Lie Group Diffusion: Stochastic Dynamics

• Lie Group Diffusion is a framework that formulates diffusion processes on Lie groups using intrinsic geometry, Lie-algebra increments, and invariant measures.
• It integrates stochastic dynamics, PDE analysis, and generative modeling with applications in lattice gauge theory, manifold learning, and quantum circuit synthesis.
• Advanced methods—such as heat kernels, Casimir operators, and symmetry-induced drift corrections—provide deep insights into non-Euclidean diffusion behavior.

Searching arXiv for the cited work and closely related papers on diffusion on Lie groups. Lie group diffusion denotes a family of constructions in which diffusion processes, diffusion equations, or diffusion-based generative dynamics are formulated using the intrinsic geometry of a Lie group. In the stochastic setting, the state variable itself is a group element, noise is injected through Lie-algebra increments, and the resulting dynamics are expressed through invariant vector fields, heat kernels, Haar measure, Laplace–Beltrami or Casimir operators, and representation theory. Current work spans compact connected Lie groups such as $\mathsf{SO}(2)$, $\mathsf{SO}(3)$, and $SU(2)$; noncompact examples such as $\mathrm{SL}(2)$; and graded or stratified Lie groups equipped with Rockland operators or sub-Laplacians. A parallel literature uses Lie group methods to analyze diffusion equations by symmetry, equivalence transformations, and martingale-problem geometry (Komijani, 17 May 2026, Mahmood et al., 14 Mar 2026, Cardona et al., 2022, Vecchi et al., 2017).

1. Stochastic dynamics on Lie groups

A basic formulation places the forward process directly on a Lie group such as $\mathrm{SU}(N)$, with diffusion time $t \in [0,1]$ and link variables $U_t = U_t(x,\mu)$. In lattice gauge theory, the forward diffusion process is written as

$$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$

where $\mathcal{T}$ is the time-ordering operator, $\sigma(\tau)$ is a scalar noise schedule, and $\mathsf{SO}(3)$0 is a Lie-algebra valued Wiener increment. Applying Itô calculus yields

$\mathsf{SO}(3)$1

so the stochastic evolution carries a deterministic drift induced by the Itô formulation itself (Komijani, 17 May 2026).

An intrinsic, coordinate-free variant on a compact connected Lie group uses the kinematic equation

$\mathsf{SO}(3)$2

and the Stratonovich SDE

$\mathsf{SO}(3)$3

For isotropic Brownian motion this becomes

$\mathsf{SO}(3)$4

The associated density evolution is expressed using the divergence and Laplace–Beltrami operators on the group together with the Haar measure $\mathsf{SO}(3)$5, rather than local coordinates or Euclidean embeddings (Mahmood et al., 14 Mar 2026).

On compact Lie groups, the canonical symmetric diffusion operator is the Laplace–Beltrami operator, also described as the Casimir operator, constructed from the Lie algebra and the Killing form. The Haar measure serves as the invariant reversible measure. In matrix realizations such as $\mathsf{SO}(3)$6, the operator acts explicitly on matrix entries; for example,

$\mathsf{SO}(3)$7

(Bakry, 2014).

A discrete-time diffusion-model formulation on $\mathsf{SO}(3)$8 uses right-multiplicative increments. For each gate variable,

$\mathsf{SO}(3)$9

and after $SU(2)$0 steps with cumulative variance $SU(2)$1, the conditional law is given by the heat kernel

$SU(2)$2

This keeps the process on the manifold at every step (Singh, 28 Jun 2026).

2. Noise schedules, drift structure, and observable evolution

The role of the noise schedule is unusually transparent in lattice-gauge diffusion. For the Wilson gauge action, if

$SU(2)$3

then the paper establishes

$SU(2)$4

The factor $SU(2)$5 comes from the four links in each plaquette. The key point is that, in expectation, only the drift term contributes to the time evolution of observables like the Wilson action because of linearity in each link variable (Komijani, 17 May 2026).

With the schedule

$SU(2)$6

and ignoring $SU(2)$7, one obtains

$SU(2)$8

Choosing

$SU(2)$9

gives

$\mathrm{SL}(2)$0

so the expectation value of the Wilson action decays linearly with diffusion time. For a generic Wilson loop containing $\mathrm{SL}(2)$1 unique links, linear decay is achieved by rescaling the normalization to $\mathrm{SL}(2)$2 (Komijani, 17 May 2026).

This behavior contrasts with standard Euclidean diffusion models, where the process is written as

$\mathrm{SL}(2)$3

There, linear decay of the mean requires an explicitly designed drift term; if $\mathrm{SL}(2)$4, then the mean decays as $\mathrm{SL}(2)$5. In the Lie-group setting discussed above, the corresponding structure emerges from the stochastic evolution itself rather than from a manually engineered drift term (Komijani, 17 May 2026).

A related generative-modeling line extends Euclidean score-based diffusion to arbitrary Lie groups through Generalized Score Matching. In that framework, the learned differential operator is induced by the Lie algebra action, the resulting Langevin dynamics decomposes as a direct sum of Lie algebra representations, standard score matching is recovered when the Lie group is the translation group, and the same formalism extends to flow matching (Bertolini et al., 4 Feb 2025).

3. Heat kernels, reverse processes, and generative modeling

On compact connected Lie groups, diffusion and control problems are naturally organized by the heat semigroup $\mathrm{SL}(2)$6 and its kernel $\mathrm{SL}(2)$7. In the Schrödinger bridge problem, the coordinate-free formulation leads, through optimality analysis and the Hopf–Cole transformation, to a pair of coupled heat equations on the group,

$\mathrm{SL}(2)$8

with nonlinear boundary coupling

$\mathrm{SL}(2)$9

The optimal interpolating density is

$\mathrm{SU}(N)$0

and the optimal geometric feedback control is

$\mathrm{SU}(N)$1

Existence and uniqueness are established using Hilbert’s projective metric and Banach’s Fixed Point Theorem, and the paper gives numerical examples on $\mathrm{SU}(N)$2 and $\mathrm{SU}(N)$3 (Mahmood et al., 14 Mar 2026).

In hardware-aware quantum circuit synthesis, the reverse process is written directly on $\mathrm{SU}(N)$4. The denoising target is the scaled score

$\mathrm{SU}(N)$5

and for relative Lie-algebra coordinate $\mathrm{SU}(N)$6 with $\mathrm{SU}(N)$7, the manifold score is

$\mathrm{SU}(N)$8

The reverse update for slot $\mathrm{SU}(N)$9 is

$t \in [0,1]$0

with $t \in [0,1]$1. The heat kernel replaces the Gaussian, and the process remains on the manifold $t \in [0,1]$2 throughout (Singh, 28 Jun 2026).

The same paper organizes quantum circuit synthesis as a hybrid problem: a circuit skeleton selector chooses an entangling circuit, while the diffusion model generates local single-qubit gates on the manifold itself. The target unitary is

$t \in [0,1]$3

fidelity is measured by

$t \in [0,1]$4

and a hardware-aware score is

$t \in [0,1]$5

The reported aggregate success rates for three-qubit Hamiltonian simulation are approximately $t \in [0,1]$6 for Diffusion+Refinement, versus approximately $t \in [0,1]$7 for a structured baseline and approximately $t \in [0,1]$8 for Haar (Singh, 28 Jun 2026).

4. PDE theory on compact, graded, and stratified Lie groups

A substantial analytic literature studies diffusion equations whose spatial operator is intrinsic to the Lie group. On compact Lie groups, drift-diffusion equations with fractional diffusion are analyzed using global pseudo-differential calculus, subelliptic Hörmander classes, and Sobolev spaces adapted to the sub-Laplacian $t \in [0,1]$9,

$U_t = U_t(x,\mu)$0

For strongly subelliptic $U_t = U_t(x,\mu)$1, the Cauchy problem

$U_t = U_t(x,\mu)$2

has a unique solution

$U_t = U_t(x,\mu)$3

together with the energy estimate

$U_t = U_t(x,\mu)$4

The theory includes quasi-geostrophic-type equations and explicit $U_t = U_t(x,\mu)$5 examples (Cardona et al., 2022).

On graded Lie groups, integro-differential diffusion with memory is written as

$U_t = U_t(x,\mu)$6

where $U_t = U_t(x,\mu)$7 is a positive Rockland operator and $U_t = U_t(x,\mu)$8 is a time kernel. Using the group Fourier transform, the solution is represented as

$U_t = U_t(x,\mu)$9

and the paper proves the $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$0-$$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$1 decay estimate

$$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$2

with $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$3 the homogeneous dimension and $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$4 the Rockland order (Restrepo et al., 2024).

A general space-time fractional diffusion problem on graded Lie groups takes the form

$$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$5

where $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$6 is a general Caputo-type time-fractional derivative and $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$7 is a positive homogeneous Rockland operator. The paper establishes global well-posedness for both homogeneous and inhomogeneous problems in the associated $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$8-Sobolev spaces, together with regularity estimates for $$U_t = K_{t,0} U_0,\qquad K_{t,t'} = \mathcal{T}\exp\!\left(\int_{t'}^t \sigma(\tau)\, dW_\tau\right),$$9 and $\mathcal{T}$0 (Dasgupta et al., 2024).

On stratified Lie groups, the transport-diffusion equation

$\mathcal{T}$1

is studied with $\mathcal{T}$2 the square root of the sub-Laplacian. The paper proves existence of weak solutions, a maximum principle, a positivity principle, and Hölder regularity for positive times by combining Hardy-space duality with a molecular method (Chamorro, 2013).

5. Hypocoercivity, effective diffusion, and representation-theoretic viewpoints

Lie group diffusion need not be elliptic in every variable. On the planar motion group $\mathcal{T}$3, a Langevin-type diffusion is driven by degenerate noise acting only in the rotational direction. With left-invariant vector fields

$\mathcal{T}$4

the generator is decomposed as

$\mathcal{T}$5

Although $\mathcal{T}$6 acts only in $\mathcal{T}$7, averaging over the compact rotation subgroup through the projection

$\mathcal{T}$8

produces the effective macroscopic operator

$\mathcal{T}$9

where $\sigma(\tau)$0. This yields exponential convergence to equilibrium in the abstract hypocoercivity framework (Grothaus et al., 12 May 2026).

A noncompact example appears in the natural diffusion on $\sigma(\tau)$1,

$\sigma(\tau)$2

where $\sigma(\tau)$3 is Brownian motion in $\sigma(\tau)$4. The Frobenius norm $\sigma(\tau)$5 satisfies

$\sigma(\tau)$6

The paper connects this process to a critical drift-diffusion problem in $\sigma(\tau)$7, emphasizing strongly non-Gaussian and intermittent behavior and a borderline super-diffusive scaling function

$\sigma(\tau)$8

This suggests that Lie-group structure can encode large-scale transport behavior that is not well described by Gaussian heuristics (Morfe et al., 2024).

Representation theory also enters discrete data analysis. For manifolds invariant under a compact Lie group $\sigma(\tau)$9, a $\mathsf{SO}(3)$00-invariant graph Laplacian is built by integrating over group orbits: $\mathsf{SO}(3)$01 with

$\mathsf{SO}(3)$02

The operator is diagonalized באמצעות irreducible unitary representation matrices of $\mathsf{SO}(3)$03, and the normalized Laplacian satisfies

$\mathsf{SO}(3)$04

The improvement in the variance term grows with $\mathsf{SO}(3)$05, reflecting the quotient geometry $\mathsf{SO}(3)$06 (Hoyos et al., 2023).

6. Lie symmetries, group classification, and terminological ambiguity

A recurrent source of ambiguity is that “Lie group diffusion” can denote either diffusion on Lie groups or the use of Lie-group methods to analyze diffusion equations. In the martingale-problem approach, a diffusion with generator

$\mathsf{SO}(3)$07

is reformulated geometrically through codiffusors and the annihilator module

$\mathsf{SO}(3)$08

A projectable diffeomorphism is a symmetry precisely when it preserves $\mathsf{SO}(3)$09, and infinitesimal symmetries satisfy

$\mathsf{SO}(3)$10

This is a theory of symmetries of diffusion processes rather than a diffusion whose state space is itself a Lie group (Vecchi et al., 2017).

The same distinction appears in PDE classification. For boundary crossing problems, if the Fokker–Planck–Kolmogorov equation has a nontrivial Lie symmetry, then a boundary crossing identity exists. For time-homogeneous diffusions, the necessary and sufficient conditions are that the drift satisfy one of four Riccati families, including

$\mathsf{SO}(3)$11

and analogous inverse-square cases. These symmetries reduce first passage time problems to Brownian-motion or Bessel-process cases (Muravey, 2018).

For nonlinear diffusion-reaction equations with gradient-dependent diffusion,

$\mathsf{SO}(3)$12

enhanced group classification is carried out via the two-step version of the method of furcate splitting, and the paper constructs a nontrivial example of a finite-dimensional effective generalized equivalence group (Opanasenko et al., 2018). A fractional counterpart based on the modified Riemann–Liouville derivative applies Lie group methods to the space-time fractional diffusion equation

$\mathsf{SO}(3)$13

using fractional characteristic equations and fractional prolongation (Wu, 2010).

This suggests a useful editorial distinction. In one branch, the diffusion process lives on $\mathsf{SO}(3)$14, $\mathsf{SO}(3)$15, $\mathsf{SO}(3)$16, $\mathsf{SO}(3)$17, compact Lie groups, graded Lie groups, or stratified Lie groups. In the other, Lie theory supplies symmetry, reduction, and classification tools for diffusion equations posed on more general spaces. Both branches are active, but they address different mathematical objects and different questions.

7. Scope, applications, and current directions

The contemporary literature connects Lie group diffusion to lattice gauge theory, stochastic optimal control, hypoelliptic kinetic models, scientific generative modeling, manifold learning, and quantum circuit synthesis. In lattice gauge theory, noise scheduling can make the expectation value of the Wilson action decay linearly in diffusion time without adding explicit drift terms (Komijani, 17 May 2026). In the Schrödinger bridge problem, the heat semigroup on a compact connected Lie group yields a geometric controller that optimally interpolates endpoint densities (Mahmood et al., 14 Mar 2026). In generative modeling, Euclidean Generalized Score Matching on Lie groups provides paired SDEs with Casimir corrections and can model any target distribution on any non-Abelian Lie group while operating in Euclidean space (Bertolini et al., 4 Feb 2025). In quantum computing, diffusion on $\mathsf{SO}(3)$18 is combined with a circuit skeleton selector to synthesize hardware-compatible circuits and to explore a fidelity-complexity frontier (Singh, 28 Jun 2026).

The analytic direction is equally broad. Compact Lie groups support drift-diffusion equations with fractional diffusion and pseudo-differential symbol calculus (Cardona et al., 2022); graded Lie groups support nonlocal-in-time diffusion driven by positive Rockland operators (Restrepo et al., 2024); general space-time fractional diffusion equations admit global well-posedness in Rockland-Sobolev spaces (Dasgupta et al., 2024); and stratified Lie groups support transport-diffusion equations with a square root of the sub-Laplacian, maximum principles, positivity, and Hölder regularity (Chamorro, 2013). On the geometric side, $\mathsf{SO}(3)$19 furnishes an intrinsic model of hypocoercive Langevin dynamics in which macroscopic diffusion on $\mathsf{SO}(3)$20 emerges by averaging over the compact rotation subgroup (Grothaus et al., 12 May 2026), while $\mathsf{SO}(3)$21 provides a natural diffusion whose non-Abelian structure is tied to intermittency and borderline super-diffusive behavior (Morfe et al., 2024).

A plausible implication is that Lie group diffusion has become less a single model class than a unifying geometric principle: stochastic evolution, PDE analysis, sampling, control, and representation-theoretic computation are all reformulated so that the group structure is not an afterthought but part of the operator itself.