The four-dimensional Singer conjecture: finite energy and boundary paths

Source and verification

This article explains family 315 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.

The four-dimensional Singer conjecture predicts that square-integrable homology of the universal cover of an aspherical closed four-manifold lives only in degree two. Family 315 claims this conclusion without a curvature, smoothness, orientability, or residual-finiteness assumption. Its argument turns a hypothetical positive first $L^2$-Betti number into a boundary continuum with incompatible properties. (OpenAI, 2026)

🏷️ Preliminaries: asphericity, duality, and $L^2$ dimension

A connected space is aspherical when its universal cover is contractible. It can still have nontrivial homology downstairs: a torus is the simplest example. An integral Poincaré complex of formal dimension four is a finite connected CW complex with a fundamental class and cap-product duality for every local coefficient system. Its orientation character records whether loops reverse orientation. The duality here is absolute, not relative duality for a pair.

Let $\Gamma$ be the deck group of the universal cover. Completing cellular chains in a square-summable norm produces Hilbert spaces with a $\Gamma$-action. Reduced $L^2$ homology takes cycles modulo the closure of boundaries. The group von Neumann algebra $\mathcal{N}\Gamma$ supplies a dimension normalized per fundamental domain; the dimension of ${\ell }^2(\Gamma )$ is one. The resulting nonnegative real numbers are the $L^2$-Betti numbers $b_j^{(2)}$.

For finite CW complexes this agrees with the algebraic definition using extended von Neumann dimension after tensoring chains with $\mathcal{N}\Gamma$. The algebraic definition also applies to singular chains of topological manifolds. It is not ordinary homology of the contractible universal cover: the completion and the requirement of square summability change the chain problem.

Manuscript's four-dimensional concentration theorem

If $Q$ is a finite connected aspherical integral Poincaré complex of formal dimension four, with any orientation character, then

$$b_j^{(2)}(\tilde{Q})=0(j\ne 2).$$

In particular this holds for every closed connected aspherical topological four-manifold. (OpenAI, 2026)

For a manifold without a triangulation, the source passes to a finite four-dimensional homotopy model; it does not claim to triangulate the manifold. Local-coefficient duality and $L^2$ invariants transport across this homotopy equivalence.

🏷️ Reduction to first-degree vanishing

Asphericity and a finite-dimensional model give a finite-length free resolution of the trivial $\mathbb{Z}\Gamma$-module. A subgroup of prime order would contradict this by its nonzero cohomology in arbitrarily high even degrees, so $\Gamma$ is torsion-free. It is infinite: otherwise it would be trivial and $Q$ contractible, contradicting its mod-two fundamental class.

A square-summable harmonic function in degree zero is constant on the connected cover and therefore zero. Poincaré duality gives

$$\begin{array}{c}{b}_0^{(2)}=b_4^{(2)}=0,\\ b_3^{(2)}=b_1^{(2)}.\end{array}$$

The finite four-dimensional model has no higher chains. For a nonorientable complex, apply duality on its orientation cover and use finite-index scaling of von Neumann dimension. Thus only $b_1^{(2)}=0$ remains.

The same duality, now over $k={\mathbb{F}}_2$, supplies

$$H^1(\Gamma ;k\Gamma )=H^2(\Gamma ;k\Gamma )=0.$$

With a permutation coefficient module, homological Shapiro gives $H^1(\Gamma ;k[\Gamma /J])\cong H_3(J;k)$. This vanishes for finite or two-ended subgroups in this torsion-free setting. These inputs control ends and obstruct certain tree actions later in the proof.

🏷️ Finite-energy functions and a compact boundary

Suppose $b_1^{(2)}>0$. Choose a finite presentation and its locally finite Cayley graph $G$. For a function $f$ on vertices, its Dirichlet energy is

$$\mathcal{E}(f)=\sum _{e=uv}|f(v)-f(u){|}^2,$$

counting each cellular edge once. Integrating a cocycle in the simply connected presentation complex shows that square-summable one-cocycles are finite-energy gradients. The positive Betti number means some such gradient is outside the closure of gradients of finitely supported functions. It does not require that the image of the coboundary operator be closed.

The source obtains a bounded finite-energy function with nonconstant behavior at infinity. Its translates, singleton indicators, and additional component-recording functions form a countable invariant family $\{F_j\}$, with every coordinate normalized to take values in $[0,1]$. Taking the closure of vertex profiles $(F_j(v){)}_j$ gives a compact metrizable space $X=\Gamma \bigsqcup B$.

Choose positive weights with ${\sum }_j{c}_j(1+\Vert d{F}_j{\Vert }_2)<\infty$. The metric and its edge lengths satisfy

$$\begin{array}{c}\rho (x,y)=\sum _j{c}_j|F_j(x)-F_j(y)|,\\ \lambda (e)=\rho (e^{-},e^{+}),\\ \sum _e\lambda (e{)}^2<\infty .\end{array}$$

The last inequality follows from Minkowski’s inequality and the chosen weights. In particular, only finitely many edges have length at least any prescribed positive number.

Singleton coordinates isolate every group vertex. Nonconstant behavior at infinity prevents $B$ from being a singleton. One-endedness makes $B$ connected: a separation of the boundary would leave two infinite vertex regions connected by only finitely many long edges, producing more than one end after a finite deletion.

The additional coordinates are crucial. For a coordinate box $O$ with finitely many graph components $U_j$, they satisfy ${\sum }_j{h}_{O,U_j}={\psi }_O$ and have disjoint nonzero supports. A positive margin at a boundary point selects exactly one component. Consequently, vertices approaching the same boundary point can be joined by paths whose vertices stay uniformly near that point. This is a graph-path statement; it is not yet local connectivity of $B$.

🏷️ Dimension, cohomology, and the cut-point reduction

Finite energy also controls translates of coordinate level cuts. The manuscript proves that the action on $B$ is a convergence action: any sequence of distinct group elements has a subsequence collapsing uniformly on compact sets away from one exceptional point. It separately proves

$$\dim B\le 1,{\check{H}}^1(B;{\mathbb{F}}_2)=0.$$

The dimension is covering dimension: finite open covers admit refinements with at most two sets through a point. The relevant cohomology is Čech cohomology, computed from nerves of covers, suitable for a boundary not yet known to be locally connected.

For the dimension bound, almost every coordinate cut has a boundary accumulation set of one-dimensional Hausdorff measure zero. A finite grid of such cuts has a small clopen partition on its accumulation set. Off that set, the near-path property makes cut signs locally constant. Small neighborhoods of the clopen pieces, together with the sign regions, give arbitrarily fine covers of multiplicity at most two.

For Čech vanishing, transfer a nerve cocycle to an edge cochain $\alpha$ on the presentation complex. Its coboundary has finite support: far-out triangular cell boundaries shrink toward single boundary points, where the nerve cocycle identity applies. The vanishing of $H^2(\Gamma ;k\Gamma )$ removes this finite coboundary by a finitely supported correction. The remaining cocycle is a gradient. Its values near the boundary give the coboundary of the original nerve class after refinement. This does not replace Čech cohomology by singular cohomology on a possibly wild continuum.

A cut point disconnects a space when removed. If $B$ has cut points, a pretree records which cut points separate which others. The manuscript combines a cut-point theorem with an additional reduction for parabolic cut points to select a nondegenerate block $L$ without cut points. It constructs a continuous retraction $\mathfrak{r}:B\to L$, locally constant off $L$.

The retraction matters: inclusion followed by retraction is the identity on $L$, so ${\check{H}}^1(L;{\mathbb{F}}_2)$ injects into the zero group ${\check{H}}^1(B;{\mathbb{F}}_2)$. Closed-subspace monotonicity gives $\dim L\le 1$. Continuity is proved using nearby graph paths and the closure of the projection’s graph, before local connectivity is available.

The resulting auxiliary graph $K$ has finitely many vertex and edge orbits under a subgroup $H$, but can have infinite valence and infinite stabilizers. Its vertices may themselves lie in $L$. These possibilities are retained in the next argument.

🏷️ Annular depth and a threshold independent of the path

An annulus is an ordered pair of disjoint closed sides. Nested annuli separate a closed set $A\subset Y=H\bigsqcup L$ from a vertex; let $D(A,g)$ be the greatest chain length for a group vertex $g$. Choose $A$ containing the identity. Depth is finite at each group vertex and tends arbitrarily high toward boundary points outside $A$.

For an auxiliary vertex $u$, let $C(u)$ be all group elements representing it from a chosen orbit representative. Define deep vertices by

$$\begin{array}{c}{\delta }_A(u)=\underset{g\in C(u)}{\min }D(A,g),\\ I_m=\{u\notin A:{\delta }_A(u)\ge m\}.\end{array}$$

The minimum imposes the depth bound on every representative. This permits moves through an infinite stabilizer without losing depth. Boundary points outside $A$ have locally constant labels $Q_m$, specifying the component of the induced graph $K[I_m]$ approached by nearby vertices.

For a deep sequence $g_n$, translated sets $g_n^{-1}A$ shrink to one boundary point $q$ after a subsequence. A uniform annular comparison shows that fixed boundary points away from $q$ acquire the same deeper component label under $g_n$.

Choose three fixed anchors $Z\subset L$. At least two avoid the one exceptional point. Their translated labels therefore define a unique majority label $J_m(g)$ at sufficiently large depth. Fix the finite list of moves coming from edge-orbit representatives, together with one contracting stabilizer move for each vertex orbit when needed. A contradiction-by-subsequence argument gives one threshold $M$ such that, for every $m\ge M$, these moves preserve the majority label whenever both representatives have depth at least $m$.

The threshold depends on the neighborhood, anchors, and finite move list. It does not depend on a path’s endpoints or length. Stabilizer powers bring the anchors toward their fixed vertex, identifying its boundary label with the majority label. Edge moves then show

$$\begin{array}{c}{Q}_m(x)=Q_m(y)\\ ⟹Q_{m+1}(x)=Q_{m+1}(y)\\ (m\ge M).\end{array}$$

Thus one small boundary neighborhood retains a common component label at every greater depth. A threshold for each individual path would not provide this conclusion. (OpenAI, 2026)

🏷️ From paths to small connected boundary sets

Take two nearby boundary points with this common label. For each depth $m$, approximate them by vertices of $I_m$ and join those vertices within their component. All path vertices stay in the prescribed neighborhood.

The metric distances between adjacent vertices tend uniformly to zero. There are only finitely many edge orbits, and translates of each representative edge shrink outside a finite set of group elements. Endpoints of the finitely many exceptional edges are excluded at large depth. This proves vanishing mesh without local finiteness of $K$.

Compactness gives a Hausdorff limit of the finite path-vertex sets. The limit contains both chosen points and lies entirely in $L$. It is connected: a separation into compact pieces at positive distance would force some approximating path to make an edge jump of that distance, contradicting vanishing mesh.

Every nearby boundary point can therefore be connected to the chosen point by a small continuum. Repeating this in each open set shows that its components are open. This is local connectivity of $L$.

🏷️ The continuum contradiction and its consequences

For a compact metric space of dimension at most one, restriction in Čech degree one to a closed subset is surjective. To see this, refine a cover to multiplicity at most two. Its nerve is a graph, and a one-cocycle on the subgraph associated to the closed subset extends by zero on the remaining edges; there are no two-simplices imposing further conditions.

Consequently $L$ cannot contain an embedded circle: the circle would be closed and have nonzero ${\check{H}}^1(-;{\mathbb{F}}_2)$, contradicting the surjection from ${\check{H}}^1(L;{\mathbb{F}}_2)=0$. A nondegenerate locally connected compact metric continuum with no embedded circle is a dendrite and has a cut point. Concretely, an interior point of an arc between two points must separate them; otherwise a second arc avoiding it would create a circle. This contradicts the construction of $L$ without cut points and forces $b_1^{(2)}=0$.

The claimed concentration yields $\chi (Q)=b_2^{(2)}(\tilde{Q})\ge 0$. The middle invariant need not vanish: a product of genus-$g$ and genus-$h$ closed surfaces, with $g,h\ge 2$, has Euler characteristic $4(g-1)(h-1)>0$, while the four-torus has zero. For oriented closed topological manifolds, the additional topological $L^2$-signature theorem gives $\chi (M)\ge |\sigma (M)|$.

If the deck group has a descending normal tower of finite-index subgroups with trivial intersection, the additional $L^2$ approximation theorem identifies normalized rational Betti-number limits with $\chi (Q)$ in degree two and zero elsewhere. This tower hypothesis is required for that consequence, not for the concentration theorem.

The selected-source review covers the duality reduction, boundary interfaces, retraction, uniform path threshold, continuum contradiction, and consequences. The full finite-energy potential construction, convergence argument, pretree reduction, and all annular estimates have not been independently reconstructed here.

🏷️ Related posts

• Quillen homology provides another instance where precise chain-level information is needed before passing to a homological conclusion; its finite posets differ from the completed chain spaces here.

🏷️ Key steps and proof outlines

Three anchor points turn sequential collapse into a uniform path estimate

The final geometric step needs a single threshold that works for every deep graph path. A statement about one convergent sequence of group elements would not suffice: the path and its endpoints change as the depth increases.

In the manuscript’s annular construction, $D(A,g)$ measures the separation depth of a translate from a fixed excluded region $A$. The label $Q_m(x)$ records which component of the depth-$m$ vertex graph approaches a boundary point $x$. Earlier lemmas show that, along a sequence of increasingly deep translates, all fixed boundary points except possibly one acquire the same eventual component label.

Fix three distinct anchor points $Z\subset L$. For a sufficiently deep $g$, at least two points of $gZ$ share one depth-$(m+1)$ label; define it as $J_m(g)$. It is unique because two distinct labels cannot each occur twice among three points.

The threshold can be made uniform. If no common threshold existed, choose violating depths $m_n\to \infty$ and translates $g_n$. Pass to the convergent subsequence supplied by the earlier dynamical lemma, with one exceptional point $q$. Two of the three fixed anchors avoid $q$, so their images eventually share a label, contradicting the violation.

The same compactness argument applies to finitely many allowed edge moves $s$. Choose two anchors in $Z$ and two in $sZ$ avoiding $q$; their four images eventually have a common label. Thus, for one threshold covering the entire finite list,

$$J_m(g)=J_m(gs)$$

whenever both endpoints are deep enough. Finitely many edge orbits supply precisely such a finite move list. In the boundary-orbit version of the graph, the list also includes stabilizing elements whose iterates move the anchors toward the represented vertex. This identifies the majority label with the component label at that vertex. (OpenAI, 2026)

Stabilizing labels pushes connectivity to arbitrarily fine depth

Constancy of $J_m$ along edges implies the refinement statement

$$Q_m(x)=Q_m(y)⟹Q_{m+1}(x)=Q_{m+1}(y)(m\ge M),$$

with $M$ independent of $x,y$. Fix a point $x$ and a prescribed neighborhood $U$. Choose an ambient neighborhood $O$ whose closure meets $L$ only inside $U$. Local constancy of $Q_M$ provides a smaller neighborhood $V$ of $x$ on which all points have the same label. The refinement implication then keeps them in the same component at every deeper level.

For each $y\in V$, take graph paths $P_m$ inside $O$ with endpoints tending to $x,y$. The maximum edge length of these paths tends to zero. To prove this uniformly, use the finite list of edge representatives: all their far translates have small length, while sufficiently deep vertex sets exclude the finitely many endpoints belonging to the remaining translates.

Regard each path as its finite vertex set and take a Hausdorff-convergent subsequence in the compact ambient space. The limit $C$ contains $x,y$ and lies in $\overline{O}\cap L\subset U$. It is connected: if it split into two compact pieces a positive distance apart, every sufficiently fine approximating path meeting both would need an edge spanning that distance, contradicting its vanishing mesh.

Thus every point of $V$ lies in the component of $x$ inside $U$. Applying this neighborhood property at each point shows that components of open sets are open, which is local connectedness. This passage from discrete paths to continua is the reason the uniform threshold matters.

Dimension one and vanishing Čech cohomology force a cut point

The concluding topological lemma also has a concrete proof. In a compact metric space $L$ of covering dimension at most one, every finite open cover has a refinement of multiplicity at most two. Its nerve is a graph. For a closed subset $C\subset L$, represent a class in ${\check{H}}^1(C;{\mathbb{F}}_2)$ on a cover and refine after extending that cover to $L$.

Retain the cover indices even when two restricted sets coincide. The restricted nerve is then a subgraph of the full nerve. Extend its one-cocycle by zero on the remaining edges. Since the full nerve has no two-simplices, this remains a cocycle. Therefore restriction

$${\check{H}}^1(L;{\mathbb{F}}_2)⟶{\check{H}}^1(C;{\mathbb{F}}_2)$$

is onto. If the left side vanishes, $L$ cannot contain an embedded circle, whose first cohomology is nonzero.

Local connectedness supplies arcs inside connected open regions. Choose an arc between two distinct points and an interior point $p$. If $p$ were not a cut point, the connected open set $L\setminus \{p\}$ would contain another arc between the same endpoints. A component of the first arc outside the second, together with the corresponding segment of the second, forms an embedded circle. This is impossible. Hence $L$ has a cut point, contradicting the cut-free block constructed from positive first $L^2$-Betti number. The finite-energy, duality, and cut-point-reduction inputs remain separate substantive parts of the source proof.

🏷️ Manuscripts and verification scope

• The Singer conjecture in dimension four — pinned manuscript and companion source.

No family scope page was present at lean/docs/315.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 Singer conjecture in dimension four.