the relation between the statistics of open ocean currents...
TRANSCRIPT
Climate Dynamics manuscript No.(will be inserted by the editor)
The relation between the statistics of open ocean1
currents and the temporal correlations of the2
wind-stress3
Golan Bel · Yosef Ashkenazy4
5
Received: date / Accepted: date6
Abstract We study the statistics of wind-driven open ocean currents. Using the7
Ekman layer model for the integrated currents, we investigate, analytically and8
numerically, the relation between the wind-stress distribution and its temporal9
correlations and the statistics of the open ocean currents. We find that temporally10
long-range correlated wind results in currents whose statistics is proportional to11
the wind-stress statistics. On the other hand, short-range correlated wind leads12
to Gaussian distributions of the current components, regardless of the stationary13
distribution of the winds, and therefore, to a Rayleigh distribution of the current14
amplitude, if the wind-stress is isotropic. We find that the second moment of the15
current speed exhibits a maximum as a function of the correlation time of the16
wind-stress for a non-zero Coriolis parameter. The results were validated using an17
oceanic general circulation model.18
Keywords Wind statistics · Ocean current statistics · Temporal correlations19
PACS 92.60.Cc · 91.10.Vr · 88.05.Np · 92.60.Gn20
1 Introduction21
Ocean currents are generated by local and remote forces and factors, including22
winds, tides, buoyancy fluxes and various types of waves. While many studies have23
investigated the distribution of the wind (Seguro and Lambert, 2000, Monahan,24
2006, 2010), focusing on its relevance to energy production, the distribution of25
ocean currents has received much less attention (Chu, 2008). Moreover, currently26
Golan BelDepartment of Solar Energy and Environmental Physics, Blaustein Institutes for Desert Re-search, Ben-Gurion University of the Negev, Sede Boqer Campus 84990, IsraelTel.: +972-8-6596845Fax: +972-8-6596921E-mail: [email protected]
Yosef AshkenazyDepartment of Solar Energy and Environmental Physics, Blaustein Institutes for Desert Re-search, Ben-Gurion University of the Negev, Sede Boqer Campus 84990, Israel
2 Bel and Ashkenazy
there is no accepted theory explaining the observed statistics of surface ocean27
currents.28
Here, we propose a simple physical theory for the distribution of wind-driven29
ocean currents and its relation to the spatially variable temporal correlations of30
the wind (see Fig. 1). We show that the distribution of wind-driven ocean currents31
strongly depends on the temporal correlations of the wind–when the wind exhibits32
long-range temporal correlations, the ocean current statistics is proportional to the33
wind-stress statistics, while for short-range correlations of the wind, the different34
components of the current vector follow Gaussian distributions. It was previously35
reported that the probability density function (PDF) of ocean currents follows the36
Weibull distribution (Gille and Smith, 1998, 2000, Chu, 2008, 2009, Ashkenazy37
and Gildor, 2011) (see Section 4.2 for the details of the Weibull distribution);38
We argue that this is not necessarily the case, even if the wind-stress magnitude39
follows the Weibull distribution.40
Oceans play a major and important role in the climate system, and ocean41
circulation underlies many climate phenomena, from scales of meters to thousands42
of kilometers and from scales of minutes to decades. The gap in our understanding43
of ocean current statistics leads to a lack in understanding the statistics of other44
related climate variables. Filling in this gap will be useful in many fields: it may45
help to predict extreme current events and, thus, may help to securely design46
maritime-associated structures. In light of the increasing efforts to find alternative47
sources of energy and the idea of using ocean currents as such a source, knowledge48
of the currents’ PDF may help to better estimate the energy production and49
to appropriately design ocean current turbines that will withstand even extreme50
current events. Moreover, knowledge regarding current statistics and, in particular,51
its relation to the wind-stress statistics may improve the parametrization of small-52
scale processes in state-of-the-art general circulation models (GCMs).53
The study of wind-driven ocean currents goes back more than 100 years, to the54
time when Ekman (1905) proposed his classical simple model to explain the effect55
of Earth’s rotation on upper ocean currents. His model predicted that the depth-56
integrated current vector is perpendicular to the wind vector, a prediction that57
was largely proven by observations. Since then, many studies have used Ekman’s58
model to propose more realistic models for ocean currents, as well as for surface59
winds (Gill, 1982, Cushman-Roisin, 1994). In what follows, we use Ekman’s model60
(Ekman, 1905) to study the statistics of wind-driven ocean currents.61
This paper is organized as follows. In section 2, we describe the Ekman layer62
model and provide a general (implicit) solution expressed in terms of the charac-63
teristics of the wind-stress statistics. In section 3, we consider two idealized cases64
of wind-stress statistics (step-like temporal behavior of the wind-stress in section65
3.1 and exponentially decaying temporal wind-stress correlation in section 3.2) and66
present analytical expressions for the second moment (and fourth moment for the67
first case) of the currents’ distribution. In section 4, we provide a description of the68
simple numerical model and of the oceanic GCM (MITgcm) used to validate and69
extend the analytical results. The numerical tests involve the Weibull distribution;70
hence, a brief review of the distribution properties and the methods we used to71
generate correlated and uncorrelated Weibull-distributed time series is provided72
in 4.2. The numerical results are presented and discussed in section 5, followed by73
a brief summary in section 6.74
The relation between current and wind statistics 3
2 The Ekman model75
In spite of the simplicity of the Ekman model, it will enable us to start investigating76
the coupling between the wind-stress and the ocean currents. The time, t, and77
depth, z, dependent equations of the Ekman model (Ekman, 1905), describing the78
dynamics of the zonal (east-west) U and meridional (south-north) V components79
of the current vector, are:80
∂U
∂t= fV + ν
∂2U
∂z2
∂V
∂t= −fU + ν
∂2V
∂z2, (1)
where f = (4π/Td) sin(φ) is the local Coriolis parameter (Td is the duration of81
a day in seconds and φ is the latitude), and ν is the eddy viscosity coefficient,82
assumed to be depth-independent. In eqs. (1), we consider U = U−Ug, V = V −Vg83
where Ug, Vg are the bottom ocean geostrophic currents determined by the pressure84
gradient and U , V are the actual surface current components. In what follows, we85
consider the statistics of U , V , where the statistics of U , V can be obtained by a86
simple transformation. The boundary conditions are chosen such that the current87
derivative, with respect to the depth coordinate z, is proportional to the integrated88
current at the bottom of the layer described by our model and to the wind-stress89
vector (τx, τy) at the surface (Gill, 1982, Cushman-Roisin, 1994),90
∂U
∂z
∣∣∣z=−h
=r
νu(t);
∂U
∂z
∣∣∣z=0
=τxρ0ν
;
∂V
∂z
∣∣∣z=−h
=r
νv(t);
∂V
∂z
∣∣∣z=0
=τyρ0ν
, (2)
where,91
u ≡0∫
−h
U(z)dz; v ≡0∫
−h
V (z)dz. (3)
Here, we introduce the following notation: h is the depth of the upper ocean layer, r92
is a proportionality constant representing the Rayleigh friction (Airy, 1845, Pollard93
and Millard, 1970, Kundu, 1976, Kase and Olbers, 1979, Gill, 1982), (τx, τy) are the94
wind-stress components, and ρ0 is the ocean water density (hereafter assumed to95
be constant, ρ0 = 1028kg/m3). The value used for r in the numerical calculations96
is based on the empirical estimate outlined in Simons (1980), Gill (1982).97
By integrating eqs. (1) over a sufficiently deep layer (i.e., h �√ν/2f), we98
obtain the equations describing the integrated currents and their coupling to the99
wind-stress,100
∂u
∂t= fv − ru+
τxρ0
,
∂v
∂t= −fu− rv +
τyρ0
. (4)
4 Bel and Ashkenazy
In the derivation above, we did not explicitly consider the pressure gradient. Ex-101
plicit inclusion of the pressure gradient would result in constant geostrophic cur-102
rents, and equations (4) describe the dynamics of the deviation from the geostrophic103
currents.104
To allow a simpler treatment of these equations, we define w ≡ u + iv. It is105
easy to show that w obeys the following equation:106
∂w
∂t= −ifw − rw +
τ
ρ0, (5)
where τ ≡ τx + iτy. The formal solution of equation (5) is107
w(t) = w(0)e−(if+r)t +1
ρ0
t∫0
τ(t′)e−(if+r)(t−t′)dt′. (6)
This formal solution demonstrates that the currents depend on the history of the108
wind-stress, and therefore, the distribution of the currents depends, not only on109
the wind-stress distribution, but on all multi-time moments of the wind-stress.110
However, the two extreme limits are quite intuitive. When the correlation time111
of the wind-stress, T , is long (i.e., T � 1/r and T � 1/f), one expects that112
the currents will be proportional to the wind-stress since the ocean has enough113
time to adjust to the wind and to almost reach a steady state. In terms of eq.114
(6), in this limit, τ(t′) can be approximated by τ(t) since its correlation time is115
longer than the period over which the exponential kernel is non-zero. In the other116
limit, when the correlation time of the wind-stress is very short (i.e., rT � 1 and117
fT � 1), namely, the wind-stress is frequently changing in a random way, the118
ocean is not able to gain any current magnitude. In this case, the central limit119
theorem (Van-Kampen, 1981) implies that each component of the current vector120
is Gaussian distributed.121
It is useful to write the formal expression for the square of the currents ampli-122
tude in terms of the wind-stress temporal correlation function. Taking the square123
of eq. (6) and averaging over all realizations (with the same statistical properties)124
of the wind stress, one obtains:125
〈|w(t)|2〉 = |w(0)|2e−2rt (7)
+w(0)∗
ρ0e−(r−if)t
t∫0
〈τ(t′)〉e−(if+r)(t−t′)dt′
+w(0)
ρ0e−(r+if)t
t∫0
〈τ(t′)∗〉e−(r−if)(t−t′)dt′
+1
ρ20
t∫0
t∫0
〈τ(t′)τ(t′′)∗〉e−if(t′′−t′)e−r(2t−t′−t′′)dt′dt′′.
The variance of the currents amplitude may be written as126
〈|w(t)|2〉 − |〈w(t)〉|2 = (8)
1
ρ20
t∫0
t∫0
C(t′, t′′)e−if(t′′−t′)e−r(2t−t′−t′′)dt′dt′′,
The relation between current and wind statistics 5
where we defined the temporal correlation function of the wind-stress as127
C(t′, t′′) = 〈τ(t′)τ(t′′)∗〉 − 〈τ(t′)〉〈τ(t′′)∗〉. (9)
To allow analytical treatment, we proceed by considering two special idealized128
cases in which the formal solution takes a closed analytical form. For simplicity,129
we only consider the case of statistically isotropic wind-stress (the direction of130
the wind-stress is uniformly distributed, Monahan, 2007), namely a case in which131
the wind-stress components are independent, identically distributed variables with132
identical temporal autocorrelation functions. The latter assumption is not always133
valid (see for example, Monahan, 2012); however, the generalization of our results134
to the case of non-isotropic wind stress is straightforward.135
3 Idealized cases136
3.1 Step-like wind-stress137
A simple way to model the temporal correlations of the wind is by assuming that138
the wind vector randomly changes every time period T while remaining constant139
between the “jumps.” By integrating the solution of w (eq. (6)) and taking into140
account the fact that the distribution of w at the initial time and final time should141
be identical, one can obtain the expressions for the second and fourth moments142
(ensemble average over many realizations of the stochastic wind-stress) of the143
current amplitude. To better relate our results to the outcome of data analysis,144
we also need to consider the time average since the records are independent of the145
constant wind-stress period. The double averaged second moment of the currents146
amplitude is defined as147
⟨|w|2
⟩≡
⟨1
T
T∫0
|w(t)|2dt⟩, (10)
and the angular brackets represent the average over different realizations of the148
wind-stress with the same statistical properties. Using eq. (7) and the fact that in149
this idealized case, the wind stress is constant over the period of the step (T ), we150
obtain for the double averaged second moment of the currents amplitude:151
⟨|w|2
⟩=
⟨|τ |2
⟩(1 + 1−A1
rT − 2 r(1−A1)+fA2
T (f2+r2)
)(f2 + r2) ρ20
, (11)
where we have introduced the notations A1 ≡ exp(−rT ) cos(fT ), and A2 ≡152
exp(−rT ) sin(fT ).153
The two extreme limits discussed in section 2 can be easily realized and under-154
stood. When the correlation time T is very long compared with 1/r, the second155
moment of the current is proportional to the second moment of the wind-stress,156
namely, 〈|w|2〉 = 〈|τ |2〉/[(f2 + r2)ρ2o]. This is exactly what one would expect. The157
long duration of constant wind-stress allows the system to fully respond and adjust158
to the driving force, and hence, the moments of the currents are proportional to159
the moments of the wind-stress. The coefficient of proportionality is given by the160
6 Bel and Ashkenazy
solution of eqs. (4) with constant wind-stress. The other limit is when rT, fT � 1161
and, in this case, 〈|w|2〉 ∼ 〈|τ |2〉T/(2rρ20).162
In this limit, the second moment of the current amplitude is very small and163
approaches zero as the correlation time approaches zero. This result is what one164
would expect for a rapidly changing wind that cannot drive significant currents.165
Using eq. (6) and the fact that wind-stress is constant over the step period (T ), we166
calculate the double averaged (over the step period and the ensemble of wind-stress167
realizations) fourth moment,168
⟨|w|4
⟩=
〈|τ |2〉2g2 (T ) + 〈|τ |4〉g4 (T )(f2 + r2)2 ρ40
, (12)
g2 (T ) =4 (D + 1− 2A1)
1−D×(
3−D − 2A1D
4rT− 2
3r (1−DA1) + fDA2
(f2 + 9r2)T
),
g4 (T ) = 1 +5− 3D + 2A1(A1 − 1−D)
2rT
− 43r (1−DA1) + fDA2
(f2 + 9r2)T
+r(1−A2
1) + rA22 + 2fA1A2
(f2 + r2)T
− 4r(1−A1) + fA2
(f2 + r2)T,
where D ≡ exp(−2rT ). Here, we again use the fact that the integrals of the time169
averaging are easily carried due to the fact that the wind-stress is constant during170
the averaged period.171
While at the limit T → 0, both the second and the fourth moments vanish, the172
ratio between the fourth moment and the square of the second moment remains173
finite and is equal to 2. This corresponds to the Rayleigh distribution which orig-174
inates in the facts that the components of the currents are independent and each175
one has a Gaussian distribution with a zero mean and the same variance.176
One can easily understand the origin of the Gaussian distribution by consid-177
ering the central limit theorem, which can be applied in this limit. Each of the178
current components is the integral of many independent, identically distributed179
random variables (the wind-stress at different times). Based on the second and180
fourth moments in the limit of T → 0, we conjecture that the overall PDF of the181
ocean current speed, in this case, is given by the Rayleigh distribution.182
In Gonella (1971), the case of constant wind-stress over some duration was183
investigated using the depth-dependent Ekman model. The existence of a wind-184
stress duration for which the current amplitude is maximal was found. The results185
of this section generalize these results by considering the randomness of the wind-186
stress and its continuity. In addition, we consider here the empirical Rayleigh187
friction.188
In order to better understand the existence of constant wind-stress duration,189
T , for which the average current amplitude is maximal, we present, in Fig. 2, two190
typical trajectories of the current velocity components u, v. The dynamics of the191
velocity field may be understood as follows. When the friction, r, is smaller than the192
The relation between current and wind statistics 7
Coriolis frequency, the velocity field undergoes circular motion and returns, with193
frequency f , to a point in the velocity space which is very close to the initial point.194
The initial current is irrelevant since even a small amount of friction ensures that195
the initial point is close to the origin at the long-time limit. The above dynamics196
is valid for any value of the temporal wind-stress. Therefore, if the wind-stress197
changes with a frequency equal to the Coriolis frequency, the current amplitude198
cannot be driven far from the origin, and its average remains small (this explains199
the observed minimum for T ≈ 2π/f). On the other hand, when the wind-stress200
changes with a frequency equal to half the Coriolis frequency, the current velocity201
reaches the largest amplitude when the wind-stress changes and the new circle202
starts around a different initial point (possibly far from the origin). Therefore,203
in the latter case, the average current speed is maximal. To better illustrate this204
picture, we present, in Fig. 3, typical time series of the current velocity components205
and amplitude. The dots represent the times when the wind-stress changed. It can206
be seen that for T = π/f , the typical velocity amplitude when the wind-stress207
changes (the dots in Fig. 3), is significantly larger than the one for T = 2π/f .208
The intuitive explanation above relies on the fact that r � f . To elucidate209
the role of the friction, we present, in Fig. 4, the analytically calculated second210
moment (eq. (11)) of the current versus the constant wind-stress duration for211
different values of r. When the viscosity is large, there is no maximum but rather212
a monotonic increase of the average current amplitude as T increases.213
3.2 Exponentially decaying temporal correlations of the wind-stress214
The second case we consider is the Orenstein-Uhlenbeck wind-stress. This pro-215
cess results in a Gaussian wind-stress with an exponentially decaying temporal216
correlation function. It is important to note that, due to the Gaussian nature of217
the force, the first and the second moments of the wind-stress, together with its218
two-point correlation function, provide all the information on the driving force.219
The correlation function of the wind-stress components is given by220
〈τi(t)τj(t′)〉 = δij〈τ2i 〉 exp(−γi|t− t′|). (13)
The long-time limits of the current components’ averages are:221
〈u〉 ∼ r〈τx〉+ f〈τy〉ρ0 (f2 + r2)
; 〈v〉 ∼ r〈τy〉 − f〈τx〉ρ0 (f2 + r2)
. (14)
The long-time limit of the current amplitude variance, s2 ≡ 〈|w|2〉 − 〈u〉2 − 〈v〉2,222
is given by223
s2 ∼∑
i=x,y
〈τ2i 〉 (r + γi)
ρ20r(f2 + (r + γi)
2) . (15)
Under the assumption of isotropic wind-stress, the variance may be expressed as224
s2 ∼ 〈|τ |2〉 (r + γ) /(ρ20r
(f2 + (r + γ)2
)). Note that in deriving the expression225
above, we assumed that the autocorrelation function of the wind-stress is isotropic.226
In general, this is not the case (Monahan, 2012); however, for simplicity, we present227
here the results for this special case, and the generalization to the more realistic228
8 Bel and Ashkenazy
case is trivial. Due to the Gaussian nature of the wind-stress and the linearity229
of the model, the current components have Gaussian distributions that are fully230
characterized by the mean and variance. Note that the result of eq. (15) holds for231
any distribution of the wind-stress components, as long as the two-point correlation232
function is given by eq. (13). Equation (15) is easily derived by substituting the233
two-point correlation function (eq. (13)) in eq. (8). Similar to the first idealized234
case discussed above, it is clear that the second moment vanishes at the limit of a235
short wind-stress correlation time, (γ � r and γ � f).236
4 Details of the numerical models237
We performed two types of numerical simulations to validate the analytical deriva-238
tions presented above. The first type was a simple integration of eqs. (4)–these239
numerical solutions are provided, both to validate the analytical results and to240
present the solutions of cases not covered by the analytical solutions. In the second241
type of simulation, we used a state-of-the-art oceanic GCM, the MITgcm (MIT-242
gcm Group, 2010), to test the applicability of our analytical derivations when the243
spatial variability of the bottom topography and the nonlinearity of the ocean244
dynamics (Gill, 1982) are taken into account.245
4.1 MITgcm setup and details246
The MITgcm solves the primitive equations (MITgcm Group, 2010) and is imple-247
mented here using Cartesian coordinates with a lateral resolution of 1 km (or 10248
km) with a 50 × 50 grid points. We consider open (periodic) and closed physical249
boundaries of the domain (the results of the closed boundary setup are presented250
only in the last figure). There is one vertical level that expresses a 2D parabolic251
basin with a maximum depth of 90m, as shown in Fig. 5. The basin is situated in a252
plateau of 100m depth. The use of the partial cell option of the MITgcm enables us253
to handle depth variability, even with a single vertical level. The integration time254
step is 10s, and the overall integration time is two years. Water temperature and255
salinity are kept constant. The horizontal viscosity is 1m2/s, and the vertical one256
is 1×10−4m2/s. We use the linear bottom drag option of the MITgcm, and the ef-257
fective drag coefficient is depth dependent where the mean bottom drag coefficient258
is 1× 10−5s−1. We also use the implicit free surface scheme of the MITgcm.259
There are different ways to introduce heterogeneity to the water flow, for ex-260
ample, by forcing the water surface with spatially variable wind stress. Changes261
in density due to changes in temperature and salinity, as a result of spatially and262
temporally variable surface heat and freshwater fluxes, can also enrich water dy-263
namics. The main goal of the MITgcm numerical experiments presented here is264
to validate the analytical results presented above when advection and horizontal265
viscosity are taken into account. To enrich water dynamics, we chose to vary the266
water depth when including (excluding) the boundaries of the domain–there is no267
particular reason for choosing this way over another. We find the depth variability268
simpler, from a numerical point of view, but still rich enough to allow us to study269
the boundary effects on the water dynamics and to compare the numerical results270
with the derived analytical expressions.271
The relation between current and wind statistics 9
4.2 The Weibull distribution272
It is widely accepted that the PDF of ocean currents follows the Weibull distribu-273
tion (Gille and Smith, 1998, 2000, Chu, 2008, 2009, Ashkenazy and Gildor, 2011);274
therefore, we use this distribution in our numerical tests. For consistency, we pro-275
vide here a brief review of its properties and the methods we used to implement276
temporally correlated and uncorrelated time series. The Weibull PDF is defined277
for positive values of the variable, x > 0, and is characterized by two parameters,278
k and λ:279
Wk,λ(x) =k
λ
(xλ
)k−1exp(− (x/λ)k). (16)
where λ and k are both positive. λ and k are usually referred to as the scale280
and shape parameters of the distribution. The cumulative Weibull distribution281
function is given by:282
Fk,λ(x) = 1− e−(x/λ)k . (17)
It is possible to express the moments of the Weibull distribution, 〈xm〉, using k283
and λ:284
〈xm〉 = λmΓ(1 +
m
k
), (18)
where Γ is the Gamma function. Thus, it is sufficient to calculate the first and285
second moments of a time series and, from them, to find the k and λ that char-286
acterize the Weibull distribution (assuming that we have a priori knowledge that287
the series is Weibull distributed). The k parameter can be calculated using any288
two different moments (n, m) of the time series by solving (numerically) a tran-289
scendental equation 〈xn〉m/n/〈xm〉 = (Γ (1 + n/k))m/n /Γ (1 + m/k), and λ can290
be estimated by using the first moment (the mean) and the estimated k. It is also291
possible to calculate the k parameter from the slope of the hazard function of292
the Weibull distribution W (x)/(1−F (x)) = (k/λ)(x/λ)k−1 when a log-log plot is293
used. In this manuscript, we use the second and the fourth moments to derive the294
values of k and λ. Previous studies used other approximations to estimate k and295
λ (Monahan, 2006, Chu, 2008).296
A special case of the Weibull distribution, called the Rayleigh distribution, is297
obtained when k = 2. It can be associated with the distribution of the magnitude298
of a vector whose components are two independent, Gaussian-distributed, random299
variables; i.e., if x and y are Gaussian-distributed independent random variables,300
then the distribution of s =√x2 + y2 is a Rayleigh distribution (Weibull distri-301
bution with k = 2).302
It is fairly easy to generate an uncorrelated Weibull-distributed time series303
(using a simple transformation rule). However, below we use time series that are304
both Weibull distributed and temporally correlated. Such time series are generated305
as follows:306
(i) Generate uncorrelated, Gaussian-distributed, time series.307
(ii) Introduce temporal correlations by applying Fourier transform to the time308
series from (i), multiply the obtained power spectrum by the power spectrum309
corresponding to the desired correlations (in our case, exponentially decaying310
temporal correlations) and apply an inverse Fourier transform.311
(iii) Generate uncorrelated Weibull-distributed time series.312
10 Bel and Ashkenazy
(iv) Rank order the time series from step (iii) according to the time series of step313
(ii).314
The resulting time series is both temporally correlated and Weibull distributed.315
5 Results316
The analytical results presented above for the idealized cases highlight the impor-317
tant role played by the temporal correlations of the wind-stress in determining318
the statistics of surface ocean currents. The behaviors at the limits of long-range319
temporal correlations (rT � 1) and short-range temporal correlations (T → 0)320
are intuitive, once derived. The existence of an optimal correlation time, at which321
the average current amplitude is maximal, is less trivial. A similar maximum was322
reported by Gonella (1971), as described above. McWilliams and Huckle (2005)323
(right panel of Fig. 11 of their paper) studied the Ekman layer rectification using324
the K-profile-parameterization and observed a maximum in the depth-averaged325
variance of the current speed as a function of the Markov memory time. In what326
follows, we present the results of the numerical tests.327
In Fig. 6, we show the second moment of the current amplitude for the isotropic,328
step-like wind-stress model. The analytical results (eq. (11)) are compared with329
the numerical solution of the Ekman model (eqs. (4)) and the MITgcm modeling330
of the currents in a simple artificial lake (details of which are provided above, using331
open boundaries with 1 km resolution). One can see that the dependence of the332
average current amplitude on the constant wind-stress duration is non-monotonic333
(due to the fact that the Coriolis effect is significant – |f | > r) and that there is334
an excellent agreement between all the results. At the equator (not shown), where335
f = 0, the second moment increases monotonically to 〈|τ |2〉/(r2ρ20
)as a function336
of T .337
It was previously argued that, under certain conditions, the wind-amplitude338
(directly related to the wind-stress) PDF is well approximated by the Weibull339
distribution (Monahan, 2010); we thus chose, in our demonstrations, a Weibull340
distribution of wind-stress. In Fig. 7, we present the Weibull kcurrent parameter341
(eq. (16)) of the current distribution versus the constant wind-stress duration for342
two different values (kwind = 1 and kwind = 2) of the wind-stress Weibull kwind343
parameter (again, we only present here the results of the isotropic case). The an-344
alytical value of kcurrent was found based on the ratio between the fourth moment345
and the square of the second moment; we note, however, that the current’s PDF is346
not necessarily a Weibull PDF, and we chose to quantify the current’s PDF using347
the k parameter of the Weibull distribution for presentation purposes only. The348
findings of Monahan (2008) are consistent with the lower range of kwind values349
used here. One can see that, for short correlation times, the current amplitude350
exhibits a Rayleigh distribution (kcurrent = 2), independent of the wind-stress dis-351
tribution. This corresponds to Gaussian distributions of the current components.352
In the other limit of long constant wind-stress periods, kcurrent converges to kwind353
(not shown for kwind = 2).354
The above results were obtained for the maximal value of the Coriolis param-355
eter, i.e., f at the pole. In Fig. 8(a,b), we show kcurrent versus kwind for different356
values of the constant wind-stress duration, T , and the two limiting values of the357
Coriolis parameter, f , at the equator and at the pole. Here again, one observes an358
The relation between current and wind statistics 11
excellent agreement between the numerical and the analytical results for both val-359
ues of f . The limits of short and long temporal correlations of the wind, at which360
kcurrent = 2 and kcurrent = kwind, correspondingly, are clearly demonstrated.361
In Figs. 9 and 10, we present similar results for the case of an exponentially362
decaying correlation function of the wind-stress. As mentioned in Sec. 3.2, the363
expression for the current’s second moment (eq. (15)) holds for any distribution364
of the wind-stress (given that the temporal autocorrelation follows eq. (13)), and365
thus it is possible to obtain analytically the current’s second moment in the case366
of Weibull-distributed wind-stress whose k value does not necessarily equal 2.367
We generated a Weibull- distributed time series that has exponentially decaying368
temporal correlations, following the procedure described in Section 4.2. In Fig.369
9, we present the second moment of the current amplitude versus the correlation370
time (1/decay rate) of the correlation function. The existence of an optimal decay371
rate, for which the average current amplitude is maximal, is demonstrated for this372
case as well. One notable difference is the absence of the secondary maxima points373
which appeared in the step-like model. For this case, one may easily find that the374
maximal average amplitude of the currents is obtained for γ = |f | − r, assuming375
that |f | ≥ r.376
In Fig. 10, we present kcurrent versus kwind. Here again, we obtain an excellent377
agreement between the predicted and the numerically obtained limiting behav-378
iors. It is important to note that the good agreement between the MITgcm results379
and the analytical results has been proven to be valid only for the setup described380
above, using a fine resolution of 1 km and open boundaries. Different behavior may381
occur for different scenarios, such as regions close to the boundary of the domain,382
spatially variable wind, complex and steep bottom topography, and vertically and383
spatially variable temperature and salinity. Using a setup with closed boundaries,384
we found that close to the boundaries of the artificial lake, the results showed a385
significant deviation due to the boundary effects that were neglected in the ana-386
lytical model. The results of the MITgcm model with closed boundary conditions387
and coarser resolution of 10 km are shown in Fig. 11. All the other figures show388
the MITgcm results for the case of open boundaries. Moreover, we used a spatially389
uniform wind-stress in our simulation and have not considered the more realistic390
case of non-uniform wind-stress.391
6 Summary392
In summary, we have shown, in two idealized cases, that the PDF of wind-driven393
ocean currents depends on the temporal correlations of the wind. For short-range394
correlations, the current speed approaches zero, and the PDF of its components395
is Gaussian. For long-range temporal correlations of the wind, the currents’ PDF396
is proportional to the wind-stress PDF. The different cases considered here and397
the MITgcm simulations described above suggest that the existence of a maximal398
current speed as a function of the temporal correlation time is not unique to the399
cases studied here, and may also be relevant in more realistic types of temporal400
correlations and setups. Analysis that is based on the space-dependent model401
(either in the horizontal or the vertical dimensions or both) is a natural extension402
of the current study and will allow us to compare the analytical results to altimetry-403
based surface currents and to study non-local phenomena.404
12 Bel and Ashkenazy
Acknowledgements The research leading to these results has received funding from the Eu-405
ropean Union Seventh Framework Programme (FP7/2007-2013) under grant number [293825].406
References407
Airy GB (1845) Tides and waves. Encyclopedia Metropolitana 5408
Ashkenazy Y, Gildor H (2011) On the probability and spatial distribution of ocean409
surface currents. J Phys Oceanogr p in press410
Chu PC (2008) Probability distribution function of the upper equatorial pacific411
current speeds. Geophys Res Lett 35:L12,606412
Chu PC (2009) Statistical characteristics of the global surface current speeds ob-413
tained from satellite altimetry and scatterometer data. IEEE J of Selected Topics414
in Applied Earth Observations and Remote Sensing 2(1):27–32415
Cushman-Roisin B (1994) Introduction to geophysical fluid dynamics, 1st edn.416
Prentice Hall417
Ekman VW (1905) On the influence of the earths rotation in ocean-currents. Arch418
Math Astron Phys 2:1–52419
Gill AE (1982) Atmosphere–ocean dynamics. Academic Press, London420
Gille ST, Smith SGL (1998) Probability density functions of large-scale turbulence421
in the ocean. Phys Rev Lett 81(23):5249–5252422
Gille ST, Smith SGL (2000) Velocity probability density functions from altimetry.423
J Phys Oceanogr 30(1):125–136424
Gonella J (1971) A local study of inertial oscillations in the up- per layers of the425
ocean. Deep Sea Res 18:775–788426
Kalnay E, Kanamitsu M, Kistler R, Collins W, Deaven D, Gandin L, Iredell M,427
Saha S, White G, Woollen J, Zhu Y, Chelliah M, Ebisuzaki W, Higgins W,428
Janowiak J, Mo KC, Ropelewski C, Wang J, Leetmaa A, Reynolds R, Jenne R,429
Joseph D (1996) The NCEP/NCAR 40-year reanalysis project. Bulletin of the430
American Meteorological Society 77(3):437–471431
Kase RH, Olbers DJ (1979) Wind-driven inertial waves observed during phase III432
of GATE. Deep Sea Res 26:191–216433
Kundu PK (1976) An analysis of inertial oscillations observed near Oregon coast.434
J Phys Oceanogr 6:879–893435
McWilliams JC, Huckle E (2005) Ekman layer rectification. J Phys Oceanogr436
36:1646–1659437
MITgcm Group (2010) MITgcm User Manual. Online doc-438
umentation, MIT/EAPS, Cambridge, MA 02139, USA,439
http://mitgcm.org/public/r2 manual/latest/online documents/manual.html440
Monahan AH (2006) The probability distribution of sea surface wind speeds. Part441
I: Theory and SeaWinds observations. J Climate 19:497–520442
Monahan AH (2007) Empirical models of the probability distribution of sea surface443
wind speeds. J Climate 20:5798–5814444
Monahan AH (2008) Probability distribution of sea surface wind stresses. Geophys445
Res Lett 34:L05,704446
Monahan AH (2010) The probability distribution of sea surface wind speeds: Ef-447
fects of variable surface stratification and boundary layer thickness. J Climate448
23(19):5151–5162449
The relation between current and wind statistics 13
Monahan AH (2012) The temporal autocorrelation structure of sea surface winds.450
J Climate 25:6684–6700451
Pollard RT, Millard RC (1970) Comparison between observed and simulated wind-452
generated inertial oscillations. Deep Sea Res 17:813–821453
Seguro JV, Lambert TW (2000) Modern estimation of the parameters of the454
Weibull wind speed distribution for wind energy analysis. J Wind Engineer-455
ing and Industrial Aerodynamics 85:75–84456
Simons TJ (1980) Circulation models of lakes and inland seas. Can Bull Fish457
Aquat Sci 203:146458
Van-Kampen NG (1981) Stochastic Processes in Physics and Chemistry. North-459
Holland460
14 Bel and Ashkenazy
Fig. 1 A map of the correlation time (in days) of the wind-stress magnitude. We define thecorrelation time as the time at which the normalized auto-correlation function first dropsbelow 1/e. Clearly, there is a large variability in the correlation time. One can also notice theremarkably shorter correlation time overland compared with that over the ocean (except forthe polar regions). The map is based on the NCEP/NCAR reanalysis six-hourly wind data forthe period of 1993-2010 (Kalnay et al, 1996). The spatial resolution of the data is 2.5◦ × 2.5◦.Other definitions of the correlation time yield qualitatively the same results. The black contourline indicates a one-day correlation time, and the white line indicates the coast line. A grid(light blue lines) of 30◦ × 30◦ was superimposed on the map.
-2 -1 0 1 2u
-2
-1
0
1
2
v
T=π/f (1/4 day)
-2 -1 0 1 2u
-2
-1
0
1
2
v
(a) (b)
T=2π/f (1/2 day)
Fig. 2 Typical trajectories of the current velocity components for the step-like wind stress.In the left panel, (a) the constant wind-stress duration, T = π/f , corresponds to changesin the wind-stress after the velocity vector completes ∼ 1/2 circle. On the right panel, (b),T = 2π/f , that is, the wind-stress changes after a full circle of the velocity field. In bothpanels, the Coriolis parameter, f , was set to its value at the pole. The symbols indicate thetimes at which the wind-stress was changed.
The relation between current and wind statistics 15
-3
-2
-1
0
1
2
3
uv|w|
0 2 4 6 8 10 12Time (days (4π/f))
-3
-2
-1
0
1
2
3
Inte
gra
ted v
eloci
ty (
m2/s
)
(a)
(b)
T=π/f
T=2π/f
Fig. 3 Typical time series of the current velocity components and amplitude. In the top panel,(a), the constant wind-stress duration, T = π/f . In the bottom panel, (b), T = 2π/f . In bothpanels, the Coriolis parameter, f , was set to its value at the pole. The symbols indicate thetimes at which the wind-stress was changed.
|w|2
(f2
+r
2)ρ
2 0
|τ|2
r=0.1f
r=0.25f
r=0.5f
r=0.75f
r=f
2π 10π 12π
6
4
14π8π0
2
0
T×f4π 6π
Fig. 4 The second moment of the current amplitude (eq. (11)) versus the constant wind-stressduration, T . The different lines correspond to different values of the Rayleigh friction, r.
16 Bel and Ashkenazy
x (km)
y (
km
)
−40 −30 −20 −10 0 10 20 30 40
−40
−30
−20
−10
0
10
20
30
40
100
110
120
130
140
150
160
170
180
Fig. 5 Contour plot of the bathymetry used in the MITgcm simulations.
The relation between current and wind statistics 17
0 2 4 6 8 10 12 14
T (104 s)
0
0.5
1
1.5
2
2.5m
2(m4 /s
2 )
kwind
=1, numerick
wind=2, numeric
kwind
=1, analytick
wind=2, analytic
kwind
=1, MITgcmk
wind=2, MITgcm
Fig. 6 The second moment of the current amplitude versus the constant wind-stress duration,T . The analytical results (solid lines) are compared with the numerical solution of the Ekmanmodel (empty symbols) and with the MITgcm modeling of the currents in a simple artificiallake with open boundaries using a 1 km resolution. The wind-stress amplitude was drawnfrom a Weibull distribution with two different values of the kwind parameter as specified inthe figure. The Coriolis frequency, f ≈ 1.45 × 10−4s−1, was set to its value in the pole, andthe constant wind duration associated with this Coriolis frequency is indicated by the verticaldashed line. The Rayleigh friction parameter is r = 10−5s−1.
18 Bel and Ashkenazy
0 2 4 6 8 10 12 14
T (104 s)
1
1.5
2
k curr
ent
kwind
=1, numerick
wind=2, numeric
kwind
=1, analytick
wind=2, analytic
Fig. 7 The Weibull kcurrent parameter of the current distribution versus the constant wind-stress duration, T . The analytical results (solid lines) are compared with the numerical solutionof the Ekman model (empty symbols). The wind-stress was drawn from a Weibull distribu-tion with two different values of the kwind parameter as specified in the figure. The Coriolisfrequency, f , was set to its value in the pole and the corresponding duration is indicated bythe vertical dashed line.
The relation between current and wind statistics 19
1
1.2
1.4
1.6
1.8
2
2.2
kcu
rren
t
1 1.2 1.4 1.6 1.8 2k
wind
1
1.2
1.4
1.6
1.8
2
2.2
kcu
rren
t
lines - analyticsymbols - numeric
T=106
s
T=104 s
T=103 s
T=5x104 s
T=104
s
T=103 s
T=106
s
T=5x104 s
(a)
(b)
Fig. 8 The Weibull kcurrent parameter of the current distribution versus the kwind parameterof the step-like wind-stress distribution. The analytical results (solid lines) are compared withthe numerical solution of the Ekman model (symbols). The Coriolis parameter, f , in panel (a),is 0, corresponding to its value at the equator. In panel (b) , f ≈ 1.45×10−4s−1, correspondingto its value at the pole and is larger than r = 10−5s−1. The different lines correspond todifferent constant wind-stress durations, T , as indicated in the figure.
20 Bel and Ashkenazy
0 2 4 6 8 10
1/γ (104 s)
0.5
1
1.5
2
m2 (
m4 /s
2 )
numericanalyticMITgcm
Fig. 9 The second moment of the current amplitude for the case of wind-stress with expo-nentially decaying temporal correlations. The analytical results (solid line) are compared withthe results of numerical integration of Ekman equations (red circles) and with the currentssimulated by the MITgcm model (blue triangles) in a simple lake with open boundaries (seethe description in section 4). The wind-stress amplitude was drawn from a Weibull distribu-tion with kwind = 1.5. The Coriolis frequency, f ≈ 7.27 × 10−5s−1, corresponds to its valueat lat = 30◦ and r = 10−5s−1. Note that the second moment is plotted versus 1/γ, to alloweasier comparison with the results of the step-like wind-stress case.
The relation between current and wind statistics 21
1 1.2 1.4 1.6 1.8 2 2.2k
wind
1
1.2
1.4
1.6
1.8
2
2.2
k curr
ent
1/γ=104 (s)
1/γ=105 (s)
1/γ=106 (s)
Fig. 10 The Weibull kcurrent parameter of the current amplitude distribution versus the kwind
parameter of the wind-stress with exponentially decaying temporal correlations for differentcorrelation times (corresponding to 1/γ as indicated in the figure). The results were obtainedusing different realizations of the wind-stress. The solid lines are drawn to guide the eye. Thevalues of the Coriolis parameter, f and the Rayleigh friction parameter, r, are the same as inFig. 9.
22 Bel and Ashkenazy
0 2 4 6 8 10
1/γ (104 s)
0
0.5
1
1.5
2
m2
theorycenterperiphery
Fig. 11 Comparison of the analytical results for the second moment of the current amplitudewith the results of the detailed MITgcm model with closed boundaries and a coarse resolutionof 10 km. It is shown that for a wide range of the wind-stress correlation time (the exponen-tially decaying temporal correlations were considered), the MITgcm results agree qualitativelywith the analytical results of the over-simplified model. On the other hand, the results in theperiphery show a significant deviation. These results, combined with the fact that for openboundaries, there was a good agreement between the analytical and the simulations results,suggest that, as expected, the simple model fails to capture the boundary effects.