betaFeedback

WorksMathematics

A Liouville theorem for Lévy operators on nilpotent and compact Lie groups

Abstract

Let L\mathcal{L} be a left-invariant Lévy operator, with characteristics (a,b,μ)(a,b,μ), on a connected and simply connected Lie group GG that is either compact or nilpotent with [G,[G,G]][G,[G,G]] relatively compact (equivalently, of step at most two). We prove that every bounded distributional solution of Lu=0\mathcal{L}u=0 is right-invariant under the closed subgroup P=⟨exp⁡h0∪suppμ⟩‾P=\overline{\langle\exp\mathfrak h_0\cup\text{supp}μ\rangle}, where h0=Lie⟨b,ran⁡(a)⟩\mathfrak h_0=\text{Lie}\langle b,\operatorname{ran}(a)\rangle, and is constant if and only if P=GP=G. In the compact case we give three proofs, one through the Itô–Kawada equidistribution theorem, one by propagation of the maximum and one by an application of the Peter-Weyl theorem, while the nilpotent case rests on the Choquet–Deny theorem of Bénard and Frisch.

Introduction

The classical theorem of Liouville states that a bounded harmonic function on Rd\mathbb{R}^d, that is a bounded function uu with Δu=0\Delta u=0, must be constant. Over the last decade this principle has been revisited in the setting of nonlocal operators. In the case of the fractional Laplacian (−Δ)s(-\Delta)^s, s∈(0,1)s\in(0,1), the Liouville property holds classically as a consequence of the Harnack inequality. Fall [16] proved that every entire ss-harmonic function is affine, and is constant precisely when s∈(0,12]s\in(0,\frac{1}{2}]. Moreover, he showed that if (−Δ)su(-\Delta)^s u equals a polynomial then that polynomial must vanish and uu is again affine.

For the more general setting of Lévy operators, which can be expressed as

Lu(x)=tr⁡(aD2u(x))+b⋅∇u(x)+∫Rd∖{0}(u(x+z)−u(x)−z⋅∇u(x)1∣z∣≤1)μ(dz),(1)\mathcal{L}u(x)=\operatorname{tr}(aD^2u(x))+b\cdot\nabla u(x)+\int_{\mathbb{R}^d\setminus\{0\}}\left(u(x+z)-u(x)-z\cdot\nabla u(x)1_{\lvert z\rvert\leq1}\right)\mu(\mathrm{d}z), \tag*{(1)}

with Lévy triplet (a,b,μ)(a,b,\mu), and the question becomes much more subtle. Concretely, one can ask for which triplets is every bounded solution of Lu=0\mathcal{L}u=0 constant?

This question was first answered by Alibaud, del Teso, Endal and Jakobsen [AdTEJ20a]. More specifically, they characterized the Liouville property for Lévy operators by using a blend of the maximum principle and propagation of a maximum across the support of the jump measure. This allowed them to obtain a sharp condition on the triplet (a,b,μ)(a,b,\mu). They showed that bounded solutions are periodic with respect to an explicit closed subgroup of Rd\mathbb{R}^d that can be characterized by the triplet (a,b,μ)(a,b,\mu). In particular, when the group is all of Rd\mathbb{R}^d they show that solutions are necessarily constant.

Shortly afterwards, Knopova and Schilling [25], Kühn [Kü21] and then Berger and Schilling [10] characterized the same property (through different assumptions on the solution) through Fourier analysis and distribution theory, recasting it in terms of the characteristic exponent (symbol) ψ\psi of the operator. Via a division-type argument, they show that the Fourier transform of a L\mathcal{L}-harmonic function is a distribution supported on the zero set {ψ=0}\{\psi=0\}, so when that set reduces to {0}\{0\} the solution must be constant, whereas any further zero ξ0≠0\xi_0\ne0 yields the bounded non-constant solution x↦cos⁡(ξ0⋅x)x\mapsto\cos(\xi_0\cdot x). Berger, Schilling and Shargorodsky [11] later re-proved the criterion for a general class of Fourier multipliers by a purely analytic argument, linking it to the coupling property of the associated semigroup. More recently, [12] extends the characterization from bounded to sub-exponentially growing solutions. A related division-type classification, valid without any boundedness or decay hypothesis on the solution, is obtained in [27] for symbols vanishing to arbitrary finite order on their zero set, and the same method yields a Liouville theorem for the logarithmic Laplacian [26].

Grzywny and Kwaśnicki [18] go further, handling positive and polynomially bounded L\mathcal{L}-harmonic functions under the same non-degeneracy {ψ=0}={0}\{\psi= 0\} = \{0\}. Their main tool is a representation theorem of Deny type. We call an exponential eλ(x)=eλ⋅xe_\lambda(x) = e^{\lambda\cdot x} L\mathcal{L}-harmonic when Leλ=0\mathcal{L}e_\lambda= 0, and write Λ\Lambda for the set of λ\lambda for which this holds. They show that every nonnegative solution is built out of these, in that any h≥0h \ge0 with Lh=0\mathcal{L}h = 0 can be written as h(x)=∫Λeλ⋅x σ(dλ)h(x) = \int_\Lambda e^{\lambda\cdot x}\,\sigma(d\lambda) for a unique nonnegative measure σ\sigma on Λ\Lambda. Since the only bounded exponentials are constant, the Liouville property drops out as a special case. For polynomial growth they show, assuming ψ\psi is smooth away from the origin (or suitable moment bounds on the Lévy measure), that L\mathcal{L}-harmonic functions of polynomial growth are polynomials with a particular degree.

Nevertheless, all of the work mentioned above takes place on Rd\mathbb{R}^d, where the Euclidean Fourier transform diagonalises the operator and translation commutes with differentiation.

This article aims to prove a Liouville theorem in the case where one works on a compact or nilpotent Lie group (under appropriate assumptions). We also compare 3 methods which have been used in the Euclidean case in this setting.

We outline each approach in reference to the existing approaches within the Euclidean setting.

  • A propagation-of-maximum proof. Alibaud, del Teso, Endal and Jakobsen [AdTEJ20a] proved the Euclidean Liouville property via the positive maximum principle, by moving a maximum across the jump support and along the diffusion and drift. The same argument runs on a compact group.

  • A Choquet–Deny proof. Both Grzywny and Kwaśnicki [18] and Berger and Schilling [10] reduce the Liouville property for a Lévy generator to the convolution equation involving the transition probability measure associated with the Lévy operator. It is from here that Berger and Schilling are able to prove their Liouville theorem via Fourier analytic methods. We apply a similar proof, in the compact setting, but we deviate from their approach by using the resolvent measure and the Itô–Kawada theorem.

  • A harmonic analysis proof. Over Rd\mathbb{R}^d, Knopova–Schilling and Berger–Schilling [25, 10] derive the Liouville property off the characteristic exponent, Berger, Schilling and Shar-gorodsky [11] handle general Fourier multipliers. On a compact group, Peter–Weyl takes the place of the Fourier transform. Expanding uu over the unitary dual, L\mathcal{L} acts on each isotypic component as an operator-valued symbol ψ(π)\psi(\pi), and ψ(π)\psi(\pi) is invertible for every non-trivial π\pi, so only the trivial component survives and uu is constant. The scalar condition {ψ=0}={0}\{\psi= 0\} = \{0\} becomes the invertibility of ψ(π)\psi(\pi) for all non-trivial π\pi.

The nilpotent case requires more consideration. In this setting we can still establish a Liouville theorem (under additional conditions on the group) via a Choquet–Deny argument.

The Choquet–Deny property has a long history. A probability measure ν\nu on a locally compact group HH is adapted if its support generates a dense subgroup, and HH is a Choquet–Deny group if for every adapted ν\nu the only bounded ν\nu-harmonic functions are constant. Choquet and Deny [13] established this property for all abelian locally compact groups, and Dynkin and Malyutov [15] proved it for finitely generated nilpotent groups. For connected Lie groups, Azencott [Aze70] and Guivarc’h [19] identified those that are Choquet–Deny for spread-out measures as the groups of polynomial growth, and Guivarc’h recorded the question, due to Furstenberg, of whether every nilpotent Lie group is Choquet–Deny. In the discrete setting the countable Choquet–Deny groups were characterized by Frisch, Hartman, Tamuz and Vahidi Ferdowsi [17] through the absence of quotients with the infinite conjugacy class property. For general measures on nilpotent Lie groups the question proved delicate, an earlier claim of Raugi [32] having been withdrawn, and it was settled only recently by Bénard and Frisch [8], who proved that a connected nilpotent Lie group is Choquet–Deny if and only if its iterated commutator subgroup [G,[G,G]][G,[G,G]] is relatively compact. It is this result, stated as Theorem 3.1, that we use to treat the nilpotent case.

Before we present our main results, we introduce some notations and conventions.

Notation and Conventions

From this point onwards, we will use the following notations and conventions. We denote GG to be a connected, simply connected, unimodular Lie group (of dimension dd) with identity ee, Lie algebra Lie⁡(G)\operatorname{Lie}(G) and a fixed basis of left-invariant vector fields X1,…,XdX_1,\ldots,X_d. Moreover, we also denote exp⁡:Lie⁡(G)→G\exp:\operatorname{Lie}(G)\to G to be the exponential map. With this choice, we fix smooth canonical coordinates x1,…,xdx_1,\ldots,x_d such that

  • xi(e)=0x_i(e)=0, for i=1,…,di=1,\ldots,d,

  • Xixj(e)=δijX_i x_j(e)=\delta_{ij}, i,j=1,…,di,j=1,\ldots,d where δij\delta_{ij} is the Kronecker delta,

  • xi(τ−1)=−xi(τ)x_i(\tau^{-1})=-x_i(\tau) for all τ∈G\tau\in G.

We denote by dτd\tau the Haar measure on GG. Due to the unimodular assumption, the Haar measure has the following properties.

  • Bi-invariance

∫Gf(τ) dτ=∫Gf(g⋆τ) dτ=∫Gf(τ⋆g) dτ,for g∈G.\int_G f(\tau)\,d\tau=\int_G f(g\star\tau)\,d\tau=\int_G f(\tau\star g)\,d\tau,\quad\text{for }g\in G.
  • Inversion-invariance

∫Gf(τ) dτ=∫Gf(τ−1) dτ.\int_G f(\tau)\,d\tau=\int_G f(\tau^{-1})\,d\tau.

We fix a relatively compact open neighborhood B0B_0 containing ee on which the exponential map exp⁡:Lie⁡(G)→G\exp:\operatorname{Lie}(G)\to G is a diffeomorphism onto its image and let y=log⁡:B0→Lie⁡(G)y=\log:B_0\to\operatorname{Lie}(G) be the canonical coordinates of the first kind. From this, we can fix a family of smooth bump functions of compact support x1,…,xdx_1,\ldots,x_d such that

xi=yi,on a neighborhood of e,supp⁡(xi)⊂B0.x_i=y_i,\quad\text{on a neighborhood of }e,\quad\operatorname{supp}(x_i)\subset B_0.

We denote Lhf=f(h⋆⋅)L_hf=f(h\star\cdot) and Rhf=f(⋅⋆h)R_hf=f(\cdot\star h) to be the left and right translations (respectively). We use two convolutions. For functions ψ,f\psi,f on GG we write the left convolution

(ψ∗f)(g):=∫Gψ(τ)f(τ−1⋆g) dτ,g∈G,(\psi\ast f)(g):=\int_G\psi(\tau)f(\tau^{-1}\star g)\,d\tau,\quad g\in G,

whenever the integral is well-defined. This is the convolution we use to mollify functions. For a bounded measurable function ff and a finite Borel measure ν\nu on GG we write the convolution of ff with ν\nu

(f∗ν)(g):=∫Gf(g⋆h) ν(dh),g∈G,(f\ast\nu)(g):=\int_G f(g\star h)\,\nu(dh),\quad g\in G,

which is the convolution appearing in the transition semigroup. Finally, for finite Borel measures μ,ν\mu,\nu on GG their convolution μ∗ν\mu\ast\nu is the push-forward of μ⊗ν\mu\otimes\nu under the group product,

∫Gf d(μ∗ν)=∫G∫Gf(h⋆h′) μ(dh) ν(dh′),\int_G f\,d(\mu\ast\nu)=\int_G\int_G f(h\star h')\,\mu(dh)\,\nu(dh'),

so that (f∗ν)∗μ=f∗(μ∗ν)(f\ast\nu)\ast\mu=f\ast(\mu\ast\nu) for every bounded measurable ff. We also adopt the following conventions for function spaces.

  • C(G)C(G) continuous functions on GG,

  • Cb(G)C_b(G) continuous and bounded functions on GG,

  • Cc∞(G)=D(G)C_c^\infty(G)=\mathcal{D}(G) smooth functions with compact support on GG (test functions),

  • D′(G)\mathcal{D}'(G) space of distributions on GG, the dual of D(G)\mathcal{D}(G).

We now introduce the operators of interest in this note.

Definition 1.1. A left-invariant Lévy operator on GG is a linear map L:Cc∞(G)→C(G)\mathcal{L}: C^\infty_c(G) \to C(G) of the form

Lf(g)=∑i,j=1daijXiXjf(g)+∑i=1dbiXif(g)+∫G(f(g⋆τ)−f(g)−∑i=1dxi(τ)Xif(g))μ(dτ),(2)\mathcal{L}f(g)=\sum_{i,j=1}^{d}a_{ij}X_iX_jf(g)+\sum_{i=1}^{d}b_iX_if(g)+\int_G\left(f(g\star\tau)-f(g)-\sum_{i=1}^{d}x_i(\tau)X_if(g)\right)\mu(d\tau), \tag*{(2)}

for g∈Gg\in G, where

  • a=(aij)a=(a_{ij}) is a symmetric, positive semidefinite matrix.

  • b=(bi)∈Rdb=(b_i)\in\mathbb{R}^d.

  • μ\mu is a Lévy measure. More specifically, μ\mu is a nonnegative Borel measure on G∖{e}G\setminus\{e\} such that

∫V∑i=1dxi(τ)2μ(dτ)<∞,andμ(Vc)<∞,\int_V\sum_{i=1}^{d}x_i(\tau)^2\mu(d\tau)<\infty,\qquad\text{and}\qquad\mu(V^c)<\infty,

for every neighborhood VV which contains ee.

We call (a,b,μ)(a,b,\mu) the characteristics of L\mathcal{L}.

We remark that left-invariant Lévy operators are precisely the left-invariant operators (with a zero killing term) which satisfy the positive maximum principle via the Courrège theorem due to Applebaum and Ngan [5, 4]. Moreover, via Hunt’s theorem on Lie groups [21] (see also [29 §3.1]), we know that there exists a unique convolution semigroup (ρt)t≥0(\rho_t)_{t\geq0} of probability measures on GG which satisfies

  • ρ0=δe\rho_0=\delta_e,

  • ρs+t=ρs∗ρt\rho_{s+t}=\rho_s\ast\rho_t, for all s,t≥0s,t\geq0,

  • ρt→δe\rho_t\to\delta_e weakly as t→0t\to0.

The corresponding Feller semigroup (on C0(G)C_0(G)) is given by Ttf=f∗ρtT_tf=f\ast\rho_t and its generator (restricted to the core Cc∞(G)C^\infty_c(G)) is given by

ddtTtf∣t=0=Lf,for f∈Cc∞(G).\left.\frac{d}{dt}T_tf\right|_{t=0}=\mathcal{L}f,\qquad\text{for }f\in C^\infty_c(G).

We will write (Xt)t≥0(X_t)_{t\geq0} to be the corresponding Lévy process in GG such that

  • X0=g∈GX_0=g\in G,

  • P(Xt∈A)=ρt(A)\mathbb{P}(X_t\in A)=\rho_t(A), for Borel sets AA,

  • E(f(Xt)∣X0=g)=(f∗ρt)(g)\mathbb{E}(f(X_t)\mid X_0=g)=(f\ast\rho_t)(g) for g∈Gg\in G and f∈Cb(G)f\in C_b(G).

We use the resolvent measure (which is a probability measure) given by

ρ(A):=∫0∞e−tρt(A) dt,for Borel measurable sets A⊂G.\rho(A):=\int_0^\infty e^{-t}\rho_t(A)\,dt,\qquad\text{for Borel measurable sets }A\subset G.

Main result

For the main result, we wish to characterize bounded solutions of the following problem

Lu=0,on G,(3)\mathcal{L}u=0,\qquad\text{on }G, \tag*{(3)}

where L\mathcal{L} is a left-invariant Lévy operator with triplet (a,b,μ)(a,b,\mu). Before we state our result we introduce the appropriate notions of solution corresponding to (3).

Definition 1.2. Let u∈L∞(G)u\in L^\infty(G) and L\mathcal{L} be a left-invariant Lévy operator. We define the distribution Lu∈D′(G)\mathcal{L}u\in\mathcal{D}'(G) by

⟨Lu,φ⟩D′(G),D(G):=∫Gu(L†φ) dτ,for all φ∈D(G),\langle\mathcal{L}u,\varphi\rangle_{\mathcal{D}'(G),\mathcal{D}(G)}:=\int_Gu\left(\mathcal{L}^{\dagger}\varphi\right)\,d\tau,\qquad\text{for all }\varphi\in\mathcal{D}(G),

where L†\mathcal{L}^{\dagger} is the Haar-adjoint of L\mathcal{L}. We say that u∈L∞(G)u\in L^\infty(G) is a bounded distributional solution of (3) if we have that

⟨Lu,φ⟩D′(G),D(G)=0,for all φ∈D(G).\langle\mathcal{L}u,\varphi\rangle_{\mathcal{D}'(G),\mathcal{D}(G)}=0,\qquad\text{for all }\varphi\in\mathcal{D}(G).

Remark 1.3. We show that the notion of bounded distributional solution is indeed valid. It can be verified by direct computation that the Haar adjoint of L\mathcal{L} is again a left-invariant Lévy operator. Namely, if the Lévy triplet corresponding to L\mathcal{L} is given by (a,b,μ)(a,b,\mu) then the Lévy triplet corresponding to L†\mathcal{L}^{\dagger} is given by (a,−b,μ^)(a,-b,\hat{\mu}) where μ^(A)=μ(A−1)\hat{\mu}(A)=\mu(A^{-1}). Here A−1A^{-1} consists of the group inverses of AA. Note that this critically requires the unimodular assumption of GG. This gives us that the adjoint is given by

L†f(g)=∑i,j=1daijXiXjf(g)−∑i=1dbiXif(g)+∫G(f(g⋆τ−1)−f(g)+∑i=1dxi(τ)Xif(g))μ(dτ),\mathcal{L}^{\dagger}f(g)=\sum_{i,j=1}^{d}a_{ij}X_iX_jf(g)-\sum_{i=1}^{d}b_iX_if(g)+\int_G\left(f(g\star\tau^{-1})-f(g)+\sum_{i=1}^{d}x_i(\tau)X_if(g)\right)\mu(d\tau),

for g∈Gg\in G. Hence to show that the notion of solution is appropriate it is sufficient to demonstrate that Lψ∈L1(G)\mathcal{L}\psi\in L^1(G) for any Lévy operator L\mathcal{L} and ψ∈D(G)\psi\in\mathcal{D}(G). From an application of the second order Taylor expansion

∣ψ(g⋆h)−ψ(g)−∑ixi(h)Xiψ(g)∣≤12max⁡i,j∥XiXjψ∥∞∑ixi(h)2,for g,h∈G,\left|\psi(g\star h)-\psi(g)-\sum_i x_i(h)X_i\psi(g)\right|\leq\frac{1}{2}\max_{i,j}\|X_iX_j\psi\|_{\infty}\sum_i x_i(h)^2,\qquad\text{for }g,h\in G,

we have that Lψ∈L1(G)\mathcal{L}\psi\in L^1(G) for any ψ∈D(G)\psi\in\mathcal{D}(G).

Before we state the main results of this article we introduce the following sets

P=⟨exp⁡h0∪supp⁡μ⟩‾,h0=Lie⁡⟨b,ran⁡(a)⟩.P=\overline{\left\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\right\rangle},\qquad\mathfrak{h}_0=\operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle.

Theorem 1.4. Let GG be a connected, simply connected Lie group which is either

  • compact,

  • nilpotent with the property that [G,[G,G]][G,[G,G]] is relatively compact; as GG is simply connected, this is equivalent to GG being nilpotent of step at most two, that is, abelian or 2-step (see Remark 1.5).

Let uu be a bounded distributional solution of (3), see Definition 1.2, for a Lévy operator with triplet (a,b,μ)(a,b,\mu). Then we have that uu is right-invariant, namely

u(g⋆p)=u(g),for almost every g∈G and p∈P.u(g\star p)=u(g),\qquad\text{for almost every }g\in G\text{ and }p\in P.

Moreover, uu is constant if P=GP=G, while if P≠GP\ne G there exist nonconstant bounded distributional solutions.

The above result is essentially a Lie group analogue of the result of Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b], however, the proof we use does not rely on their propagation-of-maximum approach. Instead, we demonstrate that it is an immediate corollary of a Choquet–Deny type theorem, due to a recent result of Bénard and Frisch [8], see Theorem 3.1. We remark that one can establish when a solution is necessarily constant. However, when P≠GP\ne G, showing that uu is right-invariant with respect to PP does not quite work when the Lie group is non-commutative. We present this in more detail in § 4.

Remark 1.5. The nilpotent hypothesis of Theorem 1.4 is more restrictive than it first appears. When GG is simply connected and nilpotent, each term of its lower central series is a closed, simply connected subgroup, hence isomorphic to some Rk\mathbb{R}^k, so [G,[G,G]][G,[G,G]] is relatively compact only when it is trivial. The simply connected nilpotent groups covered by Theorem 1.4 are therefore exactly those of step at most two. These include the abelian groups Rd\mathbb{R}^d, the Heisenberg groups Hn\mathbb{H}^n, and more generally all Heisenberg-type (Kaplan) groups and free two-step nilpotent groups, together with direct products such as Rk×Hn\mathbb{R}^k\times\mathbb{H}^n. A simply connected nilpotent group of step three or higher, such as the group of 4×44\times4 upper unitriangular matrices, has non-compact iterated commutator [G,[G,G]][G,[G,G]] and is not covered.

Examples

We first check that Theorem 1.4 recovers the known Euclidean theorem, and then illustrate the genuinely non-commutative behavior on SU(2)SU(2).

Example 1.6. Take G=RdG = \mathbb{R}^{d} with its additive group structure. It is abelian, hence nilpotent of step one, connected and simply connected, and [G,[G,G]]={0}[G,[G,G]] = \{0\} is trivially relatively compact, so Rd\mathbb{R}^{d} falls under the nilpotent case of Theorem 1.4. The left-invariant fields are the commuting partial derivatives Xi=∂xiX_i = \partial_{x_i}, the exponential map is the identity, and a left-invariant Lévy operator is the classical Lévy operator (1).

Because the group is abelian, h0=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0 = \operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle is simply the linear subspace V=span⁡{b}+ran⁡(a)V = \operatorname{span}\{b\} + \operatorname{ran}(a), and exp⁡h0=V\exp\mathfrak{h}_0 = V. Hence

P=⟨V∪supp⁡μ⟩‾=V+⟨supp⁡μ⟩‾,P = \overline{\langle V \cup\operatorname{supp}\mu\rangle} = \overline{V + \langle\operatorname{supp}\mu\rangle},

the closed subgroup of Rd\mathbb{R}^{d} generated by the subspace VV (spanned by the drift and the range of the diffusion) together with the jump vectors. Theorem 1.4 then states that every bounded solution of Lu=0\mathcal{L}u = 0 is PP-periodic, that is u(x+p)=u(x)u(x+p) = u(x) for all p∈Pp \in P, and is constant if and only if P=RdP = \mathbb{R}^{d}. This is precisely the characterization of Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b, AdTEJ20a], with PP their subgroup of periods.

The two regimes appear in the following concrete choices.

  • Fractional Laplacian. For L=−(−Δ)s\mathcal{L} = -(-\Delta)^{s} with s∈(0,1)s \in(0,1) one has a=0a = 0, b=0b = 0 and μ(dz)=cd,s∣z∣−d−2s dz\mu(dz) = c_{d,s}|z|^{-d-2s}\,dz, whose support is all of Rd∖{0}\mathbb{R}^{d} \setminus\{0\}. Thus ⟨supp⁡μ⟩\langle\operatorname{supp}\mu\rangle is dense, P=RdP = \mathbb{R}^{d}, and every bounded solution is constant. This is the classical Liouville theorem for the fractional Laplacian.

  • A proper jump lattice. On R\mathbb{R} take the pure-jump operator Lu(x)=u(x+1)+u(x−1)−2u(x)\mathcal{L}u(x) = u(x+1) + u(x-1) - 2u(x), that is a=0a = 0, b=0b = 0 and μ=δ1+δ−1\mu= \delta_{1} + \delta_{-1}. Here V={0}V = \{0\} and ⟨supp⁡μ⟩=Z\langle\operatorname{supp}\mu\rangle= \mathbb{Z}, so P=Z≠RP = \mathbb{Z} \ne\mathbb{R}. A bounded uu solves Lu=0\mathcal{L}u = 0 exactly when it is 1-periodic, for instance u(x)=cos⁡(2πx)u(x) = \cos(2\pi x), so nonconstant bounded solutions exist, in agreement with P≠RP \ne\mathbb{R}.

We illustrate Theorem 1.4 on the simplest genuinely non-commutative compact group, G=SU(2)G = SU(2), which is compact, connected and simply connected, hence covered by the compact case. Fix a basis X1,X2,X3X_1,X_2,X_3 of Lie⁡(SU(2))\operatorname{Lie}(SU(2)) with

[X1,X2]=X3,[X2,X3]=X1,[X3,X1]=X2,[X_1,X_2] = X_3,\qquad[X_2,X_3] = X_1,\qquad[X_3,X_1] = X_2,

and let T={exp⁡(θX3):θ∈R}T = \{\exp(\theta X_3) : \theta\in\mathbb{R}\} be the maximal torus generated by X3X_3, so that exp⁡(θX3)=diag⁡(eiθ/2,e−iθ/2)\exp(\theta X_3) = \operatorname{diag}(e^{i\theta/2},e^{-i\theta/2}). The two examples below form a matched pair, built on the same group and the same kind of operator and differing only in how many diffusion directions are switched on. In the first the characteristics generate only TT, so P=T≠GP = T \ne G and the Liouville property fails, while in the second one further direction makes them bracket-generate, so P=GP = G and it holds.

Example 1.7 (A compact, non-commutative operator that is not Liouville). Take the single degenerate diffusion L=X32\mathcal{L} = X_3^2, that is ran⁡(a)=RX3\operatorname{ran}(a) = \mathbb{R}X_3, b=0b = 0 and μ=0\mu= 0. The only direction present is X3X_3, and so

h0=Lie⁡⟨ran⁡(a)⟩=RX3,P=⟨exp⁡RX3⟩‾=T≠G.\mathfrak{h}_0 = \operatorname{Lie}\langle\operatorname{ran}(a)\rangle= \mathbb{R}X_3,\qquad P = \overline{\langle\exp\mathbb{R}X_3\rangle} = T \ne G.

Theorem 1.4 says that every bounded solution of X32u=0X_3^2u = 0 is right-invariant under TT, and since P=T≠GP = T \ne G it need not be constant. Indeed it is not. Writing g=(gij)∈SU(2)g = (g_{ij}) \in SU(2), set

u(g)=∣g11∣2.u(g) = |g_{11}|^2.

Right multiplication by exp⁡(θX3)=diag⁡(eiθ/2,e−iθ/2)\exp(\theta X_3) = \operatorname{diag}(e^{i\theta/2},e^{-i\theta/2}) multiplies the first column of gg by the phase eiθ/2e^{i\theta/2} and so leaves ∣g11∣2|g_{11}|^2 unchanged. Hence uu is right-TT-invariant, which gives X3u=0X_3u = 0 and so X32u=0X_3^2u = 0, while uu is plainly nonconstant. In fact a bounded function solves X32u=0X_3^2u = 0 exactly when it is right-TT-invariant, that is, when it descends to the quotient SU(2)/T≅S2SU(2)/T \cong S^2. The bounded solutions are therefore all bounded functions on the 2-sphere, a whole function space rather than the constants alone. This is the non-commutative analogue of a bounded harmonic function that depends on only some of the variables.

Example 1.8 (Switching the Liouville property back on). Keep G=SU(2)G = SU(2) and add one further diffusion direction, L=X12+X32\mathcal{L} = X_1^2 + X_3^2, so ran⁡(a)=RX1⊕RX3\operatorname{ran}(a) = \mathbb{R}X_1 \oplus\mathbb{R}X_3. The bracket [X3,X1]=X2[X_3, X_1] = X_2 recovers the missing direction, so X1X_1 and X3X_3 now bracket-generate the whole Lie algebra and

h0=Lie⁡⟨X1,X3⟩=Lie⁡(SU(2)),P=⟨exp⁡Lie⁡(SU(2))⟩‾=SU(2)=G.\mathfrak{h}_0 = \operatorname{Lie}\langle X_1, X_3\rangle= \operatorname{Lie}(SU(2)), \qquad P = \overline{\langle\exp\operatorname{Lie}(SU(2))\rangle} = SU(2) = G.

This time P=GP = G, and so Theorem 1.4 forces every bounded solution to be constant. Placing the two side by side, the bounded solutions of X32u=0X_3^2u = 0 fill L∞(S2)L^\infty(S^2), whereas adding the one bracket-generating direction X1X_1 collapses them to the constants. The obstruction in the first example is thus exactly the failure of PP to fill out GG.

Outline of the article

The remainder of the article is organized as follows. In Section 2 we reduce the problem to smooth solutions of Lu=0\mathcal{L}u = 0, taken in the pointwise sense, by a mollification argument adapted from the Euclidean work of Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b]. Section 3 proves Theorem 1.4 in both the compact and the nilpotent cases: we recast the equation as the Choquet–Deny convolution equation u=u∗ρu = u * \rho for the resolvent measure ρ\rho, apply the Itô–Kawada theorem in the compact case and the theorem of Bénard and Frisch [8] in the nilpotent case, and identify the subgroup PP in terms of the characteristics (a,b,μ)(a,b,\mu). In Section 4 we revisit the compact case by the propagation-of-maxima method of [AdTEJ20b], which recovers the part of Theorem 1.4 asserting that uu is constant if and only if P=GP = G (Theorem 4.1), and we explain why it does not yield the full right-invariance on a non-commutative or non-compact group. Finally, Section 5 gives a Fourier-analytic proof in the compact case, based on the Peter–Weyl theorem, and records some observations on a Fourier-analytic approach to the nilpotent case.

Reduction to smooth solutions

Before we commence on proving Theorem 1.4 directly, we demonstrate that we can reduce to studying smooth solutions which solve (3) in a pointwise sense. This is similar in spirit to what was considered in the Euclidean case by Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b].

The goal of this section is to prove the following proposition which is the Lie group analogue of [AdTEJ20b, Lemma 2.4]. This characterization will be useful for when we consider the approaches via propagation-via-maximum and Choquet–Deny.

Proposition 2.1. Let u∈L∞(G)u \in L^\infty(G) and let {ψε}ε>0⊂Cc∞(G)\{\psi_\varepsilon\}_{\varepsilon>0} \subset C^\infty_c(G) be an approximate identity, and set uε:=ψε∗uu_\varepsilon:= \psi_\varepsilon* u. Then the following hold.

  1. Lu=0\mathcal{L}u = 0 in D′(G)\mathcal{D}'(G) if and only if Luε=0\mathcal{L}u_\varepsilon= 0 pointwise for every ε>0\varepsilon> 0,

  1. uu is a.e. SS-periodic (i.e. Rsu=uR_su = u for almost every s∈Ss \in S) if and only if Rsuε=uεR_su_\varepsilon= u_\varepsilon for every s∈Ss \in S and every ε>0\varepsilon> 0.

The above proposition effectively allows us to consider solutions that are smooth with bounded derivatives. Moreover, under this assumption and Itô’s formula [29], we have the following identity which can be viewed as an alternative to (3).

Corollary 2.2. Let (ρt)t≥0(\rho_t)_{t \ge0} be the corresponding convolution semigroup associated with L\mathcal{L}. Suppose that u∈Cb∞(G)u \in C^\infty_b(G) and Lu=0\mathcal{L}u = 0. Then, we have that

u=u∗ρt,for all t≥0.(4)u = u * \rho_t, \qquad\text{for all } t \ge0. \tag*{(4)}

Before we prove Proposition 2.1, we remind the reader that GG is unimodular since GG is compact or nilpotent. So the modular function is Δ≡1\Delta\equiv1. We write ψ~(h):=ψ(h−1)\widetilde{\psi}(h) := \psi(h^{-1}) for the reflection and record three consequences of unimodularity that we use repeatedly.

(i) Adjoint of left convolution. For ψ∈C∞(G)\psi\in C^\infty(G), w∈L∞(G)w \in L^\infty(G) and φ∈L1(G)\varphi\in L^1(G),

∫G(ψ∗w) φ dh=∫Gw(ψ~∗φ) dh,(5)\int_G (\psi* w)\,\varphi\,dh = \int_G w(\widetilde{\psi} * \varphi)\,dh, \tag*{(5)}

which is Fubini together with the substitution h↦h−1h \mapsto h^{-1} under the (two-sided) Haar measure, here ψ∨=ψ~\psi^\vee= \widetilde{\psi} because Δ≡1\Delta\equiv1.

(ii) Left-invariance of the adjoint. Since GG is unimodular, left-invariant vector fields are formally skew-adjoint, so the formal adjoint L†\mathcal{L}^\dagger of the left-invariant generator L\mathcal{L} is again left-invariant. In particular L†\mathcal{L}^\dagger commutes with left convolution,

L†(φ∗ψ)=φ∗(L†ψ),φ,ψ∈D(G).\mathcal{L}^\dagger(\varphi* \psi) = \varphi* (\mathcal{L}^\dagger\psi), \qquad\varphi,\psi\in\mathcal{D}(G).

(iii) Left convolution commutes with right translation: Rs(ψ∗w)=ψ∗(Rsw)R_s(\psi* w) = \psi* (R_s w) for s∈Gs \in G.

We can now prove Proposition 2.1.

Proof of Proposition 2.1. First observe that each uϵ=ψϵ∗uu_\epsilon= \psi_\epsilon* u is smooth and bounded, so Luϵ\mathcal{L}u_\epsilon and Rsuϵ−uϵR_su_\epsilon-u_\epsilon are continuous functions.

  1. We consider the forward direction. Assume that

∫GuL†φ dτ=0,for all φ∈D(G).\int_G u\mathcal{L}^\dagger\varphi\,d\tau= 0, \qquad\text{for all }\varphi\in\mathcal{D}(G).

Replacing φ\varphi by ψ~ϵ∗φ∈D(G)\widetilde{\psi}_\epsilon* \varphi\in\mathcal{D}(G), then using the left-invariance of L†\mathcal{L}^\dagger and the transpose identity for left convolution, we obtain

0=∫GuL†(ψ~ϵ∗φ) dτ=∫Gu(ψ~ϵ∗L†φ) dτ=∫G(ψϵ∗u)L†φ dτ=∫G(Luϵ)φ dτ,0 = \int_G u\mathcal{L}^\dagger(\widetilde{\psi}_\epsilon* \varphi)\,d\tau= \int_G u(\widetilde{\psi}_\epsilon* \mathcal{L}^\dagger\varphi)\,d\tau= \int_G(\psi_\epsilon* u)\mathcal{L}^\dagger\varphi\,d\tau= \int_G(\mathcal{L}u_\epsilon)\varphi\,d\tau,

where the last equality is integration by parts against the smooth bounded function uϵ=ψϵ∗uu_\epsilon=\psi_\epsilon*u. Since φ∈D\varphi\in\mathcal{D} was arbitrary and Luϵ\mathcal{L}u_\epsilon is continuous, Luϵ=0\mathcal{L}u_\epsilon=0 pointwise. The other direction follows from the direct consequence of the fact that uϵ→uu_\epsilon\to u weak-* in L∞(G)L^\infty(G) as ϵ→0\epsilon\to0, and L†ψ∈L1(G)\mathcal{L}^\dagger\psi\in L^1(G) for every ψ∈D(G)\psi\in\mathcal{D}(G),

∫GuL†ψ dτ=lim⁡ϵ→0∫GuϵL†ψ dτ=lim⁡ϵ→0⟨Luϵ,ψ⟩=0.\int_G u\mathcal{L}^\dagger\psi\,d\tau= \lim_{\epsilon\to0}\int_G u_\epsilon\mathcal{L}^\dagger\psi\,d\tau= \lim_{\epsilon\to0}\langle\mathcal{L}u_\epsilon,\psi\rangle= 0.
  1. Assume Rsu=uR_su=u a.e. By (5) and property (iii) above,

∫G(Rsuϵ−uϵ)ψ dh=∫G(ψϵ∗(Rsu−u))ψ dh=∫G(Rsu−u)(ψ~ϵ∗ψ) dh=0\int_G(R_su_\epsilon-u_\epsilon)\psi\,dh = \int_G\bigl(\psi_\epsilon*(R_su-u)\bigr)\psi\,dh = \int_G(R_su-u)(\widetilde{\psi}_\epsilon*\psi)\,dh=0

for all ψ∈C∞(G)\psi\in C^\infty(G), since Rsu−u=0R_su-u=0 a.e., continuity of Rsuϵ−uϵR_su_\epsilon-u_\epsilon gives Rsuϵ=uϵR_su_\epsilon=u_\epsilon pointwise. The other direction follows similarly as in the proof of 1.

Proof via a Choquet–Deny formulation

We explain our approach for proving Theorem 1.4 via the Choquet–Deny formulation. Note that this proof allows us to cover the nilpotent case as well as the compact case. We state a result of Bénard and Frisch [8] that is critical to our proof of the nilpotent case.

Theorem 3.1 (Bénard–Frisch [8]). Let GG be a connected nilpotent Lie group, and suppose its iterated commutator subgroup [G,[G,G]][G,[G,G]] is relatively compact. If ν\nu is a probability measure on GG whose support generates a dense subgroup, then every bounded measurable h:G→Rh:G\to\mathbb{R} satisfying

h(x)=∫Gh(x⋆y) ν(dy)for all x∈G,h(x)=\int_G h(x\star y)\,\nu(dy) \qquad\text{for all }x\in G,

is necessarily constant.

Remark 3.2. Bénard and Frisch’s convolution equation is the classical Choquet–Deny equation after reflecting the measure. Starting from

h(x)=∫Gh(x⋆y) ν(dy)for x∈G,h(x) = \int_G h(x \star y)\,\nu(dy) \qquad\text{for } x \in G,

if we take νˇ\check{\nu} to be the pushforward of ν\nu under inversion y↦y−1y \mapsto y^{-1}

νˇ(A)=ν(A−1),for Borel A,\check{\nu}(A) = \nu(A^{-1}), \qquad\text{for Borel } A,

so that

∫Gφ(z) νˇ(dz)=∫Gφ(y−1) ν(dy),\int_G \varphi(z)\,\check{\nu}(dz) = \int_G \varphi(y^{-1})\,\nu(dy),

for every test function φ\varphi. Taking φ(z)=h(x⋆z−1)\varphi(z) = h(x \star z^{-1}) gives

h(x)=∫Gh(x⋆y) ν(dy)=∫Gh(x⋆z−1) νˇ(dz)=(h∗νˇ)(x),h(x) = \int_G h(x \star y)\,\nu(dy) = \int_G h(x \star z^{-1})\,\check{\nu}(dz) = (h \ast\check{\nu})(x),

with the convention (h∗σ)(x):=∫Gh(x⋆z−1) σ(dz)(h \ast\sigma)(x) := \int_G h(x \star z^{-1})\,\sigma(dz). Hence h=h∗νˇh = h \ast\check{\nu}, the original Choquet–Deny equation.

Since

supp⁡νˇ=(supp⁡ν)−1,\operatorname{supp}\check{\nu} = (\operatorname{supp}\nu)^{-1},

generates the same closed subgroup as supp⁡ν\operatorname{supp}\nu, the measure νˇ\check{\nu} is adapted, i.e. ⟨supp⁡νˇ⟩=G\langle\operatorname{supp}\check{\nu}\rangle= G, if and only if ν\nu is. Hence hh is constant when either variant of the Choquet–Deny equation is satisfied.

We now outline our approach to proving Theorem 1.4. We assume that our solution uu is continuous thanks to Proposition 2.1. Note that Corollary 2.2 enables us to consider the following problem

u=u∗ρt,for t≥0,u = u \ast\rho_t, \qquad\text{for } t \ge0,

We now explain our approach to proving Theorem 1.4.

Step 1. We show that as a consequence of (4), we have that u=u∗ρu = u \ast\rho.

We aim to show that u(g⋆p)=u(g)u(g \star p) = u(g) for g∈Gg \in G and p∈⟨supp⁡ρ⟩p \in\langle\operatorname{supp}\rho\rangle. We treat the compact and nilpotent cases separately.

Step 2. We show this under the assumption that GG is compact via the Itô–Kawada theorem, see [23] or [30].

Step 3. We directly apply the main result of Bénard and Frisch [8] to consider the case when GG is a nilpotent Lie group with the property that [G,[G,G]][G,[G,G]] is relatively compact.

We now denote P=⟨supp⁡ρ⟩P = \langle\operatorname{supp}\rho\rangle. The goal is now to show that P=⟨exp⁡h0∪supp⁡μ⟩P = \langle\exp\mathfrak{h}_0 \cup\operatorname{supp}\mu\rangle, h0=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0 = \operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle. By [29 Proposition 6.7], supp⁡ρ=⋃t>0supp⁡ρt\operatorname{supp}\rho= \bigcup_{t>0}\operatorname{supp}\rho_t, so PP is the smallest closed subgroup HH with ρt(H)=1\rho_t(H) = 1 for all tt. We therefore ask which closed subgroups carry the whole semigroup. We fix a closed subgroup HH, choose a basis X1,…,XdX_1,\ldots,X_d of Lie⁡(G)\operatorname{Lie}(G) with X1,…,XmX_1,\ldots,X_m spanning Lie⁡(H)\operatorname{Lie}(H) and Xm+1,…,XdX_{m+1},\ldots,X_d transverse, and write xi(g)x_i(g) for the ii-th coordinate of log⁡g\log g, so that near ee the subgroup HH is {xm+1=⋯=xd=0}\{x_{m+1} = \cdots= x_d = 0\}.

Step 4. We prove that ρt(H)=1\rho_t(H) = 1 for all tt forces

supp⁡μ⊆H,ran⁡(a)⊆Lie⁡(H),b∈Lie⁡(H).(6)\operatorname{supp}\mu\subseteq H,\qquad\operatorname{ran}(a) \subseteq\operatorname{Lie}(H),\qquad b \in\operatorname{Lie}(H). \tag*{(6)}

Step 5. Conversely, we show that if the above 3 conditions hold, the characteristics restrict to a Lévy triplet on HH whose generator coincides with L\mathcal{L} on C∞(G)C^\infty(G). Hunt’s uniqueness [21] then forces ρt(H)=1\rho_t(H) = 1 for all tt.

Step 6. We prove P=⟨exp⁡h0∪supp⁡μ⟩P = \langle\exp\mathfrak{h}_0 \cup\operatorname{supp}\mu\rangle, h0=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0 = \operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle which completes our proof.

Proof of Theorem 1.4 via a Choquet–Deny type approach. We are now ready to complete the proof of Theorem 1.4.

Step 1. Integrating (4) with respect to e−te^{-t}, we obtain

u(g)=∫0∞e−tu(g) dt=∫0∞e−t(u∗ρt)(g) dt=∫0∞e−t(∫Gu(g⋆h) ρt(dh)) dt=∫Gu(g⋆h) ρ(dh)=(u∗ρ)(g),g∈G,\begin{aligned} u(g) &= \int_{0}^{\infty} e^{-t}u(g)\,dt \\ &= \int_{0}^{\infty} e^{-t}(u * \rho_t)(g)\,dt \\ &= \int_{0}^{\infty} e^{-t}\left(\int_G u(g \star h)\,\rho_t(dh)\right)\,dt \\ &= \int_G u(g \star h)\,\rho(dh)=(u * \rho)(g), \qquad g \in G, \end{aligned}

which is our Choquet–Deny equation

u=u∗ρ.(7)u=u * \rho. \tag*{(7)}

Note, for the above steps to be valid, we used the Fubini theorem and the fact that uu is bounded.

Step 2. We begin with the compact setting which uses the Itô–Kawada theorem, see [23] or [30]. We present a proof using a formulation given in [23]. This is standard for readers familiar with Choquet–Deny theory, but we include it for clarity.

Using the fact that u=u∗ρ=⋯=u∗ρ∗ku=u * \rho=\cdots=u * \rho^{*k} for k∈Nk \in\mathbb{N}, we have that

u=u∗AN,AN:=1N∑k=0N−1ρ∗k(N∈N),u=u * A_N,\qquad A_N:=\frac{1}{N}\sum_{k=0}^{N-1}\rho^{*k}\qquad(N\in\mathbb{N}),

where ρ∗k\rho^{*k} is the kk-th convolution power of ρ\rho. By an application of the Itô–Kawada theorem we have that

u=u∗AN=lim⁡N→∞u∗AN=u∗mP,u=u * A_N=\lim_{N\to\infty}u * A_N=u * m_P,

where mPm_P is the Haar measure supported on supp⁡ρ‾\overline{\operatorname{supp}\rho}.

For each gg, expanding the square and using ∫Pu(g⋆p) dmP(p)=u(g)\int_P u(g \star p)\,dm_P(p)=u(g) and ∫PdmP=1\int_P dm_P=1,

∫P∣u(g⋆p)−u(g)∣2 dmP(p)=∫P∣u(g⋆p)∣2 dmP(p)−u(g)2.\int_P |u(g \star p)-u(g)|^2\,dm_P(p)=\int_P |u(g \star p)|^2\,dm_P(p)-u(g)^2.

Integrate over g∈Gg\in G against mGm_G, by Fubini and the right-invariance of mGm_G, ∫G∣u(g⋆p)∣2 dmG(g)=∥u∥22\int_G |u(g \star p)|^2\,dm_G(g)=\|u\|_2^2 for each pp, so

∫G,P∣u(g⋆p)−u(g)∣2 dmP(p) dmG(g)=∥u∥22−∥u∥22=0.\int_{G,P}|u(g \star p)-u(g)|^2\,dm_P(p)\,dm_G(g)=\|u\|_2^2-\|u\|_2^2=0.

The integrand is nonnegative, hence vanishes a.e., and we necessarily have that

u(g⋆p)=u(g)for mG-a.e. g and mP-a.e. p.u(g \star p)=u(g)\qquad\text{for }m_G\text{-a.e. }g\text{ and }m_P\text{-a.e. }p.

Since uu is continuous, we have completed the step for when GG is compact.

Step 3. We now consider the nilpotent case under the assumption that [G,[G,G]][G,[G,G]] is relatively compact.

Fix g∈Gg\in G and set ug(p)=u(g⋆p)u_g(p)=u(g \star p) for p∈Pp\in P. Since ρ\rho is carried by PP, (7) restricts to ug=ug∗ρu_g=u_g * \rho on PP, so ugu_g is a bounded ρ\rho-harmonic function on PP, and supp⁡ρ‾=P\overline{\operatorname{supp}\rho}=P by construction. As a closed subgroup of the nilpotent Lie group GG, the group PP is itself a nilpotent Lie group, and its iterated commutator satisfies [P,[P,P]]⊆[G,[G,G]][P,[P,P]]\subseteq[G,[G,G]] and is therefore relatively compact. Hence PP is a Choquet–Deny group and ugu_g is constant on PP. When PP is connected this is exactly Theorem 3.1. In general the identity component P∘P^\circ is a connected nilpotent Lie group with relatively compact iterated commutator, hence Choquet–Deny by Theorem 3.1, while the component group P/P∘P/P^\circ is finitely generated nilpotent and hence Choquet–Deny by the theorem of Dynkin and Malyutov [15] (see also [17]). Since PP is nilpotent, these two facts combine to give the Choquet–Deny property for PP. Thus u(g⋆p)=ug(e)=u(g)u(g\star p)=u_g(e)=u(g) for all p∈Pp\in P, and as gg was arbitrary, uu is right-invariant under PP.

Step 4. Suppose ρt(H)=1\rho_t(H)=1 for all tt. Any f∈Cc∞(G)f\in C^\infty_c(G) vanishing on HH then has ∫Gf dρt=0\int_G f\,d\rho_t=0 for all tt and f(e)=0f(e)=0, so

Lf(e)=lim⁡t→01t(∫Gf dρt−f(e))=0.\mathcal{L}f(e)=\lim_{t\to0}\frac{1}{t}\left(\int_G f\,d\rho_t-f(e)\right)=0.

We verify that the sets are contained in the support by choosing ff carefully.

(a) supp⁡μ⊆H\operatorname{supp}\mu\subseteq H

Take any f≥0f\geq0 with supp⁡f⊆G∖H\operatorname{supp}f\subseteq G\setminus H. As e∈He\in H, the point ee is outside the closed set supp⁡f\operatorname{supp}f, so f≡0f\equiv0 on a neighborhood of ee. In turn, we have that

f(e)=0,(Xkf)(e)=0,and(XpXqf)(e)=0.f(e)=0,\qquad(X_kf)(e)=0,\qquad\text{and}\qquad(X_pX_qf)(e)=0.

It follows that

Lf(e)=∫Gf(τ) μ(dτ)=0.\mathcal{L}f(e)=\int_G f(\tau)\,\mu(d\tau)=0.

This holds for every nonnegative ff supported in G∖HG\setminus H, which forces μ(G∖H)=0\mu(G\setminus H)=0, i.e. supp⁡μ⊆H\operatorname{supp}\mu\subseteq H.

(b) ran⁡(a)⊆Lie⁡(H)\operatorname{ran}(a)\subseteq\operatorname{Lie}(H)

Fix transverse i,ji,j and take f=xixjχf=x_ix_j\chi. Using the Leibniz rule and xi(e)=xj(e)=0x_i(e)=x_j(e)=0, we have that

f(e)=0,(Xkf)(e)=(Xkxi)(e)xj(e)+xi(e)(Xkxj)(e)=0,f(e)=0,\qquad(X_kf)(e)=(X_kx_i)(e)x_j(e)+x_i(e)(X_kx_j)(e)=0,
(XpXqf)(e)=(Xpxi)(e)(Xqxj)(e)+(Xqxi)(e)(Xpxj)(e)=δpiδqj+δpjδqi.(X_pX_qf)(e)=(X_px_i)(e)(X_qx_j)(e)+(X_qx_i)(e)(X_px_j)(e)=\delta_{pi}\delta_{qj}+\delta_{pj}\delta_{qi}.

Hence, by direct computation, we have that

Lf(e)=∑p,qapq(δpiδqj+δpjδqi)+∫Gf dμ,=(aij+aji)+∫Gf dμ,=2aij+∫Gf dμ,=2aij,\begin{aligned} \mathcal{L}f(e)=\sum_{p,q}a_{pq}\left(\delta_{pi}\delta_{qj}+\delta_{pj}\delta_{qi}\right)+\int_G f\,d\mu, \\ &=(a_{ij}+a_{ji})+\int_G f\,d\mu, \\ &=2a_{ij}+\int_G f\,d\mu, \\ &=2a_{ij}, \end{aligned}

since, from (a), we have supp⁡μ⊆H\operatorname{supp}\mu\subseteq H and ff vanishes on HH.

Since Lf(e)=0\mathcal{L}f(e)=0 we necessarily have that aij=0a_{ij}=0 for every pair of transverse indices. Taking j=ij=i shows aii=0a_{ii}=0 for every transverse ii. As aa is symmetric and positive semidefinite, each 2×22\times2 principal minor is nonnegative, so aiiakk−aik2≥0a_{ii}a_{kk}-a_{ik}^2\geq0. Hence aik=0a_{ik}=0 for all kk. Thus every transverse row of aa vanishes, which is precisely ran⁡(a)⊆Lie⁡(H)\operatorname{ran}(a)\subseteq\operatorname{Lie}(H).

(c) b∈Lie⁡(H)b\in\operatorname{Lie}(H).

Fix a transverse index ii and set f=xiχf=x_i\chi. Since χ≡1\chi\equiv1 near ee, we have f=xif=x_i on a neighborhood of ee, so f(e)=0f(e)=0 and the jet of ff at ee coincides with that of xix_i

(Xkf)(e)=(Xkxi)(e)=δki,(XpXqf)(e)=(XpXqxi)(e)=12γpqi,(X_kf)(e)=(X_kx_i)(e)=\delta_{ki},\qquad(X_pX_qf)(e)=(X_pX_qx_i)(e)=\frac{1}{2}\gamma^i_{pq},

where γpqi\gamma^i_{pq} is determined by the relation [Xp,Xq]=∑rγpqrXr[X_p,X_q]=\sum_r\gamma^r_{pq}X_r, see also [War83, Proposition 3.7]. By direct calculation,

Lf(e)=∑p,qapq(XpXqf)(e)+∑pbp(Xpf)(e)+∫G(f(τ)−f(e)−∑kxk(τ)(Xkf)(e))μ(dτ),\mathcal{L}f(e)=\sum_{p,q}a_{pq}(X_pX_qf)(e)+\sum_p b_p(X_pf)(e)+\int_G\left(f(\tau)-f(e)-\sum_k x_k(\tau)(X_kf)(e)\right)\mu(d\tau),

and evaluate its three parts. The first part is given by

∑p,qapq(XpXqf)(e)=∑p,qapq12γpqi=0,\sum_{p,q} a_{pq}(X_pX_qf)(e)=\sum_{p,q}a_{pq}\frac{1}{2}\gamma^i_{pq}=0,

since apq=aqpa_{pq}=a_{qp} while γpqi=−γqpi\gamma^i_{pq}=-\gamma^i_{qp}, so relabelling p↔qp\leftrightarrow q, the sum is equal to its own negative.

The drift term is given by

∑pbp(Xpf)(e)=∑pbpδpi=bi.\sum_p b_p(X_pf)(e)=\sum_p b_p\delta_{pi}=b_i.

The jump term is also given by

∫G(f(τ)−f(e)−∑kxk(τ)(Xkf)(e))μ(dτ)=∫G(xi(τ)χ(τ)−0−∑kxk(τ)δki)μ(dτ)=∫G(xi(τ)χ(τ)−xi(τ))μ(dτ)=∫Gxi(τ)(χ(τ)−1)μ(dτ)=0.\begin{aligned} \int_G\left(f(\tau)-f(e)-\sum_k x_k(\tau)(X_kf)(e)\right)\mu(d\tau) &=\int_G\left(x_i(\tau)\chi(\tau)-0-\sum_k x_k(\tau)\delta_{ki}\right)\mu(d\tau)\\ &=\int_G\left(x_i(\tau)\chi(\tau)-x_i(\tau)\right)\mu(d\tau)\\ &=\int_G x_i(\tau)(\chi(\tau)-1)\mu(d\tau)\\ &=0. \end{aligned}

Here we used f(τ)=xi(τ)χ(τ)f(\tau)=x_i(\tau)\chi(\tau) and f(e)=xi(e)χ(e)=0f(e)=x_i(e)\chi(e)=0 (as xi(e)=0x_i(e)=0), and the compensator ∑kxk(τ)(Xkf)(e)=∑kxk(τ)δki=xi(τ)\sum_k x_k(\tau)(X_kf)(e)=\sum_k x_k(\tau)\delta_{ki}=x_i(\tau). The final integral vanishes because, by part (a), supp⁡μ⊆H\operatorname{supp}\mu\subseteq H, and the transverse coordinate xix_i vanishes on HH, hence xi=0x_i=0 μ\mu-almost everywhere, so the integrand is 00 μ\mu-a.e.

Since Lf(e)=0\mathcal{L}f(e)=0 we get bi=0b_i=0. As ii ranges over the transverse indices, b∈Lie⁡(H)b\in\operatorname{Lie}(H).

Step 5. Conversely, assume supp⁡μ⊂H\operatorname{supp}\mu\subset H, ran⁡(a)⊆Lie⁡(H)\operatorname{ran}(a)\subseteq\operatorname{Lie}(H) and b∈Lie⁡(H)b\in\operatorname{Lie}(H), and keep the adapted basis with X1,…,XmX_1,\ldots,X_m spanning Lie⁡(H)\operatorname{Lie}(H).

The three assumptions, given in (6), tell us when the characteristics have no transverse part.

Note that ajk=0a_{jk}=0 whenever j>mj>m or k>mk>m, bi=0b_i=0 for i>mi>m, and μ\mu is carried by HH. Hence the restricted data

(a∣Lie⁡(H),b∣Lie⁡(H),μ∣H),a∣Lie⁡(H)=(ajk)j,k≤m,b∣Lie⁡(H)=(bi)i≤m,\left(a|_{\operatorname{Lie}(H)},b|_{\operatorname{Lie}(H)},\mu|_H\right),\qquad a|_{\operatorname{Lie}(H)}=(a_{jk})_{j,k\le m},\qquad b|_{\operatorname{Lie}(H)}=(b_i)_{i\le m},

is a Lévy triplet on HH since

  • a∣Lie⁡(H)a|_{\operatorname{Lie}(H)} is symmetric positive semidefinite,

  • μ∣H\mu|_H is a Lévy measure on HH, since for any neighborhood VV of ee one has ∫V∑i≤mxi2 dμ≤∫V∑ixi2 dμ<∞\int_V\sum_{i\le m}x_i^2\,d\mu\le\int_V\sum_i x_i^2\,d\mu<\infty and μ(Vc)<∞\mu(V^c)<\infty.

In turn, by Hunt’s theorem [21] this triplet generates a unique convolution semigroup (ρtH)t≥0(\rho_t^H)_{t\ge0} on HH, with generator L~\widetilde{\mathcal{L}}.

For f∈C∞(G)f\in C^\infty(G), subtracting the two generators leaves only the transverse part of each characteristic

Lf−L~f=∑j>m or k>majkXjXkf+∑i>mbiXif−∫G∑i>mxi(τ)(Xif) μ(dτ).\mathcal{L}f-\widetilde{\mathcal{L}}f= \sum_{\substack{j>m\ \mathrm{or}\ k>m}}a_{jk}X_jX_kf+ \sum_{i>m}b_iX_if- \int_G\sum_{i>m}x_i(\tau)(X_if)\,\mu(d\tau).

Observe that

∑j>m or k>majkXjXkf=0,\sum_{\substack{j>m\ \mathrm{or}\ k>m}}a_{jk}X_jX_kf=0,

since ajk=0a_{jk}=0 once jj or kk is transverse (ran⁡(a)⊆Lie⁡(H)\operatorname{ran}(a)\subseteq\operatorname{Lie}(H)). Moreover, we have that

∑i>mbiXif=0,\sum_{i>m}b_iX_if=0,

since

bi=0,for i>m, recall that b∈Lie⁡(H),b_i=0,\qquad\text{for }i>m,\text{ recall that }b\in\operatorname{Lie}(H),

and

∫G∑i>mxi(τ)(Xif) μ(dτ)=0,\int_G \sum_{i>m} x_i(\tau)(X_i f)\,\mu(d\tau)=0,

since xi=0x_i=0 on supp⁡μ⊆H\operatorname{supp}\mu\subseteq H for transverse ii, so the integrand is 00 μ\mu-a.e. Thus L=LH\mathcal{L}=\mathcal{L}^H on Cc∞(G)C^\infty_c(G).

As a convolution semigroup is determined by the restriction of its generator to Cc∞C^\infty_c (uniqueness in Hunt’s theorem [21]), ρt=ρtH\rho_t=\rho_t^H for every tt. Moreover, since each ρtH\rho_t^H is carried by HH, we conclude ρt(H)=1\rho_t(H)=1 for all tt.

Step 6. By Steps 4 and 5, for a closed subgroup HH one has ρt(H)=1\rho_t(H)=1 for all tt if and only if

supp⁡μ⊆H,ran⁡(a)⊆Lie⁡(H),b∈Lie⁡(H).\operatorname{supp}\mu\subseteq H,\qquad\operatorname{ran}(a)\subseteq\operatorname{Lie}(H),\qquad b\in\operatorname{Lie}(H).

Recalling from Step 3 that P=⟨supp⁡ρ⟩P=\langle\operatorname{supp}\rho\rangle is the smallest closed subgroup with ρt(P)=1\rho_t(P)=1 for all tt, it follows that PP is the smallest closed subgroup satisfying these three conditions. We identify it explicitly. Set h0:=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0:=\operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle, the Lie subalgebra generated by the drift and the range of the diffusion.

If a closed subgroup HH satisfies the three conditions, then ran⁡(a)⊆Lie⁡(H)\operatorname{ran}(a)\subseteq\operatorname{Lie}(H) and b∈Lie⁡(H)b\in\operatorname{Lie}(H), since Lie⁡(H)\operatorname{Lie}(H) is a subalgebra, h0⊆Lie⁡(H)\mathfrak{h}_0\subseteq\operatorname{Lie}(H), whence ⟨exp⁡h0⟩⊆H\langle\exp\mathfrak{h}_0\rangle\subseteq H, and also supp⁡μ⊆H\operatorname{supp}\mu\subseteq H. Therefore

⟨exp⁡h0∪supp⁡μ⟩⊆H.\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\rangle\subseteq H.

Conversely, ⟨exp⁡h0∪supp⁡μ⟩\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\rangle itself satisfies the three conditions, in that it contains supp⁡μ\operatorname{supp}\mu, its Lie algebra contains h0⊇ran⁡(a)\mathfrak{h}_0\supseteq\operatorname{ran}(a), and it contains bb. Being the smallest such subgroup, it equals PP

P=⟨supp⁡ρ⟩=⟨exp⁡h0∪supp⁡μ⟩,h0=Lie⁡⟨b,ran⁡(a)⟩.P=\langle\operatorname{supp}\rho\rangle=\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\rangle,\qquad\mathfrak{h}_0=\operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle.

Finally, suppose P=GP=G, that is, exp⁡h0\exp\mathfrak{h}_0 together with supp⁡μ\operatorname{supp}\mu generates a dense subgroup. Combined with the right-invariance of Step 2 (compact case) or Step 3 (nilpotent case), u(g⋆p)=u(g)u(g\star p)=u(g) for all g∈Gg\in G and p∈P=Gp\in P=G. Taking g=eg=e gives u(p)=u(e)u(p)=u(e) for every p∈Gp\in G, i.e. uu is constant.

Conversely, suppose P≠GP\ne G. Since GG is connected, a proper closed subgroup has positive codimension, so the left coset space G/PG/P is a smooth manifold of dimension at least one. Choose any nonconstant bounded smooth function φ\varphi on G/PG/P and set u(g)=φ(gP)u(g)=\varphi(gP). Since ρt\rho_t is carried by PP, every h∈supp⁡ρth\in\operatorname{supp}\rho_t lies in PP, so g⋆hg\star h and gg determine the same left coset and u(g⋆h)=u(g)u(g\star h)=u(g). Hence u=u∗ρtu=u\ast\rho_t for all t≥0t\ge0, and differentiating at t=0t=0 gives Lu=0\mathcal{L}u=0. Thus uu is a nonconstant bounded solution, so not every bounded solution is constant when P≠GP\ne G. □

A special case of Theorem 1.4 via propagation of maxima

In this section, we aim to prove a special case of Theorem 1.4 via a propagation of maxima approach due to Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b]. We aim to show why we cannot get the full generality of Theorem 1.4 via this approach. Namely, we can only prove the following weaker version of Theorem 1.4.

Theorem 4.1. Let GG be a connected compact Lie group and L\mathcal{L} a left-invariant Lévy operator with characteristics (a,b,μ)(a,b,\mu). Set h0=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0=\operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle and P=⟨exp⁡h0∪supp⁡μ⟩P=\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\rangle. Then every bounded distributional solution uu of Lu=0\mathcal{L}u=0 is constant if and only if P=GP=G.

Note that we do not claim that uu is right-invariant when P≠GP\ne G unlike in Theorem 1.4. We present our proof and highlight the limitations that we face in obtaining Theorem 1.4.

By Proposition 2.1 we may assume u∈C∞(G)u\in C^\infty(G) solves Lu=0\mathcal{L}u=0 pointwise. Since GG is compact and uu is continuous, uu attains its maximum M=max⁡GuM=\max_G u. Set F={g∈G:u(g)=M}F=\{g\in G:u(g)=M\}, a nonempty closed set.

The argument runs in four steps.

Step 1. We aim to show that, for a given gˉ∈F\bar{g} \in F, we have that

∫G(u(gˉ⋆τ)−M)μ(dτ)=0,(8)\int_G \left(u(\bar{g}\star\tau)-M\right)\mu(\mathrm{d}\tau)=0, \tag*{(8)}

and

∑i,jaijXiXju(gˉ)+∑ibiXiu(gˉ)=0.(9)\sum_{i,j}a_{ij}X_iX_ju(\bar{g})+\sum_i b_iX_iu(\bar{g})=0. \tag*{(9)}

Step 2. As a consequence of (8) we show that gˉ⋆supp⁡μ⊂F\bar{g}\star\operatorname{supp}\mu\subset F.

Step 3. On its maximum set uu is a subsolution of the local part of L\mathcal{L}. Hence, by Bony’s strong maximum principle [9], together with the compactness of GG, FF is invariant under exp⁡h0\exp h_0.

Step 4. Combining Steps 2 and 3, FF contains the right coset gˉ⋆P\bar{g}\star P. Hence, uu is constant if and only if P=GP=G.

Proof of Theorem 4.1. Step 1. Fix gˉ∈F\bar{g}\in F. At this maximum Xiu(gˉ)=0X_i u(\bar{g})=0 and the symmetric Hessian is negative semidefinite, so in Lu(gˉ)=0\mathcal{L}u(\bar{g})=0 the drift and compensator terms drop out, leaving

0=∑i,jaijXiXju(gˉ)+∫G(u(gˉ⋆τ)−M)μ(dτ).0=\sum_{i,j}a_{ij}X_iX_ju(\bar{g})+\int_G\left(u(\bar{g}\star\tau)-M\right)\mu(\mathrm{d}\tau).

The first term is ≤0\leq0 (aa is positive semidefinite and the Hessian negative semidefinite) and the second is ≤0\leq0 (as u≤Mu\leq M), so both vanish. This is (8), and, since ∑ibiXiu(gˉ)=0\sum_i b_iX_i u(\bar{g})=0, we also obtain (9).

Step 2. In (8) the integrand u(gˉ⋆τ)−Mu(\bar{g}\star\tau)-M is ≤0\leq0, so it vanishes for μ\mu-a.e. τ\tau, hence for all τ∈supp⁡μ\tau\in\operatorname{supp}\mu by continuity. Thus gˉ⋆supp⁡μ⊂F\bar{g}\star\operatorname{supp}\mu\subset F.

Step 3. Every point of FF is a maximum, so (9) gives Llocu=0\mathcal{L}^{\mathrm{loc}}u=0 on FF, where Lloc=∑i,jaijXiXj+∑ibiXi\mathcal{L}^{\mathrm{loc}}=\sum_{i,j}a_{ij}X_iX_j+\sum_i b_iX_i. In particular, uu is a subsolution of Lloc\mathcal{L}^{\mathrm{loc}} attaining its maximum on FF. By Bony’s strong maximum principle [9], FF is invariant under the diffusion flows exp⁡(tY)\exp(tY), for Y∈ran⁡(a)Y\in\operatorname{ran}(a), t∈Rt\in\mathbb{R}, and the forward drift flow exp⁡(tb)\exp(tb) (t≥0t\geq0). Since GG is compact, {exp⁡(tb):t≥0}\{\exp(tb):t\geq0\} is dense in the torus {exp⁡(tb):t∈R}\{\exp(tb):t\in\mathbb{R}\}, so the drift flow is two-sided. Closing under brackets, gˉ⋆exp⁡h0⊂F\bar{g}\star\exp\mathfrak{h}_0\subset F with h0=Lie⁡⟨b,ran⁡(a)⟩\mathfrak{h}_0=\operatorname{Lie}\langle b,\operatorname{ran}(a)\rangle.

Step 4. Being closed and right-invariant under supp⁡μ\operatorname{supp}\mu and exp⁡h0\exp\mathfrak{h}_0, FF is invariant under the subgroup they generate, so gˉ⋆P⊂F\bar{g}\star P\subset F with P=⟨exp⁡h0∪supp⁡μ⟩P=\langle\exp\mathfrak{h}_0\cup\operatorname{supp}\mu\rangle. If P=GP=G then F=GF=G and uu is constant. If P≠GP\ne G there are nonconstant bounded solutions, constructed as in the proof of Theorem 1.4 (for instance Example 1.7). Hence uu is constant if and only if P=GP=G. □

A comparison to the original approach of Alibaud, del Teso, Endal and Jakobsen

The approach of Alibaud, del Teso, Endal and Jakobsen [AdTEJ20b] establishes more than the dichotomy of Theorem 4.1. It shows that every bounded solution is right-invariant under PP, not merely constant when P=GP=G. We are not able to carry this stronger conclusion over to a general compact or nilpotent Lie group, and we explain why.

On Rd\mathbb{R}^d their argument runs as follows. By Proposition 2.1 take u∈Cb∞(Rd)u\in C_b^\infty(\mathbb{R}^d) solving Lu=0\mathcal{L}u=0 pointwise, and for a jump direction zˉ∈supp⁡μ\bar{z}\in\operatorname{supp}\mu form the increment w(x)=u(x+zˉ)−u(x)w(x)=u(x+\bar{z})-u(x). Since translation commutes with L\mathcal{L}, ww is again a bounded solution. Choosing xnx_n such that w(xn)→sup⁡ww(x_n)\to\sup w, each un:=u(⋅+xn)u_n:=u(\cdot+x_n) is again a bounded solution, and the mollified bounds ∥Dkun∥∞≤∥u∥∞∥Dkρε∥L1\lVert D^k u_n\rVert_\infty\leq\lVert u\rVert_\infty\lVert D^k\rho_\varepsilon\rVert_{L^1} are uniform in nn because, on Rd\mathbb{R}^d, translation commutes with differentiation. By Arzelà–Ascoli a subsequence converges in Cloc2C^2_{\mathrm{loc}} to a solution u∞u_\infty, and w∞=u∞(⋅+zˉ)−u∞w_\infty=u_\infty(\cdot+\bar{z})-u_\infty attains sup⁡w\sup w at the origin, so a value approached but never attained by the original increment becomes a genuine interior maximum of w∞w_\infty. Propagation of the maximum there, with a telescoping argument, forces sup⁡w=0\sup w=0, that is u(⋅+zˉ)=uu(\cdot+\bar{z})=u. The same applies in the diffusion and drift directions, giving the full right-invariance.

Two steps of this argument fail on a general group. First, it is essential that the increment w=u(⋅+zˉ)−uw=u(\cdot+\bar{z})-u again solves the equation. On Rd\mathbb{R}^{d} this is translation invariance, but on a non-commutative group the right translate RzˉuR_{\bar{z}}u solves the operator conjugated by Ad⁡(zˉ−1)\operatorname{Ad}(\bar{z}^{-1}) rather than L\mathcal{L} itself, so ww need not be a solution and the increment method breaks down. This is why, even on a compact group, propagation reaches only the maximum set and the dichotomy of Theorem 4.1, not the right-invariance. Second, manufacturing a maximizer rests on the basepoint-uniform derivative bounds above, which can fail on a nilpotent group, as the following example shows.

Example 4.2. Take the Heisenberg group H1=R3\mathbb{H}^{1}=\mathbb{R}^{3} with the group operation given by

(x,y,t)⋆(x′,y′,t′)=(x+x′,y+y′,t+12(xy′−yx′)),(x,y,t)\star(x',y',t')=\left(x+x',y+y',t+\frac{1}{2}(xy'-yx')\right),

with the left-invariant fields

X=∂x−y2∂t,Y=∂y+x2∂t.X=\partial_x-\frac{y}{2}\partial_t,\qquad Y=\partial_y+\frac{x}{2}\partial_t.

We take L=X2\mathcal{L}=X^{2} and the function uu

u(x,y,t)=sin⁡(y2),for (x,y,t)∈H1.u(x,y,t)=\sin(y^{2}),\qquad\text{for }(x,y,t)\in\mathbb{H}^{1}.

Since uu depends only on yy, Xu=0Xu=0, so uu is a bounded solution of Lu=0\mathcal{L}u=0. It is nonconstant, and right-invariant under the subgroup P={(s,0,0)}P=\{(s,0,0)\} since

u((x,y,t)⋆(s,0,0))=sin⁡(y2)=u(x,y,t).u\big((x,y,t)\star(s,0,0)\big)=\sin(y^{2})=u(x,y,t).

Run the manufacturing step above, sliding along gn=(0,yn,0)g_n=(0,y_n,0) with yn→∞y_n\to\infty and consider the translates u∘Lgnu\circ L_{g_n}. The transverse left-invariant derivative is

Yu=2ycos⁡(y2),∥Yu∥L∞(H1)=∞.Yu=2y\cos(y^{2}),\qquad\|Yu\|_{L^\infty(\mathbb{H}^{1})}=\infty.

In turn, we cannot conclude that {u∘Lgn}\{u\circ L_{g_n}\} is equicontinuous, nor extract a Cloc2C^2_{\mathrm{loc}}-convergent subsequence, as Arzelà–Ascoli fails, and no maximizer can be manufactured.

A harmonic analysis approach of Theorem 1.4 in the case of compact Lie groups

In this section, we provide a harmonic analysis approach to Theorem 1.4, in the case where GG is compact.

In this setting we follow the Fourier-analytic approach to Lévy processes on compact Lie groups developed by Liao [28] and Applebaum [6 Chapter 5] (see also [1]), in which the generator is realized as a Fourier multiplier acting on the matrix coefficients of the irreducible representations. We use the Peter–Weyl theory of harmonic analysis on compact Lie groups [31], cf. the books of Hall [20 Chapter 12] and Sepanski [33 §3.3].

The Peter–Weyl theory decomposes the L2(G)L^{2}(G) space into isotypic components

L2(G)=⨁π∈G^Vπ.L^{2}(G)=\bigoplus_{\pi\in\widehat{G}}V_{\pi}.

The Fourier transform, for f∈D(G)f\in\mathcal{D}(G), is understood as

F(f)(π)=∫Gf(g)π(g−1) dg,for π∈G^,\mathcal{F}(f)(\pi)=\int_{G}f(g)\pi(g^{-1})\,dg,\qquad\text{for }\pi\in\widehat{G},

where dgdg is the Haar measure on GG. We write S(G^)\mathcal{S}(\widehat{G}) for the Schwartz space on the unitary dual G^\widehat{G}, consisting of the matrix fields h=(h(π))π∈G^h=(h(\pi))_{\pi\in\widehat{G}}, with h(π)∈Cdim⁡(π)×dim⁡(π)h(\pi)\in\mathbb{C}^{\dim(\pi)\times\dim(\pi)}, that decay rapidly in the sense that sup⁡π∈G^(1+λπ)N∥h(π)∥<∞\sup_{\pi\in\widehat{G}}(1+\lambda_{\pi})^{N}\|h(\pi)\|<\infty for every N∈NN\in\mathbb{N}, where λπ≥0\lambda_{\pi}\geq0 is the eigenvalue by which the Laplace–Beltrami operator acts on the matrix coefficients of π\pi; the Fourier transform is a linear bijection from D(G)\mathcal{D}(G) onto S(G^)\mathcal{S}(\widehat{G}). The Fourier inverse is given as

F−1(h)(g)=∑π∈G^dim⁡(π)Tr⁡(π(g)h(π)),for g∈G, h∈S(G^).\mathcal{F}^{-1}(h)(g)=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}(\pi(g)h(\pi)),\quad\text{for }g\in G,\ h\in\mathcal{S}(\widehat{G}).

In the compact case the components VπV_\pi are all finite dimensional vector spaces, thus the representations are finite dimensional homomorphisms from G→GL⁡(Vπ)G\to\operatorname{GL}(V_\pi), (GL⁡(Vπ)\operatorname{GL}(V_\pi) is the space of invertible matrices acting on VπV_\pi).

The corresponding distributional definition of the Fourier transform is given as

⟨Fu,φ⟩G^=⟨u,(ι∘F−1)(φ)⟩G,\langle\mathcal{F}u,\varphi\rangle_{\widehat{G}}=\langle u,(\iota\circ\mathcal{F}^{-1})(\varphi)\rangle_G,

where φ∈S(G^)\varphi\in\mathcal{S}(\widehat{G}) and (ι∘φ)(x)=φ(x−1)(\iota\circ\varphi)(x)=\varphi(x^{-1}) for φ∈D(G)\varphi\in\mathcal{D}(G).

The pairing in G^\widehat{G} is given as

⟨u,φ⟩G^=∑π∈G^dim⁡(π)Tr⁡(u(π)φ(π)),\langle u,\varphi\rangle_{\widehat{G}}=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}(u(\pi)\varphi(\pi)),

where Tr⁡\operatorname{Tr} is the usual trace operator for matrices.

From [6 Theorem 5.5.1] (cf. [28]), we have that

Lf^(π)=L^(π)f^(π),for π∈G^,\widehat{\mathcal{L}f}(\pi)=\widehat{\mathcal{L}}(\pi)\widehat{f}(\pi),\quad\text{for }\pi\in\widehat{G},

where

L^(π)=∑i,j=1daij dπ(Xi)dπ(Xj)+∑i=1dbi dπ(Xi)+∫G∖{e}(π(τ)−I−∑i=1dxi(τ)dπ(Xi))μ(dτ),\widehat{\mathcal{L}}(\pi)=\sum_{i,j=1}^{d}a_{ij}\,d\pi(X_i)d\pi(X_j)+\sum_{i=1}^{d}b_i\,d\pi(X_i)+\int_{G\setminus\{e\}}\left(\pi(\tau)-I-\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)\right)\mu(d\tau),

and dπ(Xi)d\pi(X_i) is derived representation of π(Xi)\pi(X_i).

We now outline our approach proving Theorem 1.4, in the compact setting.

Step 1. We aim to show, using the notion of solution given in Definition 1.2, that we have

L^(π)u^(π)=0,for all π∈G^,\widehat{\mathcal{L}}(\pi)\widehat{u}(\pi)=0,\quad\text{for all }\pi\in\widehat{G},

where G^\widehat{G} is the unitary dual of GG.

Step 2. In this step, our goal is to characterize the kernel of L^(π)\widehat{\mathcal{L}}(\pi). Firstly, for a given v∈ker⁡L^(π)v\in\ker\widehat{\mathcal{L}}(\pi) (the kernel of L^(π)\widehat{\mathcal{L}}(\pi)) we necessarily have that

Re⁡⟨L^(π)v,v⟩=0.\operatorname{Re}\langle\widehat{\mathcal{L}}(\pi)v,v\rangle=0.

As a consequence, we determine that v∈ker⁡L^(π)v\in\ker\widehat{\mathcal{L}}(\pi) must satisfy

  • dπ(Yl)v=0d\pi(Y_l)v=0, for Yl=∑jσjlXj∈Lie⁡(G)Y_l=\sum_j\sigma_{jl}X_j\in\operatorname{Lie}(G),

  • (π(τ)−I)v=0(\pi(\tau)-I)v=0 for τ∈supp⁡(μ)\tau\in\operatorname{supp}(\mu),

More specifically, we have that vv must be invariant under the action of the subgroup H0H_0 generated by {exp⁡(tYl)∣∀l,t∈R}\{\exp(tY_l)|\forall l,t\in\mathbb{R}\} and supp⁡(μ)\operatorname{supp}(\mu).

Step 3. In this step we show that v∈ker⁡L^(π)v\in\ker\widehat{\mathcal{L}}(\pi) must satisfy dπ(X~b)v=0d\pi(\widetilde{X}_b)v=0 where

X~b:=∑i=1dbiXi−∫G∖U(∑i=1dxi(τ)Xi)μ(dτ).\widetilde{X}_b:=\sum_{i=1}^{d}b_iX_i-\int_{G\setminus U}\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)\mu(d\tau).

and UU is a neighborhood of the identity. Thus v∈ker⁡(L^(π))v\in\ker(\widehat{\mathcal{L}}(\pi)) must be invariant under the subgroup HH which is generated by H0H_0 and {exp⁡(tX~b)∣t∈R}\{\exp(t\widetilde{X}_b)|t\in\mathbb{R}\}.

Step 4. In this step, we demonstrate that X~b\widetilde{X}_b is in fact well defined and show that the subgroup HH is not dependent on the choice of neighborhood UU.

Step 5. We show HH periodicity of solutions uu and prove the converse, showing that any HH periodic distribution uu is a solution.

Proof of Theorem 1.4 in the compact case via Peter–Weyl.

Step 1. From the notion of solution, given in Definition 1.2, we have that ⟨Lu,φ⟩G=⟨u,L†φ⟩G\langle Lu,\varphi\rangle_G=\langle u,L^\dagger\varphi\rangle_G for φ∈D(G)\varphi\in\mathcal{D}(G). By then taking the distributional Fourier transform

⟨u,L†φ⟩G=0⟺⟨u^,ι∘(L†φ)^⟩G^=0.\langle u,L^\dagger\varphi\rangle_G=0\Longleftrightarrow\langle\widehat{u},\iota\circ\widehat{(L^\dagger\varphi)}\rangle_{\widehat{G}}=0.

The term

ι∘(L†φ)^(π)=∫G(ι∘(L†φ))(g)π(g−1) dg,=∫G(L†φ)(g−1)π(g−1) dg,=∫G(L†φ)(g)π(g) dg,=∫Gφ(g)L(π)(g) dg,\begin{aligned} \widehat{\iota\circ(L^\dagger\varphi)}(\pi)=\int_G(\iota\circ(L^\dagger\varphi))(g)\pi(g^{-1})\,dg, \\ &=\int_G(L^\dagger\varphi)(g^{-1})\pi(g^{-1})\,dg, \\ &=\int_G(L^\dagger\varphi)(g)\pi(g)\,dg, \\ &=\int_G\varphi(g)L(\pi)(g)\,dg, \end{aligned}

where we have used the inversion invariance of the Haar measure and the definition of the transpose. Computing the action of the operator on π\pi we see

L(π)(g)=π(g)L^(π),for (g,π)∈G×G^,L(\pi)(g)=\pi(g)\widehat{L}(\pi),\qquad\text{for }(g,\pi)\in G\times\widehat{G},

thus

∫Gφ(g)Lπ(g) dg=(∫Gφ(g)π(g) dg)L^(π).\int_G\varphi(g)L\pi(g)\,dg=\left(\int_G\varphi(g)\pi(g)\,dg\right)\widehat{L}(\pi).

Moreover, we have that

⟨u^,ι∘(L†φ)^⟩G^=⟨u^,(∫Gφ(g)π(g) dg)L^(π)⟩G^=⟨u^,ι∘φ^L^⟩G^.\langle\widehat{u},\widehat{\iota\circ(L^\dagger\varphi)}\rangle_{\widehat{G}}=\left\langle\widehat{u},\left(\int_G\varphi(g)\pi(g)\,dg\right)\widehat{L}(\pi)\right\rangle_{\widehat{G}}=\langle\widehat{u},\widehat{\iota\circ\varphi}\widehat{L}\rangle_{\widehat{G}}.

Evaluating the pairing we see

⟨u^,ι∘φ^L^⟩G^=∑π∈G^dim⁡(π)Tr⁡(u^(π)ι∘φ^(π)L^(π)),=∑π∈G^dim⁡(π)Tr⁡(L^(π)u^(π)ι∘φ^(π))=0.\begin{aligned} \langle\widehat{u},\widehat{\iota\circ\varphi}\widehat{L}\rangle_{\widehat{G}}=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\left(\widehat{u}(\pi)\widehat{\iota\circ\varphi}(\pi)\widehat{L}(\pi)\right), \\ &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\left(\widehat{L}(\pi)\widehat{u}(\pi)\widehat{\iota\circ\varphi}(\pi)\right)=0. \end{aligned}

As a consequence, we have that

L^(π)u^(π)=0,for all π∈G^.\widehat{L}(\pi)\widehat{u}(\pi)=0,\qquad\text{for all }\pi\in\widehat{G}.

Both these terms are matrices acting on the finite-dimensional representation space VπV_\pi, so this expression is 0 when the image of u^(π)\widehat{u}(\pi) is contained in the kernel of L^(π)\widehat{L}(\pi).

Step 2. Our goal is to determine the kernel of L^(π)\widehat{L}(\pi). Suppose v∈Vπv\in V_\pi is in ker⁡(L^(π))\ker(\widehat{L}(\pi)). Then necessarily, we have that

⟨L^(π)v,v⟩Vπ=0,\langle\widehat{L}(\pi)v,v\rangle_{V_\pi}=0,

where VπV_\pi is the vector space associated with the representation π\pi. From the representation theory of compact Lie groups we know that the irreducible representations are finite dimensional. The operators L^(π)\widehat{L}(\pi) are thus matrices of dimension dim⁡(π)\dim(\pi). First, we examine the real part. The term ⟨dπ(Xi)v,v⟩\langle d\pi(X_i)v,v\rangle is purely imaginary as dπ(Xi)d\pi(X_i) is skew-adjoint. As a consequence, we have that

Re⁡⟨L^(π)v,v⟩=∑j,k=1dajk⟨dπ(Xj)dπ(Xk)v,v⟩+∫G∖{e}(Re⁡⟨π(τ)v,v⟩−⟨v,v⟩)μ(dτ).\operatorname{Re}\langle\widehat{L}(\pi)v,v\rangle=\sum_{j,k=1}^{d}a_{jk}\langle d\pi(X_j)d\pi(X_k)v,v\rangle+\int_{G\setminus\{e\}}\left(\operatorname{Re}\langle\pi(\tau)v,v\rangle-\langle v,v\rangle\right)\mu(d\tau).

The d×dd \times d matrix (ajk)(a_{jk}) is nonnegative definite and symmetric, thus, along with the skew-adjointness of dπ(Xi)d\pi(X_i), we have ∑j,k=1dajk⟨dπ(Xj)dπ(Xk)v,v⟩≤0\sum_{j,k=1}^{d} a_{jk}\langle d\pi(X_j)d\pi(X_k)v,v\rangle\le0. Further, (ajk)(a_{jk}) admits a decomposition a=σσTa=\sigma\sigma^{T}, i.e. writing out the matrix multiplication element-wise: ajk=∑lσjlσkla_{jk}=\sum_l \sigma_{jl}\sigma_{kl}. The second-order term becomes

∑j,k=1dajk dπ(Xj) dπ(Xk)=∑l(∑jσjl dπ(Xj))(∑kσkl dπ(Xk))=∑ldπ(Yl)2,\sum_{j,k=1}^{d} a_{jk}\,d\pi(X_j)\,d\pi(X_k)=\sum_l\left(\sum_j\sigma_{jl}\,d\pi(X_j)\right)\left(\sum_k\sigma_{kl}\,d\pi(X_k)\right)=\sum_l d\pi(Y_l)^2,

where Yl:=∑jσjlXj∈Lie⁡(G)Y_l:=\sum_j\sigma_{jl}X_j\in\operatorname{Lie}(G). The term

⟨∑ldπ(Yl)2v,v⟩=0\left\langle\sum_l d\pi(Y_l)^2v,v\right\rangle=0

when dπ(Yl)v=0d\pi(Y_l)v=0. By applying the exponential, we see that π(exp⁡(tYl))v=v\pi(\exp(tY_l))v=v. The integrand in the second term is always negative as

Re⁡⟨π(τ)v,v⟩−∥v∥2≤∣⟨π(τ)v,v⟩∣−∥v∥2≤∥π(τ)v∥∥v∥−∥v∥2≤0,\operatorname{Re}\langle\pi(\tau)v,v\rangle-\lVert v\rVert^2\leq\lvert\langle\pi(\tau)v,v\rangle\rvert-\lVert v\rVert^2\leq\lVert\pi(\tau)v\rVert\lVert v\rVert-\lVert v\rVert^2\leq0,

as a consequence of the Cauchy–Schwarz inequality and using the fact that π\pi is a unitary representation. Since π(τ)\pi(\tau) is unitary, ∥π(τ)v∥=∥v∥\lVert\pi(\tau)v\rVert=\lVert v\rVert, thus the integrand is always ≤0\leq0. Therefore as the integrand is negative the integral must be negative. This implies that the integrand must vanish a.e. on the support of the measure, i.e.

π(τ)v=v\pi(\tau)v=v

for all τ∈supp⁡(μ)\tau\in\operatorname{supp}(\mu).

In turn, we have that

Re⁡⟨L^(π)v,v⟩=∑j,k=1dajk⟨dπ(Xj)dπ(Xk)v,v⟩⏟≤0+∫G∖{e}(Re⁡⟨π(τ)v,v⟩−⟨v,v⟩)μ(dτ)⏟≤0≤0.\operatorname{Re}\langle\widehat{L}(\pi)v,v\rangle=\underbrace{\sum_{j,k=1}^{d}a_{jk}\langle d\pi(X_j)d\pi(X_k)v,v\rangle}_{\leq0}+\underbrace{\int_{G\setminus\{e\}}\left(\operatorname{Re}\langle\pi(\tau)v,v\rangle-\langle v,v\rangle\right)\mu(d\tau)}_{\leq0}\leq0.

Step 3. We now examine L^(π)v=0\widehat{L}(\pi)v=0. Define H0H_0 to be the closed subgroup generated by the support of μ\mu and the one-parameter subgroup {exp⁡(tYl)∣∀l,t∈R}\{\exp(tY_l)\mid\forall l,t\in\mathbb{R}\}. From step 2, we see that given vv invariant under the closed subgroup H0H_0, the real parts vanish leaving

L^(π)v=∑i,j=1daij dπ(Xi)dπ(Xj)v+∑i=1dbi dπ(Xi)v+∫G∖{e}(π(τ)v−v−∑i=1dxi(τ)dπ(Xi)v)μ(dτ),=∑i=1dbi dπ(Xi)v−∫G∖{e}(∑i=1dxi(τ)dπ(Xi))vμ(dτ),=0.\begin{aligned} \widehat{L}(\pi)v=\sum_{i,j=1}^{d}a_{ij}\,d\pi(X_i)d\pi(X_j)v+\sum_{i=1}^{d}b_i\,d\pi(X_i)v+\int_{G\setminus\{e\}}\left(\pi(\tau)v-v-\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)v\right)\mu(d\tau), \\ &=\sum_{i=1}^{d}b_i\,d\pi(X_i)v-\int_{G\setminus\{e\}}\left(\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)\right)v\mu(d\tau), \\ &=0. \end{aligned}

Here, we would like to define the Lie algebra element

Xb~:=∑i=1dbiXi−∫G∖{e}(∑i=1dxi(τ)Xi)μ(dτ),(10)X_{\widetilde{b}}:=\sum_{i=1}^{d}b_iX_i-\int_{G\setminus\{e\}}\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)\mu(d\tau), \tag*{(10)}

and conclude that any vector vv in the kernel of the symbol must be dπ(Xb~)d\pi(X_{\widetilde{b}}) invariant. However the integral on the right hand side of (10) is not necessarily convergent as the Lévy measure need only integrate ∣x∣2\lvert x\rvert^2 in a neighborhood of the identity. Here we have a term that looks like ∣x∣\lvert x\rvert. Away from the identity, x(τ)x(\tau) is bounded so the integral is finite there.

We thus instead define

X~b:=∑i=1dbiXi−∫G∖U(∑i=1dxi(τ)Xi) μ(dτ),(11)\widetilde{X}_b := \sum_{i=1}^{d} b_i X_i - \int_{G\setminus U} \left(\sum_{i=1}^{d} x_i(\tau)X_i\right)\,\mu(d\tau), \tag*{(11)}

where UU is a neighborhood of the identity specified below, and show that in any direction v∈ker⁡(L^(π))v \in\ker(\widehat{L}(\pi)), the integrand vanishes.

By the closed subgroup theorem [20 Thm 3.42.], there exists a neighborhood WW of 00 in Lie⁡(G)\operatorname{Lie}(G) such that H0∩exp⁡(W)=exp⁡(Lie⁡(H0)∩W)H_0 \cap\exp(W) = \exp(\operatorname{Lie}(H_0) \cap W) and exp⁡\exp is a diffeomorphism on WW. Choosing some τ∈H0∩exp⁡(W)\tau\in H_0 \cap\exp(W), we can then write log⁡(τ)∈Lie⁡(H0)\log(\tau) \in\operatorname{Lie}(H_0), where exp⁡(tlog⁡(τ))\exp(t\log(\tau)) is an element of H0H_0 for all t∈Rt \in\mathbb{R}.

The coordinate functions xix_i are canonical, so for τ∈supp⁡(μ)⊂H0\tau\in\operatorname{supp}(\mu) \subset H_0 near the identity we have

∑i=1dxi(τ)Xi=log⁡(τ)∈Lie⁡(H0).\sum_{i=1}^{d} x_i(\tau)X_i = \log(\tau) \in\operatorname{Lie}(H_0).

So,

∫U(∑i=1dxi(τ)Xi) μ(dτ)=∫Ulog⁡(τ) μ(dτ),\int_U \left(\sum_{i=1}^{d} x_i(\tau)X_i\right)\,\mu(d\tau) = \int_U \log(\tau)\,\mu(d\tau),

and exp⁡(t∑i=1dxi(τ)Xi)∈H0\exp\left(t\sum_{i=1}^{d}x_i(\tau)X_i\right) \in H_0 for all t∈Rt \in\mathbb{R}. In turn,

dπ(∑i=1dxi(τ)Xi)v=0⟺π(exp⁡(t∑i=1dxi(τ)Xi))v=v.d\pi\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)v = 0 \Longleftrightarrow\pi\left(\exp\left(t\sum_{i=1}^{d}x_i(\tau)X_i\right)\right)v = v.

Choosing U⊂exp⁡(W)U \subset\exp(W) we then have

∫U∥(∑i=1dxi(τ)dπ(Xi))v∥ μ(dτ)=0.\int_U \left\|\left(\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)\right)v\right\|\,\mu(d\tau) = 0.

We then can conclude that

dπ(X~b)v=∑i=1dbidπ(Xi)v−∫G∖U(∑i=1dxi(τ)dπ(Xi))vμ(dτ),=∑i=1dbidπ(Xi)v−∫G∖{e}(∑i=1dxi(τ)dπ(Xi))vμ(dτ),\begin{aligned} d\pi(\widetilde{X}_b)v = \sum_{i=1}^{d}b_i d\pi(X_i)v - \int_{G\setminus U}\left(\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)\right)v\mu(d\tau), \\ &= \sum_{i=1}^{d}b_i d\pi(X_i)v - \int_{G\setminus\{e\}}\left(\sum_{i=1}^{d}x_i(\tau)d\pi(X_i)\right)v\mu(d\tau), \end{aligned}

as any contribution of τ∈U\tau\in U will vanish in the direction of vv.

Step 4. The definition (11) of X~b\widetilde{X}_b is clearly dependent on the choice of U⊂exp⁡(W)U \subset\exp(W). Without loss of generality assume a different choice of neighborhood U′U' is a subset of UU, then

∫G∖U′(∑i=1dxi(τ)Xi) μ(dτ)−∫G∖U(∑i=1dxi(τ)Xi) μ(dτ)=∫U∖U′(∑i=1dxi(τ)Xi) μ(dτ).\begin{aligned} \int_{G\setminus U'}\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)\,\mu(d\tau) - \int_{G\setminus U}\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)\,\mu(d\tau) \\ &= \int_{U\setminus U'}\left(\sum_{i=1}^{d}x_i(\tau)X_i\right)\,\mu(d\tau). \end{aligned}

See then that as the integral is finite and ∑i=1dxi(τ)Xi∈Lie⁡(H0)\sum_{i=1}^{d}x_i(\tau)X_i \in\operatorname{Lie}(H_0) then the integral defines an element Z∈Lie⁡(H0)Z \in\operatorname{Lie}(H_0). The subgroup HH is closed so Lie⁡(H)\operatorname{Lie}(H) is a subalgebra of Lie⁡(G)\operatorname{Lie}(G). As Z,XbU∈Lie⁡(H)Z, X_b^U \in\operatorname{Lie}(H) then XbU′=XbU−Z∈Lie⁡(H)X_b^{U'} = X_b^U - Z \in\operatorname{Lie}(H). By definition HH then contains {exp⁡(tXbU′)∣t∈R}\{\exp(tX_b^{U'})\mid t \in\mathbb{R}\} so the choice of UU does not change HH.

Step 5. Finally we can show HH periodicity of uu. Let RhR_h denote right translation. Then

⟨Rhu,ϕ⟩=⟨u,Rh−1ϕ⟩.\langle R_hu,\phi\rangle= \langle u,R_{h^{-1}}\phi\rangle.

Taking the Fourier transform

⟨u^,ι∘Rh−1φ^⟩.\langle\widehat{u},\widehat{\iota\circ R_{h^{-1}}\varphi}\rangle.

First note

ι∘Rh−1φ=∫G(ι∘Rh−1φ)(g)π(g−1) dg,=∫G(Rh−1φ)(g−1)π(g−1) dg,=∫Gφ(g−1h−1)π(g−1) dg.\begin{aligned} \iota\circ R_{h^{-1}}\varphi &=\int_G(\iota\circ R_{h^{-1}}\varphi)(g)\pi(g^{-1})\,dg,\\ &=\int_G(R_{h^{-1}}\varphi)(g^{-1})\pi(g^{-1})\,dg,\\ &=\int_G\varphi(g^{-1}h^{-1})\pi(g^{-1})\,dg. \end{aligned}

Take g→h−1g−1g\to h^{-1}g^{-1}

∫Gφ(g)π(gh) dg=∫Gφ(g)π(g) dgπ(h),=∫G(ι∘φ)(g−1)π(g) dgπ(h),=∫G(ι∘φ)(g)π(g−1) dgπ(h),=ι∘φ^ π(h).\begin{aligned} \int_G\varphi(g)\pi(gh)\,dg &=\int_G\varphi(g)\pi(g)\,dg\pi(h),\\ &=\int_G(\iota\circ\varphi)(g^{-1})\pi(g)\,dg\pi(h),\\ &=\int_G(\iota\circ\varphi)(g)\pi(g^{-1})\,dg\pi(h),\\ &=\widehat{\iota\circ\varphi}\,\pi(h). \end{aligned}

Thus

⟨u^,ι∘Rh−1φ^⟩=∑π∈G^dim⁡(π)Tr⁡(u^(π)ι∘Rh−1φ^(π)),=∑π∈G^dim⁡(π)Tr⁡(u^(π)ι∘Rh−1φ^(π)),=∑π∈G^dim⁡(π)Tr⁡(u^(π)ι∘φ^(π)π(h)),=∑π∈G^dim⁡(π)Tr⁡(π(h)u^(π)ι∘φ^(π)),=∑π∈G^dim⁡(π)Tr⁡(u^(π)ι∘φ^(π)),=⟨u^,ι∘φ^⟩,\begin{aligned} \langle\widehat{u},\widehat{\iota\circ R_{h^{-1}}\varphi}\rangle &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\widehat{u}(\pi)\widehat{\iota\circ R_{h^{-1}}\varphi}(\pi)\bigr),\\ &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\widehat{u}(\pi)\widehat{\iota\circ R_{h^{-1}}\varphi}(\pi)\bigr),\\ &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\widehat{u}(\pi)\widehat{\iota\circ\varphi}(\pi)\pi(h)\bigr),\\ &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\pi(h)\widehat{u}(\pi)\widehat{\iota\circ\varphi}(\pi)\bigr),\\ &=\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\widehat{u}(\pi)\widehat{\iota\circ\varphi}(\pi)\bigr),\\ &=\langle\widehat{u},\widehat{\iota\circ\varphi}\rangle, \end{aligned}

where the second last equality uses π(h)u^(π)=u^(π)\pi(h)\widehat{u}(\pi)=\widehat{u}(\pi) and inverting the Fourier transform, we determine that uu is an HH-periodic distribution.

The converse follows as a direct consequence of

⟨Lu^,Φ^⟩=∑π∈G^dim⁡(π)Tr⁡(L^(π)u^(π)Φ^(π)).\langle\widehat{Lu},\widehat{\Phi}\rangle =\sum_{\pi\in\widehat{G}}\dim(\pi)\operatorname{Tr}\bigl(\widehat{L}(\pi)\widehat{u}(\pi)\widehat{\Phi}(\pi)\bigr).

As π(h)u^(π)=u^(π)\pi(h)\widehat{u}(\pi)=\widehat{u}(\pi), then

π(h)u^(π)v=u^(π)v.\pi(h)\widehat{u}(\pi)v=\widehat{u}(\pi)v.

This implies that the image of u^\widehat{u} is fixed under the action of h∈Hh\in H. If v∈Vπv\in V_\pi is fixed by HH, then dπ(Yl)v=0d\pi(Y_l)v=0 for every ll and dπ(X~b)v=0d\pi(\widetilde{X}_b)v=0, while π(τ)v=v\pi(\tau)v=v for every τ∈supp⁡μ\tau\in\operatorname{supp}\mu. Substituting into L^(π)\widehat{L}(\pi), the diffusion term equals ∑ldπ(Yl)2v=0\sum_l d\pi(Y_l)^2v=0, the jump term ∫G∖{e}(π(τ)−I)v μ(dτ)=0\int_{G\setminus\{e\}}(\pi(\tau)-I)v\,\mu(d\tau)=0, and the remaining drift and compensator combine to dπ(X~b)v=0d\pi(\widetilde{X}_b)v=0, the contribution from UU vanishing as in Step 3. Hence the HH-fixed vectors lie in the kernel of L^(π)\widehat{L}(\pi), and we conclude that

L^(π)u^(π)=0.\widehat{L}(\pi)\widehat{u}(\pi)=0.

Thus,

⟨Lu^,φ^⟩=0,\langle\widehat{\mathcal{L}u},\widehat{\varphi}\rangle=0,

and upon Fourier inversion ⟨Lu,φ⟩=0\langle\mathcal{L}u,\varphi\rangle=0. When H=GH=G we have the Liouville property.

Some observations for the nilpotent case

It is natural to look for a Fourier-analytic proof of Theorem 1.4 where GG is a nilpotent Lie group such that [G,[G,G]][G,[G,G]] is relatively compact. The right substitution for the Peter–Weyl theorem in this case is the Kirillov orbit method [24]. The corresponding group Fourier transform turns L\mathcal{L} into a field of operators π(L)\pi(\mathcal{L}) indexed by the coadjoint orbits, and the scalar condition becomes the requirement that π(L)\pi(\mathcal{L}) be injective on every non-trivial representation (the Rockland condition). For solutions in L2L^2, or with enough decay, this already gives the Liouville property through hypoellipticity. Bounded/distributional solutions need much more consideration for the following two reasons

  • The Fourier transform of a bounded function is an operator-valued distribution on G^\widehat{G}, so one needs the full distributional calculus of the orbit method,

  • The inverse of π(L)\pi(\mathcal{L}) is not uniformly bounded as the orbits collapse to the trivial representation, which is exactly the near-abelian regime where the step-dependent behavior discussed above lives.

Making this precise would take a development well beyond the scope of the present paper. Nevertheless, we highlight one subtlety specific to the nilpotent setting. On the Heisenberg group H1\mathbb{H}^1 write L0=X2+Y2\mathcal{L}_0=X^2+Y^2 for the sub-Laplacian and fix s∈(0,1)s\in(0,1). There are two standard definitions of the fractional sub-Laplacian (−L0)s(-\mathcal{L}_0)^s. The Bochner definition subordinates the heat semigroup Pτ=eτL0P_\tau=e^{\tau\mathcal{L}_0},

(−L0)sf=cs∫0∞(f−Pτf)dττ1+s,cs=sΓ(1−s)>0,(-\mathcal{L}_0)^s f=c_s\int_0^\infty(f-P_\tau f)\frac{d\tau}{\tau^{1+s}},\qquad c_s=\frac{s}{\Gamma(1-s)}>0,

and makes sense only when this integral converges in the sense of Bochner. The Fourier-analytic definition uses the spectral resolution −L0=∫0∞λ dEλ-\mathcal{L}_0=\int_0^\infty\lambda\,dE_\lambda and sets (−L0)s=∫0∞λs dEλ(-\mathcal{L}_0)^s=\int_0^\infty\lambda^s\,dE_\lambda. Both of these definitions agree on L2(H1)L^2(\mathbb{H}^1). However, on more general distributions the two definitions can disagree. We demonstrate this on polynomials. If ff is a polynomial with L0f≠0\mathcal{L}_0f\ne0, then f−Pτf=−∑k≥1τkk!L0kff-P_\tau f=-\sum_{k\ge1}\frac{\tau^k}{k!}\mathcal{L}_0^k f is a nonzero polynomial in τ\tau, so the integrand in the Bochner formula behaves like τk−1−s\tau^{k-1-s} and the integral diverges. Hence, (−L0)sf(-\mathcal{L}_0)^sf is undefined in the Bochner sense. The Fourier-analytic definition, by contrast, remains well defined on such ff, by analytic continuation in ss [14]. In short, polynomials lie in the domain of the Fourier-analytic fractional sub-Laplacian but not of the Bochner one. They therefore fall outside the bounded setting of Theorem 1.4, and can be used only in the Fourier-analytic, distributional formulation. For the authors, this raises a few questions with regard to the correct notion of solution for the distributional case. This is a future research problem that the authors would like to revisit and understand in more detail.

Acknowledgements

The authors would like to thank Anthony Dooley for helpful discussions.

Declaration on the use of generative AI

During the preparation of this manuscript the authors used the generative AI assistant Claude (Anthropic), which brought the Choquet–Deny theorem of Bénard and Frisch [8] to the authors’ attention as the key ingredient for the nilpotent case, and which assisted with editing and exposition. The authors reviewed and verified all content and take full responsibility for it. No generative AI tool is credited as an author.

References

School of Mathematical and Physical Sciences, University of Technology Sydney (UTS) Email address: [email protected]

School of Mathematical and Physical Sciences, University of Technology Sydney (UTS) Email address: [email protected]

References

  1. [1]David Applebaum and Anthony Dooley. A generalised Gangolli-Lévy-Khintchine formula for infinitely divisible measures and Lévy processes on semi-simple Lie groups and symmetric spaces. Ann. Inst. Henri Poincaré Probab. Stat., 51(2):599–619, 2015.
  2. [4]David Applebaum and Trang Le Ngan. The positive maximum principle on Lie groups. J. Lond. Math. Soc. (2), 101(1):136–155, 2020.DOI
  3. [5]David Applebaum and Trang Le Ngan. The positive maximum principle on symmetric spaces. Positivity, 24(5):1519–1533, 2020.arxiv.org/abs/1902.03836
  4. [6]David Applebaum. Probability on compact Lie groups, volume 70 of Probability Theory and Stochastic Modelling. Springer, Cham, 2014. With a foreword by Herbert Heyer.
  5. [8]Timothée Bénard and Joshua Frisch. Classification of Choquet–Deny Lie groups. 2026. Preprint, arXiv:2609.11731.arxiv.org/abs/2609.11731
  6. [9]Jean-Michel Bony. Principe du maximum, inégalité de Harnack et unicité du problème de Cauchy pour les opérateurs elliptiques dégénérés. Ann. Inst. Fourier (Grenoble), 19(fasc. 1):277–304 xii, 1969.DOI
  7. [10]David Berger and René L. Schilling. On the Liouville and strong Liouville properties for a class of non-local operators. Math. Scand., 128(2):365–388, 2022.
  8. [11]David Berger, René L. Schilling, and Eugene Shargorodsky. The Liouville theorem for a class of Fourier multipliers and its connection to coupling. Bull. Lond. Math. Soc., 56(7):2374–2394, 2024.arxiv.org/abs/2211.08929
  9. [12]David Berger, René L. Schilling, Eugene Shargorodsky, and Teo Sharia. An extension of the Liouville theorem for Fourier multipliers to sub-exponentially growing solutions. J. Spectr. Theory, 14(2):665–695, 2024.arxiv.org/abs/2401.12876
  10. [13]Gustave Choquet and Jacques Deny. Sur l’équation de convolution μ = μ ∗ σ. C. R. Acad. Sci. Paris, 250:799–801, 1960.
  11. [14]Francesca Corni and Fausto Ferrari. The fractional powers of the sub-Laplacian in Carnot groups through an analytic continuation. J. Geom. Anal., 35(4):Paper No. 106, 2025.
  12. [15]E. B. Dynkin and M. B. Malyutov. Random walks on groups with a finite number of generators. Soviet Math. Dokl., 2:399–402, 1961.
  13. [16]Mouhamed Moustapha Fall. Entire s-harmonic functions are affine. Proc. Amer. Math. Soc., 144(6):2587–2592, 2016.arxiv.org/abs/1407.5934
  14. [17]Joshua Frisch, Yair Hartman, Omer Tamuz, and Pooya Vahidi Ferdowsi. Choquet–Deny groups and the infinite conjugacy class property. Ann. of Math. (2), 190(1):307–320, 2019.DOI
  15. [18]Tomasz Grzywny and Mateusz Kwaśnicki. Liouville’s theorems for Lévy operators. Math. Ann., 391(4):5857–5910, 2025.arxiv.org/abs/2301.08540
  16. [19]Yves Guivarc’h. Croissance polynomiale et périodes des fonctions harmoniques. Bull. Soc. Math. France, 101:333–379, 1973.
  17. [20]Brian Hall. Lie groups, Lie algebras, and representations, volume 222 of Graduate Texts in Mathematics. Springer, Cham, second edition, 2015. An elementary introduction.
  18. [21]G. A. Hunt. Semi-groups of measures on Lie groups. Trans. Amer. Math. Soc., 81:264–293, 1956.DOI
  19. [23]Yukiyosi Kawada and Kiyosi Itô. On the probability distribution on a compact group. I. Proc. Phys.-Math. Soc. Japan (3), 22:977–998, 1940.DOI
  20. [24]A. A. Kirillov. Lectures on the orbit method, volume 64 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2004.DOI
  21. [25]Victoria Knopova and René L. Schilling. On the Liouville property for non-local Lévy generators. 2019. Preprint, arXiv:1909.01237.arxiv.org/abs/1909.01237
  22. [26]David Lee. Fundamental solutions of the logarithmic Laplacian: an approach via the division problem. Proc. Amer. Math. Soc., 154(8):3479–3494, 2026.arxiv.org/abs/2506.20121
  23. [27]David Lee. A generalized Liouville theorem via division. 2026. Preprint, arXiv:2607.00863.arxiv.org/abs/2607.00863
  24. [28]Ming Liao. Lévy processes and Fourier analysis on compact Lie groups. Ann. Probab., 32(2):1553–1573, 2004.
  25. [29]Ming Liao. Lévy processes in Lie groups, volume 162 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 2004.
  26. [30]Heybetkulu Mustafayev. A note on the Kawada-Itô theorem. Statist. Probab. Lett., 181:Paper No. 109261, 6, 2022.DOI
  27. [31]F. Peter and H. Weyl. Die Vollständigkeit der primitiven Darstellungen einer geschlossenen kontinuierlichen Gruppe. Math. Ann., 97(1):737–755, 1927.
  28. [32]Albert Raugi. A general Choquet-Deny theorem for nilpotent groups. Ann. Inst. H. Poincaré Probab. Statist., 40(6):677–683, 2004. The main result is retracted in effect; see [BF26].
  29. [33]Mark R. Sepanski. Compact Lie groups, volume 235 of Graduate Texts in Mathematics. Springer, New York, 2007.

Paper details

vlnc:W_7na9r3td
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Claude

View statement
Looking for related papers…
Contents