Parabolic Anderson Model in Random Media

• PAM is a family of linear heat equations with a multiplicative random environment that models diffusion, localization, and intermittent growth in random media.
• It encompasses continuous and discrete formulations, applying tools like Feynman–Kac representations and spectral theory to capture its complex behavior.
• Research on PAM highlights phase transitions, renormalization in singular settings, and the impact of geometry on asymptotic and localization properties.

Searching arXiv for relevant Parabolic Anderson Model papers to ground the article in current literature. The Parabolic Anderson Model (PAM) is a family of linear heat equations with multiplicative random environment, formulated on discrete or continuous state spaces and used to study diffusion, growth, intermittency, localization, and spectral effects in random media. In its canonical continuous form it appears as

$$\partial_t u=\frac12\Delta u+\xi u \quad\text{or}\quad \partial_t u=\Delta u+\xi u,$$

while on $\mathbb Z^d$ it is often written as $\partial_t u=-H_\omega u$ with $H_\omega=-\Delta+V_\omega$. Across these formulations, the central observables are the solution $u$, its total mass, its moments, associated Lyapunov exponents, and the geometry of the rare regions that dominate large-time behavior (Huang et al., 2015, Metzger, 2011).

1. Core formulations

PAM encompasses several closely related models. In Euclidean space with static potential, one studies

$$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$

often with $u(0,\cdot)=\delta_0$ or $u(0,\cdot)\equiv 1$. In the stochastic-noise formulation, one instead considers multiplicative Gaussian forcing,

$\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$

where $\dot W$ may be white in time and colored in space, or homogeneous in both time and space (König et al., 2020, Balan et al., 2016).

On the lattice, a standard version is

$\mathbb Z^d$0

with $\mathbb Z^d$1. This connects PAM directly to the Anderson Hamiltonian and to integrated density of states asymptotics (Metzger, 2011). A discrete-time analogue also exists: $\mathbb Z^d$2 which the literature interprets simultaneously as directed polymer partition function recursion and as a discrete-time PAM (Caravenna et al., 2010).

The model is not tied to Euclidean geometry. It has been formulated on the torus $\mathbb Z^d$3, hyperbolic space $\mathbb Z^d$4, closed Riemannian surfaces, Galton–Watson trees, and finite locally tree-like random graphs (Chen et al., 2023, Geng et al., 25 Jun 2025, Dahlqvist et al., 2017, Hollander et al., 2020). This breadth is not merely formal: the state space changes the relevant scales, large-deviation mechanisms, and intermittency geometry.

2. Random environments, singularities, and well-posedness

The random medium in PAM ranges from static i.i.d. potentials to Gaussian fields that are white in time, colored in space, or rough in both spectral and geometric senses. For space-time homogeneous Gaussian noise on $\mathbb Z^d$5, a basic existence criterion is Dalang’s condition

$\mathbb Z^d$6

under which there is a unique random-field mild solution even for signed Borel initial measures satisfying

$\mathbb Z^d$7

(Balan et al., 2016). Continuity in law with respect to the spatial noise parameter has also been established for rough initial conditions, both in the Riesz-kernel regime and in the one-dimensional rough fractional regime $\mathbb Z^d$8 (Liang, 2023).

A major distinction is between regular and singular multiplication. In one dimension with white-in-time Gaussian noise whose spatial covariance is fractional with Hurst parameter $\mathbb Z^d$9, the product $\partial_t u=-H_\omega u$0 is interpreted through a Skorohod or Wick integral, and the solution admits an explicit Wiener chaos expansion (Hu et al., 2016). On the torus, an intrinsic colored noise with Fourier weights $\partial_t u=-H_\omega u$1 yields an Itô/Walsh theory in dimensions where spatial white noise would be too singular; the sharp torus Dalang condition is

$\partial_t u=-H_\omega u$2

(Chen et al., 2023).

Dimension $\partial_t u=-H_\omega u$3 for spatial white noise is singular in a different way. In the planar PAM with static Gaussian white-noise potential, $\partial_t u=-H_\omega u$4 is a distribution of regularity roughly $\partial_t u=-H_\omega u$5, so the product $\partial_t u=-H_\omega u$6 is ill-defined. The equation must be regularized and renormalized: $\partial_t u=-H_\omega u$7 and one studies

$\partial_t u=-H_\omega u$8

before passing to the limit $\partial_t u=-H_\omega u$9 (König et al., 2020). On closed two-dimensional Riemannian manifolds, well-posedness for the renormalized equation is obtained by extending regularity structures to graded vector bundles on curved space and by reconstructing the polynomial sector through exponential coordinates and symmetrized covariant derivatives (Dahlqvist et al., 2017).

These results collectively show that “PAM” is not a single well-posedness theorem but a collection of regimes. In regular settings one can work with mild solutions and classical stochastic integration; in singular settings one needs renormalization, paracontrolled analysis, or regularity structures.

3. Probabilistic representations and spectral structure

The standard probabilistic representation is Feynman–Kac. For a smooth static potential,

$H_\omega=-\Delta+V_\omega$0

and for total mass this often reduces to an exponential functional of Brownian motion or random walk (Metzger, 2011). In Gaussian multiplicative settings, exact moment formulas can be written in terms of independent Brownian motions or Brownian bridges; a particularly useful bridge representation separates the path interaction from the initial condition (Huang et al., 2015).

For singular white-noise potential in dimension $H_\omega=-\Delta+V_\omega$1, the classical Feynman–Kac formula is only formal. The replacement constructed in the white-noise setting is a diffusion with distributional drift obtained via a partial Girsanov transform. With

$H_\omega=-\Delta+V_\omega$2

and a suitable $H_\omega=-\Delta+V_\omega$3, the renormalized solution on a box admits a representation

$H_\omega=-\Delta+V_\omega$4

which plays the role of Feynman–Kac after renormalization (König et al., 2020).

Spectral theory is a second organizing principle. On finite domains, the Anderson operator or Hamiltonian governs the long-time behavior through its top eigenvalue. In the planar white-noise model, the renormalized Dirichlet Anderson Hamiltonian

$H_\omega=-\Delta+V_\omega$5

has principal eigenvalue $H_\omega=-\Delta+V_\omega$6, and the asymptotics of total mass are linked by

$H_\omega=-\Delta+V_\omega$7

(König et al., 2020). In the discrete Anderson model, effective-medium analysis similarly treats $H_\omega=-\Delta+V_\omega$8 and the Laplace transform of the integrated density of states as two faces of the same variational mechanism, balancing Dirichlet energy against effective potential gain (Metzger, 2011).

This spectral viewpoint is not universal in exactly the same form, but it is recurrent. Principal eigenvalues, variational formulas, and occupation-measure large deviations repeatedly encode the same competition: concentration near favorable regions yields potential gain, while localization carries kinetic cost.

4. Intermittency, localization, and front propagation

Intermittency is a defining theme of PAM. In the annealed sense, it is often expressed by strict separation of moment growth rates, for instance by divergence of

$H_\omega=-\Delta+V_\omega$9

as $u$0 (Geng et al., 25 Jun 2025). In the heavy-tail lattice setting, intermittency can be sharpened into a geometric statement: the annealed mass is carried by isolated potential peaks whose heights lie in an asymptotically narrow window

$u$1

and these relevant sites form an asymptotically Bernoulli field in space (Gärtner et al., 2010).

Localization can be much stronger in quenched settings. In the discrete-time heavy-tailed model with Pareto potential, the quenched endpoint law localizes almost surely on a single site: $u$2 and the localization site grows ballistically, with

$u$3

for an explicit law on the $u$4-unit ball (Caravenna et al., 2010). The same work shows that typical paths are nearly minimal routes to the favorite site, followed by trapping there.

At the same time, leading-order asymptotics do not always reveal intermittency. For the two-dimensional renormalized white-noise PAM, the total mass and the pointwise profile share the same logarithmic leading order: $u$5 and likewise on interior boxes $u$6. The authors explicitly note that intermittency is therefore not visible at that first logarithmic scale (König et al., 2020). A plausible implication is that localization, if present, lies in lower-order structure rather than in the leading exponent.

PAM also supports front-type observables. In the one-dimensional discrete model with bounded i.i.d. environment,

$u$7

the breakpoint

$u$8

has a quenched invariance principle after centering and $u$9-scaling, and differs from the BRWRE maximal-particle median only by $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$0 (Černý et al., 2017). This extends PAM beyond mass concentration questions to random-front propagation.

5. Asymptotic regimes and phase transitions

Large-time asymptotics in PAM are highly regime-dependent. One line of work treats annealed moment asymptotics through tail classes of the static potential. In the time-correlation analysis of lattice PAM, exact asymptotics of moments and mixed-time functionals are obtained under heavy tails with finite cumulant generating function, and ageing occurs if and only if

$$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$1

(Gärtner et al., 2010). Another line, using effective-medium methods for the discrete Anderson model, organizes the behavior into “quantum” and “classical” regimes according to the de Haan auxiliary function $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$2, with distinct optimal peak sizes and corresponding integrated-density-of-states asymptotics (Metzger, 2011).

Time-dependent diffusion produces an explicit phase diagram. In the accelerated/decelerated model

$$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$3

two critical scales emerge: the lower scale $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$4, determined by the tail thickness of the potential, and the upper scale $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$5. Below the lower scale the mass flow gets stuck; at the lower scale a new discrete variational problem $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$6 appears; between the two scales one recovers the familiar intermittent-island regimes after rescaling; and at the upper scale only boundedly favorable “super-average” potential values contribute (Konig et al., 2010).

Dynamic environments add further phase structure. For a finite number of moving catalysts,

$$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$7

the annealed Lyapunov exponents are governed by the top of the spectrum of

$$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$8

and positivity, vanishing, and intermittency depend sharply on dimension, the number of catalysts, and the diffusion rates (Castell et al., 2010). With spatially correlated Gaussian noise, large-time moment exponents are described by a variational functional $$\partial_t u(t,x)=\frac12\Delta u(t,x)+\xi(x)u(t,x),$$9, and the model exhibits a phase transition in the noise amplitude: under nonnegative locally integrable covariance, phase transition occurs if and only if

$u(0,\cdot)=\delta_0$0

(Huang et al., 2015).

Hyperbolic geometry changes the critical exponents themselves. For white-in-time colored-noise PAM on $u(0,\cdot)=\delta_0$1 with spatial correlation $u(0,\cdot)=\delta_0$2, the critical decay is $u(0,\cdot)=\delta_0$3, not $u(0,\cdot)=\delta_0$4 as in Euclidean space. If $u(0,\cdot)=\delta_0$5, the second moment stays uniformly bounded in time for sufficiently small noise; if $u(0,\cdot)=\delta_0$6, it always diverges; and for $u(0,\cdot)=\delta_0$7 with small noise one gets the new sub-exponential regime

$u(0,\cdot)=\delta_0$8

while large noise yields exponential growth (Geng et al., 7 Jul 2025).

6. Geometry, topology, and non-Euclidean PAM

The dependence of PAM on the underlying state space is now explicit across several settings. On the flat torus $u(0,\cdot)=\delta_0$9, the heat kernel converges exponentially to equilibrium and Brownian motion is ergodic. This compactness changes long-time moment behavior: under natural positivity assumptions and $u(0,\cdot)\equiv 1$0, the second moment admits exponential lower bounds, and the compact geometry suppresses the phase transition familiar from transient Euclidean dimensions (Chen et al., 2023).

On hyperbolic space, geometry has a split effect. In the annealed problem with regular stationary Gaussian potential,

$u(0,\cdot)\equiv 1$1

the second-order moment asymptotics are identical to the Euclidean case: $u(0,\cdot)\equiv 1$2 and the fluctuation functional becomes Euclidean after the rescaling $u(0,\cdot)\equiv 1$3, a mechanism described as the curvature dilation effect (Geng et al., 25 Jun 2025). In the quenched problem, by contrast, hyperbolic geometry is decisive: for regular stationary time-independent Gaussian potential,

$u(0,\cdot)\equiv 1$4

with

$u(0,\cdot)\equiv 1$5

and the optimal Brownian strategy is to reach a favorable region at distance of order $u(0,\cdot)\equiv 1$6 at time $u(0,\cdot)\equiv 1$7, then remain there (Geng et al., 25 Jun 2025). This sharp distinction between annealed and quenched hyperbolic asymptotics is one of the clearest demonstrations that geometry affects different PAM observables in different ways.

Closed two-dimensional manifolds require still another analytic layer. On a closed Riemannian surface, the renormalized PAM with spatial white noise is solved by extending regularity structures to curved space, replacing Euclidean polynomials with covariant Taylor expansions and heat-kernel-based model maps (Dahlqvist et al., 2017). On supercritical Galton–Watson trees with bounded degrees, quenched long-time asymptotics of the total mass have a second-order contribution given by a variational formula, and the analysis suggests concentration on a tree of minimal degree; the same strategy extends to finite locally tree-like random graphs in a coupled limit of time and graph size (Hollander et al., 2020).

A common misconception is that geometry only changes technical constants. The evidence across torus, hyperbolic space, manifolds, and trees indicates otherwise. Compactness can force recurrence and exponential equilibration; negative curvature can alter critical exponents and localization mechanisms; and random branching geometry can modify the effective variational problem governing concentration.

7. Representative sharp results and open directions

Several theorem-level results now serve as benchmarks for the subject. In the renormalized two-dimensional white-noise model,

$u(0,\cdot)\equiv 1$8

with

$u(0,\cdot)\equiv 1$9

linking quenched growth to the principal eigenvalue of the renormalized Anderson operator (König et al., 2020). For white-in-time spatially correlated noise, the $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$0-th moment Lyapunov exponent is exactly the variational quantity $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$1 when $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$2 (Huang et al., 2015). On hyperbolic space, the annealed fluctuation scale is $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$3 with Euclidean variational constant, whereas the quenched scale is $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$4 with a genuinely non-Euclidean optimizer (Geng et al., 25 Jun 2025, Geng et al., 25 Jun 2025).

The literature also makes clear where the main open problems lie. For the two-dimensional white-noise PAM, sharper intermittency and localization statements beyond the first $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$5 scale, as well as extension to dimension $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$6, remain open (König et al., 2020). In hyperbolic-space phase transition theory, the borderline case $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$7 is identified as critical but not treated sharply (Geng et al., 7 Jul 2025). For the torus model, the case $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$8 is more delicate, and extension from $\partial_t u=\frac12\Delta u+\lambda u\,\dot W,$9 to general compact manifolds is explicitly posed as a future direction (Chen et al., 2023). In the one-dimensional discrete front problem, an exact analogue of Bramson’s logarithmic correction remains unresolved in random environment (Černý et al., 2017). In the accelerated/decelerated model, the fastest-diffusion phase beyond the upper critical scale is conjectural rather than theorem-level (Konig et al., 2010).

Taken together, these developments present PAM not as a single asymptotic theory but as a framework in which diffusion, randomness, and geometry interact through several distinct mechanisms. Depending on the regime, the dominant object may be a principal eigenvalue, a local-time variational problem, a rare potential peak, a Brownian bridge optimization, or a geometric confinement strategy. That multiplicity of mechanisms is central to the modern understanding of the Parabolic Anderson Model.