1. The question
A group can look negatively curved from far away without arriving with a particular curved space on which it acts. Does that large-scale geometry guarantee a concrete geometric realization?
For a word-hyperbolic group, the strongest familiar examples come from trees and hyperbolic manifolds. In both cases the group acts properly and cocompactly by isometries on a space satisfying a precise curvature comparison. It has long been natural to ask whether every hyperbolic group admits such an action on some CAT(0), or even CAT(−1), space.
The manuscript claims a counterexample. Its group has a finite two-dimensional classifying space and a linear disk-filling inequality, yet admits no geometric CAT(0) action in any dimension. The obstruction survives replacing the original finite complex by any homotopy-equivalent finite simplicial complex.
2. Background and essential definitions
Large-scale negative curvature
Given a finite generating set, the Cayley graph of a group has one vertex for each element and edges for multiplication by generators. A group is word-hyperbolic if geodesic triangles in this graph are uniformly thin: each side stays within a fixed distance of the other two sides.
For finitely presented groups, an equivalent condition is a linear isoperimetric inequality. A word representing the identity can be filled by a disk diagram, a tiled disk whose faces record defining relations. If the boundary has length , linear filling means that the number of faces needed is at most
for one fixed constant . The constant may depend on the presentation. This is a combinatorial statement about how efficiently relations explain a null word.
A finite cell complex is aspherical when its universal cover is contractible. It is then a classifying space for its fundamental group: its loops encode the group, with no additional higher homotopy. The theorem here supplies a finite two-dimensional simplicial complex, built from vertices, edges and triangles, with this property.
CAT(0) geometry and actions
A geodesic space is CAT(0) if each geodesic triangle is no thicker than its Euclidean comparison triangle. This gives convexity of distance, unique geodesics, and contractibility. CAT(−1) uses comparison triangles in the hyperbolic plane and is a stronger curvature condition.
An isometric action is geometric when it is proper and cocompact. Properness means that only finitely many group elements can move a given compact set to meet itself. Cocompactness means that translates of one compact set cover the space. A metric space is proper if its closed bounded subsets are compact; completeness means that every Cauchy sequence converges.
These action hypotheses matter. The theorem does not forbid the trivial action on a point, which is not proper for an infinite group. Nor does it assert that every action on every CAT(0) space has a fixed point.
Why a finite metric model would produce an action
If a finite complex has a compatible locally CAT(0) geodesic length metric, its universal cover inherits a global CAT(0) metric. Here compatible means that the metric induces the given topology, and a length metric measures distance through path lengths.
Deck transformations of the universal cover give a geometric action of the fundamental group. Thus excluding geometric actions also excludes finite locally CAT(0) models. The same implication applies to CAT(−1) models.
3. Development history and previous partial results
Linear filling entered geometric group theory as a group-theoretic form of negative curvature in Gromov’s development of hyperbolic groups. The associated realization problem asks how far that combinatorial curvature can be upgraded to an actual metric model.
Many important classes admit such models. Free groups act on trees. Fundamental groups of closed hyperbolic manifolds act on hyperbolic space. Cubulation, which constructs actions on spaces assembled from Euclidean cubes, supplies further CAT(0) realizations.
Random groups also provide positive examples. Ollivier and Wise proved that in the standard density model, density below yields free cocompact actions on CAT(0) cube complexes with probability tending to one as relator length grows. Thus randomness and hyperbolicity do not by themselves produce a curvature obstruction. The new construction needs relations designed for an additional purpose. Ollivier–Wise
Brady and Crisp established earlier dimension obstructions. They constructed two-dimensional torsion-free hyperbolic groups that cannot act properly on proper two-dimensional CAT(0) spaces, and examples where CAT(0) and CAT(−1) dimensions differ. Such a theorem leaves open the possibility of an action in a larger dimension. The present manuscript claims an obstruction independent of dimension. Brady–Crisp
A separate line of work uses energy and averaging to study actions of random groups. Gromov and Izeki–Nayatani developed harmonic-map methods; Izeki, Kondo and Nayatani obtained fixed-point results for target classes satisfying specified nonlinear spectral bounds. Toyoda supplied relevant bounds for prescribed finite families of certain CAT(0) spaces, with constants depending on the family.
The quantifier changes in the new problem: a hypothetical target space can be chosen after the group is constructed, with no dimension bound or previously fixed spectral constant. The manuscript’s statistical estimates are designed to remain valid for every such later choice.
4. What the new manuscripts claim
The September 25, 2026 manuscript claims the existence of a finite connected two-dimensional aspherical simplicial complex and a constant such that every combinatorial loop of length in its universal cover bounds a disk diagram with at most triangular faces.
At the same time, its fundamental group
admits no proper cocompact isometric action on a proper complete CAT(0) space. The linear filling statement makes word-hyperbolic.
The paper deduces that no finite simplicial complex homotopy equivalent to admits a locally CAT(0) geodesic length metric inducing its topology. Consequently, no such finite model admits a compatible locally CAT(−1) metric either.
The construction is an existence argument. The manuscript explicitly does not give an effective size for . The filling constant belongs to the selected finite presentation; a common constant for every group in the auxiliary sequence is not required.
5. Why the advance matters
Importance: 5/5, provisional. If correct, the result resolves a fundamental geometric realization problem by separating word hyperbolicity from geometric CAT(0) actions altogether.
The finite two-dimensional classifying space strengthens the contrast. The obstruction does not come from lacking a finite topological model, nor from needing to discard higher homotopy before constructing one. Even with a finite aspherical model and efficient fillings, the desired metric geometry is claimed to be impossible.
The conclusion also allows the entire finite model to change within its homotopy type. It therefore says more than “this particular triangulation has no suitable metric.” Adding dimensions, subdividing, or choosing another finite homotopy model cannot create the excluded geometric action.
This is distinct from the Artin-group CAT(0) obstruction in family 254. Here the group is specifically word-hyperbolic, so the theorem directly addresses the gap between large-scale negative curvature and metric comparison geometry.
6. The main idea
Relations sparse enough for filling, rich enough for obstruction
The construction begins with the incidence pattern of the projective plane over the two-element field. This small geometry has seven points and seven lines, with three points on each line. The paper modifies its incidence graph by doubling selected point vertices and treating one incidence asymmetrically.
Edges receive signed generator labels. Closed walks in selected labelled graphs become relations. Small subgraphs, called fans, are selected first; a larger labelled graph is admitted only when its required fans have already been selected. Four-letter relations close certain two-edge paths into squares, while six-letter relations control individual generator displacements in hypothetical actions.
This nested sampling creates dependencies. If several large graphs share a fan, its probability cost can only be charged once. A disk diagram can also use the same relator repeatedly. The counting argument tracks both repeated relators and shared prerequisites with multiplicity weights.
Local estimates leave a positive deficit proportional to diagram area that must be paid for by boundary length. A planar-separator argument extends bounded-diagram estimates to arbitrary diagrams. This supplies the claimed linear filling inequality and asphericity without assuming any CAT(0) metric on the constructed complexes.
Statistics that apply to a space chosen later
The geometric obstruction compares labels in several matched copies of the incidence pattern. Ordinary approximate independence is not sufficient, because conditioning on shared labels can make some configurations rare.
The paper builds couplings in which selected conditioning labels agree, each graph match is nearly uniform, and additional labels are nearly independent uniform generators. The error is measured in total variation, which controls expectations of every bounded test function. Hence the same estimate applies to bounded distance-based tests from an action chosen after the random tables are revealed.
The six-letter relations give a uniform bound on each generator’s displacement after normalization. This makes the relevant tests bounded. Without that step, a small probability error could be magnified by a few arbitrarily large displacements.
Passing from energy estimates to an impossible configuration
Suppose infinitely many groups in the selected sequence admitted geometric CAT(0) actions. For each action, minimize the average squared displacement of the generators and rescale so that this positive energy has a common normalization.
Closed-walk relations realize the incidence graph by orbit points. Averaging estimates, using the statistical couplings, bound distances along short paths. Products and limits of these configurations produce a configuration in a complete CAT(0) space. The limiting space need not be proper; properness is used for the original actions.
In the limit, the square relations force exact two-step distance identities. Certain six-cycles acquire a common center, and the triangles from that center become flat. The doubled point vertices then determine four unit directions in the tangent cone at the center. A tangent cone records the infinitesimal geometry of directions and distances from a point.
Incidence constraints force one pair of these directions to have nonnegative scalar product. The exceptional incidence forces another pair to be opposite. Splitting the geometry along the resulting line exposes a flat square in a transverse part, which forces the first pair to be opposite as well. Its scalar product would then be , contradicting the earlier nonnegative bound.
The contradiction rules out infinitely many geometric actions in the chosen sequence. Thus a sufficiently late finite presentation provides the required example. The passage to a finite simplicial model preserves asphericity and transfers the linear filling inequality with a presentation-dependent constant.
7. Scope, limitations, and open questions
The theorem excludes proper cocompact actions on proper complete CAT(0) spaces. It does not classify actions that omit one of these conditions, and it is not a universal fixed-point theorem for all CAT(0) targets.
Its finite-model statement allows arbitrary finite simplicial complexes in the same homotopy type, but requires a geodesic length metric compatible with the given topology. These hypotheses are what let the universal-cover argument recover a forbidden geometric action.
The probabilistic construction chooses parameters for finitely many statistical tests before increasing the generator alphabet. A diagonal selection then handles progressively stronger lists of tests. The claimed existence result should not be read as an explicit small presentation or a practical sampling algorithm with a stated success threshold.
Finally, failure of CAT(0) realization does not remove the group’s hyperbolic features. Thin triangles, linear combinatorial fillings, and the finite aspherical model remain part of the conclusion. The point is precisely that these properties can persist while geometric CAT(0) realization fails.
8. Sources and further reading
Bridson–Haefliger’s book on nonpositive curvature develops CAT(0) geometry, group actions and covering-space consequences. Brady–Crisp’s dimension examples and Ollivier–Wise’s random cubulation theorem give important contrasting precedents.
For the energy method, see Izeki–Kondo–Nayatani’s work on random-group actions. Stark’s survey of hyperbolic boundaries records the realization questions in their recent context. The source manuscript supplies the nested sampling construction and the specific limiting incidence obstruction.
Source manuscripts
- M1. OpenAI, A hyperbolic group with no geometric CAT(0) action (September 25, 2026).
Sources consulted on 22 October 2026 at repository revision 1581d4c. Download the manuscript citations.