Kerr cosmic censorship: curvature, holonomy, and weak connections

Source and verification

This article explains family 264 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.

🏷️ Cauchy evolution and extension regularity

The Kerr solutions describe rotating vacuum black holes. Their inner Cauchy horizons permit smooth continuations beyond the region determined by their initial data. Strong cosmic censorship asks whether this failure of predictability disappears for generic perturbations. This family studies that question near each fixed rotating, subextremal Kerr bridge, at three levels of extension regularity. (OpenAI, 2026) (OpenAI, 2026a) (OpenAI, 2026c)

The prerequisites are smooth manifolds, tensor calculus, ordinary differential equations, Sobolev spaces, and the Baire category theorem. The central geometric objects are introduced here before the analytic construction.

A Lorentzian metric has signature $(-,+,+,+)$. A vector $V$ is timelike if $g(V,V)<0$; a time orientation chooses its future cone. A spacelike hypersurface has a positive-definite induced metric. A Cauchy hypersurface meets every inextendible causal curve exactly once. Its existence expresses global hyperbolicity and lets initial data determine a spacetime evolution.

Vacuum initial data are a Riemannian metric $h$ and second fundamental form $K$ on a three-manifold $\Sigma$. They satisfy the constraints

$$\begin{array}{c}R(h)+({\mathrm{tr}}_{h}K{)}^2-|K{|}_h^2=0,\\ {\mathrm{div}}_{h}K-d({\mathrm{tr}}_{h}K)=0.\end{array}$$

Their smooth evolution satisfies $\mathrm{Ric}(g)=0$. The maximal globally hyperbolic development, abbreviated MGHD, contains every globally hyperbolic development of the same data through the appropriate isometric embedding. Maximality within that class does not mean that no larger Lorentzian manifold contains it.

A future extension in these manuscripts is a smooth, time-orientation-preserving isometric embedding of the full MGHD onto a proper open subset of a connected smooth four-manifold. The ambient metric must be nondegenerate and have the specified regularity. A future-directed timelike $C^1$ curve must remain in the image before reaching its boundary. The ambient metric need not solve the vacuum equations or remain globally hyperbolic outside the image. Thus the conclusion excludes more than vacuum continuations. (OpenAI, 2026b) (OpenAI, 2026a) (OpenAI, 2026c)

🏷️ The data neighborhood and the three conclusions

Fix a Kerr mass $M>0$ and rotation parameter $\mathfrak{a}$ with $0<|\mathfrak{a}|<M$. “Subextremal” is the strict upper inequality; the lower inequality excludes the nonrotating case. The reference initial surface is the smooth complete two-ended bridge through the bifurcation sphere, with $\Sigma \simeq \mathbb{R}\times {\mathbb{S}}^2$. Denote its data by $d_{\ast }=(h_{\ast },K_{\ast })$.

Choose a smooth weight $\rho \ge 1$ equal to radius far out on each end. Using the connection and tensor norm of $h_{\ast }$, define for perturbations $(u,v)$

$$\begin{array}{rl}{p}_m(u,v)=\sum _{j=0}^m[& \underset{\Sigma }{\sup }{\rho }^{1+j}|{\nabla }_{\ast }^{j}u{|}_{\ast }\\ & +\underset{\Sigma }{\sup }{\rho }^{2+j}|{\nabla }_{\ast }^{j}v{|}_{\ast }].\end{array}$$

The data space $\mathcal{D}$ consists of smooth, positive, complete vacuum data with every $p_m(d-d_{\ast })$ finite. These conditions permit differentiated $O(r^{-1})$ metric tails and $O(r^{-2})$ second-fundamental-form tails. They impose no symmetry, parity, fixed leading coefficient, or fixed asymptotic charge. Existence of every possible charge integral is not assumed. The topology uses fixed spatial representatives and all these seminorms, rather than a quotient by arbitrary diffeomorphisms.

The neighborhood requires only

$${\mathcal{U}}_{\epsilon }=\{d\in \mathcal{D}:p_{10}(d-d_{\ast })<\epsilon \}.$$

Higher seminorms must be finite but need not be small. A set is meagre if it is a countable union of nowhere dense sets; its complement is residual. The manuscript establishes that the data space and its open neighborhoods are Baire spaces. Consequently a residual subset is dense, but need not be open and need not have a specified probability or measure.

Local generic inextendibility, as stated in the family

Near every fixed rotating subextremal Kerr center, the data whose full MGHD admits a future extension with continuous metric and locally square-integrable weak connection form a meagre subset of a sufficiently small ${\mathcal{U}}_{\epsilon }$.

The other two manuscripts establish the analogous conclusions for $C^1$ and $C^2$ ambient metrics. Their statements use positive rotation; the strongest companion explicitly treats either sign. Each result may choose its own neighborhood. (OpenAI, 2026b) (OpenAI, 2026a) (OpenAI, 2026c)

On a common restriction around a positive-rotation center, the permitted metric classes satisfy

$$C^2\subset C^1\subset C^0\cap W_{\mathrm{loc}}^{1,2}.$$

Excluding extensions in the largest class is strongest. It still does not exclude arbitrary $C^0$ extensions. The neighborhood and estimates depend on the fixed center; no uniform limit toward zero rotation or extremality, and no result for all vacuum initial data, is asserted. Exact Kerr itself illustrates why the conclusion cannot apply to every nearby datum.

🏷️ Why curvature is not enough

For a continuous nondegenerate metric, define the lowered connection coefficients distributionally by

$${\Gamma }_{ijk}^{♭}=\frac{1}{2}({\partial }_j{g}_{ik}+{\partial }_k{g}_{ij}-{\partial }_i{g}_{jk}).$$

The reverse identity is

$${\partial }_k{g}_{ij}={\Gamma }_{ikj}^{♭}+{\Gamma }_{jki}^{♭}.$$

Continuity and nondegeneracy bound $g$ and $g^{-1}$ on compact chart subsets. Hence weak $L^2$ connection coefficients are equivalent here to $g\in W_{\mathrm{loc}}^{1,2}$. The derivatives are weak derivatives, not just almost-everywhere classical derivatives. (OpenAI, 2026b)

Ambient regularity	Available control	Obstruction used
$C^2$	Curvature is locally bounded in regular frames	A sufficiently large finite tidal observation
$C^1$	Connection is continuous and locally bounded	Holonomy of short loops
$C^0\cap W_{\mathrm{loc}}^{1,2}$	Connection has a local $L^2$ bound	Holonomy averaged over families of loops

A regularity distinction

Put $f(x)=1+|x{|}^{3/2}$ and consider $-d{t}^2+d{x}^2+f(x{)}^{2}d{y}^2+d{z}^2$ near $x=0$. This is a nondegenerate $C^1$ Lorentzian metric. For $x\ne 0$, its $xy$ curvature component is, up to sign convention, $f(x)f^{\prime \prime }(x)$, which diverges as $x\to 0$. Thus unbounded curvature can coexist with a $C^1$ metric extension. This elementary metric is not offered as a vacuum solution.

For a smooth curve, parallel transport solves a matrix equation $P^{\prime }=-AP$, with $A$ the connection contracted against its velocity. Around a closed loop it gives holonomy $H$, which measures the failure of a vector to return unchanged. If $\int \Vert A\Vert \le \eta$, the integral equation and Grönwall’s inequality give $\Vert H-I\Vert \le e^{\eta }-1$.

For a small coordinate rectangle with side parameter $\ell$, the quadratic term in $H(\ell )-I$ is curvature applied to its two directions. This relates a bound on transport to curvature without assuming that ambient curvature exists. But an $L^2$ function in four dimensions has no generally defined restriction to an individually prescribed curve. The weakest-regularity proof must average before making curvewise estimates. (OpenAI, 2026a) (OpenAI, 2026b)

🏷️ From a possible extension to a finite experiment

The $C^2$ foundation constructs a comparison development containing the entire original bridge as a Cauchy hypersurface. It combines exterior estimates, control of the asymptotically flat tails, and an interior region with two optical coordinates. The resulting comparison spacetime embeds in the full MGHD by geometric uniqueness. Establishing these estimates is a major analytic input, not an elementary consequence of Kerr’s explicit formula. (OpenAI, 2026c)

In an interior branch regime, one optical coordinate tends to infinity while the other has a finite limit; in the corner regime both diverge. These are coordinate-limit descriptions, not boundary points already attached to the perturbed spacetime. The argument must locate an arbitrary future exit, rather than assume it lies on a preselected horizon patch.

A regular extension supplies a family of finite-lifetime timelike observers. For $C^2$ metrics, their curvature components in parallel frames have a common bound on suitable finite tests. The family is reduced to countably many tests using a basis of initial observer states and integer bounds. A strict violation occurs at a finite point and persists under finite-time Cauchy stability.

The perturbation is a late oscillatory packet, transported back to the original initial surface. A constraint correction makes the modified data exactly vacuum, vanishes on a protected inner region, and retains a decaying exterior tail. The tail is part of the corrected data and can change asymptotic charges. The resulting observation is attached to the development of the entire corrected bridge. An approximate packet or an isolated interior construction would not suffice for the theorem. (OpenAI, 2026c)

At $C^1$ regularity, the tests instead bound holonomy by the loop’s developed length: the integral of the velocity norm in a frame parallel-transported along that loop. Two exact perturbations with opposite linear parts cancel their common pure-packet quadratic forcing when their equations are subtracted. This controls their difference even when an individual nonlinear error need not be smaller than the linear signal. The background curvature also cancels in the signed comparison. (OpenAI, 2026a)

Let $T$ denote a late optical depth and $\kappa >0$ the fixed inner rate. The lapse coefficient is comparable to $r=e^{-\kappa T}$. The $C^1$ experiment uses amplitude $A=r^{7/8}$ and loop parameter $h=r^{1/16}$. Finite interpolation of the loop holonomy gives, under a passing transport test, a bound of the form

$$\begin{array}{rl}{Q}_T(r/h+h^{16})& =Q_T(r^{15/16}+r)\\ & =o(r^{7/8}),\end{array}$$

where $Q_T=e^{o(T)}$ denotes subexponential losses at the fixed derivative orders being used. The constructed scaled curvature signal exceeds $A$, producing a finite contradiction. The strict exponent gaps make the argument tolerate these losses. (OpenAI, 2026a)

🏷️ Averaging at the square-integrable threshold

The weakest-regularity companion first shows that every allowed future exit supplies a positive-measure family of smooth interior geodesics with finite full lifetime, controlled chart velocities, and finitely many ambient charts with uniform metric, inverse-metric, and connection $L^2$ bounds. Its argument uses a local Lipschitz first-exit graph and small averaged connection mass near that graph. The geodesics remain those of the smooth MGHD; no general weak ambient geodesic theory is assumed. (OpenAI, 2026b)

Canonical phase-space volume then provides the measure bookkeeping. On the future unit mass shell, geodesic flow is Hamiltonian. Its transverse hitting maps preserve canonical section volume. The source uses this to exclude a positive-measure family consisting only of corner tracks and to match observers between two nearby developments. It does not claim that the projection from phase space to spatial position is injective.

The new packet has larger amplitude $A=r^{1/4}$, which exceeds $\sqrt{r}$. It requires a new nonlinear comparison with error $o(A)$ through the required finite derivative orders; the earlier $C^1$ estimate alone does not supply it. The source constructs a curvature difference over a fixed proportion of an observation region, rather than at just one point. (OpenAI, 2026b)

The averaging argument can be isolated. On that region use the commuting frame

$$\begin{array}{c}{F}_0={\partial }_u,F_1=r^{-1}{\partial }_v,\\ F_2={\partial }_{{\theta }^1},F_3={\partial }_{{\theta }^2}.\end{array}$$

Its metric matrix and inverse are uniformly bounded, with the needed finite coordinate derivatives bounded by $Q_T$. The volume density in the optical coordinates is comparable to $r$.

Bounded Lorentzian metric matrices alone do not bound a change of frame: arbitrarily large Lorentz boosts preserve the same matrix. The source also controls a unit timelike observer in both frames. Adding twice the square of its metric dual to each Lorentzian form makes a positive-definite form with controlled eigenvalues, which bounds the frame change and its inverse at the base point.

Take rectangles with consecutive side displacements $\ell {\partial }_v$, $\ell rX$, and their negatives, where $X$ is either angular coordinate direction or their sum. Both sides have frame components of size $\ell r$. If $C_2$ is the quadratic holonomy coefficient, multilinearity gives

$$g(C_2{F}_1,X)=\pm R({\partial }_v,X,{\partial }_v,X).$$

The factor $r$ in the second side cancels $r^{-1}$ in $F_1$. Thus transport controls the relevant curvature block without a leftover lapse factor.

For each sampled edge with unit-parameter path $y_e(s)$, let

$$I_e={\int }_0^1{1}_{\{\text{before stop}\}}|{\tilde{\Gamma }}_c(y_e(s))|ds.$$

This uses the uncontracted chart connection array, with a stop before the frame or chart margin loses control. The actual transport coefficient also includes the velocity factor of size $\ell r$. Averaging over four position and three bounded momentum variables gives

$$\int I_e^{2}d{\nu }_T\le C/r.$$

This follows from the connection’s chart $L^2$ norm and the volume density. At each fixed edge parameter the position map is a translation followed by the smooth interior embedding and chart map. Fubini therefore avoids assigning a Sobolev trace to a prescribed loop.

Choose cutoff $L=r^{-17/32}$ and loop scale ${\ell }_T=r^{1/16}$. Markov’s inequality yields

$${\nu }_T\{I_e>L\}\le C/(r{L}^2)=C{r}^{1/16}.$$

Outside the exceptional sets for finitely many sampled loops, the frame-transport equation improves both stopping bounds. Hence no first stop occurs and each complete loop stays in its selected chart. This step justifies the chart control instead of assuming it during averaging.

To extract curvature, expand $H(\ell )-I$ through degree seventeen and evaluate it at $0,{\ell }_T,\dots ,17{\ell }_T$. The fixed Vandermonde matrix is invertible. Its inverse extracts ${\ell }_T^2{C}_2$ from the eighteen values; an eighteenth-derivative bound leaves remainder $Q_T{\ell }_T^{16}$ after division. The resulting estimate is

$$\begin{array}{rl}\Vert C_2\Vert & \le CrL/{\ell }_T+Q_T{\ell }_T^{16}\\ & =C{r}^{13/32}+Q_{T}r=o(r^{1/4}).\end{array}$$

The two angular directions and their sum recover the whole symmetric curvature block by polarization. Only finitely many sizes are sampled; no common exceptional set for every possible loop is required. (OpenAI, 2026b)

🏷️ The Baire-category conclusion

The weak-extension tests $E(O,B)$ specify a small initial seed ball $O$, an integer chart bound $B$, and an extension witness good for more than $99\%$ of its seed measure. The other geometric choices are existential witnesses, not uncountably many additional indices. The source proves that these countably many tests cover all permitted extensions. It does not assume these test sets are closed. (OpenAI, 2026b)

Suppose one test were dense in a nonempty open data set. Choose an extendible background in that set, then make a sufficiently late packet perturbation, arbitrarily small in the finite seminorm defining the chosen neighborhood. Density supplies a second extendible datum close enough to this fixed finite experiment that its curvature difference persists.

Match the two developments at the same numerical terminal positions and momenta. Canonical volume preservation leaves a positive-measure family good for both extension witnesses. The oscillatory packet signal holds on a fixed positive fraction of the observation interval, whereas the two exceptional sets from averaged holonomy have measure tending to zero. At a remaining observation, each curvature block is $o(A)$ but their difference exceeds $A$. This contradiction proves nowhere denseness.

Taking the countable union proves meagreness of the extendible data. Every persistence assertion concerns a fixed finite observation; continuity of maximal geodesic lifetime is not assumed. The order of choices is crucial: first the fixed small $p_{10}$ neighborhood, then a finite approximation order and test, then the packet parameters and sufficiently late observation. (OpenAI, 2026b)

🏷️ Key steps and proof outlines

The $C^2$ category argument needs finite witnesses

The closed sets used in the $C^2$ argument simultaneously bound the lifetime and frame curvature of geodesics launched from a fixed open set $O$ of initial positions and velocities. Write this set of data as $A(O,N)$, where every such geodesic has maximal lifetime at most $N$ and curvature magnitude at most $N$. Its closedness does not require continuity of maximal lifetime. A failure has a compact witness: either a geodesic segment exists beyond time $N$, or curvature exceeds $N$ at a point reached in finite time. Continuous dependence on a compact causal region preserves that witness under small changes of the initial data.

Why must a hypothetical extension furnish one of these bounded tests? The source first proves a fixed-endpoint homotopy lemma. In a globally hyperbolic open region $P$ of a Hausdorff ambient spacetime, a continuous family of future causal curves with fixed endpoints in $P$ cannot leave $P$ if one member stays inside. The parameters for which the curve stays inside form an open set. They also form a closed set, because each such curve lies in the intrinsic compact causal diamond between the endpoints, which is closed in the ambient space. A regular-exit construction then uses this lemma to obtain geodesic exits with controlled chart positions, velocities, and curvature. This supplies the bounded set $A(O,N)$ needed by the category argument; it avoids assuming that an arbitrary timelike curve approaching the boundary has a limiting unit tangent. These are the specific roles of the category and regular-exit lemmas in “Category” (OpenAI, 2026c).

The $C^1$ argument removes the dangerous boost powers

A frame adapted to the contracting null direction involves powers of a small parameter $a$, comparable to the packet scale $r$. If ${\tilde{e}}_{\alpha }$ is the corrected null frame, use boost weights

$$s_0=\frac{1}{2},s_1=-\frac{1}{2},s_2=s_3=0,E_{\alpha }=a^{s_{\alpha }}{\tilde{e}}_{\alpha }.$$

A direct transformation of the connection produces factors

$$a^{1/2+s_{\alpha }+s_{\beta }-s_{\mu }}{\tilde{V}}_{\alpha \beta }^{\mu }.$$

Some exponents are negative, so a uniform bound on the unboosted coefficients would be insufficient.

The proof classifies the offending indices. Null metric compatibility makes the coefficients with upper index $0$ and final index $1$ vanish exactly. The remaining potentially singular coefficients vanish for the background because its distinguished null direction is geodesic and orthogonal to the transverse directions. Their perturbations decay faster than the boost grows, leaving a decaying error. This gives controlled connection coefficients in the rescaled frame used to draw the loops. Torsion freeness then expresses frame commutators as differences of those same coefficients, so the loop-flow estimates inherit the control. This cancellation, developed in “Holonomy,” is what makes the small-loop polynomial comparison compatible with the large geometric boost (OpenAI, 2026a).

Weak connection coefficients require stopping before averaging

For a merely square-integrable connection, individual curves do not automatically have meaningful connection traces. The weak-regularity companion therefore averages a translated family of loops and works almost everywhere in its parameters. Let $M$ be the matrix of the geometric frame in the extension chart, and stop a loop if $\Vert M\Vert$ reaches a fixed large bound $K$, or if its chart position approaches the chart boundary.

Before stopping, the frame equation has the form

$$M^{\prime }=MA-{\Gamma }_c(Mf)M,|A|\lesssim \ell Q,|f|\lesssim \ell r,$$

where ${\Gamma }_c$ is the chart connection and $\ell$ is the loop parameter. The change-of-variables estimate for translated observation points follows from the isometry’s volume identity; it does not presuppose an uncontrolled bound on frame derivatives. Fubini and the $L^2$ connection bound then discard a small set of parameters and bound each retained edge integral by $L=r^{-17/32}$.

Integrating the stopped equation bounds frame variation by a constant times

$$K\ell Q+K^2\ell rL.$$

For the source’s finitely many scales, $\ell \le 17r^{1/16}$, the second term is $O(r^{17/32})$. The first is also made small in the prescribed parameter order. Both frame variation and chart displacement strictly improve their stopping thresholds, so the retained loops never stop. Only now is their full holonomy available for the polynomial contradiction. This order of operations closes a potential circularity between chart control, averaging, and transport (OpenAI, 2026b).

🏷️ Connections and proof scope

• Spacetime Penrose inequalities also begins with vacuum or matter constraint data and separates geometric hypotheses from the final spacetime conclusion.
• Coulomb ionization uses conditional averaging to turn local estimates into a global conclusion. Here the averaging is over observer positions and momenta, and its purpose is to use an $L^2$ connection bound without curve traces.

The elementary regularity identities, exponent comparisons, and interpolation mechanism explain the final obstruction. The full exterior evolution, constraint correction, beam construction, nonlinear comparison, and arbitrary-exit reduction remain substantial manuscript inputs. These notes do not independently verify those complete arguments.

🏷️ Manuscripts and verification scope

• Generic Future Inextendibility with Square-Integrable Connection Near a Fixed Kerr Spacetime — pinned manuscript and companion source.
• Generic C¹ Future Inextendibility Near Rotating Subextremal Kerr Spacetimes — pinned manuscript and companion source.
• Quantitative Near-Kerr Evolution and Generic C² Future Inextendibility — pinned manuscript and companion source.

No family scope page was present at lean/docs/264.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 2026a. Generic C¹ Future Inextendibility Near Rotating Subextremal Kerr Spacetimes.

🐻  OpenAI 2026b. Generic Future Inextendibility with Square-Integrable Connection Near a Fixed Kerr Spacetime.

🐻  OpenAI 2026c. Quantitative Near-Kerr Evolution and Generic C² Future Inextendibility.