Brownian Loop Soup

• Brownian loop soup is a conformally invariant Poisson process of Brownian loops in planar domains, capturing the scaling limits of discrete loop models and conformal loop ensembles.
• It features a well-defined loop measure with restriction properties and undergoes a phase transition in cluster formation driven by its intensity parameters.
• Couplings with the Gaussian free field and multiplicative chaos yield exact correlators and conformal fields, linking probabilistic models with conformal field theory observables.

Brownian loop soup is a conformally invariant Poisson point process of Brownian loops in a planar domain, introduced as a continuum loop ensemble that simultaneously encodes scaling limits of lattice loop gases, the geometry of conformal loop ensembles, and a substantial sector of two-dimensional conformal field theory. In a domain $D\subset \mathbb C$, it is defined from the Brownian loop measure restricted to loops staying in $D$, with an intensity parameter usually denoted $\lambda$ in CFT-oriented work and $c$ in the CLE/cluster literature; in the former notation, the associated central charge is $c=2\lambda$ (Camia, 2015, Camia et al., 2015). The object is characterized by conformal invariance, restriction to subdomains, infinitely many microscopic loops but only finitely many loops above any fixed diameter threshold in bounded domains, and a cluster structure whose outer boundaries realize $\mathrm{CLE}_\kappa$ in the subcritical and critical regime (Sheffield et al., 2010, Camia, 2015).

1. Definition, loop measure, and conformal structure

The Brownian loop soup is built from the Brownian loop measure. One standard rooted form is

$$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$

where $\mu^{br}_{z,t}$ is the Brownian bridge measure on loops rooted at $z$ with duration $t$; the unrooted measure is obtained by forgetting the root, equivalently by quotienting rooted loops by time-shifts along the loop (Camia, 2015). A Brownian loop soup in $D$0 with intensity $D$1 is then a Poisson realization from $D$2, where $D$3 is the restriction of the unrooted loop measure to loops staying in $D$4 (Camia, 2015).

Two structural properties are basic. First, the loop measure is conformally invariant: if $D$5 is conformal, then $D$6 for sets $D$7 of loops in $D$8. Second, it satisfies restriction: if $D$9, then $\lambda$0 is the restriction of $\lambda$1 to loops lying entirely in $\lambda$2. These immediately imply that the critical Brownian loop soup is conformally invariant and that restricting a soup to a subdomain yields a fresh soup with the same intensity in that subdomain (Camia, 2015, Sheffield et al., 2010).

The short-loop divergence is intrinsic. A realization in a bounded domain contains infinitely many loops, but only finitely many loops of diameter at least $\lambda$3. This separation between ultraviolet divergence and macroscopic local finiteness underlies both the renormalization of observables and the loop-soup topology used in scaling-limit results (Camia, 2015, Pang, 13 Mar 2026).

An off-critical deformation is the massive Brownian loop soup. If $\lambda$4 is a mass function, the massive loop measure is

$\lambda$5

The resulting soup is conformally covariant rather than invariant: under a conformal map $\lambda$6, the transformed mass satisfies $\lambda$7 with $\lambda$8 (Camia, 2013, Camia, 2015).

2. Discrete origins and scaling limits

A principal reason for the centrality of Brownian loop soup is that it is the continuum scaling limit of random walk loop soup. On $\lambda$9, the random walk loop soup is itself a Poisson point process of lattice loops with intensity $c$0 and loop weights derived from the random walk loop measure. Under diffusive rescaling,

$c$1

the critical random walk loop soup converges to Brownian loop soup, with a coupling that matches loops of duration at least $c$2 and controls both duration and trace: $c$3 (Camia, 2015).

Recent work strengthens this picture in two directions. First, universality has been established for random walk loop soups on suitable planar graphs, not only on $c$4. If a family of embedded planar graphs satisfies an invariance principle, bounded density, and a Russo–Seymour–Welsh type crossing estimate, then for every bounded simply connected $c$5,

$c$6

in a topology on multisets of unrooted, unparameterized, macroscopic loops (Pang, 13 Mar 2026). Second, the classical lattice-to-continuum coupling has been extended to all polynomial scales. In every dimension $c$7, one can couple Brownian and random walk loop soups so that for every $c$8, corresponding loops above time scale $c$9 are matched one-to-one with

$c=2\lambda$0

up to probability $c=2\lambda$1 for arbitrary $c=2\lambda$2 (Qian, 6 Jan 2026).

These results place Brownian loop soup in the same role for loop gases that Brownian motion occupies for random walks: it is the canonical continuum object, but with sufficiently strong couplings to retain mesoscopic and macroscopic geometry.

3. Cluster geometry, phase transition, and CLE correspondence

Loops in a Brownian loop soup intersect and therefore form clusters. In the cluster literature, two loops are adjacent if they intersect, and a cluster is a connected component for this adjacency relation. For an outermost cluster, one considers the boundary of its filling; the collection of these outer boundaries is the principal conformal-geometric observable (Sheffield et al., 2010).

The planar model has a sharp phase transition. In the normalization used by Sheffield and Werner, if $c=2\lambda$3, the outer boundaries of outermost loop-soup clusters form a random countable collection of disjoint simple loops satisfying conformal restriction, whereas if $c=2\lambda$4, there is almost surely only one cluster (Sheffield et al., 2010). These simple loop boundaries are exactly $c=2\lambda$5 loops, with

$c=2\lambda$6

In particular, the critical value $c=2\lambda$7 corresponds to $c=2\lambda$8 (Sheffield et al., 2010).

In the CFT-oriented normalization, the same CLE relation is written as

$c=2\lambda$9

and the connectivity threshold is stated at $\mathrm{CLE}_\kappa$0: if $\mathrm{CLE}_\kappa$1, the soup breaks into disjoint clusters, whereas if $\mathrm{CLE}_\kappa$2, there is a unique cluster (Camia, 2015). The two notational schemes are compatible with the relation $\mathrm{CLE}_\kappa$3 used elsewhere in the same literature (Camia et al., 2015).

The cluster geometry retains substantial rigidity under conditioning. If one conditions on a portion of the outer boundary of an outermost cluster discovered by a Markovian exploration, then after conformal normalization the loops touching the discovered boundary piece and the remaining loops decouple: the untouched part is again an independent Brownian loop soup of the same intensity, while the touching part satisfies one-sided chordal conformal restriction, more precisely chordal interior restriction, with exponent

$\mathrm{CLE}_\kappa$4

This decomposition yields a further phase transition at $\mathrm{CLE}_\kappa$5 for the connectivity of the boundary-touching subfamily of loops (Qian, 2016).

4. Conditioned clusters, excursions, and critical boundary regularity

At criticality, the conditional structure of a cluster relative to its outer boundary becomes especially explicit. If one conditions a critical loop-soup cluster on its outer boundary $\mathrm{CLE}_\kappa$6, which is a $\mathrm{CLE}_\kappa$7-type loop, then the loops inside the enclosed domain split into two independent parts: loops that stay strictly inside, forming an independent Brownian loop soup in the interior domain, and loops that touch $\mathrm{CLE}_\kappa$8. The union of excursions away from $\mathrm{CLE}_\kappa$9 generated by the boundary-touching loops is distributed like the union of excursions of a Poisson point process of Brownian excursions of intensity $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$0 in the enclosed domain (Qian et al., 2015).

This decomposition is intertwined with three classical couplings: the coupling of critical loop soup with $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$1 via cluster boundaries, the coupling of $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$2 with the Gaussian free field via level lines, and Le Jan’s coupling of loop-soup occupation times with the square of the Gaussian free field. The compatibility result that these couplings can be realized simultaneously is a major structural theorem in the subject (Qian et al., 2015).

A recent refinement concerns the fine regularity of critical cluster boundaries. Let $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$3 denote a Brownian loop soup of intensity $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$4 in the unit disc, let $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$5 be the set of soup double points, and let $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$6 be any cluster. Then almost surely

$$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$7

Thus no boundary point of any critical loop-soup cluster is a double point of the soup (Gao et al., 27 Jul 2025). This resolves an open question linked to generalized disconnection exponents. The same work recalls that, for cluster boundaries, the putative Hausdorff dimension of boundary double points is

$$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$8

and at criticality $$\mu_r := \int_{\mathbb C}\int_0^\infty \frac{1}{2\pi t^2}\,\mu^{br}_{z,t}\,dt\,d{\bf A}(z),$$9, so the dimension prediction is exactly zero. The theorem shows that the zero-dimensional exceptional set is in fact empty almost surely, not merely of Hausdorff dimension zero (Gao et al., 27 Jul 2025).

This is a recurrent phenomenon in critical Brownian geometry: exponent formulae identify delicate threshold dimensions, but dimension zero alone does not decide existence. The loop-soup result belongs to a broader class of “zero-dimensional” non-existence theorems proved by a common multiscale good-box method (Gao et al., 27 Jul 2025).

5. Conformal fields, exact correlators, and operator algebras

Brownian loop soup supports a large family of observables with the transformation laws of conformal primary fields. The foundational examples are the layering and winding operators. For layering, one assigns an independent sign $\mu^{br}_{z,t}$0 to each loop and forms the signed count $\mu^{br}_{z,t}$1 of loops whose filled interiors cover $\mu^{br}_{z,t}$2; for winding, one sums total winding numbers $\mu^{br}_{z,t}$3 around $\mu^{br}_{z,t}$4. The associated vertex operators are $\mu^{br}_{z,t}$5, and their one-point functions with ultraviolet and infrared cutoffs satisfy

$\mu^{br}_{z,t}$6

for layering, and

$\mu^{br}_{z,t}$7

for winding (Camia et al., 2015). The corresponding conformal dimensions are

$\mu^{br}_{z,t}$8

and the central charge is $\mu^{br}_{z,t}$9 (Camia et al., 2015).

The correlation functions of these operators can be computed explicitly. In the full plane, charge conservation takes the form

$z$0

and canonically normalized two- and three-point functions take the standard primary-field form. Exact closed-form expressions are also known for the full-plane four-point function and for the upper-half-plane two-point function, with nontrivial dependence on the conformal cross-ratio through hypergeometric functions. The four-point function is crossing symmetric, and its Virasoro conformal block expansion reveals an infinite tower of previously unknown primaries with dimensions

$z$1

where $z$2 (Camia et al., 2019).

The operator spectrum extends beyond layering and winding. The edge counting field $z$3, defined by a renormalized count of loop outer boundaries passing near $z$4, is a scalar primary of dimensions $z$5 with canonical normalization

$z$6

Its operator products generate higher-order edge operators $z$7 with dimensions $z$8, and charged edge operators $z$9 with dimensions $t$0 (Camia et al., 2021).

The same CFT structure includes a stress-energy tensor. In the bulk,

$t$1

so the Virasoro central charge is again $t$2. There is also a boundary stress-energy tensor in domains with boundary, obtained from a boundary edge operator and shown to generate local boundary deformations in the standard Ward-identity sense (Camia et al., 2021). More recently, conformal block analysis of half-plane correlators has produced an infinite family of boundary primary operators of nonnegative integer dimensions; the even-dimensional operators are interpreted probabilistically as inserting multiple outer boundaries of Brownian loops at boundary points (Camia et al., 6 Jan 2026). A further generalization assigns arbitrary random labels to loops, replacing $t$3 by a characteristic function $t$4 and yielding conformal dimension

$t$5

with exact two- and four-point functions in closed form (Foit et al., 2020).

6. Occupation fields, multiplicative chaos, massive deformations, and higher-dimensional extensions

A second major axis of the theory concerns occupation fields and their relation to Gaussian objects. For random walk loop soup at intensity $t$6, the occupation field satisfies

$t$7

where $t$8 is the discrete Gaussian free field with the appropriate covariance (Camia, 2015). In the critical planar continuum theory, the renormalized occupation field is correspondingly distributed like the properly defined square of a Gaussian free field, but this identification does not determine the full loop configuration. In particular, the exact visited set of the critical planar Brownian loop soup is not determined by the occupation field or GFF-square data; given these fields, a dense family of special points each has conditional probability $t$9 of belonging to the loop soup, and there are two distinct loop/excursion decompositions, each with conditional probability $D$00 (Lehmkuehler et al., 2024).

Brownian loop soup also furnishes multiplicative chaos measures. One construction produces a measure on thick points, formally by exponentiating the square root of the occupation field; at critical intensity $D$01, this measure coincides with the hyperbolic cosine of the Gaussian free field, and Liouville-typical points are then points of infinite loop multiplicity, with relative loop contributions governed by the Poisson–Dirichlet distribution with parameter $D$02 (Aïdékon et al., 2021). A different high-intensity, small-coupling limit yields imaginary Gaussian multiplicative chaos from the signed layering field, with convergence proved in negative Sobolev spaces using explicit Wiener–Itô chaos expansions (Camia et al., 2019). More recently, the real layering field, after renormalization, has been shown to converge to a subcritical Gaussian multiplicative chaos, again via Wiener–Itô chaos methods and explicit $D$03-point functions (Maitra, 25 Oct 2025).

The massive Brownian loop soup is the off-critical counterpart. It is obtained by exponentially killing loops according to the mass functional $D$04, is conformally covariant, exhibits exponential decay of clusters when the mass is bounded away from zero, and arises as the near-critical scaling limit of random walk loop soup with killing when the mass scales like $D$05 (Camia, 2013).

Although much of the literature is planar, Brownian loop soup also has nontrivial higher-dimensional theory. In dimensions $D$06, Brownian motion trace can be decomposed into a simple path $D$07 and all loops of an independent Brownian loop soup that intersect $D$08; in $D$09, any subsequential scaling limit of loop-erased random walk supplies such a simple path (Sapozhnikov et al., 2015). In $D$10, Brownian loop soup clusters exhibit a genuine phase transition for unbounded clusters, admit a one-arm exponent, and at sufficiently large intensity almost surely all loops are connected into a single cluster. The same work develops higher-dimensional Brownian loop measure tools, including rerooting decompositions and conformal invariance under inversion for $D$11 (Jego et al., 8 Jan 2026).

Brownian loop soup therefore occupies a rare position in probability and mathematical physics. It is simultaneously a scaling limit of discrete loop models, a generator of $D$12 and SLE-type interfaces, a source of exact conformal correlators and operator algebras, a vehicle for GFF and multiplicative-chaos couplings, and a framework in which fine geometric questions about clusters, boundaries, and higher-dimensional percolative behavior can be posed and resolved within a single Poissonian continuum ensemble.