Random Walk Loop Soup Overview

• Random walk loop soup is a Poisson ensemble of discrete loops sampled from Markov chains, capturing loop measures and dynamic evolution on graphs.
• It connects discrete and continuous models by linking random walk loops to Brownian loop soups, loop-erased random walks, and conformal loop ensembles.
• Generalizations such as massive, reinforced, and space-time variants extend its applicability in exploring scaling limits and field-theoretic properties.

Random walk loop soup is a Poissonian ensemble of discrete loops sampled from the natural loop measure of a random walk or, more generally, a killed or transient Markov chain on a graph. In lattice form, it is the discrete analogue of the Brownian loop soup; in probabilistic decompositions, it is the loop content erased by loop-erased random walk; and in field-theoretic formulations, it encodes occupation fields, sign clusters, and interface ensembles through isomorphism theorems and scaling limits. The model appears in rooted, unrooted, ordered, and continuous-time versions, but these formulations are structurally compatible: after forgetting auxiliary data such as roots or orderings, they describe the same underlying loop ensemble (Gu, 2024, Brug et al., 2014, Lupu et al., 2017).

1. Foundational definitions and equivalent constructions

A broad discrete setup starts from a Markov chain on a state space $A$ with transition matrix $P=[p(x,y)]$, often killed on exiting a domain. For the killed chain on $A$, the Green operator is

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$

where $L=I-P$ is the Laplacian. A path or loop $l=[l_0,\dots,l_n]$ has Markov weight

$p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$

In this setting, the intrinsic unrooted loop measure is

$m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$

and its total mass satisfies

$\sum_\ell m(\ell)=\log\det G_A.$

The random walk loop soup is then the Poisson point process with intensity $m$, or more generally $P=[p(x,y)]$0 at time $P=[p(x,y)]$1. Proposition 4.22 of the Markov-chain treatment identifies three equivalent constructions: sampling directly from the unrooted loop measure, sampling a rooted loop soup and forgetting the root, or sampling an ordered loop soup and forgetting the root (Gu, 2024).

The same paper introduces a dynamic “growing loop” formulation. At a vertex $P=[p(x,y)]$2, elementary loops have total mass

$P=[p(x,y)]$3

and the number of concatenated elementary loops evolves as a negative binomial process with rates

$P=[p(x,y)]$4

At time $P=[p(x,y)]$5, the rooted growing loop law is

$P=[p(x,y)]$6

which is exactly the law governing the loops erased in chronological loop erasure (Gu, 2024).

On $P=[p(x,y)]$7, the classical rooted loop measure assigns to a rooted loop of length $P=[p(x,y)]$8 the mass

$P=[p(x,y)]$9

and in two dimensions this becomes $A$0. Continuous-time formulations replace discrete steps by bridges of a Markov jump process; on a graph with conductances $A$1 and killing $A$2, the continuous-time loop measure is

$A$3

and the loop soup of intensity $A$4 is the Poisson point process with intensity $A$5 (Qian, 6 Jan 2026, Lupu et al., 2017).

2. Brownian scaling limits, cluster geometry, and CLE

The continuum comparator is the Brownian loop soup. Lawler–Trujillo Ferreras had shown that, after diffusive scaling, large lattice loops converge to large Brownian loops. Beneš, Černý, and Louidor advanced this from individual loops to cluster geometry: in a bounded simply connected domain $A$6, for $A$7 and $A$8, the collection of outer boundaries of outermost clusters of the cutoff soup $A$9 converges in distribution to the outer boundaries of outermost Brownian loop-soup clusters, hence to $G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$0, where

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$1

Their proof isolates the cluster-level obstacles absent from loop-by-loop convergence: near-touching instability and the approximation of infinite Brownian clusters by finite subclusters. The argument combines no-touching results for Brownian loops and soups, quantitative separation of nonintersecting loops, and topological stability of hulls and outer boundaries under uniform approximation (Brug et al., 2014).

Lupu then removed the microscopic-loop cutoff in the half-plane. On the rescaled discrete half-plane

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$2

and its metric-graph version $G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$3, he proved that for every $G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$4,

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$5

converge in law to

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$6

whose law is $G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$7. The parameter relation is

$G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$8

with $G_A=I+P_A+P_A^2+\cdots=(I-P_A)^{-1}=L_A^{-1},$9. A recurrent normalization confusion is explicitly corrected there: the central charge is not the loop-soup intensity; rather,

$L=I-P$0

The theorem shows that arbitrarily small discrete loops, although able to merge macroscopic clusters, do not change the scaling limit of outermost cluster boundaries (Lupu, 2015).

At the coupling level, Wang showed that Brownian and random walk loop soups can be put on the same probability space uniformly down to all polynomial scales. For every $L=I-P$1, every $L=I-P$2, and every $L=I-P$3, there is a coupling such that, except on an event of probability at most

$L=I-P$4

all rescaled random walk loops of duration $L=I-P$5 in a macroscopic region have Brownian partners. Paired loops satisfy

$L=I-P$6

This removes the earlier $L=I-P$7 restriction and yields control at all polynomially small loop scales (Qian, 6 Jan 2026).

A further universality result replaces the square lattice by suitable planar graphs. Pang proved that if a sequence of planar graph discretizations satisfies an invariance principle to Brownian motion, bounded density, and a Russo–Seymour–Welsh-type crossing estimate, then the associated random walk loop soup converges to Brownian loop soup in a topology on multisets of unrooted, unparameterized macroscopic loops (Pang, 13 Mar 2026).

3. Loop erasure, Wilson’s algorithm, and reconstruction of trajectories

The random walk loop soup is canonically tied to loop-erased random walk. For a self-avoiding path $L=I-P$8 from $L=I-P$9 to $l=[l_0,\dots,l_n]$0, the loop-erased random walk weight is

$l=[l_0,\dots,l_n]$1

Equivalently, if one samples a LERW path $l=[l_0,\dots,l_n]$2 and then conditionally samples independent loops attached at the vertices of $l=[l_0,\dots,l_n]$3 with the growing-loop law, concatenating these loops reconstructs the original stopped Markov chain. In this precise sense, “a random walk equals its loop erasure plus an independent soup of loops grafted back in.” The same philosophy extends to Wilson’s algorithm and the uniform spanning tree: realizations of the original walk can be reconstructed from loop-erased structures together with independent loop soups at time $l=[l_0,\dots,l_n]$4 (Gu, 2024).

This discrete identity persists in scaling limits. Aïdékon, Jego, and collaborators constructed a deterministic chronological attachment map

$l=[l_0,\dots,l_n]$5

and proved that if $l=[l_0,\dots,l_n]$6 is an independent radial $l=[l_0,\dots,l_n]$7 in a simply connected domain $l=[l_0,\dots,l_n]$8, and $l=[l_0,\dots,l_n]$9 is an independent Brownian loop soup of intensity $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$0, then chronologically inserting the loops of $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$1 that intersect $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$2 yields a continuous path $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$3 with

$p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$4

the law of Brownian motion in $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$5 stopped on exiting $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$6. On the lattice side, if $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$7 is LERW in a finite domain and $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$8 is an independent random walk loop soup, then after tie-breaking the object $p(l)=\prod_{j=1}^n p(l_{j-1},l_j).$9 has exactly the law of simple random walk stopped on exiting the domain. The continuum theorem is thus the inverse of loop erasure in the scaling limit (Berestycki et al., 4 Feb 2026).

A related trace-level decomposition was proved by Sapozhnikov and Shiraishi. In dimensions $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$0 and $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$1, if $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$2 is any subsequential scaling limit of loop-erased random walk in the unit ball and $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$3 is an independent Brownian loop soup, then the union of $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$4 with all loops from $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$5 intersecting $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$6 has the same law as Brownian trace. In dimension $m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$7, they further proved that every such subsequential limit is almost surely a simple path, with Hausdorff dimension satisfying

$m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$8

This is the continuum analogue of the discrete identity

$m(\ell)=\frac{J(\ell)\,p(\ell)}{|\ell|},$9

now expressed at the level of random compact sets (Sapozhnikov et al., 2015).

4. Gaussian free fields, metric graphs, and signed isomorphisms

At intensity $\sum_\ell m(\ell)=\log\det G_A.$0, random walk loop soups admit exact Gaussian free field couplings. Le Jan’s occupation-field identity states that for a transient Markov process,

$\sum_\ell m(\ell)=\log\det G_A.$1

where $\sum_\ell m(\ell)=\log\det G_A.$2 is the centered GFF with covariance equal to the Green function. On the metric graph $\sum_\ell m(\ell)=\log\det G_A.$3, Lupu’s signed coupling sharpens this to a geometric statement: there is a coupling between the metric-graph loop soup $\sum_\ell m(\ell)=\log\det G_A.$4 and the metric-graph GFF $\sum_\ell m(\ell)=\log\det G_A.$5 such that

$\sum_\ell m(\ell)=\log\det G_A.$6

the zero set of $\sum_\ell m(\ell)=\log\det G_A.$7 is exactly the set of points not visited by loops, and the loop clusters are exactly the sign clusters of $\sum_\ell m(\ell)=\log\det G_A.$8. Conditional on $\sum_\ell m(\ell)=\log\det G_A.$9, the sign on each connected component is an independent uniform $m$0 variable (Lupu et al., 2017).

In the half-plane convergence theorem, this metric-graph representation is the key input at criticality $m$1. The loop-soup/GFF isomorphism is used through the identity

$m$2

which identifies loop clusters with sign clusters of the metric-graph GFF. Lupu further derived explicit conditional probabilities that two marked points lie in different clusters, expressed in terms of effective conductances, and used excursion measures together with SLE restriction martingales to control continuum connectivity events (Lupu, 2015).

On the discrete graph, the signed coupling is encoded by random open edges. Conditional on $m$3, an edge $m$4 is declared closed with probability

$m$5

Conditional on the GFF $m$6, the corresponding edge configuration satisfies

$m$7

Thus same-sign regions generate FK-type random subgraphs, while opposite-sign edges are deterministically closed (Lupu et al., 2017).

The signed coupling can be inverted. Starting from a signed GFF $m$8, Lupu and collaborators constructed a self-interacting process $m$9 with depleted amplitudes

$P=[p(x,y)]$00

and Poisson edge stacks

$P=[p(x,y)]$01

The time-reversal of this process reconstructs, in law, the path content underlying the Ray–Knight coupling; iterated over vertices and combined with Poisson–Dirichlet splitting, it reconstructs the full random walk loop soup in the exact coupling law of Lupu’s discrete isomorphism. The same construction yields an inversion of the Lupu–Werner coupling between random current and FK-Ising random cluster models (Lupu et al., 2017).

5. Percolation, arm events, vacant sets, and cover phenomena

The loop soup has a sharp percolative phase structure in the discrete half-plane. On

$P=[p(x,y)]$02

the critical intensity for existence of an infinite loop cluster is

$P=[p(x,y)]$03

Non-percolation at $P=[p(x,y)]$04 was known from the GFF coupling; the complementary statement, proved using Brownian loop-soup crossings, robust convergence of loop intersections, and a 1-dependent block percolation construction, is that every $P=[p(x,y)]$05 yields an infinite cluster. Moreover, in the supercritical regime one can delete all loops above a sufficiently large diameter cutoff and still retain percolation (Lupu, 2014).

In dimensions $P=[p(x,y)]$06, the vacant set

$P=[p(x,y)]$07

has long-range dependencies that are not covered by the standard decoupling framework for Bernoulli percolation, random interlacements, or GFF level sets. Chang, Drewitz, and coauthors established a weaker but natural decoupling inequality: if $P=[p(x,y)]$08 depend on boxes of side $P=[p(x,y)]$09 whose centers are at distance $P=[p(x,y)]$10, then for increasing $P=[p(x,y)]$11,

$P=[p(x,y)]$12

with the analogous decreasing-event bound obtained by replacing $P=[p(x,y)]$13 with $P=[p(x,y)]$14. They proved that this weaker inequality still suffices for the full large-scale geometry theory in the strongly supercritical regime. In particular, there exists $P=[p(x,y)]$15 such that for $P=[p(x,y)]$16, the infinite vacant cluster satisfies local uniqueness with stretched-exponential probability, and therefore also the quenched invariance principle, Gaussian heat-kernel bounds, Harnack inequalities, and the local CLT (Alves et al., 2018).

In two dimensions, recent work studies rare connectivity events for loop clusters. For $P=[p(x,y)]$17, the central event is that at least two outermost clusters cross an annulus $P=[p(x,y)]$18. Using separation lemmas for random walks in the presence of loop soup, loop surgery, locality and reversed locality, and an upper quasi-multiplicativity estimate,

$P=[p(x,y)]$19

Abraham and collaborators transferred CLE/Brownian loop-soup arm exponents to the discrete model. The resulting four-arm upper bound is

$P=[p(x,y)]$20

with analogous estimates for interior two-arm, boundary two-arm, and boundary four-arm events (Gao et al., 2024).

Massive variants generate a different set of observables. For the two-dimensional massive random walk loop soup with killing rate $P=[p(x,y)]$21, the cover time $P=[p(x,y)]$22 of a finite set $P=[p(x,y)]$23 satisfies, whenever

$P=[p(x,y)]$24

the asymptotic law

$P=[p(x,y)]$25

where $P=[p(x,y)]$26 is Gumbel distributed and $P=[p(x,y)]$27. This identifies a broad regime in which cover-time fluctuations remain of independent-extremes type despite nontrivial loop correlations (Broman et al., 2024).

6. Massive, space-time, reinforced, and other generalizations

A natural off-critical perturbation introduces killing. If $P=[p(x,y)]$28 is a mass function, the massive random walk loop soup is defined by the killing field

$P=[p(x,y)]$29

so that each loop acquires the weight

$P=[p(x,y)]$30

The massive soup is obtained from the critical soup by independent exponential thinning. Under diffusive scaling on $P=[p(x,y)]$31, a fixed positive mass kills all macroscopic loops, while a mass $P=[p(x,y)]$32 with $P=[p(x,y)]$33 recovers the critical theory. The nontrivial near-critical scaling is

$P=[p(x,y)]$34

and with this choice the massive random walk loop soup converges to the massive Brownian loop soup (Camia, 2013).

A different generalization adds a discrete time torus. For a space-time random walk on $P=[p(x,y)]$35 with generator

$P=[p(x,y)]$36

the projection of the space-time loop measure to $P=[p(x,y)]$37 converges, as $P=[p(x,y)]$38, to the Bosonic loop measure

$P=[p(x,y)]$39

Unlike the standard Markovian loop measure, this assigns mass only to loops whose durations are integer multiples of $P=[p(x,y)]$40. The corresponding occupation-field Laplace transform is

$P=[p(x,y)]$41

and the total mass equals $P=[p(x,y)]$42, which is the logarithm of the grand-canonical partition function of the ideal Bose gas (Adams et al., 2017).

Reinforcement leads away from Poissonian ensembles. In the VRJP/$P=[p(x,y)]$43 setting, the reinforced loop soup is generated by a reinforced Wilson algorithm: a single vertex reinforced jump process alternates between invisible phases inside the current tree and visible phases whose chronological loop erasures are recorded as loops. Quenched on the random environment $P=[p(x,y)]$44, this is just the standard Markovian Wilson algorithm for jump rates $P=[p(x,y)]$45; annealed over $P=[p(x,y)]$46, it produces a genuinely non-Poissonian loop ensemble. Its occupation field $P=[p(x,y)]$47 satisfies the analogue of Le Jan’s theorem,

$P=[p(x,y)]$48

so the reinforced loop soup is the $P=[p(x,y)]$49/VRJP counterpart of the standard Markovian loop soup (Chang et al., 6 Mar 2025).

High-intensity observables exhibit additional limit laws. On a finite graph, if $P=[p(x,y)]$50 is a one-form and $P=[p(x,y)]$51 is the loop soup of intensity $P=[p(x,y)]$52, then

$P=[p(x,y)]$53

converges in distribution to a centered Gaussian with explicitly computable variance

$P=[p(x,y)]$54

For planar graphs this yields a Gaussian limit for the winding field. By contrast, in the Brownian loop soup the total winding around a point of all loops of diameter at least $P=[p(x,y)]$55, scaled by $P=[p(x,y)]$56, converges to a centered Cauchy law with scale parameter $P=[p(x,y)]$57 (Camia et al., 2020).

Taken together, these developments show that random walk loop soup is not a single narrowly defined object but a flexible probabilistic formalism. Its standard Poissonian version links killed Markov chains, LERW, GFFs, CLE, and Brownian motion; its massive, reinforced, and space-time variants connect near-critical scaling, supersymmetric models, Bosonic statistics, and non-Poissonian loop ensembles; and its continuum scaling limits carry explicit conformal field theoretic data, including the Brownian loop-soup central charge relation $P=[p(x,y)]$58 (Camia et al., 2021).