Cosmetic knot surgery: repetition and an oriented boundary count

Source and verification

This article explains family 306 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

Dehn surgery fills the boundary torus of a knot exterior in a chosen direction. Can two different directions produce the same oriented three-manifold? Family 306 claims that this never happens for a nontrivial knot in the three-sphere. After importing a reduction to slopes $-2$ and $2$, the manuscript constructs a nonzero instanton count and a one-dimensional spinor count whose boundary would force that integer to vanish. (OpenAI, 2026)

🏷️ Preliminaries: slopes and oriented fillings

Let $K\subset S^3$ be a smooth knot and remove the interior of a tubular neighborhood to form its exterior $E_K$. Its boundary is a torus. The meridian $\mu$ bounds a disk in the removed solid torus; the preferred longitude $\lambda$ follows the knot and is nullhomologous in the exterior.

A slope is an unoriented primitive torus homology class $p\mu +q\lambda$, with $\gcd (p,q)=1$, taken up to simultaneous sign. We label it $p/q\in \mathbb{Q}\cup \{\infty \}$. Dehn filling glues in a solid torus so that this slope bounds its meridian disk, producing $S_{p/q}^3(K)$ with the orientation extending that of the exterior.

Two distinct slopes are purely cosmetic if their fillings are orientation-preservingly homeomorphic. An orientation-reversing coincidence is a different question.

Manuscript claim

For a smooth nontrivial knot $K\subset S^3$, an orientation-preserving homeomorphism

$$S_r^3(K)\cong S_s^3(K)$$

implies $r=s$. The homeomorphism need not preserve the surgery core or the original knot exterior. (OpenAI, 2026)

Nontriviality matters: surgeries of slopes $1/q$ on the unknot give the three-sphere. The meridional filling on any knot also gives $S^3$, but the imported sphere-filling theorem excludes a second such filling for a nontrivial knot.

For a finite slope, abelianizing the filling relation gives first homology $\mathbb{Z}/|p|$ if $p\ne 0$, and $\mathbb{Z}$ if $p=0$. Thus zero cannot pair with any other slope. The manuscript imports the combined Floer-theoretic reduction that every remaining distinct purely cosmetic pair must be $\{-2,2\}$ on a genus-two knot. Genus is the minimum genus of an orientable surface in $S^3$ bounded by the knot.

Write $Y_j=S_j^3(K)$. Three-dimensional smoothing permits an assumed oriented homeomorphism $Y_{-2}\to Y_2$ to be replaced by an oriented diffeomorphism $\varphi$. The new argument seeks a contradiction from this arbitrary $\varphi$.

🏷️ Gauge-theoretic counts and their prerequisites

A connection describes parallel transport in a vector bundle, and its curvature measures the failure of transport around small loops to be trivial. On an oriented Riemannian four-manifold, the Hodge star splits two-forms into self-dual and anti-self-dual parts. An instanton satisfies the anti-self-dual equation for the trace-free curvature.

Gauge transformations change a bundle’s local descriptions. A moduli space is the solution space modulo these transformations. Its Fredholm index gives an expected dimension; actual counts additionally require regularity, orientations, and control of limits. A reducible connection preserves a splitting into line bundles, and has extra symmetry that must be handled separately.

Instanton Floer theory packages critical connections on a three-manifold into a chain complex. Its differential counts trajectories, and four-dimensional cobordisms give maps by counting solutions with specified end behavior. Incidence conditions impose further constraints, lowering expected dimension.

The manuscript fixes determinant connections and uses determinant-one gauge transformations throughout. At its ordinary rational ends, first homology has order at most two, so reducible flat connections have central normalized holonomy. At zero surgery it chooses a determinant evaluating oddly on the capped Seifert surface, excluding reducible projectively flat connections.

A unit below means a map inducing an isomorphism on Floer homology modulo two. It does not mean an invertible matrix on the chain group or an integral homology isomorphism.

The lengthy constructions of ordinary representatives, surgery maps, symplectic caps, and their compactifications are manuscript inputs in this post. The finite repetition argument and final link multiplicity can be explained once these inputs are stated precisely.

🏷️ A negative cobordism and two determinant lifts

The difficult negative step is a cobordism from $Y_2$ to $Y_{-2}$ containing a sphere $S$ with $S^2=-4$. Its metric interval runs between a three-sphere splitting and a splitting that separates a neighborhood of $S$, whose boundary is a lens space of order four.

Two determinant choices differ by $\mathrm{PD}(S)$ and agree outside that neighborhood. A degree-one sweep of $S$ by loops imposes normalized holonomy $-1\in SU(2)$. Holonomy has codimension three, while the sweep contributes one parameter, so this is a degree-two sphere insertion. With the metric interval, the dimension calculation is

$$i+1+1-3=0,i=1.$$

The manuscript chooses a local relative sign for the two determinant sectors. Their signed counts give an integral chain map

$$B:C(Y_2;\mathbb{Z})⟶C(Y_{-2};\mathbb{Z})$$

whose reduction modulo two is a unit. The two minimal lens-cap contributions cancel with these signs. This same sign comparison is retained in the later spinor boundary argument; it is not chosen afresh at the end.

🏷️ Finite repetition preserves a nonzero count

Fixed caps supply an integral Floer cycle and functional with nonzero pairing $q$, after clearing fixed rational denominators. The functional kills torsion, so work on the free lattice

$$L=H(C(Y_2);\mathbb{Z})/\mathrm{Tor}.$$

Choose $N$ so that $q\not\equiv 0(\mathrm{mod}{2}^N)$. Every integral chain map used in the repetition that is a quasi-isomorphism modulo two induces an automorphism of $L/2^{N}L$.

Here is the algebraic reason. Its mapping cone is a finite free complex with zero mod-two homology. Over the local ring ${\mathbb{Z}}_{(2)}$, the universal coefficient sequence and Nakayama’s lemma make the cone homology zero. The induced map on free homology therefore has odd determinant, and its adjugate makes it invertible modulo every $2^N$. Integral Floer homology need not be torsion-free.

Let $\mathcal{A}={\varphi }_{\ast }B$ be the returned negative operation and $\mathcal{D}$ the positive-return operation. The chronological four-interval pattern acts as ${\mathcal{A}}^3\mathcal{D}$. The finite group

$$\mathrm{GL}(L/2^{N}L)$$

has an exponent $E$: every element raised to $E$ is the identity. Enlarge $E$ to be divisible by eight. Choose outside padding lengths and the repetition number as sufficiently large multiples of $E$. Use literally periodic local data so each repeated operation really is a power of one map.

The closed count then satisfies

$$\Omega \equiv \pm q(\mathrm{mod}{2}^N),\Omega \ne 0.$$

This uses a manuscript gluing and comparison theorem identifying the actual count in the final metrics with the composition. A formal product of operators alone would not establish that geometric identity.

The same repetition increases the complex coupled Dirac index. With padding and cap data fixed, the source formula has the form

$$n_D=\frac{q_0}{8}+C,$$

where $q_0$ is the repetition number and $C$ is fixed. Thus $n_D>0$ for sufficiently large allowed $q_0$, while nonvanishing persists. Divisibility by eight also preserves the bundle-degree congruences. All topological integers and determinant lifts are chosen before the later metric scales and perturbations.

🏷️ Spinors and the doubled projective link

A spinor is a section of a bundle associated to a spin${}^c$ structure. The coupled Dirac operator is an elliptic first-order operator on such sections; its complex index is kernel dimension minus cokernel dimension. The nonabelian monopole equations couple a connection to a spinor. Their zero-spinor solutions include the ordinary instantons.

Multiplying a spinor by a unit complex number gives a phase action. Set $r=n_D>0$ and impose $\eta =r-1$ complex phase conditions. Starting with a zero-dimensional ordinary count, the expected dimension becomes

$$2r-2(r-1)-1=1.$$

The last subtraction takes the effective phase quotient. This is only an index calculation until the analytic regularization is supplied.

At a regular instanton the manuscript makes the Dirac operator surjective, so its kernel has dimension $r$. A normal spinor link is $S^{2r-1}/\{\pm 1\}$; quotienting by phase gives ${\mathbb{CP}}^{r-1}$. The chosen phase sections have weight two and restrict to $\mathcal{O}(2)$, whose first Chern class is $2h$, with $h$ the hyperplane class.

The resulting oriented link multiplicity is

$$⟨(2h{)}^{r-1},[{\mathbb{CP}}^{r-1}]⟩=2^{r-1}.$$

One can see the number directly. In homogeneous coordinates take equations $z_j^2-a_j{z}_0^2=0$, with $a_j\ne 0$, for $1\le j<r$. Every solution has $z_0\ne 0$, and each affine coordinate has two choices. There are $2^{r-1}$ regular solutions, all with the complex orientation. For $r=1$ there is one point and no phase cuts.

The center quotient does not turn the line into $\mathcal{O}(1)$: it changes the effective circle’s transition functions, while the original weight-two section still has class $2h$. Tracking that distinction is essential to the factor $2^{\eta }\Omega$.

🏷️ Which ends must be excluded

A one-dimensional moduli space can have ends from broken trajectories, reducible fields, or curvature concentrating into bubbles. A boundary argument must account for every such limit. The manuscript establishes three major interfaces.

First, metric periods and curvature estimates bound possible parallel line splittings uniformly for each fixed finite construction. Constants for an old cap are chosen before the number of repetitions; other constants may depend on the whole finite problem. An energy bound concerns total charge and nonnegative ideal losses, not a claim that every component charge is positive.

Second, incidence conditions have exact support, permitting an index budget for lost conditions at particles and separating faces. In a mixed candidate, the necessary projection discards zero-spinor reducible fields, their particle positions, and their tests, but retains their parameter variables. The retained equations must be independent of the discarded fields before a negative projected dimension can exclude the candidate. Arbitrary transversality of the discarded reducers is not assumed.

Third, finite equivariant branch perturbations regularize the surviving problems while preserving the ordinary count. Their rational weights have compatible oriented germs and sum to one where the additional perturbation is zero. A weighted Stokes theorem then applies after compactness and regular end gluing have been established.

The resulting exclusion theorem leaves only legitimate instanton links and a single minimal lens cap attached to one free main component, with no particles or other losses. This theorem is conditional on the regularity and incidence interfaces supplied later in the manuscript; an index inequality by itself is not a compactness proof.

The knot-surgery seams and auxiliary isolation collars remain long but finite. The genuine infinite faces are specified three-sphere cuts and lens boundaries with positive scalar curvature. The proof does not invoke a general spinor Floer theory on the knot surgeries or gluing at positive bubble charge.

🏷️ The oriented boundary contradiction

The manuscript glues exhaustive regular collars at the surviving minimal lens ends. The two determinant sectors there have opposite weighted orientations using the same local sign convention that defined $B$. They therefore cancel pairwise.

Remove small neighborhoods of the finitely many instanton links and truncate these lens collars. The remaining weighted one-dimensional space is compact by the preceding exclusion and exhaustion results. Its total oriented boundary is zero. The instanton links contribute the ordinary signs multiplied by $2^{\eta }$, while the lens pairs contribute zero. Thus

$$0={\sigma }_0{2}^{\eta }\Omega ,{\sigma }_0\in \{1,-1\}.$$

The single sign ${\sigma }_0$ is a common boundary-orientation convention. The equality contradicts the nonzero integer $\Omega$, even though the later branch count is rationally weighted. This excludes the remaining $-2,2$ pair, and the imported reduction gives the claimed theorem.

The selected-source audit covers the introduction and reduction, stated ordinary and negative units, the cap-pairing interface, the finite-lattice repetition proof, selected metric and analytic interfaces, the projective-link calculation, and the final boundary argument. The complete ordinary Floer, symplectic-cap, metric, index-exclusion, compactness, regularization, orientation, and gluing proofs were not independently verified. Their compatibility is the main unresolved verification task for an independent audit of the entire manuscript.

Links

• Family 304 also keeps an integral invariant under control before changing coefficients; here a finite quotient preserves a nonzero cap pairing while the geometric construction grows.
• Family 305 explains another four-dimensional obstruction where finite algebra is useful only after the geometric and analytic input conditions have been retained precisely.

🏷️ Key steps and proof outlines

Repetition preserves an ordinary nonzero count

The construction must satisfy two competing requirements: make the stack long enough for a positive coupled Dirac index, while keeping its ordinary instanton pairing nonzero. The finite-lattice lemma explains why repeating blocks does not destroy that pairing.

Clear the fixed cap denominators to obtain a nonzero integer pairing $q$. It factors through the free quotient $L$ of integral Floer homology, since an integer-valued functional kills torsion. Choose $N$ with $q\not\equiv 0(\mathrm{mod}{2}^N)$.

If a map between finite free chain complexes is a quasi-isomorphism modulo two, its mapping cone is acyclic modulo two. The universal coefficient sequence and Nakayama’s lemma imply that the cone homology vanishes after localization at two. The induced map on the free homology quotients consequently has odd determinant, and the adjugate formula makes it invertible modulo $2^N$. This reasoning allows integral torsion; it does not remove torsion from a mod-two homology group.

The negative and positive return operations therefore give automorphisms of $L/2^{N}L$. Let $E$ be a common exponent for the finite group of all such automorphisms, enlarged to be divisible by eight. The four-interval pattern has operation ${\mathcal{A}}^3\mathcal{D}$. Repeat it a multiple of $E$ times, and choose the end padding lengths as multiples of $E$ too. The entire operation is then the identity modulo $2^N$, so its cap pairing remains nonzero.

The repetition number can be arbitrarily large. Moreover, $E$ depends only on the lattice rank and $N$, so the argument survives changes of the particular operators required by the later small perturbations. The block data must be literally periodic within each repetition: a product of unrelated automorphisms would not be a power of one map. Divisibility by eight also enforces the bundle-degree congruences. This supplies the nonzero-count part of the stack theorem; the separate index telescope supplies positivity of the Dirac index. (OpenAI, 2026)

A long-neck inverse may grow, provided the error decays faster

At a surviving lens end, pregluing the main solution to the minimal cap along a neck of length $D$ produces an approximate solution with residual

$$\Vert {\mathcal{F}}_D(0)\Vert \le C{e}^{-bD}.$$

The proof uses an asymmetric weight $w_D(t)=e^{\delta t}$, with $t$ measured from the main body. At the cap, the outward coordinate is $s=D-t$; after dividing by the cap-core factor $e^{\delta D}$, the weight becomes $e^{-\delta s}$. Thus the two ends see the required opposite completion weights.

On the cylindrical middle, conjugation gives an operator ${\partial }_t+B-\delta$. Choose $\delta >0$ away from the spectrum of the boundary operator $B$. Every spectral mode then has a decaying integral inverse with rate $|\lambda -\delta |$, including the mode $\lambda =0$. The scalar mode is controlled by the weight, rather than discarded. Patch this cylinder inverse to the two completed-piece inverses. First choose wide cutoff transitions to make their commutators small; then increase $D$ to absorb the splicing errors.

After undoing the weights, the inverse need only satisfy

$$\Vert Q_D\Vert \le C{e}^{a\delta D}.$$

For the nonlinear remainder $R_D$, the estimate on a radius-$r$ ball is

$$\Vert R_D(v)-R_D(v^{\prime })\Vert \le C{e}^{a^{\prime }\delta D}r\Vert v-v^{\prime }\Vert .$$

Choose $\delta$ small enough that $(2a+a^{\prime })\delta <b/2$. On a ball of radius comparable to $e^{-(b-a\delta )D}$, the map

$$v⟼-Q_D({\mathcal{F}}_D(0)+R_D(v))$$

is a contraction for large $D$. The exponentially small residual pays for the inverse loss. This proves existence and local uniqueness of a glued solution on each prescribed finite branch.

Exhaustion is needed before Stokes can count the ends

A gluing map alone does not show that every escaping solution is counted by it. The manuscript’s exhaustion argument starts with an arbitrary exact sequence approaching the matching pair. Exponential neck decay allows both halves to be flattened into solutions with exponentially small residuals on the completed pieces. Their inverse estimates improve initially qualitative $o(1)$ closeness to exponential closeness.

The global gauge correction loses at most a polynomial in $D$, which exponential closeness absorbs. The exact solution therefore enters the contraction’s uniqueness ball and must coincide with the constructed glued solution. Branch labels, multiplicities, and rational weights are retained throughout; there is one match per cap crossing, with no additional stabilizer factor.

After the separate compactness and exclusion theorems have ruled out the other ideal limits, these exhaustive collars and the instanton-link neighborhoods leave a compact one-dimensional weighted space. Lens ends cancel in pairs because the determinant-sector flip has exactly the relative orientation sign already used in the ordinary count. The spinor determinant contributes its complex orientation and no new sign.

The remaining boundary contribution is the instanton-link multiplicity times the nonzero ordinary count:

$$0={\sigma }_0{2}^{r-1}\Omega ,{\sigma }_0\in \{1,-1\},\Omega \ne 0.$$

This is the contradiction. It depends on three distinct inputs: preservation of the nonzero count, exclusion of unwanted limits, and exhaustive oriented gluing of the allowed ones. An expected-dimension calculation alone supplies none of these conclusions.

🏷️ Manuscripts and verification scope

• Purely cosmetic surgery on knots in the three-sphere — pinned manuscript and companion source.

No family scope page was present at lean/docs/306.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. Purely cosmetic surgery on knots in the three-sphere.