Treatment of observational uncertainty

A particular fit between model and observation is deemed acceptable and the associated estimate of β z_s and z_b is deemed plausible, when X² is reduced to a value consistent with random measurement error in the observations of u _o, s and b. Measurement error can be characterized by the random vector variable ξ _u and the random scalar variables ξ_s and ξ_b, which define the deviation of the observed fields from their true values, i.e.

and

where the subscript t signifies the true fields. We assume that the probability distributions of ξ_u, ξ_s and ξ_b are Gaussian and that their means and variances are

where

and

Under the above assumptions, a new positive definite random variable χ² can be defined:

The probability distribution of this new random variable is known, and is referred to as the χ² distribution with three degrees of freedom (Reference Menke,Menke, 1989, p. 32).

Our effort to fit model to observation is guided by the assumption that X ² and χ² are closely related. In practical terms, we accept a given fit of model to observation and its associated estimate of β, z_s and z_b , if X ² is reduced below the bound which χ² can exceed only with a probability of 10%. The rationale behind this acceptability criterion is that further improvements to the fit would not be differentiated from the effects of random measurement error. This bound is approximately χ² ≈ 6.25 (Reference Mendenhall, and Schaeffer,Mendenhall and Schaeffer, 1973, p. A32).

An additional unknown which affects ice-stream surface velocity is the flow-law rate constant B involved in the relation between strain rate and deviatoric stress. This constant depends on temperature and crystal fabric. Direct observation of Β is possible only in the laboratory, therefore its variation on Ice Stream Ε must be inferred from assumptions about the temperature distribution and ice fabric of the ice stream. We lack this information, so we are forced either to include B as an unknown variable to he determined by the control method or accept simple assumptions about the variation of B across the ice stream (e.g. that Β is a constant).

As described below, we shall try both strategies in a series of sensitivity tests which provide alternative estimates of β, z_s and z_b . Experience reveals that lack of uniqueness demonstrated by these sensitivity tests does not bar us from strong conclusions concerning the dynamics of the ice stream and the role of basal friction.

Control-method algorithm

The control method is implemented by an iterative algorithm (see Reference MacAyeal,MacAyeal, 1993a) described as follows. A first guess of the spatial distribution of β, z_s and z_b , is chosen arbitrarily. This guess is then used with the finite-element model to produce an associated velocity field u. The production of the model velocity field may itself involve several iterations of the model-solution procedure to ensure satisfaction of the flow law.) The model/observation misfit, i.e. u – u ₀, is then used to force a model of the adjoint trajectory equations (see Reference MacAyeal,MacAyeal (1993a) for a description of this nomenclature) which produces the vector function Λ. The Λ is then used to evaluate the gradient of X ² in the space of unknown parameters (i.e. nodal values of the three unknown fields β z_s and z_b ). A conjugate-gradient method (Reference Press,, Flannery,, Teukolsky, and Vetterling,Press and others, 1989, p. 301) is then used to search for an improved guess of the unknown variables in a direction down the gradient of X ².

The conjugate-gradient method is an improvement of the steepest-descent method used previously by Reference MacAyeal,MacAyeal (1993a). We use the International Mathematical Sub-routine Library (IMSL) implementation of the conjugate-gradient method provided by the FORTRAN sub-routine UMCGG. The guess-improvement process is repeated many times until X² converges to a minimum. If this minimum is below the acceptability criterion of 6.25 defined above, the derived β, z_s and z_b are accepted as being consistent with the observed flow of the ice stream. The derived basal stress τ is then determined from β and u using Equation (1).

Model equations

The model equations are written

and

where h = z_s — z_b, g = 9.81 m s⁻² is the acceleration of gravity, x and y are horizontal coordinates on a horizontal plane approximating the Earth’s surface (we use a Lambert equal-area projection centered on 90° S) and ρ = 917 kg m⁻³ is the density of ice (assumed constant).

The model makes use of an effective viscosity υ to represent Glen’s flow law. This representation involves a temperature-dependent rate constant Β and a flow-law exponent n = 3

The velocity in our model is assumed independent of z, thus the rate constant B used in the above expression represents a depth-averaged value of the flow-law rate constant derived from laboratory studies of ice (e.g. Reference Barnes,, Tabor, and Walker,Barnes and others, 1971; Reference Thomas, and MacAyeal,Thomas and MacAyeal, 1982).

Boundary conditions

The boundary condition sufficient for solution of the above equations is u - u ₀ at the edges of the finite-element mesh. For further description of the model equations, boundary conditions and their implementation in the control method, we refer to Reference MacAyeal,MacAyeal (1993a). Our specification of velocity as the boundary condition ensures that the stresses at the boundaries are determined as output of the control-method algorithm. This means that aspects of the ice-stream force balance, including the forces transmitted through the sides of the ice stream, are free to vary within the context of fitting the finite-element model to the observed velocity field.

Finite-element algorithm

The finite-element mesh used to cover the study area Γ is shown in Figure 1. This mesh involves 8595 triangular elements (within which fields are linearly interpolated) of variable size and 4563 nodes. Near the edges of the finite-element domain, the mesh resolution is 0.5 km to account for strong gradients in horizontal velocity. In the center of the mesh, the resolution (i.e. the length of a triangular element’s shortest side) is either 2 km or 8 km depending on the density of surface-velocity data points. The regions having the lowest resolution (8 km) are chosen based on the density of velocity-measurement data points. Low resolution in data-poor regions is justified by the need to damp unwanted structure (i.e. wiggles) in the derived fields which cannot be sufficiently constrained by observation.

The finite-element rendition of the forward and adjoint problems described by Reference MacAyeal,MacAyeal (1993a) produced a banded symmetric, positive-definite matrix which is solved using the Cholesky factorization (we used the Lapack software routine SPBTRF on the U.S. National Aeronautics and Space Administration’s Cray C90 sited at Goddard Space Flight Center) and back substitution (Lapack’s SPBTRS).

Observational uncertainty

The uncertainties σ|u|, σ_s and σ_b are also specified in the definition of X ². This specification accounts for random measurement error, interpolation error and, in the case of s and b, map-digitization error. For σ|u| we take 6 m a⁻¹, the estimate of standard deviation of point-displacement and image-co-registration errors cited by Reference Bindschadler, and Scambos,Bindschadler and Scambos (1991). Errors associated with s and b are more difficult to estimate. We take σ_s, = 10 m and σ_b, = 30 m. These values are consistent with the error analysis of radio-echo sounding surveys of Ice Streams A, Β and C reported by Reference Retzlaff,, Lord, and Bentley,Retzlaff and others (1993).

The error analysis presented by Reference Retzlaff,, Lord, and Bentley,Retzlaff and others (1993) does not apply directly to our problem, because it involves an airborne-radar survey of a different ice stream and with higher spatial resolution than that used to determine s and h in the Antarctic folio maps (Reference Rose,Rose, 1979; Reference Drewry,Drewry, 1983). Our estimates of σ_s and σ_b may therefore be too low. However, experience dictates that we accept these low estimates, because σ_s and σ_b are also important in characterizing the amplitude of irregularly distributed surface and basal undulations (such as seen in the Landsat image of Figure 2). Such undulations are necessary to achieve a close fit between the model and observed velocity fields. If σ_s and σ_b are chosen to be too large, this fit exacts a price of unrealistically rough surface and basal topography. By using the σ_s and σ_b values consistent with analysis of undulations reported by Reference Retzlaff,, Lord, and Bentley,Retzlaff and others (1993), we restrict the amplitude of undulations derived by our control method to be similar to that found on Ice Streams A, Β and C.

Derived fields

Derived fields associated with our standard inversion are displayed in Figures 3–7 along with the data used to generate them. A statistical summary is presented in Table 2. A close-up view of the derived fields for a limited 47 km by 33 km region covering the middle section of the ice stream is displayed in Figure 8 (see Figure 7 for the location of this sub-region).

Fig. 3.

Observed surface-velocity magnitude |u_o| (left), model surface-velocity magnitude (right) (c.i. 50 m a⁻¹) derived from the standard inversion.

Fig. 4.

Velocity misfit |u — u_o| for the standard inversion. The standard deviation of observed velocity error is approximately 6 m a⁻¹.

Fig. 5.

Surface topographies. Previous estimates (left), result of standard inversion z_s (right) (c.i. 50 m).

Sticky spots

The basal-friction field derived by our standard inversion (Fig. 7) displays the result we expected from previous study (Bindschsdler and Scambos, 1991; Reference MacAyeal,MacAyeal, 1992); namely, that basal friction is dominated by an irregular distribution of basal sticky spots. (We do not show a map of the basal-coupling parameter β², because of its similarity in appearance to the map of |τ|.) To emphasize this irregularity, we give area-distribution functions and cumulative area-distribution functions of |τ| and β² in Figures 9 and 10. The relative importance of basal stress in resisting; the local driving stress is also shown in Figure 9. A comparison of the area-distribution functions associated with |τ| and pgh|∇z_s | indicates that basal stress eaten a weds wider range of variation than the driving stress and is, on average, smaller in magnitude than the driving stress. This indicates that horizontal deviatoric-stress gradients are an important resistive element in the force balance of the ice stream. Approximately 30% of Γ possesses a basal stress less than 10³ Pa. This result is consistent with the widespread presence of deformable basal till as a lubricant beneath the ice stream (Reference Kamb,Kamb, 1991). Sticky spots, taken here to be regions where |τ| exceeds 5 × 10⁴Pa, comprise about 15% Γ.

Fig. 6.

Basal topographies. Previous estimate b (left), result of standard inversion z_s (right) (c.i. 50 m).

Fig. 7.

Derived |τ| (Pa) from the standard inversion. Magnitudes less than 10 Pa are denoted by clear (white) color. Box indicates sub-region shown in detail in Figure 8.

The area-average of |τ| in the standard inversion is 1.41 × 10⁴Pa (Table 2). The area-average driving stress is 4.87 × 10⁴Pa. Thus, on average, about 29% of the driving stress is balanced by basal friction. The role of basal stress in balancing driving stress in this general sense remains somewhat uncertain because our knowledge of flow-law rate parameter variation is limited. The sensitivity experiments described in the following sections consider alternative specifications of Β and provide some sense of the uncertainty of the area average of |τ| (i.e. ±10³ Pa; Table 3). These sensitivity experiments do not explore the uncertainty generated by systematic variations of Β that could accompany strain heating and crystal-fabric development in the shear margins of the ice stream.

Ice softening in shear margins has been studied by Reference MacAyeal,(MacAyeal 1989; see Table 2 and Figure 11 of this reference) in the context of Ice Stream B. MacAyeal’s sensitivity experiments using the same finite-element model as that used here indicate that ice softening by a factor of 10 could increase the horizontal velocity of Ice Stream Β by a factor of 2. This result implies that our 29% figure for the proportion of driving stress balanced by basal stress is a lower bound. A reasonable guess of the upper bound based on MacAyeal’s (1989) sensitivity experiments is 60%. At present, we do not pursue this question further because we believe that a reliable answer can be reached only when more accurate surface and basal topography data are available, shear margins are better resolved m the velocity data and model physics can account for conditions where [H]/[L] ≈ 1 at the ice stream sides.

Table 3.

Effect of flow-law rate constant an derived basal friction

Sticky spots in some regions appear to be linked into narrow bauds several tens of kilometers in length and only several kilometers in width. We call these features sticky strings. We have no explanation for sticky strings. In some circumstances, they correspond with surface lineations visible in Landsat imagery (e.g. the linear features which cut obliquely across the Landsat image in Figure 2). A possible explanation of this correspondence is that structure in the bedrock serves to generate an efficient basal-water conduit which locally decreases the water pressure at the ice/bed contact and within pores of the subglacial sediment (personal communication from R. C. A. Hindmarsh, 1993). Such a decrease would be expected to make the ice/bed contact more rigid and lead to a sticky string aligned with bedrock structure.

To examine our results for the suggestion of a basal-sliding law, we present in Figure 11 a scatter plot of |u| vs |τ| where each point represents conditions in one of the 8595 triangular elements. The lack of coherent structure in Figure 11 suggests that basal friction is not a simple function of ice velocity such as would be the case if the subglacial bed could be idealized as a subglacial till of uniform viscosity and thickness. Our results suggest that the essential physical controls on the basal friction of Ice Stream Ε (e.g. water pressure) have yet to be determined.

Test of model assumptions

As shown in Table 2, the area-average of |τ| in the standard inversion is 1.41 × 10⁴Pa. This value is relatively small compared to the criterion [τ] ≪ 10⁵ which justifies the assumptions on which our treatment of ice-stream dynamics is based. We therefore regard our model as satisfying the relationship expressed in Equation (8) in a large-scale, average sense.

Fig. 8.

Detail of derived results of standard inversion for sub-region indicated by the box in Figure 7. The dimensions of this sub-region are 47 km by 33 km. Basal stress magnitude |τ| (panel a), model velocity magnitude |u| (panel b) (c.i. 20 m a⁻¹), derived basal topography Z_b (panel c) (c.i. 50 m), velocity misfit |u — u_o| (panel d) (c.i. 20 m), observed surface-velocity magnitude (panel f) (c.i. 20 m a⁻¹).

The maximum local value of |τ| produced by our standard inversion, however, is 2.19 × 10⁵ Pa and is not small compared to the criterion [τ] ≪ 10⁵ which justifies the assumptions on which our model is based. The results of our standard inversion are therefore not consistent with the assumptions used to generate them in several limited regions surrounding the strongest sticky spots. In view of the limited extent of these regions and the close agreement between model and observed velocity attained generally throughout the ice stream, we choose to accept model errors generated by this inconsistency and defer further consideration of such errors to some future study in which the small-scale flow variation near a particular sticky spot where |τ| exceeds 10⁵ Pa is of concern.

Fig. 9.

Area-distribution function (upper) and cumulative area-distribution function (lower) for |τ| and pgh|∇z_s| determined by the standard inversion. The value of the area-distribution function (upper) for a given stress magnitude expresses the fraction of total ice-stream area which possesses a basal traction or driving stress within a small range of the given stress. The size of the stress range is variable and is designed to cover the range between 1 and 10⁶ Pa with 100 intervals of equal length on a logarithmic scale. The value of the cumulative area-distribution function (lower) for a given stress represents the fraction of total ice-stream area which possesses a basal traction or driving stress with a magnitude less than the given stress.

Surface and basal topography

As expected, z_s and z_b , derived from our standard inversion displays roughness with characteristics similar to roughness found on Ice Stream Β (Reference Retzlaff,, Lord, and Bentley,Retzlaff and others, 1993) (Figs 5 and 6). In the sub-region shown in Figure 8 (panels c and e), z_s tends to intensify over a site of increased basal friction. Examination of other regions suggests that surface-slope steepening can also occur over bumps in basal topography which may lack strong basal friction. No account is made of mass-balance considerations in this study. It is therefore possible that the relationship between surface topography, basal topography and horizontal flow derived here may be inconsistent with steady state. Further analysis of this question is deferred to the future when more accurate observations of s and b may become available through satellite altimetry and additional radio-echo sounding.

Simulated satellite image

An independent assessment of the quality of our standard inversion is made by examining the correspondence between surface features seen in the Landsat image (Fig. 2) and surface features associated with z_s . Simulated satellite images of both the derived surface topography z_s and the previous estimate s are given in Figure 12. The simulated images were constructed from the surface slope in the direction roughly parallel to the Sun azimuth of the Landsat image (Fig. 2). These images were then smoothed to reduce the prismatic visual appearance (i.e. constant gray level in each triangular element) introduced by the finite-element discretization.

Fig. 10.

Area-distribution function (upper) and cumulative area-distribution function (lower) for β² derived by the standard inversion. See the caption of Figure 9 for an explanation of the meaning of these functions.

Fig. 11.

Points represent average values of |τ| and |u| for the standard inversion in each of the 8595 elements comprising the finite-element mesh

The most striking resemblance between the Landsat image (Fig. 2) and its simulated counterpart of z_s (Fig. 12) is the general mottled appearance due to small-scale undulations. The apparent basal escarpments which slash obliquely across the ice stream’s mid-section in the Landsat image appear to have counterparts in the simulated image of z_s . In the simulated image, these oblique linear features are associated with steps in the ice-stream surface above sticky strings at the root of the obliquely slashing features in the real Landsat image.

Fig. 12.

Simulated satellite images associated with the previous estimate of surface topography s (left) and the surface topography resulting from the standard inversion z_s (right). These images represent gray-scale renderings of -∂s/∂y and -∂z_s/∂y where y is the horizontal coordinate which runs vertically (positive upward) in the map plane. Brightness of the images corresponds to illumination from the positive y direction and is approximately equivalent to the Sun angle of the Landsat image in Figure 2. The images were smoothed to eliminate patchiness associated with the finite-element discretization. Wide boundaries of relatively uniform color at the edges of the active finite-element domain are artifacts of the smoothing algorithm. A linear feature indicated by the arrow in the image of the model surface corresponds in position and geometry with the prominent linear feature of the real Landsat image displayed in Figure 2.

In contrast with the simulated image of z_s , the simulated image of s bears little resemblance to the real Landsat image. We regard the appearance of the simulated image of z_s to be more realistic than that of s, and interpret this realism as an independent check on the results of our standard inversion.

Relative importance of basal friction, surface topography and basal topography

A crucial sequence of alternative inversions (runs 2–4) were run to examine the question of whether variation in any or all of the three fields β, z_s and z_b could be arbitrarily replaced with a simple, intuitively motivated prior estimate (i.e. a guess) without significantly altering the fit between model and observed surface velocity. In these runs,

as in our standard inversion.

In run 2, we computed X ² under the assumption that the control method provides no added value to the prospect of fitting model to observation, i.e. that a good guess of the unknown fields would be sufficient. For this guess, we take β = 1.67 × 10⁴ (Pa m s⁻¹)^½, z_s = s and z_b , = b. The uniform value of β for this run is the area average of the β field derived in our standard run (Table 2). As indicated by its associated X ² value (Table 1), the hypothesis tested by run 2, i.e that a good guess of unknown ß, z_s and z_b , is sufficient, is false.

In runs 3 and 4, we tested the assumptions that the observed velocity field could be explained, in the first case, without basal friction (β = 0) and, in the second case, without adjustments of surface and basal topography from their previous estimates (z_s = s and z_b = b). Again, the X ² values obtained for these runs rule out their associated hypotheses. We are therefore forced to conclude that spatial variability in basal friction, surface topography and basal topography are equally important influences on the surface velocity of Ice Stream E. This result motivates additional future measurements of surface and basal topography that may be possible using satellite altimetry and additional airborne radio-echo sounding.

Flow-law rate-factor variation

To examine the consequences of flow-law rate-factor uncertainty, we present runs 5-8. In run 5, we assume ice to be a viscous fluid with constant viscosity ν = 10¹⁴ Pa s. In runs 6 and 7 we use uniform Β with values of 1.5 and

, respectively. The range of Β variation between runs 6 and 7 is inspired by the derived pattern of B variation on the Ross Ice Shelf (Reference Thomas, and MacAyeal,Thomas and MacAyeal, 1982). As mentioned above, we do not consider the effects of ice softening restricted to the shear margins of the ice stream.

Fig. 13.

Area-distribution functions (upper) and cumulative area-distribution functions (lower) for |τ| for runs with various treatments of the flow-law rate factor Β (runs 1 and 6–8). See the caption of Figure 9 for an explanation of the meaning of these functions. Similarity among the functions plotted here suggests that uncertainty in the specification of Β does not affect the derived τ in a qualitatively important way.

Run 5 was found to be incompatible with the observations (Table 1). A linear viscous flow law is therefore ruled out and a non-linear (e.g. power) flow law is found to be mandatory. Runs 6 and 7 were found to be compatible with observations (Table 1). These runs, plus our standard inversion

, suggest that flow-law variation of order is compatible with observations and cannot be resolved further by our control method. In all three cases, the area distribution of |τ| and the locations of prominent sticky spots were found to be comparable to that determined by our standard inversion (Fig. 13). Differences in the force balance of runs 6 and 7 (Table 3) suggest that uncertainty in Β translates to an uncertainty of approximately 10³ Pa in the area-averaged |τ|.

In run 8, we employ the control method as a means to select a variation of B, in addition to the other three unknown fields, to fit the data better. In this circumstance, we add an additional term to X ² as follows

where

and The values of B ₀ and σ_B are chosen arbitrarily but are within the range of Β variation suggested by runs 6 and 7. Run 8 produced the best fit between model and observed velocity of all the inversions we tried but its physical significance is not clear.

A map of Β determined in run 8 is displayed in Figure 14. The area-distribution function and cumulative area-distribution function of B are displayed in Figure 15. The pattern of Β derived in run 8 has greater variability in regions where velocity data points are sparse. We are therefore tempted to conclude that this pattern is determined primarily by errors in the velocity data.

Basal drag tied to bumps in basal topography

Reference Alley,Alley (1993) suggested that sticky spots may be caused by large bedrock bumps which stick into the base of the ice stream or which compromise an intervening layer of a lubricating substance such as water or basal till. The results of our standard inversion (run 1) were generated under the assumption that z_b and β² were completely independent. A scatter plot of elemental values of z_b vs β² (not shown) derived from our standard inversion fails to suggest a relationship between β² and z_b . The observations we invert therefore support an hypothesis that z_b and β² need not be related. This does not mean that the observations we invert definitively rule out such an hypothesis.

Fig. 14.

Variable, flow-law rate factor Β (Ρa s^1/3 ) derived by run 8.

Fig. 15.

Area-distribution function (upper) and cumulative area-distribution function (lower) for Β derived by run 8. See the caption of Figure 9 for an explanation of the meaning of these functions.

To examine Alley’s suggestion further, we ran run 9 under the assumption

where S = 5.0 × 10⁹ Pa s. We require the perturbation of the basal elevation above its previous estimate b to be positive and to have an amplitude of σ_b, = 30 m when the bed-coupling parameter β² rises to approximately half of a representative sticky-spot value (i.e. 5 × 10⁹ Pa s). The hypothetical relationship expressed in Equation (13) is arbitrary and bears little resemblance to the scale analysis presented by Reference Alley,Alley (1993).

For run 9, X ² < 6.25; thus, the relationship represented by Equation (13), and presumably many other such relationships which tie β² to z_b , is consistent with the observations. This result, in concert with the results of the standard inversion where β² to z_b , were assumed independent, suggests that our analysis is unable to resolve the question of how basal friction should relate to basal topography. This question perhaps can be answered when more accurate observations of s and b become available.

Comparison with previous results

Our results confirm previous suggestions that basal shear stress of the West Antarctic ice streams is extremely low (< 10⁴ Pa) except in isolated sticky spots (Reference Kamb,Kamb, 1991). These characteristics are compatible with the notion that soft, unconsolidated sediments lubricate the base of the ice stream. Our results further suggest that sticky spots, not intervening lubricated zones which are nearly stress free, appear to determine the overall contribution of basal friction to the forces which resist ice-stream discharge. While significant in the overall force budget of the ice stream, friction induced by sticky spots is not, however, the dominant resistive stress. Drag introduced at the ice-stream’s sides and at the grounding line (i.e. ice-shelf back-pressure) are also significant resistive forces. We therefore suggest that the question which is most important, basal friction or side and grounding-line drag, has no definitive answer. Both can be important.

The pattern and location of sticky spots we find in our standard inversion suggest that bedrock topography, rock type or large-scale structure such as faults or escarpments which affect water pressure can cause sticky spots below ice streams. This suggestion is in agreement with theory (Reference Alley,Alley, 1993) but a further test of theory is beyond the resolving power of our methods and data. We therefore suggest that additional measurements, perhaps involving satellite-altimeter profiling, airborne-radar sounding and passive seismology (e.g. Reference Anandakrishnan, and Bentley,Anandakrishnan and Bentley, 1993), be focused in small-scale regions where sticky spots appear on Ice Stream E.

Previous studies of ice-stream force balance (Reference Whillans, and van der Veen,Whillans and Van der Veen, 1993; Reference Whillans,, Jackson, and Tseng,Whillans and others, 1993) failed to find direct evidence for sticky spots on Ice Stream B. We do not consider our results to be incompatible with conditions found on Ice Stream B. The study areas on Ice Stream Β which appear to lack sticky spots are comparable in size with some of the sticky-spot-free zones found near the grounding line of Ice Stream Ε (e.g. as suggested by Figure 7).

Lessons for ice-stream modeling

Our results suggest that basal friction below ice streams is not determined by purely glaciological considerations. We suspect that basal friction depends crucially on the interplay between ice conditions, geologic properties of the bedrock, subglacial hydrologic conditions and the disposition of unconsolidated basal sediments. Up to now, ice-sheet modelers have attempted to rely on various basal-sliding laws as a means to characterize ice-stream flow (e.g. Reference Bentley,Bentley, 1987). The simplest of these laws express τ as a function of basal-ice velocity. Our results suggest that these simplest laws may oversimplify. The origin of sticky spots may depend as much on the geophysical properties and history of the glacial bed as with the ice stream above. Our results suggest that the intricate pattern of basal coupling that controls the present flow of ice Stream E arises from some as yet inscrutable property of the bed that escapes our understanding. Ice-stream modelers are urged to heed signs of this inscrutability and to recognize the virtue of simple basal-sliding laws that mimic the statistics of conditions found below Ice Stream E.