Mean Velocity Flow: Concepts & Applications

• Mean velocity flow is defined as the averaged transport component obtained via temporal, spatial, ensemble, or directional averaging to separate coherent motion from fluctuations.
• It appears in varied contexts—from wall turbulence and pipe flow scaling to thermal convection and machine learning—with each application dictating unique averaging techniques and profile characteristics.
• Advanced methods including Bayesian inversion and analytical reconstruction demonstrate its role in accurately modeling transport dynamics and guiding both theoretical and data-driven studies.

Searching arXiv for the cited papers and closely related "mean velocity flow" work to ground the article in the supplied literature. Mean velocity flow denotes the averaged or coherent transport component of a system’s motion, as distinguished from instantaneous fluctuations, stochastic dispersion, or microscale variability. Across the literature, the term refers to several non-equivalent objects: a long-time and space-averaged wall-parallel velocity in thermal convection (Samuel et al., 2024); the mean streamwise profile in canonical wall turbulence (Bretheim et al., 2014); a reconstructed time-averaged three-dimensional velocity field in inverse Reynolds-averaged Navier–Stokes (RANS) problems (Kontogiannis et al., 2024); a centerline mean velocity law in pipe flow (Nagib et al., 9 Nov 2025); an average velocity over a finite time interval in one-step generative modeling (Geng et al., 19 May 2025); and a mean-flow field constrained by wave dynamics, porous-media coupling, curvature, or rotation in other physical settings (Yang et al., 2015). Taken together, these uses show that mean velocity flow is not a single universal object but a family of averaged kinematic descriptions whose meaning depends on the averaging operator, geometry, governing equations, and the relation between coherent motion and fluctuations.

1. Averaging, definition, and conceptual scope

In fluid-mechanical usage, mean velocity flow is typically an averaged velocity field obtained from temporal, spatial, ensemble, or directional averaging. In plane Rayleigh–Bénard convection, the mean wall-parallel field is defined by a combined horizontal area and time average,

$$\langle u_i\rangle_{A,t}(z)=\frac{1}{A N_t}\sum_{k=1}^{N_t}\int_A u_i(x,y,z,t_k)\,dA, \qquad i=x,y,$$

with statistics gathered over at least $100$ free-fall times and $N_t=20$ statistically independent snapshots (Samuel et al., 2024). In half-channel turbulence modeled by the restricted nonlinear framework, the total velocity is decomposed as

$\mathbf{u}_T = \mathbf{U} + \mathbf{u},$

where $\mathbf{U}$ is the streamwise-averaged mean flow and $\mathbf{u}$ contains streamwise-varying perturbations (Bretheim et al., 2014). In inverse RANS, the inferred quantity is explicitly the time-averaged mean velocity field $\mathbf{u}$ of a turbulent jet (Kontogiannis et al., 2024).

The same phrase acquires a different formal meaning in generative modeling. MeanFlow defines an average velocity over a finite interval,

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$

contrasted with the instantaneous velocity field $v(z_t,t)$ used in Flow Matching (Geng et al., 19 May 2025). In reinforcement learning, Mean Velocity Policy defines an analogous interval-average field for action generation,

$$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$

with the state $100$0 entering as conditioning (Zhan et al., 14 Feb 2026). These constructions are not Reynolds means; they are finite-time average displacements per unit time.

A plausible implication is that “mean velocity flow” functions less as a single term of art than as a structural motif: it denotes whichever reduced velocity descriptor is intended to capture organized transport after eliminating some class of unresolved variability.

2. Wall-bounded turbulence and canonical mean profiles

In canonical wall turbulence, mean velocity flow is often identified with the mean streamwise profile and its asymptotic structure. The band-limited restricted nonlinear model shows that the prediction of this mean profile depends strongly on the retained perturbation support. Baseline RNL sustains only a few streamwise-varying modes and produces a logarithmic region with nonstandard fitted constants, including $100$1 at $100$2 and $100$3 at $100$4, whereas carefully chosen band limitation can recover the standard logarithmic law

$100$5

including with a single perturbation mode $100$6 up to about $100$7 (Bretheim et al., 2014). The paper’s central claim is therefore that a very small number of streamwise-varying modes can sustain a realistic turbulent mean profile.

Pipe flow work addresses mean velocity flow through centerline scaling. In the CICLoPE facility, the normalized centerline mean velocity $100$8 follows

$100$9

leading to a centerline von Kármán constant $N_t=20$0 (Nagib et al., 9 Nov 2025). The same study argues that meaningful high-Reynolds-number behavior is only clearly reached when $N_t=20$1, and more convincingly beyond $N_t=20$2 (Nagib et al., 9 Nov 2025). This separates asymptotic mean-velocity scaling from lower-$N_t=20$3 transitional trends in centerline turbulence statistics.

A different pipe-flow theory derives an explicit outer mean velocity profile from Lie-group similarity. There the mean velocity defect is written

$N_t=20$4

with

$N_t=20$5

and the overlap asymptotic becomes

$N_t=20$6

That analysis identifies $N_t=20$7 as a universal Kármán constant across pipe and channel flows and reports $N_t=20$8 accuracy for high-$N_t=20$9 experimental data up to $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$0 (Chen et al., 2012). The coexistence of $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$1 for a centerline law and $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$2 for an outer-profile theory should not be conflated: the former pertains to a centerline logarithmic relation in $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$3, while the latter arises from an outer similarity description of the full mean velocity profile.

Analytical work on near-wall turbulent flow complicates any single-law picture. Using a dispersion-relation approach in two-dimensional channel flow, Kazakov derives families of asymptotic mean profiles including

$\mathbf{u}_T = \mathbf{U} + \mathbf{u},$4

with arbitrary $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$5 and, in general, arbitrary $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$6 (Kazakov, 2014). Under the sharpened assumption that the leading near-wall asymptotic is independent of correlations between near-wall and bulk fluctuations, the power-law family is ruled out, whereas the logarithmic and power-log families remain compatible, with $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$7 as $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$8 (Kazakov, 2014). This directly addresses a recurrent misconception: the logarithmic law may be asymptotically selected, but the mathematics does not render alternative local asymptotics impossible in principle.

3. Thermal convection and the absence of a sustained mean wind

In plane thermal convection, mean velocity flow becomes a question of whether a statistically persistent wall-parallel wind exists in the boundary region. Direct numerical simulations of three-dimensional Boussinesq Rayleigh–Bénard convection in a Cartesian slab with no-slip plates, horizontal periodic boundary conditions, and primarily aspect ratio $\mathbf{u}_T = \mathbf{U} + \mathbf{u},$9 show that such a sustained mean flow is practically absent over $\mathbf{U}$0 at $\mathbf{U}$1 (Samuel et al., 2024).

The defining evidence is cumulative. First, long-time and area-averaged mean wall-parallel profiles remain extremely small and continue drifting toward zero as averaging time increases, with the longest study at $\mathbf{U}$2 extended from $\mathbf{U}$3 to $\mathbf{U}$4 (Samuel et al., 2024). Second, near-wall Blasius fits yield only tiny effective free-stream velocities,

$\mathbf{U}$5

in units of the free-fall velocity, together with

$\mathbf{U}$6

even at the largest $\mathbf{U}$7 (Samuel et al., 2024). Third, the magnitude of the velocity fluctuations exceeds the mean by up to two orders of magnitude (Samuel et al., 2024).

The boundary region is therefore not described as a weakly fluctuating perturbation about a robust mean shear. Instead, it consists of differently oriented local shear-dominated coherent patches interspersed with shear-free incoherent regions. The incoherent regions occupy about $\mathbf{U}$8 of the area for all Rayleigh numbers studied and can span the whole cross section (Samuel et al., 2024). The thermal boundary layer remains identifiable through

$\mathbf{U}$9

but “no momentum boundary layer can be easily identified,” so the velocity field is characterized through rms profiles rather than a classical momentum boundary layer (Samuel et al., 2024). The velocity fluctuation thickness exceeds the thermal one by a ratio growing from roughly $\mathbf{u}$0 at $\mathbf{u}$1 to about $\mathbf{u}$2 at $\mathbf{u}$3 (Samuel et al., 2024).

This directly opposes an extrapolation from small-aspect-ratio convection. In closed cylindrical cells of aspect ratio $\mathbf{u}$4, a large-scale circulation can create a recognizable mean shear near the wall. In the horizontally periodic plane layer, the full-domain statistics are dominated by incoherent regions, and any apparent mean wind over short windows is interpreted as a residual of large, slowly varying fluctuations rather than a persistent wall-parallel flow (Samuel et al., 2024). A plausible implication is that local instability, plume formation, and heat transport in large-aspect-ratio convection should be analyzed through fluctuation-dominated, spatially heterogeneous near-wall dynamics rather than through a classical mean-shear paradigm.

4. Mean-flow reconstruction and inference from data

Mean velocity flow can also denote an inferred or reconstructed field rather than a directly measured average. A Bayesian inverse RANS framework assimilates mean flow MRI data to jointly reconstruct the mean velocity field and learn unknown RANS parameters in a turbulent jet through an FDA nozzle (Kontogiannis et al., 2024). The forward model solves

$\mathbf{u}$5

with no-slip, inlet Dirichlet, and outlet stress boundary conditions (Kontogiannis et al., 2024).

The inverse problem is posed through a Bayesian map

$\mathbf{u}$6

with Gaussian measurement noise and posterior

$\mathbf{u}$7

and a MAP estimate obtained by minimizing the negative log-posterior (Kontogiannis et al., 2024). The turbulence closure is a Boussinesq-type algebraic eddy viscosity,

$\mathbf{u}$8

with jointly learned parameters $\mathbf{u}$9 (Kontogiannis et al., 2024).

In the FDA nozzle experiment, the inversion learns the inlet condition and turbulence parameters while geometry and outlet conditions are fixed. The model is optimized with adjoint-based gradients, the MAP point is found in 48 iterations, and most error reduction occurs in the first 20 (Kontogiannis et al., 2024). Reported uncertainties from the Laplace approximation include $\mathbf{u}$0 mPa·s, $\mathbf{u}$1, $\mathbf{u}$2, $\mathbf{u}$3 cm, $\mathbf{u}$4 cm, and $\mathbf{u}$5 cm (Kontogiannis et al., 2024).

The experimental setting is a confined turbulent jet with 10 mm throat, sudden expansion to 3 cm diameter, flow rate 200 mL/s, Reynolds number $\mathbf{u}$6, and 4D flow MRI on a 3T scanner at isotropic resolution $\mathbf{u}$7 mm$\mathbf{u}$8 (Kontogiannis et al., 2024). The study reports that the MAP reconstruction reduces the data-model discrepancy for all velocity components and reproduces both the jet breakdown region and the pipe-flow region without overfitting (Kontogiannis et al., 2024). Here, mean velocity flow is a latent field inferred from noisy observations under physics-based regularization, rather than a profile obtained by direct averaging alone.

5. Mean-flow organization under rotation, curvature, and wave interaction

In rotating and curved systems, mean velocity flow is constrained by dynamical compatibility with waves, vorticity, or geometry. In a rotating pipe or circular annulus with axial flow, inertial-wave analysis implies that the bulk mean profiles must satisfy radial conditions equivalent to

$\mathbf{u}$9

leading to Bessel-function mean profiles (Yang et al., 2015). The central prediction is

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$0

with excellent agreement to DNS in the bulk, especially for moderate rotation $$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$1 (Yang et al., 2015). Large-scale vortex clusters in the bulk are interpreted as the coherent structures associated with the same inertial-wave organization (Yang et al., 2015).

In strongly turbulent Taylor–Couette flow, curvature modifies the mean profile through a Monin–Obukhov-like length,

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$2

This length separates a shear-dominated near-wall region from a curvature-dominated one (Berghout et al., 2020). The angular velocity, rather than the streamwise velocity, becomes the natural mean variable, and a second logarithmic regime appears,

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$3

with $$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$4, before transition to a constant-angular-momentum bulk state (Berghout et al., 2020). This indicates that a single Prandtl–von Kármán log law is insufficient for the mean velocity structure in cylindrical shear flow with strong curvature.

Wave–mean-flow interaction gives mean velocity flow a different role. In a spherical shallow-water model, subtropical Rossby waves resonantly excite equatorial Kelvin waves through interactions with the balanced zonal mean flow, not merely through direct forcing (Holube et al., 2024). The mean flow Doppler-shifts both Rossby and Kelvin frequencies and alters wave structure; the dominant Kelvin tendency is the Rossby–mean flow contribution, which accounts for about $$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$5 of the Kelvin energy tendency in the reference case, whereas Rossby–Rossby interactions remain below $$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$6 (Holube et al., 2024). The velocity-tendency term

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$7

and the corresponding depth tendency are the key interaction channels (Holube et al., 2024). In this setting, the mean flow is not merely a background average; it is an active dynamical participant that makes resonance possible.

Equatorially trapped water waves with constant underlying current provide yet another formulation. For the exact Lagrangian solution considered there, the mean Lagrangian velocity over one period is

$$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$8

independent of depth and latitude (Henry et al., 2024). The mean Eulerian velocity is given by a transformed depth-fixed integral and is always westward for $$u(z_t,r,t)\triangleq \frac{1}{t-r}\int_r^t v(z_\tau,\tau)\,d\tau,$$9, while for $v(z_t,t)$0 its sign depends on parameter regime (Henry et al., 2024). The resulting Stokes drift,

$v(z_t,t)$1

is eastward throughout the fluid when $v(z_t,t)$2 (Henry et al., 2024). This case underscores that different averaging frames—Lagrangian, Eulerian, and drift—yield distinct mean velocity flows even for the same exact wave field.

6. Coupled media, transport problems, and non-fluid extensions

In coupled channel–porous systems, mean velocity flow enters through cross-sectional averages and flow partition. In a vanadium redox flow battery model comprising a single serpentine flow channel over a porous electrode, the inlet condition is expressed through the mean inlet velocity $v(z_t,t)$3,

$v(z_t,t)$4

which sets the total electrolyte feed rate (Malkov, 2016). The channel obeys a simplified steady laminar Navier–Stokes balance,

$v(z_t,t)$5

while the porous electrode is modeled by Darcy–Brinkman,

$v(z_t,t)$6

with continuity of velocity and normal stress at the interface (Malkov, 2016). The fully developed velocity in the channel is quadratic, the porous-layer velocity is exponential, and total flow satisfies

$v(z_t,t)$7

The porous-layer flow rate increases almost linearly with mean inlet velocity, and at $v(z_t,t)$8 cm s$v(z_t,t)$9 the analytical maximum current density is $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$0 mA cm$$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$1, compared favorably with an experimental value of about $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$2 mA cm$$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$3 (Malkov, 2016).

Granular heap flow uses mean velocity flow to describe a strongly depth-localized profile. The mean motion consists of a fast-flowing surface layer and a slower creep layer beneath, with a transition at about $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$4 cm and a profile that decays roughly exponentially with depth (Yu et al., 16 Mar 2025). The fluctuation spectrum

$$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$5

has $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$6 in the surface layer and approaches $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$7 by about $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$8 cm depth, roughly 55 particle diameters (Yu et al., 16 Mar 2025). The paper links this layered mean flow to a self-organized criticality interpretation with open-boundary behavior near the surface and closed-boundary behavior in the deep creep region (Yu et al., 16 Mar 2025). A plausible implication is that the mean profile and the fluctuation spectrum are not separate observables but coupled signatures of the same depth-dependent organization.

Beyond continuum transport, mean velocity flow appears in molecular communication. A molecular flow velocity meter infers the medium’s mean flow velocity from Poisson counts at a transparent receiver, with the flow entering the advection–diffusion equation and the received count means $$u(a(t), t, r, s) \triangleq \frac{1}{r-t}\int_t^r v(a(\tau),\tau,s)\, d\tau,$$9 through the drift-induced displacement of the concentration field (Farahnak-Ghazani et al., 2020). The work derives MAP detection and MAP/MMSE estimation rules, Bayesian Cramér–Rao and expected Cramér–Rao bounds, and optimum sampling-time criteria (Farahnak-Ghazani et al., 2020). Here, mean velocity flow is a channel parameter to be statistically detected or estimated, rather than a velocity field resolved in space.

Dark matter halo modeling extends the concept further. In axisymmetric rotating and growing halos, mean flow is decomposed into radial, polar, and azimuthal components, with $100$00 and the azimuthal flow linked directly to velocity dispersion structure (Xu, 2022). Large halos exhibit a self-similar radial flow—outward in the core and inward in the outer region—whereas small halos have vanishing radial flow, $100$01 (Xu, 2022). The paper interprets fictitious stress as a Reynolds-stress-like mechanism transferring energy between coherent mean flow and random motion (Xu, 2022). Although physically remote from laboratory turbulence, this usage retains the same structural opposition between organized transport and dispersion.

7. Mean velocity flow in generative modeling and policy learning

In machine learning, mean velocity flow has become a formal alternative to instantaneous velocity fields in flow-based generation. MeanFlow introduces the average velocity field

$100$02

and derives the MeanFlow Identity,

$100$03

with

$100$04

computed by Jacobian-vector product (Geng et al., 19 May 2025). The training objective regresses a network $100$05 to a stop-gradient target based on this identity, and when $100$06, the correction vanishes so that MeanFlow reduces to standard Flow Matching (Geng et al., 19 May 2025). The practical consequence is one-step generation,

$100$07

without numerical ODE integration (Geng et al., 19 May 2025). On ImageNet $100$08, the paper reports FID $100$09 at 1-NFE for MeanFlow-XL/2 trained from scratch, together with $100$10 FID for XL/2+ at 2-NFE (Geng et al., 19 May 2025).

Mean Velocity Policy transfers the same idea to reinforcement learning. It learns a mean velocity field for one-step action generation and imposes an instantaneous velocity constraint

$100$11

which serves as a boundary condition for the mean-flow identity (Zhan et al., 14 Feb 2026). The mean-flow loss is

$100$12

The paper proves that without such a boundary condition, the learned field can differ from the true mean velocity by $100$13, whereas IVC forces the ambiguity to zero (Zhan et al., 14 Feb 2026). On nine sparse-reward Robomimic and OGBench tasks, the reported average success rate is $100$14, with training speed $100$15 iter/s and inference time $100$16 ms (Zhan et al., 14 Feb 2026).

Transition Flow Matching then positions mean velocity methods as an intermediate step between local-velocity flow matching and direct transition prediction. It defines

$100$17

and states the relation

$100$18

before introducing a full transition map $100$19 and the Transition Flow Identity

$100$20

(Ma, 16 Mar 2026). This suggests that mean velocity flow in generative models is a globalized representation of transport that sits conceptually between infinitesimal vector fields and full state-transition operators.

The cross-domain continuity is striking. In both turbulence and generative modeling, the central issue is whether one should model local instantaneous dynamics and integrate them, or instead learn an averaged transport law over a finite interval. The objects are mathematically different, but the reduction principle is analogous.

8. Statistical-mechanical mean velocity equations and closure

A more formal interpretation of mean velocity flow arises in statistical-mechanical derivations of hydrodynamics. Piest interprets the hydrodynamic velocity derived via the Zwanzig–Mori projection-operator technique as the mean velocity field $100$21 in turbulent flow (Piest, 2013). The exact generalized momentum equation is

$100$22

with the stress tensor represented by a nonlocal memory integral (Piest, 2013). By expanding the kernel around equilibrium and applying multilinear mode coupling theory, the second-order term vanishes and the first non-Navier–Stokes correction is a third-order nonlinear friction force expressed as a convolution integral containing higher-order velocity gradients (Piest, 2013).

A numerical follow-up applies this mean velocity equation for fluctuating flow to low-Reynolds-number circular jet flow (Piest, 2016). Using a stream-function formulation and treating the fluctuation-induced term as a source $100$23 in the vorticity equation,

$100$24

the study computes approximate corrections by iteration from a Navier–Stokes jet solution (Piest, 2016). The main result is that deviations from the ordinary laminar solution become distinct for $100$25 (Piest, 2016). This program differs from Reynolds-averaged closure in that the mean velocity equation is presented as a formally derived transport equation with a nonlinear dissipation term rather than as an averaged equation requiring phenomenological closure from the outset.

A plausible implication is that the phrase “mean velocity equation” names two different intellectual projects. One is operational and data-driven, as in inverse RANS or profile fitting. The other is foundational, seeking an exact or asymptotically justified governing equation for the mean flow itself. The closure problem remains central in both.

9. Common themes and recurrent misconceptions

Several themes recur across these otherwise disparate literatures. First, the existence of a mean velocity flow should not be assumed from short-window visual coherence. Plane thermal convection shows that apparent wall-parallel organization can disappear under longer and wider averaging, leaving a fluctuation-dominated field with no sustained mean wind (Samuel et al., 2024). Second, a mean velocity profile can remain robust even when other statistics have not yet reached asymptotic form, as illustrated by the CICLoPE centerline law and the slower convergence of turbulence intensity and normal stress (Nagib et al., 9 Nov 2025). Third, the “correct” mean variable may depend on geometry: angular velocity is more natural than azimuthal velocity in turbulent Taylor–Couette flow (Berghout et al., 2020), and different averages produce distinct transport measures in equatorial water waves (Henry et al., 2024).

A further misconception is to equate mean velocity with weak fluctuations around a dominant coherent background. In several settings the opposite holds. Plane convection exhibits fluctuations up to two orders of magnitude larger than the mean (Samuel et al., 2024). Granular heap flow links a layered mean profile to depth-dependent fluctuation spectra rather than to a single coherent bulk motion (Yu et al., 16 Mar 2025). In machine learning, the mean velocity field is not a smoothed estimate of an underlying trajectory sample but a deliberately constructed global transport target distinct from the instantaneous field (Geng et al., 19 May 2025).

The literature therefore supports a disciplined usage of the term. Mean velocity flow is best understood as a reduced transport descriptor defined by a specified averaging or marginalization procedure and validated, or invalidated, by the physics of the system at hand. Whether it appears as a logarithmic profile, a vanishing long-time average, a Bessel-structured rotating-flow field, a Bayesian posterior mean, or a one-step generative displacement, its scientific content lies in the relation between the chosen mean and the fluctuations, structures, and dynamics that remain unresolved.