WorksMathematics
An example of a non-sofic hyperbolic group
Abstract
We construct a torsion-free, two-dimensional, word-hyperbolic group which is not sofic. Our construction builds on OpenAI's construction of a non-residually finite hyperbolic group, see OpenAI 252.
Introduction
For a positive integer , let denote the symmetric group on . Its normalized Hamming metric and the fixed-point fraction of a permutation are given by
We write id for the identity permutation. A group is sofic if, for every finite subset containing and every , there are a positive integer and a map , sending to the identity permutation, such that
Thus multiplication of elements in is preserved on most points of the finite set, while each specified nonidentity element fixes only a small proportion of the points. The notion was introduced by Gromov [8]; see also Weiss [19]. All amenable groups and all residually finite groups are sofic; see [16]. The question whether all groups are sofic was open for a long time.
The first proof of the existence of a non-sofic group was given by OpenAI [15]. Kun and the second author [11] constructed non-sofic generalized wreath products and doubles of residually finite groups. Chapman and Peled [4] gave a conditional construction of a non-sofic hyperbolic group using random simplicial complexes. The aim of this paper is to construct a torsion-free word-hyperbolic group which is not sofic. We build on OpenAI’s construction of a non-residually finite torsion-free hyperbolic group [14]. Recall that a finitely generated group is word-hyperbolic if geodesic triangles in a Cayley graph for a finite generating set are uniformly thin: each side lies within a fixed distance of the other two sides, see [3] for details.
We obtain the group as the fundamental group of a finite complex. By a triangle complex we mean a two-dimensional cell complex obtained from a graph by attaching triangles, each side being mapped homeomorphically onto an edge. Several edges may have the same endpoints. In our construction, each closed face is embedded. We assign a length to every edge and give the faces the metrics of triangles in the hyperbolic plane of curvature , with the prescribed side lengths. The complex then carries the induced path metric.
A geodesic metric space is if distances between points on the sides of a geodesic triangle are at most the distances between the corresponding points of a triangle in the hyperbolic plane with the same side lengths. Corresponding points have the same distances from the endpoints of the corresponding sides. The metric is locally if each point has a neighbourhood which, with its intrinsic path metric, is . We refer to [3] for these notions and their basic properties. Our main result is the following.
Theorem A. There exists a finite connected triangle complex with a locally metric such that is torsion-free, word-hyperbolic, and non-sofic.
2020 Mathematics Subject Classification. Primary 20F67; Secondary 20E26, 20F65, 05C38, 15A03.
Key words and phrases. Sofic groups, hyperbolic groups, negatively curved complexes, matrix rank, finite-field interpolation, graphs of large girth.
Put . The presentation of in § 3 has
generators,
relations.
The group in Theorem A admits no proper cocompact action by cubical automorphisms on a CAT(0) cube complex. Indeed, by Agol’s theorem [1 Theorem 1.1] and the work of Haglund and Wise [9 Theorem 4.4], a word-hyperbolic group admitting such an action is residually finite and therefore sofic.
Corollary B. There exists a nontrivial torsion-free word-hyperbolic group which has no nontrivial sofic quotient.
Proof. If every nontrivial torsion-free word-hyperbolic group had a nontrivial sofic quotient, the theorem of Kapovich and Wise [10 Theorem 3.4], applied to the class of sofic groups, would imply that every torsion-free word-hyperbolic group is residually sofic and hence sofic by [7]. This contradicts Theorem A. □
The construction starts with a family of affine lines in , where is the field with elements and is prime. Its incidence graph has one vertex for each point and each line, and an edge whenever the point lies on the line. This is the graph of Lazebnik and Ustimenko [12]. Its girth, the length of its shortest simple cycle, is at least ten [13 Proposition 2.1(iii)]. In § 2, we select a family of lines with distinct distinguished points, called marks, and delete the incidence with its own mark for each line in a specified subfamily.
The retained incidences determine a triangle complex with three vertices. Each retained line gives an array of edges between two of these vertices. Choosing two rows and two columns gives a closed four-edge path, traversed with alternating orientations.
In § 3, we combine polynomial interpolation with the rank estimate of Alon–Hall–Knauer–Pinchasi–Yuster [2 Theorem 6]. We prove that every assignment of permutations to the oriented edges has either a triangular boundary word moving a fixed positive fraction of points or one of the four-edge words fixing such a fraction. The bound is independent of the degree of the permutations.
In § 4, we use the girth bound to construct a locally CAT() metric and prove that the four-edge paths are non-nullhomotopic. Soficity would allow both fractions in the preceding paragraph to be arbitrarily small. § 5 derives the contradiction.
Acknowledgments. This paper is based on an AI-generated construction obtained by Harjas Sandhu on October 7, 2026, which was in turn based on the OpenAI breakthrough [14]. The second author worked through the proof to make it readable and to reduce unnecessary overhead. OpenAI’s GPT Astra helped draft this document. We simplified various constructions and related them to previous work. The aim of this note is to give a reasonably self-contained presentation of the result. An independent AI-assisted proof was given by Douglas–Mghirbi [5].
From the construction in [14] we use the marked lines, triangular relations associated to Cartesian products, interpolation identity, and comparison of real and finite-field ranks. We prove these ingredients and combine them with the rank theorem and the incidence relations of to obtain the inequality for approximate permutation assignments. We thank Ian Agol for pointing out the relevance of Dvir’s work [6] to the crucial interpolation argument, see Lemma 2.1. We view this document as a commentary on OpenAI 252 [14].
Marked lines from
We fix
The integer is prime.1 We consider two copies of , points and labels. We write points as and labels as . The incidence relation between points and labels is defined by
These equations define a bipartite graph as in [13 Section 2]. By [13 Proposition 2.1], this graph is -regular, has vertices in each part, and has girth at least ten. These properties also follow from [12 Proposition 2.1 and Theorem 3.3].
For a fixed label, the incident points form the full affine line
Different labels determine different lines. Indeed, the normalized direction determines , and the point with first coordinate zero then determines .
Consider the set of incidence edges satisfying . Given a point , the condition and the incidence equations uniquely determine a label by
Thus each point is contained in exactly one of these incidence edges. Conversely, for a fixed label , the parametrization (2) gives exactly one incident point with first coordinate . Thus each label is also contained in exactly one of these incidence edges. These incidence edges therefore form a perfect matching. Since each matched pair satisfies (1), the point matched to a label lies on the corresponding affine line. We henceforth index each line by its matched point, writing for the line marked at .
Put , using the standard integer representatives, and set
We retain all lines with and distinguish those with . For each distinguished line, we delete only its incidence with its own mark. All incidences of the other retained lines remain. Thus, for , the remaining incidences at are indexed by
The bipartite graph on , whose two parts represent points and retained lines, has an edge from to precisely when . It is a subgraph of and therefore has girth at least ten.
For a fixed point and a prescribed first label coordinate , (1) uniquely determines . The first coordinate of the matched point is , so the retained lines are precisely those with . Thus retained lines pass through each point before deletion. A point loses one incidence if it belongs to , and none otherwise. Consequently,
For each point , put if and otherwise. All these factors are at least . Choose a bijection
The sets may overlap, but their bijections are chosen separately for each point.
The following interpolation argument is adapted from [14 Section 2.5, equation (2.13)] to the grid . It uses low-degree polynomials on affine lines, as in Dvir’s proof of the finite-field Kakeya conjecture [6], applied here to the entries of rank-one matrices.
Lemma 2.1. There are matrices , indexed by , such that
Proof. For , let be the Lagrange basis polynomial associated with the node in , given by
It is the unique polynomial of degree at most satisfying
For and , set
We have . For , equals 1 if and 0 otherwise, so for . Each has total degree at most . It follows that every entry of , restricted to an affine line, is a polynomial of degree at most .
It is elementary to see that the sum of the values of a polynomial of degree less than over is zero. Applying this observation entrywise gives
For , deleting the term at its mark gives . For , the full line is retained, so its sum remains zero. This proves (5). □
A finite permutation obstruction
Our next aim is to define a connected triangle complex whose fundamental group will have all the required properties. We use the construction in § 2 and write . We follow the construction in [14 Section 2.3]. Consider three vertices and define oriented edges
Here , , , , and . For each point , , and , attach a triangle with boundary, read in path order,
Denote the resulting finite triangle cell complex by . There are - and -edges, and hence is connected.
Let . Choose a point . The edges and form a spanning tree of the graph of vertices and edges of . Let be the constant path at . For an oriented edge , define
Denote the elements associated with , , and by , , and , respectively. We multiply loops in path order. Collapsing the spanning tree gives a finite presentation of with generators , with the index ranges specified above, and relations
for every point , , and , together with
Our next aim is to understand how this presentation interacts with the assignment of permutations to the generators. We identify a permutation with the matrix defined by , where is the -th standard coordinate vector. Thus inverse matrices are transposes, and . Assign arbitrary permutations of degree to the oriented edges of . Denote the permutations assigned to , , and by , , and , respectively. Reversing an edge replaces its permutation by its inverse. For a triangular face indexed by , define
This is exactly the defect of the corresponding relation being satisfied in the permutation group. For , and , define the rectangular product
Let be the average of over all ordered tuples with and .
The key technical result is the following theorem. It gives a lower bound on the sum of defects of the defining relations together with a measure of non-triviality of the rectangular words . We will later show that all these rectangular words are non-trivial in .
Theorem 3.1. Every such assignment, in every degree , satisfies
The proof of this theorem occupies the remainder of this section and needs various lemmas. The following result appears in [14 Lemma 2.4]. It is also a consequence of Gabber’s bound on torsion in cokernels; see [18 Lemma 1] and [17 Lemma 3.1]. We include a direct proof.
Lemma 3.2. If an integer matrix has row norms at most and is prime, then
Proof. Set and choose a nonsingular minor if . By Hadamard’s inequality, we have . Let and choose rows forming a basis modulo . Subtract integer lifts of their linear combinations from the remaining rows. These row operations preserve the determinant and make rows divisible by . Hence divides , so
Since , this implies . If , there is nothing to prove. □
Lemma 3.3. For , let , an integer block matrix. For every ,
Proof. We apply [2 Theorem 6]. For a nonzero 0-1 matrix , this theorem gives the following estimate in terms of the bipartite graph whose vertices are the rows and columns and whose edges correspond to entries equal to 1:
Here is the number of edges, the vertex degrees, and the number of unoriented four-cycles. For , each of the vertices has degree and . Every four-cycle uses two different row blocks and two different column blocks: two rows of one permutation block cannot share a column, and the analogous assertion holds for columns. For each ordered choice of these four blocks, the cycles are counted by the fixed points of (10). Each unoriented cycle is counted four times, giving . By (13), we get
Each row of has exactly entries equal to 1, so Lemma 3.2 proves the first bound in (12). For the second inequality, use for to obtain
The following lemma extends the exact factorization and rank bounds in [14 equations (2.11)–(2.12)] to permutation assignments with triangular defects.
Lemma 3.4. Over , let , an matrix. For every ,
Proof. Permutation matrices differing on points have at most nonzero columns in . Thus over every field. Bi-invariance gives
Fix a point . In , replace each permutation matrix by the corresponding product in the preceding display. For fixed , the pairs , with , traverse the whole Cartesian product once. The resulting matrix is
Both factors have rank at most , so their product has rank at most . Adding the difference in one matrix block increases the rank by at most the rank of that difference. Subadditivity of rank now gives (14).
Proof of Theorem 3.1. Recall that and . We use the identifications
Thus the matrices have blocks, each of size . The symbol denotes the Kronecker product: the block of is . We write for the block diagonal matrix with diagonal blocks and zero off-diagonal blocks. From (5), we obtain the tensor identity of [14 equation (2.14)]:
Indeed, expanding and changing the order of summation makes the coefficient of equal to for , and zero for . The rank on the left of (16) is the sum of the individual ranks. Since by Lemma 2.1, each summand on the right satisfies . We apply Lemma 3.4 at every point . Each triangular face has a unique associated point, so its defect is counted exactly once. Rank subadditivity therefore gives
Using the second bound of Lemma 3.3 and cancelling the positive real number , we obtain
Since and , we have . This proves (11).
The negatively curved complex
We equip every face of the complex from § 3 with the hyperbolic triangle of curvature with angles
The angle sum is , so this triangle exists. Assign to all - and -edges the equal side lengths opposite and , respectively, and to all -edges the side length opposite . This includes edges contained in no face. All gluings are therefore isometric.
An edge germ at a vertex is an incident edge together with that endpoint. The angular link has one vertex for each edge germ and one edge for each face corner. The length of a link edge is the angle at the corresponding corner. A path in a graph is reduced if it has no immediate edge reversal. A reduced closed path is cyclically reduced if the last edge is not the inverse of the first edge. We set the distance between directions in different link components equal to infinity.
We adapt the link argument of [14 Lemma 2.3] to the present girth bound and angles.
Lemma 4.1. Every nonempty cyclically reduced closed path in any vertex link has length at least . At either or , two different -directions with the same index have distance at least , or belong to different components.
Proof. At , the link is simple and bipartite: a corner between and determines uniquely by and . Every nonempty cyclically reduced closed path therefore has at least four edges and length at least .
We map the link at to by sending the direction to the line vertex and the direction to the point . Each corner maps to the corresponding incidence edge. We claim that this is an immersion, that is, the map is injective on incident edges at every vertex. Indeed, at , a fixed point determines at most one corner, namely that of the face indexed by if . At , a fixed index determines modulo . There is at most one such , since . At , the same argument uses modulo and .
Under these immersions, both a nonempty cyclically reduced closed link path and a reduced path joining distinct -directions with the same index map to nonempty closed walks in with no immediate reversal at an internal vertex. Every such walk contains a simple cycle and hence has at least ten edges. Since link edges at and have length , both assertions follow. □
Proposition 4.2. The universal cover is a proper CAT() space. The group is torsion-free and word-hyperbolic, and is a finite two-dimensional classifying space for .
Proof. We first subdivide the complex so that cells meet in single faces. Subdivide each edge at its midpoint and each face by joining an interior point to its vertices and edge midpoints by geodesics. Since the closed cells embed and each edge and face has a distinct new vertex, the subdivision is a finite simplicial complex with the same intrinsic metric. The angular link at a new face-interior vertex is a circle of length . At an edge-interior vertex it consists of length- paths between the two edge directions.
The original angular links have the same metric after subdivision. By Lemma 4.1, their injective circuits have length at least , and all circuits in the new links have length at least . It follows from the metric graph criterion [3 Lemma II.5.6] and the link condition [3 Theorem II.5.2] that is locally CAT().
The lifted complex is locally finite and has finitely many shapes. Its length metric is complete by [3 Theorem I.7.13]. By local compactness and [3 Proposition I.3.7], it is proper and geodesic. The Cartan–Hadamard theorem [3 Theorem II.4.1] now implies that the universal cover is globally CAT(). Deck transformations act freely, properly and cocompactly by isometries. A finite group of isometries of a complete CAT(0) space fixes a point [3 Corollary II.2.8]. Since the deck action is free, no nonidentity deck transformation has finite order. The space is hyperbolic [3 Chapter III.H], so the Švarc–Milnor lemma [3 Proposition I.8.19] implies that is word-hyperbolic. Since is contractible, is a classifying space. □
It remains to show that the rectangular words studied before are all non-trivial in . This is the content of the following lemma. We adapt the argument following [14 equation (2.6)] to the CAT() metric constructed above.
Lemma 4.3. For , and , the element is nonidentity in .
Proof. The edge path is a loop based at , and is the class of the loop based at . At every turn of , including the closing turn, the two -directions are distinct and have the same index . By Lemma 4.1, their link distance is greater than . It follows that is a positive-length closed local geodesic, by the link description of angles [3 Chapter I.7]. If this loop were nullhomotopic, it would lift to a positive-length closed local geodesic in . Every local geodesic in a CAT() space is a geodesic [3 Proposition II.1.4]. This is impossible for a positive-length closed path. □
Proof of the main result
Let be a sofic group and let . Given finitely many words in a finite alphabet of elements of and their inverses, choose a finite set containing the alphabet, its inverses, all word prefixes, and 1. Let satisfy the inequalities in the definition of soficity from the introduction. For a word of length with inverse letters, evaluate it using the assigned permutations of the positive letters. Its distance from the image under of the corresponding group element is at most . Indeed, each inverse substitution changes the evaluated permutation by less than , since is within of . Each successive prefix multiplication introduces an error of less than . The asserted bound follows from bi-invariance and the triangle inequality.
Proof of Theorem A. The geometric assertions follow from Proposition 4.2. By Lemma 4.3, all rectangular elements are nonidentity. Assume that is sofic. Apply the preceding estimate in to all triangular and rectangular words, with the elements as the alphabet. Assign to each edge symbol. Each triangular word has length three, one inverse letter, and group value . Hence . Each rectangular word has length four and two inverse letters, so its evaluated permutation is within of . Since , the latter permutation has fixed-point fraction less than . Thus the evaluated rectangle has fixed-point fraction less than , and . These estimates also hold when some alphabet elements equal .
Each point indexes triangular faces. By the degree count in § 2, there are triangle indices in total. The data are fixed independently of and . Applying (11) from Theorem 3.1 to the permutations just constructed, and using the distinguished lines, gives
Choosing gives a contradiction to our choices for and .
References
Harjas Sandhu, The University of Queensland, St Lucia, QLD 4072, Australia
Email address: [email protected]
Andreas Thom, Faculty of Mathematics, TU Dresden, 01062 Dresden, Germany
Email address: [email protected]
References
- [1]I. Agol, The virtual Haken conjecture, with an appendix by I. Agol, D. Groves, and J. Manning, Doc. Math. 18 (2013), 1045–1087.arxiv.org/abs/1204.2810
- [2]N. Alon, H. T. Hall, C. Knauer, R. Pinchasi, and R. Yuster, On graphs and algebraic graphs that do not contain cycles of length 4, J. Graph Theory 68 (2011), no. 2, 91–102.DOI
- [3]M. R. Bridson and A. Haefliger, Metric Spaces of Non-Positive Curvature, Grundlehren der mathematischen Wissenschaften, vol. 319, Springer-Verlag, Berlin, 1999.DOI
- [4]M. Chapman and Y. Peled, Cocycle stability in permutations of random simplicial complexes, preprint, 2025, arXiv:2509.21566v2.arxiv.org/abs/2509.21566
- [5]S. Douglas and N. Mghirbi, Non-soficity of OpenAI’s torsion-free hyperbolic group, preprint, 2026, https://github.com.com/Asiape/hyperbolic-group-not-sofic/blob/main/paper/paper.pdf.DOI
- [6]Z. Dvir, On the size of Kakeya sets in finite fields, J. Amer. Math. Soc. 22 (2009), no. 4, 1093–1097.DOI
- [7]G. Elek and E. Szabó, On sofic groups, J. Group Theory 9 (2006), no. 2, 161–171.
- [8]M. Gromov, Endomorphisms of symbolic algebraic varieties, J. Eur. Math. Soc. 1 (1999), no. 2, 109–197.DOI
- [9]F. Haglund and D. T. Wise, Special cube complexes, Geom. Funct. Anal. 17 (2008), no. 5, 1551–1620.DOI
- [10]I. Kapovich and D. T. Wise, The equivalence of some residual properties of word-hyperbolic groups, J. Algebra 223 (2000), no. 2, 562–583.DOI
- [11]G. Kun and A. Thom, Nonsofic wreath products of residually finite groups, preprint, 2026, arXiv:2608.06222v3.arxiv.org/abs/2608.06222
- [12]F. Lazebnik and V. A. Ustimenko, Explicit construction of graphs with an arbitrary large girth and of large size, Discrete Appl. Math. 60 (1995), nos. 1–3, 275–284.DOI
- [13]F. Lazebnik, V. A. Ustimenko, and A. J. Woldar, A new series of dense graphs of high girth, Bull. Amer. Math. Soc. (N.S.) 32 (1995), no. 1, 73–79.arxiv.org/abs/math/9501231
- [14]OpenAI, A torsion-free hyperbolic group that is not residually finite, family 252, manuscript dated 23 September 2026, https://github.com/openai/math/blob/main/preprints/a-torsion-free-hyperbolic-group-that-is-not-residually-finite-September-23-2026/paper.pdf.hub.valency.io/works/W_38ukwq7p
- [15]OpenAI, Nonsofic groups exist, Chapter 3 of Ten Advances in Mathematics and Theoretical Computer Science, 2026, pp. 78–95, updated 6 August 2026, https://cdn.openai.com/pdf/ten-proofs-oai.pdf.
- [16]V. G. Pestov, Hyperlinear and sofic groups: a brief guide, Bull. Symbolic Logic 14 (2008), no. 4, 449–480.arxiv.org/abs/0804.3968
- [17]R. Sauer, Volume and homology growth of aspherical manifolds, Geom. Topol. 20 (2016), no. 2, 1035–1059.DOI
- [18]C. Soulé, Perfect forms and the Vandiver conjecture, J. Reine Angew. Math. 517 (1999), 209–221.
- [19]B. Weiss, Sofic groups and dynamical systems, Sankhyā Ser. A 62 (2000), no. 3, 350–359.