A random walk on a group can be stepped forward once, twice, or a hundred times. A natural question is whether it can be stepped forward half a time: is there a probability measure whose two-fold convolution is the original step distribution, and what about 0.3 or 1.7 steps? On the real line and on commutative groups this is the theory of partly divisible distributions. That theory tracks the set $\Lambda(\mu)$ of exponents $t$ for which "the $t$-th power of $\mu$" is still a genuine probability measure. It sits between the two textbook extremes of infinitely divisible laws (every $t$ works) and indecomposable ones (only whole numbers work). The theory stops at commutative groups. There the Fourier transform of a measure is a scalar function, and taking its $t$-th power just means choosing a branch of the logarithm.
This note proposes extending the theory to finite and compact non-commutative groups, such as permutation groups and $SL(2,\mathbb{Z}_p)$. On these groups the Fourier transform is a family of matrices, one per irreducible representation. The main claim is that $\Lambda(\mu)$ keeps its semigroup structure under a principal-branch definition. Its two ends are then controlled by two different edges of the walk's spectrum. The top edge, the eigenvalues just below 1, governs the large-$t$ tail. The edge near the negative real axis governs which small fractional exponents fail. The genuinely non-commutative part is that non-normal matrix blocks make the relevant top edge a pseudospectral one: transient growth of $\|A^t\|$ beyond what the eigenvalues predict. Commutative or conjugation-invariant theory cannot see this effect at all.
Core Claim
Albeverio et al. defined admissible measures on locally compact Abelian groups whose dual group is arcwise connected (Partly Divisible Probability Measures on Locally Compact Abelian Groups). An admissible measure has a Fourier transform that never vanishes and lifts to a continuous logarithm $\psi$ with $\hat\mu = e^{\psi}$. For such measures, $\Lambda(\mu)$ is a closed semigroup containing 1, extending their earlier result on the real line (Partly divisible probability distributions).
The same paper shows the construction breaking on $\mathbb{Z}_2$. There the dual is discrete, so the branch of the logarithm can be chosen independently at each point. $\Lambda$ then becomes an infinite union of closed semigroups. Every finite non-Abelian group has a discrete dual, so naively it inherits this breakdown, with the added complication that the Fourier coefficients $\hat\mu(\pi)=\sum_g \mu(g)\pi(g)$ are $d_\pi\times d_\pi$ matrices.
Our first observation is that the semigroup argument never needed commutativity between different measures. It only needs powers of a single matrix to commute, and they always do. We call $\mu$ principally admissible if no block $A_\pi=\hat\mu(\pi)$ has an eigenvalue on $(-\infty,0]$. In that case the principal power $A_\pi^t=\exp(t\,\mathrm{Log}\,A_\pi)$ is defined for every irreducible representation $\pi$, and $A_\pi^sA_\pi^t=A_\pi^{s+t}$. We define $\mu_t$ by Fourier inversion,
$$\mu_t(g)=\frac{1}{|G|}\sum_{\pi}d_\pi\,\mathrm{tr}\big(\pi(g)^{*}A_\pi^{t}\big),$$
and set $\Lambda_p(\mu)=\{t>0:\mu_t\ge 0\}$. Here $d_\pi$ is the dimension of $\pi$ and the sum runs over irreducible representations. Fixing the principal branch picks out one canonical member of the $\mathbb{Z}_2$-style union. $\Lambda_p(\mu)$ is then closed under addition, because the convolution of two nonnegative measures is nonnegative. It contains 1 and, for aperiodic walks, a tail $[T^*,\infty)$, because $\mu_t$ tends to the uniform measure.
The second, spectral claim concerns the two ends of $\Lambda_p$. Bounding each nontrivial term of the inversion formula by its trace norm gives the sufficient condition
$$\sum_{\pi\neq 1} d_\pi^{2}\,\big\|A_\pi^{t}\big\|_{\mathrm{op}}<1 \;\Longrightarrow\; t\in\Lambda_p(\mu).$$
For normal blocks, $\|A^t\|=\rho(A)^t$ exactly, so $T^*$ is set by the top nontrivial spectral edge and its multiplicities. For non-normal blocks, which arise exactly when $\mu$ is not conjugation-invariant, $\|A^t\|$ can exceed $\rho^t$ for a long transient. $T^*$ should then track the pseudospectral edge instead.
The lower end behaves differently. As $t\to 0$, $\mu_t\approx\delta_e+t\Psi$, where $\Psi$ is the inverse Fourier transform of the logarithm. So $\inf\Lambda_p=0$ exactly when $\Psi$ is nonnegative off the identity, that is, when $\mu$ embeds in a continuous-time walk driven by a group-invariant generator. When it does not, the failing fractional exponents are driven by eigenvalues whose argument approaches $\pm\pi$. These are the eigenvalues near the cut where the principal branch winds. This is the non-commutative analogue of the zero-crossings that end infinite divisibility in the generalized Gaussian family past $p=2$ (Analytical properties of generalized Gaussian distributions).
Research Program
The theoretical work has three parts. First, we prove the semigroup and tail statements for principally admissible measures on finite groups. We then extend them to compact connected Lie groups such as $SU(2)$ and $SO(3)$, where the infinitely many representations require uniform control of the blocks' spectra. Second, we prove the two-edge bounds: an upper bound on $T^*$ in terms of $\sup_t$ of weighted block norms, and a lower obstruction for fractional $t$ in terms of the angular distance of block eigenvalues from the negative axis. Third, we characterize when the principal-branch semigroup is the largest of the branch-indexed semigroups. This would turn the $\mathbb{Z}_2$ pathology from Partly Divisible Probability Measures on Locally Compact Abelian Groups into a statement about which branch is canonical.
The computational work builds an exact solver. It block-diagonalizes the walk via irreducible representations, computes principal matrix powers per block, and inverts the transform to test $\mu_t\ge 0$ on a fine grid of $t$, plus root-finding at sign changes. We deliberately avoid Krylov-type matrix-function approximations. Determinant Estimation under Memory Constraints and Neural Scaling Laws reports that stochastic Lanczos quadrature reaches 55% relative error on operators with near-zero spectral tails. Our question lives at exactly such edges, so smoothing them would corrupt the answer.
Infinite groups are a stated extension, not the core. Markovian Linearization of Random Walks on Groups turns finitely supported walks on free products into nearest-neighbour walks on coloured state spaces and lists the spectral radius as open. Whether our edge bounds transfer through that linearization is a follow-up question.
Baselines
The naive reference point is "integer powers only." It asserts $\Lambda=\{1,2,3,\dots\}$, which is trivially true, and offers no information about fractional steps or the tail. Any useful method must at least locate $T^*$ and the fractional gaps.
The strongest existing theory is the commutative admissibility theory of Partly Divisible Probability Measures on Locally Compact Abelian Groups. It can be applied to a non-Abelian walk in two legitimate ways. The first is to push the measure forward to the abelianization $G/[G,G]$. The second is to restrict to conjugation-invariant measures, whose Fourier blocks are scalar multiples of the identity, so the character table reduces the problem to a commutative one. We run both on exactly the same measures, grids of $t$, and numerical precision as our solver.
For the spectral side, the competitor is a classical spectral-gap estimate of $T^*$ from the second-largest eigenvalue modulus alone. It ignores multiplicities and non-normality. Beating it means predicting $T^*$ with materially smaller error on non-conjugation-invariant walks. A comparable error would itself be a result, showing non-normality is irrelevant in practice.
Ground truth comes from dense brute force on small groups: the $|G|\times|G|$ transition matrix and its principal power. We use this to validate the solver to machine precision. There is no reported result found on Lacuna for divisibility exponents of measures on any non-Abelian group, so there is no published number to beat. The comparison is against the baselines above, computed by us.
Two ablations isolate the key ingredients. The first replaces the principal branch with independently chosen branches per representation, to show the union-of-semigroups breakdown reappears. The second replaces the pseudospectral norm $\|A^t\|$ with $\rho(A)^t$ in the $T^*$ predictor, to measure how much non-normality matters.
Why not just…?
Why not just abelianize the group? Pushing forward to $G/[G,G]$ only gives an outer bound. If $\mu_t$ is a probability measure on $G$, its image on the abelianization is one too, so $\Lambda(\mu)\subseteq\Lambda(\text{push}(\mu))$. Equality fails generically, and for $S_n$ the abelianization is $\mathbb{Z}_2$, where the commutative theory itself degenerates. The broader literature points the same way. On the Expansion of Group-Based Lifts proves Abelian structure cannot reproduce near-optimal expansion. Differentiable Algebra Discovery Accelerates Grokking via a Non-Fourier Mechanism finds Fourier-style models fail completely on $S_3, D_4, A_4$. We will report the measured gap between the outer bound and the true $\Lambda_p$ for every group in the suite.
Why not just restrict to conjugation-invariant measures, where everything is scalar? This is the serious threat. For such measures our theory reduces to the commutative one on class functions, and nothing new is claimed there. The answer is that the walks people actually run are not conjugation-invariant. Cayley walks on $SL(2,\mathbb{Z}_p)$ with standard generators, as used in Cayley Graph Propagation, are one example; $S_n$ generated by a transposition and an $n$-cycle is another. The experiments test this directly. If the non-normality ablation shows that $\rho^t$ predicts $T^*$ as well as $\|A^t\|$ on these walks, the distinctively non-commutative content is weaker than claimed. We would then reframe the contribution as the branch theory alone.
Why not just compute $P^t$ numerically and check positivity? For small groups this is exactly our ground truth, and we use it. Block diagonalization buys the scaling to groups of order $10^6$ and beyond. More importantly, brute force gives a yes/no answer per $t$, while the edge bounds explain which eigenvalues cause each failure. That explanation is what transfers to families of growing groups.
Why not just use the spectral gap? On normal walks it should essentially suffice for the tail, and we expect to confirm that. It cannot see the small-$t$ obstruction, which lives at the opposite edge. On non-normal walks, any transient growth of block norms would make it underestimate $T^*$.
Proposed Experiments
Calibration comes first. On the circle and on cyclic groups $\mathbb{Z}_n$, we check that the solver recovers the commutative semigroup structure of Partly Divisible Probability Measures on Locally Compact Abelian Groups. For wrapped generalized Gaussian steps, we check that infinite divisibility is lost past $p=2$, as Analytical properties of generalized Gaussian distributions predicts on the line.
The main testbed is the suite of 20 finite groups up to order 2,520 used by Differentiable Algebra Discovery Accelerates Grokking via a Non-Fourier Mechanism. It spans cyclic, dihedral, quaternionic, symmetric and alternating groups, including $A_7$. We extend it with $S_n$ up to $n=10$ and $SL(2,\mathbb{Z}_p)$ for primes up to a few hundred. On each group we sample three classes of measures: conjugation-invariant, lazy Cayley walks, and non-lazy non-symmetric walks. For each measure we compute $\Lambda_p$, $T^*$, and $\inf\Lambda_p$.
The metrics are the relative error of the predicted $T^*$ and the precision and recall of predicted fractional-gap locations against exact computation. We also report the Hausdorff distance between the abelianization outer bound and the true $\Lambda_p$.
There are three decisive checks. The first is whether the pseudospectral predictor beats the eigenvalue predictor on non-conjugation-invariant walks while tying on conjugation-invariant ones; this is the signature of a genuinely non-commutative effect. The second is whether the angular-distance statistic for block eigenvalues near $-1$ predicts which non-integer $t$ fail, better than a bulk spectral summary such as low-order spectral moments. The third is whether the per-representation branch ablation reproduces a $\mathbb{Z}_2$-type union of semigroups.
Finally, in a more exploratory experiment, we draw random generator sets in $SL(2,\mathbb{Z}_p)$ and $S_n$ and study how $T^*$ fluctuates as the group grows. Edge universality is known to be fragile and much harder than bulk universality (RANDOM MATRICES WITH SLOW CORRELATION DECAY), and no local laws exist for group-structured walks. We therefore ask only an empirical question: do rescaled $T^*$ fluctuations look closer to a Tracy–Widom-type law, as found for structured pseudo-random matrices in On the spectral norms of pseudo-wigner and related matrices, or to the Gumbel law that arises when extreme eigenvalues decouple, as in On the rightmost eigenvalue of non-Hermitian random matrices? A clear answer either way would identify which edge theory the group setting should borrow.
References
- Partly Divisible Probability Measures on Locally Compact Abelian Groups
- Partly divisible probability distributions
- Analytical properties of generalized Gaussian distributions
- Determinant Estimation under Memory Constraints and Neural Scaling Laws
- Markovian Linearization of Random Walks on Groups
- On the Expansion of Group-Based Lifts
- Differentiable Algebra Discovery Accelerates Grokking via a Non-Fourier Mechanism
- Cayley Graph Propagation
- RANDOM MATRICES WITH SLOW CORRELATION DECAY
- On the spectral norms of pseudo-wigner and related matrices
- On the rightmost eigenvalue of non-Hermitian random matrices