1. Introduction

Measurements of ocean turbulence using neutrally buoyant floats have produced empirical forms for the spectra of velocity (Lien et al. 1998) and density (Lien et al. 2002) following a Lagrangian trajectory in stratified or unstratified turbulence. These spectra predict the dispersion of particles and thus imply mixing rates. In this paper, we compare these mixing rates to those derived by other methods and thus attempt to clarify the relationship between Lagrangian and Eulerian turbulence measurements.

Taylor (1921) pioneered the study of turbulent mixing through the examination of the statistics of single particle dispersion. In a stationary and shear-free flow, the dispersion coefficient K_z is defined as

View Expanded

where R_w(τ) = 〈w(t)w(t + τ)〉 is the Lagrangian velocity covariance function at a time lag τ, z is the displacement of a particle from its initial position, and angle brackets denote an average over an ensemble of particles. Two fundamental results were obtained by Taylor without invoking explicit properties of turbulence. First, at time scales much smaller than the integral time scale of velocity T = ∫^∞₀ R_w(τ) dτ/R_w(0), K_z grows linearly with time. Second, for t ≫ T, K_z converges to a constant value. The temporal evolution of K_z depends on the details of R_w(τ).

The turbulent transport of a scalar c—for example, temperature or density—is often parameterized by an eddy diffusivity defined as the ratio of the flux F_c to its mean gradient ∂_zc; that is, K_e = −F_c/∂_zc. Dispersion predicts the flux of particles, and so, if c is the particle concentration and K_z is the eddy diffusivity for particles, then K_e should equal K_z. Here, we will compare estimates of K_z from Lagrangian models with estimates of K_e from Eulerian models.

For stratified turbulence, two parameterizations of K_e are commonly used, particularly in oceanic measurements. The Osborn and Cox (1972) method is based on measuring χ_c, the dissipation rate of scalar variance, and a mean scalar gradient (∂_zc); that is,

View Expanded

This expression was originally derived from the scalar variance equation by assuming a balance between the turbulence flux and dissipation. In this derivation, K_h is the vertical eddy diffusivity and (∂_zc) is the average vertical gradient. Winters and D'Asaro (1996) derive (2) exactly, but with χ_c describing the diascalar flux, that is, the flux of scalar across isoscalar surfaces, and (∂_zc) is the gradient of the profile of c after it has been monotonically sorted. For a well-stratified fluid, the difference between these definitions does not create large differences in the computed eddy coefficient (D'Asaro et al. 2004). For the rest of this paper, (2) will be used for density flux, that is, c = ρ.

The Osborn (1980) method is based on ε, the dissipation rate of turbulent kinetic energy,

View Expanded

Here Γ is the mixing efficiency, closely related to the flux Richardson number, and N is the buoyancy frequency. The value of Γ is typically 0.2 or less in the stratified ocean (Osborn 1980). The values of K_ρ and K_h, when sufficiently averaged, generally agree with each other and with other estimates of diapycnal diffusivity (Gregg 1998).

This paper will 1) revisit the dispersion coefficient K_z using up-to-date Lagrangian velocity spectra in unstratified turbulence (section 2), internal waves (section 3), and stratified turbulence (sections 4 and 5); 2) investigate diapycnal dispersion by defining a diapycnal dispersion coefficient K∗ based on Lagrangian scalar spectra (section 5); and 3) reconcile the dispersion coefficients K_z and K∗ with the diapycnal eddy diffusivity coefficients K_h and K_ρ.

2. Unstratified turbulence

Taylor's dispersion coefficient K_z (1) can be expressed in terms of Lagrangian velocity spectrum Φ_w(ω) or structure function D(τ) = 〈[w(t + τ) − w(t)]²〉; for example,

View Expanded

In the inertial subrange, Kolmogorov scaling suggests forms for both the structure function D(τ) = C₀ετ and the spectrum Φ_w(ω) = βεω⁻². Monin and Yaglom (1975) show that the two constants are related as C₀ = βπ. Lien and D'Asaro (2002) find that the value C₀ = 6 ± 0.5 and β = 2 ± 0.2 are consistent with existing data at high Reynolds number.

Three Lagrangian velocity spectra (Fig. 1) are chosen to study the behavior of K_z. The first spectrum is proposed by Tennekes and Lumley (1972):

View Expanded

The second form (Mordant et al. 2001) assumes an exponential form e^−τ/T for the Lagrangian velocity covariance function; the velocity spectrum has a Lorentzian form,

View Expanded

The third form is constructed empirically using observations taken in the oceanic surface mixed layer (Lien et al. 1998):

View Expanded

The three spectral forms are plotted in Fig. 1a. In each, the spectrum transitions from a flat slope at low frequencies to an inertial subrange with a −2 slope. The transition occurs at a large-eddy frequency ω^TL_o, ω^M_o, or ω_o. The large- eddy frequency is defined as ω_o = ε^1/3k^2/3_o, where k_o is the large-eddy wavenumber (Lien et al. 1998). The transition is most abrupt for Φ_TL and the smoothest for Φ_L. The inertial subrange Kolmogorov constants are β_TL = 1.2, β_M = 1.8 and β_L = 1.8. Given these values and requiring all three forms to have the same velocity variance, σ²_w = 1.6ε/ω_o, requires ω^TL_o = 1.6ω_o, and ω^M_o = 1.9ω_o.

Correlation and structure functions are computed from the spectrum and plotted in Figs. 1b and 1c, respectively. For these plots, the inertial subrange has been extended to a Nyquist frequency 10⁶ω_o; our results are insensitive to this factor as long as the inertial subrange is sufficiently wide. The three correlation functions have a similar exponential form. The three structure functions normalized by ετ have a plateau at τω_o ≪ O(1). The level of the plateaus gives C₀(=πβ), yielding 3.8 for Φ_TL and 5.7 for Φ_M and Φ_L, consistent with the results of previous work (Yeung 2002). Figure 2 shows the particle dispersion coefficients K_z(t) computed from the three Lagrangian spectra using (4). The curves are consistent with Taylor's results. K_z(τ) increases linearly as σ²_wτ at τ ≪ ω⁻¹_o and reaches a constant level K_z (∞) = (π/2)α at τ > ω⁻¹_o, where α is the spectral level at ω ≪ ω_o. Thus we scale K_z(∞) with εω⁻²_o, the spectral level at low frequencies; K_z(∞)ω²_oε⁻¹ equals 0.73, 0.78, and 0.58 for Φ_TL, Φ_M, and Φ_L, respectively.

The model of Mellor and Yamada (1974) will be used as a summary of the rates of mixing in unstratified turbulence. They parameterized the vertical eddy diffusivity K_M = 0.4ql, where q² = 3σ²_ω and l is the master length scale. Assuming l = k⁻¹_o and using σ²_w = 1.6εω⁻¹_o and ω_o = ε^1/3k^2/3_o, predicts K_M = 0.88εω⁻²_o. Thus for unstratified homogeneous turbulence the coefficients for dispersion K_z and diffusion K_M are very close.

3. Oceanic internal waves—Vertical dispersion

Internal gravity waves are the dominant cause of vertical motion in the ocean for frequencies greater than f, the inertial frequency. A large body of literature starting with Fofonoff (1969) and reviewed by Garrett and Munk (1979) finds the vertical displacements due to these motions to be well represented by linear wave theory with a “Garrett–Munk” (henceforth GM) vertical velocity spectrum, which can be approximated as

View Expanded

where ϕ₀ sets the internal wave energy level. Figures 3a and 3b show Φ_GM and the associated dispersion function K_GM(t) for a typical value of N/f = 100. Since internal waves are adiabatic, they cannot cause mixing and K_GM(∞) = 0. However, they do cause a net displacement of water from its initial position, resulting in an apparent vertical dispersion with a maximum value of about (π/2)ϕ₀ near Nt = 2. The value decreases to zero for Nt = 200 and then oscillates around zero with decreasing amplitude. Smoothing Φ_GM near f and N results in a decreased amplitude of oscillations and a more rapid decay of K_GM(t) for Nt ≫ 100.

4. Stratified turbulence—Vertical dispersion

D'Asaro and Lien (2000a) argue that in stratified turbulence the large-eddy frequency ω_o should be approximately N. They support this by oceanic observations in which Lagrangian vertical velocity spectra exhibit a form nearly indistinguishable from (7) and N = 2(±0.3)ω_o. The spectra are isotropic for ω ≫ N, supporting the idea that these are turbulent motions. The spectra are highly anisotropic in the same sense as internal waves for ω ≪ N, supporting the idea that these motions are due to internal waves. These results should apply for flows with gradient Richardson numbers of order 1; for sufficiently larger Richardson numbers the spectral form near N changes (D'Asaro and Lien 2000b); for sufficiently smaller Richardson numbers the flow geometry determines ω_o.

The large value of K_z(∞) is explained by adiabatic internal waves being responsible for most of the vertical excursions. As shown in Fig. 3 internal waves produce a large apparent vertical dispersion for 1 < Nt < N/f. However, if the vertical motions of stratified turbulence are dominated by internal waves for ω < N, then the spectrum must also reflect the low frequency limit of internal waves. In the open ocean, this is due to f. In a confined sea or a laboratory vessel it might be set by the lowest mode of oscillation. In either case, the vertical velocity spectrum cannot remain white at low frequency as in (5), (6), and (7); instead, it must decrease by a factor Γ/Γ_d so as to bring K_z(∞) into agreement with K_ρ.

An example of a Lagrangian spectrum with a low frequency cutoff is shown in Fig. 4. The dispersion function K_z(t) is similar to those in Fig. 2 for Nt ≤ 10 but decreases as in Fig. 3 for large Nt, ultimately reaching a value of Γε/N² at long times. The initial large apparent dispersion is due to the wave part of the flow; the final smaller value is the true dispersion due to diapycnal mixing.

5. Stratified turbulence—Diapycnal dispersion

Following Taylor's dispersion theory, as in (1), we define a diapycnal dispersion coefficient as

View Expanded

Here, w∗ is the diapycnal velocity, R_w* is the diapycnal velocity covariance function, D_w* is the corresponding structure function, Φ_w* is the diapycnal velocity spectrum, and Φ_ρ̇ is the spectrum of Dρ/Dt, the Lagrangian rate of change of density.

To evaluate (12), Lagrangian spectra of density change are needed. There is only scant theoretical or measurement guidance available. Kolmogorov scaling suggests Φ_ρ̇(ω) = β_ρχ in an inertial subrange extending from ω > ω_o ≈ N/2 to ω < ω_ν = ε^1/2ν^−1/2. Here β_ρ is a Kolmogorov constant, ω_ν is the Kolmogorov frequency at which viscous effects become important, and ν is the kinematic viscosity. Lagrangian observations in an oceanic shear-stratified flow provide vague evidence of white Φ_ρ̇ spectrum in the inertial subrange (Lien et al. 2002). Yeung (2001) performed direct numerical simulations but did not observe a white Φ_ρ̇ in the inertial subrange.

We will assume a white spectrum Φ_ρ̇ = β_ρχ and apply a cutoff of ω⁻² for ω > ω_ν (Fig. 5). Figure 6 shows that K∗(t) attains a constant value within the inertial subrange:

View Expanded

The true diapycnal dispersion rate is given by K∗(∞), which depends on the spectral level of Φ_ρ̇ at low frequency. All contributions to Dρ/Dt and thus Φ_ρ̇ must occur at diffusive scales since adiabatic motions cannot affect Dρ/Dt. Diffusive motions have frequencies near ω_ν and thus contribute to the increase in K∗ for small times. Motions within the inertial subrange are adiabatic and do not contribute to K∗. Motions at lower frequencies should also be adiabatic and will also not contribute to K∗. Thus we expect (12) to be accurate at all frequencies and K∗(∞) to be given by (13).

Expression (13) is identical to (2), except for the numerical factor. Equating K∗ and K_h implies β_ρ = 1/π. This is similar to the previous estimates of Yeung (2001) (β_ρ ≈ 0.5) and Lien et al. (2002) (β_ρ ≈ 1).

6. Discussion

Figure 7 helps to illustrate the above ideas. Consider the motion of an isopycnal surface within a given volume in a turbulent stratified fluid. The volume is chosen to be somewhat larger than the overturning scale of the turbulence. The average position of the surface (black) moves vertically in response to internal waves. If there is no mixing, the internal waves move dye vertically from its initial position, causing apparent vertical dispersion at short times. The motion is bounded by the wave amplitude so that the dispersion decreases to zero for long times (Fig. 3).

Suppose dye is placed on the surface at time t = 0 and spreads away from the isopycnal by mixing. The gray shading indicates the spread of the dye. At short times, waves are still the dominant cause of vertical motion of the dye. At longer times, the spread of the dye exceeds the amplitude of the waves, resulting in a finite dispersion that is smaller than the initial wave- induced apparent dispersion. This sequence is described by Fig. 4. The transition time between these two regimes depends on the amplitude and frequency of the internal waves, even if these play no role in the mixing.

The Lagrangian vertical velocity spectrum, or equivalently the dispersion function K_z(t), contains information on both diabatic and adiabatic motions. The diabatic processes can be isolated by taking a sufficiently long time average, that is, finding K_z(∞). In some systems this may not be possible. For example, we have several measurements of Φ_w in stratified fluids, but none extend to low enough frequency to see the expected decrease in level. This issue is not unique to Lagrangian spectra, but is common to many mixing diagnoses that use an Eulerian coordinate system (Winters and D'Asaro 1996).

These problems can be mostly overcome by working in an isopycnal coordinate system. Obviously, the mixing rate is more easily determined by measuring the diapycnal spreading of the dye around the isopycnal than by measuring its vertical spreading. Thus it is not surprising that the Lagrangian spectrum of density, which measures the motion of particles across isopycnals, gives a measurement of diapycnal diffusivity (13) that is completely insensitive to adiabatic motions of the isopycnal and therefore does not need to be averaged over long time scales to provide useful information.

Measurements of Φ_w and Φ_ρ̇ can, in principle, be made using water-following floats (D'Asaro 2003). Practically, however, these floats only follow water at high frequencies. At low frequencies they are usually isopycnal; that is, they return to a given density surface. Isopycnal floats would follow the black line in Fig. 7, while truly Lagrangian floats, like dye, would undergo random walks filling the shaded region. This distinction is critical for the estimation of K_z(∞) from Φ_w since it is the low frequency value of Φ_w that determines K_z(∞). As discussed in D'Asaro (2003), it becomes increasingly difficult to design floats that are accurately Lagrangian at frequencies very much smaller than N. Float technology may be a better match for estimation of K∗ from Φ_ρ̇. If, as hypothesized here, the spectrum is indeed white for all frequencies below the Kolmogorov frequency, then measurements of Φ_ρ̇ within the inertial subrange will be sufficient. Existing floats can be accurately Lagrangian for these frequencies. However, a better understanding of the form and scaling of Φ_ρ̇, gained by both theory and observations, will be required before diapycnal mixing rates can be reliably computed using Lagrangian measurements of high frequency density fluctuations.

7. Conclusions

This study has examined the relationship between diapycnal mixing and Lagrangian spectra of vertical velocity and temperature change through the concept of single particle dispersion [(1)]. This is motivated by recent measurements that define empirical forms for these spectra in turbulent flows.

• In unstratified homogeneous turbulence, Lagrangian velocity spectra predict dispersion coefficients that closely match known diffusion coefficients.

• Adiabatic internal gravity waves in density stratified flows produce large apparent dispersions at time lags somewhat larger than 1/N, but no average dispersion at long times.

• In density stratified turbulence with Richardson numbers of O(1), the existing empirical Lagrangian velocity spectra predict vertical dispersion coefficients with the Osborn (1980) [(3)] form but with a mixing efficiency Γ_d much larger than observed. This is explained by the presence of a large adiabatic gravity wave component in these flows that leads to a large apparent dispersion at short times. If gravity waves and other adiabatic vertical motions become sufficiently weak at low frequency, then the Lagrangian velocity spectrum can be used to measure diapycnal mixing. The existing spectral forms must be modified to represent these low frequency changes. Similarly, measurements of the Lagrangian velocity spectrum must extend to sufficiently low frequency to be useful for the estimation of diapycnal mixing.

• Inertial subrange theory and limited observations predict a white Lagrangian inertial subrange spectrum for density change with a level β_ρχ, where χ is the dissipation rate for density. This form gives a diapycnal dispersion coefficient with the Osborn and Cox (1972) [(2)] form and thus predicts a value β_ρ = 1/π for the Kolomogorov constant. This formulation is insensitive to adiabatic motions such as gravity waves and is thus a promising approach for estimation of diapycnal mixing from neutrally buoyant floats.