Evidence for and implications of a Bering plate based on geodetic measurements from the Aleutians and western Alaska

Transcription

1 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 113,, doi: /2007jb005136, 2008 Evidence for and implications of a Bering plate based on geodetic measurements from the Aleutians and western Alaska Ryan S. Cross 1,2 and Jeffrey T. Freymueller 1 Received 26 April 2007; revised 12 February 2008; accepted 25 March 2008; published 15 July [1] Global Positioning System (GPS) measurements are used to examine the hypothesis of a clockwise rotating Bering plate. Originally proposed on the basis of seismicity, the Bering plate encompasses the Bering Sea, western Alaska, and the Aleutian Islands. GPS measurements from the Bering plate s interior show south to southwest motions of 3 5 mm/a (where a is years). We construct elastic dislocation models to determine the spatial distribution and intensity of locked patches on the Aleutian subduction interface. We use these to remove interseismic strain from the GPS observation and determine an arc translation velocity for each region of the Aleutians, revealing south and southwest motions of 4 8 mm/a. We combine the arc translation rates with measurements from the Bering plate s interior sites and estimate the Euler pole for the Bering plate relative to North America to be located at 42.5 N, E with an angular speed of 6.0 /Ma. The clockwise rotation of the Bering plate may cause left lateral faulting in interior Alaska. The Bering plate s interaction with south-central Alaska may be responsible for the decreased slip rate on the western Denali fault and for contraction across the central Alaska Range. We analyze slip partitioning along the Aleutian arc on the basis of both GPS measurements and slip azimuths of thrust earthquakes and find a systematic discrepancy between plate convergence direction and slip azimuths. We conclude that slip partitioning in the back arc only develops west of Amchitka Pass, whereas slip partitioning in the fore arc is present throughout the arc. Citation: Cross, R. S., and J. T. Freymueller (2008), Evidence for and implications of a Bering plate based on geodetic measurements from the Aleutians and western Alaska, J. Geophys. Res., 113,, doi: /2007jb Introduction [2] Plate tectonics is the unifying theory behind all geology and geophysics. The recognition of rigid tectonic plates and delineation of their boundaries has resulted in great improvements in our understanding of Earth. However, there remains at least one significant region for which there is not agreement about the plate configuration: northeast Asia, including the Bering Sea and its margins. [3] The Bering Sea is bordered by Alaska to the east, east Siberia to the west, and the Aleutian Islands to the south. This body of water covers a great expanse of continental crust that extends hundred of kilometers west from the coast of western Alaska (Figure 1). The geology of this submerged land, an area nearly the size of Alaska itself, is poorly understood. West of Alaska s Bering Sea Islands lies a passive margin connecting the submerged continental crust to the oceanic crust underlying the Aleutian Basin. The Bering Seafloor is seismically quiet with only a few events and no clear spatial pattern. 1 Geophysical Institute, University of Alaska Fairbanks, Fairbanks, Alaska, USA. 2 Now at Tetra Tech, Bothell, Washington, USA. Copyright 2008 by the American Geophysical Union /08/2007JB005136$09.00 [4] A Bering plate was proposed in early plate tectonic models, based almost entirely on seismic observations. Minster et al. [1974] proposed a Bering plate to explain the misfit of earthquake slip vectors along the Aleutian arc. However, this misfit is due to slip partitioning of oblique subduction, so Minster and Jordan [1978] and later plate motion models did not include a Bering plate. Lander et al. [1996] proposed a Beringia plate to explain the seismicity of the Koryak Highlands, Russia. Mackey et al. [1997] presented a more complete argument for a Bering plate based on the pattern and focal mechanisms of seismic events around the Bering Sea, clarified by a newly compiled Russian and U.S. seismological catalog. A broad area of deformation extends across the Seward Peninsula to the Chukotka Peninsula and from there southwest through the Koryak Highlands to Kamchatka, which may define the western boundary of the Bering plate. Focal mechanisms indicate normal faulting in the Seward Peninsula that transitions into right lateral strike-slip faulting to the west, while the Koryak Highlands are dominated by thrust events. The southern boundary was defined as the Aleutian subduction zone by Mackey et al. [1997]. However, the western Aleutian Islands migrate westward because of slip partitioning of oblique subduction [Geist et al., 1988; Ekstrom and Engdahl, 1989; Avé Lallemant and Oldow, 2000], so it is uncertain which islands may lie on the Bering plate. 1of19

2 Figure 1. Overview map of the study area, showing important geographic features and the locations of other figures in this paper. GPS survey locations are shown as follows: squares are sites from the Bering plate interior, inverted triangles are sites used in geodetic models for the Aleutians, hexagons are sites in the far western Aleutians, and all other sites are shown as blue diamonds. The white vectors show the velocity of the Pacific plate relative to stable North America. Blue vectors are velocities of sites that are considered to move as part of the Bering plate; gray vectors are sites that have a velocity significantly different from that of the Bering plate. The southern boundary of the Bering plate is the Aleutian subduction zone, shown as a solid black line. The eastern boundary of the Bering plate involves the Kaltag (K) and Denali (D) faults. [5] Geodetic measurements using GPS provide a way to quantitatively analyze the Bering plate hypothesis. With the exception of the Bering plate margins and a few remote islands, the Bering plate is subaqueous, so geodetic observations are concentrated on the edges of the plate, where plate interactions may result in strains in addition to the plate s rigid body motion. These strains must be removed from the GPS observations by modeling to properly determine the plate s velocity at a given location. This is especially true for the Aleutian arc, which has the highest number of observations anywhere on the proposed Bering plate (Figure 1). [6] Previous geodetic studies on the lower Alaska Peninsula, where Pacific North America convergence direction is almost entirely trench normal, recorded velocities of 4 mm/a (where a is years) to the southwest [Freymueller and Beavan, 1999; Mann and Freymueller, 2003; Fournier and Freymueller, 2007]. These studies found little indication of strain associated with subduction, and interpreted the velocities as a southwestward translation of those regions relative to stable North America. Cross and Freymueller [2007] used elastic dislocation modeling to remove interseismic strain from GPS measurements in the Andreanof region of the Aleutian arc, and determined the spatial pattern of coupling on the subduction interface while also finding that the Andreanof region is translating southwest relative to North America. [7] GPS measurements have also been recorded in western Alaska, and on the Bering Sea Islands of St. Paul (Pribilof Islands) and St. Lawrence (Figure 1). We quantitatively evaluate the Bering plate hypothesis using these data, combined with modeled arc translation rates from the Aleutians, and investigate the boundaries of the proposed Bering plate. In this paper we first derive estimates for the motion of the arc and fore arc at two locations along the Aleutian arc, and then test the Bering plate hypothesis and estimate the motion of the Bering plate relative to North America. 2. Dislocation Modeling Methodology [8] Strain accumulation at a subduction boundary is modeled using elastic dislocation theory following the approach of Savage [1983] and the computational methods of Okada [1992]. The earth is represented by a uniform elastic half space and the plate interface by one or more rectangular planar faults, and the strain accumulation rate is assumed to be constant through the interseismic period. The interseismic deformation rate is computed from the superposition of steady state subduction along the entire plate interface and steady normal slip (back slip) in the main thrust zone at the plate convergence rate, resulting in a plate interface that has a locked main thrust zone and is slipping freely above and below this zone [Savage, 1983]. Appropriate strike-slip and dip-slip components are determined by the convergence direction of the subducting plate and the strike of the trench. Because the Aleutian arc is moving relative to North America, the convergence velocity is the Pacific plate velocity minus the North American plate velocity minus the velocity of the Aleutian Arc relative to stable North America. 2of19

3 Figure 2. (a) Diagram of vector components. The arc translation velocity (arc velocity) may represent the motion of the Bering plate, but to determine this velocity we must remove the component of subduction strain from the GPS measurements. (b) Dc 2 contour plot versus arc velocity for Andreanof region. Dc 2 is contoured on the basis of minimum misfit models for a range of east and north arc translation velocities. Contour lines are equally spaced at intervals of 10 above the minimum c 2 value. The best arc translation velocity is 0.4 mm/a west and 7.9 mm/a south. The 95% confidence region is outlined by the thick blue line, and the red ellipse is the best fit ellipse to the 95% confidence region. [9] We extend the method of Savage [1983] by allowing the main thrust zone to be either fully locked or partially creeping. We define the slip deficit (rate) as the plate convergence rate minus the actual slip rate on the plate interface. There is no slip deficit if the slip on the interface occurs at the plate convergence rate, and if the plate interface is completely locked (no slip) the slip deficit would be equal to the convergence rate. We parameterize the slip deficit by a coupling coefficient, which is one minus the slip on the interface, divided by plate convergence rate. If there is no slip the coupling coefficient is one, and if the slip between the two plates is equal to the rate of plate convergence, the coupling coefficient will be zero. The coupling coefficient can represent the percentage of the total interface that is locked. [10] All measurements made at sites near a subduction zone are composed of two components, a component of interseismic strain caused by the subducting plate and coupling on the main thrust zone, and a component of rigid body translation that represents the long-term motions of the sites due to slip partitioning of oblique subduction or due to other motion of the overriding plate (Figure 2a). To isolate the arc translation velocity we must model the interseismic strain with an elastic dislocation model and remove this component from our measurements. [11] Estimation of the arc translation velocity and coupling coefficients on the fault planes is a nonlinear inversion, which we solve using a grid search inversion procedure. First we subtract from our data an arc translation velocity from a grid of candidate arc velocities. If we have selected the correct arc velocity, then the adjusted data will result entirely from interseismic strain and can be explained using the elastic dislocation model. If our candidate arc velocity is not correct, then the dislocation source will be unable to explain the adjusted data. We calculated the 3D surface displacements at each station assuming 100% coupling on all fault planes to obtain the Green s functions needed for the inversion. We invert for the coupling coefficients (m) using a MATLAB script "lsqlin," which solves for m using linear least squares and allows for inequality constraints on model parameters, and calculate misfit. We repeat this procedure for each candidate arc velocity and use the best chi-square values for each candidate to estimate the optimal model (Figure 2b). The fault plane coupling coefficients associated with the best arc translation velocity represent the best estimate of subduction interface behavior. [12] We estimated a 95% confidence region for the arc velocity on the basis of the increase in the chi-square above the best fitting model. We contoured the chi-square for the best fitting model for each candidate arc velocity (Figure 2), and approximated this region by an ellipse centered on the best estimate. This method implicitly accounts for tradeoffs between the arc velocity and the plate coupling parameters. We also tested the impact of varying the arc velocity on the plate coupling estimates, and the impact of the assumed model geometry on the arc velocity estimate. 3. GPS Data [13] Western Alaska is an area with few geodetic measurements prior to this study. We estimated site velocities for 7 sites on the Bering Sea coast and islands, plus 68 sites along the Alaska Peninsula and Aleutian arc, and 4 sites in eastern Russia (Figure 1 and Tables S1 and S2 in the auxiliary material). 1 Almost all of these sites are campaign (survey mode) sites. We incorporated the data of Avé Lallemant and Oldow [2000] from the central and western Aleutians, and added new surveys of their sites. Additional data in the Fairbanks area come from J. T. Freymueller et al. (Active deformation processes in Alaska, based on 15 years of GPS measurements, submitted to Geophysical Monograph Series, 2008), and these velocities are also given in Tables S1 and S2; because of the 2002 Denali fault earthquake, these data are preearthquake only. Most GPS sites in the Aleutian Islands have been surveyed as part of an effort to record volcanic deformation. In most cases there is not enough information to confidently remove the volcanic signal, so we omitted sites near active volcanoes. [14] We used the GIPSY/OASIS II software version GOA4 to obtain daily coordinate and covariance estimates of our stations and regionally distributed stations in the ITRF2000 reference frame, IGb00 realization (all references to ITRF2000 in this paper mean specifically the IGb00 realization of ITRF2000). See J. T. Freymueller et al. (Active deformation processes in Alaska, based on 15 years of GPS measurements, manuscript in preparation, 2008) for more details of the analysis strategy. We estimated site velocities in ITRF2000, and then converted to velocities relative to the North America plate by subtracting the motion of North America [Sella et al., 2007]. In our previous study we used the REVEL model of Sella et al. [2002], but that model is inconsistent with ITRF2000 because the origin (geocenter) rate in ITRF2000 and ITRF97 differ by 2 3 mm/a, and this difference in origin translation rate affects the plate angular velocities [Argus, 1 Auxiliary materials are available in the HTML. doi: / 2007JB of19

4 Table 1. Fault Plane Parameters Used for Dislocation Modeling a Name Longitude Latitude Length Width Dip Strike Depth Coupling 95% Range Andreanof Islands Region Adak upper % 0 100% Adak middle % % Adak lower % % Adak bottom % 61 2% Atka upper % 0 100% Atka middle % 0 80% Atka lower % 0 65% Atka bottom % 0 100% Fox Islands Region Fox % 0 85% Fox % 0 35% Fox % % Fox % 10 35% Fox % 0 10% Alaska Peninsula Region Alaska Peninsula % 63 94% Alaska Peninsula % 24 92% Alaska Peninsula % 9 57% Alaska Peninsula % 0 73% Near Islands Region Near Islands % 40 82% a The longitude and latitude describe the southeastern most corner of each plane. Depth is the vertical distance to the top of each plane. Length, width, and depth are in units of kilometers; strike and dip are in unit of degrees, and the strike direction is listed such that dip is down to the right. Coupling is the best value of coupling, and 95% range is the range of coupling at the 95% confidence limits and within the realistic physical values (0 100% except for Adak bottom). 2007; Kogan and Steblov, 2008]. This new realization of the North America reference frame changes the measured velocities for sites in this region by 2.5 mm/a compared to REVEL, which changes our estimate arc velocity from that of Cross and Freymueller [2007] by a similar amount. [15] Slow slip events and other transient slip events, with accompanying seismic tremor, have now been observed at most subduction zones where instruments are sufficient to detect them [e.g., Schwartz and Rokosky, 2007], but we do not yet know how frequently such events occur along the Alaska Peninsula and the Aleutian arc. The limited PBO and other continuous GPS data from the Alaska Peninsula and Aleutians do not show evidence for any significant transient slip events, although the records at most sites are still too short to draw firm conclusions. Because our data come from surveys carried out 2 10 years apart, we cannot tell whether any transient slip events have occurred along this section of the Aleutian arc during our study period. If there have been transient slip events, we would underestimate the width of the locked zones and/or the slip deficit, compared to what we would have estimated in a time period between transients. However, because the site measurement history for all sites in a given arc segment is nearly the same, transient slip events are unlikely to bias the estimate of the arc velocity. In this paper we focus primarily on the arc velocity and the first-order along-strike variations in the plate coupling, which should be unaffected by small transient events. 4. Model Results 4.1. Andreanof Islands [16] Using the new reference frame realization we reanalyze the velocities from the Andreanof Islands region, updating the results of Cross and Freymueller [2007]. We use the same eight fault plane model as Cross and Freymueller [2007], which is based on seismic studies by Engdahl et al. [1989], Ekstrom and Engdahl [1989], and Engdahl and Gubbins [1987]. The fault geometry parameters are listed in Table 1. We find the optimal boundary between the Adak and Atka "middle" and lower planes by testing models with the boundaries in different locations along strike. The boundary condition 0 m 1 is applied to all fault planes except the bottom plane on the Adak side where m is allowed to range between 1 and 1. This negative coupling coefficient condition allows for modeling afterslip and/or viscous relaxation that could be present below the main thrust zone because of the 1986 or 1996 earthquakes; we find that a small negative value fits better than requiring m 0 for the bottom planes [Cross and Freymueller, 2007]. [17] We find an arc translation velocity of 7.9 ± 2.4 mm/a nearly due south (Table 2, Figure 2b). The 95% confidence ellipse of arc translation velocity is elongated in the plate convergence direction. This orientation mainly results from the tradeoff between coupling on the fault planes and arc translation in the direction of plate convergence. Arc translation in this direction would be better constrained if we had more sites with a greater range of trench normal distances. Compared to the old arc translation velocity for the Andreanof region, the new arc velocity is different by a vector 3.9 mm/a toward 108. This change is slightly larger than the changes to the velocities caused by the reference frame (2.4 mm east). [18] The estimated arc velocity is relatively insensitive to reasonable variations in the geometric model for subduction. If we change the dip angles on the subducting planes by ±5, the arc velocity changes by less than 0.5 mm/a. Changing the width of the lower plane by ±10 km changes the arc velocity by 1 mm/a. These changes are small 4of19

5 Table 2. Calculated Arc Translation Velocities Relative to North America a Region East North East s North s Correlation Arc Parallel Arc Perpendicular Andreanof Alaska Peninsula a Given in mm/a. compared to the uncertainty estimate in the arc velocity, so we do not include them further. If we remove the bottom plane, which Cross and Freymueller [2007] introduced to account for possible postseismic deformation following the 1986 and 1996 earthquakes, the estimated arc velocity increases by abut 3 mm/a trenchward. This may be evidence for significant postseismic deformation in the near field from these recent M w 7.9 earthquakes. [19] Although there is a significant change in the arc velocity from the previous analysis by Cross and Freymueller [2007], due to the improved North American frame, the agreement between the areas of high coupling and the rupture zones of past major earthquakes is not changed. We find little to no coupling in Atka (eastern Andreanof Islands) and very strong coupling in the Adak region (western Andreanof Islands); this result is consistent with the moment release distribution and seismicity patterns of previous major earthquakes (Figure 3, [Ekstrom and Engdahl, 1989; Houston and Engdahl, 1989; Boyd and Nabelek, 1988]). We are able to resolve both the arc velocity and plate coupling variations because there is such a strong along-strike variation in the slip deficit along the plate interface. [20] We determined the 95% confidence limits for each coupling coefficient (Table 1) by fixing the other coupling coefficients at their best value and allowing the arc translation velocity to vary within its 95% confidence region. The coupling coefficients of the upper planes are the most poorly determined and have an uncertainty equal to their physical range of values. In other words, we cannot confidently say anything about the coupling on the upper planes except that coupling on these planes does not significantly affect the model results. In general, the coupling coefficients of the planes in the Adak region are better determined than the Atka region planes; this is due to a greater range of trench normal measurements in the Adak region Alaska Peninsula [21] The Alaska Peninsula region includes data from Chirikof Island (west of Kodiak) at 155 W to Sanak Island at 163 W, an along strike distance of approximately 500 km (Figure 4). This region of the Aleutian arc has the greatest trench normal range of data of any region and many of these sites have well determined velocities (Table S1). A geodetic study by Fletcher et al. [2001] used data from the Chirikof Island region to estimate an interseismic coupling value of Figure 3. Updated dislocation modeling results for Andreanof Islands region. Measured velocities are shown in blue, and modeled velocities are shown as white vectors. All velocities are relative to stable North America (Sella et al. s (2007) realization). Bold numbers denote the percent of unit coupling for the associated fault plane (see Table 1 for uncertainty range). The error ellipse for the arc translation velocity is based on the best fit ellipse to the 95% confidence region shown in Figure 2b. The component of the arc translation velocity in the direction of plate convergence is less certain because of tradeoffs between the arc velocity and plate coupling. Epicenters of major earthquakes are shown as stars, with the rupture areas shaded. The depth to the top of each set of fault planes in kilometers is shown at the west end; the top of the upper fault planes corresponds to the trench location. GPS sites are shown as white inverted triangles if they are used in this study, and other GPS sites are indicated by black dots. 5of19

6 Figure 4. Dislocation modeling results for the Alaska Peninsula. Color coding is the same as Figure 3 and all velocities are relative to North America. Measured velocities in the lower or western Alaska Peninsula are primarily explained by the arc translation velocity relative to North America shown by the red vector. The very low coupling coefficients for this region indicate the subduction zone interface is slipping freely. 80%. With an improved data set, Fournier and Freymueller [2007] used the method of simulated annealing to estimate the optimal fault model for this region. Fournier and Freymueller [2007] assumed a fixed arc translation velocity and found the coupling on the four fault planes to be 90, 70, 30, and 0% from east to west. This finding is in agreement with seismicity patterns. Specifically, the 1938 M w 8.3 ruptured the eastern two planes, whereas the western region is dominated by creep and has seen less moment release over the last 100 years. This western region is often referred to as the Shumagin seismic gap. [22] We use the same four fault planes but reestimate plate coupling and determine an arc translation velocity using the grid search inversion procedure described above with an improved North American reference frame. The estimated coupling coefficients and arc translation velocity are shown in Figure 4. We find an arc translation velocity of 4.0 ± 1.2 mm/a toward 211 (1.9 mm/a arc normal and 3.6 mm/a arc parallel) (Table 2). This is similar to the fixed arc velocity of 5.3 mm/a at 241 used by Fournier and Freymueller [2007] but shifted by the amount expected for the new North America reference frame [Sella et al., 2007]. Fournier and Freymueller [2007] discussed the constraints on the fault model, and evaluated tradeoffs between the geometric model parameters and coupling coefficients for this region. The dip angle here is shallow, and tightly constrained by seismic refraction lines. A 3 reduction in dip angle increases the misfit to the GPS data by an order of magnitude, and such an error can be ruled out. A 3 dip angle increase changes the arc velocity by less than 1 mm/a, although a 5 increase in dip could change the arc velocity by about 3 mm/a. However, such an error in dip angle is unlikely, as it would correspond to a nearly 50% error in the depth to the subducting plate. Changing the width of the fault planes by ±20 km changes the arc velocity by only 0.1 mm/a. Because the arc velocity is not sensitive to reasonable variations in the geometry, we do not include any additional uncertainty in the arc velocity estimate from this source. [23] The coupling coefficients illustrate that the subduction interface is nearly fully locked (90%, 63 95% at 95% confidence) at the Semidi Islands (eastern region), decreasing to about 30 ± 12% locked at the Shumagin Islands, and freely slipping to the west of the Shumagin Islands near Sanak Island. (Where the uncertainty range is strongly asymmetric about the best estimate, we give the full 95% confidence range, and where it is symmetric we give a standard 1s error bound.) Thus, the measurements recorded in the western area, where very little strain is observed, are almost entirely the result of the southwestward translation of the arc. 5. Possible Effects of Postseismic Deformation [24] Postseismic deformation from great subduction zone earthquakes can affect a large area for decades after the event [e.g., Freymueller et al., 2000; Cohen and Freymueller, 2004], but the record for smaller events is much less clear. The typical pattern of postseismic deformation following a great earthquake is illustrated by the model of Hu et al. 6of19

7 Table 3. Bering Plate Euler Pole Comparison Table a Pole (Data Used) Symbol Latitude Longitude w (degree/ma) sw Reduced c 2 x y z Pole Vector Bering plate Interior sites only Black square Bering sites with Andreanof Red circle arc velocity Bering sites w/alaska Peninsula Blue hexagon arc velocity Bering sites with both Aleutian Pink triangle Regions Full covariance matrix for the best pole (xyz) a Latitude and Longitude are in decimal degrees and w is in degrees per million years. The pole vector is also given in Cartesian coordinates in units of degree/ma. The best Euler pole is shown in bold. The full covariance matrix for the best pole is given in units of (degree/ma) 2. [2004] for the 1960 Chile earthquake. Within a zone extending landward from the downdip limit of the rupture, postseismic deformation results in trenchward velocities, which reach a peak in magnitude a few hundred km from the trench and decay with distance beyond that. To either side of the rupture, postseismic velocities are much smaller and are oriented toward the rupture zone. Suito et al. [2003] developed a detailed finite element model for postseismic deformation after the 1964 Alaska earthquake, which predicts no significant deformation at any of the Bering plate interior sites because they are far to the west of the area affected by the 1964 earthquake. The other great earthquakes along the Aleutian arc, the 1957 Central Aleutian earthquake and the 1965 Rat Islands earthquake, were much smaller than the 1964 Alaska earthquake (average slip was a factor of 2 3 smaller), but closer to our sites. [25] For multiple reasons, we argue that these earthquakes do not significantly affect our present velocities, despite the generally trenchward motion of the Bering Sea sites (Figures 1 and 6). The first reason involves the characteristic decay of postseismic deformation with distance. The site on St. Paul Island (500 km from the trench) has the same trenchward component of velocity as sites located km from the trench, but postseismic deformation would decay significantly between these two distances. For example, in the case of the 1964 earthquake, the present horizontal velocity peaks at mm/a at a distance of 350 km from the trench, but decays to 6 8 mm/a at 500 km from the trench and 2 3 mm/a at a distance of 800 km from the trench [Suito et al., 2003]. Furthermore, the arc velocities we estimated from sites close to the trench are also of the same order as the velocities of the sites in the Bering plate interior, which would not be the case for postseismic deformation. Postseismic deformation at sites near the arc should be several times faster than for remote sites, on the basis of the 1964 Alaska and 1960 Chile examples. Second, given the smaller slip of the 1957 and 1965 earthquakes, we expect that the postseismic deformation will be proportionately smaller. Finally, the orientation of the observed velocities is not actually trenchward, but is rotated clockwise from that direction and even more clockwise from the slip vectors of the 1957 and 1965 Aleutian earthquakes. [26] We calculated postseismic models using the spherical viscoelastic code VISCO1D, version 3 [Pollitz, 1997]. Even for an earthquake the size of the 2004 Sumatra-Andaman earthquake, velocities for sites >500 km from the trench are predicted to be <1.5 mm/a 40 years after the earthquake (assuming an upper mantle with a uniform viscosity of Pa s, similar to that estimated by Suito et al. [2003] for the 1964 earthquake). Average slip in the 1957 and 1965 events was about half as much as the 2004 Sumatra case, and we estimate postseismic velocities in the Bering Sea to be <1 mm/a. Postseismic deformation thus is likely to cause, at most, a small southward bias in the arc velocities, and it is probably insignificant at the Bering Sea Island sites. As noted in section 4.2, the Andreanof Island data supports the existence of a small postseismic transient following the 1986 Andreanof Islands and/or 1996 Delarof Islands earthquakes. 6. Western Alaska and Bering Sea Island Data and Angular Velocity of the Bering Plate [27] Sites in western Alaska and the Bering Sea Islands (Figure 1) show southward to south-southwestward velocities of 3 5 mm/a (Table S1). We first test whether data from these sites can be explained by rotation of a rigid Bering plate relative to North America. We select sites for this test only if they lie far enough away from the subduction zone to not record postseismic deformation from the 1964 Great Alaska Earthquake and are not affected by other known sources of interseismic strain. This eliminates sites in central Alaska or any sites within a few hundred km of the subduction zone (shown in gray on Figure 1), leaving only the sites on the west coast of Alaska and on the Bering Sea Islands. These sites are well fit by a single angular velocity (Table 3). We excluded the site BETC in Bethel, because an F test revealed that this site has a motion significantly different from others at the 90% confidence level. BETC has only two surveys, and is difficult to know whether this represents site instability, a bad survey, or real motion relative to the other sites. Removing BETC does not significantly affect the estimated angular velocity. We refer to the remaining set of 6 sites as the Bering plate interior sites. 7of19

8 Figure 5. Map showing Bering plate Euler pole locations. Locations of possible Euler poles for the Bering plate relative to the North America plate. Ellipses represent 95% confidence regions for pole location. Euler pole location depends on which subset of GPS sites are included in the inversion. A quantitative comparison of Euler poles is giving in Table 3. The preferred model is shown by the large pink triangle. [28] On the basis of only data from the Bering plate interior sites, we estimate that the Bering plate rotates clockwise relative to North America about a pole located in northern China (Figure 5, Table 3). If we use this angular velocity to predict the motion of the Bering plate along the Aleutian arc, the predicted motion is very similar to the arc velocities estimated in the previous section. There is no obvious active structure dividing the Bering plate interior from the Aleutians, so we propose that the Alaska Peninsula and eastern and central Aleutian Islands also lie on the Bering plate. To test this hypothesis we take various combinations of the data from the Aleutians, western Alaska, and the Bering Sea Islands, and for each we estimate the angular velocity of the Bering plate relative to North America and the misfit of the rigid plate model (Figure 5, Table 3). The translation velocities are based on many measurements, but we consider each as a single observation, with a weight based on the estimated confidence region. The estimated angular velocities are very similar to each other for all combinations of data (Figure 5). The best fitting Bering plate Euler pole uses all of the data (pink triangle on Figure 5), and has the Bering plate rotating clockwise relative to North America at 6.4 ± 3.3 /Ma about a pole located at 44.1 N, E. The predictions of this model are compared to the data in Figure 6. All predicted velocities fall within the 95% confidence regions of the measured velocities. [29] We use the F test to analyze the significance of the Bering plate motion. The F test compares the misfit of two models, where the first model represents a null hypothesis and the second model includes additional adjustable parameters. Because the misfit to the data is expected to decrease when there are more adjustable parameters, even if those parameters are just fitting noise, the calculated F ratio for the pair of models can be compared to tabulated values to assess the significance of the improvement in misfit. Our first null hypothesis is that all sites lie on the North American plate, and we compare this model to one in which all sites lie on a rigid Bering plate. The calculated F ratio far surpasses the minimum value at the 99% confidence level to reject this null hypothesis, so we conclude that the motion of these sites relative to North America is significant, and that a two-plate model (North America and Bering) fits much better than a single plate model. Note that data from sites on stable North America are not explicitly used here, but our velocities for sites in the stable, eastern part of North America are consistent with zero motion. [30] Next, we test models in which one or both Aleutian arc locations are assumed to move independently from the Bering plate. The reduction in misfit for both of these models is not significant compared to the model that assumes both locations lie on the Bering plate. We also test a two-plate model in which the two Aleutian arc locations are assumed to lie on a different plate than sites to the north, and an F test reveals that this model is not a significant improvement either. In this last model, the angular velocity of the Aleutian sites is poorly constrained, because there are barely enough data (4 observations) to estimate the 3 components of the angular velocity, and because the uncertainties in the arc velocities are large. [31] In the remaining sections of this paper, we use the estimate of the Bering plate angular velocity based on all data (best estimate in Table 3). However, the results for slip 8of19

9 Figure 6. Map of measured and modeled velocities for the Bering plate. Velocities for the Bering plate and surrounding areas are relative to North America. See Figure 1 for geographic names. Blue vectors are measured velocities on the Aleutian arc and the Alaska Peninsula. Pink vectors are velocities for sites located on the Bering plate interior that are used in the Euler pole inversion. Estimated arc translation velocities are shown as red vectors. Yellow vectors are the velocities predicted for the Bering plate using the best angular velocity. Long white vectors represent Pacific plate velocities. All other measured sites are shown as gray vectors. Note that not all sites in the Aleutians and central Alaska are shown to avoid clutter. Sites west of the Andreanof Islands (And) region show a clear westward acceleration of the arc, indicating that slip partitioning is an important mechanism contributing to the measured velocities of sites west of Amchitka Pass (180 W). Thrust events are clustered in the Koryak highlands, possibly the result of convergence between the Bering plate and eastern Russia (sites KMS and BILI). Approximate western and northern boundaries of the Bering plate are shown by a thick dashed line. partitioning are nearly identical if we use any of the angular velocities in Table Plate Coupling and Slip Partitioning Along the Rest of the Aleutian Arc [32] In this section, we investigate plate coupling along the remainder of the Aleutian arc. In the Fox Islands of the eastern Aleutians, we do not have enough data to estimate an arc translation velocity, because the sites are all located at about the same distance from the trench. But if we assume that this region lies on the Bering plate, like the segments to the east and west of it, we can make a first-order estimate of the distribution of slip deficit along this section of the arc. In the western Aleutians, slip partitioning of oblique subduction is likely to give rise to significant translation of the arc, and we test this by estimating the plate coupling and arc translation velocities for the western Aleutians and comparing these estimates to the predicted motion of the Bering plate. Avé Lallemant and Oldow [2000] found slip partitioning to be much more significant in the Near Islands group than in the Andreanof Islands, but did not develop quantitative models to separate the long-term motion of the arc from the elastic strain from subduction, as we have done. For the western Aleutians, we first consider the Near Islands, because we have enough data to optimize a simple fault model for the plate interface. We then apply this model to the Rat and Komandorsky segments on either side of it, where we have only a single data point from each, and test whether the arc velocities are significantly different. Finally, we compare the arc velocities measured by GPS to the slip vector orientations for underthrusting earthquakes to investigate the nature of slip partitioning Fox Islands Region [33] Data from the Fox Islands region include sites on Umnak, Unalaska, and Akutan Islands (Figure 7). On Umnak Island, the vast majority of GPS measurements have been established to record deformation associated with Okmok Caldera on the northeast end of the island, which is actively deforming [Miyagi et al., 2004; Lu et al., 2005], but the site ROWD on the western end of Umnak is unaffected by volcanic deformation. On Unalaska Island, all of the sites are located within 8 km of each other near the town of Dutch Harbor. On Akutan Island, all but four of the sites are affected by volcanic deformation associated with Akutan Volcano. Sites on Akutan show velocities very similar to Unalaska. [34] To construct the fault model in the Fox Islands region, we use the location of the trench, the volcanic axis, focal mechanisms from large thrust events, and the location of smaller earthquakes to constrain the fault geometry. We use a five fault plane model as shown in Figure 7. The fault geometry parameters are listed in Table 1. Because there are 9of19

10 Figure 7. Dislocation modeling results for the Fox Islands region. Color coding is the same as Figure 3, and all velocities are relative to North America. Note the higher coupling in plane 3 reflecting the more inboard (northward) directed measurements for sites ROWD and UNAL, but coupling is very weak in most of this area. along-strike differences in velocity of up to 7.4 mm/a, we assume coupling varies along strike and subdivide the fault plane model accordingly. We optimize the boundary between fault planes 2 and 3 by testing models with the boundary at different locations along strike and finding the model with the best fit to the data. Because there is a limited spatial distribution of data, we assume that planes 1, 4, and 5 have uniform coupling along strike. [35] The results of the Fox Islands region modeling are displayed in Figure 7 and Table 1. Coupling is zero in the shallowest plane and very low in the two deepest planes. Coupling is highest in plane 3 (47%, 25 95% at 95% confidence) and zero in plane 2, which allows for a better fit to sites ROWD and UNAL. Despite being in the rupture zone of the 1957 M w 8.6 earthquake, coupling throughout the Fox Islands region is low except for a region offshore of southwest Umnak. This agrees with the rupture model for the 1957 earthquake determined from tsunami waveform inversion, which found low slip in the eastern half of the rupture area, except for a concentration at southwest Umnak [Johnson et al., 1994]. [36] In the eastern Fox Islands region we find greater coupling on plane 4 at a depth range of 30 to 47 km than on plane 2 at a depth of 14 to 30 km. This pattern is in agreement with the rupture area for the 1957 earthquake based on aftershock distribution [Sykes, 1971], in which the rupture was inferred to have continued farther east at depth than in the shallower region. Despite their limited resolution, our results show a general correlation between regions of high slip during the 1957 earthquake and areas of higher coupling measured today Near Islands [37] The 7 sites in the Near Islands (Figure 8) provide sufficient data to develop a simple fault model and estimate the arc velocity. These sites have been surveyed over as much as a 10-year timespan, but a M w 7.7 earthquake occurred 100 km southwest of Amchitka Island in November of 2003 and caused large displacements on Amchitka and small displacements on Attu. We have considered only preearthquake data in this analysis. [38] The convergence direction between the Pacific plate and the North American plate in the Near Islands region has an obliquity of 80 ; the convergence direction is only 10 away from being arc parallel. Yet very large thrust events such as the 1965 (M w 8.7) Rat Islands earthquake ruptured both the Rat Islands and Near Islands regions, an along strike distance of 600 km, indicating that subduction is active throughout this region despite a trench normal convergence rate of only 14 mm/a. We construct a simple fault geometry with one fault plane (Table 1). The top of the fault plane is located at the trench, the dip is set to 14 and the width is optimized at 120 km. Using the same grid search inversion procedure as before, we solve for the coupling on the fault plane and an arc translation velocity of the Near Island region relative to North America using the seven sites from Attu and Shemya. [39] We find a coupling coefficient of 62 ± 10% and an arc translation velocity of 14.4 mm/a west and 2.3 mm/a north (Figure 8). By subtracting the arc translation velocity from the Pacific-North America convergence velocity, we find the actual trench normal convergence rate is 19 mm/a. This higher rate of trench normal convergence helps explain the existence of large thrust events in a region that has highly oblique subduction. Beck and Christensen [1991] used P waves to identify three regions of concentrated moment release for the 1965 Rat Island earthquake, and these regions were interpreted as asperities. One of these regions, 60 km wide, was located south of Agattu Island in 10 of 19

11 Figure 8. Fault plane geometry and dislocation modeling results for the Near Islands. Measured (blue vectors) and modeled (white vectors) velocities are relative to North America. The red vector is the translation velocity of the arc relative to North America calculated using the grid search inversion procedure. The yellow vector is the Bering plate motion relative to North America. Pink vector is the arc translation velocity minus the Bering plate velocity and represents the motion of the arc relative to the Bering plate. The dashed line indicates the approximate location of a seismically active right lateral strike-slip fault in the back arc, which is probably the main structure between the arc and the Bering plate interior. Inset shows residual velocities for sites on Attu and Shemya; note the velocity scale is larger. the Near Islands region and corresponds to the region of our data (Figure 8). [40] The Near Islands arc velocity shows that slip partitioning of oblique subduction is important in this region. It has been proposed, and our data support, that slip partitioning with right lateral strike-slip faulting in the back arc is a dominant factor in arc translation west of Amchitka Pass at 180 W [e.g., Geist et al., 1988; Ekstrom and Engdahl, 1989]. Thus, the translation velocity determined for the Near Islands represents both the motion of the Bering plate and the translation of the arc relative to the Bering plate. We subtract the predicted Bering plate velocity for the Near Islands from the arc translation velocity to obtain the velocity of the arc relative to the Bering plate, 15.4 mm/a toward 288. As expected, this velocity is nearly arc parallel (14.7 mm/a arc parallel, 3.2 mm/a arc normal), and it represents the slip rate on a strike-slip fault in the back arc. The active strike-slip faulting in the back arc is most likely located at the bathymetric break or lineation only 20 km northeast of Shemya, and may pose a significant hazard to the defense facilities on this island Rat Islands and Komandorsky Islands [41] In the Rat Islands, there is only one pre-2003 velocity (preearthquake). This velocity of site BKEB on Amchitka Island is nearly twice the velocity measured on Adak and Kanaga Islands 250 km to the east, indicating that slip partitioning has become an important mechanism here in the translation of the arc. We use the fault plane geometry for the Near Islands and adjust the strike for the Rat Islands region, and estimate an arc translation velocity of 16.1 mm/a toward 254, with the coupling coefficient for the single fault plane being 24%. Because the number of data and model parameters are equal, we cannot estimate a reliable uncertainty range from the misfit to the data. Instead, we test the hypothesis that the slip partitioning rate for the Near Islands can be applied to the Rat Islands region (i.e., there is no arc parallel extension between the two locations). We can still fit the data on Amchitka within its 95% confidence limits under this hypothesis, so a model in which the Near Islands and the Rat Islands translate as a single block is permitted by our data. This does not necessarily rule out arc parallel extension between the Rat Islands and the Near Islands, but such extension in not required by the limited data we have. [42] On Bering Island in the Komandorsky Islands, the site BKI has been operating almost continuously since 1997, and has a trench parallel velocity component of 49.5 mm/a, 19 mm/a faster than the site MURD in the Near Islands (Figure 9). We apply the same methodology used on the Rat Islands to the Komandorsky Islands, and find an arc translation velocity of 40 mm/a toward 311, with a coupling coefficient of 40%. The arc translation velocity is 80% of the observation at site BKI, so the other 20% is caused by interseismic strain. We again test whether the slip partitioning rate calculated for the Near Island region can be applied to the Komandorsky Islands, but in this case we find we can no longer fit the data within its 95% confidence regions. Thus there must be significant arc parallel extension between the Near Islands and the Komandorsky Islands. 11 of 19

12 Figure 9. Measured GPS velocities (blue vectors) in the western Aleutians and estimated arc velocities (red vectors). All velocities are relative to North America, including the velocity of the Pacific plate (white vector) and Bering plate (yellow vectors). Moment tensors of strike-slip events are from the Harvard CMT catalog. The cluster of events between the Rat Islands and the Near Islands may indicate a strike-slip fault that jumps from the fore arc to the back arc. The approximate locations of strike-slip faults are shown with a dashed line, and black vectors indicate the relative motion across the faults. Thrust faults are shown with teeth on the overriding plate; open teeth drawn for the far western Aleutians indicate a thrust fault where there is no trench normal convergence. [43] We use the arc parallel components of the arc translation velocities to calculate the strain rates between regions in the western Aleutians. Between the Komandorsky Islands and the Near Islands, the average arc parallel strain rate is a 1 (extension positive) and between the Rat Islands and the Andreanof Islands the strain rate is a 1. This strain must be accommodated either by normal faulting in the arc or by strike-slip faults that transition from the fore arc to the back arc, or by both mechanisms. Structural analysis of geological faults and folds supports the idea of longitudinal stretching of the arc in addition to strike-slip faulting on faults that roughly parallel the trench [Avé Lallemant and Oldow, 2000]. We have emphasized the role of strike-slip faults (Figure 9) here, but do not rule out significant longitudinal extension as an additional factor. [44] An investigation of strike-slip earthquakes along the arc west of the Andreanof region, recorded in the Harvard centroid moment tensor (CMT) catalog, reveals the predominance of strike-slip events located in the back arc (Figure 9). One cluster of strike-slip events located at 175 to 178 E shows an orientation rotated clockwise from the axis of the arc, which we believe may represent a fault or series of faults that cross from the fore arc to the back arc. The CMT catalog reveals no normal events at the appropriate depths and orientations to indicate extension in the arc by this mechanism Earthquake Slip Azimuths [45] An independent method for examining slip partitioning is to use slip azimuths from thrust earthquakes along the Aleutian arc. We use all Aleutian thrust events in the Harvard CMT catalog from 1976 to 2007 with the following constraints: depth<50 km, dip <30, strike , rake , and all events must locate along the arc. These thrust azimuths are then compared to the Pacific-North America [Sella et al., 2002] and the Pacific-Bering convergence directions (Figure 10). The earthquake slip vectors are systematically rotated from both convergence directions. [46] We use the same model of slip partitioning employed by Ekstrom and Engdahl [1989], in which the relative plate motion is accommodated by slip on the main thrust zone and strike-slip motions on a separate vertical fault plane. In the case of full partitioning, the strike-slip motion would be equal to the projection of the relative plate motion vector onto a line following the strike of the arc. In the case of partial partitioning, we can define V ARC BR ¼ k cosðf qþv PC BR ; ð1þ where V ARC-BR is the velocity of the arc relative to the Bering plate, V PC-BR is the Pacific-Bering plate convergence velocity with azimuth f, q is strike of the arc, and k is a proportionality constant that specifies what fraction of the arc parallel relative plate motion is partitioned on to the strike-slip faults. For no partitioning, k = 0, and for full partitioning, k = 1. The azimuth of convergence across the main thrust zone becomes y ¼ p ð1 kþcosðf qþ þ q tan 1 : ð2þ 2 sinðf qþ 12 of 19

13 Figure 10. Plots examining Aleutian arc parallel translation. (top) Thrust earthquake slip azimuths and plate convergence directions versus longitude for thrust events. The green line is the Pacific-North America convergence direction and the red line is the Pacific-Bering plate convergence direction. The thick black line is the convergence direction based on the model described in the text with k = (middle) Arc parallel translation of the arc relative to Bering plate based on slip azimuth, convergence direction, and arc orientation. Circles are the arc parallel translation velocities based on GPS measurements (relative to the Bering plate). (bottom) Slip azimuth minus Pacific-Bering plate convergence direction versus longitude. Average is shown for 5 bins of longitude with 1 standard deviation error bars. Note the 15 jump at Amchitka Pass and 9 jump at Amukta Pass. This should be the same as the slip azimuths for thrust events. A value of k = 0.55 ± 0.03 provides the best fit to the earthquake slip azimuths, essentially equivalent to the result of Ekstrom and Engdahl [1989]. [47] Using this value of k we can predict the velocity of the arc relative to the Bering plate, given the fraction of partitioning derived from earthquake slip vectors (Figure 10). This model predicts that the velocity of the arc relative to the Bering plate should be zero at 158 W, where Pacific- Bering plate convergence is normal to the trench. At 170 W, slip azimuths indicate that slip partitioning should be approximately 20 mm/a, yet our geodetic modeling of the Andreanof region data revealed trench parallel arc velocity of only 0.8 mm/a relative to the Bering plate, and the GPS velocities themselves are much smaller than 20 mm/a. The lack of evidence for partitioning in the Andreanof Islands was previously noted by Avé Lallemant and Oldow [2000]. In fact, nowhere do these two estimates of the slip partitioning rate agree (Figure 10 middle). [48] The large discrepancy between the geodetic observations and the slip vector azimuths can be explained if slip partitioning in the eastern Aleutians involves strike-slip faults entirely in the fore arc. Ryan and Scholl [1989] used seismic reflection data to identify major arc parallel shear zones in the fore arc, such as the Hawley ridge shear zone south of the Andreanof Islands. GPS measurements cannot record the rapid arc parallel motions in the fore arc because these faults are inaccessible beneath the ocean. If the faults in the fore arc are slipping freely, we would expect to record a step in the trench parallel strain across these faults. If these strike-slip faults in the fore arc are locked interseismically, they might still be difficult to detect using GPS. Darby and Beavan [2001] used a dense network of GPS measurements on the southern tip of North Island, New Zealand, to show that strain in the upper crust resulting from oblique subduction can be explained by an elastic deformation model with oblique slip on the plate interface even when there is geologic evidence for slip partitioning on strike-slip faults. In their model, the instantaneous interseismic GPS measurements recorded the same deformation in the case of no slip partitioning as in the case of locked strike-slip faults in the fore arc. [49] We have constructed a simple dislocation model to show how this is possible. The assumptions of the Savage model are that the interseismic deformation can be represented by the superposition of steady state subduction along the entire plate interface (forward slip), and steady normal slip (back slip) in the main thrust zone at the plate convergence rate, resulting in a plate interface with a locked main thrust zone and steady slip above and below this zone [Savage, 1983]. Savage argued that the forward slip produces no deformation and thus it is not necessary to construct this part of the model. In the case of slip partitioning in the fore arc, we can no longer make this assumption as the forward slip does produce deformation, and we have included forward slip in the model using the elastic subducting plate model proposed by Kanda and Simons [2006] (Figure 11). Our simple model follows the cartoon of Figure 11 and consists of a dipping slab with a thickness of 22 km, on the basis of the effective elastic thickness of 60 Ma oceanic crust [Watts et al., 1980], with a slip rate of 50 mm/a in both the trench normal and trench parallel directions. A vertical trench parallel strike-slip fault is located 50 km from the trench with a locking depth of 5 to 23 km. The slip rate on this fault is half of the trench parallel convergence (25 mm/a) (i.e., 50% partitioning in the fore arc or k = 0.5). We divide the thrust interface into two parts, the upper section where there is only half the trench parallel slip rate and a lower section where there is full plate convergence velocity. Furthermore, in the forward slip part of the model we include slip at the convergence velocity at the bottom of the elastic lithosphere. This lower dislocation approximates the motion of the downgoing plate in a frame in which the overriding plate is fixed. Within the region of potential observations (130 km or more from the trench) there is a constant 2 mm/a difference between this model and a simple oblique back slip model with no strike-slip 13 of 19

14 Figure 11. Schematic cross section of the elastic slip partitioning model. This model is used to examine the effect of a locked strike-slip fault in the fore arc. The model is slightly different from previous models in that we calculate the effects of forward slip of the subducting slab relative to the overriding plate. The standard back slip model is identical to the back slip component of this model but without the strike-slip fault. The red region in the elastic model represents the locked zone where no slip occurs in the interseismic periods. faults; such a small difference between modeling strategies is not likely to be distinguishable in practice, given the number of estimated parameters. [50] In the case of the Aleutians, we have evidence from thrust-earthquake slip azimuths that there is slip partitioning in the fore arc, but our data cannot resolve whether these faults are locked or creeping. We can say with certainty that east of Amchitka Pass, slip partitioning is not active in the back arc (Figure 12). This is different from previous studies such as Ekstrom and Engdahl [1989] that suggested slip partitioning in the back arc was a major tectonic component of the central and eastern Aleutians. [51] We have calculated an arc parallel translation rate of the fore arc on the basis of slip vector azimuths and an arc parallel translation of the main arc on the basis of GPS measurements (Figure 10 middle). The difference between Figure 12. Map of slip partitioning regions of the Aleutian arc. The red vectors are the arc parallel components of the arc translation velocities relative to the Bering plate. Red dots indicate regions in which there is no motion relative to the Bering plate. Pacific plate velocities are shown as the white vectors. The trench of the Aleutian subduction zone is shown as a thick black line. Strike-slip faults are indicated by thinner black lines and are dashed where they are inferred. 14 of 19

15 these velocities is the slip rate on the strike-slip faults inferred to exist in the fore arc. We can calculate approximate recurrence intervals by assuming that these faults are locked and rupture periodically in large earthquakes that produce 2 5 m of displacement (M w 7 7.5). In the central Aleutians, the slip rate would be approximately 20 mm/a, giving a recurrence interval of years, and in the eastern Aleutians, the slip rate would be 10 mm/a, giving a recurrence interval of years. Both of these intervals are longer than the complete record of seismicity in the Aleutians, so it is plausible that we simply have not recorded one of these large strike-slip events yet. These strike-slip faults also might rupture with an oblique component during great subduction earthquakes, thus acting as splay faults. [52] We subtract the earthquake slip azimuths from the Pacific-Bering convergence direction and group events into bins of 5 longitude and calculate the mean difference and standard deviation (Figure 10 bottom). This calculation reveals a jump in slip azimuth of 15.4 at approximately 180 W, corresponding to Amchitka Pass and Sunday Basin, where we believe slip partitioning becomes active in the back arc as well as the fore arc on the basis of GPS measurements. The next largest change in slip azimuth is 9.1 and occurs at approximately 170 W (Amukta Pass). This region corresponds to the Amlia and Amukta Basins, which like the Sunday Basin are interpreted to have evolved from the westward dismemberment of the arc via block translation and rotation [Geist et al., 1988]. [53] A model that may explain our observations would be one in which slip partitioning in the fore arc increases with the steady increase of obliquity of subduction, but slip partitioning also involving the back arc begins at discrete locations and thus results in step like changes in thrustearthquake slip azimuth when plotted against their along strike location (Figure 12). 8. Boundaries of the Bering Plate [54] With the recognition of a new plate comes that realization that some regions previously considered to be part of a stable plate interior are in fact active plate boundaries. The previous sections were devoted to discussing the tectonics of the Aleutians, the southern boundary of the Bering plate. In the following sections, we focus on the remaining boundary zones in western Alaska and eastern Russia Eastern Bering Plate Boundaries [55] Measurements from western Alaska show a southwestward velocity that is clearly distinct from the North American plate and represents coherent Bering plate motion. Where is the eastern boundary of the Bering plate? It must lie to the west of Fairbanks, because the Bering plate motion does not predict the observed GPS site velocities (Figure 13). However, active faults identified from seismicity patterns and focal mechanisms are aligned with the predicted Bering-North America motion direction. Multiple NNE trending seismic zones have been identified north of the Denali fault [Page et al., 1995], and have been responsible for several M w 6 and larger events (Figure 13). The westernmost and most prominent of these seismic lineations, the Minto Flats seismic zone (MFSZ), consistently produces shallow earthquakes with left lateral slip on near vertical, NNE striking fault planes, consistent with predicted Bering North America relative motion. In 1995, the north segment of the MFSZ ruptured in a M w 6.0 earthquake, and the best moment tensor for this event shows a strike of 208 and a dip of 74 E (Figure 13). West of the MFSZ, Ratchkovski and Hansen [2002] found SH min to be oriented 41, whereas in the MFSZ it is oriented at 83, and east of the MFSZ SH min is 105. Velocities predicted for the Bering plate relative to North America in interior Alaska are parallel to these seismic zones and to the SH min west of the MFSZ, and Bering-North America relative motion would produce left lateral faulting here. [56] We conclude on the basis of the data available that the Bering plate is moving to the south-southwest relative to the North American plate, and that these seismic zones represent the eastern limit of a diffuse North America- Bering plate boundary zone. There are no reliable GPS measurements in the 500 km region between here and the west coast of Alaska, so further details of the plate boundary zone will have to be determined by future studies. [57] The Denali fault is a major tectonic boundary that isolates the counterclockwise block rotation of south-central Alaska from the rest of the state. GPS measurements from south of the Denali fault can be fit by an Euler pole located at N, W with an angular velocity of 0.77 /Ma ([Fletcher, 2002], Figure 14). The Denali fault continues west of the 2002 rupture area for over 400 km (as the Farewell fault), but the Holocene slip rate of the Denali fault is thought to decrease west of 150 W, and there have been no major earthquakes recorded along this western section [Doser, 2004]. One possible explanation for this behavior is the similarity of the velocities predicted for the Bering plate and the south-central block in this region (Figure 14). If the Denali fault separates the Bering plate on the north from the south-central Alaska block on the south, the rate of strikeslip motion would be very slow. On the Denali Fault at 155 W and 62 N, the difference between the Bering plate prediction and the south-central Alaska prediction is 1.8 mm/a parallel to the Denali fault, with a right lateral sense of motion. East of this point, the models predict a component of compression across the Denali fault, and east of 152 W there is a well-developed foreland fold and thrust belt north of the Alaska Range and Denali fault, which may accommodate this fault normal convergence (Figure 14). [58] A reasonable analogy is interlocking gears, where two gears of different sizes rotate about different axes at different angular velocities but their interaction causes little strain because they are moving at the same rate along their mutual boundary. This type of tectonic interaction exists in other locations, for example, the Sierra Nevada block with the Oregon coast block [Wells et al., 1998; Williams et al., 2006], and the South China block with North China block [Heki et al., 1999] Western Bering Plate Boundary [59] The western boundary of the Bering plate is assumed to be located in the Koryak Highlands, Russia, on the basis of mapped faults and large earthquakes [Mackey et al., 1997, Fujita et al., 2002]. Using the Harvard CMT catalog, we have compared earthquake mechanisms to the few 15 of 19

16 Figure 13. Map of pre-denali earthquake seismicity and GPS velocities in Interior Alaska. Earthquakes (gray dots) show three clear NNE lineations that are almost parallel to the velocities of the Bering plate relative to North America, shown as white vectors. Thin black lines indicate mapped fault traces. Pre- Denali earthquake measured velocities are shown by the gray vectors. The Minto Flats Seismic Zone (MFSZ) is the most prominent zone of seismicity in the interior of Alaska [Page et al., 1995]. available GPS measurements from eastern Russia (Figure 6). Sites BILI and KMS show an east-southeast motion relative to North America, but the predicted motion of the Bering plate is south-southwest. If these two sites are representative of the motion of this region, we would expect contraction and northeast-southwest right lateral shearing between here and the Bering plate. A cluster of primarily thrust events is located near the top of the Kamchatka Peninsula in the southern Koryak highlands. The GPS site TIL has a velocity of 7.7 mm/a in a direction very similar to that predicted for the Bering plate. It is possible that the sites KMS and TIL are located on opposite sides of the plate boundary, although TIL moves faster than predicted for the Bering plate. The motion of TIL is nearly opposite in direction to the velocity of a campaign site on the neck of the Kamchatka Peninsula to the south [Bürgmann et al., 2005], so details of the plate boundary in this region will require further study Northern Bering Plate Boundary [60] The northern boundary of the Bering plate is an extensional setting trending east-west through the Seward Peninsula [Fujita et al., 2002]. Earthquakes here are primarily tensional and a breadth of geologic and geophysical data suggests the onset of continental rifting [e.g., Turner and Swanson, 1981; Dumitru et al., 1995; Page et al., 1991]. The original evidence for young to active extension on the Seward Peninsula came from geologic mapping and the recognition of young normal faults with 4 to 10 m of Holocene offset [Hudson and Plafker, 1978]. Turner and Swanson [1981] proposed that these faults are part of an incipient rift through the Seward Peninsula on the basis of their association with geothermal anomalies and young basalt flow deposits. [61] The boundary extends to the west and southwest through eastern Russia where faulting is primarily right lateral strike-slip eventually reaching the Koryak Highlands, which are dominated by thrust events [Mackey et al., 1997; Fujita et al., 2002]. To the east of the Seward Peninsula, there is less certainty in the Bering plate s boundary. We speculate that the plate boundary involves the Kaltag fault (Figure 13), as it is the most obvious feature that connects the Seward Peninsula to seismicity in Interior Alaska. The Kaltag fault has had at least two M w 5 earthquakes in the last 30 years; both of these events were right lateral with a minor component of extension. [62] There are no GPS measurements in the vicinity of the Kaltag fault and all measurements on the Seward Peninsula are south of the proposed boundary. The nearest reliable 16 of 19

17 Figure 14. Plate velocity comparison map for interior Alaska. Yellow vectors represent the predicted velocity of the Bering plate, and white vectors represent the south-central Alaska block. Blue vectors are measured velocities of sites on the Bering plate, and gray vectors are pre-denali earthquake velocities. The velocities of the Bering plate and the south-central block are similar in the region of the ellipse near the western segment of the Denali fault. Measured velocities for south-central Alaska are not shown because of complications involving interseismic strain and postseismic deformation from the 1964 Great Alaska Earthquake [Zweck et al., 2002]. measurement to the north is site SG27 in Barrow over 800 km away. SG27 has a well-determined velocity of 3.1 ± 0.3 mm/a at an azimuth of 155 (Figure 1). Although this site still has a significant southward velocity relative to North America, it is only half of the predicted Bering plate motion or the velocity of site MELS on the southern Seward Peninsula. As with the eastern boundary, the northern boundary may be a diffuse zone of deformation with localized regions of more intense deformation such as the Seward Peninsula. 9. Summary [63] Our GPS measurements support the existence of a rigid Bering plate that rotates clockwise relative to North America about a pole in eastern Asia. Given the boundaries of the plate, this pole indicates that most of the plate moves in a southward direction resulting in extension on the Seward Peninsula and a southward migrating Aleutian subduction zone. [64] Elastic dislocation modeling of GPS measurements along the Aleutian arc show an arc parallel increase in velocity of the arc west of Amchitka Pass, indicating slip partitioning has developed in the back arc. East of Amchitka Pass, the arc moves uniformly as part of the Bering plate. Discrepancies between thrust earthquake slip azimuths and plate convergence directions can be explained by active strike-slip faulting in the fore arc. Our interpretation of the GPS results is that the rapid westward translation of the western Aleutians is a result of strike-slip faults crossing from the fore arc to the back arc. Our dislocation models also show a good first-order correlation between areas of high- and low-slip deficit today, and areas of high and low slip in the 1957 Aleutian earthquake, extending the result of Cross and Freymueller [2007]. [65] The eastern boundary of the Bering plate lies between the west coast of Alaska and interior Alaska, with the most likely location being in interior Alaska on the basis of seismicity patterns and focal mechanisms. Southwest motion of the Bering plate relative to North America may be responsible for the left lateral faulting in the Minto Flats seismic zone that connects the Denali fault to the south with the Tintina fault to the north, as part of a diffuse Bering- North America plate boundary zone. Interaction between a clockwise rotating Bering plate and a counterclockwise rotating south-central Alaska block may be responsible for the reduced slip rate and lack of seismic activity on the western Denali fault, and for the development of a prominent foreland fold and thrust belt in the central Alaska Range. [66] The western boundary of the Bering plate probably lies in eastern Russia and involves thrusting in the Koryak Highlands and strike-slip faulting farther to the northeast that connects with the extension on the Seward Peninsula. GPS measurements from eastern Russia are sparse but support this hypothesis. [67] This study has identified a plate boundary zone in western Alaska, a region previously considered to lie within 17 of 19