Wall’s finite D(2) problem: cyclic towers and twisted homology

Source and verification

This article explains family 321 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.

Wall’s finite $D(2)$ problem asks whether a finite complex that looks two-dimensional to suitable homological tests must have a finite two-dimensional homotopy model. This manuscript claims a counterexample with a finite three-dimensional cell structure and infinite fundamental group. Its proof combines an explicit twisted boundary matrix with a geometric obstruction obtained by repeatedly lifting maps through infinite cyclic covers. (OpenAI, 2026)

🏷️ Preliminaries: cells, covers and local coefficients

A finite CW complex is built from finitely many cells by attaching their boundary spheres to earlier stages. Its dimension is the largest cell dimension. A homotopy equivalence can change the cell structure substantially, so the existence of three-cells in one presentation does not itself obstruct a two-dimensional model.

Let $G={\pi }_1(X)$ and let $\tilde{X}$ be the universal cover. Its cellular chains are modules over the group ring $\mathbb{Z}G$, with the action given by deck transformations. A local coefficient system allows coefficients to change by parallel transport around loops. A left $\mathbb{Z}G$-module $M$ gives cellular cochains ${\mathrm{Hom}}_{\mathbb{Z}G}(C_{\ast }(\tilde{X}),M)$.

The $D(2)$ condition asks for

$$\begin{array}{c}{H}_i(\tilde{X};\mathbb{Z})=0(i>2),\\ H^3(X;M)=0\end{array}$$

for every finitely generated $\mathbb{Z}G$-module $M$. The finite problem asks whether a finite connected complex satisfying these conditions is homotopy equivalent to a finite complex of dimension at most two. Asphericity is not part of the question.

Manuscript's finite $D(2)$ counterexample

There is a finite connected three-dimensional CW complex $X$ satisfying the displayed vanishing, with $H^3(X;M)=0$ in fact for every coefficient module $M$, but with no finite two-dimensional homotopy model. Its fundamental group is infinite. (OpenAI, 2026)

The finite qualification matters on both sides. The argument below does not establish that every possible infinite two-dimensional model is excluded. It also does not contradict stable realization results obtained after adjoining sufficiently many two-spheres.

🏷️ A rank-one character that detects torsion

A character $\rho :G\to {\mathbb{C}}^{\times }$ defines a rank-one complex local system $L_{\rho }$. Operationally, a vector transported around a loop representing $g$ is multiplied by $\rho (g)$. With the right action $v\cdot g=\rho (g)v$, its homology is computed by

$${\mathbb{C}}_{\rho }{\otimes }_{\mathbb{Z}G}{C}_{\ast }(\tilde{X}).$$

The central obstruction is the following statement.

Manuscript's two-dimensional character obstruction

If $Y$ is a finite connected complex of dimension at most two and $H_2(Y;L_{\rho })=0$, then $\rho (g)=1$ for every finite-order element $g\in {\pi }_1(Y)$. (OpenAI, 2026)

The source proves more: if a finite connected simplicial complex $F$ has $b_1(F;\mathbb{Q})=0$, every map $F\to Y$ pulls $L_{\rho }$ back to a system with trivial monodromy. Here the first rational Betti number is the dimension of $H_1(F;\mathbb{Q})$.

We first explain why the vanishing hypothesis survives the two operations in its proof.

🏷️ Why twisted second homology survives a cyclic cover

For a subcomplex $A\subseteq Y$ in dimension at most two, $H_2(A;L{|}_A)\to H_2(Y;L)$ is injective. Both groups are kernels of the second cellular boundary, since there are no three-cells. This needs no injectivity of fundamental groups.

Now take a connected infinite cyclic cover $\hat{A}\to A$ with $A$ finite and two-dimensional. Give the pulled-back local system its canonical deck action: a deck transformation changes the point upstairs while leaving its coefficient in the downstairs fiber fixed. With $R=\mathbb{C}[z,z^{-1}]$, the cellular chain groups upstairs are finite free $R$-modules. Setting $z=1$ recovers the downstairs twisted cellular complex.

Let $D(z)$ be the second boundary matrix and let it have $m$ columns. If $H_2(A;L)=0$, then $D(1)$ is injective. For $m>0$ it has a nonzero $m\times m$ minor. The same minor of $D(z)$ is a nonzero Laurent polynomial, so $D(z)$ has full column rank over $\mathbb{C}(z)$ and hence is injective over $R$. Thus

$$H_2(\hat{A};p^{\ast }L)=0.$$

This is ordinary homology with finite-support chains; no completed or square-summable chain complex is involved. (OpenAI, 2026)

🏷️ The tower must grow inside a fixed finite domain

Suppose $f:F\to Y$ has nontrivial pullback monodromy and $b_1(F;\mathbb{Q})=0$. Replace $Y$ by a finite simplicial model of dimension at most two and make $f$ simplicial after a sufficiently fine initial subdivision of $F$. Fix this triangulation for the remainder of the argument. Let $A_0=f(F)$.

Inductively suppose $f_i:F\to A_i$ is onto its finite simplicial image, twisted second homology vanishes on $A_i$, and the pullback monodromy remains nontrivial. Nontrivial rank-one monodromy gives $H_0(A_i;L_i)=0$. Euler characteristic therefore yields

$$\begin{array}{rl}\chi (A_i)& =-\dim H_1(A_i;L_i)\le 0,\\ \chi (A_i)& =1-b_1(A_i;\mathbb{Q})+b_2(A_i;\mathbb{Q}).\end{array}$$

Thus $b_1(A_i;\mathbb{Q})\ge 1$, so ${\pi }_1(A_i)$ surjects onto $\mathbb{Z}$. The map from $F$ lifts to the associated infinite cyclic cover: every homomorphism ${\pi }_1(F)\to \mathbb{Z}$ factors through the finite group $H_1(F;\mathbb{Z})$ and is zero.

Let $A_{i+1}$ be the image of that lift. The cyclic-cover lemma and subcomplex injection preserve twisted second-homology vanishing; the pullback system on $F$ is unchanged. The lift remains simplicial on the same fixed domain triangulation.

Let $s(A)$ count nonempty simplices. The covering projection maps $A_{i+1}$ onto $A_i$ and maps every simplex isomorphically onto a simplex. Hence $s(A_{i+1})\ge s(A_i)$. Equality would give exactly one compatible lift of each simplex, hence a section of the connected nontrivial cyclic cover. That would make the covering subgroup all of ${\pi }_1(A_i)$, a contradiction. Consequently

$$\begin{array}{c}s(A_0)<s(A_1)<s(A_2)<\cdots ,\\ s(A_i)\le s(F)\text{for every }i.\end{array}$$

which is impossible. This proves the tower obstruction.

For a finite-order element $g$ of order $n$, a loop representing $g$ extends to the Moore complex $S^1{\cup }_{\mathrm{deg}n}{D}^2$, whose first rational Betti number is zero. If $\rho (g)\ne 1$, this map violates the tower obstruction. The order-two case uses ${\mathbb{RP}}^2$. (OpenAI, 2026)

🏷️ The finite plus construction supplies $D(2)$

Let $B$ be a finite two-dimensional complex with fundamental group $P$, and let $K$ be the normal closure of $r$ elements in $P$. Assume $K$ is perfect, meaning $K/[K,K]=0$, and put $G=P/K$. Normal finite generation does not require $K$ to be finitely generated as an abstract group.

Attach $r$ two-cells to kill those normal generators, obtaining $B^{\prime }$ with fundamental group $G$. In the universal cover of $B^{\prime }$, the preimage $\overline{B}$ has fundamental group $K$, so $H_1(\overline{B};\mathbb{Z})=0$. With $R=\mathbb{Z}G$, the relative chain group consists of $R^r$ in degree two. The relative homology sequence therefore gives a surjection

$$H_2(\widetilde{B^{\prime }};\mathbb{Z})\twoheadrightarrow R^r.$$

Choose preimages of its basis. Since the cover is simply connected, the degree-two Hurewicz theorem represents them by spheres. Attach $r$ three-cells along their projections to get $X$. If $q$ projects two-chains onto the newly added two-cells, the choices give the exact identity

$$q{\partial }_3={\mathrm{id}}_{R^r}.$$

Thus ${\partial }_3$ is injective, proving the universal-cover homology condition. For any module $M$ and cochain $\phi :R^r\to M$, the cochain $\phi q$ extends it across ${\partial }_3$. Hence the third cochain boundary is onto and $H^3(X;M)=0$ for every $M$.

Moreover, the relative complex in degrees three and two is $R^r\stackrel{\mathrm{id}}{\to }{R}^r$, which is contractible. The absolute-relative sequence is degreewise split. Tensoring with any right $R$-module preserves its exactness and this contraction, so $B\to X$ preserves all homology with coefficients pulled back from $G$. In particular, it preserves the character test needed above. (OpenAI, 2026)

🏷️ An explicit presentation and perfect kernel

The source takes five generators $x_1,a_1,x_2,a_2,s$ with four relations

$$\begin{array}{rl}{x}_1{a}_1{x}_1^{-1}& =a_1^4,\\ x_2{a}_2{x}_2^{-1}& =a_2^3,\\ x_1& =a_2^{13},\\ s{x}_2{s}^{-1}& =a_1^5.\end{array}$$

Let $B$ be the presentation complex, let $t_1=x_1^2$, $t_2=x_2^3$, and let $K$ be their normal closure. In the quotient $G=P/K$, the map sending $s$ to $1\in \mathbb{Z}$ and the other generators to zero is onto, so $G$ is infinite.

Perfectness is an exact calculation. With $[g,h]=gh{g}^{-1}{h}^{-1}$, the relations give

$$\begin{array}{rl}{t}_1=a_2^{26}& =[t_2,a_2],\\ s{t}_2{s}^{-1}=a_1^{15}& =[t_1,a_1].\end{array}$$

In the abelianization $K_{\mathrm{ab}}$, let $u_i=[t_i]$. Conjugation factors through $G$, giving

$$\begin{array}{rl}{u}_1& =(1-a_2)u_2,\\ s{u}_2& =(1-a_1)u_1.\end{array}$$

The element $a_2$ fixes $u_1$, and $a_2^{26}$ acts trivially on all of $K_{\mathrm{ab}}$. Multiplying the first relation by $1+a_2+\cdots +a_2^{25}$ gives $26u_1=0$. Similarly, $a_1$ fixes $s{u}_2$ and its fifteenth power acts trivially, so $15s{u}_2=0$, hence $15u_2=0$ and $15u_1=0$. Since $15$ and $26$ are coprime, $u_1=0$; then invertibility of $s$ gives $u_2=0$. Their conjugates generate the abelianization, so $K$ is perfect.

The finite plus construction now adds two two-cells and two three-cells. Thus $X$ has one vertex, five edges, six two-cells and two three-cells, and satisfies $D(2)$. The remaining task is to exclude a finite two-dimensional model. (OpenAI, 2026)

🏷️ The twisted matrix and the final contradiction

Choose a primitive cube root of unity $\zeta$. Define a character by

$$\begin{array}{c}\rho (x_1)=-1,\rho (a_1)=\zeta ,\\ \rho (x_2)={\zeta }^2,\rho (a_2)=-1,\\ \rho (s)=1.\end{array}$$

These values satisfy the four relations and kill $t_1,t_2$, so the character descends to $G$. In particular, $x_1$ has order exactly two in $G$: its square is one, but its character value is $-1$.

The twisted second boundary on $B$ maps ${\mathbb{C}}^4$ to ${\mathbb{C}}^5$. The four boundary vectors, written as rows with columns $x_1,a_1,x_2,a_2,s$, are

$$\begin{array}{c}D=(A\mid B),\\ A=\left(\begin{array}{ccc}1-\zeta & -2& 0\\ 0& 0& 2\\ 1& 0& 0\\ 0& -(1+\zeta )& 1\end{array}\right),\\ B=\left(\begin{array}{cc}0& 0\\ {\zeta }^2-1& 0\\ -1& 0\\ 0& 1-{\zeta }^2\end{array}\right).\end{array}$$

One can compute each row by following the attaching word, weighting a traversed edge by the character of the path already traversed. Equivalently, use the Fox product rule $\partial (uv)=\partial u+u\partial v$, with $\partial g/\partial g=1$ and $\partial g^{-1}/\partial g=-g^{-1}$, and evaluate at $\rho$. For example, the relation $xa{x}^{-1}{a}^{-n}$ contributes $1-\rho (a{)}^n$ in its $x$ coordinate and $\rho (x)-{\sum }_{j=0}^{n-1}\rho (a{)}^j$ in its $a$ coordinate.

The minor on columns $x_1,a_1,x_2,s$ is

$$-4(1-{\zeta }^2)\ne 0.$$

Thus the rows are independent, the second boundary is injective, and $H_2(B;L_{\rho })=0$. The plus construction preserves this homology, so $H_2(X;L_{\rho })=0$ as well.

A finite two-dimensional model $Y\simeq X$ would inherit both this vanishing and the character taking value $-1$ on an element of order two. That contradicts the tower obstruction. Equivalently, the corresponding loop would extend to ${\mathbb{RP}}^2\to Y$ with nontrivial pulled-back monodromy, the precise configuration the tower excludes. (OpenAI, 2026)

All four source sections were read for this exposition, and the displayed minor was independently checked symbolically. This checks the matrix and explains the proof architecture; it does not constitute a formal verification of the manuscript or its imported topological results.

🏷️ Related notes

Cyclic covers and Borel rigidity also turns properties of covers into an obstruction to geometric realization. Its proper height-translation argument differs from the fixed-domain simplex count used here.

🏷️ Key steps and proof outlines

The construction separates a chain-level cancellation from a geometric realization problem. The following diagram shows where the distinction enters.

flowchart TD
  B["Finite presentation complex B"] --> P["Kill the perfect normal subgroup with two 2-cells"]
  P --> X["Add two 3-cells with split boundary"]
  X --> D["D(2) holds for every coefficient module"]
  B --> T["Injective twisted second boundary"]
  T --> H["Twisted H2 of X vanishes"]
  H --> Y["Suppose a finite 2-dimensional model Y exists"]
  Y --> F["Order-two loop gives a map from RP2"]
  F --> C["Cyclic lifts strictly increase image size"]
  C --> N["Contradiction with the fixed finite domain"]

A split boundary is not a geometric cell cancellation

In the plus construction, $q{\partial }_3=\mathrm{id}$ says that each new three-cell has coefficient one on its designated new two-cell after projection. It proves injectivity of ${\partial }_3$ and gives an explicit extension of every third cochain. It also contracts the relative chain complex, even after arbitrary changes of coefficients.

This does not assert that the three-cells can be removed by geometric elementary collapses. An elementary collapse requires an actual free face in the attaching geometry. The matrix identity records the projected cellular incidence, while the attaching spheres can also involve the old cells. The point of the example is that these strong homological cancellations coexist with the obstruction to every finite two-dimensional homotopy model. (OpenAI, 2026)

The tower consumes a resource that cannot be replenished

The tower proof uses a particularly concrete finite resource: the simplices in one fixed triangulation of the source $F$. Each stage maps this same $F$ onto its image $A_i$. Lifting to a cyclic cover does not subdivide $F$, so the number of simplices of $A_i$ is bounded by the number in $F$ at every stage.

Why is growth strict? The projection $A_{i+1}\to A_i$ is onto, with each simplex mapped isomorphically to a simplex. If both image complexes had the same total number of nonempty simplices, every target simplex would have exactly one lift in the image. Face compatibility would make these lifts a simplicial section of the cyclic cover over all of $A_i$. On fundamental groups, a section would make the image of the covering subgroup equal to the whole group. That is impossible for the connected cover defined by an epimorphism to $\mathbb{Z}$.

Thus every lift spends at least one additional simplex while the supply stays fixed. This explains both why the source is triangulated once and why allowing a new subdivision at every stage would break the contradiction.

The twisted matrix can be checked without numerical roots

The nonzero minor can be verified symbolically using $1+\zeta +{\zeta }^2=0$. Restrict the displayed matrix to columns $x_1,a_1,x_2,s$. Its last column has only one nonzero entry, $1-{\zeta }^2$, in the last row. The remaining determinant is

$$\det \left(\begin{array}{ccc}1-\zeta & -2& 0\\ 0& 0& 2\\ 1& 0& 0\end{array}\right)=-4.$$

Hence the full minor is $-4(1-{\zeta }^2)$, nonzero because a primitive cube root does not have square one. No numerical approximation of $\zeta$ is needed. The same exact character makes $x_1$ nontrivial while the quotient relation makes $x_1^2=1$. These two uses of the character are complementary: one supplies homological vanishing, the other supplies the forbidden torsion monodromy.

🏷️ Manuscripts and verification scope

• A Counterexample to Wall’s D(2) Problem — pinned manuscript and companion source.

No family scope page was present at lean/docs/321.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. A Counterexample to Wall’s D(2) Problem.