Invariant Random Subgroups Overview

• Invariant Random Subgroups (IRS) are conjugation-invariant probability measures on subgroup spaces that generalize normal subgroups, lattices, and stabilizer distributions.
• IRS theory bridges geometric group theory, ergodic theory, and operator algebras to study rigidity, spectral approximation, and entropy in group actions.
• Applications of IRSs span free groups, Lie groups, and hyperbolic actions, offering insights into subgroup dynamics, amenability, and random walk behavior.

Invariant random subgroups (IRS) are conjugation-invariant Borel probability measures on the space of subgroups of a group. For a countable discrete group $\Gamma$, or more generally for a locally compact second countable group $G$, the ambient parameter space is the Chabauty space $\operatorname{Sub}(\Gamma)$ or $\operatorname{Sub}(G)$, and IRSs encode random closed subgroups in a way that simultaneously generalizes normal subgroups, lattices, and stabilizer distributions of probability-measure-preserving actions (Bowen, 2012, Gelander, 2018). Across geometric group theory, ergodic theory, measured group theory, and operator algebras, IRS theory organizes questions about subgroup structure, orbit equivalence, Benjamini–Schramm limits, entropy, spectral approximation, rigidity, and stability; the contrast between the “large zoo” of IRSs in free groups and the strong rigidity phenomena in higher-rank or geometrically constrained settings is one of the subject’s central themes (Bowen, 2012, Gelander, 2015).

1. Foundational framework

For a countable group $\Gamma$, the Chabauty space $\operatorname{Sub}(\Gamma)$ is the compact space of all subgroups of $\Gamma$ endowed with the topology induced by the product topology on $\{0,1\}^\Gamma$; for a locally compact second countable group $G$, $\operatorname{Sub}(G)$ is the space of closed subgroups equipped with the Chabauty topology, which is compact and metrizable (Eisenmann et al., 2014, Gelander, 2018). Conjugation defines a continuous action on this space, and an IRS is precisely a conjugation-invariant Borel probability measure:

$G$0

In the locally compact setting this definition is often stated for closed subgroups; in the discrete setting every subgroup is closed (Biringer et al., 2014, Gelander, 2015).

The basic construction of an IRS comes from stabilizers. If a group acts on a standard probability space by probability-measure-preserving transformations, then the stabilizer map is measurable, and pushing forward the measure produces an IRS. In the free-group setting this can be written explicitly: if $G$1 is a homomorphism into the full group of a measured equivalence relation, then

$G$2

and $G$3 is conjugation invariant because the stabilizer map is $G$4-equivariant and $G$5 is invariant under the action (Eisenmann et al., 2014). Conversely, every IRS of a countable group arises as the stabilizer distribution of some probability-measure-preserving action (Bowen, 2012, Gelander, 2018).

The space of IRSs is compact in the weak-* topology, and its ergodic points are the extreme points of the compact convex set of invariant measures (Gelander, 2015, Gelander, 2018). This convex structure is central in several directions: ergodic decomposition, approximation by finite-index IRSs or lattice IRSs, and the distinction between Bauer-type rigidity and Poulsen-type abundance.

A second foundational axis is the relation between IRSs and homogeneous spaces. For $G$6 locally compact second countable and $G$7 an invariant subgroup measure, every $G$8 in the support of $G$9 admits a nontrivial $\operatorname{Sub}(\Gamma)$0-invariant Borel measure on $\operatorname{Sub}(\Gamma)$1, equivalently

$\operatorname{Sub}(\Gamma)$2

almost surely (Biringer et al., 2014). This “unimodularity of invariant random subgroups” supplies measurable families of invariant measures on coset spaces and leads to mass transport principles for IRSs supported on discrete or compact subgroups. In the unimodular case, the mass transport principle takes the form

$\operatorname{Sub}(\Gamma)$3

linking IRSs to unimodular random Schreier graphs (Biringer et al., 2014).

2. Free groups: abundance, genericity, and the stabilizer model

Free groups provide the most expansive IRS landscape in the supplied corpus. Bowen showed that each nonabelian free group has a large “zoo” of IRSs, and that the convex set $\operatorname{Sub}(\Gamma)$4 of IRSs supported on infinite-index subgroups is a Poulsen simplex; in particular, ergodic infinite-index IRSs form a dense $\operatorname{Sub}(\Gamma)$5 and are homeomorphic to $\operatorname{Sub}(\Gamma)$6 (Bowen, 2012). He also showed that the conjugation action on $\operatorname{Sub}(\Gamma)$7 is universal for ergodic aperiodic discrete p.m.p. equivalence relations of cost less than $\operatorname{Sub}(\Gamma)$8, and that symbolic dynamics can be encoded inside subgroup dynamics on $\operatorname{Sub}(\Gamma)$9 (Bowen, 2012).

A complementary model arises from measured equivalence relations and full groups. Fix a measure-preserving equivalence relation $\operatorname{Sub}(G)$0 with countable classes on $\operatorname{Sub}(G)$1, equip the full group $\operatorname{Sub}(G)$2 with the uniform metric

$\operatorname{Sub}(G)$3

and consider $\operatorname{Sub}(G)$4, a Polish space (Eisenmann et al., 2014). A Baire generic homomorphism yields a “generic IRS” via random stabilizers, while the lean aperiodic model

$\operatorname{Sub}(G)$5

forces infinite orbits (Eisenmann et al., 2014).

This generic model exhibits a sharp hyperfiniteness dichotomy. For $\operatorname{Sub}(G)$6, the following are equivalent for generic $\operatorname{Sub}(G)$7: $\operatorname{Sub}(G)$8 is hyperfinite, and the index random variable

$\operatorname{Sub}(G)$9

is finite almost surely (Eisenmann et al., 2014). In particular, for the standard ergodic aperiodic hyperfinite relation $\Gamma$0, generic IRSs in $\Gamma$1 are finite-index almost surely, whereas for $\Gamma$2 the generic action spans $\Gamma$3 and yields infinite index almost surely (Eisenmann et al., 2014).

The lean aperiodic model shifts the generic picture. For ergodic $\Gamma$4, a residual set of $\Gamma$5 yields IRSs that are co-amenable almost surely, and a generic $\Gamma$6 also satisfies

$\Gamma$7

almost surely, so the IRS is supported on core-free subgroups (Eisenmann et al., 2014). For $\Gamma$8, high transitivity is generic in the lean aperiodic model, and the corresponding IRS is supported on maximal subgroups (Eisenmann et al., 2014).

Free groups also provide explicit one-parameter IRS families with analytic applications. “Intersectional IRSs” are constructed from a subgroup $\Gamma$9 with infinite conjugacy class by taking Bernoulli-$\operatorname{Sub}(\Gamma)$0 percolation on the coset space $\operatorname{Sub}(\Gamma)$1 and pushing forward via

$\operatorname{Sub}(\Gamma)$2

The resulting ergodic IRSs $\operatorname{Sub}(\Gamma)$3 interpolate between two normal cores $\operatorname{Sub}(\Gamma)$4 and $\operatorname{Sub}(\Gamma)$5, with

$\operatorname{Sub}(\Gamma)$6

and, under a Liouville hypothesis at $\operatorname{Sub}(\Gamma)$7, yield continuous entropy profiles for Poisson bundles (Hartman et al., 2017). For finitely supported generating measures on free groups, this machinery gives full realization of the interval $\operatorname{Sub}(\Gamma)$8 by ergodic stationary actions arising from IRSs (Hartman et al., 2017).

3. Geometric density, hyperbolic actions, and subgroup rigidity

In geometric settings, IRSs are often forced to inherit the ambient action’s largeness. For countable groups acting on hyperbolic spaces of general type, an ergodic IRS satisfies a sharp dichotomy: it is either geometrically dense or supported on the elliptic radical (Osin, 2015). Here geometric density means equality of limit sets together with the absence of boundary fixed points. In the acylindrically hyperbolic case, this yields the statement that an ergodic IRS is either supported on subgroups of the finite normal subgroup $\operatorname{Sub}(\Gamma)$9 or is geometrically dense; correspondingly, stabilizers in ergodic p.m.p. actions are either finite almost surely or acylindrically hyperbolic almost surely (Osin, 2015).

An analogous Borel-density phenomenon holds for CAT(0) spaces. If $\Gamma$0 is irreducible, $\Gamma$1, and either has finite telescopic dimension or is proper with finite-dimensional Tits boundary, then a faithful, continuous, geometrically dense action $\Gamma$2 forces any nontrivial IRS to act geometrically densely almost surely (Duchesne et al., 2014). In product spaces, the same holds provided the pushed-forward IRSs to all factors are nontrivial (Duchesne et al., 2014). This extends the classical density paradigm for lattices and finite-covolume subgroups to IRSs in nonpositively curved geometry.

Groups acting on rooted trees exhibit a different rigidity pattern tied to boundary actions. For $\Gamma$3, a nontrivial ergodic IRS acting without fixed points on the boundary contains a level stabilizer, hence is the law of a random conjugate of a finite-index subgroup (Bencs et al., 2018). For finitary regular branch groups, an ergodic IRS contains the derived subgroup of a generalized rigid level stabilizer, while weakly branch groups admit continuum many distinct atomless ergodic IRSs (Bencs et al., 2018). This combination of large-subgroup containment and profuse atomless examples is characteristic of the rooted-tree setting.

Semidirect products display yet another rigidity/structure interface. For $\Gamma$4, every nontrivial ergodic IRS is either of the form $\Gamma$5 for an IRS $\Gamma$6 of $\Gamma$7, or is induced from an IRS of $\Gamma$8 for some lattice $\Gamma$9 (Biringer et al., 2017). For parabolic subgroups $\{0,1\}^\Gamma$0 of $\{0,1\}^\Gamma$1, the ergodic IRSs are exactly

$\{0,1\}^\Gamma$2

with $\{0,1\}^\Gamma$3 an ergodic IRS of a specific kernel subgroup $\{0,1\}^\Gamma$4 determined by the action on $\{0,1\}^\Gamma$5 (Biringer et al., 2017). This classification is driven by cocycle analysis on semidirect products and transverse-IRS constraints.

4. Lie groups, lattices, and semisimple rigidity

A recurring theme is that IRSs generalize lattices while retaining much of lattice rigidity in semisimple settings. In connected noncompact simple Lie groups, proper IRSs are discrete and almost surely either trivial or Zariski-dense, a measure-theoretic extension of Borel density (Gelander, 2018, Gelander, 2015). In higher rank with property (T), the Stuck–Zimmer theorem implies that irreducible ergodic IRSs are exactly $\{0,1\}^\Gamma$6, $\{0,1\}^\Gamma$7, and the lattice IRSs $\{0,1\}^\Gamma$8 (Gelander, 2018). This rigidity underlies higher-rank Benjamini–Schramm convergence and limit multiplicity theorems.

The lattice-induced IRS

$\{0,1\}^\Gamma$9

is the basic bridge between subgroup geometry and random-subgroup compactifications (Gelander, 2018). In semisimple Lie groups, weak uniform discreteness of discrete IRSs strengthens Kazhdan–Margulis: for every $G$0 there exists an identity neighborhood $G$1 such that

$G$2

for every discrete IRS $G$3 (Gelander, 2015). Reformulated for p.m.p. actions, almost every stabilizer is locally trivial on a set of measure at least $G$4 (Gelander, 2015).

This IRS perspective is central in Benjamini–Schramm convergence. For locally symmetric spaces $G$5, convergence $G$6 is equivalent to geometric local convergence to the universal cover, and in higher rank implies normalized Betti number and relative Plancherel convergence (Gelander, 2018). Over non-Archimedean local fields, an analogous theory holds for higher-rank semisimple analytic groups: any accumulation point of the IRSs associated to pairwise non-conjugate irreducible lattices is central, yielding Benjamini–Schramm convergence of finite quotients of Bruhat–Tits buildings and convergence of normalized Betti numbers (Gelander et al., 2017).

Rank one behaves differently. In $G$7 there are uncountable families of exotic IRSs not induced from lattices, produced by random bi-infinite gluings of compact hyperbolic manifolds with totally geodesic boundary, by random trees of pants in dimension $G$8, and by limits of mapping tori in dimension $G$9 (Abert et al., 2016). These examples underscore the gap between higher-rank rigidity and rank-one abundance.

Linear groups supply another decisive rigidity theorem. If $\operatorname{Sub}(G)$0 is countable, non-amenable, and has simple center-free Zariski closure, then there exists a free subgroup $\operatorname{Sub}(G)$1 and a non-discrete group topology $\operatorname{Sub}(G)$2 such that for every nontrivial IRS $\operatorname{Sub}(G)$3, $\operatorname{Sub}(G)$4-almost every subgroup is open in $\operatorname{Sub}(G)$5, $\operatorname{Sub}(G)$6 almost surely, and the map $\operatorname{Sub}(G)$7 gives an $\operatorname{Sub}(G)$8-invariant isomorphism of probability spaces between $\operatorname{Sub}(G)$9 and its image in $G$00 (Gelander et al., 2014). The same paper shows that an IRS supported on amenable subgroups of a linear group is supported on the amenable radical (Gelander et al., 2014).

5. Random walks, entropy, stability, and other structural applications

IRSs interact strongly with random walks. The IRS version of Kesten’s theorem states that for a countable group $G$01 generated by a finite symmetric set $G$02 and an IRS $G$03, the equality

$G$04

holds almost surely if and only if $G$05 is amenable almost surely (Abert et al., 2012). This generalizes Kesten’s theorem for normal subgroups and implies strict spectral increase on Schreier graphs for nonamenable IRSs. One consequence is that, for a Cayley graph of a linear group with no amenable normal subgroups, any sequence of finite quotients that spectrally approximates the Cayley graph must converge to it in Benjamini–Schramm sense; in particular, infinite sequences of finite Ramanujan Schreier graphs have essentially large girth (Abert et al., 2012).

Entropy realization is another major application. For an IRS $G$06 of a countable group $G$07 and a generating probability measure $G$08, Bowen’s formula for the Poisson bundle is

$G$09

and $G$10 is ergodic if and only if $G$11 is ergodic (Hartman et al., 2017). Combined with intersectional IRSs, this yields continuum families of ergodic stationary actions with controlled Furstenberg entropy, including full realization for free groups and several other classes (Hartman et al., 2017).

IRSs also organize permutation stability. For a finitely generated amenable group $G$12, P-stability is equivalent to the statement that every $G$13 is co-sofic, meaning a weak-* limit of finite-index IRSs (Becker et al., 2018). This equivalence is mediated by local statistics of finite Schreier graphs, the metrics $G$14 and $G$15, and hyperfiniteness phenomena for amenable actions (Becker et al., 2018). The same framework yields positive results for virtually polycyclic groups and $G$16, and negative results for specific amenable groups such as Abels’ group (Becker et al., 2018).

The theory also extends to IRSs supported on conjugacy orbits. For connected Lie groups, the existence of an IRS supported on the orbit of a semisimple Levi subgroup, a maximal compact subgroup, or a maximal diagonalizable subgroup is equivalent to stringent normality or centrality conditions; for connected real algebraic groups, an IRS supported on the orbit of a Borel subgroup exists if and only if the group is amenable (Choudhuri et al., 11 Jun 2025). In the concrete case $G$17, the paper determines orbit closures of several natural subgroup orbits in $G$18, displaying explicit degeneration patterns (Choudhuri et al., 11 Jun 2025).

6. Compact IRSs, operator-algebraic directions, and current frontiers

A recent rigidity direction concerns compact support. An IRS is compact if it gives full measure to compact subgroups. In real Lie groups, compactly generated $G$19-adic Lie groups, locally compact hyperbolic groups, and infinitely ended groups, ergodic compact IRSs are forced into compact normal subgroups; in general $G$20-adic Lie groups they are contained in the locally elliptic radical, and in totally disconnected locally compact groups they are contained in

$G$21

the intersection of Levi subgroups of inner automorphisms (Cohen et al., 17 Mar 2026). This sharply constrains compact stabilizer distributions and yields triviality in simple settings with no nontrivial compact normal subgroup (Cohen et al., 17 Mar 2026).

Operator-algebraic applications pass through traces associated to IRSs. For an IRS $G$22 of a discrete group $G$23, define

$G$24

This packages the IRS as a trace and a von Neumann algebra (Manzoor, 21 Aug 2025). In free groups, co-sofic IRSs satisfy an IRS version of Lück’s determinant conjecture, and there exist IRSs satisfying the determinant conjecture that are not co-hyperlinear and hence not co-sofic (Manzoor, 21 Aug 2025). The construction uses non-local games, $G$25, and a computable hierarchy for “Det-IRS strategies,” producing evidence that the determinant conjecture is weaker than soficity in the IRS framework (Manzoor, 21 Aug 2025).

A parallel complexity-theoretic frontier concerns surjunctivity. There exists a surjunctive IRS of a free group that is not co-sofic, showing that surjunctivity does not characterize co-soficity for IRSs (Bowen et al., 10 Nov 2025). The argument combines subgroup tests, local outer approximations, and an IRS analogue of Seward’s maximal Rokhlin entropy of Bernoulli shifts; IRSs satisfying this RBS property are surjunctive, all co-sofic IRSs satisfy RBS, and yet RBS does not force co-soficity (Bowen et al., 10 Nov 2025).

Several open problems recur across these directions. It is unknown in general whether generic IRSs in the lean aperiodic model are highly transitive for all ergodic equivalence relations (Eisenmann et al., 2014). Gelander’s survey asks whether every IRS is co-sofic, with the free-group case linked to the Aldous–Lyons conjecture (Gelander, 2015). The determinant-conjecture paper asks whether all IRSs of a free group satisfy the IRS determinant conjecture (Manzoor, 21 Aug 2025). The compact-IRS work raises the broader question whether, for every compactly generated lcsc group, the normal closure of an ergodic compact IRS must be compact (Cohen et al., 17 Mar 2026). These questions suggest that IRS theory remains a meeting point for ergodic, geometric, spectral, and algorithmic phenomena rather than a completed classification program.