Four-dimensional disk embedding and a duality group with no manifold model

← All digests

Family 305 · Topology
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-22

1. The question

In four dimensions, algebraically cancelling intersections need not be easy to remove geometrically. Two surfaces meet at isolated points, and a disk intended to cancel a pair of intersections can itself intersect the surfaces. This is a central difficulty in four-manifold topology.

The disk-embedding problem asks whether suitable intersection data and auxiliary dual spheres allow immersed disks to be replaced by disjoint embedded disks. The theorem is known under restrictions on the fundamental group. The first two manuscripts claim that removing those restrictions is impossible, even if the replacement disks may change their relative homotopy classes.

The third manuscript turns a related marked obstruction into a different counterexample: a group with the integral duality and finite classifying space expected of a closed aspherical four-manifold, but with no such manifold realization. The common theme is that strong algebraic information can still fail to produce the required four-dimensional geometry.

2. Background and essential definitions

Immersed disks and algebraic duals

An immersed disk is locally embedded but may have self-intersections. A proper disk in a manifold with boundary has its boundary on the ambient boundary and its interior in the ambient interior. A locally flat embedded disk has a neighborhood locally homeomorphic to the standard pair $(\mathbb R^4,\mathbb R^2)$; it need not be smooth.

The fundamental group $\pi=\pi_1(M)$ records loops up to homotopy. To count intersections of surfaces, choose paths from them to a common basepoint. Each intersection then carries both a sign and a loop label in $\pi$. The equivariant pairing $\lambda$ adds these contributions in the group ring $\mathbb Z[\pi]$, whose elements are finite integer sums of group elements.

Sphere maps $g_j$ are algebraically dual to disks $f_i$ when

$$\lambda(f_i,g_j)=\delta_{ij}\,1.$$

Here $\delta_{ij}$ is one when $i=j$ and zero otherwise, and $1$ is the identity group element. Algebraic duality allows cancelling intersection pairs. Geometric duality would require exactly one actual intersection with the designated disk and none with the others.

A framing is a trivialization of the normal two-plane bundle. It records how a neighborhood twists. The reduced Wall self-intersection invariant $\widetilde\mu$ records signed self-intersections with group labels, modulo the standard changes from reversing a loop and the identity contribution. Its target is an additive quotient, not a quotient by a group-ring ideal.

Good groups

In Freedman–Quinn theory, a group is called good when specified local disk operations can be performed with double-point loops becoming trivial in that group. This is a technical property designed to make the disk-embedding construction work.

A capped grope is a tree of surfaces attached along curves, with disks capping its terminal curves. Such objects organize repeated attempts to resolve intersections. Their double-point loops explain why the ambient fundamental group enters a problem that initially looks local.

Duality groups and manifold models

A closed oriented $n$-manifold has Poincaré duality, relating homology in complementary dimensions. A Poincaré complex satisfies the corresponding homological duality, including local coefficients, without necessarily being a manifold.

An aspherical space has contractible universal cover. It is a classifying space for its fundamental group, so its homotopy type is determined by that group. An oriented integral Poincaré duality group of dimension four has the algebraic finiteness and duality conditions summarized by

$$H^i(G;\mathbb ZG)=0\ (i\ne4), \qquad H^4(G;\mathbb ZG)\cong\mathbb Z,$$

with trivial group action on the final copy of $\mathbb Z$, together with a finite resolution by finitely generated projective modules. Group-ring coefficients retain information from the universal cover; ordinary Betti-number symmetry alone would be much weaker.

3. Development history and previous partial results

Casson’s handle construction and Freedman’s recognition theorem established the simply connected topological foundation. A Casson handle is an infinite tower of immersed disks; Freedman’s theorem identifies it topologically, relative to its attaching region, with a standard open two-handle. This provides topological disks where smooth intersection-removal methods fail.

Freedman and Quinn developed the general framework using capped gropes and good groups. Freedman–Teichner proved that groups of subexponential growth are good, and Krushkal–Quinn gave another proof while recording a correction to the earlier height-raising argument. The known good-group class also has closure properties that include many exponential-growth groups, such as solvable groups. Growth rate alone therefore does not identify the boundary of the theory.

Powell, Ray and Teichner supplied a missing construction of output geometric dual spheres in the classical theorem. This correction concerns the dual spheres, separately from the height-raising correction. Their modern formulation gives the precise positive theorem that the new manuscripts seek to obstruct without the good-group assumption.

Link and surgery formulations provide further context. Weak and homotopy versions of the A–B slice problem admit important positive results by Krushkal and Freedman–Krushkal. They relax conditions that matter to actual embeddings. There are also positive surgery theorems for free fundamental groups under additional hypotheses on intersection pairings. The claimed counterexamples do not erase these restricted results.

The manifold-realization problem has its own finiteness distinctions. Wall asked whether Poincaré duality groups come from closed aspherical manifolds. Davis constructed integral examples that are not finitely presented, already excluding manifold realization in that unrestricted formulation. Fowler constructed finitely presented rational duality examples without finite classifying spaces.

M3 claims to avoid both explanations: its group has integral duality, is finitely presented, and already has a finite four-dimensional classifying space. Its obstruction concerns turning that finite Poincaré complex into a topological manifold of the same dimension.

4. What the new manuscripts claim

All three manuscripts are dated September 24, 2026.

M2 gives a framed disk counterexample. In a compact connected oriented smooth four-manifold, it constructs finitely many proper generically immersed disks with disjoint embedded boundaries and framed algebraic dual spheres satisfying

$$\lambda(f_i,g_j)=\delta_{ij}\,1, \qquad \lambda(g_i,g_j)=0, \qquad \widetilde\mu(g_i)=0.$$

Nevertheless, it claims there are no pairwise disjoint locally flat replacement disks with the same boundary maps and induced boundary normal framings. No relative homotopy classes for the output disks are prescribed.

M1 gives a stronger boundary-only obstruction. Its specified boundary circles bound no pairwise disjoint proper locally flat disks at all, even when output framings and output dual spheres are not required. The input dual spheres satisfy the same algebraic conditions; each can individually be chosen smoothly embedded, although different spheres may intersect.

The vanishing conditions concern the dual spheres. The theorem does not require vanishing intersections among the input disks. M1 deduces that the free group $F_2$ is not good, and neither is any group containing it. This deduction uses subgroup and quotient closure of goodness; it does not identify the counterexample manifold’s fundamental group with $F_2$.

M1 also concludes that the unrestricted topological surgery and five-dimensional topological s-cobordism assertions for four-manifolds cannot both hold. That is a statement about their conjunction, not an identification of which separately formulated assertion fails.

M3 gives a manifold-realization counterexample. It claims a finitely presented oriented integral duality group $G$ of dimension four with a finite oriented four-dimensional aspherical Poincaré complex $Q$, but no closed aspherical topological four-manifold with fundamental group $G$.

M3 specifically imports M2’s marked geometric and tensor package. The boundary-only headline of M1 is not a substitute for that more detailed input.

5. Why the advance matters

Importance: 5/5, provisional. If correct, the family identifies fundamental limits to two major four-dimensional existence principles: unrestricted disk embedding and realization of finitely presented integral duality groups by manifolds.

The boundary-only statement is particularly strong. It cannot be avoided by choosing different homotopy classes or allowing a different normal framing for the replacement disks. The obstruction lies in the existence of any disjoint locally flat filling of the given circles.

The group-realization result is similarly stronger than a failure of smoothability. It excludes a closed aspherical topological manifold, despite providing a finite aspherical Poincaré complex with the expected integral duality.

The claims are dimension-specific. A counterexample in dimension four refutes an assertion quantified over all dimensions, but does not establish counterexamples in every dimension or invalidate surgery theorems under their established hypotheses.

6. The main idea

Encode boundary information before choosing disks

The first two constructions install a finite family of commutator relations in a smooth plumbing model. A commutator is $[x,y]=xyx^{-1}y^{-1}$; triple commutators record interactions among three inputs. The words become trivial if any input is deleted, and independent conjugating paths are included in the fixed data.

Capped gropes realize these relations geometrically. Removing neighborhoods of their base surfaces produces the ambient disk problem; compressed product tori supply algebraic dual spheres. The markings record based loops and paths in this initial construction. They do not impose extra unannounced conditions on hypothetical output disks.

If the requested embedded disks existed, their images would allow the manifold to be cut into pieces separated by smooth three-dimensional walls with torus boundaries. The cuts miss the locally flat disks, so ordinary smooth differential forms can still be used on the cut pieces.

Convert the cuts into incompatible algebra

The papers study deformations of flat connections, equivalently compatible matrix holonomies around loops, using $\mathrm{SL}_2$, the determinant-one two-by-two matrices. Wall data give formal potentials whose derivative equations describe flatness. The potentials arise from Chern–Simons transgression, which relates boundary expressions to curvature in the interior.

Restriction from the two sides supplies complementary tangent data. The installed triple commutators force products to belong to actual gradient ideals. This is stronger than saying that the products vanish at every ordinary solution.

The distinction matters because the parameter rings contain nilpotent elements. For example, in a ring with $s^2=t^2=0$, the mixed term $st$ may still be nonzero. Passing only to ordinary points would discard precisely such information. The arguments retain the mixed terms and their ideal-membership witnesses.

M2 uses a comparison with algebraic critical charts that preserves the marked holonomies. M1 develops a boundary-polarized replacement and correction argument suited to the unframed boundary obstruction. These are related strategies, but their input requirements and algebraic packages are not identical.

After algebraic replacement and specialization to positive characteristic, truncating coordinate powers makes the relevant differential-form complexes finite. The transverse data determine classes with nonzero pairing. The two families of cubic ideal relations force that same pairing, or a nonzero parameter product times it, to vanish. With sufficiently many wall directions, the claimed finite tensor obstruction gives the contradiction.

Turn a marked chamber into a closed duality complex

M3 starts with a chamber whose boundary has a specified map to a finite graph. Attaching three-cells along a hyperbolic basis of middle-dimensional spheres produces a Poincaré pair with graph homotopy type. The imported marked obstruction says that its boundary cannot be filled by a topological four-manifold with that graph homotopy type while preserving the boundary map.

A reflection construction assembles copies of the chamber into a closed aspherical Poincaré complex $Q$. An auxiliary CAT(0) model supplies the assembly-theorem inputs needed later; it does not itself provide a four-manifold model for $Q$.

Assume a closed aspherical manifold model existed. Stable surgery and a product comparison would then make prescribed sphere pairs standard after stabilization by copies of $S^2\times S^2$. The sphere markings must be retained: an unmarked stable equivalence would not supply the needed geometric data.

The paper lifts these data to a finite union of reflected chambers, caps all but the central boundary and removes the standard inner pairs. Direction control far from the central chamber arranges the remaining loop systems inside suitable contractible patches. This supplies the particular null-loop step needed for the remaining disk construction, rather than assuming goodness of the free chamber group.

Surgery would then produce exactly the forbidden marked graph-type filling. That contradiction rules out the hypothetical closed manifold model.

7. Scope, limitations, and open questions

The input manifolds in the disk examples are smooth, but the forbidden output disks are topological and locally flat. Thus the claims are not merely smooth disk obstructions.

M2 preserves boundary framings; M1 removes that output requirement. Both allow changes of relative homotopy class. The marked loops in their constructions are essential to the arguments and should not be confused with additional restrictions on prospective disks.

The nongoodness conclusion for $F_2$ does not explicitly produce a counterexample manifold with that exact fundamental group. The surgery/s-cobordism consequence also retains its logical limitation: the two unrestricted assertions cannot both be true.

M3 concerns integral, oriented duality in dimension four and excludes every closed aspherical topological manifold with the specified group. It does not classify all four-dimensional duality groups or answer the corresponding realization problem in each other dimension. Its finite Poincaré complex remains a valid homotopy-theoretic realization; the missing structure is the manifold structure.

8. Sources and further reading

Freedman–Quinn’s Topology of 4-manifolds develops the disk and surgery framework. Powell–Ray–Teichner’s disk theorem with dual spheres gives the modern positive statement and its group hypothesis.

For the realization problem, compare Davis’s duality constructions, Fowler’s rational examples, and Davis–Hillman’s relative realization results. The source manuscripts give the exact marked interfaces and finite algebraic obstructions summarized here.

Source manuscripts

• M1. OpenAI, A boundary-only obstruction to four-dimensional disk embedding (September 24, 2026).
• M2. OpenAI, A marked tensor obstruction to four-dimensional disk embedding (September 24, 2026).
• M3. OpenAI, A PD4 group without an aspherical manifold model (September 24, 2026).

Sources consulted on 22 October 2026 at repository revision 1581d4c. Download the manuscript citations.