Baum-Connes Conjecture

Baum-Connes Conjecture

OpenAI's claim

Manuscript 1 (family 285, 2026-09-23): For one finitely generated group without torsion, the coefficient-free reduced assembly map annihilates an infinite-order degree-zero K-homology class; this is claimed to refute rational injectivity. Manuscript 2 (family 285, 2026-09-23): OpenAI claims the existence of a nontrivial scalar projection inside the reduced C*-algebra of a finitely generated torsion-free group, contradicting the Kadison--Kaplansky assertion. Manuscript 3 (family 285, 2026-09-23): An irrational-trace scalar projection is constructed for a finitely generated discrete group, yielding a K_0 class that reduced Baum--Connes assembly with trivial coefficients cannot reach.

Mathematical statement of the claimed result

Mathematical claim 1.

OpenAI claims a discrete group, torsion-free and finitely generated, $G$, an injection $i:\mathbb Z^2\hookrightarrow G$, and $c\in H^2(BG;\mathbb Z)$ such that \[ \langle(Bi)^*c,[T^2]\rangle=1,\qquad i_*\beta=0\text{ in }K_0(C_r^*(G)), \] here $\beta\in K_0(C_r^*(\mathbb Z^2))$ is the Bott generator chosen to have rank zero. The class $h=(Bi)_*[D_{T^2}]\in K_0(BG)$ has infinite order but satisfies $\mu_G^r(h)=0$. Here $\mu_G^r:K_*(BG)\to K_*(C_r^*(G))$ is reduced coefficient-free analytic assembly and $[D_{T^2}]$ is the oriented torus spin-Dirac class. This asserts failure of reduced assembly injectivity without coefficients, even rationally.

Mathematical claim 2.

OpenAI claims a discrete group, torsion-free and finitely generated, $G$ whose reduced group $C^*$-algebra contains a scalar projection $e=e^*=e^2$ with \[ 0<\tau_G(e)<\tfrac12,\qquad\tau_G(x)=\langle x\delta_1,\delta_1\rangle. \] Thus $e$ is neither $0$ nor $1$. This concerns a scalar projection in the operator-norm completion of the left regular group algebra, and supplies a counterexample to the torsion-free projection assertion. The manuscript also uses the integral trace restriction on reduced assembly for torsion-free groups to place this projection class outside the assembly image.

Mathematical claim 3.

For some finitely generated discrete group $G$, OpenAI claims a projection $b_0$ in its reduced group $C^*$-algebra satisfying \[ \tau_G(b_0)\notin\mathbb Q,\qquad [b_0]\notin\mu_0^G\left(K_0^G(\underline EG)\right)\subset K_0(C_r^*(G)). \] Here $\tau_G$ is the canonical trace, $\underline EG$ the universal proper $G$-space and $\mu_0^G$ reduced coefficient-free analytic assembly. The obstruction uses the rational trace of every assembly class. No torsion-free hypothesis is asserted for this group.

Outline of the argument

Manuscript 1 (family 285, 2026-09-23): The construction embeds a torus subgroup with an integral cohomology detector while arranging disappearance of its Bott class analytically. Operators assembled on almost disjoint Cayley-graph sheets retain controlled spectral supports; estimates for the leftover forest permit the required homotopies in the reduced completion. A finite witness then restricts the construction to a finitely generated subgroup.

Manuscript 2 (family 285, 2026-09-23): A magnetic tower model produces a separated spectral band carrying a matrix projection of positive trace below one half. Graph sampling and small-cancellation recovery preserve this band after the model is represented on the group, with intersection losses bounded by trace estimates. A free-product compression converts the matrix projection into a scalar one while retaining torsion-freeness.

Manuscript 3 (family 285, 2026-09-23): Finite-field rules generate graph components on which a group-ring operator has kernels of explicitly counted pattern probabilities. Expander estimates separate those kernels from the remaining spectrum, putting their projection in the reduced C*-algebra rather than merely the von Neumann completion. Its irrational trace conflicts with the rational trace values available to assembly classes.

Formal verification

Family 285 has no Lean scope page in this snapshot; no formalization of these manuscript results is listed.

This site has not executed the Lean files or checked their dependencies. These descriptions report the scope of the linked declarations, rather than a new verification of the proofs.

Status caveat

OpenAI attributes these results to an internal model. Its release notice cautions that results lacking full formalization can contain mistakes. This site has not independently assessed the proofs; a solved label records the scope of the reported claim.

Sources

• OpenAI, A torsion-free counterexample to reduced Baum–Connes injectivity, 2026-09-23; citation key OAI:A-Torsion-Free-Counterexample-to-Reduced-Baum-Connes-Injectivity-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI, A torsion-free counterexample to the Kadison–Kaplansky projection conjecture, 2026-09-23; citation key OAI:A-Torsion-Free-Counterexample-to-the-Kadison-Kaplansky-Projection-Conjecture-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main.
• OpenAI, An irrational-trace counterexample to reduced Baum–Connes, 2026-09-23; citation key OAI:An-Irrational-Trace-Counterexample-to-Reduced-Baum-Connes-September-23-2026. PDF (GitHub); PDF (direct); Manuscript page and citation; Latest version; Statement source: thm:main; Statement source: thm:trace-image.
• OpenAI catalogue entry.

Baum-Connes Conjecture

Definitions and mathematical statement

Let $\underline EG$ be a classifying space for proper $G$-actions: fixed-point sets for compact subgroups are contractible and those for noncompact subgroups are empty. Topological $K$-theory here is the compact-support/direct-limit equivariant $K$-homology of $\underline EG$. The reduced group $C^*$-algebra $C_r^*(G)$ is the operator-norm completion of convolution operators in the left regular representation. The assembly map is \[ \mu_G:K_*^{\rm top}(G)\longrightarrow K_*(C_r^*(G)),\qquad *=0,1. \] The conjecture says it is an isomorphism for every second-countable locally compact $G$. This is the coefficient-free target; adding arbitrary coefficient algebras changes the conjecture.

Formulation and concept references: [S1].

Short English statement

Does geometric equivariant $K$-homology compute the operator-algebraic $K$-theory of every group in the stated class?

Sources

• S1 W. Lück, H. Reich, in Handbook of K-theory, Springer, 2005. https://arxiv.org/abs/math/0402405. Catalogue formulation source.
• S2 There are no known counter-examples to the Baum-Connes conjecture. https://www.fields.utoronto.ca/talks/There-are-no-known-counter-examples-to-Baum-Connes-conjecture. Catalogue research/status reference; not a proof endorsement.
• S3 Expanders and K-theory for group C* algebras. https://www.fields.utoronto.ca/talk-media/2/78/29/slides.pdf. Catalogue research/status reference; not a proof endorsement.
• S4 Controlled Analytic Properties and the Quantitative Baum-Connes Conjecture. https://arxiv.org/abs/1908.02131. Catalogue research/status reference; not a proof endorsement.
• S5 The Baum-Connes conjecture for extensions. https://arxiv.org/abs/2508.05726. Catalogue research/status reference; not a proof endorsement.
• E1 Shintaro Nishikawa, Crossed product approach to equivariant localization algebras, arXiv:2108.11235v2, Theorem A, Definition 5.2 and Corollary 13.4; see also the introduction on gamma elements. https://arxiv.org/abs/2108.11235v2. Added research source: Localization/crossed-product formulation for second-countable locally compact groups, including the kernel criterion and the semisplit evaluation sequence. Consulted 24 September 2026.
• E2 Martin Finn-Sell, Controlled Analytic Properties and the Quantitative Baum--Connes Conjecture, arXiv:1908.02131v2, author withdrawal notice dated 12 May 2026. https://arxiv.org/abs/1908.02131v2. Added research source: Withdrawn: the notice reports a missing norm-control argument in the K-theory component. Cited for source status only, not as an established quantitative theorem. Consulted 24 September 2026.
• E3 Ralf Meyer, The Baum--Connes conjecture for extensions, arXiv:2508.05726v2, 29 August 2025, two-page note. https://arxiv.org/abs/2508.05726v2. Added research source: Counterexample to an overstrong extension-permanence assertion with coefficients. This is not a counterexample to the coefficient-free target. Consulted 24 September 2026.

Baum-Connes Conjecture

Definitions and mathematical statement

Let $\underline EG$ be a classifying space for proper $G$-actions: fixed-point sets for compact subgroups are contractible and those for noncompact subgroups are empty. Topological $K$-theory here is the compact-support/direct-limit equivariant $K$-homology of $\underline EG$. The reduced group $C^*$-algebra $C_r^*(G)$ is the operator-norm completion of convolution operators in the left regular representation. The assembly map is \[ \mu_G:K_*^{\rm top}(G)\longrightarrow K_*(C_r^*(G)),\qquad *=0,1. \] The conjecture says it is an isomorphism for every second-countable locally compact $G$. This is the coefficient-free target; adding arbitrary coefficient algebras changes the conjecture.

Formulation and concept references: [S1].

Short English statement

Does geometric equivariant $K$-homology compute the operator-algebraic $K$-theory of every group in the stated class?

Research attempt

Current frontier and source audit.

The non-discrete formulation was checked against Section 4.1.6 of Lück--Reich. Their Sections 5.1.1--5.1.2 record the a-T-menable and almost-connected cases, among other positive results. The target here still quantifies over every second-countable locally compact group and uses only the coefficient algebra ${\mathbb{C}}$. [S1] Nishikawa's Theorem A and Corollary 13.4 provide a controlled-algebra reformulation for this full locally compact scope, not just for discrete groups. [E1]

Editorial status note. The paper listed as [S4] was withdrawn in arXiv version 2 on 12 May 2026. Its author reports a missing norm-control argument in the $K$-theory component. Its advertised conclusion is therefore not used as an established theorem here. [S4], [E2] Meyer's 2025 note concerns the hypotheses of an extension permanence theorem with coefficients. It disproves an attempted simplification of those hypotheses, not the coefficient-free conjecture in this section. [S5], [E3]

Attack route.

Three approaches lead to different missing statements. In the Dirac--dual-Dirac approach, an appropriate gamma element gives a split injection, but surjectivity still requires that its descended action be the identity on $K_*(C_r^*(G))$. Requiring a gamma element equal to the identity in equivariant $KK$-theory for every group would be stronger: it would give arbitrary coefficients as well. One cannot silently add that stronger requirement, in view of coefficient counterexamples. [E1], [E3]

An induction using group extensions would need the full finite-subgroup preimage conditions in the appropriate permanence theorem; the shortcut of testing only a normal subgroup is specifically unsafe. [E3] The route tested below is instead to regularize representatives of $K$-classes until geometric control becomes small, then recover exact projections or unitaries by functional calculus. Its missing ingredient is quantitative: the spectral gap needed for the recovery must not collapse during regularization. This is tested by an exact example, not by an assumption that norm-small coefficients imply controlled $K$-theory.

Attempt: identify the obstruction algebra in the full scope.

Here is the established bridge used, with coefficients specialized to ${\mathbb{C}}$. For a proper $G$-space $X$ and a universal $X$-$G$ Hilbert module $H_X$ in Nishikawa's Definition 5.2, $RL_c^*(H_X)$ is the completion of bounded, norm-continuous, $G$-continuous compact-operator-valued paths $T_t$, $t\ge1$, with uniformly compact support and $\|[T_t,\phi]\|\to0$ for $\phi\in C_0(X)$. Let $RL_c^0(H_X)$ be its evaluation-at-one kernel. Set \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-6b-65-72-6e-65-6c} I_G=RL_c^0(H_{\underline EG})\rtimes_r G. \end{equation} For non-locally-compact models of $\underline EG$, use the source's direct-limit convention over $G$-compact proper subspaces and compatible modules. Corollary 13.4 identifies the required condition as \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-6f-62-73-74-72-75-63-74-69-6f-6e} \mu_G\text{ is an isomorphism} \quad\Longleftrightarrow\quad K_0(I_G)=K_1(I_G)=0. \end{equation} This is a known reformulation, not a newly proved reduction. The semisplit evaluation sequence is what permits the reduced-crossed-product argument; exactness of reduced crossed products for arbitrary sequences is not being assumed. [E1]

A concrete sufficient analytic programme is now visible. For every projection $p$ in a matrix unitization of $I_G$, with scalar part $e$, construct a norm-continuous path $a_s$, $s\in[0,\infty)$, in the same matrix unitization, satisfying \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-70-72-6f-67-72-61-6d-30} \begin{gathered} a_0=p,\qquad a_s=a_s^*,\qquad a_s-e\in M_N(I_G),\qquad \lim_{s\to\infty}\|a_s-e\|=0,\\ \sup_{s\ge0}\|a_s^2-a_s\|\le\varepsilon_0<\tfrac14. \end{gathered} \end{equation} Stabilization is allowed. For every unitary $u$ with scalar part one, seek norm-continuous paths $b_s$, starting at $u$, tending in norm to one, and satisfying \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-70-72-6f-67-72-61-6d-31} b_s-1\in M_N(I_G),\qquad \sup_s\max\{\|b_s^*b_s-1\|,\|b_sb_s^*-1\|\} \le\varepsilon_1<1. \end{equation} The $s$ here is a homotopy variable. It is different from the localization variable $t$ hidden in $I_G$; controlling each separately is not enough unless the resulting paths are continuous in the algebra norm.

Attempt: prove the functional-calculus step.

Lemma 53.1 (uniform spectral recovery). In a unital $C^*$-algebra, a self-adjoint element $a$ with $\|a^2-a\|\le\varepsilon<1/4$ has spectrum separated from $1/2$ by at least $\sqrt{1-4\varepsilon}/2$. Its spectral projection $p(a)=\mathbf1_{(1/2,\infty)}(a)$ is well defined by continuous functional calculus on its spectrum and \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-72-65-63-6f-76-65-72-79} \|p(a)-a\|\le r_\varepsilon:=\frac{1-\sqrt{1-4\varepsilon}}2. \end{equation} The assignment is norm-continuous on any norm-bounded family with the same error bound. It respects a quotient in which $a$ maps to a projection.

Proof. The spectral theorem implies $|\lambda^2-\lambda|\le\varepsilon$ for every spectral value $\lambda$. Inside $[0,1]$ this excludes the interval $(r_\varepsilon,1-r_\varepsilon)$; the possible excursions below zero or above one have size $(\sqrt{1+4\varepsilon}-1)/2\le r_\varepsilon$. This proves the spectral gap and the distance bound. Choose a single continuous function equal to zero on the lower allowed band and one on the upper band. Its functional calculus gives $p(a)$ for every element of the family. Uniform polynomial approximation of this function on a common compact interval proves norm-continuity. Functoriality under a quotient, and the values zero and one at those scalars, prove the last assertion. $\square$

The analogous recovery of a unitary is \[ v(b)=b(b^*b)^{-1/2}. \] Both inequalities in (ra:1479:program1) ensure invertibility; the uniform lower bound $b^*b\ge(1-\varepsilon_1)1$ makes inverse square roots continuous with uniform control. These statements follow from scalar continuous functional calculus and do not require an assembly theorem.

Applying Lemma 53.1 to (ra:1479:program0) would turn $a_s$ into a projection homotopy from $p$ to $e$, with the endpoint at infinity supplied by norm convergence. Applying the polar construction to (ra:1479:program1) would give a unitary homotopy from $u$ to one. Thus the proposed programme would kill both $K$-groups in (ra:1479:obstruction). This implication is proved; the paths in (ra:1479:program0)--(ra:1479:program1) have not been constructed.

In a discrete equivariant Roe/localization model there is a useful additional check. Products of $k$ finite-propagation operators have propagation at most the sum of their propagations. Hence a fixed-degree polynomial in a path whose propagation tends to zero again has that property. The uniform polynomial approximation in the proof above therefore keeps spectral recovery inside the completed localization algebra. It does not show that an exact recovered projection itself has literal finite propagation. Confusing the dense controlled subalgebra with its norm completion would be an error.

Attempt: test a natural damping regularization.

Consider already $G={\mathbb{Z}}$, so $C_r^*(G)\cong C(S^1)$, and write $u(z)=z$. For $s\ge0$, the Poisson averaging operator satisfies \[ M_s(u^k)=e^{-s|k|}u^k. \] It is unital and completely positive: for $s>0$ it is averaging of rotations against the nonnegative Poisson kernel of radius $e^{-s}$, whose integral is one, and the same averaging works entrywise on matrices. The identity case is $s=0$. It therefore looks attractive as a regularizer that suppresses nontrivial group coefficients.

The following calculation breaks that proposal's spectral-control step. In $M_2(C(S^1))$ put \[ p=\frac12\begin{pmatrix}1&u\\u^*&1\end{pmatrix},\qquad a_s=M_s^{(2)}(p)=\frac12 \begin{pmatrix}1&e^{-s}u\\e^{-s}u^*&1\end{pmatrix}. \] Direct multiplication, using $u^*u=uu^*=1$, gives \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-64-61-6d-70-69-6e-67} p^2=p,\qquad a_s^2-a_s=-\frac{1-e^{-2s}}4 I_2, \qquad \operatorname{spec}(a_s)= \left\{\frac{1-e^{-s}}2,\frac{1+e^{-s}}2\right\}. \end{equation} For each finite $s$ the idempotence error is strictly below $1/4$. Nevertheless its supremum is $1/4$, and the spectral gap is $e^{-s}$, which tends to zero. Functional calculus does not retain the coefficient suppression: \begin{equation}\label{wsref-0-72-61-3a-31-34-37-39-3a-75-6e-64-61-6d-70-69-6e-67} \mathbf1_{(1/2,\infty)}(a_s)=p\qquad(s<\infty). \end{equation} The exact spectral projection restores the original off-diagonal $u$. Moreover $a_s\to I_2/2$, which is not a projection. For $K_1$ the same phenomenon is even simpler: $M_s(u)=e^{-s}u$ is invertible at every finite $s$, but its polar unitary is always $u$, while the inverse norm grows as $e^s$.

This is not a counterexample to Baum--Connes, nor even a nonzero obstruction class. Indeed the displayed projection is explicitly homotopic to the constant rank-one projection through \[ p_\theta= \begin{pmatrix} \cos^2\theta&\cos\theta\sin\theta\,u\\ \cos\theta\sin\theta\,u^*&\sin^2\theta \end{pmatrix},\qquad 0\le\theta\le\pi/4. \] Every $p_\theta$ is a projection, $p_0=\operatorname{diag}(1,0)$ and $p_{\pi/4}=p$. The example rejects the specific coefficient-damping argument, even in a group for which the conjecture is established; it does not reject the existence of a different geometric homotopy.

Stress test.

There are three distinct uniformities. Spectral separation must be uniform along the proposed homotopy. Localization estimates must hold in the completed algebra, rather than separately for each $t$. Finally, a direct-limit argument may enlarge the proper subspace for an individual class, but cannot let that subspace run off to infinity during a purported homotopy without proving that the homotopy represents a class in the required direct limit. None is supplied by pointwise decay of matrix coefficients.

The exact example (ra:1479:damping) is a counterexample to the intermediate claim that pointwise almost-idempotence below $1/4$ is sufficient for uniform spectral recovery during regularization. Its failure is independent of numerical accuracy. The withdrawn norm-control argument in [S4] is a separate source warning, not evidence that the calculation here reproduces that paper's particular missing step. [E2]

A proof only lifting $K$-classes to an assembly image would address surjectivity, not injectivity. The obstruction-algebra formulation and both proposed contractions were included to avoid that omission. Likewise, calculations with ${\mathbb{Z}}$ test a mechanism but do not reduce all locally compact groups to discrete ones. The general formulation remains (ra:1479:kernel)--(ra:1479:obstruction).

Outcome.

Outcome: Exploratory approach with unresolved gap.

The functional-calculus recovery lemma and the explicit damping failure were proved. Together they identify a precise defect in one possible controlled-algebra attack: spectral flattening can erase the very gap needed to pass back from approximate to exact $K$-theory representatives. The known localization equivalence is used as a framework, not presented as new progress.

The unresolved step is the construction, for every group in the original scope, of geometric homotopies with genuinely uniform norm and spectral control, or another method that kills the same obstruction classes. No such construction, universal assembly inverse, or coefficient-free counterexample was obtained. In particular, neither an extension claim with altered hypotheses nor a withdrawn quantitative theorem is used to fill that gap.

Sources

• S1 W. Lück, H. Reich, in Handbook of K-theory, Springer, 2005. https://arxiv.org/abs/math/0402405. Catalogue formulation source.
• S2 There are no known counter-examples to the Baum-Connes conjecture. https://www.fields.utoronto.ca/talks/There-are-no-known-counter-examples-to-Baum-Connes-conjecture. Catalogue research/status reference; not a proof endorsement.
• S3 Expanders and K-theory for group C* algebras. https://www.fields.utoronto.ca/talk-media/2/78/29/slides.pdf. Catalogue research/status reference; not a proof endorsement.
• S4 Controlled Analytic Properties and the Quantitative Baum-Connes Conjecture. https://arxiv.org/abs/1908.02131. Catalogue research/status reference; not a proof endorsement.
• S5 The Baum-Connes conjecture for extensions. https://arxiv.org/abs/2508.05726. Catalogue research/status reference; not a proof endorsement.
• E1 Shintaro Nishikawa, Crossed product approach to equivariant localization algebras, arXiv:2108.11235v2, Theorem A, Definition 5.2 and Corollary 13.4; see also the introduction on gamma elements. https://arxiv.org/abs/2108.11235v2. Added research source: Localization/crossed-product formulation for second-countable locally compact groups, including the kernel criterion and the semisplit evaluation sequence. Consulted 24 September 2026.
• E2 Martin Finn-Sell, Controlled Analytic Properties and the Quantitative Baum--Connes Conjecture, arXiv:1908.02131v2, author withdrawal notice dated 12 May 2026. https://arxiv.org/abs/1908.02131v2. Added research source: Withdrawn: the notice reports a missing norm-control argument in the K-theory component. Cited for source status only, not as an established quantitative theorem. Consulted 24 September 2026.
• E3 Ralf Meyer, The Baum--Connes conjecture for extensions, arXiv:2508.05726v2, 29 August 2025, two-page note. https://arxiv.org/abs/2508.05726v2. Added research source: Counterexample to an overstrong extension-permanence assertion with coefficients. This is not a counterexample to the coefficient-free target. Consulted 24 September 2026.