3.1 The topography

If the topographic height has an r.m.s. value of order 200 m, typical of abyssal hills (Goff Reference Goff2010), then $\unicode[STIX]{x1D702}_{\mathit{rms}}^{-1}$ is less than two days. Thus, even rather small topographic features produce a topographic PV with a time scale that is much less than that of the typical drag coefficient in table 2. This order-of-magnitude estimate indicates that the form stress is likely to be large. To say more about form stress we must introduce the model topography with more detail.

Figure 2.

The structure and spectrum of the topography used in this study. (a) The structure of the topography for a quarter of the full domain. The solid curves are positive contours, the dashed curves are negative contours and the thick curves mark the zero contour. (b) The 1D power spectrum. The topography has power only within the annulus $12\leqslant |\boldsymbol{k}|L\leqslant 18$ .

Figure 3.

The structure of the geostrophic contours, $\unicode[STIX]{x1D6FD}y+\unicode[STIX]{x1D702}$ , for the monoscale topography of figure 2 and for various values of $\unicode[STIX]{x1D6FD}_{\ast }$ : (a) $\unicode[STIX]{x1D6FD}_{\ast }=0.10$ , (b) $\unicode[STIX]{x1D6FD}_{\ast }=0.35$ and (c) $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . It is difficult to visually distinguish the geostrophic contours with $\unicode[STIX]{x1D6FD}_{\ast }=0$ from those with $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ in (a).

The topography is synthesized as $\unicode[STIX]{x1D702}(x,y)=\sum _{\boldsymbol{k}}\text{e}^{\text{i}\boldsymbol{k}\boldsymbol{\cdot }\boldsymbol{x}}\unicode[STIX]{x1D702}_{\boldsymbol{k}}$ , with random phases for $\unicode[STIX]{x1D702}_{\boldsymbol{k}}$ . We consider a homogeneous and isotropic topographic model illustrated in figure 2. The topography is constructed by confining the wavenumbers $\unicode[STIX]{x1D702}_{\boldsymbol{k}}$ to a relatively narrow annulus with $12\leqslant |\boldsymbol{k}|L\leqslant 18$ . The spectral cutoff is tapered smoothly to zero at the edges of the annulus. In addition to being homogeneous and isotropic, the topographic model in figure 2 is approximately monoscale, i.e. the topography is characterized by a single length scale, determined, for instance, by the central wavenumber $|\boldsymbol{k}|\approx 15/L$ in figure 2(b). To assess the validity of the monoscale approximation, we characterize the topography using the length scales

(3.2a,b ) $$\begin{eqnarray}\ell _{\unicode[STIX]{x1D702}}\stackrel{\text{def}}{=}\sqrt{\left\langle \unicode[STIX]{x1D702}^{2}\right\rangle /\left\langle |\unicode[STIX]{x1D735}\unicode[STIX]{x1D702}|^{2}\right\rangle }\quad \text{and}\quad L_{\unicode[STIX]{x1D702}}\stackrel{\text{def}}{=}\sqrt{\left\langle |\unicode[STIX]{x1D735}\unicode[STIX]{x1D6FB}^{-2}\unicode[STIX]{x1D702}|^{2}\right\rangle /\left\langle \unicode[STIX]{x1D702}^{2}\right\rangle }.\end{eqnarray}$$

For the model in figure 2,

(3.3a,b ) $$\begin{eqnarray}\ell _{\unicode[STIX]{x1D702}}=0.0690L\quad \text{and}\quad L_{\unicode[STIX]{x1D702}}=0.0707L.\end{eqnarray}$$

(It should be recalled that the domain size is $2\unicode[STIX]{x03C0}L\times 2\unicode[STIX]{x03C0}L$ .) Because $\ell _{\unicode[STIX]{x1D702}}\approx L_{\unicode[STIX]{x1D702}}$ , we conclude that the topography in figure 2 is monoscale to a good approximation, and we use the slope-based length $\ell _{\unicode[STIX]{x1D702}}$ as the typical length scale of the topography.

The isotropic homogeneous monoscale topographic model adopted here has no claims to realism. However, the monoscale assumption greatly simplifies many aspects of the problem because all relevant second-order statistical characteristics of the model topography can be expressed in terms of the two-dimensional quantities $\unicode[STIX]{x1D702}_{\mathit{rms}}$ and $\ell _{\unicode[STIX]{x1D702}}$ , e.g. $\langle (\unicode[STIX]{x1D6FB}^{-2}\unicode[STIX]{x1D702}_{x})^{2}\rangle =\ell _{\unicode[STIX]{x1D702}}^{2}\unicode[STIX]{x1D702}_{\mathit{rms}}^{2}/2$ . The main advantage of monoscale topography is that, despite the simplicity of its spectral characterization, it exhibits the crucial distinction between open and closed geostrophic contours; see figure 3.

3.2 Non-dimensionalization

There are four time scales in the problem: the topographic PV $\unicode[STIX]{x1D702}_{\mathit{rms}}^{-1}$ , the dissipation $\unicode[STIX]{x1D707}^{-1}$ , the period of topographically excited Rossby waves $(\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}})^{-1}$ and the advective time scale associated with the forcing $\ell _{\unicode[STIX]{x1D702}}\unicode[STIX]{x1D707}/F$ . From these four time scales, we construct the three main non-dimensional control parameters:

(3.4a-c ) $$\begin{eqnarray}\unicode[STIX]{x1D707}_{\ast }\stackrel{\text{def}}{=}\unicode[STIX]{x1D707}/\unicode[STIX]{x1D702}_{\mathit{rms}},\quad \unicode[STIX]{x1D6FD}_{\ast }\stackrel{\text{def}}{=}\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}/\unicode[STIX]{x1D702}_{\mathit{rms}}\quad \text{and}\quad F_{\ast }\stackrel{\text{def}}{=}F/(\unicode[STIX]{x1D707}\unicode[STIX]{x1D702}_{\mathit{rms}}\ell _{\unicode[STIX]{x1D702}}).\end{eqnarray}$$

The parameter $\unicode[STIX]{x1D6FD}_{\ast }$ is the ratio of the planetary PV gradient over the r.m.s. topographic PV gradient. There is a fourth parameter $L/\ell _{\unicode[STIX]{x1D702}}$ which measures the scale separation between the domain and the topography. We assume that as $L/\ell _{\unicode[STIX]{x1D702}}\rightarrow \infty$ , there is a regime of statistically homogeneous two-dimensional turbulence. In other words, as $L/\ell _{\unicode[STIX]{x1D702}}\rightarrow \infty$ , the flow becomes asymptotically independent of $L/\ell _{\unicode[STIX]{x1D702}}$ , so that the large-scale flow $\bar{U}$ and other statistics, such as $\overline{E_{\unicode[STIX]{x1D713}}}$ , are independent of the domain size $L$ .

Besides the control parameters in (3.4), additional parameters are required to characterize the topography. For example, in the case of a multiscale topography, the ratio $L_{\unicode[STIX]{x1D702}}/\ell _{\unicode[STIX]{x1D702}}$ characterizes the spectral width of the power-law range. Another simplification of the monoscale topography is that we do not have to contend with these additional topographic parameters.

3.3 Geostrophic contours

We refer to the contours of constant $\unicode[STIX]{x1D6FD}y+\unicode[STIX]{x1D702}$ as the geostrophic contours. Closed geostrophic contours enclose isolated pools within the domain – see figure 3(a) – while open geostrophic contours thread through the domain in the zonal direction, connecting one side to the other – see figure 3(c). The transition between the two limiting cases is controlled by $\unicode[STIX]{x1D6FD}_{\ast }$ . Figure 3(b) shows an intermediate case with a mixture of both closed and open geostrophic contours.

It is instructive to consider the extreme case $\unicode[STIX]{x1D6FD}=0$ . Then, only the geostrophic contour $\unicode[STIX]{x1D702}=0$ is open and all other geostrophic contours are closed. This intuitive conclusion relies on a special property of the random topography in figure 2: the topography $-\unicode[STIX]{x1D702}$ is statistically equivalent $+\unicode[STIX]{x1D702}$ . In other words, if $\unicode[STIX]{x1D702}(x,y)$ is in the ensemble, then so is $-\unicode[STIX]{x1D702}(x,y)$ . A more detailed discussion of this conclusion is given in the review paper by Isichenko (Reference Isichenko1992).

If $\unicode[STIX]{x1D6FD}$ is non-zero but small, in the sense that $\unicode[STIX]{x1D6FD}_{\ast }\ll 1$ , then most of the domain is within closed contours; see figure 3(a). The planetary PV gradient $\unicode[STIX]{x1D6FD}$ is too small relative to $\unicode[STIX]{x1D735}\unicode[STIX]{x1D702}$ to destroy local pools of closed geostrophic contours. However, $\unicode[STIX]{x1D6FD}$ dominates the long-range structure of the geostrophic contours and opens up narrow channels of open geostrophic contours.

The other extreme is $\unicode[STIX]{x1D6FD}_{\ast }\gg 1$ . In this case, illustrated in figure 3(c), all geostrophic contours are open. Because of its geometric simplicity, the situation with $\unicode[STIX]{x1D6FD}_{\ast }\gg 1$ is the easiest to analyse and understand. Unfortunately, the difficult case in figure 3(a) is the most relevant to oceanic conditions. In §§ 3.4 and 3.5, we illustrate the two cases using numerical solutions of (1.1) and (2.3).

Figure 4.

A solution with closed geostrophic contours: $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ , $F_{\ast }=2.20$ and $\unicode[STIX]{x1D707}_{\ast }=10^{-2}$ . (a) The evolution of the large-scale zonal flow $U(t)$ (dashed) and the form stress $\langle \unicode[STIX]{x1D713}\unicode[STIX]{x1D702}_{x}\rangle$ (solid). (b) The evolution of $E_{\unicode[STIX]{x1D713}}$ (solid) and $E_{U}$ (dashed). The dash-dotted line in (b) is the energy level of the standing eddies, $E_{\bar{\unicode[STIX]{x1D713}}}=\langle |\unicode[STIX]{x1D735}\bar{\unicode[STIX]{x1D713}}|^{2}\rangle /2$ . (c) A snapshot of the relative vorticity, $\unicode[STIX]{x1D701}$ (shaded), at $\unicode[STIX]{x1D707}t=10$ in one quarter of the domain overlying the topographic PV (the solid contours are positive $\unicode[STIX]{x1D702}$ and the dashed contours are negative). (d) The time mean $\bar{\unicode[STIX]{x1D701}}$ . A movie showing the evolution of $q=\unicode[STIX]{x1D701}+\unicode[STIX]{x1D702}$ and $\unicode[STIX]{x1D713}-Uy$ from rest is found in the supplementary material (movie 1) available at https://doi.org/10.1017/jfm.2017.482.

Figure 5.

A solution with $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ , $F_{\ast }=2.20$ and $\unicode[STIX]{x1D707}_{\ast }=10^{-2}$ . (a) The speed of the time-mean flow, $|\bar{\boldsymbol{U}}|$ , is indicated by colours; the geostrophic contours $\unicode[STIX]{x1D6FD}y+\unicode[STIX]{x1D702}$ are shown as white curves. (b) Surface plot of the total time-mean streamfunction, $\bar{\unicode[STIX]{x1D713}}-\bar{U}y$ .

3.4 An example with mostly closed geostrophic contours: $\unicode[STIX]{x1D6FD}_{\ast }=0.1$

Figures 4 and 5 show a numerical solution for a case with mostly closed geostrophic contours; this is the $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ ‘boxed’ point indicated in figure 1. In this illustration, we use the Southern Ocean parameter values given in table 2 with $1024^{2}$ grid points. The system is evolved using the ETDRK4 time-stepping scheme of Cox & Matthews (Reference Cox and Matthews2002) with the refinement of Kassam & Trefethen (Reference Kassam and Trefethen2005).

Figure 4(a) shows the evolution of the large-scale flow $U(t)$ and the form stress $\langle \unicode[STIX]{x1D713}\unicode[STIX]{x1D702}_{x}\rangle$ . After a spin-up of duration ${\sim}\unicode[STIX]{x1D707}^{-1}$ , the flow achieves a statistically steady state in which $U(t)$ fluctuates around the time mean $\bar{U}$ . In figure 4(a), the form stress $\langle \bar{\unicode[STIX]{x1D713}}\unicode[STIX]{x1D702}_{x}\rangle$ balances almost 98 % of $F$ , so that $U(t)$ is very much smaller than $F/\unicode[STIX]{x1D707}$ in (3.1). The time mean of the large-scale flow is $\bar{U}=1.70~\text{cm}~\text{s}^{-1}$ , which is 2.2 % of $F/\unicode[STIX]{x1D707}$ . Figure 4(b) shows the evolution of the energy. The eddy energy $E_{\unicode[STIX]{x1D713}}$ is approximately 50 times greater than the large-scale energy $E_{U}$ . With the decomposition of $\unicode[STIX]{x1D713}$ in (2.6), the time-mean eddy energy $\overline{E_{\unicode[STIX]{x1D713}}}$ is decomposed into $E_{\bar{\unicode[STIX]{x1D713}}}+\overline{E_{\unicode[STIX]{x1D713}^{\prime }}}$ ; the dash-dotted line in figure 4(b) is the energy of the standing component $E_{\bar{\unicode[STIX]{x1D713}}}$ . The transient eddies are less energetic than the standing eddies. This is also evident by comparing the snapshot of the relative vorticity $\unicode[STIX]{x1D701}$ in figure 4(c) with the time mean $\bar{\unicode[STIX]{x1D701}}$ in figure 4(d). Many features in the snapshot are also seen in the time mean.

Figures 4(c) and 4(d) also show that there is anticorrelation between the time-mean relative vorticity and the topographic PV: $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})=-0.53$ , where the correlation between two fields $a$ and $b$ is

(3.5) $$\begin{eqnarray}\text{corr}(a,b)\stackrel{\text{def}}{=}\langle ab\rangle /\sqrt{\langle a^{2}\rangle \langle b^{2}\rangle }.\end{eqnarray}$$

Another statistical characterization of the solution is that $\text{corr}(\bar{\unicode[STIX]{x1D713}},\unicode[STIX]{x1D702}_{x})=0.06$ , showing that a weak correlation between the standing streamfunction $\bar{\unicode[STIX]{x1D713}}$ and the topographic PV gradient $\unicode[STIX]{x1D702}_{x}$ can produce a form stress balancing approximately 98 % of the applied wind stress.

The most striking characterization of the time-mean flow is that it is very weak in most of the domain. Figure 5(a) shows that most of the flow through the domain is channelled into a relatively narrow band centred very roughly at $y/(2\unicode[STIX]{x03C0}L)=0.25$ . This ‘main channel’ coincides with the extreme values of $\unicode[STIX]{x1D701}$ and $\bar{\unicode[STIX]{x1D701}}$ evident in figure 4(c,d) (it should be noticed that figure 4(c,d) shows only a quarter of the flow domain). Outside the main channel, the time-mean flow is weak. We emphasize that $\bar{U}=1.70~\text{cm}~\text{s}^{-1}$ is an unoccupied mean that is not representative of the larger velocities in the main channel; the channel velocities are 40–50 times larger than  $\bar{U}$ .

Figure 5(b) shows the streamfunction $\bar{\unicode[STIX]{x1D713}}(x,\!y)-\bar{U}y$ as a surface above the $(x,y)$ -plane. The time-mean streamfunction appears as a terraced hillside; the mean slope of the hillside is $-\bar{U}$ , and stagnant pools, with constant $\bar{\unicode[STIX]{x1D713}}-\bar{U}y$ , are the flat terraces carved into the hillside. The existence of these stagnant dead zones is explained by the closed-streamline theorems of Batchelor (Reference Batchelor1956) and Ingersoll (Reference Ingersoll1969) (see the discussion in § 6.2). The dead zones are separated by boundary layers; the strongest of these boundary layers is the main channel which appears as the large cliff roughly at $y/(2\unicode[STIX]{x03C0}L)=0.25$ in figure 5(b). The main channel is determined by a narrow band of geostrophic contours that are opened by the small- $\unicode[STIX]{x1D6FD}$ effect; this provides an open path for flow through the disordered topography.

3.5 An example with open geostrophic contours: $\unicode[STIX]{x1D6FD}_{\ast }=1.38$

Figures 6 and 7 show a solution for the same parameters as in figures 4 and 5, except that $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ ; this is the $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ ‘boxed’ point indicated in figure 1. The geostrophic contours are open throughout the domain. The most striking difference when compared with the example in § 3.4 is that there are no dead zones; the flow is more evenly spread throughout the domain. This can be seen by comparing figure 7 with figure 5. The time-mean streamfunction in figure 7(b) is not ‘terraced’. Instead, $\bar{\unicode[STIX]{x1D713}}-\bar{U}y$ in figure 7(b) is better characterized as a bumpy slope.

Figure 6.

A solution with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ and open geostrophic contours. All other parameters are as in figure 4. The panels are also as in figure 4. It should be noted that the colour scale is different between (c) and (d). A movie showing the evolution of $q=\unicode[STIX]{x1D701}+\unicode[STIX]{x1D702}$ and $\unicode[STIX]{x1D713}-Uy$ from rest can be found in the supplementary material (movie 2).

Figure 7.

A solution with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . All other parameters are as in figure 5. (a) The speed of the time-mean flow, $|\bar{\boldsymbol{U}}|$ (colours); the geostrophic contours $\unicode[STIX]{x1D6FD}y+\unicode[STIX]{x1D702}$ are shown as white curves. (b) Surface plot of the total time-mean streamfunction, $\bar{\unicode[STIX]{x1D713}}-\bar{U}y$ .

The large-scale flow is $\bar{U}=4.68~\text{cm}~\text{s}^{-1}$ , which is again much smaller than the flow that would exist in the absence of topography; $\bar{U}$ is only 6 % of $F/\unicode[STIX]{x1D707}$ . The eddy energy $E_{\unicode[STIX]{x1D713}}$ is roughly 15 times larger than the large-scale flow energy $E_{U}$ . Moreover, the energy of the transient eddies, shown in figure 6(b), is in this case much larger than that of the standing eddies. This is also apparent by comparing the instantaneous and time-mean relative vorticity fields in figure 6(c,d). In anticipation of the discussion in § 8, we remark that these strong transient eddies act as PV diffusion on the time-mean QGPV (Rhines & Young Reference Rhines and Young1982).

In contrast to the example of § 3.4, the relative vorticity is positively correlated with the topographic PV: $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})=0.23$ . Because of the strong transient eddies, the sign of $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})$ is not apparent by visual inspection of figure 6(c,d). The form-stress correlation is $\text{corr}(\bar{\unicode[STIX]{x1D713}},\unicode[STIX]{x1D702}_{x})=0.15$ . Again, this weak correlation is sufficient to produce a form stress balancing 94 % of the wind stress.

4.1 Flow regimes: the lower branch, the upper branch, eddy saturation and the drag crisis

Keeping $\unicode[STIX]{x1D6FD}_{\ast }$ fixed and increasing the wind forcing $F_{\ast }$ from very small values, we find that the statistically equilibrated solutions show either one of the two characteristic behaviours depicted in figure 8.

For $\unicode[STIX]{x1D6FD}=0$ , or for values of $\unicode[STIX]{x1D6FD}_{\ast }$ much less than 1, we find that the equilibrated time-mean large-scale flow $\bar{U}$ scales linearly with $F_{\ast }$ when $F_{\ast }$ is very small. On this lower branch, the large-scale velocity is

(4.2) $$\begin{eqnarray}\bar{U}\approx F/\unicode[STIX]{x1D707}_{\mathit{eff}},\quad \text{with }\unicode[STIX]{x1D707}_{\mathit{eff}}\gg \unicode[STIX]{x1D707}.\end{eqnarray}$$

In § 6, we provide an analytic expression for the effective drag $\unicode[STIX]{x1D707}_{\mathit{eff}}$ in (4.2); the $\bar{U}$ based on this analytic expression for $\unicode[STIX]{x1D707}_{\mathit{eff}}$ is shown by the dashed lines in figure 8. As $F_{\ast }$ increases, $\bar{U}$ transitions to a different linear relation with

(4.3) $$\begin{eqnarray}\bar{U}\approx F/\unicode[STIX]{x1D707}.\end{eqnarray}$$

On this upper branch, the form stress is essentially zero and $F$ is balanced by the bare drag $\unicode[STIX]{x1D707}$ .

Figure 8.

(a) The equilibrated large-scale mean flow $\bar{U}$ as a function of $F_{\ast }$ for cases with $\unicode[STIX]{x1D6FD}_{\ast }=0$ and $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . Results are shown for three different monoscale topography realizations (each denoted with a different marker symbol: *, ▵, ○), all with the spectrum in figure 2(c). Other parameters are given in table 2, e.g. $\unicode[STIX]{x1D707}_{\ast }=10^{-2}$ . (b) A detailed view of the transition from the lower- to the upper-branch solution for the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . (c) The hysteretic solutions for one of the topography realizations with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . The dashed lines in (a) correspond to asymptotic expressions derived in § 6 and the dash-dotted lines in all panels mark the solution $U=F/\unicode[STIX]{x1D707}$ .

For the $\unicode[STIX]{x1D6FD}=0$ case shown in figure 8, the transition between the lower and upper branches occurs in the range $0.6<F_{\ast }<3$ ; the equilibrated $\bar{U}$ increases by a factor of more than 200 within this interval. On the other hand, for $\unicode[STIX]{x1D6FD}_{\ast }$ larger than 0.05, we find a quite different behaviour, illustrated in figure 8 by the runs with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . On the lower branch, $\bar{U}$ grows linearly with $F$ with a constant $\unicode[STIX]{x1D707}_{\mathit{eff}}$ , as in (4.2). However, the linear increase in $\bar{U}$ eventually ceases and instead $\bar{U}$ then grows at a much more slower rate as $F$ increases. For the case $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ shown in figure 8, $\bar{U}$ only doubles as $F$ is increased over 150-fold from $F_{\ast }=0.2$ to 30. We identify this regime, in which $\bar{U}$ is insensitive to changes in $F$ , with the ‘eddy saturation’ regime of Straub (Reference Straub1993). As $F$ increases further, the flow exits the eddy saturation regime via a ‘drag crisis’ in which the form stress abruptly vanishes and $\bar{U}$ increases by a factor of over 200 as the solution jumps to the upper branch (4.3). In figure 8, this drag crisis is a discontinuous transition from the eddy saturated regime to the upper branch. The drag crisis, which requires non-zero  $\unicode[STIX]{x1D6FD}$ , is qualitatively different from the continuous transition between the upper and lower branches which is characteristic of flows with small (or zero)  $\unicode[STIX]{x1D6FD}_{\ast }$ .

Figure 8 shows results obtained with three different realizations of monoscale topography, namely the topography illustrated in figure 2(a) and two other realizations with the monoscale spectrum of figure 2(b). The large-scale flow $\bar{U}$ is insensitive to these changes in topographic detail. In this sense, the large-scale flow is ‘self-averaging’. However, the location of the drag crisis depends on differences between the three realizations. Figure 8(b) shows that the location of the jump from lower to upper branch is realization-dependent: the three realizations jump to the upper branch at different values of  $F_{\ast }$ .

The case with $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ , which corresponds a value close to realistic (cf. table 2), does show a drag crisis, i.e. a discontinuous jump from the lower to the upper branch at $F_{\ast }\approx 3.9$ ; see figure 1. However, the eddy saturation regime, i.e. the regime in which $\bar{U}$ grows with wind-stress forcing at a rate less than linear, is not nearly as pronounced as in the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ shown in figure 8(a).

4.2 Hysteresis and multiple flow patterns

Starting with a severely truncated spectral model of the atmosphere introduced by Charney & DeVore (Reference Charney and Devore1979), there has been considerable interest in the possibility that topographic form stress might result in multiple stable large-scale flow patterns which might explain blocked and unblocked states of atmospheric circulation. Focusing on atmospheric conditions, Tung & Rosenthal (Reference Tung and Rosenthal1985) concluded that the results of low-order truncated models are not reliable guides to the full nonlinear problem: although multiple stable states still exist in the full problem, these occur only in a restricted parameter range that is not characteristic of the Earth’s atmosphere.

With this meteorological background in mind, it is interesting that in the oceanographic parameter regime emphasized here, we easily found multiple equilibrium solutions on either side of the drag crisis. After increasing $F$ beyond the crisis point, and jumping to the upper branch, we performed additional numerical simulations by decreasing $F$ and using initial conditions obtained from the upper-branch solutions at larger values of $F$ . Thus, we moved down the upper branch, past the crisis, and determined a range of wind-stress forcing values with multiple flow patterns. Figure 8(c), with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ , shows that multiple states coexist in the range $11\leqslant F_{\ast }\leqslant 29$ . It should be noted that for the quasirealistic case with $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ , multiple solutions exist only in the limited parameter range $2.9\leqslant F_{\ast }\leqslant 3.9$ . These coexisting flows differ qualitatively. The lower-branch flows, being near the drag crisis, have an important transient-eddy component and almost all of $F$ is balanced by form stress; the example discussed in connection with figures 4 and 5 is typical. On the other hand, the coexisting upper-branch solutions are steady (that is, $\unicode[STIX]{x1D713}^{\prime }=U^{\prime }=0$ ), and nearly all of the wind stress is balanced by bottom drag, so that $\unicode[STIX]{x1D707}U/F\approx 1$ .

4.3 A survey

In this section, we present a suite of solutions, all with $\unicode[STIX]{x1D707}_{\ast }=10^{-2}$ . The main conclusion from these extensive calculations is that the behaviour illustrated in figure 8 is representative of a broad region of parameter space.

Figure 9(a) shows the ratio $\unicode[STIX]{x1D707}\bar{U}/F$ as a function of $F_{\ast }$ for seven different values of $\unicode[STIX]{x1D6FD}_{\ast }$ . The three series with $\unicode[STIX]{x1D6FD}_{\ast }\leqslant 0.10$ are ‘small- $\unicode[STIX]{x1D6FD}$ ’ cases in which closed geostrophic contours fill most of the domain. The other four series, with $\unicode[STIX]{x1D6FD}_{\ast }\geqslant 0.35$ , are ‘large- $\unicode[STIX]{x1D6FD}$ ’ cases in which open geostrophic contours fill most of the domain. For small values of $F_{\ast }$ in figure 9(a), the flow is steady ( $\unicode[STIX]{x1D713}^{\prime }=U^{\prime }=0$ ) and $\unicode[STIX]{x1D707}\bar{U}/F$ does not change with $F$ . This is the lower-branch relation (4.2) in which $\bar{U}$ varies linearly with $F$ with an effective drag coefficient $\unicode[STIX]{x1D707}_{\mathit{eff}}$ . As $F_{\ast }$ is increased, this steady flow becomes unstable and the strength of the transient-eddy field increases with  $F$ .

Figure 9(b) shows a detailed view of the eddy saturation regime and the drag crisis. The dashed lines on the left of figure 9(a,b) show the analytic results derived in § 6. For the large- $\unicode[STIX]{x1D6FD}$ cases, the form stress makes a very large contribution to the large-scale momentum balance prior to the drag crisis. We emphasize that, although the drag  $\unicode[STIX]{x1D707}$ does not directly balance  $F$ in this regime, it does play a crucial role in producing non-zero form stress  $\langle \bar{\unicode[STIX]{x1D713}}\unicode[STIX]{x1D702}_{x}\rangle$ . In all of the solutions summarized in figure 9, non-zero $\unicode[STIX]{x1D707}$ is required so that the flow is asymmetric upstream and downstream of topographic features; this asymmetry induces non-zero $\langle \bar{\unicode[STIX]{x1D713}}\unicode[STIX]{x1D702}_{x}\rangle$ .

Figure 9.

(a) The ratio $\unicode[STIX]{x1D707}\bar{U}/F$ as a function of the non-dimensional forcing,  $F_{\ast }$ , for seven values of $\unicode[STIX]{x1D6FD}_{\ast }$ . The dashed lines are asymptotic results in (6.3). (b) A detailed view of the shaded lower part of (a), showing the eddy saturation regime and the drag crisis. The dashed line is the asymptotic result in (6.5).

Figure 10.

(a) The index in (4.4) measures the strength of the transient eddies as a function of the forcing $F_{\ast }$ . (b) A detailed view of the shaded lower-left part of (a). The onset of transient eddies is signalled by the large jump in the fluctuation index.

Figure 11.

(a) The equilibrated large-scale flow, $\bar{U}$ , scaled with $\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}$ as a function of the non-dimensional forcing for various values of $\unicode[STIX]{x1D6FD}_{\ast }$ . The dashed curves indicate the upper-branch analytic result from § 7 (also scaled with $\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}$ ). (b) An expanded view of the shaded part of (a) which shows the eddy saturation regime.

In figure 10, we use

(4.4) $$\begin{eqnarray}\sqrt{\overline{{U^{\prime }}^{2}}}/\bar{U}\end{eqnarray}$$

as an indication of the onset of the transient-eddy instability and as an index of the strength of the transient eddies. Remarkably, the onset of the instability is roughly at $F_{\ast }=1.5\times 10^{-2}$ for all values of $\unicode[STIX]{x1D6FD}_{\ast }$ . The onset of transient eddies is the sudden increase in (4.4) by a factor of approximately $10^{4}$ or $10^{5}$ in figure 10(b). The transient eddies result in a reduction of $\unicode[STIX]{x1D707}\bar{U}/F$ . For the large- $\unicode[STIX]{x1D6FD}$ runs, this is the eddy saturation regime. In the presentation in figure 9(a), the eddy saturation regime is the decease in $\unicode[STIX]{x1D707}\bar{U}/F$ that occurs once $0.03<F_{\ast }<0.3$ (depending on $\unicode[STIX]{x1D6FD}_{\ast }$ ). The eddy saturation regime is terminated by the drag-crisis jump to the upper branch where $\unicode[STIX]{x1D707}\bar{U}/F\approx 1$ . This coincides with vanishing of the transients. On the upper branch, the flow becomes steady, $\unicode[STIX]{x1D713}^{\prime }=U^{\prime }=0$ ; see figure 10(a).

Figure 12.

(a) Correlation of the standing-eddy vorticity $\bar{\unicode[STIX]{x1D701}}$ with the topographic PV $\unicode[STIX]{x1D702}$ for $\unicode[STIX]{x1D6FD}_{\ast }=0$ , 0.10 and 1.38. For $\unicode[STIX]{x1D6FD}_{\ast }=0$ , the correlation $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})$ is always negative; for $F_{\ast }=10^{-3}$ , $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})=-1.35\times 10^{-3}$ . (b) Correlation of $\bar{\unicode[STIX]{x1D713}}$ with $\unicode[STIX]{x1D702}_{x}$ .

Figure 11 shows the eddy saturation regime that is characteristic of the three series with $\unicode[STIX]{x1D6FD}_{\ast }\geqslant 0.35$ . Eddy saturation occurs for forcing in the range

(4.5) $$\begin{eqnarray}0.1\lessapprox F_{\ast }\lessapprox 30.\end{eqnarray}$$

In this regime, the large-scale flow is limited to the relatively small range

(4.6) $$\begin{eqnarray}0.06\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}\lessapprox \bar{U}\lessapprox 0.25\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}.\end{eqnarray}$$

In anticipation of analytic results from the next section, we note that in (4.6) $\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}$ is the speed of Rossby waves excited by topography with typical length scale $\ell _{\unicode[STIX]{x1D702}}$ .

Figure 12 shows the correlations $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})$ and $\text{corr}(\bar{\unicode[STIX]{x1D713}},\unicode[STIX]{x1D702}_{x})$ as a function of the forcing $F_{\ast }$ for three values of $\unicode[STIX]{x1D6FD}_{\ast }$ . In the most weakly forced cases, $\bar{\unicode[STIX]{x1D701}}$ is positively correlated with $\unicode[STIX]{x1D702}$ . As the forcing $F$ increases, $\bar{\unicode[STIX]{x1D701}}$ and $\unicode[STIX]{x1D702}$ become anticorrelated. However, for $\unicode[STIX]{x1D6FD}_{\ast }=0$ , the correlation $\text{corr}(\bar{\unicode[STIX]{x1D701}},\unicode[STIX]{x1D702})$ is negative for all values of $F$ . For the monoscale topography used here, the term  $\langle \unicode[STIX]{x1D702}\text{D}\bar{\unicode[STIX]{x1D701}}\rangle$ , which is the only source of enstrophy in the time average of (A 1b ) if $\unicode[STIX]{x1D6FD}=0$ , can be approximated as $(\unicode[STIX]{x1D707}+\unicode[STIX]{x1D708}/\ell _{\unicode[STIX]{x1D702}}^{2})\langle \bar{\unicode[STIX]{x1D701}}\unicode[STIX]{x1D702}\rangle$ . Therefore, in this case, $\langle \bar{\unicode[STIX]{x1D701}}\unicode[STIX]{x1D702}\rangle$ must be negative (see the discussion in appendix A).

6.1 The case with either $\unicode[STIX]{x1D707}_{\ast }\gg 1$ or $\unicode[STIX]{x1D6FD}_{\ast }\gg 1$

Assuming that lengths scale with $\ell _{\unicode[STIX]{x1D702}}$ , the ratio of the terms on the left-hand side of (6.1) is

(6.2a,b ) $$\begin{eqnarray}\unicode[STIX]{x1D6FD}\unicode[STIX]{x1D713}_{x}/\text{J}\left(\unicode[STIX]{x1D713},\unicode[STIX]{x1D702}\right)=O(\unicode[STIX]{x1D6FD}_{\ast })\quad \text{and}\quad \unicode[STIX]{x1D707}\unicode[STIX]{x1D701}/\text{J}\left(\unicode[STIX]{x1D713},\unicode[STIX]{x1D702}\right)=O(\unicode[STIX]{x1D707}_{\ast }).\end{eqnarray}$$

If $\unicode[STIX]{x1D707}_{\ast }\gg 1$ or if $\unicode[STIX]{x1D6FD}_{\ast }\gg 1$ , then $\text{J}\left(\unicode[STIX]{x1D713},\unicode[STIX]{x1D702}\right)$ is negligible relative to one, or both, of the other two terms on the left-hand side of (6.1). In that case, one can neglect the Jacobian in (6.1) and adapt the QL expression (5.5) to determine the effective drag of monoscale topography as

(6.3) $$\begin{eqnarray}\unicode[STIX]{x1D707}_{\mathit{eff}}=\unicode[STIX]{x1D707}+\frac{\unicode[STIX]{x1D707}\unicode[STIX]{x1D702}_{\mathit{rms}}^{2}\ell _{\unicode[STIX]{x1D702}}^{2}}{\unicode[STIX]{x1D707}^{2}\ell _{\unicode[STIX]{x1D702}}^{2}+\unicode[STIX]{x1D6FD}^{2}\ell _{\unicode[STIX]{x1D702}}^{4}+\unicode[STIX]{x1D707}\ell _{\unicode[STIX]{x1D702}}\sqrt{\unicode[STIX]{x1D707}^{2}\ell _{\unicode[STIX]{x1D702}}^{2}+\unicode[STIX]{x1D6FD}^{2}\ell _{\unicode[STIX]{x1D702}}^{4}}}.\end{eqnarray}$$

In simplifying the QL expression (5.5) to the linear result (6.3), we have neglected $U$ relative to either $\unicode[STIX]{x1D6FD}\ell _{\unicode[STIX]{x1D702}}^{2}$ or $\unicode[STIX]{x1D707}\ell _{\unicode[STIX]{x1D702}}$ . This simplification is appropriate in the limit $F_{\ast }\rightarrow 0$ . The expression in (6.3) is accurate within the shaded region in figure 14. The dashed lines in figures 8 and 9(a) that correspond to the series with $\unicode[STIX]{x1D6FD}_{\ast }\geqslant 0.35$ indicate the approximation $\bar{U}\approx F/\unicode[STIX]{x1D707}_{\mathit{eff}}$ with $\unicode[STIX]{x1D707}_{\mathit{eff}}$ in (6.3).

Figure 14.

Schematic for the three different parameter regions in the weakly forced regime. The shaded region depicts the parameter range for which the QL theory gives good predictions. For $\unicode[STIX]{x1D6FD}_{\ast }<1$ and $\unicode[STIX]{x1D707}_{\ast }<1$ , the form stress and the large-scale flow largely depend on the actual geometry of the topography and $\text{J}(\unicode[STIX]{x1D713},\unicode[STIX]{x1D702})$ cannot be neglected. The expressions show the behaviour of ${\unicode[STIX]{x1D707}_{\mathit{eff}}}_{\ast }=\unicode[STIX]{x1D707}_{\mathit{eff}}/\unicode[STIX]{x1D702}_{\mathit{rms}}$ in each parameter region.

6.2 The thermal analogy – the case with $\unicode[STIX]{x1D707}_{\ast }\lessapprox 1$ and $\unicode[STIX]{x1D6FD}_{\ast }\lessapprox 1$

When both $\unicode[STIX]{x1D707}_{\ast }$ and $\unicode[STIX]{x1D6FD}_{\ast }$ are order one or less, the term $\text{J}(\unicode[STIX]{x1D713},\unicode[STIX]{x1D702})$ in (6.1) cannot be neglected. As a result of this Jacobian, the weakly forced regime cannot be recovered as a special case of the QL approximation. In this interesting case, we rewrite (6.1) as

(6.4) $$\begin{eqnarray}\text{J}(\unicode[STIX]{x1D702}+\unicode[STIX]{x1D6FD}y,\unicode[STIX]{x1D713}-Uy)=\unicode[STIX]{x1D707}\unicode[STIX]{x1D6FB}^{2}\unicode[STIX]{x1D713}\end{eqnarray}$$

and rely on intuition based on the ‘thermal analogy’. To apply the analogy, we regard $\unicode[STIX]{x1D702}+\unicode[STIX]{x1D6FD}y$ as an effective steady streamfunction advecting a passive scalar $\unicode[STIX]{x1D713}-Uy$ . The planetary PV gradient $\unicode[STIX]{x1D6FD}$ is analogous to a large-scale zonal flow $-\unicode[STIX]{x1D6FD}$ , and the large-scale flow $U$ is analogous to a large-scale tracer gradient. The drag $\unicode[STIX]{x1D707}$ is equivalent to the diffusivity of the scalar. The form stress $\left\langle \unicode[STIX]{x1D713}\unicode[STIX]{x1D702}_{x}\right\rangle$ is analogous to the meridional flux of tracer $\unicode[STIX]{x1D713}$ by the meridional velocity $\unicode[STIX]{x1D702}_{x}$ . The geostrophic contours are equivalent to streamlines in the thermal analogy. Usually, in the passive-scalar problem, the large-scale tracer gradient $U$ is imposed and the main goal is to determine the flux $\left\langle \unicode[STIX]{x1D713}\unicode[STIX]{x1D702}_{x}\right\rangle$ (equivalently, the Nusselt number). However, here, $U$ is unknown and must be determined by satisfying the steady version of the large-scale momentum equation (5.2).

With the thermal analogy, we can import results from the passive-scalar problem. For example, in the passive-scalar problem, at large Péclet number, the scalar is uniform within closed streamlines (Batchelor Reference Batchelor1956; Rhines & Young Reference Rhines and Young1983). The analogue of this ‘Prandtl–Batchelor theorem’ is that in the limit $\unicode[STIX]{x1D707}_{\ast }\rightarrow 0$ , the total streamfunction, $\unicode[STIX]{x1D713}-Uy$ , is constant within any closed geostrophic contour, i.e. all parts of the domain contained within closed geostrophic contours are stagnant; see also Ingersoll (Reference Ingersoll1969). This ‘Prandtl–Batchelor theorem’ explains the result in figure 15, which shows a weakly forced small-drag solution with $\unicode[STIX]{x1D6FD}_{\ast }=0$ . The domain is packed with stagnant eddies (constant $\unicode[STIX]{x1D713}-Uy$ ) separated by thin boundary layers. The ‘terraced hillside’ in figure 15(c) is even more striking than in figure 5(b). The solution in figure 5 has transient eddies, resulting a blurring of the terraced structure. The weakly forced solution in figure 15 is steady and the thickness of the steps between the terraces is limited only by the small drag, $\unicode[STIX]{x1D707}_{\ast }=5\times 10^{-3}$ .

Figure 15.

Snapshots of the flow fields for the weakly forced solution at $F_{\ast }=10^{-3}$ with dissipation $\unicode[STIX]{x1D707}_{\ast }=5\times 10^{-3}$ and $\unicode[STIX]{x1D6FD}_{\ast }=0$ . (a) The total flow speed $|\boldsymbol{U}|$ . The flow is restricted to a boundary layer around the dashed $\unicode[STIX]{x1D702}=0$ contour. (b) The total streamfunction $\unicode[STIX]{x1D713}-Uy$ . (c) Surface plot of the total streamfunction, $\unicode[STIX]{x1D713}-Uy$ . The terraced hillside structure is apparent. (In (a,b), only one quarter of the domain is shown.)

Figure 16.

The large-scale flow for weakly forced solutions ( $F_{\ast }=10^{-3}$ ) with $\unicode[STIX]{x1D6FD}=0$ as a function of $\unicode[STIX]{x1D707}_{\ast }$ . The dashed line shows the scaling law (6.5) with $c=1$ .

Isichenko et al. (Reference Isichenko, Kalda, Tatarinova, Tel’kovskaya and Yan’kov1989) and Gruzinov, Isichenko & Kalda (Reference Gruzinov, Isichenko and Kalda1990) discussed the effective diffusivity of a passive scalar due to advection by a steady monoscale streamfunction. Using a scaling argument, Isichenko et al. (Reference Isichenko, Kalda, Tatarinova, Tel’kovskaya and Yan’kov1989) showed that in the high-Péclet-number limit, the effective diffusivity of a steady monoscale flow is $D_{\mathit{eff}}=DP^{10/13}$ , where $D$ is the small molecular diffusivity and $P$ is the Péclet number; the exponent $10/13$ relies on critical exponents determined by percolation theory. Applying Isichenko’s passive-scalar results to the $\unicode[STIX]{x1D6FD}=0$ form-stress problem, we obtain the scaling

(6.5a,b ) $$\begin{eqnarray}\unicode[STIX]{x1D707}_{\mathit{eff}}=c\unicode[STIX]{x1D707}^{3/13}\unicode[STIX]{x1D702}_{\mathit{rms}}^{10/13}\quad \text{and}\quad \frac{\unicode[STIX]{x1D707}\bar{U}}{F}=\frac{1}{c}\left(\frac{\unicode[STIX]{x1D707}}{\unicode[STIX]{x1D702}_{\mathit{rms}}}\right)^{10/13},\end{eqnarray}$$

where $c$ is a dimensionless constant. Numerical solutions of (1.1) and (2.3) summarized in figure 16 confirm this remarkable ‘10/13’ scaling and show that the constant  $c$ in (6.5) is close to unity. The dashed lines in figures 8 and 9(b) corresponding to the solution suites with $\unicode[STIX]{x1D6FD}_{\ast }\leqslant 0.1$ show the scaling law (6.5) with  $c=1$ .

To summarize, the weakly forced regime is divided into the easy large- $\unicode[STIX]{x1D6FD}$ case, in which $\unicode[STIX]{x1D707}_{\mathit{eff}}$ in (6.3) applies, and the more difficult case with small or zero $\unicode[STIX]{x1D6FD}$ . In the difficult case, with closed geostrophic contours, the thermal analogy and the Prandtl–Batchelor theorem show that the flow is partitioned into stagnant dead zones; Isichenko’s $\unicode[STIX]{x1D6FD}=0$ scaling law in (6.5) is the main result in this case. The value of $\unicode[STIX]{x1D6FD}_{\ast }$ separating these two regimes in the schematic of figure 14 is identified with the value of $\unicode[STIX]{x1D6FD}$ below which (6.3) underestimates $\unicode[STIX]{x1D707}U/F$ compared with (6.5). For the topography used in this work, and taking $c=1$ in (6.5), this is $\unicode[STIX]{x1D6FD}_{\ast }=0.17$ . (If we choose $c=0.5$ , the critical value is $\unicode[STIX]{x1D6FD}_{\ast }=0.24$ .) This rationalizes why the $\unicode[STIX]{x1D6FD}=0$ result in (6.5) works better than $\unicode[STIX]{x1D707}_{\mathit{eff}}$ in (6.3) for $\unicode[STIX]{x1D6FD}_{\ast }<0.35$ ; see figure 9.

8.1 Eddy saturation regime

As wind stress increases transient eddies emerge; in figure 10, this instability of the steady solution occurs very roughly at $F_{\ast }=1.5\times 10^{-2}$ for all values of $\unicode[STIX]{x1D6FD}$ . The power integrals in appendix A show that the transient eddies gain kinetic energy from the standing eddies $\bar{\unicode[STIX]{x1D713}}$ through the conversion term $\langle \bar{\unicode[STIX]{x1D713}}\unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{E}\rangle$ , where

(8.1) $$\begin{eqnarray}\boldsymbol{E}\stackrel{\text{def}}{=}\,\overline{\boldsymbol{U}^{\prime }q^{\prime }}\end{eqnarray}$$

is the time-averaged eddy PV flux.

Figure 18 compares the numerical solutions of (1.1) and (2.3) with the prediction of the QL approximation (asterisks versus the solid QL curve) for the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . The QL approximation has a stronger large-scale flow than that of the full system in (1.1) and (2.3). Moreover, the full system is more impressively eddy saturated than its QL approximation. There are at least two causes for these failures of the QL approximation: (i) the QL approximation assumes steady flow and has no way of incorporating the effect of transient eddies on the time-mean flow and (ii) the QL approximation neglects the term $\text{J}(\bar{\unicode[STIX]{x1D713}},\bar{q})$ .

We address these points by following Rhines & Young (Reference Rhines and Young1982) and approximating the effect of the transient eddies as PV diffusion:

(8.2) $$\begin{eqnarray}\unicode[STIX]{x1D735}\boldsymbol{\cdot }\boldsymbol{E}\approx -\unicode[STIX]{x1D705}_{\mathit{eff}}\unicode[STIX]{x1D6FB}^{2}\bar{q}.\end{eqnarray}$$

In the discussion surrounding (A 6), we determine $\unicode[STIX]{x1D705}_{\mathit{eff}}$ using the time-mean eddy energy power integral (A 5b ). According to this diagnosis, the PV diffusivity is

(8.3) $$\begin{eqnarray}\unicode[STIX]{x1D705}_{\mathit{eff}}=\unicode[STIX]{x1D707}\overline{\langle |\unicode[STIX]{x1D735}\unicode[STIX]{x1D713}^{\prime }|^{2}\rangle }/\left\langle \bar{\unicode[STIX]{x1D701}}\,\bar{q}\right\rangle .\end{eqnarray}$$

Figure 19(a) shows $\unicode[STIX]{x1D705}_{\mathit{eff}}$ in (8.3) for the solution suite with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ .

Figure 18.

The eddy saturation regime for $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ (shaded). The asterisks indicate numerical solutions of (1.1) and (2.3). The circles show the numerical solutions of (1.1) and (2.3) with the added PV diffusion, $\unicode[STIX]{x1D705}_{\mathit{eff}}\unicode[STIX]{x1D6FB}^{2}q$ . The solid curve is the QL prediction (5.4) and the dash-dotted curve is the QL prediction with added PV diffusion.

Figure 19.

(a) The effective PV diffusivity, $\unicode[STIX]{x1D705}_{\mathit{eff}}$ , diagnosed from (8.3) for the series of solutions with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ . (b) The correlation $\text{corr}(\bar{\unicode[STIX]{x1D713}},\unicode[STIX]{x1D702}_{x})$ for this series of solutions (asterisks) and for the solutions with parameterized transient eddies (circles). (c) The same as (b), but showing the strength of the transient eddies using the index (4.4). (c) The same as (b), but showing the energy of the standing eddies $\bar{\unicode[STIX]{x1D713}}$ . Also shown are the power laws $F_{\ast }^{1}$ and $F_{\ast }^{2}$ as dashed black lines.

Abernathey & Cessi (Reference Abernathey and Cessi2014), in their study of baroclinic equilibration in a channel with topography, developed a two-layer QG model that incorporated the role of standing eddies in determining the transport. Abernathey & Cessi (Reference Abernathey and Cessi2014) also used an effective PV diffusion to parameterize transient eddies. However, they specified $\unicode[STIX]{x1D705}_{\mathit{eff}}$ , rather than determining it diagnostically from the energy power integral as in (8.3).

With $\unicode[STIX]{x1D705}_{\mathit{eff}}$ in hand, we can revisit the QL theory and ask for its prediction when the term $\unicode[STIX]{x1D705}_{\mathit{eff}}\unicode[STIX]{x1D6FB}^{2}q$ is added on the right-hand side of (5.1). In this way, we include the effect of the transients on the time-mean flow but do not include the effect of the term $\text{J}(\bar{\unicode[STIX]{x1D713}},\bar{q})$ . The QL prediction is only slightly improved – see the dash-dotted curve in figure 18.

To include also the effect of the term $\text{J}(\bar{\unicode[STIX]{x1D713}},\bar{q})$ , we obtain solutions of (1.1) and (2.3) with added PV diffusion in (2.3) with $\unicode[STIX]{x1D705}_{\mathit{eff}}$ as in figure 19(a). With the added PV diffusion, we find that the strength of the transient eddies is dramatically reduced; see figure 19(c). Thus, the approximation (8.2) with the PV diffusivity supplied by (8.3) is self-consistent in the sense that we do not both resolve and parameterize transient eddies. Moreover, the large-scale flow $\bar{U}$ with parameterized transient eddies is in much closer agreement with $\bar{U}$ from the solutions with transient eddies – see figure 18. This striking quantitative agreement as we vary $F_{\ast }$ shows that at least in the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ the transient eddies act as PV diffusion on the time-mean flow.

Thus, we conclude that, in addition to $\unicode[STIX]{x1D6FD}$ , the main physical mechanisms operating in the eddy saturation regime are PV diffusion via the transient eddies and the mean advection of mean PV, i.e. the term $\text{J}(\bar{\unicode[STIX]{x1D713}},\bar{q})$ .

There are three remarkable aspects of this success. First, it is important to use $\unicode[STIX]{x1D705}_{\mathit{eff}}\unicode[STIX]{x1D6FB}^{2}(\bar{\unicode[STIX]{x1D701}}+\unicode[STIX]{x1D702})$ in (8.2); if one uses only $\unicode[STIX]{x1D705}_{\mathit{eff}}\unicode[STIX]{x1D6FB}^{2}\bar{\unicode[STIX]{x1D701}}$ , then the agreement in figure 18 is degraded. Second, PV diffusion does not decrease the amplitude of the standing eddies; see figure 19(d). Furthermore, PV diffusion quantitatively captures $\text{corr}(\bar{\unicode[STIX]{x1D713}},\unicode[STIX]{x1D702}_{x})$ ; see figure 19(b).

Unfortunately, the success of the PV diffusion parameterization does not extend to cases with closed geostrophic contours (small $\unicode[STIX]{x1D6FD}_{\ast }$ ), such as the case with $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ . For small $\unicode[STIX]{x1D6FD}_{\ast }$ , the flow is strongly affected by the detailed structure of the topography. The solution described in § 3.4 shows that the flow is channelled into a few streams and thus a parameterization that does account for the actual structure of the topography is, probably, doomed to fail. In fact, for $\unicode[STIX]{x1D6FD}_{\ast }=0.1$ , the $\unicode[STIX]{x1D705}_{\mathit{eff}}$ diagnosed according to (8.3) is negative because $\langle \bar{\unicode[STIX]{x1D701}}\bar{q}\rangle <0$ .

In conclusion, the PV diffusion approximation (8.2) gives good quantitative results provided that the flow does not crucially dependent on the structure of the topography itself, i.e. for large $\unicode[STIX]{x1D6FD}_{\ast }$ , so that the geometry is dominated by open geostrophic contours. In the context of baroclinic models, eddy saturation is not captured by standard parameterizations of transient baroclinic eddies (Hallberg & Gnanadesikan Reference Hallberg and Gnanadesikan2001). Only very recently have Mak et al. (Reference Mak, Marshall, Maddison and Bachman2017) proposed a parameterization of baroclinic turbulence that successfully produces baroclinic eddy saturation. Thus, the success of (8.2) in the barotropic context, even though it depends on diagnosis of $\unicode[STIX]{x1D705}_{\mathit{eff}}$ via (8.3), is significant.

8.2 Drag crisis

In this section, we provide some further insight into the drag crisis. We argue that the requirement of enstrophy balance among the flow components leads to a transition from the lower to the upper branch as wind-stress forcing increases. We make this argument by constructing lower bounds on the large-scale flow $\bar{U}$ based on energy and enstrophy power integrals.

We consider a ‘test streamfunction’ that is efficient at producing form stress,

(8.4) $$\begin{eqnarray}\unicode[STIX]{x1D713}^{\mathit{test}}=\unicode[STIX]{x1D6FC}\unicode[STIX]{x1D702}_{x},\end{eqnarray}$$

with $\unicode[STIX]{x1D6FC}$ a positive constant to be determined by satisfying either the energy or the enstrophy power integrals from appendix A. A maximum form stress corresponds to a minimum large-scale flow $\bar{U}^{\mathit{min}}$ , which in turn can be determined by substituting (8.4) into the time-mean large-scale flow equation (5.2):

(8.5) $$\begin{eqnarray}\unicode[STIX]{x1D707}\bar{U}^{\mathit{min}}=F-\unicode[STIX]{x1D6FC}\langle \unicode[STIX]{x1D702}_{x}^{2}\rangle .\end{eqnarray}$$

We can determine $\unicode[STIX]{x1D6FC}$ so that the eddy energy power integral, that is the sum of (A 5a ) and (A 5b ) is satisfied:

(8.6) $$\begin{eqnarray}0=\bar{U}^{\mathit{min}}\,\unicode[STIX]{x1D6FC}\langle \unicode[STIX]{x1D702}_{x}^{2}\rangle -\unicode[STIX]{x1D707}\unicode[STIX]{x1D6FC}^{2}\langle |\unicode[STIX]{x1D735}\unicode[STIX]{x1D702}_{x}|^{2}\rangle -\unicode[STIX]{x1D708}\unicode[STIX]{x1D6FC}^{2}\langle (\unicode[STIX]{x1D6FB}^{2}\unicode[STIX]{x1D702}_{x})^{2}\rangle .\end{eqnarray}$$

The averages above are evaluated using properties of monoscale topography, e.g. $\langle (\unicode[STIX]{x1D6FB}^{2}\unicode[STIX]{x1D702}_{x})^{2}\rangle =\unicode[STIX]{x1D702}_{\mathit{rms}}^{2}\ell _{\unicode[STIX]{x1D702}}^{-6}/2$ . Solving (8.5) and (8.6) for $\unicode[STIX]{x1D6FC}$ and $\bar{U}^{\mathit{min}}$ , we obtain a lower bound on the large-scale flow based on the energy constraint,

(8.7) $$\begin{eqnarray}\bar{U}\geqslant \bar{U}_{E}^{\mathit{min}}\stackrel{\text{def}}{=}\frac{F}{\unicode[STIX]{x1D707}}\left[1+\frac{\unicode[STIX]{x1D702}_{\mathit{rms}}^{2}}{2(\unicode[STIX]{x1D707}+\unicode[STIX]{x1D708}/\ell _{\unicode[STIX]{x1D702}}^{2})^{2}}\right]^{-1}.\end{eqnarray}$$

Alternatively, one can determine $\unicode[STIX]{x1D6FC}$ and $\bar{U}^{\mathit{min}}$ by satisfying the eddy enstrophy power integral; the sum of (A 8a ) and (A 8b ). This leads to a second bound,

(8.8) $$\begin{eqnarray}\bar{U}\geqslant \bar{U}_{Q}^{\mathit{min}}\stackrel{\text{def}}{=}\frac{F}{\unicode[STIX]{x1D707}}\left[1-\frac{\unicode[STIX]{x1D6FD}\unicode[STIX]{x1D702}_{\mathit{rms}}^{2}\ell _{\unicode[STIX]{x1D702}}^{2}}{2(\unicode[STIX]{x1D707}+\unicode[STIX]{x1D708}/\ell _{\unicode[STIX]{x1D702}}^{2})F}\right].\end{eqnarray}$$

Thus,

(8.9) $$\begin{eqnarray}U\geqslant \max (\bar{U}_{E}^{\mathit{min}},\bar{U}_{Q}^{\mathit{min}}).\end{eqnarray}$$

The test function in (8.4) does not closely resemble the realized flow, so the bound above is not tight. Nonetheless, it does capture some qualitative properties of the turbulent solutions.

(Using a more elaborate test function with two parameters, one can satisfy both the energy and the enstrophy power integrals simultaneously and obtain a single bound. However, the calculation is much longer and the result is not much better than the relatively simple (8.9).)

The lower bound (8.9) is shown in figure 20 for the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ together with the numerical solution of the full nonlinear equations (1.1) and (2.3) and the QL prediction. Although it cannot be clearly seen, the energy bound $\bar{U}_{E}^{\mathit{min}}$ does not allow $\bar{U}$ to vanish completely, e.g. for $\unicode[STIX]{x1D707}_{\ast }=10^{-2}$ we have that $\unicode[STIX]{x1D707}\bar{U}_{E}^{\mathit{min}}/F=2\times 10^{-4}$ . On the other hand, the dominance of the enstrophy bound $\bar{U}_{Q}^{\mathit{min}}$ at high forcing values explains the occurrence of the drag crisis: the enstrophy power integral requires that the large-scale flow transitions from the lower to the upper branch as $F_{\ast }$ is increased beyond a certain value.

Figure 20.

The numerical solutions of (1.1) and (2.3) (asterisks) and the QL prediction (5.4) (solid curve). The dashed line shows the lower bound (8.9).

The bounds (8.7) and (8.8) provide a qualitative explanation for the existence of the drag crisis. However, the critical forcing predicted by the enstrophy bound dominance overestimates the actual value of the drag crisis. For example, for the case with $\unicode[STIX]{x1D6FD}_{\ast }=1.38$ shown in figure 20, $\bar{U}_{Q}^{\mathit{min}}$ becomes the lower bound at a value of $F_{\ast }$ that is approximately 240 times larger than the actual drag-crisis point. We have no reason to expect these bounds to be tight: they do not depend on the actual structure of the topography itself but only on gross statistical properties, e.g. $\unicode[STIX]{x1D702}_{\mathit{rms}}$ , $\ell _{\unicode[STIX]{x1D702}}$ , $L_{\unicode[STIX]{x1D702}}$ . For example, a sinusoidal topography

(8.10) $$\begin{eqnarray}\unicode[STIX]{x1D702}=\sqrt{2}\unicode[STIX]{x1D702}_{\mathit{rms}}\cos (x/\ell _{\unicode[STIX]{x1D702}})\end{eqnarray}$$

has identical statistical properties to the random monoscale topography used in this paper and, therefore, imposes the same bounds. However, with the topography in (8.10), there is a laminar solution with $\unicode[STIX]{x1D713}_{y}=0$ and, as a result, also $\text{J}(\unicode[STIX]{x1D713},q)=0$ . In this case, the QL solution (5.3) is an exact solution of the full nonlinear equations (1.1) and (2.3) and the bound (8.9) is tight.