4.1. Plate Rigidity and Error Model
 Rigid plate behavior on timescales of several million years is demonstrated by the success of rigid plate geologic models, and the exceptions are well described and spatially limited [Gordon, 1998]. Space geodesy can test the assumption of rigid plates on shorter timescales [Argus and Gordon, 1996; Dixon et al., 1996]. The existence of intraplate seismicity such as the New Madrid seismic zone [Nuttli, 1973; Schweig and Ellis, 1994; Weber et al., 1998; Newman et al., 1999] argues that some local intraplate deformation must occur. Plate rigidity over several years at the level of a few millimeters per year is suggested by the generally good agreement between geologic models and space geodetic estimates of plate motion [Smith et al., 1990]. On the other hand, the spatial sampling afforded by techniques such as VLBI and SLR is limited, and until recently, the precision of GPS, which in principle can provide better spatial sampling, has also been limited. Dixon et al.  noted that with the data then available, GPS data lacked the sensitivity to detect a clear postglacial rebound signal and provided only a crude upper limit to plate rigidity, as apparent deviations from a rigid plate model (e.g., the velocity residuals in Tables 1 and 2) mainly reflect GPS velocity uncertainties rather than true nonrigid plate processes. While the current data set is a significant improvement relative to the data available in 1996, and apparently has some sensitivity to postglacial rebound (see section 4.2), the same limitation still applies, although we can now place a tighter upper bound on plate rigidity from the GPS data. Useful constraints are available in the current data set for the North American, Eurasian, and possibly Australian plates. For the remaining plates, the number of data is too small to derive rigorous estimates of plate rigidity, although we note that the velocity data fit a rigid plate model at about the expected level for most of the plates studied (Figure 2b).
 Using all 84 available sites to define stable North America, comprising sites on the plate interior that have observations spanning as short as 2.0 years, the mean rate residual is 0.99 mm/yr. Below, we describe a 64-site solution that eliminates sites most sensitive to postglacial rebound, resulting in a mean rate residual of 0.86 mm/yr. In either case the residual magnitude is roughly the same as the GPS velocity error, suggesting that the rigid plate model is appropriate and that the residuals likely reflect GPS observation errors rather than nonrigid plate processes, as found in earlier studies [e.g., Dixon et al., 1996]. Since GPS velocity errors are a strong function of total observing time, one way to check this conclusion is to restrict the data set to sites with longer time spans and to investigate the effect on residuals. The residual magnitudes as a function of mean observing time are consistent with the error model illustrated in Figure 3. For example, if we base the angular velocity estimate for North America on the 36 sites in our database that are not sensitive to postglacial rebound and have time spans of 4 years or longer, the mean rate residual is 0.71 mm/yr, compared to 0.86 mm/yr for the larger data set, supporting the inference that the rate residuals reflect GPS velocity errors rather than nonrigid plate processes.
 Rate residuals are positively biased (Argus and Gordon  and Argus et al.  give a complete discussion) implying that the actual level of plate nonrigidity could be higher than the mean rate residual. If we arbitrarily assume that an upper bound on plate rigidity is 3 times the mean rate residual, then the GPS data set composed of 4 years or longer time series suggests that the stable interior of North America is rigid to better than 2.2 mm/yr. Presumably, this bound will be better constrained in future years as GPS velocity uncertainties decline further and new sites are added. Of the 19 plates and continental blocks examined here, all but three (Anatolia, Arabia, Nazca) have mean rate residuals of 1.8 mm/yr or less (Table 2). These three plates or blocks have a very limited distribution of sites, suggesting that their angular velocities are not well determined. The main point for this paper is that the velocity residuals, whether reflecting true nonrigid plate processes, local site effects, or (our preferred explanation) GPS velocity uncertainties, are small enough for the rigid plate approximation to be assumed valid for the data set under consideration.
 The relatively uniform nature of space geodetic data makes it feasible to apply a consistent error model, facilitating statistical tests of plate rigidity. If our error model for GPS velocities is approximately correct, χ2 per degree of freedom (hereafter χv2) for the individual plates should approximately equal 1.0, provided that the data fit the rigid plate model, and that the data set is large enough to be statistically representative. The latter criterion is met for the North American plate (128 data; χv2 = 1.05) and the Eurasian plate (30 data; χv2 = 1.02). However, the majority of plates tested also appear to approximately satisfy this “rule of thumb,” despite the relatively sparse data (Figure 2b). All but three plates or blocks (Amuria, Australia, Okhotsk) have χv2 < 2.0, and two of these (Amuria and Okhotsk) are probably related to limited data (Table 2). Since the error model is derived independently of any rigid plate criteria, this tends to confirm the joint hypothesis that the great majority of plates and blocks tested are rigid within the velocity uncertainty, that the sites used to define the various plates lie on the rigid portion of the plate interior, and that the error model is approximately correct. The fact that χv2 for North America and Eurasia is slightly larger than 1.00 might reflect the fact that we have neglected random walk noise. However, the effect is small and is ignored here.
 For several plates the number of space geodetic data is small enough that the solutions are sensitive to outliers at one or two sites, the statistical effect of which is diminished when a large number of sites is available. For example, Amuria and Okhotsk, with only three sites and five sites, respectively, have χv2 = 2.3 and 4.4, suggesting that these data are not well fit by the rigid plate model. Possible explanations include that (1) the number of data on the plate are too small to be statistically representative; (2) one or more site velocities have a systematic error not reflected in the error model; (3) our choice of sites is inappropriate, and one or more sites lie in a deforming boundary zone rather than the rigid plate or block interior; and (4) the plate is not rigid. We suspect that explanations 1, 2, and 3 together explain such misfits, reflecting limitations in available data.
4.2. Glacial Isostatic Adjustment
 The lithosphere in North America, Eurasia, and Antarctica is isostatically adjusting from mass loading and unloading of the last glaciation, reflecting the delayed response associated with viscous flow in the Earth's mantle. This adjustment imparts both vertical and horizontal motions to the surface velocity field [e.g., Peltier, 1998a] and needs to be considered when using geodetic data to derive plate motion models because the velocity effects are not representative of longer-term motions. That part of the surface velocity field due to glacial isostatic adjustment (GIA) is a function of both ice loading history and mantle viscosity structure, neither of which is well known, so the corresponding model predictions can vary significantly [e.g., Argus et al., 1999; Mitrovica et al., 2000]. However, most models agree on the region of maximum glacial isosatic effect, even if they do not agree on the magnitude or direction of predicted motion. Rather than correcting each site velocity by a specific model prediction, we take a more conservative approach and simply eliminate the subset of sites most likely to be affected by GIA. For North America, maximum GIA-related motion is centered in a broad region around Hudson Bay. We use the ICE-4G model [Peltier, 1994] to define sites with horizontal rates due to GIA ≥ 0.7 mm/yr and compare solutions with and without this data subset. Thus we compare a solution for North America using all available data (84 sites, χv2 = 1.63) versus a model that omits 20 sites most affected by GIA (algo, chb1, chur, det1, dubo, flin, kew1, mil1, neb3, nrc1, oro_, sag1, sch2, stb1, stp1, vcap, whp1, wis1, yell, you1) giving a 64-site model with χv2 = 1.05. Of these 20 sites, 9 exhibit statistically significant uplift (greater than one standard error), and 5 exhibit uplift rates greater than 3.0 mm/yr (Table 1). Sites may also subside due to GIA, associated with collapse of the peripheral bulge, but associated horizontal velocities are thought to be small compared to our observational error.
 For Eurasia we compared the solution using all available sites (19 sites, χv2 = 3.73) versus a solution that omits four sites on the Fennoscandian platform (kiru, mets, onsa, trom) most likely to be affected by GIA. The omitted sites were chosen using the same criteria for horizontal motions as for North America [Peltier, 1998b], approximately equivalent to eliminating sites with vertical motion >3 mm/yr, as predicted by the uplift model of Davis et al. . In this case the fit is significantly improved (15 sites, χv2 = 1.02). We conclude that the current GPS velocity field is sensitive to GIA and that correcting for this effect or eliminating sites sensitive to it is important for obtaining accurate plate angular velocity estimates for comparison to geologic values. For Antarctica we cannot do this as all sites are experiencing GIA uplift (see Table 1).
4.3. Comparison to Geologic and Geodetic Data
 In this section we discuss the criteria for inclusion or omission of key sites by plate (listed alphabetically, except South America, included in the Nazca plate discussion), compare our results to independent data and the NUVEL-1A geologic model [DeMets et al., 1994], and discuss some implications. The discussion below focuses on those plates or blocks where limited available data make the results sensitive to the selection criteria, where differences appear to exist with NUVEL-1A, and on the relative velocities of several of the “new” plates or continental blocks that have not been included in previous global models. We omit discussion of plates or blocks where our data are limited or where our results do not differ significantly from previous publications: Anatolia [Reilinger et al., 1997; McClusky et al., 2000], Okhotsk [Seno et al., 1996; Wei and Seno, 1998], Sierra Nevada [Dixon et al., 2000a], and Sunda [Walpersdorf et al., 1998; Rangin et al., 1999; Chamot-Rooke and Le Pichon, 1999].
 Angular velocities for the Amurian plate have been reported by Zonenshain and Savostin , Wei and Seno , Heki et al. , and Holt et al. . Following Heki et al. , we use three sites to define this plate, DAE*, SUWN, and VLAD (Figure 5). The limited geographic distribution results in correspondingly large uncertainties. Site bjfs is not included in our rigid plate definition since it is located in an area of relatively high seismicity and active faulting [Shen et al., 2000] (the 4-site solution including bjfs, 62.64°N, −128.76°E, 0.319°/Myr, σmaj = 17.1, σmin = 1.3, σω = 0.036°/Myr, mean rate residual (MMR) = 1.2 mm/yr, is similar to the 3-site solution listed in Table 2 and improves χv2 from 2.30 to 1.46).
 The formation of the Baikal rift reflects relative motion between Amuria and Eurasia [Zonenshain and Savostin, 1981]. Our relative angular velocity predicts south-southeast extension across the rift at 6–7 mm/yr. At 47.5°N, 106.5°E, close to where Calais and Amarjargal  measure a velocity of 6.4 ± 1.6 mm/yr at an azimuth of 125 ± 30° from continuous GPS data, we predict (Amuria relative to Eurasia) 7.0 ± 3.2 mm/yr, 165 ± 19°.
 We define the Antarctic plate using seven sites (Table 1), excluding ohig and palm because of their proximity to the complex plate boundary with the Scotia plate [Pelayo and Wiens, 1989; Klepeis and Lawver, 1996]. The Antarctic-Australia and Antarctic-Pacific boundaries are spreading ridges, and their velocities are well determined in geologic models such as NUVEL-1A. Pacific and Australia velocities are also well determined in REVEL-2000 (Tables 1, 2, and 3). Any difference in velocity between the two models probably indicates real velocity changes. Figures 6 and 7 indicate that the rates and azimuths predicted by the two models for both plate pairs are virtually identical along most of the respective plate boundaries. For Australia-Antarctica, REVEL-2000 and NUVEL-1A agree to better than 0.6 mm/yr in rate for >80% of the entire plate boundary, implying a remarkable steadiness of plate motion for this plate pair over the last 3 Myr.
 We define the Arabian plate using a combination of two GPS sites (BAHR, KATZ), and one SLR site (7832) (Figure 5). KATZ is within 30 km of the active Dead Sea fault and thus does not meet our criterion for minimum distance from the plate boundary. However, the amount of strain accumulation is believed to be small here [Pe'eri et al., 2002]. The north striking Dead Sea fault is a left-lateral strike-slip fault separating Arabia and the Sinai block, a small continental block that may move relative to Nubia. Young (<140 kyr) offset geomorphic features along the Dead Sea fault at 30.8°N, 35.4°E indicate a slip rate of 4 ± 2 mm/yr at a strike of 028° [Klinger et al., 2000]. Our predicted Arabia-Nubia velocity at this location is very similar, 3.9 ± 0.8 mm/yr at an azimuth of 009 ± 10°, implying very slow motion of Sinai relative to Nubia.
 If motion of Arabia relative to Nubia has been steady over the last few million years, we expect agreement between the REVEL-2000 prediction and the rate of seafloor spreading across the Red Sea. However, measured spreading rates averaged over the last 3 Myr [Chu and Gordon, 1998], as well as the model rates of Jestin et al.  based on a similar time period, are both systematically higher than REVEL-2000 (8 and Table 5). Similarly, NUVEL-1A's predicted Arabia-Eurasia rate is higher than REVEL-2000 (Figure 9). McClusky et al.  obtained a similar result. These rate differences may reflect gradual slowing of the Arabian plate as it moves north and collides with Eurasia. Associated crustal thickening forms the Zagros [Alavi, 1994] and Caucasus Mountains, increasing gravitational body forces that oppose Arabia's northward motion, perhaps resulting in gradual slowing of Arabian-Eurasia convergence and Red Sea spreading.
 We define Australian plate motion using 11 sites on Australia and Tasmania (Table 1 and Figure 10). Sites noum and auck have velocities relative to stable Australia that are larger than the corresponding velocities of stable interior sites, perhaps reflecting strain accumulation associated with nearby subduction zones, suggesting that they should not be used for the Australian plate definition. The parameter χv2 is 3.18 for a 13-site solution that includes noum and auck versus 2.78 for the 11-site solution; χv2 is 3.31 for a 12-site solution that includes auck. However, even omitting noum and auck results in misfits and a χv2 that is higher than expected. The velocity residual pattern is very similar to the direction of maximum horizontal compression in the region. Recent compilations of stress data [Hillis and Reynolds, 2000] suggest east-west compression in western Australia, rotating to northwest-southeast compression in southeastern Australia and northeast-southwest compression in northeast Australia (e.g., in vicinity of TOW2) (Figure 10). This may explain why χv2 for this plate is high. In other words, the rigid plate assumption may not be strictly valid for the continental portion of the plate, and Australia may be undergoing intraplate deformation in response to these compressional stresses. Alternatively, we may have underestimated the site velocity errors for Australia. We prefer the former explanation because the error model gives consistent results for most other plates (Figure 2b).
 Separate Indian and Australian plates have been recognized for some time [Wiens et al., 1985; DeMets et al., 1988], and a Capricorn plate has also been proposed in the central western Indian Ocean with a broad, diffuse boundary [Royer and Gordon, 1997; Gordon et al., 1998]. We are unable to calculate relative angular velocities for plate pairs involving the Capricorn plate because of insufficient geodetic data. However, we can use the χ2 test to evaluate the fit of the velocity data for coco (Cocos Island, on the southern edge of India-Australia boundary), dgar (Diego Garcia, near the inferred boundary between the India and Capricorn plates), HYDE (Hyderbad, India), IISC (Bangalore, India), and MALD (Male, Maldives) to our rigid Australian plate model, with 11 sites and χv2 = 2.78 (Figure 5). The χv2 for the 11 + 5 sites is 15.25, i.e., significantly higher, indicating a very poor fit. If we exclude HYDE, IISC, and MALD (11 + 2 sites), χv2 drops to 2.74, consistent with separate Indian and Australian plates (e.g., using the F ratio test of Stein and Gordon ). However, when we exclude either dgar or coco in an 11 + 1 site solution for Australia, we get little or no improvement (excluding dgar gives χv2 = 2.72, excluding coco gives χv2 = 2.86). The small change from excluding coco or dgar reflects the fact that these two sites have velocities that are negligibly different from stable Australia, consistent with very slow relative motion between Capricorn and Australia [Royer and Gordon, 1997; Conder and Forsyth, 2001].
 The Caribbean plate has been the focus of kinematic studies for at least 25 years [Jordan, 1975; Stein et al., 1988; Deng and Sykes, 1995; DeMets et al., 2000]. Geologic estimates of the velocity of the Caribbean plate with respect to its neighbors are hampered by the paucity of relevant data (especially rate data from spreading centers) and the tectonic complexity of the boundary zones around the plate (see discussions by DeMets et al.  and Dixon et al. [1991b]). The velocity data set used here is similar to that presented by Weber et al. , adding only additional position data at the continuous sites BARB and CRO1. Also, we take a slightly more conservative approach, omitting two sites in Puerto Rico (isab, pur3) because of the possibility of independent motion of the Puerto Rico block [Jansma et al., 2000]. The resulting Caribbean–South America angular velocity is nevertheless essentially identical to that presented by Weber et al.  (Table 5). The Caribbean-North America angular velocity is similar to that presented by DeMets et al. , although the current vector is constrained by additional position data at SANA, ROJO, and CRO1, which improves their velocity estimate, as well as by velocities from new sites at BARB and TDAD. As pointed out by Dixon et al. , Pollitz and Dixon , DeMets et al. , Weber et al. , and Perez et al. , the motion of the Caribbean plate with respect to both North and South America is considerably faster than predicted by NUVEL-1A, probably reflecting systematic errors in the geologic model (Table 5).
 With the benefit of hindsight, it is clear why NUVEL-1A systematically underestimates the speed of the Caribbean plate (the predicted azimuths for NUVEL-1A and REVEL-2000 are very similar along much of the Caribbean boundary). Sykes et al. , Rosencrantz and Mann , and Mann et al.  discuss the importance of the Gonave microplate, separating the North American and Caribbean plates. This microplate is defined by the Cayman spreading center on the west [Macdonald and Holcombe, 1978; Rosencrantz et al., 1988], the Oriente-Septentrional fault zone on the north, and the Enriquillo-Plantain Garden fault zone on the south [Leroy et al., 2000; Pubellier et al., 2000] (Figure 11). Dixon et al.  estimated 8 ± 4 mm/yr of motion along the Enriquillo fault zone in the Dominican Republic, a significant fraction of overall North America–Caribbean motion. NUVEL-1A underestimates Caribbean–North America (and, by implication, Caribbean–South America) motion by roughly the slip rate on this fault because this rate represents that portion of plate motion not accommodated on the Cayman spreading center, NUVEL-1A's only Caribbean rate datum. REVEL-2000 provides a more accurate estimate of Caribbean–North America and Caribbean–South America motion, not only for the decade timescale but perhaps for the last few million years as well, because it represents the total relative plate motion, not just motion accommodated at the spreading center.
 We define the Eurasian plate using 15 sites (Figure 5 and Table 1) and obtain a well-defined solution (χv2 = 1.02, mean rate residual of 0.71 mm/yr). We exclude all sites south of significant seismicity reflecting the Nubia–Eurasia collision (e.g., south of the Pyrenees we exclude casc, ebre, and vill [Ribeiro et al., 1996; Herraiz et al., 2000]), all sites in or south of the Alps and south of the Carpathians (geno, noto, penc, sofi) [Calais, 1999; Grenerczy et al., 2000], all sites in central Asia south of 45°N because they may be affected by the India–Eurasia collision (kit3, pol2, sumk, urum), and all sites east of 130°W that may be on the North American plate or in the deforming zone between North America and Eurasia (bili). We also exclude sites in or near the Rhine Graben and Roer Graben (dour, kosg, wsrt) and sites west of the Rhine Graben and south of the Roer Graben (brst, brus, hers, mans, mlvl, sjdv, toul) similar to Nocquet et al. . These grabens were active in late Pliocene time and are also the current locus of significant seismicity [Plenefisch and Bonjer, 1997]. In addition, irkt is excluded because it is <100 km from the active Baikal Rift [Doser, 1991; Delvaux et al., 1997] (see section 4.3.1), and yakz is excluded because of problems with its time series that we do not understand (our time series for this site shows a large seasonal variation in all three components). We also omitted kiru, mets, onsa, and trom due to sensitivity to GIA, as discussed earlier. Site brst shows a large residual motion with respect to Eurasia, possibly reflecting neotectonic effects in the area [van Vliet-Lanoe et al., 1997]. A Eurasia solution that excludes TIXI, near the possible plate boundary with North America, results in an angular velocity that is negligibly different (14-site solution 58.37°N, −102.06°E, 0.258°/Myr, σmaj = 1.7, σmin = 0.4, σω = 0.004°/Myr, χv2 = 1.10, MRR = 0.7 mm/yr).
 Inclusion of the available six sites in stable western Europe (west of the Rhine Graben) in the Eurasia plate solution increases misfit (χv2 = 1.70 for 21 sites compared to our preferred 15-site solution with χv2 = 1.02), but the difference is small enough that separate western Europe and Eurasian plates are not supported by an F ratio test [Stein and Gordon, 1984] (angular velocity for the six sites west of the Rhine Graben with respect to ITRF-97 is 45.64°N, −117.08°E, 0.215°/Myr, σmaj = 18.1, σmin = 0.9, σω = 0.016°/Myr, χv2 = 1.46, MRR = 0.8 mm/yr). The two plates may exist, but our data are insufficient to reliably resolve their relative motion. Given the observed seismicity and surface faulting in the Rhine Graben [Camelbeeck and Meghraoui, 1998; Meghraoui et al., 2000], we have taken the more conservative approach and used only sites east of the rift in our definition of Eurasia.
 We can nevertheless calculate relative motion across the Rhine Graben, using the admittedly noisy relative angular velocity vector between Asia and western Europe as defined above (58.48°N, 2.35°E, 0.076°/Myr, σmaj = 9.4, σmin = 2.7, σω = 0.005°/Myr). Calculated at a point on the upper Rhine Graben (49°N, 8°E) we predict west-southwest/east-northeast extension, 1.5 ± 0.4 mm/yr at 73 ± 17°. Note that this calculation does not require that relative motion between western Europe and Eurasia is accommodated exclusively in the Rhine Graben. Our calculated direction agrees with the orientation of minimum principal stress from earthquake focal mechanisms in the region [Plenefisch and Bonjer, 1997] and is consistent with Nocquet et al. . Westward motion of western Europe may reflect extrusion tectonics associated with the northward motion of Italy and the Nubian plate relative to Eurasia.
 Following Kogan et al. , we evaluate if site bili in northeastern Russia lies on the North American or Eurasian plate by comparing the size of the corresponding velocity residual. This site has a residual rate of 6.4 ± 2.0 mm/yr with respect to Eurasia and 3.4 ± 1.6 mm/yr with respect to North America, suggesting that bili lies either on the North American plate or on the diffuse boundary between the two plates [Chapman and Solomon, 1976; Zonenshain and Savostin, 1981; Cook et al., 1986].
 We can also compare the predictions of our angular velocity estimate for Eurasia-North America with measured spreading rates and transform fault azimuths along the Mid-Atlantic Ridge and with the NUVEL-1A model. Figure 12 shows this comparison and also the predictions of Kogan et al. . REVEL-2000 shows good agreement with the geologic data along the Mid-Atlantic Ridge north of about 65°N but is systematically faster than a cluster of geologic data around 40°–45°N by ∼2 mm/yr. At 64°N, 20.5°W on the plate boundary in Iceland we predict spreading at a rate of 19.9 ± 0.2 mm/yr at an azimuth of 102 ± 1°, which compares well with the local measured velocity using episodic GPS data, 21 ± 4 mm/yr at azimuth of 117 ± 11° [Sigmundsson et al., 1995].
 We can determine what fraction of the total plate motion is accommodated within Iceland by comparing the velocity of hofn in the southeastern corner of the island relative to stable North America (21.2 ± 0.8 mm/yr at 104.1 ± 2.0°) with the relative plate velocity computed at the same location (19.7 ± 0.2 mm/yr at 104.2 ± 0.9°) (hofn is not used in the definition of Eurasia, so this is an independent test). Thus hofn's velocity is consistent with stable Eurasia's velocity within 95% confidence. Similarly reyk in southwestern Iceland is not used in the definition of stable North America and moves at 20.2 ± 0.6 mm/yr at 281.4 ± 1.8° relative to stable Eurasia, compared to the calculated relative plate velocity here of 19.9 ± 0.2 mm/yr at 280.8 ± 0.9°, again equivalent within uncertainties. The baseline between hofn and reyk is more precise than the site velocities described above and is independent of any definition of stable North America or Eurasia. Its length rate of change, 20.3 ± 0.2 mm/yr, is very similar to the predicted rate of change from REVEL-2000, 19.8 ± 0.2 mm/yr, calculated at the point where the baseline between these two sites crosses the plate boundary (64.2°N, 18.8°W). These results confirm that essentially all of the plate motion between North America and Eurasia is accommodated within the island, consistent with earlier studies [Sigmundsson et al., 1995; Jonsson et al., 1997; Hreinsdottir et al., 2001].
 We define the Indian plate using three sites, HYDE, IISC, and MALD. Our mean rate residual (1 mm/yr) is compatible with earlier rigidity studies of Paul et al.  and Malaimani et al. . We agree with earlier findings suggesting that the Indian plate is moving slower than predicted by NUVEL-1A [Chen et al., 2000; Shen et al., 2000; Holt et al., 2000; Paul et al., 2001; Kreemer et al., 2000]. These studies show a range of velocity estimates (e.g., rates at IISC relative to Eurasia of 34.8, 41.9, 36, 43.7, and 34.3 mm/yr, respectively), but all are slower than the corresponding NUVEL-1A estimate (47.8 mm/yr). Our Eurasia-India angular velocity predicts a rate of 35.2 mm/yr at this location. It is unclear whether the difference between the geodetic and geologic rates reflects a true deceleration in the relative velocity of this plate pair or a systematic error in NUVEL-1A [Gordon et al., 1999].
4.3.8. Nazca and South America
 Owing to the paucity of GPS data on the Nazca plate (EISL, GALA) we include all available space geodetic data, adding SLR (7097) and DORIS (EASA, GALD) (Figure 13). Previous estimates for South America were also hampered by relatively sparse data. For example, Norabuena et al.  had only four continuous GPS sites with a total of ∼3000 station days of data to define stable South America, while Norabuena et al.  had 6200 station days of data, also at four sites. For this study we have 11 GPS sites with 14,200 station days of data; hence South America's velocity is much better defined. Our resulting Nazca-South America rate is nevertheless consistent with previous results [Norabuena et al., 1998, 1999; Angermann et al., 1999] and significantly slower than the NUVEL-1A estimate. Nazca-Antarctica and Nazca-Pacific are also significantly slower than the geologic model (Figures 14 and 15). As has been shown in previous studies, both Nazca-Pacific and Nazca-South America have been decelerating during the last 25 Myr [Tebbens and Cande, 1997; Somoza, 1998]. A similar deceleration applies to Nazca-Antarctica (Figure 16). These decelerations are sufficiently rapid that they can be observed as differences between geodetic plate motion estimates and NUVEL-1A predictions.
 The REVEL-2000 velocity azimuths for Nazca-Antarctica and Nazca-Pacific also differ from NUVEL-1A (Figure 16). Although the Nazca site distribution is limited, the difference may be real, perhaps reflecting changing plate direction over time. Stage pole reconstructions back to 30 Ma [Tebbens and Cande, 1997] allow us to look at a much longer time record and suggest changes in Nazca-Antarctica and Nazca-Pacific direction in the last 15 Myr in the same sense as we infer from the shorter epochs “sampled” by NUVEL-1A and REVEL-2000 (compare Figures 14 and 16).
4.3.9. North America
 The rigid North American plate is defined using 64 sites (Figure 13 and Table 1). We exclude all sites located in the Gulf Coast because of possible subsidence and all sites west of the Rio Grande Rift and west of the Rocky Mountain front because of possible tectonic effects. Sites mem2 in Memphis, Tennessee, and cha1 in Charleston, South Carolina, are also omitted as they have both been the focus of major intraplate earthquakes in the last 200 years [Nuttli, 1973; Bollinger, 1972]. As discussed earlier, we omit 20 sites likely to be most affected by glacial isostatic adjustment in the northern part of the plate. A more detailed discussion of North America is in preparation.
 Fairbanks, Alaska (fair), has been used in our previous definitions of stable North America [e.g., Dixon et al., 1996]. It is excluded here because it is too close to the seismically active plate boundary zone and moves 2.2 ± 0.5 mm/yr at 171 ± 11.4° relative to stable North America (Table 1), somewhat faster than would be expected for a “stable plate” interior site.
 We define an angular velocity for Nubia using five sites (GOUG, HAR*, HRAO, MAS*, SUTH). HRAO and HAR* in South Africa are located near an area of intense microseismicity (Figure 17). A three-site solution that omits these two sites has an angular velocity of 52.65°N, −82.79°E, 0.249°/Myr, σmax = 1.9, σmin = 0.6, σω = 0.004, χv2 = 0.27, MRR = 0.79 mm/yr, very similar to our preferred five-site solution (Table 2), and suggests that sites HRAO and HAR* can be included in the stable plate solution.
 Our predicted Nubia-South America velocity should not be directly comparable to the NUVEL-1A model, since NUVEL-1A is based on a composite African plate, while we define separate Nubia and Somalia plates. However, we can still compare with the underlying geologic data (Table 5). Despite this difference, REVEL-2000 and NUVEL-1A azimuths in the South Atlantic are indistinguishable from each other and from the geologic data (Figure 18). However, the REVEL-2000 and NUVEL-1A rates differ significantly, with REVEL-2000 slower than both NUVEL-1A and the geologic data upon which it is based, by ∼4 mm/yr through a large range of latitudes (Figure 18). The simplest interpretation is that this reflects true deceleration rather than a systematic bias in either model, since both models are well constrained in this region. Longer-term geologic data support this inference. Figure 19, modified from Cande and Kent , shows a remarkable agreement between the REVEL-2000 geodetic rate and a longer-term trend of decelerating spreading in the south Atlantic going back ∼25 Myr. This is roughly the time of initiation of the current phase of Andean crustal shortening (see discussion of slowing Nazca-South America convergence by Norabuena et al. [1999, and references therein]). It is tempting to speculate that the same process that has slowed convergence between South America and Nazca (e.g., formation of the Andes, associated crustal thickening, and possible increased resistance to subduction) may also contribute to a gradual slowing of South America's westward component of motion and consequent slowing of spreading in the south Atlantic. For example, growth of the Andes and high topography would increase South America's east directed gravitational body force, opposing the west directed body force associated with spreading at the Mid-Atlantic Ridge.
 We define the Pacific plate using 9 sites (Table 1 and Figure 13), excluding two sites (mkea and upo1) on the Big Island of Hawaii that may be affected by deformation associated with the active Mauna Loa and Kilauea volcanoes. We agree with the conclusions of DeMets and Dixon  that the NUVEL-1A angular velocity for Pacific-North America significantly underestimates present-day Pacific-North American motion. DeMets  discussed possible biases in the NUVEL-1A estimate for Pacific-North America, in particular, the rate data based on magnetic anomalies from the Alarcon Rise in the southern Gulf of California, which do not reflect the full plate motion due to faults to the west that have been active for most of the last 3 Myr. These may also be active today [Dixon et al., 2000b]. Thus the bias in NUVEL-1A's estimate of Pacific-North America motion has essentially the same explanation as biases in Caribbean-North America and Caribbean-South America motion, namely, incorporation of rate data in the geologic model that, because of tectonic complexity, do not reflect total plate motion.
 Spreading in the southern Gulf of California shows a trend of increasing rate since 2.5 Ma [DeMets, 1995], reflecting a combination of increased “focusing” of plate motion as spreading at the Alarcon Rise increasingly reflects the full plate motion and perhaps a small amount of acceleration. REVEL-2000's estimate of Pacific-North America motion is 1.4 ± 0.5 mm/yr faster than the geologic estimate of DeMets and Dixon  for the southern Gulf of California (Figure 20).
 Six Pacific sites near the plate boundary zone with North America (cicz, farb, scip, sni1, spmx, vndp) have velocities slower than expected if they were on the rigid Pacific plate. Dixon et al. [2000b] and Beavan et al.  explain these discrepancies by a combination of strain accumulation on locked faults of the San Andreas system, plus slip on additional faults offshore to the west [e.g., Sorlien et al., 1999]. Figure 20 suggests that such faults in the region of southern Baja California have a total slip rate less than ∼2 mm/yr (the difference between the REVEL-2000 rate and estimated present-day spreading rate in the Alarcon Rise based on a linear fit to the magnetic anomaly rate data younger than 3 Ma).
 We are unable to directly verify our Australia-Pacific angular velocity against independent geologic data owing to the lack of transform faults or spreading ridges. Australia-Pacific stage pole data suggest little or no change over the last 11 Myr [Sutherland, 1995]. At a point on the Alpine fault in New Zealand (43.5°S, 170°E), REVEL-2000 predicts 42.0 ± 0.6 mm/yr at an azimuth of 248 ± 1°, giving 41 ± 1 mm/yr of fault-parallel slip (azimuth 236°) and 8 ± 1 mm/yr of shortening. The fault-parallel component is larger than the rate measured with episodic GPS, although the fault-normal component is comparable: Beavan et al.  obtain 29.7 ± 1.4 mm/yr of fault-parallel slip and 9.9 ± 1.8 mm/yr of shortening. Beavan et al.  and Sutherland et al.  conclude that the Alpine fault and subparallel faults in New Zealand take up only ∼70–80% of total plate motion.
 Angular velocity estimates for the Philippine plate by conventional geologic approaches are problematic because no spreading ridges bound the plate, and therefore geologic rate data are not available. This problem has been attacked with earthquake slip vector data [Seno et al., 1993] and GPS data [Kato et al., 1996, 1998; Kotake et al., 1998]. We use four sites (GSI1, GSI3, OKTO, and PAL*; χv2 = 1.03) to define the rigid Philippine plate (Figure 5). To test if sites close to the subducting Pacific plate experience strain accumulation or other nonrigid plate effects, we calculated residual velocities for ccjm, gsi2, and haci with respect to the rigid Philippine plate as defined above. Sites ccjm and gsi2, both ∼100 km from the trench, have residual velocities of 5.0 ± 1.7 mm/yr at 274 ± 16° and 3.0 ± 1.8 mm/yr at 142 ± 32°, respectively. Site haci, 200 km from the trench, but within 100 km of Amuria (mainland Japan), has a residual of 11.1 ± 1.7 mm/yr at 119.9 ± 5.4°. The eastward direction of the residuals at gsi2 and haci suggests that these sites are not strongly influenced by subduction strain accumulation from a seismically coupled Pacific plate (otherwise they would move west). Their eastward velocities may represent postseismic effects, or more likely slow spreading of a back arc basin behind the subduction zone. We cannot exclude the possibility that both effects occur, are of opposite sign, and sum to give the observed residual velocity. Site ccjm's westward residual velocity may reflect seismic coupling. Inclusion of any of these sites increases misfit (4 + ccjm, χv2 = 2.05; 4 + gsi2, χv2 = 1.19; 4 + haci, χv2 = 7.01). Site s102, located just offshore Taiwan, clearly does not represent the stable interior of the Philippine plate (4 + s102, χv2 = 21.5) in agreement with Yu et al. [1997, 1999].
 The velocity of guam with respect to the Philippine plate is a measure of spreading across the Mariana Trough, a back arc basin known to be actively extending. We predict a spreading velocity across the Mariana Trough at guam (13.6°N, 144.9°E) of 46.6 ± 1.5 mm/yr at an azimuth of 096.3 ± 1.8°, in broad agreement with published geologic rates (full spreading rate), 30–60 mm/yr [Bibee et al., 1980; Hussong and Uyeda, 1981; Ishihara et al., 2001]. Part of this range may reflect a change in spreading rate with latitude, with rate increasing to the south [Stern et al., 1984]. The subduction velocity of the Pacific plate relative to the overlying plate at the southern Mariana trench is defined by guam's velocity relative to the Pacific, 63.4 ± 1.1 mm/yr at an azimuth of 104.9 ± 0.7°. The velocity of guam could be affected by strain accumulation from the subduction zone; thus the subduction rate estimate is a lower limit. However, seismic coupling in this region is believed to be low owing to the old age of the subducting plate [Uyeda and Kanamori, 1979].
 A similar analysis, and with the same caveats, can be applied to the western margin of the Philippine plate along the central Ryukyu arc, comparing Philippine and South China angular velocities with site oknw. The subduction rate of the Philippine plate in the central Okinawa Trough is defined by oknw's velocity relative to the Philippine plate, 98.1 ± 1.3 mm/yr at 131.9 ± 0.8°. We predict a rate of spreading across the back arc basin at oknw of 21.4 ± 1.4 mm/yr at 173.7 ± 4.1° using our South China angular velocity. This is consistent with estimates of extension from seismic reflection data, 10–20 mm/yr [Park et al., 1998], and earthquake slip vectors azimuths (∼150°) in the central part of the Okinawa Trough [Fabbri and Fournier, 1999].
 For the Somalia (East Africa) plate we have only two sites (MALI and SEY*) (Figure 17). SEY* is site SEY1 omitting the first year of data that is particularly noisy (Table 1) (this is the only example in this study where data were excluded). The resulting Somalia–Nubia angular velocity predicts extension across the East African Rift in an east-southeast to west-northwest direction, at rates and azimuths that are in approximate agreement with geologically based estimates [Jestin et al., 1994; Chu and Gordon, 1999] (Figure 21 and Table 5). For example, at the equator on the active western rift (30°E) we predict 5.5 ± 0.4 mm/yr of extension at an azimuth of 98.3 ± 4.3°. At 10°N, 40°E, we predict 6.9 ± 0.4 mm/yr of extension at 107.9 ± 3.7°.
4.3.14. South China
 Three sites are used to define the angular velocity of the South China block: S012, SHAO, and WUHN (Figure 5). We can estimate the relative motion of adjoining plates or blocks by determining the velocity of selected sites with respect to South China or by predicting relative velocities across faults based on our angular velocity, for comparison to geologic data. Site kunm located southwest of the South China block, across the north-south striking, left-lateral Xiaojiang fault forming the block's southwest boundary, has a velocity of 10.2 ± 2.2 mm/yr with an azimuth of 200 ± 16°. This is consistent with the geologic estimate of 5 ± 3 mm/yr for the Xiaojiang fault [England and Molnar, 1997] and with GPS measurements of Chen et al.  that suggest ∼10 mm/yr motion across the fault. The east-west striking left-lateral Qinling fault defines the northern boundary with Amurian plate. At a point on the western part of the fault (34.5°N, 109°E) our Amuria–South China angular velocity predicts 4.1 ± 1.4 mm/yr at 260 ± 49°, compatible with geologic observations [Zhang et al., 1995] showing 7.2 ± 2.2 mm/yr of motion. On the eastern part of the fault (30°N, 116°E) we predict 3.4 ± 2.5 mm/yr at 278 ± 33°, where Zhang et al.  measure 2 ± 1 mm/yr.