Hilbert–Smith: character partitions and a fixed sphere lattice

Source and verification

This article explains family 304 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

Overview

Hilbert–Smith asks whether a locally compact group acting faithfully and continuously on a manifold must be a Lie group. Family 304 claims an affirmative answer in every finite dimension. Its obstruction to a faithful $p$-adic action compares arbitrarily fine character partitions with a single fixed denominator in a sphere invariant. The denominator must be fixed before the partition; a rational calculation alone would not suffice. (OpenAI, 2026)

🏷️ Manifold actions and the claimed theorem

A topological manifold here is Hausdorff, second countable, finite dimensional, and without boundary. Its charts are homeomorphisms to open subsets of Euclidean space; neither smoothness nor a triangulation is assumed.

An action is jointly continuous when the map $(g,x)\mapsto gx$ is continuous. It is faithful when the only element fixing every point is the identity. This allows individual points to have nontrivial stabilizers. A Lie group has a finite-dimensional smooth manifold structure compatible with its group operations.

Manuscript claim

A locally compact second-countable Hausdorff group acting faithfully and jointly continuously on a connected topological manifold is a Lie group. The manifold can be noncompact, nonorientable, or nontriangulable. (OpenAI, 2026)

The zero-dimensional connected manifold is a point, for which faithfulness forces the trivial group. The substantive case has dimension $n\ge 1$.

The classical Hilbert–Smith reduction, imported by the manuscript, reduces the problem to excluding faithful actions of

$${\mathbb{Z}}_p=\underset{j}{\underleftarrow{\lim }}\mathbb{Z}/p^j\mathbb{Z}.$$

This is the compact group of $p$-adic integers. Its nonzero closed subgroups are $p^j{\mathbb{Z}}_p$. Consequently, proving that every such action has nonzero kernel makes it factor through a finite quotient.

The post explains selected parts of the proposed proof. The sheaf-duality and localization foundations are substantial manuscript inputs, not independently established here.

🏷️ An invariant chart and an odd sphere

Suppose a faithful ${\mathbb{Z}}_p$-action exists. A local reduction using Newman’s theorem supplies an open subgroup $G\cong {\mathbb{Z}}_p$ acting faithfully on a connected invariant open set $O\subset {\mathbb{R}}^n$. Newman’s theorem says that a periodic homeomorphism of a connected manifold which is the identity on an open set is the identity everywhere.

The key local point is that some point has no neighborhood fixed by an open subgroup. Otherwise the interiors of subgroup fixed sets become both open and closed, by applying Newman’s theorem to a local finite quotient action. Connectedness would contradict faithfulness. A small subgroup then keeps a chart neighborhood inside the chart. Replacing it by its subgroup of squares preserves faithfulness and makes the action orientation preserving; for $p=2$ this can require index two.

Fix an odd integer $d>n$, with $d\ge 3$, and put

$$W=O\times {\mathbb{R}}^{d-n}\subset {\mathbb{R}}^d.$$

Let $G$ act trivially on the added coordinates. Write $W^{+}$ for its one-point compactification and $Y=W^{+}/G$. The collapse map $c:S^d\to W^{+}$ sends the complement of $W$ to infinity; $q:W^{+}\to Y$ is the orbit map.

After shrinking $G$ once more, there is a continuous $a:Y\to S^d$ with

$$\mathrm{deg}(aqc)=1.$$

To construct it, start with a degree-one pinch supported in a ball compactly contained in $W$. A sufficiently small subgroup moves every value of this sphere-valued map by less than $1/2$. Average in ${\mathbb{R}}^{d+1}$ over that subgroup and normalize. The average remains nonzero, and normalized linear interpolation gives a homotopy to the original pinch. The normalized average is invariant and therefore descends to $Y$.

🏷️ Forms, sheaves, and the fixed lattice

A sheaf records local data and its restriction to smaller open sets. A derived direct image $R{f}_{\ast }{\mathbb{R}}_B$ also records the cohomology of fibers. The allowed test spaces are finite polyhedra, spaces homeomorphic to them, open subsets of those spaces, and finite products of these; points and the empty space are included. The manuscript uses the category $\mathcal{T}(X)$ generated by these images for compact finite polyhedra $B$ and arbitrary continuous maps $f:B\to X$.

“Generated” allows finite sums, shifts, mapping cones, and direct summands. The stalks need not be finite dimensional or constructible with respect to a finite stratification. Finiteness is imposed on the source and on the number of category operations.

Verdier duality $D_X$ is the sheaf-theoretic dual adapted to the local topology. With shifted duality $[-r]D_X$, a nonsingular symmetric form is an object identified with this shifted dual by a symmetric pairing. Its Witt class identifies orthogonal sums additively and sets neutral forms, those possessing the appropriate lagrangian, equal to zero. Write

$$E_r(X)=W_r(\mathcal{T}(X)).$$

The manuscript proves that these groups are covariant and homotopy invariant. All sheaves have real coefficients. Calling a Witt class integral means a class in this abelian group before tensoring with $\mathbb{Q}$; it does not mean that the sheaves have integer coefficients.

Fixed sphere lattice

Let $u_d$ be the orientation class of an odd sphere. There is a power of two $L_d$, depending only on $d$, such that

$$\begin{array}{c}{E}_d(S^d)\otimes \mathbb{Q}=\mathbb{Q}{\bar{u}}_d,\\ \mathrm{im}(E_d(S^d)\to E_d(S^d)\otimes \mathbb{Q})\\ \subseteq L_d^{-1}\mathbb{Z}{\bar{u}}_d.\end{array}$$

A degree-$s$ sphere map acts on this rational line by multiplication by $s$. (OpenAI, 2026)

The proof compares arbitrary continuous source maps with semialgebraic ones. At a point the Witt groups are $\mathbb{Z}$ in degrees divisible by four and zero otherwise, by the signature calculation for real symmetric forms. Open-cover localization and a doubling theorem for dense subcategories propagate a bounded power-of-two kernel and cokernel across the sphere dimension.

The exponent is uniform over all classes. The manuscript’s unoptimized bookkeeping permits

$$\begin{array}{c}{e}_0=0,e_j=9e_{j-1}+6,\\ L_d=2^{e_d}.\end{array}$$

A rational isomorphism alone would not give this: $\mathbb{Z}\subset \mathbb{Z}[1/2]$ becomes an isomorphism over $\mathbb{Q}$, but its denominators are unbounded. The complete duality, germ-localization, and dense-doubling proofs are inputs to this post; the fixed-lattice deduction from their stated interfaces is the portion explained here.

🏷️ Character partitions without splitting their images

The characters of $G$ take values in

$$\Lambda =\bigcup _{j\ge 0}{\mu }_{p^j},$$

where ${\mu }_m$ denotes the $m$th roots of unity. Locally constant functions on an orbit factor through finite quotients, so their complex-valued sheaf decomposes into character eigensheaves. Complex coefficients are regarded as a real plane with positive pairing $B(z,w)=\mathrm{Re}(z\bar{w})$.

For a compact test map $t:Y\to X$, let $b=t(q(\infty ))$ and assume $b$ has a basis of contractible open neighborhoods. Push the oriented real-plane form on $S^d$ along $tqc$ to obtain $(A_t,{\alpha }_t)$.

A finite partition $\Lambda =D_0\bigsqcup \cdots \bigsqcup D_{m-1}$ gives orthogonal self-adjoint projectors $e_i$ away from $b$. These projectors need not extend as global direct summands. Instead choose a small contractible neighborhood $U$ of $b$ and use the quotient category

$${\mathcal{Q}}_U=\mathcal{T}(X)/\mathcal{T}(U).$$

Here $\mathcal{T}(U)$ is identified with objects supported compactly inside $U$. A germ-localization theorem puts the projectors in this quotient. A separate integral localization theorem gives an isomorphism

$${\ell }_U:E_d(X)⟶W_d({\mathcal{Q}}_U).$$

Oddness makes $E_d(U)=0$. The next map $E_{d-1}(U)\to E_{d-1}(X)$ is split injective by projection to a point, which completes the exact-sequence argument. One must not assume that this adjacent point group always vanishes.

Define the character class $z_i(t)$ by

$$\begin{array}{rl}{\ell }_U(z_i(t))=& [A_t,{\alpha }_t]\\ & +[A_t,{\alpha }_t(2e_i-1)].\end{array}$$

The operator $2e_i-1$ is a self-adjoint involution, so the second form is nonsingular. If the projector had an image object, this construction would double its form and cancel the complementary form. The definition needs only the endomorphism, and the integral localization isomorphism makes the lifted class independent of $U$.

The manuscript proves the integral sum identity

$$\sum _{i=0}^{m-1}{z}_i(t)=2[A_t,{\alpha }_t].$$

Its useful algebraic mechanism is explicit. On $A_t^{\oplus m}$ let $E_j$ apply $e_j$ to every coordinate and let $P_j$ swap coordinates zero and $j$. Then $S={\sum }_j{E}_j{P}_j$ is a self-adjoint involution. It conjugates $\mathrm{diag}(1,-1,\dots ,-1)$ to $\mathrm{diag}(2e_i-1{)}_i$. Thus the twisted forms sum to $(2-m)[A_t,{\alpha }_t]$; adding the $m$ untwisted forms gives the displayed identity.

🏷️ Finite quotient sheets and concentration

For any $k\ge 1$, set $m=p^k$. Faithfulness supplies a point whose stabilizer lies in $p^{k}G$: otherwise $p^{k-1}G$ would fix every point. The finite quotient action therefore has a free point, and disjoint neighborhoods of its finite orbit give a nonempty open $V\subset Y\setminus \{q(\infty )\}$ over which

$$W/(p^{k}G)⟶W/G$$

has exactly $m$ separate homeomorphic sheets. This is a local assertion about the finite quotient; the original action need not be free.

Label these sheets by powers of $\zeta =\exp (2\pi i/m)$. The resulting locally constant unit-modulus function $\beta$ shifts characters by a generator of ${\mu }_m$. Choose $D_0$ as one representative in each multiplication orbit of ${\mu }_m$ on $\Lambda$, and put $D_i={\zeta }^i{D}_0$. For a test $h:Y\to S^d$ constant outside $V$, multiplication by $\beta$ is a real isometry cyclically permuting the character projectors. Hence all $z_i(h)$ are equal integrally.

The fixed degree-one test $a$ is not initially concentrated in $V$. The manuscript constructs a positive-degree map $D_s:S^d\to S^d$ and a concentrated $h$ with

$$h\simeq D_{s}a.$$

The degree $s$ may depend on $k$. First move the original pinch inside $q^{-1}(V)$. Rational orbit pullback is injective: zero-dimensional compact orbits have no higher sheaf cohomology, and averaging locally constant rational functions gives a rational sheaf retraction. Thus $a$ is rationally trivial on $Y\setminus V$.

For an odd sphere, Serre’s homotopy results allow a rationally trivial map from a compact metrizable space to be killed by some positive-degree postcomposition. The proof first passes to a finite nerve; it does not impose a dimension bound on $Y$. Extending the resulting nullhomotopy from the closed complement gives $h$.

🏷️ The denominator contradiction

Covariance and homotopy invariance identify $z_i(h)$ with $(D_s{)}_{\ast }{z}_i(a)$. On the rational sphere line this is $s{\bar{z}}_i(a)$. Since $s>0$ and the left-hand classes are equal, all ${\bar{z}}_i(a)$ are equal. This cancellation takes place over $\mathbb{Q}$, with no assumption that $s$ is coprime to $p$.

The real coefficient plane has signature two, and $\mathrm{deg}(aqc)=1$. The sum identity therefore gives

$$\sum _{i=0}^{m-1}{\bar{z}}_i(a)=4{\bar{u}}_d,{\bar{z}}_i(a)=\frac{4}{m}{\bar{u}}_d.$$

Each $z_i(a)$ was already an integral class, so the fixed lattice forces $4L_d/m$ to be an integer. Choose $k$ so large that $m=p^k>4L_d$. Then

$$0<\frac{4L_d}{p^k}<1,$$

the contradiction. It applies also to $p=2$: bounded denominators, rather than merely inverting two, are essential.

The logic fixes $d$ and its lattice first, then varies $k$, its sheet neighborhood, and the concentrating degree. Verdier duality is used on actual manifold sources and finite-dimensional test spaces. The orbit space enters through ordinary direct images and rational cohomology, without an asserted dimension bound.

Links

• Family 302 also passes through finite polyhedral models while retaining support conditions; here finite nerves avoid imposing a dimension bound on an orbit space.
• Collection guide places this family in the topology reading sequence and distinguishes manuscript claims from the selected checks performed for these posts.

The selected-source review covers the introduction, local-test construction, character-class construction and final contradiction, plus selected category, localization, and lattice passages. The complete categorical foundations, sign appendix, semialgebraic signature argument, and cited external theorems remain unaudited inputs. No independent proof of Hilbert–Smith is claimed by this exposition.

🏷️ Key steps and proof outlines

A quantitative exact-sequence argument fixes all denominators at once

The fixed-lattice theorem is stronger than a rational calculation. Its proof compares two versions of the invariant: $F_r$, built from semialgebraic source maps, and the larger $E_r$, allowing the continuous source maps needed by the action. The useful induction statement is that one exponent $e_j$ kills both kernel and cokernel of

$$F_r(S^j)⟶E_r(S^j)$$

for every degree $r$ and every class. Uniformity in the class is what later permits arbitrarily many character pieces.

The elementary engine is a quantitative five-lemma argument. In a commutative diagram of five-term exact sequences, suppose multiplication by $q$ kills the kernels and cokernels of the four outer vertical maps. Then $q^3$ kills those of the middle map. For example, a middle kernel element $x$ first satisfies that $qx$ lifts from the preceding group. A second multiplication allows its image to lift from the first group; a third kills the remaining discrepancy. Applying the horizontal map gives $q^{3}x=0$. The cokernel argument reverses the lifting procedure. Keeping these multipliers records information lost by tensoring with $\mathbb{Q}$ at the start.

Cover $S^j$ by the complements $U,V$ of two poles. Both are contractible, and $U\cap V$ has the homotopy type of $S^{j-1}$. The manuscript’s localization and dense-doubling theorems then give three successive comparisons:

• The localization sequence for $(V,U\cap V)$ changes the induction exponent $e$ to $3e$.
• Passing to the quotient associated with $(S^j,U)$ costs another factor $4$, giving $3e+2$.
• The final localization sequence changes this to $9e+6$.

Starting with the point calculation gives $e_0=0$ and $e_j=9e_{j-1}+6$. Put $L_d=2^{e_d}$. Every class $x\in E_d(S^d)$ then has $L_{d}x$ in the image of $F_d(S^d)$. The semialgebraic signature calculation says that this latter class has an integral orientation coefficient. Therefore $x$ has denominator dividing $L_d$, regardless of how $x$ was constructed. This outlines the induction after the manuscript’s substantial localization and duality inputs. (OpenAI, 2026)

Removing one neighborhood preserves the integral class

The character projectors are initially available only away from the exceptional point. It matters that localization lifts their associated forms integrally, not merely rationally. For a nonempty contractible open set $U\subset X$, the relevant sequence is

$$\begin{array}{rl}0⟶E_d(X)& ⟶W_d(\mathcal{T}(X)/\mathcal{T}(U))\\ & ⟶E_{d-1}(U)⟶E_{d-1}(X).\end{array}$$

The initial zero comes from oddness of $d$ and the point calculation. The last map is injective for a different reason: compose inclusion with projection $X\to \mathrm{pt}$. On $E_{d-1}(U)$ this composite is the isomorphism induced by the contraction of $U$. It gives a left inverse to inclusion. Exactness now makes the middle localization map an integral isomorphism. This also covers $d\equiv 1(\mathrm{mod}4)$, when the adjacent point group is $\mathbb{Z}$ rather than zero.

Equal pieces are compared before their concentrating degree is canceled

The last contradiction has a delicate order of choices. First fix the odd dimension $d$, its lattice denominator $L_d$, and the degree-one test $a$. Only then choose the character partition size $m=p^k$. A sheet neighborhood for that partition gives equality of the concentrated character classes $z_i(h)$. Concentration may require a positive degree $s$ depending on $k$, so it would be incorrect to require a common degree for every partition or a degree prime to $p$.

Naturality gives

$${\overline{z}}_i(h)=s{\overline{z}}_i(a)\text{in }{E}_d(S^d)\otimes \mathbb{Q}.$$

Since $s\ne 0$, equality can be canceled on this rational line. The original $z_i(a)$ are already integral classes. Their integral sum identity, together with the real-plane signature two, gives rational coefficient $4/m$ for each one. The fixed lattice therefore requires $4L_d/m\in \mathbb{Z}$. Taking $m>4L_d$ contradicts positivity and integrality. Rational cancellation is used only to prove equality; the denominator bound survives because integrality was established earlier.

🏷️ Manuscripts and verification scope

• The Hilbert–Smith conjecture in every finite dimension — pinned manuscript and companion source.

No family scope page was present at lean/docs/304.md in the pinned snapshot. This does not establish that all related formalization code is absent. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. The Hilbert–Smith conjecture in every finite dimension.