Kervaire Conjecture: How OpenAI Proved One Relation Cannot Erase a Group

“Can one new generator and one relation erase an entire group?” The Kervaire conjecture says no. Start with any nontrivial group, adjoin a generator, and impose one relation. The resulting group still has an element other than the identity. OpenAI’s September 24, 2026 manuscript proves something stronger in the difficult case. When the new generator occurs with total exponent +1+1 or −1-1, the original group embeds in the quotient, so every original element survives with its identity and multiplication intact.

The proof starts with a group element that supposedly vanishes and turns that equality into a planar surface. A trace built from unitary spectral angles then shows that the surface cannot have exactly one nontrivial boundary word.

1. What Is the Kervaire Conjecture in Group Theory?

Let AA be the coefficient group and tt the new generator. The free product H=A*⟨t⟩H=A*\langle t\rangle keeps all the laws of AA and imposes no relation between AA and the infinite cyclic group generated by tt. Pick a word ww in HH. Imposing w=1w=1 kills every conjugate of ww and w−1w^{-1}, and every finite product of those conjugates. They form the normal closure N=⟨⟨w⟩⟩HN={\langle\!\langle w\rangle\!\rangle}_H. The quotient Q=H/NQ=H\mathbin{/}N is the group obtained by imposing the relation.

The Kervaire Result at a Glance
QuestionAnswer in the September Manuscript
Starting groupAny group, including infinite groups and groups with torsion.
ConstructionAdjoin one generator and impose one relation.
NontrivialityEvery relator leaves a nontrivial quotient if the starting group is nontrivial.
EmbeddingThe original group embeds when the exponent sum is plus or minus one.

A presentation specifies generators and the relations they satisfy. For a small example, take AA to be the cyclic group of order three, generated by aa, and impose ta−1=1ta^{-1}=1. The new relation gives t=at=a. Substituting this into the presentation leaves the original relation a3=1a^3=1, so QQ is again cyclic of order three.

2. Why Is the Exponent Sum ±1 the Difficult Case?

	Kervaire conjecture diagram showing when the quotient maps onto a cyclic group and when the starting group embeds.
Kervaire conjecture diagram showing when the quotient maps onto a cyclic group and when the starting group embeds.

The homomorphism p:H→ℤp:H\to\mathbb Z sends every coefficient to zero and sends tt to one. Thus d=p(w)d=p(w) counts occurrences of tt minus occurrences of t−1t^{-1} after powers are split into individual letters. The composite map to ℤ/dℤ\mathbb Z\mathbin{/}d\mathbb Z kills ww and therefore descends to a surjection from QQ. For d=0d=0 the target is ℤ\mathbb Z. For |d|≥2|d|\ge 2 it is a finite cyclic group with at least two elements.

When d=±1d=\pm1, the cyclic target is trivial. These are the unimodular equations over groups. OpenAI proves survival in this case by embedding the coefficient group itself. For comparison, imposing a=1a=1 in the order-three example kills the entire coefficient group, while the free generator tt still gives an infinite cyclic quotient.

3. What Exactly Does OpenAI’s Kervaire Theorem Prove?

Theorem 1.1 of the Kervaire theorem for groups asserts that for every group AA and every word w∈A*⟨t⟩w\in A*\langle t\rangle with p(w)=±1p(w)=\pm1, the following canonical homomorphism is injective.

𝑨⟶(𝑨∗⟨𝒕⟩)/⟨⟨𝒘⟩⟩𝑯\boldsymbol{A\longrightarrow (A*\langle t\rangle)\mathbin{/}{\langle\!\langle w\rangle\!\rangle}_H}

Corollary 1.2 says that if A≠1A\ne1, then Q≠1Q\ne1 for every ww. It combines the embedding at exponent sum ±1\pm1 with the cyclic quotient at all other sums. Coefficient injectivity also solves the equation in an overgroup. Take QQ as that overgroup and the image of tt as the solution. The theorem permits the solution to lie outside AA.

4. How Far Did Earlier Approaches Reach?

The problem grew out of Kervaire’s characterization of high-dimensional knot groups, where normal generation by one element is one of the defining conditions. The group-theoretic question asks whether a free product with a new infinite cyclic factor can have that property. Gerstenhaber and Rothaus established solvability of nonsingular equations over finite groups in finite overgroups.

For a residually finite coefficient group, each nonidentity element stays nonidentity in some finite quotient, so the finite result prevents that element from vanishing in the one-relator quotient. Hyperlinear methods extended nonsingular solvability to groups that embed in metric ultraproducts of unitary groups. Klyachko proved unimodular injectivity for torsion-free groups, and Lvzhou Chen gave a proof in that setting using estimates on the complexity of surfaces.

The September manuscript also discusses Akio Kawauchi’s published proposed resolution via knot exteriors and ribbon sphere-link groups. OpenAI gives an independent argument using a faithful trace from the regular representation of every group. Its applicability covers coefficient groups with torsion, regardless of their finite quotients or approximation properties. The proof also accommodates infinite presentations, since any proposed vanishing coefficient can be witnessed by a finite product of conjugates.

5. How the Proof Turns a Vanishing Group Element Into a Planar Diagram

Suppose g∈Ag\in A maps to the identity in QQ. Membership in NN means there are finitely many conjugating words bj∈Hb_j\in H and signs εj∈{1,−1}\varepsilon_j\in\{1,-1\} with

𝒈=∏𝒋=1𝒓𝒃𝒋𝒘𝜺𝒋𝒃𝒋−1.\boldsymbol{g=\prod_{j=1}^{r}b_jw^{\varepsilon_j}b_j^{-1}.}

Picture a disk with one hole per factor. Its outer circle reads gg, each hole reads ww or w−1w^{-1} in the orientation of its missing disk, and the paths from the outer basepoint to the holes read the conjugating words. Mark each tt-letter as a slot. Disjoint bands pair slots with opposite traversal signs. To index the slots uniformly, cyclically conjugate the relator so it starts with a tt-letter, then split powers into single letters. This gives w=ts1a1⋯tsnanw=t^{s_1}a_1\cdots t^{s_n}a_n, with si∈{1,−1}s_i\in\{1,-1\} and coefficients ai∈Aa_i\in A that may be the identity. A negative disk reverses the reading, inverting each letter and each corner label. The pieces between slots carry the original coefficients, and following the edge of a disk-and-band component reads a word entirely in AA. Section 10 explains how to obtain the bands from the given equality.

6. How Spectral Phase Bounds a Product

	Kervaire conjecture phase diagram showing two three-quarter turns wrapping to a half turn and the spectral-phase bound.
Kervaire conjecture phase diagram showing two three-quarter turns wrapping to a half turn and the spectral-phase bound.

Use the Hilbert space ℓ2(A)\ell^2(A) with basis vectors δx\delta_x indexed by group elements. Left translation acts by L(a)δx=δaxL(a)\delta_x=\delta_{ax} and commutes with right translations. On the algebra of operators commuting with all right translations, the trace τ\tau takes the diagonal entry at δ1\delta_1. It is faithful because a positive operator with zero trace has a square root that kills δ1\delta_1 and, by translation, every other basis vector. This group von Neumann algebra works for arbitrary AA. The trace τb\tau_b on bb-by-bb operator matrices sums the traces of the diagonal entries, giving τb(I)=b\tau_b(I)=b.

For a unitary UU, take the fractional phase h(e2πix)=x−⌊x⌋h(e^{2\pi ix})=x-\lfloor x\rfloor, measured in turns. Set ℓ(U)=τb(h(U))\ell(U)=\tau_b(h(U)) using spectral calculus. The cut is at 11, where h(1)=0h(1)=0. Because the phase operator is positive and the trace faithful, ℓ(U)\ell(U) vanishes exactly when U=IU=I. Lemma 2.1 supplies the estimate used to compare two ways of following a surface.

ℓ(𝑼𝑽)≤ℓ(𝑼)+ℓ(𝑽).\boldsymbol{\ell(UV)\le \ell(U)+\ell(V).}

Write B=h(V)B=h(V) and follow the unitary path Wu=Ue2πiuBW_u=Ue^{2\pi i uB} from UU to UVUV. For smooth ff on the circle, differentiating the trace along this path gives τb(f′(Wu)B)\tau_b(f'(W_u)B), with the derivative taken in turns. The formula follows for powers of WuW_u from the product rule and cyclicity of the trace, and extends by Fourier expansion. If f′≤1f’\le1, positivity bounds the derivative by τb(B)\tau_b(B).

The fractional phase itself jumps at the cut. Averaging small positive shifts smooths it while keeping the derivative at most one, since a forward move increases fractional phase by at most the length of the move. The smoothed function converges to zero at the cut as well as to the correct value elsewhere. In the scalar example U=V=e2πi(3/4)U=V=e^{2\pi i(3/4)}, the two phases add to 32\frac{3}{2} turns, while the product’s phase is 12\frac{1}{2} turn. In the operator argument the unitary path and trace give the inequality even when the two operators fail to commute. Integrating the smooth estimate and taking the limit under the finite trace gives the inequality.

7. Where Topology Enters the Fixed-Space Lemma

Write kk for the number of disks of each sign. If there are nn slots on a relator, their signs sis_i sum to ±1\pm1. Lemma 3.1 says that for any P∈U(nk)P\in U(nk), a single matrix X∈U(k)X\in U(k) can mix the disk copies so that

𝑭(𝑿)=diag⁡(𝑿𝒔1,…,𝑿𝒔𝒏)𝑷\boldsymbol{F(X)=\operatorname{diag}(X^{s_1},\ldots,X^{s_n})P}

fixes a complex subspace of dimension at least kk. Here U(j)U(j) is the group of jj-by-jj unitary matrices.

The argument keeps a unitary WW together with a complex kk-plane EE that it fixes pointwise. These pairs form a compact incidence manifold. Over each EE, the action on its orthogonal complement is an arbitrary unitary. For m=nkm=nk the incidence space has real dimension m2−k2m^2-k^2. Adding the variable XX gives a domain of dimension m2m^2, matching the target U(m)U(m) for the map taking (X,W,E)(X,W,E) to F(X)W−1F(X)W^{-1}.

A homotopy first moves PP to the identity. Rotating the two coordinate planes in each positive and negative pair cancels the corresponding blocks of XX and X−1X^{-1}. Their constant blocks are then moved to minus the identity. One unpaired block with sign ε=∑isi\varepsilon=\sum_i s_i remains. At that endpoint the map hits the identity only when X=IX=I, EE is the unpaired plane and WW is the resulting endpoint matrix.

Variations of XX, the plane and the complementary unitary give independent tangent directions covering the target, so this preimage is regular. The unpaired plane is singled out because minus the identity fixes no vector on its complement other than zero. The regular fiber therefore has odd cardinality. Homotopy preserves that mod-two degree, so the original map also hits the identity, where F(X)=WF(X)=W fixes EE.

8. Why a Planar Surface Cannot Have Exactly One Nontrivial Boundary

	Kervaire conjecture diagram of two word disks joined by three bands, forming three boundary circles with trivial words.
Kervaire conjecture diagram of two word disks joined by three bands, forming three boundary circles with trivial words.

Label a corner by cd∈Ac_d\in A when it follows slot dd on the same disk. Let σ\sigma advance along a disk to its next slot, and let θ\theta exchange the two slots of a band. Each boundary step crosses the band first, then passes the corner cθdc_{\theta d} before reaching σθd\sigma\theta d. The cycles of σθ\sigma\theta are therefore the boundary circles. Multiply the corner labels along a cycle zz to obtain its boundary word rzr_z. Moving the starting point conjugates the word and preserves whether it is the identity.

Each positive disk contributes the relator’s exponent sum, and each negative disk contributes its opposite. Every band cancels a pair of traversal signs, so the counts of positive and negative disks match. The connected surface has 2k2k disks and nknk bands. Since it has genus zero, its Euler characteristic gives the boundary count qq.

2−𝒒=2𝒌−𝒏𝒌,𝒒=𝒏𝒌−2𝒌+2.\boldsymbol{2-q=2k-nk,\qquad q=nk-2k+2.}

For example, join one positive three-slot disk to its negative copy by matching slot indices. The opposite cyclic orders give a planar attachment with three boundary circles. In each circle, a negative corner contributes ai−1−1a_{i-1}^{-1} and its positive partner contributes ai−1a_{i-1}, so the two labels multiply to the identity. Theorem 4.1 says that if all but one boundary word are trivial, that last word must be trivial too. The count is what lets the phase comparison isolate its contribution.

9. The Trace Comparison That Forces the Boundary Word to Vanish

The unitary SS advances around the disks with weights L(cd−1)L(c_d^{-1}). The unitary involution Θ\Theta crosses the bands with opposite scalar phases at the two ends. Their product SΘS\Theta therefore travels around the boundary cycles. To bound its phase, insert a scalar involution JJ that exchanges positive and negative disk arrays via XX and X−1X^{-1}. The corner conventions give JSJ=S−1JSJ=S^{-1}. Thus SJSJ is a self-adjoint unitary with trace zero: it interchanges disk kinds. Its negative spectral projection has trace nknk, giving ℓ(SJ)=nk2\ell(SJ)=\frac{nk}{2}. The fixed-space estimate acts on the finite slot coordinates, while the coefficient weights continue to act on the Hilbert space of the original group.

The fixed-space lemma chooses XX so that JΘJ\Theta fixes at least kk directions on the positive-traversal slots. Conjugation by JJ exchanges traversal signs and sends JΘJ\Theta to its inverse, giving at least kk fixed directions on the other half. Away from eigenvalue 11, inverse eigenvalues contribute one turn per pair. Hence ℓ(JΘ)≤nk−k\ell(J\Theta)\le nk-k, and subadditivity gives

ℓ(𝑺𝚯)≤32𝒏𝒌−𝒌.\boldsymbol{\ell(S\Theta)\le \frac{3}{2}nk-k.}

For the second calculation, a boundary cycle of length dzd_z contributes dz−12\frac{d_z-1}{2} plus the trace of the fractional phase of its circuit weight e2πiηzL(rz−1)e^{2\pi i\eta_z}L(r_z^{-1}). Here ηz\eta_z is its total scalar phase. Rephasing the slot coordinates makes the cycle operator unitarily conjugate to its rotation by a dzd_zth root of unity. Summing fractional phases over these rotations produces the constant dz−12\frac{d_z-1}{2} and leaves the phase of one complete circuit.

The band phases realize any boundary phases of total zero by routing a flow along a spanning tree of the connected graph of boundary circles. Put phase −ϵ-\epsilon for ϵ>0\epsilon>0 on each trivial circle and the compensating nonnegative phase on the exceptional circle z0z_0. The trivial circuits approach the cut from below and contribute one each. The exceptional circuit retains its nonnegative phase in the limit. With just one boundary circle its phase shift is zero throughout. The length terms sum to 2nk−q2\frac{2nk-q}{2}, because the cycles partition all 2nk2nk slots. The q−1q-1 trivial circuits contribute another q−1q-1 in the limit. Substituting the Euler identity makes the phase limit

lim𝝐↓0⁡ℓ(𝑺𝚯)=32𝒏𝒌−𝒌+𝝉(𝒉(𝑳(𝒓𝒛0−1))).\boldsymbol{\lim_{\epsilon\downarrow 0}\ell(S\Theta)=\frac{3}{2}nk-k+\tau\bigl(h(L(r_{z_0}^{-1}))\bigr).}

The upper bound holds throughout the limiting process, so the extra term must be zero. Faithfulness then implies L(rz0−1)=IL(r_{z_0}^{-1})=I. Applying that translation to δ1\delta_1 shows rz0=1r_{z_0}=1.

10. Completing the Proof From Planar Surfaces to Group Injectivity

Realize the punctured disk as a map into a presentation space for AA with one circle wedged on for tt. Project onto the circle. The outer circle projects to the wedge point, and the coefficient intervals do too. A piecewise-affine approximation that agrees with the boundary lets the authors choose a level away from vertex values and the wedge point. Each tt-letter meets that level once. The level arcs pair opposite traversal signs. Thin bands around the arcs stay where the original map takes values in the added circle, so their projection into the coefficient space is constant. The intervening corner labels retain the original coefficients.

Now restore the missing disks and consider the components of the disks and bands. Choose an innermost component, whose interior complementary disks contain no other component. The coefficient map already fills those inner circles, so their words are trivial. The planar boundary theorem makes the exterior word trivial too. Fill that exterior loop in the coefficient space and remove the component. The replacement agrees with the old map on the exterior boundary. The other components retain their labels, and repeating the step eventually fills the original outer loop in the coefficient space. Thus g=1g=1 already in AA.

11. What the OpenAI Proof Establishes and What Lean Verifies

The September manuscript proves nontriviality for arbitrary relators and coefficient injectivity for unimodular relators. The broader Kervaire–Laudenbach question asks for injectivity at every nonzero exponent sum. A separate October companion preprint proves nonsingular solvability in overgroups, including that broader one-variable case.

The selected Lean comparator for the September result checks the canonical map’s injectivity for arbitrary coefficient groups at exponent sum ±1\pm1. The Lean implementation also proves the all-relator nontriviality corollary by splitting off the cyclic-quotient case. OpenAI’s October 6 release announcement describes the broader collection, while the manuscript and Lean files give the statements specific to Kervaire.

The Kervaire conjecture proof shows how planar topology can keep a newly imposed relation from erasing an original element: if a nonidentity coefficient were erased, the trace would require more phase than the surface allows. Follow Binary Verse AI for more proof walkthroughs in group theory and formal mathematics, including the work on Thompson’s group F and Lean formalization of Fermat’s Last Theorem.

What is the Kervaire conjecture in group theory?

The Kervaire conjecture states that adjoining one new generator and imposing one relation on a nontrivial group cannot make the resulting group trivial. Equivalently, no free product of a nontrivial group with an infinite cyclic group can be normally generated by a single element.

Has the Kervaire conjecture been solved?

Yes. OpenAI’s September 2026 proof establishes the conjecture for arbitrary nontrivial groups. Its main theorem proves that the original group embeds into the one-relator quotient whenever the new generator has exponent sum plus or minus one.

Why is exponent sum ±1 important in the Kervaire proof?

For every other exponent sum, the quotient maps onto a nontrivial cyclic group. At exponent sum plus or minus one that cyclic target is trivial, so the proof instead establishes coefficient-group injectivity using spectral phase, unitary matrices and planar topology.

Does the Kervaire theorem apply to non-sofic groups and groups with torsion?

Yes. The theorem imposes no restrictions on torsion, cardinality or presentation. Its proof uses the left regular representation and a faithful trace available for every group, without requiring soficity or hyperlinearity.

Does imposing one relation always preserve the original group?

Preservation can fail when the exponent sum is zero. For example, with an order-three coefficient group, imposing that its generator equals the identity erases the coefficients and leaves the new generator free. The September theorem guarantees an embedding for unimodular relations, and the October companion extends that guarantee to all nonzero exponent sums.

Is the Kervaire conjecture the same as the Kervaire–Laudenbach conjecture?

The Kervaire conjecture asks whether the resulting group remains nontrivial. Kervaire–Laudenbach asks for the original group to embed whenever the exponent sum is nonzero. The September paper proves injectivity in the unimodular case and resolves the original conjecture. The October companion covers the broader nonsingular case.

What part of OpenAI’s Kervaire proof is formally verified in Lean?

The selected Lean comparator checks coefficient-group injectivity for arbitrary groups when the relator has exponent sum plus or minus one in the new generator. The accompanying implementation also proves the nontriviality corollary for every relator, using the separate cyclic-quotient argument for the other exponent sums.

Leave a Comment