Tits Alternative for CAT(0) Groups

Tits Alternative for CAT$0$ Groups

Definitions and mathematical statement

A CAT$0$ space is a geodesic metric space whose geodesic triangles are no thicker than Euclidean comparison triangles. It is proper if closed bounded sets are compact. A group action by isometries is proper if only finitely many translates of a compact set meet that set, and cocompact if the quotient is compact. Suppose $G$ admits such an action on a proper CAT$0$ space. The target is the alternative \[ H\le G,\ H\text{ finitely generated} \quad\Longrightarrow\quad [H\text{ virtually solvable}\ \text{or}\ F_2\hookrightarrow H]. \] Here virtually solvable means having a solvable subgroup of finite index, and $F_2$ is the free group on two generators. No cubical structure, finite-dimensionality, or geodesic completeness of the space is assumed in the recorded question.

Formulation and concept references: [S1].

Short English statement

Must every finitely generated subgroup of a CAT$0$ group either be solvable up to finite index or contain a nonabelian free group?

Research attempt: elementary actions, free subgroups, and the flat obstruction

Status and geometric hypotheses. This attempt does not prove the Tits alternative for every group in the statement. Properness, cocompactness, and properness of the CAT$(0)$ space are retained; no cubical, two-dimensional, visibility, or geodesic-completeness condition is added to the original target. The inherited paper [S1] treats specified two-dimensional complexes. The primary manuscript of Ji and Wu [E1], checked on 27 September 2026, treats visibility CAT$(0)$ spaces with a bounded packing property. Those hypotheses are not consequences established here for an arbitrary proper CAT$(0)$ space.

The arguments below handle bounded orbits, invariant lines, and torsion-free tree actions. They also give the exact dynamical implication behind a free-subgroup construction and explain why it does not cover all CAT$(0)$ actions.

Lemma 313.1 (bounded subgroups are finite). Suppose a discrete group $H$ acts properly by isometries on a proper CAT$(0)$ space $X$. If one $H$-orbit is bounded, then $H$ is finite.

Proof. Let $Y=Hx_0$ be bounded and consider the radius function $r(x)=\sup_{y\in Y}d(x,y)$. It is continuous and tends to infinity as $d(x,x_0)$ tends to infinity, by the triangle inequality. Properness of $X$ implies that it attains a minimum. This minimizer is unique. Indeed, if $x,x'$ both have minimal radius $R$ and $m$ is their midpoint, the CAT$(0)$ midpoint inequality gives \[ d(m,y)^2\leq\tfrac12d(x,y)^2+\tfrac12d(x',y)^2 -\tfrac14d(x,x')^2 \leq R^2-\tfrac14d(x,x')^2 \] for every $y\in Y$. Distinct minimizers would produce a smaller radius. Every element of $H$ preserves $Y$ and hence fixes its unique circumcenter. A proper action has finite point stabilizers: the set of elements carrying the compact singleton consisting of that point to intersect itself is finite. Therefore $H$ is finite. $\square$

The lemma requires a bounded orbit of the entire subgroup. Knowing separately that each element has some bounded orbit does not, without another theorem, give a common bounded orbit for an infinitely generated or even an arbitrary finitely generated subgroup.

Lemma 313.2 (an invariant line gives a virtually cyclic subgroup). Under the same proper-action assumptions, if $H$ preserves a geodesic line $L\cong\mathbb R$, then $H$ is virtually cyclic.

Proof. Restriction gives a homomorphism $\rho:H\to\operatorname{Isom}(\mathbb R)$ whose kernel fixes each point of $L$ and is finite. Let $H^+$ be the subgroup preserving its orientation, of index at most two. Its image is a subgroup $\Lambda$ of the additive translation group $\mathbb R$.

Properness implies that for every $R$ there are only finitely many elements of $H$ moving a fixed point of $L$ distance at most $R$. To see this, take the compact ball of radius $R$ around that point; each such element carries the point, and thus this ball, to a set intersecting the ball. Hence $\Lambda$ has only finitely many elements in each bounded interval. If it is nonzero, its positive elements have a least member $a>0$: start with any positive element, and use finiteness below it. Division with remainder shows $\Lambda=a\mathbb Z$.

Choose $t\in H^+$ mapping to $a$. Every element of $H^+$ is a product of an element of the finite kernel and a power of $t$. Thus $\langle t\rangle$ has finite index, and it is infinite cyclic. If $\Lambda=0$, $H^+$ and hence $H$ are finite. Either way $H$ is virtually cyclic and therefore virtually solvable. $\square$

This proves the desired alternative for subgroups preserving a line, including the finite-kernel issue which would be missed by simply identifying $H$ with its image in the line isometry group.

A complete tree subcase. Suppose in addition that the ambient space is a simplicial tree and $G$ is torsion-free. Subdivide edges if necessary to remove inversions. Properness makes every vertex stabilizer finite, and torsion-freeness makes it trivial. Thus $G$ acts freely on the tree, as does every subgroup $H$.

The quotient map of graphs $T\to H\backslash T$ is a covering map, since no nonidentity element identifies incident half-edges at a vertex. The tree is simply connected, so $H$ is the fundamental group of the quotient graph. Choosing a maximal tree in that graph proves that $H$ is free. If $H$ is finitely generated, its rank is finite: its abelianization is free abelian on a free basis and could not be finitely generated if that basis were infinite. Rank zero or one gives a solvable group; rank at least two gives a subgroup freely generated by two basis elements. This proves the selected alternative throughout this additional tree-action setting.

Lemma 313.3 (a concrete free-subgroup criterion). Let a group act on a set $Y$. Suppose there are four pairwise disjoint subsets $A_+,A_-,B_+,B_-$ and a point outside their union, and elements $a,b$ satisfying \[ a(Y\setminus A_-)\subseteq A_+, \quad a^{-1}(Y\setminus A_+)\subseteq A_-, \] \[ b(Y\setminus B_-)\subseteq B_+, \quad b^{-1}(Y\setminus B_+)\subseteq B_-. \tag{313.1} \] Then $a,b$ freely generate a group of rank two.

Proof. Take a nonempty freely reduced word in the letters $a^{\pm1},b^{\pm1}$ and apply it, starting from its rightmost letter, to the chosen point outside the four sets. The first letter sends the point into its attracting set. Each next letter has a repelling set different from the current attracting set, because the word is freely reduced. Disjointness permits application of $313.1$, sending the point into the next attracting set. The final result is in one of the four sets, whereas the original point was outside. Thus the word is not the identity. $\square$

This isolates a useful implication for boundary dynamics. If two isometries have disjoint attracting and repelling fixed points on a compact boundary and uniform north--south dynamics away from the repelling point, choose four small disjoint neighborhoods and a point outside them. Sufficiently large powers satisfy $313.1$, so the group contains $F_2$. The uniform north--south property is an additional premise of this argument; it is not established for all axial isometries of all CAT$(0)$ spaces.

Why the general boundary shortcut fails. Translations of the Euclidean plane fix every point of its visual boundary. They have infinite order and unbounded orbits, but do not have north--south boundary dynamics. The group $\mathbb Z^2$ acting by integer translations is a proper cocompact CAT$(0)$ example. It satisfies the solvable side of the desired alternative, yet disproves the inference from "infinite-order isometry" to the dynamical hypotheses of Lemma 313.3.

Nor does the absence of rank-one axes imply virtual solvability. The product $F_2\times F_2$ acts properly and cocompactly on the product of two locally finite trees. It contains a free subgroup as either factor. An axis of an element translating in both factors lies in a Euclidean flat obtained from the product of its two tree axes. If an element translates in only one factor, its axis bounds a flat half-plane obtained by taking a ray in the other tree. Thus this standard action has no rank-one group element, despite satisfying the free-subgroup side. A dichotomy based solely on finding rank-one dynamics would miss product phenomena.

There is a further subgroup issue. Although $G$ acts cocompactly, a finitely generated subgroup $H$ need not act cocompactly on the same space. For instance, $F_2\times\mathbb Z$ acts geometrically on $T\times\mathbb R$, while $F_2\times\{0\}$ has quotient $(F_2\backslash T)\times\mathbb R$, which is not compact. Therefore an argument applying the ambient cocompactness hypothesis anew to every finitely generated subgroup requires justification that is absent in general.

An algebraic stress test for free words. Independently of the CAT$(0)$ geometry, the integer matrices \[ U=\begin{pmatrix}1&2\\0&1\end{pmatrix},\qquad V=\begin{pmatrix}1&0\\2&1\end{pmatrix} \] illustrate the kind of separated projective dynamics behind free-subgroup proofs. On real projective coordinates, nonzero powers of $U$ send $|x|<1$ into $|x|>1$, while nonzero powers of $V$ send $|x|>1$ into $|x|<1$: the second assertion follows from \[ |2nx+1|\geq2|x|-1>|x|\qquad(|x|>1, |n|\geq1). \] The accompanying finite diagnostic multiplies reduced words in these matrices and checks that none in the tested range is the identity. This finite enumeration is only a check of the example; the general injectivity criterion is proved by the set-action argument, not by testing a bounded word length.

Outcome and first gap. Proper CAT$(0)$ actions yield the desired alternative for bounded subgroups, line-preserving subgroups, and the stated torsion-free tree actions. Suitable separated boundary dynamics also force $F_2$. For a general finitely generated subgroup in the selected setting, the missing step is proving virtual solvability or producing a free-subgroup mechanism without assuming visibility, a small dimension, or a special cubical structure. Flat and product examples show why two natural attempts to force that dichotomy are incomplete. The selected assertion remains unresolved by this work.

Sources

• S1 Osajda and Przytycki, Tits Alternative for 2-dimensional CAT$0$ complexes. https://www.cambridge.org/core/journals/forum-of-mathematics-pi/article/tits-alternative-for-2dimensional-mathrm-cat0-complexes/E1B5B96E5D4324F2F3707B4A05B2D86C. Formulation source.
• E1 Ran Ji and Yunhui Wu, The Tits alternative for visibility spaces, arXiv:2510.01008 $2025$. https://arxiv.org/abs/2510.01008. Primary source consulted 27 September 2026; visibility and bounded packing are additional hypotheses in the cited theorem.