Quenched Local Limit Theorem

• Quenched Local Limit Theorem is a precise asymptotic result for transition probabilities in fixed random environments, showing local Gaussian behavior after diffusion scaling.
• It refines the quenched invariance principle by determining the exact shape of heat kernels or mass functions through techniques like Sobolev inequalities, Moser iteration, and parabolic Harnack estimates.
• The theorem applies to various models—including degenerate diffusions, reflecting processes, and random dynamical systems—under sharp moment conditions and model-specific corrections.

A quenched local limit theorem is an almost-sure asymptotic statement, for a fixed realization of a random environment or driving sequence, about the local behavior of transition probabilities, heat kernels, or analogous observables after central-limit scaling. In Markov-process settings, the “local” qualifier refers to pointwise asymptotics of the transition density itself rather than only to averaged quantities or weak convergence, and the “central” qualifier specifies that the limit has Gaussian form (Takeuchi, 2023). In this sense, a quenched local limit theorem refines a quenched invariance principle: the latter identifies Brownian scaling limits in distribution, whereas the former identifies the asymptotic shape of the heat kernel or local mass function in a fixed medium (Chiarini et al., 2015).

1. Definition and basic meaning

A local limit theorem for Markov processes and diffusions is a pointwise asymptotic statement for transition probabilities or densities. A local central limit theorem strengthens this by showing that the limiting kernel is Gaussian, with the canonical $t^{-d/2}$ normalization and a quadratic form determined by an effective covariance matrix (Takeuchi, 2023). In continuous-space settings, this means that a heat kernel $p_t(x,y)$ or a density $q_t(x,y)$ behaves asymptotically like a Gaussian heat kernel after diffusive rescaling.

The distinction between quenched and annealed formulations is structural. Annealed statements average over the environment law, whereas quenched statements hold for almost every fixed realization of the environment. In ergodic random media, quenched results typically show that the local asymptotic profile is deterministic at leading order, even though the underlying medium is random (Chiarini et al., 2015).

For symmetric diffusions with effective covariance matrix $\Sigma$, a standard Gaussian reference kernel is

$$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$

This kernel appears as the universal limit in several distinct models, although the observable converging to it may require a deterministic normalization, an invariant-density correction, or a multiplicative stationary random field depending on the model class (Chiarini et al., 2015).

2. Canonical asymptotic forms

Across the literature, the local object and the limiting profile vary with the reference measure, reversibility structure, and geometry of the state space.

Model	Local quantity	Limiting profile
Degenerate divergence-form diffusion	$$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$	$\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$
Reflecting diffusion on $W'(\omega)$	$$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$	$k_t^\Sigma(x)$
Dynamic random conductance model	$p_t(x,y)$0	$p_t(x,y)$1
Balanced time-dependent RWRE	$p_t(x,y)$2	$p_t(x,y)$3
Long-range reversible conductance model	$p_t(x,y)$4	$p_t(x,y)$5

In the divergence-form diffusion with speed measure $p_t(x,y)$6, the heat kernel $p_t(x,y)$7 is defined with respect to $p_t(x,y)$8. The quenched local central limit theorem therefore carries the deterministic prefactor $p_t(x,y)$9. If one instead uses the Lebesgue density $q_t(x,y)$0, that prefactor disappears and the limit is the standard Gaussian kernel with covariance $q_t(x,y)$1 (Chiarini et al., 2015).

Other models require different corrections. In time-dependent balanced environments, the natural local object is the adjusted kernel $q_t(x,y)$2, where $q_t(x,y)$3 is the invariant density for the environment seen from the particle (Deuschel et al., 2017). In directed random walk on the backbone of supercritical oriented percolation, the quenched probability is asymptotically the annealed probability multiplied by $q_t(x,y)$4, the density of the invariant measure for the environment seen from the particle (Bethuelsen et al., 2024). In stochastic flows, the quenched density is asymptotically a deterministic Gaussian multiplied by a spacetime-stationary random field $q_t(x,y)$5, and only in the incompressible case does one recover a pure Gaussian local central limit theorem with $q_t(x,y)$6 (Dunlap et al., 2021).

3. Analytical mechanisms

A recurring theme is that a quenched invariance principle alone is not sufficient for a quenched local limit theorem. What is additionally required is a local regularity theory strong enough to upgrade averaged convergence to pointwise convergence.

In degenerate divergence-form diffusions, the key tools are weighted Sobolev and Nash inequalities, Moser iteration adapted to measurable unbounded and degenerate coefficients, a local parabolic Harnack inequality, on-diagonal heat-kernel bounds, and Hölder-type equicontinuity on diffusive scales. The deterministic local central limit theorem is first proved under convergence of spatial averages and a QIP-type integral convergence, and then ergodicity identifies the normalizing constant as $q_t(x,y)$7 and yields the quenched statement (Chiarini et al., 2015).

Reflecting diffusions on a continuum percolation cluster use a related but geometrically sharper scheme. Very regularity, relative isoperimetry, hole-size control, and intrinsic–Euclidean distance comparability yield Sobolev and weighted Poincaré inequalities on intrinsic balls; these feed Moser iteration, parabolic Harnack inequality, and Hölder continuity of the heat kernel. The pointwise theorem is then obtained by combining the invariance principle with an averaging-to-pointwise upgrade based on projection to the accessible set via $q_t(x,y)$8 (Takeuchi, 2023).

For time-dependent degenerate discrete media, the local regularity step is obtained differently. In the dynamic random conductance model on $q_t(x,y)$9, the proof relies on Hölder continuity estimates for solutions to the heat equation for discrete finite difference operators in divergence form with time-dependent degenerate weights, obtained by De Giorgi’s iteration technique rather than by a parabolic Harnack inequality (Andres et al., 2020). In long-range conductance models, classical local Harnack theory is replaced by a weak parabolic Harnack inequality with an $\Sigma$0-tail term, together with maximal inequalities and on-diagonal heat-kernel bounds (Chen et al., 2024).

In random dynamical systems, the local theorem is not phrased in terms of a heat kernel. Instead, twisted transfer operator cocycles encode characteristic functions of Birkhoff sums. The decisive objects are the top Lyapunov exponent of the twisted cocycle, its regularity near the origin, and exponential decay of twisted operators away from the origin in the twist parameter. This spectral mechanism yields quenched large deviations, central limit, and local central limit theorems for random expanding, hyperbolic, and expanding-on-average cocycles (Dragicevic et al., 2017, Dragičević et al., 2018, Dragičević et al., 2021).

4. Degenerate and unbounded diffusions in divergence form

A central continuous model is the symmetric diffusion on $\Sigma$1 in divergence form studied in “Local Central Limit Theorem for diffusions in a degenerate and unbounded Random Medium” (Chiarini et al., 2015). The environment is a stationary ergodic probability space $\Sigma$2 with measure-preserving spatial shifts $\Sigma$3. The coefficient field is a symmetric nonnegative-definite random matrix $\Sigma$4, controlled by random scalar fields $\Sigma$5 through

$\Sigma$6

No differentiability in $\Sigma$7 is imposed, and the coefficients may be both degenerate and unbounded (Chiarini et al., 2015).

The formal generator is

$\Sigma$8

with reference measure $\Sigma$9. Because the coefficients are merely measurable, the process is constructed via Dirichlet forms rather than by stochastic differential equation methods. If $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$0, Fukushima’s theory gives a conservative symmetric diffusion $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$1 with semigroup $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$2 and heat kernel $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$3 with respect to $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$4 (Chiarini et al., 2015).

The quenched local central limit theorem is proved for $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$5 under the moment assumptions

$$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$6

together with a quenched invariance principle with deterministic positive-definite covariance matrix $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$7. For any compact $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$8 and $$k_t^\Sigma(x)=\frac{1}{\sqrt{(2\pi t)^d\det\Sigma}\exp\!\Big(-\frac{x\cdot\Sigma^{-1}x}{2t}\Big)}.$$9,

$$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$0

for $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$1-almost every $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$2 and almost every starting point $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$3; if $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$4, the statement holds for all $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$5 (Chiarini et al., 2015).

In large-time form this becomes uniform convergence on diffusive scales $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$6. The result does not claim explicit rates. Its main novelty is that the classical uniformly elliptic framework is replaced by weighted Sobolev/Nash inequalities and Moser iteration under integrability moment conditions only. The condition $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$7 is presented as essentially sharp; the paper further notes that if $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$8, ergodic environments where the quenched LCLT fails are known in the discrete random conductance literature (Chiarini et al., 2015).

The same framework extends to general speed measures $$\varepsilon^{-d}p^\omega_{t/\varepsilon^2}(o,x/\varepsilon)$$9 with formal generator

$\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$0

under moment assumptions involving $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$1, $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$2, and $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$3. This places the theorem within a broader homogenization program for degenerate media in divergence form (Chiarini et al., 2015).

5. Geometric, discrete, and long-range variants

The geometric content of a quenched local limit theorem becomes especially visible for reflecting diffusions on random domains. In the modified continuum percolation cluster $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$4, the diffusion is constructed from a uniformly elliptic Dirichlet form on $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$5 with Neumann boundary conditions. Under stationarity and ergodicity of the underlying point process, very regularity, global isoperimetry, hole-size control, and intrinsic–Euclidean distance comparability, the rescaled heat kernel satisfies

$\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$6

for $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$7-almost every $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$8 when $\mathbb{E}_\mu[\Lambda]^{-1}k_t^\Sigma(x)$9. The projection $W'(\omega)$0 compensates for the fact that the random domain is not all of $W'(\omega)$1 (Takeuchi, 2023).

For discrete walks among time-dependent ergodic degenerate conductances on $W'(\omega)$2, the variable-speed random walk has generator

$W'(\omega)$3

Assuming space-time stationarity and ergodicity and the moment condition

$W'(\omega)$4

the quenched local CLT takes the form

$W'(\omega)$5

for almost every environment. The proof is based on De Giorgi iteration and a general local-CLT criterion combining a CLT, volume regularity, distance control, and a Hölder-type modulus of continuity for the heat kernel (Andres et al., 2020).

A related non-reversible model is the continuous-time random walk in a uniformly elliptic time-dependent balanced random environment. There the relevant local object is the adjusted kernel

$W'(\omega)$6

where $W'(\omega)$7 is the invariant density of the environment process seen from the walker. The quenched local central limit theorem states that

$W'(\omega)$8

for almost every environment. Here the balance condition replaces reversibility, and the analytical core is a parabolic Harnack inequality for the adjoint operator together with volume-doubling and $W'(\omega)$9 properties for $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$0 (Deuschel et al., 2017).

Long-range conductance models permit jumps of arbitrary length. Under stationary ergodic symmetric conductances, strengthened moment assumptions, and either a higher moment condition on $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$1 or $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$2, the quenched LLT states that for $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$3,

$$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$4

almost surely. The proof is based on weak parabolic Harnack inequalities and on-diagonal heat-kernel estimates for long-range random walks on general ergodic environments (Chen et al., 2024).

Directed walk on the backbone of supercritical oriented percolation exhibits yet another form of quenched local asymptotics. In that setting,

$$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$5

for $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$6-almost every $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$7, where $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$8 is the density of the invariant measure for the environment seen from the particle. The 2024 result extends the previously known quenched theorem from $$\varepsilon^d p^\omega_{t/\varepsilon^2}(0,g^\omega(x/\varepsilon))$$9 to all $k_t^\Sigma(x)$0 (Bethuelsen et al., 2024).

6. Random modulation and dynamical-systems formulations

Not every quenched local limit theorem converges to a purely deterministic Gaussian. In the stochastic-flow model

$k_t^\Sigma(x)$1

where $k_t^\Sigma(x)$2 is a mean-zero Gaussian velocity field that is white in time and smooth in space and $k_t^\Sigma(x)$3 is an independent Brownian motion, the annealed law of $k_t^\Sigma(x)$4 is exactly Gaussian with covariance $k_t^\Sigma(x)$5. Conditionally on the velocity field, however, the quenched density $k_t^\Sigma(x)$6 is asymptotically

$k_t^\Sigma(x)$7

where $k_t^\Sigma(x)$8 is a positive, spacetime-stationary random field with $k_t^\Sigma(x)$9. Only in the incompressible case does one have $p_t(x,y)$00 almost surely, which yields a genuine local central limit theorem in Gaussian form (Dunlap et al., 2021). This shows that quenched local asymptotics may retain a stationary multiplicative memory of the environment.

In random dynamical systems, the local theorem concerns Birkhoff sums rather than a heat kernel. For random expanding, hyperbolic, and expanding-on-average cocycles, one studies

$p_t(x,y)$01

under a fixed realization $p_t(x,y)$02. The spectral approach introduces twisted transfer operators and analyzes the top Lyapunov exponent of the twisted cocycle. Under centering, positivity of the variance, and a non-arithmetic or aperiodicity condition, one obtains interval-level local Gaussian asymptotics. In the expanding-on-average setting, the local limit theorem takes the form

$p_t(x,y)$03

uniformly in $p_t(x,y)$04, for almost every $p_t(x,y)$05 (Dragičević et al., 2021). Earlier spectral formulations established quenched non-lattice and periodic local central limit theorems for random expanding maps and then extended them to random hyperbolic systems, including some classes of billiards (Dragicevic et al., 2017, Dragičević et al., 2018).

These examples show that “quenched local limit theorem” is not restricted to reversible diffusion theory. The common feature is a local Gaussian asymptotic under fixed disorder, but the local observable may be a heat kernel, a mass function, or an interval probability for a Birkhoff sum.

7. Sharpness, limitations, and conceptual boundaries

Several limitations recur across the subject. Dimension restrictions are often intrinsic to the analytic mechanism: the degenerate divergence-form diffusion and the dynamic random conductance model are treated for $p_t(x,y)$06, whereas the reflecting-diffusion theorem on continuum percolation clusters requires $p_t(x,y)$07 because the local analytic scheme uses Sobolev embeddings and Moser iteration exponents in that range (Chiarini et al., 2015, Takeuchi, 2023, Andres et al., 2020).

A second limitation concerns the strength of regularity statements. In degenerate or geometrically irregular media, genuine two-sided Gaussian bounds are often not claimed. Instead, the proofs use on-diagonal bounds, parabolic Harnack inequalities, and Hölder-type equicontinuity sufficient for the local theorem (Chiarini et al., 2015, Takeuchi, 2023). This is one reason the passage from a quenched invariance principle to a quenched local limit theorem is nontrivial: local continuity estimates are an additional ingredient, not a formal consequence of weak convergence.

A third boundary is the persistence of model-dependent corrections. The factor $p_t(x,y)$08 in weighted diffusions, the invariant density $p_t(x,y)$09 or $p_t(x,y)$10 in non-reversible walks, and the stationary field $p_t(x,y)$11 in stochastic flows show that the limiting local profile need not be a bare Gaussian kernel (Chiarini et al., 2015, Deuschel et al., 2017, Dunlap et al., 2021, Bethuelsen et al., 2024). A plausible implication is that the correct reference measure, or the correct environment-seen-from-the-particle stationary state, is part of the theorem’s content rather than an auxiliary normalization.

The threshold conditions can also be close to optimal. In the degenerate divergence-form diffusion, the strict inequality $p_t(x,y)$12 is presented as essentially sharp, and the paper notes that when $p_t(x,y)$13 there are ergodic environments where the quenched LCLT fails in the discrete random conductance literature (Chiarini et al., 2015). For long-range conductance models, the Gaussian quenched LLT is confined to the diffusive regime under finite-variance-type moment assumptions; heavy-tailed regimes with $p_t(x,y)$14-stable scaling are explicitly outside the scope (Chen et al., 2024).

Open problems remain model-specific. For reflecting diffusions in continuum percolation, the theorem is proved on the modified cluster $p_t(x,y)$15, while a direct quenched LCLT on the original cluster $p_t(x,y)$16 is left open because traps and bottlenecks may invalidate quenched invariance (Takeuchi, 2023). In dynamical systems, non-arithmetic conditions are decisive, and periodic corrections replace the non-lattice Gaussian form when arithmetic obstructions are present (Dragicevic et al., 2017, Dragičević et al., 2018).

Taken together, these results identify the quenched local limit theorem as a high-resolution homogenization statement. It combines ergodicity, a macroscopic Gaussian scaling limit, and a local regularity theory strong enough to resolve the heat kernel or local mass function itself. In the strongest form, the theorem shows that even in degenerate, time-dependent, reflecting, non-reversible, or randomly driven settings, the local law at diffusive scale is asymptotically Gaussian, up to the precise deterministic or stationary correction dictated by the model.