gpt 6 astra pro attempt on Math-196

Grothendieck–Serre Conjecture for Regular Local Rings

Definitions and mathematical statement

A Noetherian local ring $(R,\mathfrak m)$ is regular when $\dim R=\dim_{R/\mathfrak m}\mathfrak m/\mathfrak m^2$; it is a domain here, with fraction field $K$. A reductive $R$-group scheme is smooth affine with connected reductive geometric fibers. A $G$-torsor is a principal homogeneous space locally trivial for the ëtale topology. The assertion is \[ \ker\left(H^1_{\mathrm{et}}(R,G)\to H^1_{\mathrm{et}}(K,G)\right)=\{1\}, \] where these are pointed sets and the kernel is the preimage of the trivial torsor. Thus generic triviality forces triviality over $R$. The exact statement does not impose a field inside $R$, excellence, isotropy, or a constant group scheme.

Formulation and concept references: [S1].

Short English statement

If a principal bundle under a reductive group over a regular local ring is trivial over the fraction field, must it already be trivial over the ring?

Research attempt

Current frontier.

The cited survey distinguishes the full regular-local conjecture from field-containing and unramified cases [S1]. A recent theorem covers unramified regular semilocal rings with totally isotropic reductive groups, still not the unrestricted pair $(R,G)$ here [E1]. The discrete-valuation-ring case is an established input, as recorded in the survey. We investigate whether codimension-one trivializations can be synchronized into one generic section. This separates the easy normality extension from the genuinely nonabelian patching problem.

Attack route.

There are three plausible directions: descend a trivialization from a completion, use approximation in $G(K)$ to correct generic denominators, or apply a geometric presentation/patching theorem. Completion descent requires compatible descent data, not just a point after base change. We pursue the second route and identify an exact simultaneous-coset condition; the last route would have to prove that condition for the torsor-derived data in the remaining generality.

Attempt.

Let $P$ be a generically trivial right $G$-torsor. Since $G$ is affine, affineness descends along an ëtale trivializing cover, so $P=\operatorname{Spec}B$ is affine over $R$. Choose $s\in P(K)$.

Hartogs extension lemma. If this same point $s$ extends to $P(R_{\mathfrak p})$ for every height-one prime $\mathfrak p$, then it extends uniquely to $P(R)$. A regular local domain is normal, and a Noetherian normal domain satisfies \begin{equation}\label{wsref-0-65-71-31-39-32-3a-69-6e-74-65-72-73-65-63-74-69-6f-6e} R=\bigcap_{\operatorname{ht}\mathfrak p=1}R_{\mathfrak p}\subset K. \end{equation} The coordinate homomorphism $B\to K$ corresponding to $s$ therefore has image in $R$, proving existence; injectivity $R\hookrightarrow K$ proves uniqueness. In dimension zero the ring is already a field and the assertion is immediate. Equation (eq192:intersection) is the standard height-one description of a normal Noetherian domain [E2].

Because $B$ is finitely presented, $s$ extends over $R[1/f]$ for some nonzero $f$. Only finitely many height-one primes contain $f$. For each of those primes, the discrete valuation ring theorem supplies a local point $t_{\mathfrak p}\in P(R_{\mathfrak p})$. Since $P_K$ is a torsor, there is a unique $g_{\mathfrak p}\in G(K)$ such that \[ t_{\mathfrak p}=s\cdot g_{\mathfrak p}. \] For primes away from the bad set put $g_{\mathfrak p}=1$. The set of corrections making $s\cdot g$ integral at $\mathfrak p$ is exactly \[ g_{\mathfrak p}G(R_{\mathfrak p})\subset G(K). \] Indeed, an integral point trivializes the torsor over $R_{\mathfrak p}$, so every other integral point differs by an element of $G(R_{\mathfrak p})$.

Thus a sufficient and necessary patching condition for this torsor and this generic section is \begin{equation}\label{wsref-0-65-71-31-39-32-3a-70-61-74-63-68} \bigcap_{\operatorname{ht}\mathfrak p=1} g_{\mathfrak p}G(R_{\mathfrak p})\ne\varnothing. \end{equation} For $g$ in the intersection, $s\cdot g$ extends everywhere in codimension one and the Hartogs lemma gives a section of $P$ over $R$, hence triviality. Conversely any global section determines such a $g$. Notice the order $g_{\mathfrak p}^{-1}g\in G(R_{\mathfrak p})$; swapping factors would change a nonabelian condition.

This is not a claim that arbitrary collections of local cosets can always be synchronized. The collection here is constrained by coming from a single torsor over $R$. Proving (eq192:patch) just for those collections is enough; making it an unrestricted approximation assertion would be stronger and could be false.

Abelian calibration. For $G=\mathbf G_m$, regular local factoriality permits an explicit correction. Write the prescribed valuations at the finitely many bad primes as $e_{\mathfrak p}=v_{\mathfrak p}(g_{\mathfrak p})$, choose generators $\pi_{\mathfrak p}$ of these height-one primes, and take \[ g=\prod_{\mathfrak p\text{ bad}}\pi_{\mathfrak p}^{e_{\mathfrak p}}\in K^\times. \] It has the required valuation at every height-one prime, including zero at all others; hence it lies in every coset. The same argument works coordinatewise for a split torus. This rederives an elementary special case: invertible modules over a local ring are free. For $\mathrm{GL}_r$, torsors are finite locally free rank-$r$ modules and local freeness likewise makes every torsor trivial. These calibrations show that nonabelian structural constraints, not the normality intersection itself, contain the general difficulty.

Stress test.

The false quantifier swap would be \[ \left[\forall\mathfrak p\ \exists t_{\mathfrak p}\in P(R_{\mathfrak p})\right] \quad\Longrightarrow\quad \left[\exists s\in P(K)\ \forall\mathfrak p\ (s\text{ integral at }\mathfrak p)\right]. \] The Hartogs lemma proves only the conclusion after the single section has been found. Weak approximation at finitely many primes may introduce new bad denominators elsewhere, whereas (eq192:patch) includes integrality at all the other primes. A proof that ignores them is incomplete. Nor does an $R$-torsor point after completion descend automatically to $R$ without a descent argument.

No excellence, coefficient field, isotropy or unramifiedness was inserted into the Hartogs calculation. Such hypotheses do appear in available theorems that might supply the missing patching; they cannot be dropped when applying those theorems. The current attempt does not produce a patching theorem in the arbitrary ramified and anisotropic setting.

Outcome.

Outcome: Conditional reduction. The full target for a given torsor is reduced to a precise intersection of torsor-derived cosets in $G(K)$. The extension step and the split-torus calibration are proved, and the difference between local existence and a simultaneous generic correction is explicit. The unresolved ingredient is nonabelian synchronization in the full regular-local generality, not a further application of normality.

Sources

• S1 Problems about torsors over regular rings, with an appendix by Yifei Zhao; arXiv2201.06424v4,5May2025; HTML internal29March2023. https://arxiv.org/pdf/2201.06424v4. Catalogue formulation source.
• E1 Kęstutis Česnavičius, Unramified Grothendieck--Serre for isotropic groups, arXiv:2311.08660, main theorem and its unramified/totally isotropic hypotheses. https://arxiv.org/abs/2311.08660. Research reference; consulted 27 September 2026.
• E2 The Stacks Project Authors, The Stacks Project, Algebra, Lemma on a Noetherian normal domain as the intersection of its height-one localizations (Tag 031T), and regular local factoriality. https://stacks.math.columbia.edu/tag/031T. Research reference; consulted 27 September 2026.