Structure functions and intermittency

Choose a separation vector $\mathbf r=r\hat{\mathbf r}$ and define the longitudinal increment as the velocity difference projected along it. The absolute pth-order structure function averages the pth power of its magnitude. It depends on separation, direction, flow position, and time unless statistical symmetries eliminate some of those variables. A mean over independent realizations is an ensemble average. Replacing it by a spatial or time average requires a justified sampling assumption; taking many nearby samples does not make them independent.

For even integer p, absolute and signed powers coincide. For odd p they generally do not. In particular, the signed longitudinal third moment can be negative while the absolute third moment cannot. Losing that sign erases information carried by the exact four-fifths relation in its inertial-range regime. Always identify which moment has been plotted before comparing a measured exponent or coefficient with theory. A label saying only “third-order structure function” leaves a consequential ambiguity.

A simple inertial-range similarity estimate gives $S_p(r)\propto r^{p/3}$, with coefficients carrying the required dimensions. Actual high-order statistics can depart from these exponents, reflecting scale-dependent distribution shapes and concentration of intense events. A useful diagnostic is flatness, the fourth moment divided by the second moment squared. A Gaussian increment has flatness three, but a different constant value alone does not establish scale-dependent turbulent intermittency. Examine how the normalized distribution and its tails change with separation.

For homogeneous isotropic incompressible three-dimensional turbulence, under the high-Reynolds-number inertial-range and suitable stationarity conditions, the signed third-order law is $\langle(\delta u_L)^3\rangle=-4\varepsilon r/5$. Its sign is linked to the forward energy flux with this increment convention. The relation does not imply that every third-order absolute moment has that coefficient, nor does it determine all higher-order exponents. Proposed intermittency models can organize observations, but no general exact derivation of every turbulent scaling exponent is established.

A fourth or sixth power can make a few intense events dominate an estimate. A record containing many ordinary fluctuations may therefore give a stable variance while its high-order moment continues changing dramatically. Assess convergence by comparing independent blocks, examining which amplitudes contribute to the integral, and checking sensor saturation and noise. Quietly clipping large values as outliers can remove the very physical events under investigation. Conversely, isolated electronic spikes can imitate a heavy tail and must be distinguished using the measurement chain.

Intermittency also has several contextual meanings. A turbulent patch alternating with a nearly nonturbulent region is not identical to small-scale internal intermittency within an already turbulent region. State the conditioning, separation, geometry, Reynolds number, and measured component. In a finite data range, a fitted slope has uncertainty and may mix forcing, inertial, and dissipation behavior. Reporting that limitation is part of the physical result. A good statistical model explains which features it captures and which rare-event probabilities remain insufficiently constrained.