gpt 6 astra ultra attempt on Math-17

Smooth 4-Dimensional Poincaré Conjecture

Definitions and mathematical statement

A closed manifold is compact without boundary. A smooth homotopy $4$-sphere is a smooth closed $4$-manifold $M$ for which there exist continuous maps $M\leftrightarrows S^4$ whose compositions are homotopic to the identities. The conjecture is \[ M\simeq S^4\quad\Longrightarrow\quad M\cong_{\rm diff} S^4. \] A diffeomorphism must be smooth with smooth inverse. A homeomorphism alone does not settle this smooth-category question.

Formulation and concept references: [S1].

Short English statement

Is every smooth four-dimensional space with the homotopy type of a sphere smoothly the standard sphere?

Research attempt

This unresolved attempt by GPT-6 Astra Ultra was prepared on 27 September 2026, using explicit reductions, source checks, and adversarial verification.

Scope, checked inputs, and two routes.

The target is a diffeomorphism, not a homeomorphism, for every smooth closed homotopy four-sphere. This attempt pursues a proof route through geometric handle cancellation and a disproof route through a compact contractible manifold with boundary $S^3$. It derives precise sufficient certificates for each route and tests what information would be lost by replacing geometry with matrices or a finite group computation.

Gerig [S1], Theorems 1.2 and 1.3, records two inputs used below: an orientation-preserving diffeomorphism of $S^3$ extends across $D^4$ (Cerf), and equi-oriented embedded closed four-balls in a connected four-manifold are ambient isotopic (Cerf--Palais). His invariant result is not a classification of all smooth homotopy spheres. The source repository [S2], under its review boundary, explicitly leaves geometric interfaces as hypotheses of its formalized implication. No external Lean code was executed and no counterexample is accepted here on that basis.

The handle and knot strategies have substantial precedents, notably Freedman--Gompf--Morrison--Walker [E1]. What follows is a written derivation and adversarial check of these routes, without a novelty claim.

Lemma 22.1: an equivalent compact-boundary problem.

The smooth four-dimensional Poincaré conjecture is equivalent to: every smooth compact contractible four-manifold $W$ with $\partial W\cong S^3$ is diffeomorphic to $D^4$.

Proof. First let $M$ be a homotopy four-sphere, choose a smooth embedded closed ball $B$, and put $W=M\setminus\operatorname{int}B$. Van Kampen, applied using boundary collars, gives $\pi_1(M)\cong\pi_1(W)$ because $B$ and $S^3$ are simply connected. Thus $W$ is simply connected. Excision identifies $H_*(M,W;\mathbb Z)$ with $H_*(D^4,S^3;\mathbb Z)$. In the long exact sequence, the map $H_4(M)\to H_4(M,W)$ is an isomorphism: it sends the oriented fundamental class to the local relative fundamental class. Together with the homology of $M$, this gives \[ H_0(W;\mathbb Z)=\mathbb Z,\qquad \widetilde H_j(W;\mathbb Z)=0\quad(j\ge0). \] A simply connected acyclic CW complex is contractible. Indeed any first nonzero higher homotopy group would, by Hurewicz, give nonzero homology; if there is none, Whitehead applies to its map to a point. Smooth compact manifolds have CW homotopy type, so this applies to $W$; see [E2], Section 4.1. If the stated compact-boundary assertion holds, $W\cong D^4$. The remaining boundary gluing of two standard balls gives $S^4$ by Cerf's extension theorem. For an orientation-reversing gluing, compose with a standard reflection extending over a ball and then use the orientation-preserving theorem.

Conversely, cap such a $W$ by a standard ball to obtain $M=W\cup_{S^3}D^4$. Van Kampen makes $M$ simply connected and Mayer--Vietoris gives the integral homology of $S^4$. Hurewicz gives $\pi_2(M)=\pi_3(M)=0$ and $\pi_4(M)\cong H_4(M)\cong\mathbb Z$. A map $S^4\to M$ representing a generator is a homology equivalence, and the simply connected homology form of Whitehead makes it a homotopy equivalence. If the original conjecture holds, $M\cong S^4$. The capping ball maps to a smooth embedded standard ball in $S^4$; Cerf--Palais moves it to a hemisphere. Its closed complement is a standard $D^4$, and hence so is $W$.$\square$

The boundary condition cannot be replaced by "homology sphere" in this argument. An acyclic manifold can have a quite different boundary; without an actual boundary diffeomorphism to $S^3$, the capping step does not even have the asserted input. Nor is a diffeomorphism of an open interior by itself the boundary-compatible compact statement proved equivalent above.

Attempt A: what a cancellation certificate must contain.

A smooth homotopy sphere admits a finite handle description. Removing a small ball leaves a compact handlebody, but this does not justify assuming that every such description has only indices zero, one and two. Elimination of the three-handles is not an input supplied by simple connectivity or homology alone.

When there is a description of $W$ with one zero-handle, $r$ one-handles, and $r$ two-handles and no higher handles, its cellular chain complex has the form \[ 0\longrightarrow\mathbb Z^r \xrightarrow{A}\mathbb Z^r \xrightarrow{0}\mathbb Z\longrightarrow0. \] Rows of $A$, up to the transpose convention, are exponent sums of the attaching words in the one-handle generators. If $W$ is contractible, $A$ is unimodular. Integer row and column operations can turn it into the identity; corresponding handle slides change the algebraic intersection data. This observation alone does not give the geometric position required to cancel handles.

The cancellation theorem requires the attaching sphere of the higher handle and the belt sphere of the lower handle to meet transversely in exactly one point, with their actual attaching data; [E3], Lemma 3.8. An algebraic intersection number one could instead be the signed sum $1+1-1$ of three intersections. A matrix entry does not certify that the extra intersections have been removed by permitted isotopies. Likewise, altering a word or an integer framing without recording its geometric realization is not a handle move.

A sufficient proof certificate is therefore a finite list of ambient isotopies of the framed attaching data, handle slides, introductions of standard cancelling pairs, and geometric cancellations, ending with a single zero-handle for $W$. Every step must be legal on the current boundary and account for the remaining handles. Reversing the resulting diffeomorphisms gives $W\cong D^4$ and Lemma 22.1 gives $M\cong S^4$. This proves the conditional implication. No theorem ensuring such a list for every candidate is supplied here. Failure of a bounded search for such a list is also not an invariant of the manifold.

An exact test of the homology-only shortcut.

Consider the balanced group presentation \begin{equation}\label{wsref-0-63-6f-6e-74-32-32-3a-70-72-65-73-65-6e-74-61-74-69-6f-6e} P=\langle s,t\mid s^3t^{-5},\ (st)^2s^{-3}\rangle. \end{equation} Its exponent-sum matrix is \[ A=\begin{pmatrix}3&-5\\-1&2\end{pmatrix},\qquad\det A=1. \] The associated presentation complex, with one vertex, two edges and two two-cells, is consequently acyclic: its $H_1$ is the cokernel of this invertible matrix and its $H_2$ is the kernel. Yet its fundamental group is nontrivial. In permutations of five letters, let \[ s=(1\ 2\ 3),\qquad t=(1\ 3\ 4\ 2\ 5), \] and compose right to left. Then $s^3=t^5=1$ and $st=(2\ 5)(3\ 4)$, so $(st)^2=1$. Both relators map to the identity, whereas $s$ does not. This defines a nontrivial finite image of the presented group.

Thus even a balanced presentation with unimodular matrix cannot certify simple connectivity. The example is not offered as a homotopy four-sphere, and no boundary-$S^3$ claim is made for an associated handlebody. It rejects only the proposed inference from the chain calculation to contractibility. For an actual candidate whose fundamental group has already been proved trivial, the further geometric-cancellation problem still remains. A perfect group example does not prove that any particular simply connected candidate is exotic.

Attempt B: turn cocores into a potential disproof certificate.

Suppose a candidate $W$ has been independently certified contractible, has an explicit boundary identification $\partial W\cong S^3$, and has a handle decomposition using only indices zero, one and two. A two-handle $D^2\times D^2$ contains a cocore disk $\{0\}\times D^2$. Its boundary is a circle in the final boundary. Cocores of different two-handles give disjoint properly embedded disks.

Choose some $k\ge1$ of these disks and join their boundary circles by $k-1$ disjoint oriented bands in the boundary, arranged in a tree. Push band interiors slightly into a collar of $\partial W$, smoothing the joins to the disks. This gives an embedded connected orientable surface in $W$. The tree condition joins different components at each step, leaving one boundary component, and its Euler characteristic is \[ \chi=k-(k-1)=1. \] Hence the surface is a disk. Under the boundary identification, its boundary is an explicit knot $K\subset S^3$ which is smoothly slice in $W$.

Let $s(K)$ denote Rasmussen's knot invariant, normalized so that the unknot has value zero. The established slice-genus inequality [E4], Theorem 1, is \begin{equation}\label{wsref-0-63-6f-6e-74-32-32-3a-73-6c-69-63-65} |s(K)|\le2g_4(K), \end{equation} where $g_4$ is the minimum genus of a smooth connected orientable surface in the standard $D^4$ with boundary $K$.

Proposition 22.2. If the preceding construction produces a knot with a certified value $s(K)\ne0$, then $W\not\cong D^4$, and capping $W$ gives a counterexample to smooth four-dimensional Poincaré.

Proof. If $W\cong D^4$, transfer the constructed disk to $D^4$. Its boundary identification may differ from the prescribed one by a diffeomorphism of $S^3$. Cerf's extension theorem, with a reflection if necessary, extends that difference over $D^4$ and restores the prescribed knot. Thus $g_4(K)=0$, contradicting (cont22:slice). The cap is a homotopy sphere by Lemma 22.1, and would force $W$ standard if it were standard.$\square$

This is the established type of test in [E1], Fact 2.1, written out to show its dependencies. It is not a new counterexample. No candidate with all the required properties and nonzero $s$ is produced here. In particular, simply finding a nonslice knot and declaring it to be a cocore band sum in a homotopy ball would assume the key construction.

Stress tests for the knot route.

The input disk must be embedded, not merely an immersed disk with double points. A disk with self-intersections does not give $g_4=0$ after being transferred to a standard ball. The boundary identification must track the actual knot, not just its homology class: every knot in $S^3$ has trivial class in $H_1(S^3)$, which supplies no sliceness information. The band construction explicitly keeps both properties.

The tree condition on the bands also matters. An additional band can change the number of boundary components or the genus; Euler characteristic must be recomputed before claiming a disk. If the constructed surface has genus $g>0$, the same reasoning supplies an obstruction only when $|s(K)|>2g$. A nonzero value smaller than that threshold would not suffice.

Conversely $s(K)=0$ is compatible with both slice and nonslice knots; (cont22:slice) has no converse in this argument. Zero values on a finite collection of band sums do not prove $W$ standard. A failed cancellation search and failed invariant search can therefore coexist without resolving the candidate in either direction.

Verification, next steps, and outcome.

The accompanying finite check verifies the determinant and the two permutation relators in (cont22:presentation), using explicit permutation composition. It does not certify any four-dimensional diagram. No knot-homology computation or candidate sphere is claimed to have been verified. The compact-boundary reduction retains the smooth extension and isotopy theorems as named external inputs.

UNRESOLVED. The strongest conclusions are the compact-boundary equivalence, an exact obstruction to a homology-only certificate, and a fully specified conditional knot certificate. The proof route stops at geometric simplification of an arbitrary candidate; the disproof route stops at producing a certified homotopy ball and embedded knot disk with a nonzero standard-ball obstruction. Concrete next work is to track framed handle moves on an actual diagram, verify its boundary as $S^3$, and compute a suitable knot invariant with reproducible input and grading conventions. Neither a matrix normal form nor a conditional formal implication alone completes those tasks.

Sources

• S1 Chris Gerig, No homotopy 4-sphere invariants using ECH = SWF, Algebraic & Geometric Topology 21 $2021$, Conjecture 1.1. https://arxiv.org/pdf/1905.10938. Formulation source.
• S2 smooth4pc-t73-lean: Independent Review. https://github.com/toffee-desuwa/smooth4pc-t73-lean/blob/main/docs/INDEPENDENT_REVIEW.md. Further reference; not a proof endorsement.
• S3 The Smooth Four-Dimensional Poincare Conjecture. https://epoch.ai/frontiermath/open-problems/smooth-poincare-4d. Further reference; not a proof endorsement.
• E1 Michael Freedman, Robert Gompf, Scott Morrison and Kevin Walker, Man and machine thinking about the smooth 4-dimensional Poincaré conjecture, arXiv:0906.5177, Fact 2.1 and Sections 2--3. https://arxiv.org/pdf/0906.5177. Precedent for the cocore band-sum and Rasmussen-invariant route.
• E2 Allen Hatcher, Algebraic Topology, Chapter 4, Section 4.1. https://pi.math.cornell.edu/~hatcher/AT/ATch4.pdf. Hurewicz and Whitehead theorems used in the compact-boundary reduction.
• E3 Mark Powell, MAT993D Topology of Manifolds: Surgery Theory, Section 3.3, Lemma 3.8. https://www.maths.gla.ac.uk/~mpowell/surgery-notes.pdf. Geometric handle-cancellation criterion; no high-dimensional surgery conclusion is asserted in dimension four.
• E4 Jacob Rasmussen, Khovanov homology and the slice genus, arXiv:math/0402131, Theorem 1. https://arxiv.org/pdf/math/0402131. The inequality for the smooth slice genus in the standard ball. Additional references consulted 27 September 2026.