Annealed Exponential Mixing in Random Dynamics

• Annealed exponential mixing is the averaged decay of mixing observables in random systems, distinguishing it from the pathwise quenched mixing approach.
• It applies to a variety of models—from random wedge shears and cellular flows to stochastic Navier–Stokes advection—using tools like hypocoercivity, two-point PDEs, and Lyapunov methods.
• Recent findings reveal that while randomness does not always enhance the mixing rate, it crucially underpins exponential convergence by averaging over random driving forces.

Searching arXiv for the cited works and closely related papers on annealed exponential mixing. Annealed exponential mixing is the averaged-over-randomness form of exponential mixing for random or randomly driven dynamics. In the modern literature it is not tied to a single diagnostic: depending on the model, the exponentially decaying quantity may be a geometric mixing scale, a negative Sobolev norm, a two-point correlation, or a distance between transition laws. The common feature is that the decay is formulated after averaging over the driving randomness, in contrast with quenched statements, which hold along almost every realization. Recent work places the notion at the intersection of passive scalar advection, random dynamical systems, stochastic fluid models, time-inhomogeneous Markov processes, and annealing-based sampling algorithms (Cheng et al., 2021, Navarro-Fernández et al., 24 Feb 2025, DeWitt et al., 5 Sep 2025).

1. Definitions, observables, and the annealed–quenched distinction

In random transport and random dynamical systems, annealed exponential mixing is usually formulated as exponential decay in expectation of a mixing observable. For randomized alternating wedge shears on $\mathbb{T}^2$, the observable is a grid-adapted geometric mixing scale $\epsilon(k)$, and annealed exponential mixing is implemented as

$$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$

equivalently $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$ over the regression window used in the numerics (Cheng et al., 2021). In random cellular flows, the initial annealed statement is instead decay of a deterministic two-point correlation PDE, which is then upgraded to quenched exponential decay of $\|\theta(t)\|_{\dot H^{-1}}$ (Navarro-Fernández et al., 24 Feb 2025). In the abstract IID conservative setting, annealed exponential mixing is stated for the two-point motion:

$$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$

with corresponding quenched one-point bounds derived under additional structure (DeWitt et al., 5 Sep 2025).

The observable varies substantially across subfields. This is not a notational accident: each choice encodes a different coarse-graining of filamentation, correlation loss, or semigroup contraction.

Setting	Annealed quantity	Exponential statement
Random wedge shears	$\mathbb{E}[\epsilon(k)]$	$$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$
Random cellular flow	Two-point hypoelliptic PDE / $\|f(t)\|_{\mathcal H^1}$	$$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$
Stochastic Navier–Stokes advection	$\epsilon(k)$0 or Wasserstein contraction	$\epsilon(k)$1 decay / semigroup spectral gap
IID conservative random systems	Annealed two-point correlation	$\epsilon(k)$2 decay
Dynamical percolation	Block-time expected TV contraction	$\epsilon(k)$3 across blocks

A central distinction is that annealed and quenched mixing are not interchangeable notions. Quenched mixing permits a random prefactor and is pathwise; annealed mixing is averaged over the environment or driving noise. Several papers obtain annealed bounds first and then upgrade to quenched statements, but the upgrade requires additional ingredients such as two-point estimates, moment bounds, Borel–Cantelli arguments, or Harris-type contraction mechanisms (Navarro-Fernández et al., 24 Feb 2025, Cooperman et al., 2024, DeWitt et al., 5 Sep 2025).

2. Alternating shears and passive scalar mixing

A concrete numerical realization appears in the study of alternating horizontal and vertical wedge shears on $\epsilon(k)$4, governed by the inviscid advection equation

$\epsilon(k)$5

with $\epsilon(k)$6 throughout. The horizontal and vertical shears are

$\epsilon(k)$7

with exact maps

$\epsilon(k)$8

The numerical study compares fixed-shift/fixed-time, random-shift/fixed-time, fixed-shift/random-time, and random-shift/random-time protocols, always with 100 independent runs per case and a grid resolution $\epsilon(k)$9 (Cheng et al., 2021).

For $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$0, all randomized protocols exhibit annealed exponential mixing in the sense of linear decay of $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$1 and hence exponential decay of $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$2. The observed rates are largest for $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$3 or $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$4; for example, at $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$5 the mean base-2 rates are $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$6 for RSFT, $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$7 for FSRT, and $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$8 for RSRT, while the deterministic FSFT value is $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k},$$9. The same study reports that randomization does not enhance mixing in the exponentially mixing regime and instead slightly lowers the mean rate, with reductions of about $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$0 for RSFT and about $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$1 to $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$2 for FSRT and RSRT at $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$3 (Cheng et al., 2021).

The same model also isolates a failure mechanism for too-short switching times. For $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$4, invariant line segments exist in the fixed-shift Poincaré map, and for $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$5 a $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$6-cycle of segments appears; in both cases local Jacobians are described as Jordan blocks with eigenvalue $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$7, precluding exponential mixing in the deterministic fixed-shift case. Randomization can remove these specific invariant sets, but it still does not recover the higher rates observed for $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$8 (Cheng et al., 2021).

A rigorous counterpart appears for analytic Hamiltonian shears with random switching durations. There the advection equation is again inviscid on $\mathbb{E}[-n_k]\approx a-\bar\lambda_2 k$9, but the shears are analytic:

$\|\theta(t)\|_{\dot H^{-1}}$0

and one step of random alternation is

$\|\theta(t)\|_{\dot H^{-1}}$1

Under analyticity and Lie-bracket spanning assumptions, the one-point, projective, and two-point chains are uniformly geometrically ergodic; the top Lyapunov exponent is positive almost surely; and the two-point chain satisfies annealed exponential correlation decay. The paper then derives quenched exponential decay of the geometric mixing scale in Bressan’s sense, including the randomized-time Pierrehumbert model and a random analogue of the Chirikov standard map (Zhang, 13 Feb 2025).

These two alternating-shear literatures exhibit a shared structural fact: exponential mixing is linked not to randomness alone, but to repeated transverse stretching. In the wedge-shear numerics, coherent deterministic alternation with moderate $\|\theta(t)\|_{\dot H^{-1}}$2 is fastest; in the analytic Hamiltonian theory, random durations furnish the controllability and cocycle structure needed for positive Lyapunov exponents and geometric ergodicity (Cheng et al., 2021, Zhang, 13 Feb 2025).

3. Two-point PDEs, hypocoercivity, and stochastic fluid models

A different realization of annealed exponential mixing arises in passive scalar advection by a cellular flow with a Brownian-moving center:

$\|\theta(t)\|_{\dot H^{-1}}$3

Here $\|\theta(t)\|_{\dot H^{-1}}$4 is the canonical cellular flow generated by $\|\theta(t)\|_{\dot H^{-1}}$5, and $\|\theta(t)\|_{\dot H^{-1}}$6 is Brownian motion on $\|\theta(t)\|_{\dot H^{-1}}$7. The key annealed object is the deterministic two-point correlation function $\|\theta(t)\|_{\dot H^{-1}}$8, which solves a hypoelliptic PDE on $\|\theta(t)\|_{\dot H^{-1}}$9:

$$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$0

The proof uses a modified hypocoercive functional $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$1, an enlarged Hörmander commutator chain $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$2, and additional orthogonal complements $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$3 that recover coercivity near the diagonal degeneracy set. This yields a deterministic exponential rate $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$4 independent of $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$5, with

$$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$6

and then an almost-sure quenched bound

$$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$7

with finite moments of all orders for $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$8. The same framework yields enhanced dissipation in $$\Bigg|\mathbb{E}\Big[\iint_{M\times M} A(x,y)\, B(F^N_\omega(x),F^N_\omega(y))\,dx\,dy\Big]\Bigg| \le C e^{-\alpha N}\|A\|\,\|B\|,$$9 and the asymptotic relation $\mathbb{E}[\epsilon(k)]$0 as $\mathbb{E}[\epsilon(k)]$1 (Navarro-Fernández et al., 24 Feb 2025).

In passive scalar advection by the 2D stochastic Navier–Stokes equations with finitely many forced modes, annealed exponential mixing is formulated through exponential contraction of the two-point Markov semigroup in a weighted Wasserstein metric. The forcing satisfies a finite-dimensional hypoellipticity condition, and the proof combines asymptotic strong Feller estimates, controllability, positivity of the top Lyapunov exponent, tangent-process approximations near the diagonal, and the weak Harris theorem. The resulting two-point spectral gap implies annealed scalar mixing in expectation,

$\mathbb{E}[\epsilon(k)]$2

for large times, together with an almost-sure quenched $\mathbb{E}[\epsilon(k)]$3 decay with a random prefactor having controlled moments (Cooperman et al., 2024).

The cellular-flow and stochastic-Navier–Stokes papers represent two complementary routes to annealed exponential mixing. The former is purely Eulerian and PDE-based, built around a deterministic two-point hypoelliptic equation and weighted hypocoercivity. The latter is probabilistic and semigroup-based, centered on ASFP smoothing, Lyapunov repulsion from the diagonal, and Harris-type contraction of the two-point process (Navarro-Fernández et al., 24 Feb 2025, Cooperman et al., 2024).

4. From annealed two-point mixing to quenched one-point mixing

The most explicit abstract formulation is given for conservative IID random dynamical systems on a compact manifold $\mathbb{E}[\epsilon(k)]$4. Random maps are sampled from a probability measure on volume-preserving $\mathbb{E}[\epsilon(k)]$5 diffeomorphisms, and the two-point motion uses the same randomness in both coordinates:

$\mathbb{E}[\epsilon(k)]$6

If the two-point motion is annealed exponentially mixing on $\mathbb{E}[\epsilon(k)]$7 or on Hölder spaces, then for every $\mathbb{E}[\epsilon(k)]$8 there exists $\mathbb{E}[\epsilon(k)]$9 such that for almost every $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$0,

$$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$1

with a polynomial tail bound $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$2. The same work proves that annealed two-point exponential mixing plus an annealed CLT with polynomial rate of convergence implies a quenched CLT for the one-point motion, and it does so uniformly across all Sobolev and Hölder spaces of positive index (DeWitt et al., 5 Sep 2025).

This result clarifies a frequent misconception. Annealed one-point mixing by itself does not imply quenched mixing. The paper gives counterexamples: random translations on the torus exhibit annealed mixing trivially, yet quenched mixing fails; conversely, there are examples with quenched mixing but no annealed mixing. What is structurally decisive is not annealing alone, but annealed exponential mixing of the two-point motion (DeWitt et al., 5 Sep 2025).

The cellular-flow analysis provides a concrete realization of the same bridge. There the proof begins with annealed exponential decay for the deterministic two-point PDE and then upgrades to quenched pathwise mixing by a Borel–Cantelli argument along integer times together with stability estimates for continuous times (Navarro-Fernández et al., 24 Feb 2025). The stochastic-Navier–Stokes setting follows a similar architecture at the semigroup level: annealed contraction of the two-point process is the input from which annealed and quenched scalar mixing are derived (Cooperman et al., 2024).

5. Other manifestations and terminological extensions

Annealed exponential mixing also appears in one-dimensional random dynamics. For a finite family of orientation-preserving $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$3 interval diffeomorphisms $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$4 on $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$5, applied IID with probabilities $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$6, the annealed Koopman operator is

$$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$7

Under the conditions that each interior point can be moved both left and right by some maps and that the expected boundary Lyapunov exponents

$$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$8

are positive, there exists a unique stationary measure $$\mathbb{E}[\epsilon(k)] \le C\,2^{-\bar\lambda_2 k}$$9 supported in the interior, and the key estimate is exponential synchronization in average:

$\|f(t)\|_{\mathcal H^1}$0

This implies exponential convergence of $\|f(t)\|_{\mathcal H^1}$1 to $\|f(t)\|_{\mathcal H^1}$2 in weighted Hölder and $\|f(t)\|_{\mathcal H^1}$3 senses, hence annealed exponential decay of correlations with respect to $\|f(t)\|_{\mathcal H^1}$4 (Czudek, 23 Apr 2025).

For random walk on supercritical dynamical percolation, the chain is time-inhomogeneous, so the natural annealed statement is a block-time contraction rather than a single semigroup estimate. On blocks of length

$\|f(t)\|_{\mathcal H^1}$5

the analysis of the volume-biased evolving set process and giant-cluster isoperimetry yields, with probability at least $\|f(t)\|_{\mathcal H^1}$6 over the environment, a contraction of total variation distance by a factor at most $\|f(t)\|_{\mathcal H^1}$7. Iterating over $\|f(t)\|_{\mathcal H^1}$8 blocks gives annealed exponential decay across blocks:

$\|f(t)\|_{\mathcal H^1}$9

leading to mixing by time of order $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$0 up to polylogarithmic factors (Peres et al., 2017).

The phrase also appears in annealing-based sampling, but there it has a distinct meaning. In nearest-neighbor weighted random walks governed by $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$1 on $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$2, multimodality causes exponential-in-$$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$3 slow mixing without annealing, whereas a simulated annealing schedule in $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$4 yields exponential-in-time convergence at each temperature with a polynomially bounded overall mixing time. In unimodal landscapes the base chain already mixes in $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$5, while in multimodal settings an annealing schedule reduces the total complexity to $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$6 or, with varying step size, to $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$7 (Jonasson et al., 2021). Reweighted ALPS extends this algorithmic use: it combines cold-to-warm tempering, teleportation at the coldest level, and dynamic reweighting to maintain component and level balance, and proves polynomial total-variation mixing under warm-start tilt-dominance and local Poincaré assumptions (Lee et al., 19 Dec 2025).

These sampling papers use “annealed” in the temperature-schedule sense rather than the expectation-over-randomness sense that dominates the random-dynamics and passive-scalar literature. The shared term does not imply a shared definition (Jonasson et al., 2021, Lee et al., 19 Dec 2025).

6. Mechanisms, misconceptions, and structural themes

Several mechanisms recur across the literature. Alternating shears generate transverse stretching and filamentation; random Hamiltonian splitting produces positive Lyapunov exponents and controllability; Brownian motion of a cellular center creates a hypoelliptic two-point PDE; finite-mode stochastic Navier–Stokes forcing propagates randomness through a Hörmander-type bracket structure; random interval diffeomorphisms mix through exponential synchronization in average; and dynamical percolation uses conductance growth of evolving sets via giant-cluster isoperimetry (Cheng et al., 2021, Zhang, 13 Feb 2025, Navarro-Fernández et al., 24 Feb 2025, Cooperman et al., 2024, Czudek, 23 Apr 2025, Peres et al., 2017).

One misconception is that randomness always accelerates mixing. The wedge-shear computations show the opposite in the exponentially mixing regime: randomized phases and randomized times are consistently slightly slower than the best deterministic periodic schedule with the same base $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$8, especially near the optimal $$\|f(t)\|_{\mathcal H^1}\lesssim e^{-\lambda t}\|f(0)\|_{\mathcal H^1}$$9 or $\epsilon(k)$00 (Cheng et al., 2021). Another misconception is that diffusion is necessary for exponential decay. Several results are explicitly inviscid or $\epsilon(k)$01-independent: the wedge-shear study sets $\epsilon(k)$02 throughout; the random cellular flow proves a deterministic exponential rate independent of $\epsilon(k)$03, including $\epsilon(k)$04; and the stochastic-Navier–Stokes scalar results focus on pure transport (Cheng et al., 2021, Navarro-Fernández et al., 24 Feb 2025, Cooperman et al., 2024).

A further misconception is that “exponential mixing” refers to one canonical norm. In fact, the surveyed papers use a grid-adapted geometric mixing scale, Bressan’s geometric scale, $\epsilon(k)$05 or $\epsilon(k)$06 norms, weighted two-point correlations, Wasserstein contraction of semigroups, and total variation contraction across blocks. This suggests that annealed exponential mixing is best regarded as a structural principle rather than a single formula: randomness, when combined with an appropriate coupling, commutator, synchronization, or conductance mechanism, yields an exponential spectral-gap-like decay for an averaged mixing observable (Navarro-Fernández et al., 24 Feb 2025, Cooperman et al., 2024, DeWitt et al., 5 Sep 2025).

Open directions are likewise model-dependent. The Hamiltonian-shear theory relies crucially on analyticity and bounded switching times, and extending it to lower regularity remains open (Zhang, 13 Feb 2025). The finite-mode stochastic Navier–Stokes analysis focuses on $\epsilon(k)$07, with extension to uniform-in-$\epsilon(k)$08 advection–diffusion left for future work (Cooperman et al., 2024). The abstract equivalence theory is established in the IID conservative setting, and extension to non-IID bases is identified as a natural next step (DeWitt et al., 5 Sep 2025).

Across these settings, annealed exponential mixing functions as a bridge between randomness and deterministic decay laws. Whether expressed through a two-point PDE, a Wasserstein spectral gap, a geometric mixing scale, synchronization in average, or block-time total-variation contraction, it formalizes the emergence of exponential loss of coherence after averaging over the random environment.