent blocks: $D(\texttt{recTExists})$ lies entirely inside the support of the Galois representations, as do 289 of the 291 declarations under TLiftK, so both enter the proof as ingredients of the construction of Galois representations, which alone holds 78,765 exclusive declarations, or 61% of the proof. Second, the Selmer bound VWin reaches 28,257 declarations, of which 14,315 are shared with the Galois block and 13,934 are its own, concentrated in archimedean base change, block decompositions and inertia eigenvalues. Third, the bridge from VWin to nonvanishing is almost disjoint from the rest of the proof: 18,825 of its 19,182 declarations (98%) are exclusive, and they lie in arithmetic geometry, namely embedded resolution of singularities, the arithmetic Hilbert–Samuel theorem and ampleness of heights. If the proof contains a genuinely new argument, these two blocks, the Selmer bound and the height-theoretic bridge, are where this new argument might lie.
Row $a$, column $b$: the share of $D(a)$ that also lies in $D(b)$.
I compiled a series of documents that record the formalization at different levels of detail and that can be used to accompany the above visualization. The two “statements” files give all of the theorems and statements from the first part of the archive, grouped by file; and the table of declarations reached from riemannHypothesis gives the kind, module, depth and statement of all of the key results. This makes dependencies searchable and sortable. The last file gives the principal statements that are closest to the main theorem.