Random Averaging Operator Ansatz

• Random Averaging Operator Ansatz is a framework that models metastable dynamics by approximating the infinite-dimensional Perron–Frobenius cocycle with an averaged finite-state Markov operator.
• It shows that the random invariant density converges to a convex combination of deterministic invariant densities, with weights determined by averaged escape rates and the limit of the second Oseledets space.
• The approach extends to systems with multiple metastable states, reducing complex transfer operators to effective lower-dimensional Markov dynamics for clearer analysis.

Random Averaging Operator Ansatz denotes, in the setting of random metastable dynamics, the statement that the infinite-dimensional Perron–Frobenius cocycle acts, on the metastable decomposition, like an averaged finite-state Markov operator, so that the quenched random invariant density is approximated by an averaged linear combination of deterministic invariant densities, with weights obtained from the stationary law of a Markov chain that emerges from the random transfer operator cocycle (González-Tokman et al., 2024). In “Averaging for random metastable systems,” this ansatz is developed for one-dimensional piecewise smooth expanding interval maps with two initially invariant subintervals, and it identifies both the small-perturbation limit of the random absolutely continuous invariant measure and the limit of the second Oseledets space, or coherent structure, as the perturbation shrinks to zero (González-Tokman et al., 2024).

1. Metastable random maps and transfer-operator cocycles

The basic deterministic object is a one-dimensional, piecewise $C^2$, uniformly expanding map

$T^0:I\to I,\qquad I=[-1,1],$

with a critical partition $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$ such that each branch $T^0|_{(c_i,c_{i+1})}$ extends to a $C^2$ function on a neighbourhood of $[c_i,c_{i+1}]$, and

$\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$

A boundary point $b\in(-1,1)$ determines two invariant subintervals

$I_L=[-1,b],\qquad I_R=[b,1],$

in the sense that $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$ for $T^0:I\to I,\qquad I=[-1,1],$0. Each restricted map $T^0:I\to I,\qquad I=[-1,1],$1 is a piecewise expanding Lasota–Yorke map and has a unique ergodic absolutely continuous invariant measure $T^0:I\to I,\qquad I=[-1,1],$2 with density

$T^0:I\to I,\qquad I=[-1,1],$3

Accordingly, any ACIM of $T^0:I\to I,\qquad I=[-1,1],$4 is a convex combination of $T^0:I\to I,\qquad I=[-1,1],$5 and $T^0:I\to I,\qquad I=[-1,1],$6 (González-Tokman et al., 2024).

The metastable interface is encoded by the infinitesimal holes

$T^0:I\to I,\qquad I=[-1,1],$7

The assumptions ensure that $T^0:I\to I,\qquad I=[-1,1],$8 and $T^0:I\to I,\qquad I=[-1,1],$9 are continuous and strictly positive at the points in $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$0 and $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$1. This regularizes the eventual leakage mechanism between the two deterministic components.

Randomness is introduced through a semi-invertible random dynamical system

$\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$2

where $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$3 is an ergodic, invertible, measure-preserving transformation, and $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$4 is a small $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$5-perturbation of $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$6. For each fibre $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$7, the perturbation creates random holes

$\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$8

which destroy the invariance of $\mathcal{C}^0=\{-1=c_0<c_1<\dots<c_d=1\}$9 and $T^0|_{(c_i,c_{i+1})}$0. These holes converge, in the Hausdorff metric, to $T^0|_{(c_i,c_{i+1})}$1, uniformly in $T^0|_{(c_i,c_{i+1})}$2 outside a $T^0|_{(c_i,c_{i+1})}$3-null set, and their sizes scale linearly in $T^0|_{(c_i,c_{i+1})}$4: $T^0|_{(c_i,c_{i+1})}$5 The coefficients $T^0|_{(c_i,c_{i+1})}$6 and $T^0|_{(c_i,c_{i+1})}$7 are the escape rates from the two metastable sets.

Associated with each map $T^0|_{(c_i,c_{i+1})}$8 is the Perron–Frobenius operator

$T^0|_{(c_i,c_{i+1})}$9

and the cocycle

$C^2$0

Under the stated assumptions there is a uniform Lasota–Yorke inequality in $C^2$1, $C^2$2 is $C^2$3-continuous in operator norm, and for $C^2$4 the system admits a unique random absolutely continuous invariant measure $C^2$5 with density $C^2$6 satisfying

$C^2$7

(González-Tokman et al., 2024).

2. Convex-combination limit for the invariant density

Before perturbation, the transfer operator $C^2$8 has a two-dimensional top Oseledets space spanned by $C^2$9 and $[c_i,c_{i+1}]$0. These two densities describe the two metastable states of the deterministic system. After perturbation, the top Lyapunov exponent remains $[c_i,c_{i+1}]$1 but becomes simple, the corresponding top Oseledets space becomes one-dimensional, and it is spanned by the random invariant density $[c_i,c_{i+1}]$2. The second Lyapunov exponent $[c_i,c_{i+1}]$3 is simple, its Oseledets space is one-dimensional, and as $[c_i,c_{i+1}]$4 the plane $[c_i,c_{i+1}]$5 converges to $[c_i,c_{i+1}]$6 (González-Tokman et al., 2024).

Any accumulation point of $[c_i,c_{i+1}]$7 therefore lies in $[c_i,c_{i+1}]$8, so

$[c_i,c_{i+1}]$9

for some $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$0. The central assertion of the ansatz is that these limiting weights are non-random and are determined by averaged escape rates. In the two-state case, the limiting density is

$\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$1

uniformly in $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$2, outside a $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$3-null set, provided

$\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$4

(González-Tokman et al., 2024).

Equivalently, the asymptotic mixture coefficients are

$\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$5

The constants do not depend on $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$6; randomness in the limit affects fluctuations at finite $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$7, not the asymptotic mixture weights.

The proof uses a law of large numbers type argument applied to products

$\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$8

and sums involving $\inf_{x\in I\setminus\mathcal{C}^0}|(T^0)'(x)|>1.$9, together with moving-average ergodic theorems. The key limit is

$b\in(-1,1)$0

which defines the limiting probability of being in $b\in(-1,1)$1 (González-Tokman et al., 2024).

3. Operator-theoretic interpretation and coherent structures

In the random setting, the relevant spectral picture is expressed in terms of Lyapunov exponents and Oseledets spaces rather than isolated eigenvalues and eigenfunctions. The cocycle $b\in(-1,1)$2 is quasi-compact, with

$b\in(-1,1)$3

the top Oseledets space $b\in(-1,1)$4 is spanned by $b\in(-1,1)$5, and the second Oseledets space $b\in(-1,1)$6 is spanned by $b\in(-1,1)$7, a coherent structure that decays at rate $b\in(-1,1)$8 (González-Tokman et al., 2024).

The ansatz becomes especially transparent after reducing the transfer-operator dynamics to an effective metastable Markov description. On the level of metastable sets, the dynamics induces a 2-state Markov chain in random environment with transition matrices

$b\in(-1,1)$9

As $I_L=[-1,b],\qquad I_R=[b,1],$0, the random invariant measure of this chain converges to

$I_L=[-1,b],\qquad I_R=[b,1],$1

Hence the infinite-dimensional random cocycle behaves, on the metastable decomposition, like an effective $I_L=[-1,b],\qquad I_R=[b,1],$2 averaging operator whose stationary vector determines the convex combination

$I_L=[-1,b],\qquad I_R=[b,1],$3

(González-Tokman et al., 2024).

The second Oseledets space records the slow exchange between the two metastable components. Choosing the sign so that $I_L=[-1,b],\qquad I_R=[b,1],$4, one has

$I_L=[-1,b],\qquad I_R=[b,1],$5

for $I_L=[-1,b],\qquad I_R=[b,1],$6-almost every $I_L=[-1,b],\qquad I_R=[b,1],$7. The limiting coherent structure is therefore a signed combination that distinguishes the two metastable regions, positive on $I_L=[-1,b],\qquad I_R=[b,1],$8, negative on $I_L=[-1,b],\qquad I_R=[b,1],$9, normalized to have zero integral and unit $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$0-norm. In the metastable interpretation, the slow mode is precisely the difference between the two metastable densities (González-Tokman et al., 2024).

4. Random paired tent maps and the extension to $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$1 metastable sets

The theory is applied to random paired tent maps

$(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$2

defined piecewise by

$(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$3

When $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$4, the map consists of two disjoint tent maps on $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$5 and $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$6. For small positive $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$7 and $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$8, there is leakage between the two halves. With measurable $(T^0|_{I_\star})^{-1}(I_\star)\subset I_\star$9, the random perturbation is

$T^0:I\to I,\qquad I=[-1,1],$00

The holes are explicitly

$T^0:I\to I,\qquad I=[-1,1],$01

Since the unperturbed ACIMs are Lebesgue restricted to each half,

$T^0:I\to I,\qquad I=[-1,1],$02

so $T^0:I\to I,\qquad I=[-1,1],$03 and $T^0:I\to I,\qquad I=[-1,1],$04. The invariant density therefore satisfies

$T^0:I\to I,\qquad I=[-1,1],$05

and the second Oseledets vector converges to

$T^0:I\to I,\qquad I=[-1,1],$06

(González-Tokman et al., 2024).

The same mechanism extends to $T^0:I\to I,\qquad I=[-1,1],$07 initially invariant intervals $T^0:I\to I,\qquad I=[-1,1],$08, each supporting a unique ACIM with density $T^0:I\to I,\qquad I=[-1,1],$09. For neighbouring intervals, the holes are

$T^0:I\to I,\qquad I=[-1,1],$10

with

$T^0:I\to I,\qquad I=[-1,1],$11

or zero when there is no direct transition. The induced random Markov chain has transition matrices

$T^0:I\to I,\qquad I=[-1,1],$12

where $T^0:I\to I,\qquad I=[-1,1],$13 is diagonal with rates out of each state and $T^0:I\to I,\qquad I=[-1,1],$14 collects off-diagonal transition rates. If $T^0:I\to I,\qquad I=[-1,1],$15 is invertible, then

$T^0:I\to I,\qquad I=[-1,1],$16

with coefficients $T^0:I\to I,\qquad I=[-1,1],$17, $T^0:I\to I,\qquad I=[-1,1],$18, independent of $T^0:I\to I,\qquad I=[-1,1],$19, and the coefficient vector $T^0:I\to I,\qquad I=[-1,1],$20 solves

$T^0:I\to I,\qquad I=[-1,1],$21

with normalization $T^0:I\to I,\qquad I=[-1,1],$22. The limiting coefficients are therefore the stationary vector of an effective averaged $T^0:I\to I,\qquad I=[-1,1],$23 Markov operator on metastable states (González-Tokman et al., 2024).

5. Conditions, scope, and limitations

The ansatz is proved under a rigid perturbative and operator-theoretic framework. The dynamics on each branch is piecewise $T^0:I\to I,\qquad I=[-1,1],$24 and uniformly expanding; the unperturbed system has two, or more generally $T^0:I\to I,\qquad I=[-1,1],$25, initially invariant sets with unique ergodic ACIMs and no further splitting; the random perturbations are small in $T^0:I\to I,\qquad I=[-1,1],$26; the base dynamics $T^0:I\to I,\qquad I=[-1,1],$27 is ergodic and invertible; and the transfer operators satisfy uniform Lasota–Yorke inequalities together with the continuity assumptions needed for the Oseledets splitting (González-Tokman et al., 2024).

Equally important are the regularity hypotheses at the leakage interface. The infinitesimal holes must be regular enough that the unperturbed densities are continuous and strictly positive there, and the boundary condition ensures that leakage occurs away from the boundary point $T^0:I\to I,\qquad I=[-1,1],$28, so that the geometry is stable. The hole sizes must admit linear expansions in $T^0:I\to I,\qquad I=[-1,1],$29, because the averaged escape rates are exactly the coefficients that survive in the limiting convex combination.

The mechanism is therefore not a generic averaging principle for arbitrary random dynamical systems. It is an asymptotic statement for random metastable systems with rare communication between initially invariant components. If

$T^0:I\to I,\qquad I=[-1,1],$30

then finer control of error terms is needed and the convex combination may degenerate to a single metastable component. Within the stated assumptions, however, the theory gives a fully rigorous justification of the claim that random metastable dynamics can be understood through a random averaging operator acting on the invariant densities of the unperturbed system (González-Tokman et al., 2024).

6. Broader operator-averaging patterns

This suggests a broader operator-theoretic pattern in which a random evolution is replaced, after averaging or after many weak random interactions, by an effective deterministic or finite-rank operator. In “Self–averaging of random quantum dynamics,” many independent random sudden quenches on a finite-dimensional Hilbert space produce a random unitary evolution whose Frobenius-norm variance scales as $T^0:I\to I,\qquad I=[-1,1],$31, and, for protocols that commute in the statistical sense, the averaged unitary converges to the unitary generated by the averaged Hamiltonian (Łobejko et al., 2018).

A different version appears in random iterations of $T^0:I\to I,\qquad I=[-1,1],$32-averaged operators on Hilbert space. There the residual recursion

$T^0:I\to I,\qquad I=[-1,1],$33

is controlled by the single geometric parameter

$T^0:I\to I,\qquad I=[-1,1],$34

yielding exponential mean-square decay, almost-sure convergence, and random nonlinear fusion frames with exact synthesis and frame-type energy bounds in expectation (Tian, 12 Sep 2025).

In consensus and distributed averaging, the random operator is a stochastic matrix. “On Endogenous Random Consensus and Averaging Dynamics” studies

$T^0:I\to I,\qquad I=[-1,1],$35

for balanced adapted random stochastic matrices with uniformly positive diagonal and proves almost sure convergence together with a limiting clustering law governed by the infinite flow graph (Touri et al., 2014). “A Random Adaptation Perspective on Distributed Averaging” studies a random adaptation process

$T^0:I\to I,\qquad I=[-1,1],$36

with $T^0:I\to I,\qquad I=[-1,1],$37, and shows that ergodicity of the deterministic chain $T^0:I\to I,\qquad I=[-1,1],$38 is equivalent to almost sure finite-time agreement attainment in the random adaptation dynamics (Parasnis et al., 2022).

In random unitary circuits with unitary-invariant gate distributions, the ensemble-averaged Pauli-string weights

$T^0:I\to I,\qquad I=[-1,1],$39

obey a classical Markovian evolution on Pauli strings, and the long-time operator-spreading front is governed by a drift–diffusion equation characterized by the butterfly velocity $T^0:I\to I,\qquad I=[-1,1],$40 and diffusion constant $T^0:I\to I,\qquad I=[-1,1],$41. Relative to the Haar case, general unitary-invariant ensembles introduce a finite binary time $T^0:I\to I,\qquad I=[-1,1],$42 and a finite domain-wall width $T^0:I\to I,\qquad I=[-1,1],$43 (Tan et al., 7 Jan 2025).

The phrase is also used explicitly in probabilistic PDE. In “Gauge transforms, random averaging operator ansatz and improved probabilistic well-posedness for the radial NLS on the $T^0:I\to I,\qquad I=[-1,1],$44 ball,” the rough high-frequency part of the solution is represented by dyadic Gaussian blocks multiplied by random time-dependent phases $T^0:I\to I,\qquad I=[-1,1],$45, and this frequency-by-frequency random averaging structure is combined with a gauge transform and refined modulation analysis to construct probabilistic strong solutions for the cubic radial NLS on the three-dimensional ball in a supercritical probabilistic regime (Burq et al., 5 Jun 2026).

These uses do not define a single universal theorem. They do, however, exhibit the same structural move: random microscopic dynamics is compressed into an averaged operator, or into averaged coefficients on a reduced state space, and the effective evolution is then analyzed through the resulting deterministic or lower-dimensional object.