Le Jan Isomorphism: Loop Soup & GFF

• Le Jan isomorphism is a probabilistic equivalence relating loop soup occupation fields of Markov processes with squared Gaussian free fields and 1-permanental vectors.
• It uses analytic techniques like Laplace transforms and determinant formulas to derive explicit density formulations in both discrete and continuous settings.
• The isomorphism underpins practical applications in random sampling, uniform spanning trees, cover time bounds, and comparison inequalities for permanental processes.

The Le Jan isomorphism establishes a precise distributional equivalence between the occupation field of a Markovian loop soup and the squared Gaussian free field (GFF) or, in a more general and non-reversible setting, a permanental vector. This isomorphism provides a probabilistic and analytic bridge connecting random walk local times, loop measures, permanental processes, and the structure of the GFF on discrete (and by extension continuous) spaces. Originally formulated for reversible Markov processes, Le Jan’s isomorphism has since been extended to encompass non-reversible chains, with the occupation field explicitly realized as a 1-permanental process governed by the Green’s function of the generator.

1. Definitions, Notation, and Key Objects

Let $A$ (or $X$) denote a finite state space, $Q=(Q(x,y))$ an edge-weight or generator matrix (which may be symmetric or not), and $\Delta = I - Q$, with the associated Green’s function $G = \Delta^{-1}$ for discrete time or $G = -Q^{-1}$ for continuous time. The following elements are central:

• Rooted loop: A sequence $\omega=[\omega_0, \ldots, \omega_k]$, with $\omega_k = \omega_0$, representing a closed path.
• Rooted loop measure: $\mu^{\text{root}}(\omega) = Q(\omega)/k$ where $$Q(\omega) = \prod_{j=1}^k Q(\omega_{j-1}, \omega_j)$$.
• Unrooted loop measure: Orbits under cyclic re-rooting, with induced measure $$\mu(\ell) = \sum_{\omega\in\ell}\mu^{\text{root}}(\omega)$$.
• Loop soup: For intensity $t > 0$, a Poisson point process on unrooted loops with intensity $t\,\mu$, i.e., the random multiset $\mathcal{L}_t = \sum_\ell N_\ell \delta_\ell$ with independent $N_\ell \sim \text{Poisson}(t\,\mu(\ell))$.
• Occupation field/local times: For each state $x$, the cumulative local time from the loops at $x$, $\mathcal{L}_t(x) = \sum_\ell N_\ell\,\ell_x(\ell)$.
• Gaussian free field (GFF): Vector $(\phi_x)_{x\in A}$ with mean zero and covariance $G$, i.e., $\mathbb{E}[\phi_x\phi_y]=G(x,y)$.
• 1-permanental vector: A nonnegative random vector $(\ell_x)$ with joint Laplace transform $$\mathbb{E}[\exp(-\sum_x \lambda_x \ell_x)] = |\mathbf{I} + \text{diag}(\lambda) G|^{-1}$$.

Further, a continuous-time correction incorporates trivial loops (i.e., exponential holding times at each vertex) by superposing independent Gamma processes of parameter $t$ at each vertex, resulting in increments that are independent $\Gamma(t,1)$ random variables.

2. Statement of Le Jan’s Isomorphism (Reversible and Non-reversible Forms)

Reversible Setting

Let $Q=Q^T$ be an acceptable symmetric edge-weight matrix with spectral radius $\rho(|Q|) < 1$. The GFF $(\phi_x)$ with covariance $G$ and the loop-soup occupation field at intensity $t = 1/2$ satisfy

$$\{\phi_x^2 : x \in A\} \overset{d}{=} \{\mathcal{L}_{1/2}(x): x \in A\}$$

Moreover, the Laplace transforms are related by

$$\mathbb{E}\exp\left\{-\frac{1}{2}\sum_x f(x)\phi_x^2 \right\} = \mathbb{E}\exp\left\{-\sum_x f(x)\mathcal{L}_{1/2}(x)\right\} = \det(I + G D_f)^{-1/2}$$

where $D_f$ is the diagonal matrix with entries $f(x)$ (Lawler et al., 2014).

Non-reversible Generalization

For an irreducible (generally non-symmetric) generator $Q$, with potential kernel $G = -Q^{-1}$ (assumed to have a positive-definite symmetric part), the nonnegative occupation field $\ell = (\ell_i)$ of the loop soup is a 1-permanental vector with Laplace transform $$\mathbb{E}\left[\exp(-\sum_i \lambda_i \ell_i)\right] = |\mathbf{I} + \text{diag}(\lambda) G|^{-1}$$. Le Jan’s isomorphism asserts that for any subexponential function $F$,

$$\mathbb{E}[F(\ell)] = C_n \int_{\mathbb{C}^n} F(|\varphi|^2) \exp\{-(\varphi, Q \varphi)\} d\varphi d\bar\varphi$$

with normalization constant $C_n = (2\pi i)^{-n}|G|^{-1}$ (Qinghua et al., 17 Mar 2025).

3. Derivation and Density Formulas

The derivation proceeds via an explicit calculation of Laplace transforms or density formulas for the occupation field:

• For the loop-soup occupation field, the Laplace transform is given by

$$\mathbb{E}\exp\bigl\{ -\sum_x f(x)\mathcal{L}_t(x) \bigr\} = \det(I + G D_f)^{-t}$$

after incorporating the continuous-time correction (Lawler et al., 2014).

• For non-reversible chains, the law of the 1-permanental vector admits a density on $\mathbb{R}_+^n$ (with respect to Lebesgue measure) as a twisted Gaussian integral:

$$p(\ell) = \frac{1}{(2\pi)^n |G|} \int_{[0,2\pi]^n} \exp\Bigl\{ -\sum_{i,j} Q_{ij} \sqrt{\ell_i\ell_j}\, e^{i(\theta_j - \theta_i)} \Bigr\} d\theta_1\cdots d\theta_n$$

Equivalently, via change of variables $\varphi_i = \sqrt{\ell_i}e^{i\theta_i}$,

$$p(\ell)\,d\ell = C_n\,\exp\{-(\varphi, Q\varphi)\} d\varphi d\bar\varphi \quad \text{subject to } |\varphi_i|^2 = \ell_i$$

(Qinghua et al., 17 Mar 2025).

A crucial feature in the non-reversible case is that the twisted complex Gaussian density $\exp\{-(\varphi,Q\varphi)\}$ replaces the standard real GFF density, yielding a “twisted” measure rather than a decoupled real structure.

4. Extensions and Related Isomorphism Theorems

The Le Jan isomorphism encompasses several fundamental probabilistic identities, many of which generalize longstanding classical isomorphisms:

• Non-reversible Dynkin isomorphism: For a killed chain (with additional killing rates $h_i>0$), the occupation field relates to the local time process of the killed Markov chain (Qinghua et al., 17 Mar 2025).
• Second Ray–Knight Theorem: The remaining local times after spending a specified local time at a root are equal in law to the permanental field conditioned on its value at the root.
• Eisenbaum’s isomorphism: An extension involves functionals of the occupation field shifted by local times at independently exponentially distributed random times.

For complex or Hermitian transition matrices, the isomorphism continues to hold for complex GFFs, with the squared moduli $\{|h(x)|^2\}$ equidistributed with the loop-soup occupation field at $t=1/2$ via analytic continuation (Lawler et al., 2014).

5. Applications in Probability, Statistical Physics, and Combinatorics

Le Jan’s isomorphism underpins several notable results and algorithms:

• Uniform Spanning Trees and Wilson’s Algorithm: The loop measure precisely matches the determinant factors in Kirchhoff’s matrix-tree theorem, explaining the correctness of Wilson’s algorithm in generating uniform spanning trees (Lawler et al., 2014).
• Bounds on Cover Time for Markov Chains: The isomorphism enables comparison inequalities for non-reversible Markov chains, bounding cover times via interpolation between kernels and the use of symmetrized chains. Notably, the expected cover time of a non-reversible chain can be bounded in terms of the cover time for the symmetrized chain and the determinant ratio $y = |\det(G)|/|\det(\frac{1}{2}(G+G^T))|$, with $t_{\text{cov}} = O(t'_{\text{cov}} \log y)$ (Qinghua et al., 17 Mar 2025).
• Inequalities for Permanental Processes: The isomorphism provides a framework for monotonicity and comparison inequalities, such as Kahane-type and Slepian-type inequalities, across families of kernels (Qinghua et al., 17 Mar 2025).

6. Significance and Generalizations

Le Jan isomorphism situates loop soups, GFFs, and permanental fields within a common analytic framework, revealing their deep structural equivalence via Laplace and density transforms. Its fundamental role in establishing connections between Markovian path structures and Gaussian processes has led to advances in the analysis of cover times, algorithmic random sampling, and the study of random fields in both discrete and continuous settings.

The extension to non-reversible processes broadens the isomorphism’s reach to general generators, with the twisted complex Gaussian density as a unifying measure. Such extensions enable the recovery of the entire suite of probabilistic isomorphisms—Ray–Knight, Dynkin, Eisenbaum—across symmetric and non-symmetric models (Qinghua et al., 17 Mar 2025, Lawler et al., 2014).

The Le Jan isomorphism thus serves as a cornerstone of modern probabilistic analysis of Markovian structures, random permutations, and field theories, with ongoing developments in both mathematical theory and algorithmic applications.