---
title: Lévy Laplacian in Infinite-Dimensional Analysis
url: https://www.emergentmind.com/topics/levy-laplacian
type: topic
---

# Lévy Laplacian in Infinite-Dimensional Analysis

Searching arXiv for recent and foundational papers on the Lévy Laplacian to ground the article.
The Lévy Laplacian is an infinite-dimensional differential operator defined so as to retain a distinguished diagonal or Cesàro-averaged component of a second derivative rather than the full Hilbert-space trace. In its classical form on linear spaces, it is extracted from the “Lévy part” of the Hessian; on manifolds of paths it admits covariant and basis/Cesàro formulations; in stochastic analysis and white-noise settings several nonequivalent operators occur under the same name. Its importance in geometric analysis derives from precise correspondences with Yang–Mills equations, Yang–Mills heat flow, and, in modified four-dimensional settings, instanton and anti-instanton equations [1905.01223], [2205.14351], [1709.09221], [2507.13013].

## 1. Classical operator and the Lévy part of the second derivative

The classical starting point is a twice Fréchet differentiable function \(f\) on an infinite-dimensional linear space such as \(L_2([0,1],\mathbb R)\), whose second derivative is decomposed into a nonlocal Volterra term and a diagonal Lévy term. In the form recalled in several later treatments, one writes
\[
\langle f''(x)u,v\rangle
=
\int\!\!\int K_V(x;t,s)u(t)v(s)\,dt\,ds
+
\int K_L(x;t)u(t)v(t)\,dt,
\]
and then defines
\[
\Delta_L f(x)=\int_0^1 K_L(x;t)\,dt.
\]
An equivalent-looking but conceptually distinct representation is the basis/Cesàro form
\[
\Delta_L^{\{e_n\}} f(x)
=
\lim_{n\to\infty}\frac1n\sum_{k=1}^n
\langle f''(x)e_k,e_k\rangle,
\]
for a suitable orthonormal basis \(\{e_n\}\) [1905.01223], [2507.13013].

The relation between these two descriptions is controlled by the class of bases used. For weakly uniformly dense orthonormal bases, the Cesàro average isolates precisely the Lévy part of the second derivative. One formulation is
\[
\lim_{n\to\infty}\int_0^1 h(t)\left(\frac1n\sum_{k=1}^n e_k(t)^2-1\right)\,dt=0
\]
for every \(h\in L_\infty([0,1],\mathbb R)\). The sine basis \(e_n(t)=\sqrt2\sin(n\pi t)\) is a standard example; Sturm–Liouville eigenfunctions with Dirichlet boundary conditions also appear as examples in the manifold setting [2205.14351], [1709.09221].

A recurrent point in the literature is that “the Lévy Laplacian” is not a unique object unless the ambient space, trace prescription, and basis class are fixed. One paper develops a one-parameter chain
\[
\Delta_L^{\{e_n\},s} f(x)
=
\lim_{n\to\infty}\frac1n \sum_{k=1}^n\sum_{\mu=1}^d
k^{\,1-s}\,
\left\langle f''(x)p_\mu e_k,p_\mu e_k\right\rangle,
\]
with the classical Lévy Laplacian identified as the order \(s=1\) element, while the order \(s=-1\) operator later becomes central in stochastic/Hida comparisons [1709.09221]. This establishes that the classical operator is only one member of a broader hierarchy of Lévy-type traces.

## 2. Covariant formulation on manifolds of paths

A major geometric development is the intrinsic definition of the Lévy Laplacian on the Hilbert manifold of \(H^1\)-paths in a Riemannian manifold. In one formulation, the path space is
\[
\Omega=\{\gamma:[0,1]\to M \text{ of Sobolev class } H^1\},
\]
or \(\mathcal{P}ath\) in later notation, with submanifolds such as \(\Omega_x\), \(\Omega_{x,x}\), \(\mathcal{P}ath_m\), and \(\mathcal{L}oop_m\). The tangent fiber at \(\gamma\) is \(H^1_\gamma(TM)\), while an auxiliary bundle \(\mathcal H^0\) has fiber \(H^0_\gamma(TM)\), the \(L_2\)-vector fields along \(\gamma\), equipped with
\[
G_0(X(\gamma),Y(\gamma))
=
\int_0^1 g(X(\gamma;t),Y(\gamma;t))\,dt.
\]
The Levi-Civita connection on \(M\) induces a canonical connection on \(\mathcal H^0\) by
\[
\nabla^{\mathcal H^0}_Y X(\gamma;\tau)
=
d_YX(\gamma;\tau)+\Gamma(\gamma(\tau))(X(\gamma;\tau),Y(\gamma;\tau)).
\]
This induced geometry is the basis of the covariant theory [1905.01223], [2507.13013].

The intrinsic operator is defined as
\[
\Delta_L\varphi=\operatorname{div}_L(\operatorname{grad}_{H^0}\varphi),
\]
or in the later notation,
\[
\Delta_L\varphi=\mathrm{div}_L(\mathrm{grad}_{H^0}\varphi).
\]
Its crucial input is an AGV-type decomposition of the covariant derivative of the \(H^0\)-gradient. For tangent fields \(X,Y\), the relevant bilinear form has the structure
\[
\begin{aligned}
K(\gamma)\langle X,Y\rangle
&=
\int_0^1\!\!\int_0^1 K^V(\gamma;\tau_1,\tau_2)
\langle X(\gamma;\tau_1),Y(\gamma;\tau_2)\rangle\,d\tau_1d\tau_2 \\
&\quad+
\int_0^1 K^L(\gamma;\tau)
\langle X(\gamma;\tau),Y(\gamma;\tau)\rangle\,d\tau \\
&\quad+
\frac12\int_0^1 K^S(\gamma;\tau)\langle \nabla X(\gamma;\tau),Y(\gamma;\tau)\rangle\,d\tau \\
&\quad+
\frac12\int_0^1 K^S(\gamma;\tau)\langle \nabla Y(\gamma;\tau),X(\gamma;\tau)\rangle\,d\tau .
\end{aligned}
\]
Here \(K^V\), \(K^L\), and \(K^S\) are the Volterra, Lévy, and singular kernels. The divergence then extracts only the Lévy kernel:
\[
\operatorname{div}_L\psi(\gamma)
=
\int_0^1 \operatorname{tr}_g(K^L_\psi(\gamma;\tau))\,d\tau.
\]
This makes explicit that the operator is not the ordinary Hilbert trace of a Hessian, but a Lévy-type trace that discards the off-diagonal Volterra contribution and the singular term [1905.01223], [2507.13013].

A separate but equivalent presentation arises on \(\mathcal{P}ath_m\), where Levi-Civita parallel transport trivializes the tangent bundle. There one defines an AGV Lévy trace \(\mathrm{tr}^{AGV}_L\) of the parallelized second derivative \(\widetilde D^2\varphi\), and also a direct basis/Cesàro operator
\[
\Delta^{\{e_{\mu,n}\}_L}\varphi(\gamma)=
\lim_{n\to\infty}\frac 1n\sum_{k=1}^n\sum_{\mu=1}^d
\left.\frac {d^2}{ds^2}\right|_{s=0}
\varphi\left(\mathrm{Exp}_\gamma(s\widetilde{e_{\mu,k}})\right).
\]
A later paper proves that, on their common domain, the covariant definition, the AGV trace definition, and the Cesàro definition coincide on the manifold of \(H^1\)-paths [2507.13013]. This resolves a substantial part of the basis-versus-intrinsic ambiguity in the geometric setting.

## 3. Parallel transport, Yang–Mills equations, and heat flow

The most developed geometric application concerns parallel transport of a connection. Let \(E\to M\) be a complex vector bundle with structure group \(G\subset SU(N)\), and let \(A=A_\mu dx^\mu\) be a connection with curvature
\[
F_{\mu\nu}
=
\partial_\mu A_\nu-\partial_\nu A_\mu+[A_\mu,A_\nu].
\]
The Yang–Mills equations are
\[
\nabla^\mu F_{\mu\nu}=0,
\]
and the Yang–Mills heat equation for a time-dependent connection is
\[
\partial_s A_\nu(s,x)=\nabla^\mu F_{\mu\nu}(s,x).
\]
If \(U_{t,s}(\gamma)\) denotes parallel transport along \(\gamma\), then the first path-space variation formula gives
\[
d_X U_{d,c}(\gamma)
=
-\int_c^d U_{d,t}(\gamma)F_{\mu\nu}(\gamma(t))X^\mu(\gamma;t)\dot\gamma^\nu(t)U_{t,c}(\gamma)\,dt
+\text{endpoint terms},
\]
and for the full transport one obtains
\[
\operatorname{grad}_{H^0}U_{1,0}(\gamma;t)^\mu
=
-
U_{1,t}(\gamma)F^\mu{}_\nu(\gamma(t))\dot\gamma^\nu(t)U_{t,0}(\gamma).
\]
Thus curvature appears already at first order in the path variable [1905.01223].

The central formula for the covariant Lévy Laplacian is
\[
\Delta_L U_{1,0}(\gamma)
=
-\int_0^1
U_{1,t}(\gamma)\,\nabla^\mu F_{\mu\nu}(\gamma(t))\,\dot\gamma^\nu(t)\,U_{t,0}(\gamma)\,dt.
\]
Hence the Lévy Laplacian of parallel transport is exactly the pathwise insertion of the Yang–Mills operator \(\nabla^\mu F_{\mu\nu}\) [1905.01223]. A closely related Euclidean/Minkowski path-space calculation yields, for the Lévy d’Alembertian or Laplacian in the older path-space formalism,
\[
\Box_L U^A_{1,0}(\sigma)
=
-\int_0^1
U^A_{1,t}(\sigma)\,\eta^{\mu\nu}\nabla_\mu F_{\nu\lambda}(\sigma(t))
\dot\sigma^\lambda(t)\,U^A_{t,0}(\sigma)\,dt,
\]
and similarly in the Euclidean case with \(\delta^{\mu\nu}\) [1612.00310].

This yields a path-space characterization of gauge equations. In the manifold/covariant theory, the main equivalence is
\[
\partial_s A_\nu=\nabla^\mu F_{\mu\nu}
\quad\Longleftrightarrow\quad
\partial_s U_{1,0}(s,\gamma)=\Delta_LU_{1,0}(s,\gamma).
\]
Thus the nonlinear Yang–Mills heat flow on the finite-dimensional bundle is equivalent to a linear heat equation for the Lévy Laplacian on the associated path-space parallel transport [1905.01223]. A later paper rederives this correspondence in a broader theory of path-space Lévy heat equations and uses it as one of the principal motivations for the operator [2507.13013].

A common misconception is that every Lévy-Laplace equation on path space is automatically equivalent to Yang–Mills. The deterministic path-space correspondence does hold in the geometric formulations just described, but stochastic analogues show that this equivalence is definition-sensitive. In the stochastic Cesàro/Malliavin setting, the Lévy Laplacian of stochastic parallel transport contains an additional curvature-square term, so the equation \(\Delta_LU=0\) is generally stronger than Yang–Mills [1605.06024].

## 4. Modified, stochastic, and white-noise variants

One important extension is the modified Lévy Laplacian on a path manifold \(\Omega_m\), where tangent directions are rotated by a time-dependent curve \(W\in C^1([0,1],SO(d))\). The operator is defined by
\[
\Delta_L^{W,\{e_{\mu,n}\}}f(\gamma)=
\lim_{n\to\infty}\frac 1n\sum_{k=1}^n\sum_{\mu=1}^d
\left.\frac {d^2}{ds^2}\right|_{s=0}
f\bigl(\mathrm{exp}_\gamma(s\,W(\widetilde{e_{\mu,k}}(\gamma)))\bigr).
\]
For constant \(W\), the second term in the corresponding parallel-transport formula vanishes and one recovers the ordinary Lévy-Laplacian relation to Yang–Mills. For nonconstant \(W\), the infinitesimal generator
\[
L_W(t)=W^{-1}(t)\dot W(t)\in\mathfrak{so}(4)
\]
couples to curvature. In dimension \(4\), the decomposition
\[
\mathfrak{so}(4)=\mathrm{Lie}(S^3_L)\oplus \mathrm{Lie}(S^3_R)
\]
allows the modified operator to detect self-dual or anti-self-dual curvature components. Under the stated subgroup and nondegeneracy assumptions, together with the vanishing-along-a-sequence condition on \(F_+\) or \(F_-\), the equation
\[
\Delta_L^{W,\{e_{\mu,n}\}}U^A_{1,0}=0
\]
is equivalent to anti-self-duality or self-duality, respectively [2205.14351]. This gives a path-space characterization of instantons and anti-instantons rather than merely Yang–Mills fields.

In stochastic analysis, another major divergence of definitions appears. A Malliavin-calculus Lévy Laplacian is defined on \(W^{2,2}(P,M_N(\mathbb C))\) by
\[
\Delta_L f(b)
=
\lim_{n\to\infty}\frac1n \sum_{k=1}^n\sum_{\mu=1}^d
\partial_{p_\mu h_k}\partial_{p_\mu h_k}f(b),
\]
with strong \(L_2\)-convergence over Wiener measure. For stochastic parallel transport \(U^A(b,1)\), one has
\[
\Delta_L U^A(b,1)
=
U^A(b,1)\int_0^1 U^A(b,t)^{-1}F_{\mu\nu}(b_t)F^\mu{}_\nu(b_t)U^A(b,t)\,dt
-
U^A(b,1)\int_0^1 U^A(b,t)^{-1}\nabla^\mu F_{\mu\nu}(b_t)U^A(b,t)\,db_t^\nu.
\]
The first term has no analogue in the deterministic path-space formula, and it is exactly why the stochastic Lévy-Laplace equation is not equivalent to Yang–Mills in this model [1709.09221], [1605.06024].

The white-noise/Hida picture introduces yet another distinction. One paper shows that the Malliavin operator corresponds, under the canonical embedding \(\mathcal J\) of Wiener Sobolev functionals into Hida generalized functionals, not to the classical order-\(1\) Hida Lévy Laplacian but to the nonclassical order-\((-1)\) operator:
\[
\mathcal J \Delta_L \Psi
=
\pi^2\,\widetilde\Delta_L^{\{l_n\},-1}\mathcal J\Psi.
\]
At the same time, the classical order-\(1\) Hida Lévy Laplacian vanishes on square-integrable Hida functionals for weakly uniformly dense bases [1709.09221]. This is an explicit resolution of a naming ambiguity: identical terminology in Malliavin and Hida calculi need not refer to the same operator.

A different generalization acts on square roots of measures on infinite-dimensional spaces. In the Wiener-space specialization, the Lévy Laplacian is defined on root measures by
\[
\Delta_L f
=
\lim_{n\to\infty}\frac1n\sum_{k=0}^n \partial_{e_k}^2 f,
\]
and its Fourier multiplier is shown to be the quadratic variation:
\[
\mathcal F(\Delta_L f)
=
-(2\pi)^2 \frac{\langle \phi\rangle_T}{T}\times \hat f.
\]
Here \(\langle\phi\rangle_T\) is the quadratic variation of the path \(\phi\) on \([0,T]\). This identifies the symbol of the Lévy Laplacian with a pathwise stochastic quantity rather than a deterministic curvature operator [1212.4205].

## 5. Differential forms, heat equations, and long-time behavior on path space

A recent path-manifold treatment studies the heat equation
\[
\partial_t F=\Delta_LF
\]
for the covariant/path-manifold Lévy Laplacian and constructs explicit solutions from finite-dimensional heat flows of functions and \(1\)-forms on a compact Riemannian manifold [2507.13013]. The simplest functionals are obtained from a smooth function \(\mathfrak f\) by
\[
\mathfrak L_f(\gamma)=\int_0^1 \mathfrak f(\gamma(\tau))\,d\tau,
\]
for which
\[
\Delta_L\mathfrak L_f(\gamma)=\int_0^1 \Delta \mathfrak f(\gamma(\tau))\,d\tau.
\]
This is the pathwise lift of the Laplace–Beltrami operator [1905.01223], [2507.13013].

For a \(1\)-form \(a\in\Omega^1(M)\), one defines
\[
\Theta_a(\gamma)=\int_\gamma a=\int_0^1 a(\gamma(\tau))\dot\gamma(\tau)\,d\tau.
\]
Then
\[
\Delta_L\Theta_a(\gamma)=-\int_\gamma \delta\,da,
\]
and on loop space,
\[
\Delta_L\Theta_a(\gamma)=\int_\gamma \Delta a.
\]
For the abelian parallel transport functional
\[
U^a(\gamma)=e^{-i\Theta_a(\gamma)},
\]
one correspondingly has
\[
\Delta_LU^a(\gamma)=-iU^a(\gamma)\int_\gamma \delta\,da
\]
on path space and
\[
\Delta_LU^a(\gamma)=ie^{-i\Theta_a(\gamma)}\int_\gamma \Delta a
\]
on loop space [2507.13013]. These identities make the \(U(1)\) case a direct model of the nonabelian Yang–Mills correspondence.

The same paper exploits the unusual chain rule
\[
\Delta_L(\varphi_1\varphi_2)=\varphi_1\Delta_L\varphi_2+\varphi_2\Delta_L\varphi_1,
\]
and more generally
\[
\Delta_L(\mathcal F\circ\Phi)
=
(\nabla\mathcal F(\Phi),\Delta_L\Phi)_{\mathbb R^N},
\]
with no second-derivative term of \(\mathcal F\) [2507.13013]. This first-order-like behavior distinguishes the Lévy Laplacian sharply from ordinary finite-dimensional Laplacians and explains why composite path functionals built from heat flows of \(0\)-forms and \(1\)-forms again solve the Lévy heat equation.

The long-time behavior is also governed by finite-dimensional Hodge theory. If \(M\) is compact, orientable, and without boundary, then heat flows of functions converge to harmonic functions, hence to constants on a compact connected manifold, while heat flows of \(1\)-forms converge to harmonic \(1\)-forms. Consequently, functionals built from \(\mathfrak L_f\) and \(\Theta_a\) converge pointwise on loop space to locally constant functionals, taking the same value on homotopy equivalent curves [2507.13013]. This suggests a projection of this class of Lévy heat solutions onto a topological sector determined by harmonic \(1\)-forms and loop homotopy classes.

## 6. Lévy differential operators, basis dependence, and conceptual issues

The Lévy Laplacian is often best viewed as one member of a larger family of Lévy differential operators. In one abstract formulation, given a trace-type functional \(S\) on multilinear maps, one sets
\[
D_{n,S}f(x)=S(f^{(n)}(x)).
\]
The Lévy Laplacian is then \(D_{2,\operatorname{tr}_L^\delta}\), the Lévy d’Alembertian is \(D_{2,\operatorname{tr}_L^\eta}\), and the Lévy divergence is \(D_{1,\operatorname{tr}_L^g}\) [1612.00310]. This framework is used to derive path-space systems equivalent not only to Yang–Mills equations but also to Yang–Mills–Higgs and Yang–Mills–Dirac equations, with parallel transport playing the role of an infinite-dimensional chiral field [1612.00310].

In that setting, a path-space \(1\)-form
\[
B^A(\sigma)u=U^A_{0,1}(\sigma)\,\partial_u U^A_{1,0}(\sigma)
\]
satisfies
\[
\operatorname{div}_L^\eta B^A(\sigma)
=
U^A_{0,1}(\sigma)\,\Box_L U^A_{1,0}(\sigma),
\]
and a closedness condition of Maurer–Cartan type. The resulting system is equivalent to the Yang–Mills equations, and analogous systems with endpoint derivatives encode Yang–Mills–Higgs and Yang–Mills–Dirac fields [1612.00310]. This broadens the role of the Lévy Laplacian from a single trace operator to part of a full path-space differential calculus.

Several conceptual cautions are standard across the literature. First, basis dependence is real at the raw Cesàro-definition level, though weakly uniformly dense and uniformly bounded bases frequently recover the intrinsic Lévy trace on the natural domain [1709.09221], [1612.00310], [2507.13013]. Second, stochastic and deterministic operators with the same name need not coincide, and may have different gauge-theoretic content [1709.09221], [1605.06024]. Third, in manifold settings the operator is intrinsically tied to the Levi-Civita connection, parallel transport, and curvature; it is not a generic Hilbert-manifold Laplacian of ordinary full-trace type [1905.01223], [2507.13013].

Taken together, these developments define the Lévy Laplacian not as a single universally fixed operator, but as a class of infinite-dimensional trace constructions centered on one principle: only a special diagonal or asymptotically diagonal part of the second derivative is retained. In deterministic path geometry this principle isolates the Yang–Mills operator; in modified four-dimensional settings it detects self-duality; in Malliavin and Hida calculi it separates inequivalent stochastic operators; and in root-measure analysis it produces quadratic variation as Fourier symbol [1905.01223], [2205.14351], [1709.09221], [1212.4205].

Source: https://www.emergentmind.com/topics/levy-laplacian