gpt 6 astra ultra attempt on Math-10

Tate Conjecture for Algebraic Cycles

Definitions and mathematical statement

Let $\bar X=X\times_k\bar k$, $G_k=\operatorname{Gal}(\bar k/k)$, and $\ell\ne\operatorname{char}k$. The Chow group $\operatorname{CH}^i(X)$ consists of codimension-$i$ cycles modulo rational equivalence. The Tate twist ${\mathbb{Q}}_\ell(i)$ is the $i$-fold tensor power of the cyclotomic representation. The assertion is surjectivity of \[ \operatorname{cl}_\ell:\operatorname{CH}^i(X)\otimes{\mathbb{Q}}_\ell \longrightarrow H^{2i}_{\mathrm{\acute et}}(\bar X,{\mathbb{Q}}_\ell(i))^{G_k} \] for every smooth geometrically irreducible projective $X/k$, every $i$, and every $k$ finitely generated over its prime field. The domain uses cycles over $k$, and the coefficients are ${\mathbb{Q}}_\ell$, not ${\mathbb{Z}}_\ell$.

Formulation and concept references: [S1].

Short English statement

Are all arithmetic-symmetry-invariant even cohomology classes generated by algebraic cycles defined over the base field?

Research attempt

Scope and source check (27 September 2026).

The target is surjectivity over the stated base field, for all codimensions and finitely generated fields. It is not the additional conjecture that homological and numerical equivalence agree. Nor may the rational coefficient field be replaced by an integral lattice. Li's formulation [S1] and Milne's account [E1] distinguish these assertions. Kahn [S2, S4] explicitly describes three incomplete approaches for surfaces over finite fields. Milne [S3] discusses failure of an integral Tate assertion for abelian varieties; that does not disprove the displayed rational map. None of these references is treated as a general solution.

The approach here separates three tasks often mixed together: finding geometric algebraic classes, descending classes to the base field, and certifying that enough independent classes have been found. The descent task admits an exact argument. The cycle-construction task remains open. For finite fields, a Frobenius calculation locates the relevant eigenspaces but also exposes a failure of counting eigenvalues alone. The elementary deductions below are not claimed as new results in the theory of cycles.

Boundary cases that fix the coefficient convention.

Write $d=\dim X$ and $V_i=H^{2i}_{\mathrm{\acute et}}(\bar X,\mathbb Q_\ell(i))$. For $i=0$, geometric irreducibility gives $V_0=\mathbb Q_\ell$, generated by $[X]$. For $i=d$, the trace identifies $V_d$ with $\mathbb Q_\ell$ with trivial Galois action. Any closed point $P$ has finite degree $n=[k(P):k]>0$, and its cycle class has trace $n$. Thus $[P]/n$ generates the target. No $k$-rational point is required. Codimensions outside $0\leq i\leq d$ contribute nothing.

This proves all codimensions for curves, but not divisors on arbitrary higher-dimensional varieties. More generally, for $X=\prod_{j=1}^s\mathbb P_k^{n_j}$ the projective-space cohomology and Künneth formulas give \[ H^{2i}(\bar X,\mathbb Q_\ell(i)) =\bigoplus_{\substack{a_1+\cdots+a_s=i\\0\leq a_j\leq n_j}} \mathbb Q_\ell h_1^{a_1}\cdots h_s^{a_s}. \] Each monomial is the class of a product of linear subspaces over $k$. Thus the selected map is surjective for this complete family. This calculation supplies a check on twists and degrees; a general projective variety does not have this cohomology presentation.

Lemma 15.1: finite-extension descent of an invariant class.

Let $L/k$ be finite Galois, with group $\Gamma$ and degree $n$. Suppose $v\in V_i^{G_k}$ and that $v$ is the class of $z\in\operatorname{CH}^i(X_L)\otimes\mathbb Q_\ell$ under the geometric identification of cohomology. Then \begin{equation}\label{wsref-0-74-63-3a-74-72-61-63-65} z_0=\frac1n f_*z\in\operatorname{CH}^i(X)\otimes\mathbb Q_\ell, \qquad\operatorname{cl}_\ell(z_0)=v, \end{equation} where $f:X_L\to X$ is the finite projection.

Proof. Finite flat pullback and proper pushforward satisfy $f^*f_*z=\sum_{\sigma\in\Gamma}\sigma z$. After applying the cycle map, its right side is $\sum_\sigma\sigma v=nv$, because the original class is fixed by all of $G_k$. The geometric pullback identifies the target cohomology spaces, so division by $n$ proves the claim. This argument is linear over $\mathbb Q_\ell$ and does not require the coefficients of $z$ to be rational numbers or its individual components to be invariant. $\square$

In particular, surjectivity over a finite Galois extension implies surjectivity over $k$: first restrict a $G_k$-invariant class, produce its cycle over $L$, and apply the lemma. A non-Galois separable extension is handled by pulling its cycle to a Galois closure. The denominators are legitimate even when $\ell\mid n$, since the coefficients are a field.

There is a qualification in imperfect characteristic. A cycle over $\bar k$ may require a purely inseparable extension in its field of definition, so it is not enough to say that every cycle descends to a finite separable extension. Finite purely inseparable base change is a universal homeomorphism and does not change geometric $\ell$-adic cohomology for $\ell\ne\operatorname{char}k$. For its finite flat projection of degree $p^e$, both push--pull composites on cycle groups multiply by $p^e$: there is a unique component above each component, and the residue degree times its flat multiplicity is $p^e$. Equivalently, the geometric fiber has that multiplicity. Dividing by $p^e$ gives the same class-descent statement. Combining this with separable descent handles a general finite algebraic field of definition. These are the usual cycle-map push--pull properties [E2], not an assumption that $k$ is perfect.

Every individual algebraic cycle has such a finite field of definition, and a finite linear combination has a common one. It follows that if a $G_k$-fixed class is known to lie in the geometric algebraic span, then it is represented by a cycle over $k$ with $\mathbb Q_\ell$ coefficients. This isolates the real gap: invariance has not been shown to imply membership in that geometric span.

The invariance assumption in Lemma 15.1 cannot be omitted. In the simple permutation representation with basis $e_1,e_2$ and an involution interchanging them, averaging $e_1$ gives $(e_1+e_2)/2$, not $e_1$. The anti-invariant vector $e_1-e_2$ is annihilated. Thus an orbit sum of a cycle does not preserve an arbitrary geometric class; it preserves the given class precisely when that class was already invariant. Also, averaging is performed on a known cycle representative. Writing the same averaging operator on cohomology only projects onto invariants and gives no procedure to lift an invariant vector through the cycle map. This distinction prevents the descent lemma from being read as an existence proof in disguise.

An actual integral countertest: the anisotropic conic.

Take $k=\mathbb Q$ and the smooth projective conic \[ X:\ x^2+y^2+z^2=0\subset\mathbb P^2_{\mathbb Q}. \] It has no real point. A closed point of odd degree would have residue field of odd degree over $\mathbb Q$, hence at least one real embedding; applying that embedding would give a real point of $X$. Therefore every closed point has even degree. The point $[1:i:0]$ over $\mathbb Q(i)$ yields a closed point of degree two, so the subgroup of degrees of zero-cycles is exactly $2\mathbb Z$.

Consequently, for the coefficient-replaced version of the displayed map, \[ \operatorname{CH}^1(X)\otimes\mathbb Z_2 \longrightarrow H^2(\bar X,\mathbb Z_2(1))^{G_{\mathbb Q}} \cong\mathbb Z_2, \] the image is $2\mathbb Z_2$, and the map is not surjective. With $\mathbb Q_2$ coefficients it is surjective, by dividing the degree-two class by two. This proves that the field and coefficient qualifications are doing real work. It is a counterexample only to that integral replacement over the fixed base field, not to the rational target or every convention called an integral Tate conjecture.

Finite-field attack: separate invariants from eigenvalue counts.

Now suppose $k=\mathbb F_q$. Let $F$ denote geometric Frobenius on untwisted $H^{2i}$ and put $T=q^{-i}F$ on $V_i$. This is the action after the Tate twist; geometric Frobenius acts on $\mathbb Q_\ell(1)$ by $q^{-1}$. Fixed classes over $\mathbb F_{q^m}$ are $\ker(T^m-I)$. The Weil eigenvalue bounds alone do not make all eigenvalues of $T$ roots of unity: algebraic numbers of complex absolute value one need not be algebraic integers after this normalization.

The following linear algebra requires no semisimplicity. Over a splitting field, on the generalized $\lambda$-eigenspace, $T^m-I$ is invertible unless $\lambda^m=1$. In the latter case \[ x^m-1=(x-\lambda)Q(x),\qquad Q(\lambda)=m\lambda^{m-1}\ne0. \] Thus $Q(T)$ is invertible on that generalized eigenspace and $\ker(T^m-I)=\ker(T-\lambda I)$ there. Generalized eigenvectors are not accidentally fixed by passing to a finite extension. Let $M$ be the least common multiple of the orders of the finitely many root-of-unity eigenvalues, taking $M=1$ if there are none. Then \begin{equation}\label{wsref-0-74-63-3a-73-74-61-62-69-6c-69-7a-65} \bigcup_{m\geq1}\ker(T^m-I)=\ker(T^M-I). \end{equation} The equality descends to $\mathbb Q_\ell$, since both sides are defined over it. The union is a vector space: a common multiple of two extension degrees fixes sums of the corresponding vectors.

This is a useful finite-dimensional reduction for a fixed variety, but it constructs no cycles spanning the stabilized space. It also does not identify its dimension with the total multiplicity of the relevant roots in the characteristic polynomial. For example, \[ T_1=\begin{pmatrix}1&0\\0&1\end{pmatrix},\qquad T_2=\begin{pmatrix}1&1\\0&1\end{pmatrix} \] have the same characteristic polynomial and the same traces of every positive power. Yet their fixed spaces have dimensions two and one, respectively, and $T_2^m-I=m(T_2-I)$ has rank one for every $m>0$. These are test representations, not asserted examples of geometric Frobenius. They show that characteristic-polynomial data alone do not justify replacing a fixed-space calculation by a multiplicity count. Milne [E1, Section 1] keeps the relevant semisimplicity condition separate.

A finite intersection certificate.

Suppose an upper bound $r$ for $\dim V_i^{G_k}$ is established, and one finds cycles $z_1,\ldots,z_r$ of codimension $i$ and $w_1,\ldots,w_r$ of codimension $d-i$, all over $k$, such that \begin{equation}\label{wsref-0-74-63-3a-6d-61-74-72-69-78} B_{ab}=\deg(z_a\cdot w_b),\qquad \det B\ne0. \end{equation} Then the Tate map in codimension $i$ is surjective. Indeed, a relation $\sum_a c_a\operatorname{cl}_\ell(z_a)=0$ pairs with each $w_b$ to give $\sum_a c_aB_{ab}=0$. Invertibility forces every $c_a=0$. These $r$ invariant classes therefore fill a space whose dimension is at most $r$. The products are Chow intersection products on the smooth variety; no unproved transversality of chosen representatives is needed.

For finite fields, the multiplicity of $1$ in the characteristic polynomial of $T$ is always an upper bound, so it can be used in this sufficient certificate without assuming semisimplicity. Finding that many paired cycles would itself force equality of the dimensions. But a universal supply of complementary cycles with an invertible pairing is extra work. Even a nondegenerate full pairing need not restrict nondegenerately to invariant subspaces: for $T_2$ above and its contragredient, the fixed vectors are $e_1$ and $e_2^*$, whose pairing is zero. Therefore the certificate is sufficient, not an automatic consequence of duality.

Verification, first gap, and outcome.

The descent proof was checked at degrees divisible by $\ell$ and at purely inseparable degrees. The conic provides a concrete failure when the permitted denominators are removed. The Frobenius argument was checked on a nontrivial Jordan block rather than assuming diagonalizability. Invariant dimensions, intersection ranks, and eigenvalue multiplicities were kept distinct. No finite point count or cycle search is being reported as a proof for all varieties.

Three next targets are to construct cycles spanning the stabilized Frobenius space for a specified finite-field family; to obtain actual intersection matrices meeting (tc:matrix); and to identify the geometric algebraic span for Galois-invariant classes over other finitely generated fields. These cannot be replaced by averaging a class that has no known cycle representative. The first unresolved step is exactly that geometric algebraicity, before the proved descent lemma applies.

UNRESOLVED. The output proves descent and certification lemmas, elementary complete subcases, and explicit obstructions to integral and spectral shortcuts. It does not construct the missing cycles in general. Written by GPT-6 Astra Ultra with primary-source scope checks and direct algebraic verification; no formal proof certification is claimed.

Sources

• S1 Li, What is the Tate conjecture?, map $2$ and Conjecture 1 $2013$. https://www.math.columbia.edu/~chaoli/docs/TateConjecture.html. Formulation source. Consulted 24 September 2026: Cycle-class map and Conjecture 1.
• S2 An approach to the Tate conjecture for surfaces over a finite field — Bruno Kahn $2025$. https://arxiv.org/abs/2505.07337. Further reference; not a proof endorsement.
• S3 On the integral Tate conjecture for abelian varieties — J. S. Milne (September 8, 2025). https://arxiv.org/abs/2509.06707. Further reference; not a proof endorsement.
• S4 An approach to the Tate conjecture for surfaces over a finite field — Bruno Kahn, full text. https://arxiv.org/html/2505.07337v1. Further reference; not a proof endorsement.
• E1 J. S. Milne, The Tate conjecture over finite fields (AIM talk), revised 7 May 2008, especially Section 1. https://www.jmilne.org/math/articles/2007e.pdf. Primary author account separating cycle-surjectivity, numerical equivalence, and semisimplicity; consulted 27 September 2026.
• E2 The Stacks Project, Chow Homology, proper pushforward, flat pullback, and their compatibility. https://stacks.math.columbia.edu/tag/02P3. Also Topological invariance of the small étale site: https://stacks.math.columbia.edu/tag/04DY. References for finite-extension push--pull and purely inseparable invariance; consulted 27 September 2026.