Frozen Turbulence Approach

• Frozen Turbulence Approach is a family of asymptotic modeling strategies that treats turbulent fields as approximately unchanged over short time scales, enabling reduction of complex spatio-temporal problems into translational models.
• It underpins diverse applications such as adaptive optics, aeroacoustics, particulate collisions, and MHD by incorporating corrections for decorrelation and boiling when pure frozen flow fails.
• Mathematically, the approach exploits time-scale separation by assuming advection times are shorter than intrinsic turbulence evolution, facilitating state-space prediction and causal isolation of physical mechanisms.

Searching arXiv for recent and relevant papers on the frozen turbulence approach across major application areas. Across the cited literature, the frozen turbulence approach denotes a family of asymptotic modeling strategies in which a turbulent field is treated as approximately unchanged over the time scale of advection, sensing, or local interaction. In its classical form, this is Taylor’s frozen-flow or frozen-in hypothesis: the structure is transported by a mean velocity, and the dominant dynamics are translational rather than intrinsic. In later developments, the same idea appears as quasi-steady local closure, frozen-in transport of divergence-free fields, or controlled freezing of a base flow to isolate causal mechanisms. The approach is therefore not a single formalism but a recurrent reduction principle, used in adaptive optics, aeroacoustics, particulate collisions, magnetohydrodynamics, reconnection physics, and turbulence dynamics, often with explicit correction terms for decorrelation or “boiling” when perfect freezing fails (Guesalaga et al., 2014, Srinath et al., 2015, Tian et al., 2023, Jiang et al., 2024, Verma, 2022).

1. Canonical hypothesis and spectral interpretation

In the classical trailing-edge-noise formulation, frozen turbulence means that the wall-pressure field convects downstream without changing shape: $Q_{pp}(\xi,\eta,\tau)=Q_{pp}(\xi-U_c\tau,\eta,0),$ so that the streamwise coherence is identically unity,

$\gamma(\xi,0,\omega)=1,$

and the streamwise cross-spectral content collapses to

$\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$

This is the analytical simplification used in frozen versions of serrated-edge models: for a given frequency, only the single convective wavenumber contributes (Tian et al., 2023).

The same reduction underlies Taylor’s hypothesis in hydrodynamics and its MHD generalization. In hydrodynamics, a strong mean flow $\mathbf{U}_0$ advects turbulent fluctuations past a probe sufficiently rapidly that temporal variation is dominated by sweeping rather than local evolution. In the MHD extension, the natural variables are the Elsässer fields $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$, and the effective advection speeds become

$\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$

The derived frequency spectra are then

$$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$

for a Kolmogorov-like phenomenology, and

$$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$

for the Iroshnikov–Kraichnan case, with the analogous minus-field expression obtained by replacing $|\mathbf{U}_0-\mathbf{B}_0|$ by $|\mathbf{U}_0+\mathbf{B}_0|$ (Verma, 2022).

These formulations all rely on the same asymptotic ordering: the advection time past the observer or interaction region is shorter than the intrinsic evolution time of the turbulence. This suggests that the essential role of the frozen approximation is to convert a spatio-temporal problem into a translational one with a reduced spectral support.

2. Mathematical realizations and controlled departures from perfect freezing

In wide-field adaptive optics, the frozen approximation is implemented through spatio-temporal cross-correlations of Shack–Hartmann slopes. For two wavefront sensors, $\gamma(\xi,0,\omega)=1,$0 and $\gamma(\xi,0,\omega)=1,$1, the time-delayed cross-correlation is

$\gamma(\xi,0,\omega)=1,$2

with overlap compensation and deconvolution by the average autocorrelation,

$\gamma(\xi,0,\omega)=1,$3

The deconvolved estimator is preferred because it is more robust when the ground layer is contaminated by dome seeing, which is often non-Kolmogorov. Real data do not exhibit purely translating peaks: the peaks broaden and lose amplitude, so the phase evolution is modeled as

$\gamma(\xi,0,\omega)=1,$4

where $\gamma(\xi,0,\omega)=1,$5 is a decay rate and $\gamma(\xi,0,\omega)=1,$6 is a colored-noise term representing residual non-frozen effects such as shear or buoyancy (Guesalaga et al., 2014).

A closely related synthesis appears in autoregressive phase-screen generation for AO simulation. Each Fourier mode evolves according to

$\gamma(\xi,0,\omega)=1,$7

where the first term carries forward frozen-flow memory and the second injects stochastic boiling. Wind translation is encoded in the phase of $\gamma(\xi,0,\omega)=1,$8,

$\gamma(\xi,0,\omega)=1,$9

while $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$0 partitions power between deterministic transport and stochastic renewal. The limiting cases are explicit: $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$1 gives pure frozen flow, and smaller $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$2 gives more stochasticity. The same construction also reduces the practical impact of periodicity in translating Fourier screens because the added noise progressively destroys memory of the initial state; for example, with $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$3, after 500 steps only $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$4 of the original phase remains (Srinath et al., 2015).

These developments preserve the core frozen-flow kinematics while replacing strict immutability by controlled temporal decorrelation. A plausible implication is that, in modern usage, frozen turbulence is often a leading-order transport model supplemented by a low-dimensional relaxation law.

3. Wide-field adaptive optics and predictive control

The adaptive-optics literature treats frozen flow as both a physical hypothesis and an operational profiling tool. Using GeMS telemetry, a SLODAR-like triangulation based on Shack–Hartmann slope cross-correlations retrieves the number of turbulence layers, their associated velocities, altitudes, strengths, and dome-seeing contribution. At zero delay, peaks in the cross-correlation map identify turbulent layers; as $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$5 increases, tracking the moving peaks yields layer altitude, $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$6 strength, wind speed, and wind direction. Reported examples include a dome or ground component at zero or near-zero speed, a ground layer moving at $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$7, a layer around $\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$8–$\phi_x(k_1,\omega)=\delta(k_1-\omega/U_c).$9 moving at $\mathbf{U}_0$0, and a jet-stream layer around $\mathbf{U}_0$1–$\mathbf{U}_0$2 moving at $\mathbf{U}_0$3. By combining multiple LGS pairings with different altitude resolutions, the profiler can build extended profiles up to more than $\mathbf{U}_0$4. The same analysis shows that the decorrelation rate can be characterized by a single parameter that is independent of altitude and turbulence strength but dependent on wind speed, with a linear relation between decay and layer speed. Typical reported decay rates are about $\mathbf{U}_0$5 for dome seeing, $\mathbf{U}_0$6 for a ground layer at $\mathbf{U}_0$7, $\mathbf{U}_0$8 for a mid-altitude layer at $\mathbf{U}_0$9, and $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$0 for a jet-stream layer at $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$1. Over a typical 2-frame delay of about $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$2, the correlation remains above $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$3, and for common $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$4 layers it remains around $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$5, which is why predictive control remains viable despite measurable boiling (Guesalaga et al., 2014).

That operational use of frozen flow was subsequently turned into controller design. “Predictive Fourier Control” implements an LQG or Kalman-filter architecture that identifies frozen-flow peaks in Fourier-mode temporal PSDs and predicts the disturbance one or more steps ahead. The method uses a dual-loop architecture: a fast AO loop for sensing and correction, and a slow analysis loop operating on roughly $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$6–$\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$7 second telemetry blocks for Fourier Wind Identification and model update. In the reported laboratory verification, PSD peaks near $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$8, $\mathbf{z}^{\pm}=\mathbf{u}\pm\mathbf{b}$9, and $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$0 timesteps$\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$1 were consistent with a simulated wind of $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$2 subapertures/timestep and $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$3, and $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$4 of detectable modes showed evidence of frozen flow. The predictive controller reduced RMS residual wavefront error from $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$5 to $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$6 and improved measured Strehl from $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$7 to $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$8, with the time-delay term in the error budget reduced from $\mathbf{Z}^{\pm}=\mathbf{U}_0 \mp \mathbf{B}_0.$9 to $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$0 (Rudy et al., 2015).

In this setting, the frozen turbulence approach is neither merely diagnostic nor merely philosophical. It is the basis for estimating $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$1 and wind profiles, quantifying decorrelation, and constructing state-space predictors that target servo-lag directly.

4. Quasi-steady encounter models in particle collisions and aeroacoustics

In bubble–particle collisions, the frozen turbulence approach is a local quasi-steady closure. During a bubble–particle encounter, the flow field near the bubble is assumed to be approximately stationary and uniform; the interaction time

$$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$2

is assumed shorter than the correlation time of turbulence fluctuations, $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$3, and the flow correlation length is of order $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$4, comparable to or larger than the bubble size. The turbulent collision kernel is then written as an ensemble average over instantaneous quiescent-flow problems: $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$5 Here the hydrodynamic collision efficiency remains the deterministic quiescent-flow quantity $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$6, while turbulence enters through the fluctuating instantaneous slip velocity $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$7 and the local particle density $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$8. The dominant mechanism is the nonlinear dependence of $$E^{+}(f)\propto \left|\mathbf{U}_0-\mathbf{B}_0\right|^{2/3} f^{-5/3}, \qquad E^{-}(f)\propto \left|\mathbf{U}_0+\mathbf{B}_0\right|^{2/3} f^{-5/3}$$9 on $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$0 and $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$1, not direct random forcing of the particles. Reported consequences include about a $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$2 increase in $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$3 from the Reynolds-number effect in the present case, and collision-kernel enhancement up to about $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$4 relative to quiescent flow, with the largest enhancement near $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$5, especially for $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$6. For freely rising bubbles, an additional offset $$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$7 accounts for slip-velocity fluctuations during the encounter (Jiang et al., 2024).

A later predictive formulation preserves the same logic but replaces the slip-velocity PDF by an a priori model built from bubble, particle, and turbulence properties. The collision kernel becomes

$$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$8

with validity assessed by the criterion

$$E^{+}(f)=A_{IK}\,(\varepsilon B_0 |\mathbf{U}_0-\mathbf{B}_0|)^{1/2} f^{-3/2}$$9

Good agreement with simulations is reported for Froude number $|\mathbf{U}_0-\mathbf{B}_0|$0, and the model indicates that smaller bubbles, larger particles, and stronger turbulence increase the overall collision rate (Chan et al., 18 Jul 2025).

In serrated trailing-edge noise prediction, by contrast, the frozen assumption is shown to be too strong. LES of a fully developed turbulent boundary layer demonstrates that wall-pressure fluctuations lose coherence in the streamwise direction; the coherence is not equal to unity, the space-time correlation contours are elliptic or narrow-band rather than straight, and for $|\mathbf{U}_0-\mathbf{B}_0|$1 coherence decays quickly with separation. The relaxed model replaces $|\mathbf{U}_0-\mathbf{B}_0|$2 by

$|\mathbf{U}_0-\mathbf{B}_0|$3

so that the streamwise spectral factor becomes the Lorentzian

$|\mathbf{U}_0-\mathbf{B}_0|$4

centered at $|\mathbf{U}_0-\mathbf{B}_0|$5. A correction coefficient $|\mathbf{U}_0-\mathbf{B}_0|$6 then shifts the effective wavenumber in the far-field formula. The decisive control parameter is

$|\mathbf{U}_0-\mathbf{B}_0|$7

the ratio of serration half-amplitude to streamwise frequency-dependent correlation length. When $|\mathbf{U}_0-\mathbf{B}_0|$8, the frozen model is closer to valid; when $|\mathbf{U}_0-\mathbf{B}_0|$9, it overpredicts destructive interference and therefore overpredicts noise reduction. The paper reports that, in many serrated trailing-edge cases, the serration amplitude is $|\mathbf{U}_0+\mathbf{B}_0|$0–$|\mathbf{U}_0+\mathbf{B}_0|$1 times larger than the streamwise correlation length, which explains the discrepancy with experiment (Tian et al., 2023).

The contrast between these two domains is instructive. In bubble collisions, freezing is justified by a short interaction time and local hydrodynamic dominance. In trailing-edge acoustics, the same simplification fails because the relevant observable depends directly on streamwise phase coherence, which real boundary-layer turbulence loses.

5. Frozen-in fields, reconnection transport, and frozen-base-flow causality tests

A distinct use of the frozen turbulence approach appears in systems where the key quantity is a frozen-in field rather than a convected scalar pattern. In freely decaying two-dimensional hydrodynamic turbulence, the transported field is the di-vorticity

$|\mathbf{U}_0+\mathbf{B}_0|$2

which satisfies a frozen-in-fluid equation

$|\mathbf{U}_0+\mathbf{B}_0|$3

Although the carrier flow is incompressible, the mapping associated with motion normal to $|\mathbf{U}_0+\mathbf{B}_0|$4 is compressible, because

$|\mathbf{U}_0+\mathbf{B}_0|$5

so $|\mathbf{U}_0+\mathbf{B}_0|$6 can shrink strongly and even approach zero. The resulting concentration of $|\mathbf{U}_0+\mathbf{B}_0|$7 into narrow ribbons is interpreted as folding of a divergence-free frozen field. Numerically, the maximum di-vorticity grows as $|\mathbf{U}_0+\mathbf{B}_0|$8 with rate about $|\mathbf{U}_0+\mathbf{B}_0|$9, the transverse thickness decays as $\gamma(\xi,0,\omega)=1,$00 with rate about $\gamma(\xi,0,\omega)=1,$01, the longitudinal scale grows at about $\gamma(\xi,0,\omega)=1,$02, and the amplitude-thickness relation follows

$\gamma(\xi,0,\omega)=1,$03

This $\gamma(\xi,0,\omega)=1,$04 law is interpreted not as a Kolmogorov inertial-range relation but as evidence of folding for a frozen divergence-free field (Kuznetsov et al., 2018).

In guide-field reconnection at the magnetopause, the notion of freezing concerns electrons rather than eddy morphology. Along the separatrices, a variant of the lower hybrid drift instability produces electric-field fluctuations much larger than the reconnection electric field, yet electrons remain magnetized or frozen-in to the fluctuating fields because the turbulence is low frequency compared with the electron gyrofrequency and the electron $\gamma(\xi,0,\omega)=1,$05 drift is usually too small for resonance. The perpendicular particle flux across $\gamma(\xi,0,\omega)=1,$06 is

$\gamma(\xi,0,\omega)=1,$07

and the turbulent flux is large enough to balance the laminar inflow,

$\gamma(\xi,0,\omega)=1,$08

At the same time, in a frozen-in region the turbulent electric and motional terms nearly cancel,

$\gamma(\xi,0,\omega)=1,$09

Fluctuation amplitudes around $\gamma(\xi,0,\omega)=1,$10 can therefore coexist with an almost frozen-in electron response. Near the X-line, however, electrons are not frozen-in, classic LHDI is stabilized by magnetic shear, and the turbulence makes a significant net contribution to the generalized Ohm’s law through anomalous viscosity in a region extending only about $\gamma(\xi,0,\omega)=1,$11 downstream (Price et al., 2019).

A third variant uses freezing as a controlled intervention rather than as a physical approximation. In DNS of wall turbulence, the flow is decomposed as

$\gamma(\xi,0,\omega)=1,$12

with $\gamma(\xi,0,\omega)=1,$13 given by the streamwise average. Freezing $\gamma(\xi,0,\omega)=1,$14 at an instant $\gamma(\xi,0,\omega)=1,$15,

$\gamma(\xi,0,\omega)=1,$16

removes feedback from $\gamma(\xi,0,\omega)=1,$17 to $\gamma(\xi,0,\omega)=1,$18 and allows causal testing of linear energy-injection mechanisms. After stabilizing the operator to remove exponential instabilities and freezing the base flow to remove parametric instabilities, more than 500 frozen-base-flow simulations show that transient growth alone is sufficient to sustain realistic wall turbulence. Further interventions show that removing lift-up still permits sustained turbulence, whereas removing push-over or Orr mechanisms causes the turbulence to decay quickly. The key sustaining mechanism is therefore identified as the Orr/push-over route associated with spanwise variations of the streak (Lozano-Durán et al., 2020).

These examples broaden the meaning of frozen turbulence well beyond Taylor advection. The common element is a transport or amplification mechanism that remains analyzable when the carrier structure is treated as frozen, even though the observable consequences differ sharply across domains.

6. Breakdown, alternative meanings, and scope of the term

The literature also shows that “frozen turbulence” is not semantically uniform. In surface gravity waves on a finite periodic domain, a frozen turbulence state does not refer to frozen advection at all; it means that exact-resonant energy transfer stops before reaching the boundary of the discrete wavenumber set $\gamma(\xi,0,\omega)=1,$19. For a direct cascade with $\gamma(\xi,0,\omega)=1,$20, the paper reports that $\gamma(\xi,0,\omega)=1,$21 yields $\gamma(\xi,0,\omega)=1,$22, $\gamma(\xi,0,\omega)=1,$23 yields $\gamma(\xi,0,\omega)=1,$24, and $\gamma(\xi,0,\omega)=1,$25 reaches $\gamma(\xi,0,\omega)=1,$26, with a sharp transition around $\gamma(\xi,0,\omega)=1,$27. In this usage, frozen turbulence denotes an arrested cascade in a finite resonant network rather than a frozen-flow closure (Zhang et al., 2022).

Neighboring methods can also resemble frozen-turbulence reasoning without actually assuming frozen turbulence. In reduced-order snow-transport modeling, particle motion is driven by externally supplied DNS or LES velocity fields $\gamma(\xi,0,\omega)=1,$28, and the reduced model writes

$\gamma(\xi,0,\omega)=1,$29

The turbulence is externally supplied and time-varying, and the method is explicitly described as better characterized by externally forced inertial-particle transport in prescribed unsteady turbulence than by a frozen-flow closure (Aksamit et al., 6 May 2025). Likewise, field observations of snow settling interpret preferential sweeping using quasi-instantaneous vortex reconstructions, but the study does not present a formal frozen-turbulence model (Li et al., 2021).

Terminological confusion is reinforced by unrelated uses of “frozen.” “Frozen Waves” in optical trapping are structured non-diffracting beams obtained by superposing co-propagating Bessel beams and have no explicit connection to turbulence-based control (Suarez et al., 2020). A plausible implication is that the encyclopedia meaning of the frozen turbulence approach must be delimited by mechanism, not by the adjective “frozen” alone.

Taken together, these works show that the frozen turbulence approach is best understood as a hierarchy of reductions based on time-scale separation. It is powerful when advection, local encounter, or base-state persistence dominates intrinsic evolution; it becomes inadequate when streamwise coherence loss, boiling, instability-driven restructuring, or finite-network disconnection governs the observable. The most mature applications therefore retain the frozen approximation only as a leading-order model and then add explicit decorrelation, stochastic renewal, or validity criteria to delimit its range.