Notice: Wiley Online Library will be unavailable on Saturday 30th July 2016 from 08:00-11:00 BST / 03:00-06:00 EST / 15:00-18:00 SGT for essential maintenance. Apologies for the inconvenience.
 Based on global earthquake catalogs, the hypocenters, nodal planes, and seismic moments of worldwide subduction plate interface earthquakes were extracted for the period between 1900 and 2007. Assuming that the seismogenic zone coincides with the distribution of 5.5 ≤ M < 7 earthquakes, the subduction interface seismogenic zones were mapped for 80% of the trench systems and characterized with geometrical and mechanical parameters. Using this database, correlations were isolated between significant parameters to identify cause-effect relationships. Empirical laws obtained in previous studies were revisited in light of this more complete, accurate, and uniform description of the subduction interface seismogenic zone. The seismogenic zone was usually found to end in a fore-arc mantle, rather than at a Moho depth. The subduction velocity was the first-order controlling parameter for variations in the physical characteristics of plate interfaces, determining both the geometry and mechanical behavior. As such, the fast subduction zones and cold slabs were associated with large and steep plate interfaces, which, in turn, had large seismic rates. The subduction velocity could not account for the potential earthquake magnitude diversity that was observed along the trenches. Events with Mw ≥ 8.5 preferentially occurred in the vicinity of slab edges, where the upper plate was continental and the back-arc strain was neutral. This observation was interpreted in terms of compressive normal stresses along the plate interface. Large lateral ruptures should be promoted in neutral subduction zones due to moderate compressive stresses along the plate interface that allow the rupture to propagate laterally.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 A significant portion of global seismicity occurs at convergent margins either within each deforming plate or along the plate interface [Byrne et al., 1988]. In particular, most Mw ≥ 8.0 earthquakes are shallow depth events that originate along the frictional interfaces between subducting and overriding plates [e.g., Scholz, 2002]. The subduction plate interface seismicity thus accounts for approximately 90% of the total seismic moment that has been globally released during the last century [Pacheco and Sykes, 1992]. This finding highlights the primary role played by the long-term dynamics of convergent margins in short-term seismogenic processes.
 The contribution of each subduction toward the total released worldwide seismic moment is uneven, and large events are restricted to a subset of convergent margins. Despite the short length of instrumental time record data on earthquakes [McCaffrey, 2008], several authors have described the variability of subduction plate interfaces, both in terms of mechanical behavior and geometry [e.g., Ruff and Kanamori, 1980; Peterson and Seno, 1984; Kanamori, 1986; Jarrard, 1986; Tichelaar and Ruff, 1993; Pacheco et al., 1993; Oleskevich et al., 1999; Conrad et al., 2004], and have attempted to correlate these results with related tectonic parameters. In particular, possible relationships between the characteristics of subduction interface seismogenic zones and the subduction/convergence velocity, slab age and upper plate absolute motion have been widely explored, providing insight into how the subduction plate interface behavior mirrors the complexities that occur at convergent margins.
Ruff and Kanamori  and Jarrard  were the first to note that the earthquake magnitude potential of subduction zones is positively correlated with relative plate motions. Faster subduction zones enhance the triggering of the largest events and vice versa. McCaffrey [1994, 1997b] explained that such a correlation is predictable using the Gutenberg and Richter  recurrence relation. However, these data do not explain why the fastest subduction zones (e.g., Tonga or the New Hebrides) are not associated with large earthquakes. Furthermore, these results do not elucidate why the seismic coupling coefficients calculated by Pacheco et al.  are not correlated with subduction velocities.
 Contradictory results also appear when investigating the role played by seismogenic zone geometry on seismic activity. Kelleher et al.  found that the size of events increases with the width of the contact zone; however, Pacheco et al.  found no significant correlation when considering the seismic coupling coefficient, largest event magnitude, or cumulated seismic moment magnitude.
Ruff and Tichelaar  observed a correlation between the coastline location and downdip limit of the major circum-Pacific subduction interface seismogenic zones. These authors inferred that both of these characteristics are determined by the intersection of the overlying plate's Moho with the slab top and downdip edge of the seismogenic zone. However, the seismogenic zones of several subduction zones (e.g., Japan and Sumatra) have recently been found to extend into the fore-arc mantle [Seno, 2005; Dessa et al., 2009].
 In this study, the diversity of the subduction interface seismogenic zones was analyzed using data from global seismicity catalogs that covered a time period of 107 years. Parameters describing the plate interface seismogenic zone geometry (e.g., dip, maximal and minimal depths and maximal and minimal distances to trench) and mechanical behavior (e.g., seismic rate, moment release rate, largest magnitude and seismic coupling coefficient) were calculated and added to the SubductionZones database [Heuret, 2005] to enable statistical comparisons with a wide range of updated tectonic parameters related to worldwide convergent margins. The continuous development of global earthquake catalogs, such as the CMT Harvard catalog [Dziewonski and Woodhouse, 1981], EHB hypocenter catalog [Engdahl et al., 1998] and Centennial catalog [Engdahl and Villaseñor, 2002], as well as 15 additional years of earthquake records, are expected to improve the completeness, accuracy and uniformity of this data set.
 In the present section, we describe the parameters used to characterize the subduction interface seismogenic zone (see the notation section) and the adopted methodology used for their sampling and determination.
2.1. Sampling the Subduction Seismogenic Zone
 The location of the seismogenic zone can be defined using different approaches: (1) the area of the coseismic slip, as estimated for a selected Mw ≥ 8.5 earthquake (this can be estimated from seismic and tsunami waveform data, as well as from geodetic data sets [e.g., Satake, 1993]; this parameter is often assumed to be approximately equivalent to the area of the aftershock distribution [e.g., Schwartz and DeShon, 2007]); (2) the area locked during the interseismic period, as inferred from geodetic measurements of strain accumulation induced at the surface by plate interface locking [e.g., Savage, 1983; Thatcher and Rundle, 1984; Dixon, 1993; Lundgren et al., 1999; Norabuena et al., 1998, 2004] (it is commonly assumed that the fault area that is fully locked during the interseismic period is the same area that will subsequently rupture in the next extreme event); and (3) the area defined by the distribution of all of the Mw 5.5–7.0 thrust earthquake nucleation locations (i.e., not that of a single event) [Pacheco et al., 1993].
 In the present study, following Pacheco et al. , we assumed that the seismogenic zone of subduction boundaries coincides with the distribution of shallow (depth ≤ 70 km) and moderately sized (5.5 ≤ Mw < 7.0) subduction thrust fault earthquakes. The seismogenic zone was mapped by assessing its location, dip, depth, downdip limit and updip limit to determine the parameters that control seismic activity along the plate interface.
 Three complementary global earthquake catalogs were utilized to describe the location, geometry and seismic activity of the subduction interface seismogenic zone: the Harvard CMT catalog, the EHB catalog and the Centennial catalog (see auxiliary material). The 1976–2007 Harvard CMT catalog was used to identify the Mw ≥ 5.5 subduction plate interface thrust earthquakes. The mapping of the seismogenic zone was improved by using, for each of the identified thrust events, the location given in the EHB catalog. The Centennial catalog was used to extend this study to the 1900–1975 time period for Mw ≥ 7.0 earthquakes.
 For each subduction zone, a set of 2° wide trench-normal transects were constructed that were spaced by 1° along the trench. For each of the 505 resulting transects, the subduction interface seismogenic zone was mapped by selecting from the Harvard CMT catalog the shallow (depth ≤ 70 km) thrust earthquakes for which one nodal plane was consistent with the plate interface geometry and orientation (see auxiliary material). The events selected from the Centennial catalog included all of the shallow and Mw ≥ 7.0 events during the 1900–1975 time period that were located on the slab between the volcanic arc and 50 km in front of the trench. All of these events were supposed to occur along the plate interface.
 The 505 transects were merged into 62 segments so that the rupture areas inferred for Mw ≥ 8.0 earthquakes were included in single segments (Figure 1; see auxiliary material).
2.1.2. Parameters of the Subduction Interface Seismogenic Zone
22.214.171.124. Geometrical Parameters
 Distributions of the earthquake dip, distance to trench and depth for 5.5 ≤ Mw < 7.0 events were used as mapping tools for delimitation of the seismogenic zone, with the assumption that all of the identified thrust earthquakes occurred on plate interfaces and not on secondary faults (Figure 2; see auxiliary material). The cutoff for the moment magnitude Mw ≥ 5.5, as shown by Pacheco et al. , enabled a reliable determination of the earthquake source and depth from the Harvard CMT data. When less than ten events with Mw 5.5–7.0 were recorded, the seismicity was considered as too sparse to allow any reasonable mapping from the global seismicity catalogs alone. Finally, 42 of the initial set of 62 segments were mapped using the global catalogs, and Cascadia, E-Alaska, southern Chile and Nankai were constrained using results presented by Oleskevich et al. . The geometries of the 16 remaining N5.5–7.0 < 10 event segments were not measured. Examples of the trench-normal hypocenter distributions are presented in Figure 2 (the entire set is in the auxiliary material).
 Several previous studies focused on the determination of seismogenic zone depth limits [e.g., Pacheco et al., 1993; Tichelaar and Ruff, 1993], whereas other reports have focused on the characterization of distance-to-trench limits [e.g., Ruff and Tichelaar, 1996]. Here, both of these limits were analyzed together for the first time. The horizontal (x) and vertical (z) coordinates of the seismogenic zone updip (U) and downdip (D) limits were calculated. In a similar manner to Pacheco et al. , the vertical limits (Uz and Dz) were taken to correspond to the 5th and 95th percentiles, respectively, of the thrust event depth distribution. In the same way, the trenchward and arcward horizontal limits of the seismogenic zone (Ux and Dx) were defined as the 5th and 95th percentiles, respectively, of the trench-earthquake distance distribution (Figure 2). This percentile choice enabled a reduction in uncertainties of the depth and distance-to-trench determinations due to outliers in the distribution caused by unknown errors.
 We assumed [Ux; Uz] and [Dx; Dz] as the coordinates of the updip (U) and downdip (D) limits, respectively, of the seismogenic zone. Thus, the seismogenic zone downdip width W is given by the distance between U and D, and the dip angle θ is determined by the angle between the horizontal and the line (UD) (Figure 2). The horizontal and vertical extents of the seismogenic zone (Wx and Wz) were also constrained with this method.
126.96.36.199. Mechanical Parameters
 A given trench segment is characterized by its along-strike length L, its downdip width W and by N plate interface thrust events that have been recorded during a period T. Based on these events, the mechanical behavior of the segment can be described using quantitative measures of the seismicity, i.e., the seismic rate, released seismic energy, and seismic coupling coefficient (see auxiliary material). For the Mw calculation, the moment-magnitude relation by Kanamori  was utilized:
 The seismic rate τ was defined as the number of subduction interface events recorded during a century and along 1000 km of a trench. This parameter was computed across the selection of Mw ≥ 5.5 plate interface events from the Harvard CMT catalog.
 The amount of seismic energy released and the seismic coupling parameters were computed for the entire period of 107 years covered by the 1900–1975 Centennial catalog and the 1976–2007 Harvard catalog. The seismic energy released along the trench segment is described by Mmax, the maximum earthquake magnitude observed along the segment, and the moment release rate MRR. MRR is defined as the cumulated seismic moment released by N events of the segment during a century and along 1000 km of the trench [Peterson and Seno, 1984]:
where Moi is the seismic moment released by an individual earthquake. Subsequently, MRR was converted into an equivalent representative magnitude MMRR using equation (1).
 The seismic coupling coefficient χ [Peterson and Seno, 1984; Pacheco et al., 1993; Scholz and Campos, 1995; McCaffrey, 1997a; Scholz, 2002] is determined by the ratio between the seismic slip rate vss and the subduction velocity vs, as computed from global plate models. The seismic coupling coefficient was estimated by taking into account either the total subduction rate vs or its trench-normal component vs(n), yielding the corresponding coefficients of χ and χ(n), respectively. The seismic slip rates were computed according to the relationship given by Brune  for a given subduction zone:
where the plate interface rigidity μ is taken as an average (5 × 1010 N/m2).
2.2. Other Subduction Parameters
 Most of the other subduction parameters (e.g., subducting plate age, absolute and relative plate motions, upper plate strain, slab geometry, arc-trench distance and upper plate nature) were extracted from the global database already presented by Heuret , Heuret and Lallemand , and Lallemand et al. . For the present study, the number of transects was doubled with respect to earlier studies (one transect per trench degree, approximately 100 km, along the subduction zones), and the data were averaged for each of the 62 segments.
 The relative plate motions are described in the NUVEL1A model [DeMets et al., 1990], and the absolute plate motions were estimated using the HS3 hot spot reference frame [Gripp and Gordon, 2002]. The values of vs, vc, Vup, Vt and Vsub were defined as the velocities of the subduction, convergence, upper, trench and subducting plates, respectively (with lowercase letters used for relative velocities and uppercase letters for absolute velocities). The value of Vt was calculated by subtracting the back-arc deformation velocity vd (estimated using geodetic measurements) from Vup while neglecting erosion and accretion at the toe of the margin's wedge. The subduction velocity vs is the relative motion of the subducting plate with respect to the trench (vs = Vsub − Vt), and vc is the relative motion of the subducting plate with respect to the upper plate (vc = Vsub − Vup; vs = vc if vd = 0).
 The type of upper plate strain (UPS) was determined from the focal mechanisms of shallow earthquakes occurring at depths less than 40 km from the surface of the upper plate, far from the subduction interface. The simplified strain classification system of three classes, as described by Lallemand et al. , was utilized: extensional (rifting or spreading), neutral (no significant deformation or strike-slip), and compressive (shortening). Upper plate nature (UPN) differentiate between oceanic and continental fore arcs.
 The slab age A was extracted from the Müller et al.  digital age grid. The thermal parameter ϕ was calculated from the initial temperature of the subducting plate (proportional to A) and its warming velocity at depth [Kirby et al., 1996].
 The mean arc-trench distance darc-t was calculated normal to the trench. The location of the active volcanoes was taken from the Smithsonian Institution Database [Siebert and Simkin, 2002].
Wada and Wang  compiled estimates of the upper plate Moho depth below the fore arcs zMoho for 17 of the selected 62 trench segments. These values were used to calculate Δz, the depth difference between the fore-arc Moho and the downdip limit of the subduction interface seismogenic zone:
3. Statistical Analysis
 There were two main objectives for the statistical analysis carried out on the subduction plate interfaces: (1) to describe the diversity of the subduction plate interface physical characteristics for the 62 worldwide subduction zones (Table 1) and (2) to determine quantitative relationships between the subduction plate interface parameters and independent subduction parameters that have already been compiled in previous publications [Heuret and Lallemand, 2005; Lallemand et al., 2005; Heuret, 2005; Wu et al., 2008; Wada and Wang, 2009], including the arc-trench distance, relative and absolute plate kinematics, upper plate strain, age and thermal parameter of the subducted plate, upper plate nature and fore-arc Moho depth.
Table 1. Seismogenic Zone Parameters of the Worldwide Subduction Zonesa
Symbols are as in the notation section. Trench segments are mapped in Figure 1.
 The diversity of the physical characteristics of the subduction interface was described using various statistical indicators (e.g., mean and median values, standard deviation and minimum and maximum values), histograms and maps of the along-trench variations for each subduction interface parameter.
 The significance of the bivariate relationships between the subduction plate interface parameters and independent variables was statistically tested using linear regression analysis. The potential to fit the data using this type of linear regression analysis was initially quantified using the Pearson's product-moment correlation coefficient R. The correlation coefficients are presented in Figure 3. This analysis shed light on the first-order parameters, identifying the most relevant relationships between the subduction interface characteristics and other subduction parameters. As the R values were quite low on average (mean R value was 0.25 ± 0.21), section 3 only describes the most statistically significant relationships, as well as the most meaningful ones, in terms of the physics and mechanisms of the system. To enable a more detailed test of the reliability of each described linear regression, we also considered the standard error (σe) and the results of t tests for the slope and intercept values at a 95% confidence level. We have arbitrarily considered as significant the relations that had ∣R∣ > 0.30 and p values ≤ 0.05 (5%). The corresponding trends were then represented in Figures 7, 9, and 12. Figures 7, 9, and 12 account for the standard errors of the y intercept and slope (see auxiliary material for the detailed analysis).
 Outliers were frequent in most of the bivariate relationships that have been analyzed. Some of this outliers were constant from one relation to the other. This was particularly verified for the relationships between vs (and ϕ) and most of the subduction interface parameters (Figures 7 and 9). Most of these relationships have appeared nonlinear for the fastest subductions, i.e., for northern Tonga and northern New Hebrides. Nonlinearity and alternative relationships (e.g., exponential or second-degree polynomial) were difficult to test statistically because only a few fast subductions exist (80% of vs values were lower than 80 mm yr−1 and only northern Tonga and northern New Hebrides vs values were larger than 130 mm yr−1). Thus, for Figures 7 and 9, we have decided to test the reliability of linear relationships both considering or excluding northern Tonga and northern New Hebrides (see auxiliary material). These two segments are however neglected in the correlations presented in Figures 7 and 9.
3.1. Geometry of the Subduction Interface Seismogenic Zone
3.1.1. Global Variability
 Most of the seismogenic zones exhibited shallow dipping geometries (Figure 4a). The mean θ value was 23 ± 8° and the largest value was 35° (central Aleutians). Outliers with exceptionally steep dips, however, were observed for the incipient subduction zones of Cotobato (51°), Wetar (58°) and Flores (71°). It is possible that these zones have not yet reached a steady state configuration. Their positions are highlighted in Figures 4 and 5 for completeness; however, they are excluded from the following statistical analysis.
 The average global view (Figure 6) shows that U exhibited a limited variability with respect to D, especially in depth (Uz mean location was at 11 ± 4 km, whereas Dz was at a depth of approximately 51 ± 9 km). In turn, Wz accounted mainly for the variability in Dz (R = 0.90 for the [Dz, Wz] relation; Figures 3, 4b, and 4c).
 The geometrical relations between θ, Wz and Wx imply that the variability in Wx (and W) is defined by the combined variations in θ and Wz (i.e., Dz). These relations are illustrated in Figure 6 using the observation that Uz was approximately constant at a 10 km depth. At first order, Wx and W were determined by θ (R = −0.81; Figure 3). In this case, flat geometries allowed for large W values (mean value was 112 ± 40 km, but values larger than 180 km were observed for subduction zones such as the Andaman and western Aegean; Figure 4 and Table 1) that were on average 12 km larger than Wx (R = 0.99; Figure 3). However, this effect was self-limited by the opposite and second-order correlation between θ and Wz (R = 0.46; Figure 3) so that flat geometries were associated at the same time with large values of Wx and small values of Wz.
3.1.2. Correlations With Subduction Parameters
 The assessed geometrical parameters were found to be fairly uncorrelated with most of the subduction parameters (∣R∣ mean value was 0.21 ± 0.16), in particular the fore-arc Moho depth, slab age, and absolute plate and trench motions (∣R∣ < 0.20; Figure 3). However, significant correlations were observed between the relative plate motions and ϕ.
188.8.131.52. Location of the Volcanic Arc and the Fore-Arc Moho
 The parameter Dx was found to extend up to 65 ± 20% of the trench-arc distance, and Wx was found to represent 50 ± 20% of this distance (Figure 6). A strong relation was observed between these two structural features, and significant correlations were observed between darc-t and Dx (R = 0.67; Figure 3) or Wx (R = 0.60; Figure 3). However, these parameters were derived from a first-order geometrical control (that is, R = 0.69 for the [θ, darc-t] relation and R = 0.69 for the [θ, Dx] relation; Figure 3).
 The upper plate Moho has often been presented as a structural depth limit for the seismogenic zone [e.g., Ruff and Tichelaar, 1996; Oleskevich et al., 1999]. This hypothesis was tested for the following systems: (1) the whole set of segments, using the upper plate nature as a proxy for the fore-arc Moho depth and (2) for the set of 17 segments for which updated fore-arc Moho depth data were available [Wada and Wang, 2009]. Continental upper plates represent approximately 80% of the Wada and Wang  database. This proportion reflects the presence of continental upper plates in the complete subduction data set. In intraoceanic subduction zones, the fore-arc Moho is as shallow as approximately 10 km in depth. However, in continental fore arcs, it is reported at a depth of 34 ± 11 km. No relation has been found between the Moho depth (or the upper plate nature) and Dz, Wz or any other seismogenic zone geometry parameter. In fact, 70% of the seismogenic zones extend more than 10 km below the upper plate Moho (i.e., in contact with the fore-arc mantle; Figure 6c). The mean Δz value was −23 ± 20 km. The largest Δz values have been observed for intraoceanic subductions and reach −40 km.
184.108.40.206. Subduction Velocity
 The relative plate motions vs and vc exhibited some of the most significant correlations with the subduction interfaces geometrical parameters. As the subduction velocity increased, the values of θ, Dz and Wz increased and the value of W decreased. However, the correlations with θ (R = 0.55; Figure 7a) and W (R = −0.45; Figure 7d) and, to a lower extent, with Wz (R = 0.30; Figure 7c) and Dz (R = −0.34; Figure 7b) were improved by 25%–50% if the fastest subduction zones (northern Tonga and northern New Hebrides) were removed from the statistical analysis or if second-order polynomial regressions were used. This finding hints at possible inversions in the tendencies over vs values of approximately 80–150 mm yr−1. The parameter Uz was found to be uncorrelated with vs (R = −0.04; Figure 7b).
220.127.116.11. Thermal Parameter of the Subducting Plate
 The thermal parameter exhibited a significant correlation coefficient with the subduction interfaces geometrical parameters. The values of Dz, Wz and θ increased as the slab became colder (that is, ϕ became lower) whereas Wx. However, the correlations with θ (R = 0.49; Figure 7e), Dz (R = −0.42; Figure 7f) and Wz (R = 0.40; Figure 7g) and, to a lower extent, with W (R = −0.30; Figure 7h) were improved by 25%–70% if northern Tonga and northern New Hebrides segments were removed from the statistical analysis or if second-order polynomial regressions were used. This finding hints at possible inversions in the tendencies over ϕ values of approximately 4500 km. There was no correlation between Uz and ϕ (R = −0.23; Figure 7b).
3.2. Mechanical Behavior of the Subduction Interface Seismogenic Zone
3.2.1. Global Variability
 For the time period from 1976 to 2007, a total of 3283 Mw ≥ 5.5 plate interface earthquakes have been identified, 4.2% of which were Mw ≥ 7.0 events. For the time period from 1900 to 2007, a total of 726 large subduction plate interface events were considered. Among these, 49 were Mw ≥ 8.0 earthquakes and 13 were Mw ≥ 8.5 events. The distributions of the seismic rate (Figure 8a) and largest magnitude events (Figure 8b) were heterogeneous along the trench segments.
 The mean τ value was approximately 210 events when normalized to 100 years and 1000 km of trench length (Figure 8a), but the distribution was lognormal. In other words, the lowest rates were the most widely represented (65% of the segments had τ values lower than 200 events). The largest values reached 1000 events (e.g., north Kermadec, Bougainville and New Britain; Table 1).
 The Mmax mean value was 7.8 ± 0.7, and 75% of the segments experienced at least one Mw 7.5 event. The south Chile event (Mw 9.5, 1960) was the largest ever observed. The 14 recorded Mw ≥ 8.5 earthquakes were confined to only four different trench systems: South America (5 events), Japan-Kuril (3 events), Aleutians-Alaska (3 events) and Sunda (3 events) which also experienced 28%, 22%, 10% and 8%, respectively, of the 50 subduction interface Mw ≥ 8.0 earthquakes.
 The seismic moment released by the subduction interface during the time period of 1900–2007 was approximately 6.9 × 1023 N·m, representing 80.2% of the total seismic moment released worldwide over all geodynamical environments. The value of MRR was variable along the trenches (Figure 8c), and its mean value was 1022 N·m when normalized to 100 years and 1000 km of trench length (MMRR = 8.2 ± 0.6). The seismic slip rates associated with the moment rates for most trench segments were low with respect to their subduction velocities, and the seismic coupling coefficients were generally low (Figure 8d). The parameter χ(n) exhibited a similar distribution; however, it was shifted slightly toward higher values, as vs(n) was lower than vs. The mean χ value was 0.25, and 56% of the segments had values lower than 0.2.
 For a given trench segment, MRR and χ were functions of both Mmax and τ. However, the largest MRR values were controlled by the occurrence of Mw ≥ 8.0 earthquakes, rather than by the number of events. Thus, the sole south Chile event (Mw 9.5, 1960) represents 33% of the seismic energy released along the subduction interface. As a consequence of the MMRR dependence on Mmax, χ was also well correlated with Mmax (R = 0.74; Figure 3). Further, the occurrence of at least one Mw ≥ 8.5 earthquake is a necessary condition for a given segment to have a fully coupled interface (χ ≥ 0.8; 12% of the cases). In fact, the segments where the largest events occurred were those where the largest MRR and χ were measured. Because Mmax was found to highly influence the moment rate and extent of seismic coupling, we focused our analysis on the variability in Mmax.
3.2.2. Correlations With Subduction Parameters
 The mechanical parameters were fairly uncorrelated with most of the subduction parameters (∣R∣ mean value was 0.21 ± 0.16). However, except for χ, significant correlations were observed with the subduction velocity (0.30 < ∣R∣ < 0.60; Figure 3). Seismic rate was the only parameter that correlated with the subducting plate thermal state (R ≈ 0.60; Figure 3).
18.104.22.168. Subduction Velocity
 Among all the subduction parameters, vs and vsn were the most significantly correlated with the mechanical parameters and, in particular, the seismic rate (R = 0.61; Figure 9d). Better correlation coefficients were observed when taking into account the cumulated seismic moments that were released by all of the events registered in the segment (R = 0.50 for [MMRR; vs] relation; Figure 3), rather than only the moment released by the largest event (R = 0.34 for [Mmax; vs] relation; Figure 3). No correlation was observed between χ and the subduction velocity (R = 0.04; Figure 3).
 The relationships between vs and the mechanical parameters were not strictly linear and the correlations were improved when northern Tonga and northern New Hebrides were not considered. Indeed, τ increased with increasing vs slowly for the low-velocity plates (vs < 80 mm yr−1) and sped up for the higher-velocity plates (Figure 9d). An exponential regression only moderately improved the correlation coefficient (approximately 6%). The largest Mmax (or MMRR) values were not associated with the fastest subduction zones (Figures 9a and 9b). For example, northern Tonga (vs = 220 mm yr−1), had never experienced an Mw > 8.1 interface event during the time span of the seismic records used in this work. In fact, Mmax (or MMRR) increased linearly with the subduction velocity, up to approximately 80 mm yr−1. For larger velocity values, Mmax and MMRR seemed to gradually decrease (Mmax ≤ 8.1 and MMRR ≤ 8.8). A second-degree polynomial regression improved the correlation of these parameters with vs (R = 0.42 and 0.59 for Mmax and MMRR, respectively). A strong contrast was also found between the slowest and fastest subduction zones in terms of seismic coupling (Figure 9c). All of the subduction zones with vs < 80 mm yr−1 had χ > 0.2. In contrast, for larger velocities, all of the subduction zones exhibited low seismic coupling (χ < 0.2).
 Among the other kinematical parameters, τ was also reasonably correlated with Vsub (R = 0.57; Figure 3), but not with Vup or Vt. Vup, and Vt also exhibited negligible correlations with Mmax (R ≈ −0.15; Figure 3). Advancement of the upper plate did not systematically promote Mw ≥ 8.5 earthquakes, and upper plate retreat was not always associated with a low Mmax. In fact, the fastest upper plates (or trenches) were never associated with Mw ≥ 8.5 earthquakes, as has been argued in previous studies [e.g., Uyeda and Kanamori, 1979; Conrad et al., 2004]. Events with Mmax ≥ 8.5 were limited to relatively fixed trenches and upper plates that had velocities ranging between −30 mm yr−1 and 50 mm yr−1 in the Pacific hot spot reference frame HS3 [Gripp and Gordon, 2002].
22.214.171.124. Thermal Parameter of the Subducting Plate
 The subduction velocity was not able to explain the diversity in Mmax, and high scattering remained inside the general trend. To explain this scattering, second-order parameters should be involved. Ruff and Kanamori  observed a negative correlation between Mmax and the slab age. They argued that a large Mmax can be explained by the combined influence of a young subducting plate and fast subduction velocity. However, in the revised data set obtained here (Figure 10), no significant correlation was observed between the Mmax and plate age (R = 0.05; Figure 3). Thus, it is necessary to review the theory proposed by Ruff and Kanamori . Recently, a similar conclusion was reached by Stein and Okal  and Gutscher and Westbrook  thanks to a detailed study of past Mw ≥ 8.0 earthquakes [e.g., Clague, 1997; Satake and Atwater, 2007] and the occurrence of the peculiar recent 2004 Sumatra event [Lay et al., 2005]. The thermal parameter ϕ, which is an alternative way to combine vs and the plate age influence, was not correlated with Mmax (R = 0.09; Figure 3), MMRR (R = −0.10; Figure 3) or χ (R = 0.18; Figure 3). Despite the seismic rate was uncorrelated with the slab age (R = 0.17; Figure 3), it was found that large ϕ (i.e., cold slabs) promote a large number of plate interface earthquakes (R = 0.54; Figure 3).
126.96.36.199. Upper Plate Nature
 A strong contrast was observed between the plate interface mechanical behavior of intraoceanic subduction zones and those associated with continental upper plates (Figure 11a). All of the observed Mw ≥ 8.5 events were associated with continental upper plates (the Mmax mean value was 7.9 ± 0.7). However, the largest event experienced at an intraoceanic plate interface during the last century was only 8.1 (the Mmax mean value was 7.5 ± 0.6). The same trend was observed with χ (Figure 11b) and MMRR. The oceanic upper plates were associated with a low seismic moment release and with low seismic coupling values (χ ≤ 0.2). No particular relationship was observed with τ (Figure 11c).
188.8.131.52. Upper Plate Strain
 Based on the study by Uyeda and Kanamori , it is commonly thought that Mw ≥ 8.5 earthquakes are associated with compressive upper plate strain and that Mmax decreases when the strain becomes extensive. However, the correlation between Mmax and UPS was found to be low (R = 0.19; Figure 3), and as previously noticed by Hayes and Conrad , Mw ≥ 8.5 earthquakes were preferentially associated with a neutral UPS (Figure 11b). The compressive back arcs exhibited intermediate values (Mmax ≤ 8.7), and high values were never observed where back-arc spreading or rifting occurred (Mmax ≤ 8.1). The Sumatra event (2004; Mw = 9.0) was not an exception to this trend. Even when the ruptured area extended from northern Sumatra up to the Andaman extensive margin, the earthquake nucleation occurred in an area characterized by a neutral tectonic regime.
 Similar trends were observed for χ and MMRR. In particular, the largest χ values were preferentially associated with neutral back-arc deformations (Figure 11c), whereas compressive back arcs showed intermediate values. Large values were never observed where back-arc spreading or rifting was found to occur (χ ≤ 0.20). No particular relationship was observed with τ (Figure 11f).
3.3. Relationships Between the Mechanical Behavior and the Subduction Fault Geometry
3.3.1. Downdip Width and Dip of the Seismogenic Zone
 Theoretically, the earthquake magnitude potential of a given trench segment should increase with the surface of the potential rupture area and, in particular, with the downdip width of the coupled zone. Correlations of Mmax with θ (R = 0.32; Figure 12a) and W (R = −0.19; Figure 12b) were found to be weak. The downdip width was not the key parameter in determining the occurrence of Mw ≥ 8.5 earthquakes. This observation was confirmed by the fact that some of the segments that expressed Mw ≥ 8.5 earthquakes were associated with steep and/or narrow seismogenic zones (e.g., central Aleutians: θ = 35°, W = 75 km; Table 1). Moreover, despite the fact that Mw ≥ 8.5 earthquakes were often associated with the flattest and widest seismogenic zones, some segments with similar seismogenic zone geometries did not express any Mw ≥ 8.5 events during the last century (e.g., Antilles, Java and the western Aegean; Table 1).
 As with Mmax, similar trends were observed for MMRR and χ. However, better correlations were observed for the seismic rate. For example, the value of τ was found to increase with θ (R = 0.50; Figure 12c) and decrease with W (R = −0.31; Figure 12d).
3.3.2. Along-Trench Variability of the Subduction Interface Mechanical Behavior
 The values of τ and Mmax were uncorrelated at first order (R = 0.25; Figure 3). However, noteworthy characteristics of the [Mmax; τ] relation and for the occurrence of Mw ≥ 8.5 events were observable when focusing on their variability along isolated trench systems. The longest subduction zones (i.e., Sunda, South America, Aleutians-Alaska and NE Japan-Kuril) expressed strong lateral variabilities from their edges to centers with respect to their plate interface mechanical parameters, namely in terms of τ and Mmax (Figures 8a and 8b). Figure 13a shows how τ varied along the half-trench systems, from one slab edge toward the slab center (for each trench system, the slab center was defined as the maximum τ value location). Similarly, Figure 13b shows the Mmax along-trench variations for the half-trench systems (the slab center was then determined by the minimum Mmax value locations).
 It appeared that the variability was not randomly organized. Comparison of the contrast in slab edges with slab centers yielded opposite trends when considering τ or Mmax. The lowest τ values were located near slab edges, and they progressively increased toward slab centers (Figure 13a). Conversely, the largest Mmax occurred near slab edges and progressively decreased toward slab centers (Figure 13b). Thus, large seismic energy releases in a few large events at slab edges contrasted with slab centers, where lower seismic energy levels were obtained in numerous events of relatively smaller sizes. As a result, slab edges appeared much more seismically coupled than the centers.
 An original database was built to describe the subduction interface seismogenic zone variability in terms of geometry and mechanical behavior, as well as to compare these constraints to other subduction parameters, revisiting some empirical laws that were obtained in previous studies [e.g., Ruff and Kanamori, 1980; Kanamori, 1986; Jarrard, 1986; Tichelaar and Ruff, 1993; Pacheco et al., 1993; Conrad et al., 2004]. Correlation coefficients were often low (that is, the R value was 0.25 ± 0.21), and the lack of a strong relationship between parameters emerged from the statistical analysis. This complexity in the statistical bivariant relations between collected parameters reflects, to some extent, the intricate subduction related physical processes. Subduction was settled by the interplay of numerous possible variables so that a single parameter was never able to explain the whole diversity observed at subduction interfaces. However, the statistical scattering was also largely related to the limited time span and accuracy of the collected seismic observations, such that most of the estimated parameters (e.g., Dz, W, Mmax and χ) have to be considered as lower bounds; thus, it was necessary to approach their interpretation with caution. As an example, the instrumental record gives low values for the seismicity parameters of the Cascadia subduction (Mmax = 7.6 and χ = 0.01; Table 1), whereas there was evidence for a Mw = 9.0 earthquake in that location approximately 300 years ago [Satake et al., 2003]. Under such conditions, it is unfair to discuss accurate empirical equations between these parameters. It is, however, possible and useful to discuss the most reliable statistical trends and the most meaningful physical relations.
4.1. Global Variability
 Our analysis confirms that most of the seismogenic zones exhibited shallow dipping geometries (Figures 4 and 5) with dip angles (θ) between 10° and 35° and a mean dip value at 23 ± 8°. The seismogenic zone downdip width (W) distribution spanned over a wide range, with most of the values lying between 60 km and 180 km and with a mean downdip extent of 112 km ± 40 km. These estimates are consistent with previous work by Pacheco et al.  (W = 95 km ± 27 km), considering the adopted limits of the different data sets, selection criteria and computation methods. Our estimates for the downdip limit depth (Dz = 51 km ± 8 km; Figure 6) are also consistent with that of Pacheco et al.  (Dz = 49 km ± 11 km), but they were on average 10 km deeper than those found by Tichelaar and Ruff , who used a smaller set of subduction zones and constrained the depth limit by considering only the rupture area of large events. Though the fore-arc Moho has often been presented as a structural depth limit for the seismogenic zone [e.g., Ruff and Tichelaar, 1996], we found that 70% of the seismogenic zones extended more than 10 km below the fore-arc Moho, lying in contact with the fore-arc mantle (Figure 6). Because the lower aseismic limit of the seismogenic zone has often been explained as being caused by mantle serpentinization [e.g., Reinen et al., 1992; Hyndman et al., 1997; Moore et al., 1997], our results suggest that the extent of serpentinization of the mantle wedge may differ locally [e.g., Seno, 2005].
4.1.2. Mechanical Parameters
 In the present study, we confirmed that the seismic potential of the subduction plate interface was large (the Mmax and MMRR mean values were 7.8 ± 0.7 and 8.2 ± 0.6, respectively; Figure 4b) with respect to other geodynamical environments [e.g., Pacheco and Sykes, 1992]. The plate interface released 80% of the worldwide seismic moment. Regardless of the subduction conditions, this was able to generate large earthquakes (75% of the segments experienced at least one Mw 7.5 event).
 In agreement with previous studies [e.g., Pacheco et al., 1993; McCaffrey, 1997a], the average seismic coupling coefficient was generally found to be low (χ = 0.25 ± 0.30; Figure 8d), indicating that the slip over the last 107 years was either mostly aseismic or that the strain was accumulating. However, Bird and Kagan  found a higher average seismic coupling (χ = 0.69), probably due to their different approach; the shallow seismicity was cumulated over all subduction zones, leading to an overestimation of the seismic coupling due to the following reasons: (1) there was no distinction between the interface and intraplate events and (2) at subduction zones where Mw ≥ 8.5 events occurred, χ was often larger than 1.0, which is physically impossible but not detectable in the Bird and Kagan  view (e.g., χ = 3.5 in southern Chile; Table 1). In our work, such values were fixed at 1.0.
4.2. Subduction Interface Variability and Subduction Velocity
4.2.1. Geometry and Seismic Rate
 Among all the subduction parameters, vs and ϕ were the most significantly correlated with the plate interface parameters (Figures 7 and 9), especially with the geometry and seismic rate (0.40 < ∣R∣ < 0.60; Figure 3). These correlations depict a coherent portrait of the seismogenic zone variability, in which fast and cold subduction zones produced a large number of moderate earthquakes over a narrow, deep and steeply dipping seismogenic interface (Figure 14). Physically, this means that the subduction velocity vs may have affected the seismogenic zone behavior through a combination of stress accumulation and temperature-related processes (the faster a slab penetrates into the hot mantle, the slower it warms; that is, fast subduction zones are associated with cold slabs).
 The increase of Dz and Wz with vs may be mainly explained by temperature-related processes, as supported by the improved correlation when ϕ, rather than vs, was considered. Thus, in fast and cold subduction zones, the frictional properties of the plate interface are likely maintained at large depths. As the seismogenic zone often ends in the fore-arc mantle, the process could imply slab dehydration. This confirms the theory put forth by Wada and Wang , which highlights how slab dehydration occurs deeper in the mantle in cold compared to warm subduction zones. This results in a lower amount of fore-arc mantle serpentinization and a greater tendency for the subduction interface to be seismogenic. The fundamental role that the mechanical contribution can play in the subduction interface seismic behavior has also to be considered to explain the fact that the Moho falls above the downdip limit of the seismogenic zone (F. Corbi et al., Seismic variability of subduction thrust faults: Insights from laboratory models, submitted to Journal of Geophysical Research, 2010). In contrast with our observations, Pacheco et al.  did not find any significant dependence of θ or W on subduction velocity. The improved and increased earthquake records adopted in this study are likely the key to explaining such differences.
 The positive exponential correlation observed between vs and τ (that is, τ increased slowly for sluggish plates and sped up for fast plates; Figure 9d) is coherent with the Gutenberg and Richter  recurrence law, confirming a common feature of frictional dynamics observed in laboratory experiments [e.g., Baumberger et al., 1994; Corbi et al., submitted manuscript, 2010] and theoretical calculations of fault slip rates [Molnar, 1979; McCaffrey, 1997b]. Moreover, this observation is consistent with the nonlinear relation found by Bird et al.  between earthquake production and relative plate velocity in subduction zones (slower subduction zones with vs ≤ 66 mm yr−1 are found to represent only 20% of earthquake productivity and 35% of the cumulative tectonic moment rate). The positive correlation between τ and ϕ opens the possibility of a significant influence from additional temperature-related processes (low temperatures may favor unstable sliding of the plate interface).
4.2.2. Largest Observed Magnitudes, Moment Release, and Seismic Coupling
 The relationships between vs and Mmax, MMRR and χ were difficult to interpret because the period record analyzed was shorter than the recurrence time of most Mw ≥ 8.5 earthquakes. On the one hand, this implies uncertainties in the observed trend that could partly explain the data scattering. On the other hand, the recurrence time of Mw ≥ 8.5 earthquakes is known to be a function of vs (see Gutscher and Westbrook  for a review). It is thus possible that the observed trend is just a probabilistic consequence of the [recurrence time; vs] relation (that is, the probability for a Mw ≥ 8.5 earthquake to be expressed is larger during a given time period for fast subduction than for slow subduction) and it may disappear on a larger time scale. Following a similar hypothesis, McCaffrey  argued that every subduction zone in the world is able to generate Mw ≥ 8.5 earthquakes, and their occurrence should just be a question of time.
 The subduction velocity is, at first order, positively correlated with Mmax (R = 0.42) and MMRR (R = 0.59) (Figures 9a and 9b), as expected from the Gutenberg and Richter  recurrence law. Better correlations were observed for the [vs; MMRR] relationship than for the [vs; Mmax] relationship because, as a consequence of the Gutenberg and Richter  recurrence law, the cumulated seismic moment is much more representative of the long-term moment rates than Mmax. Indeed, as the time interval for the available seismic data is short, the largest possible events have likely not been observed yet in every subduction zones, especially the slowest ones, whereas lower size events are more frequent. The seismic coupling χ was found to be extremely low and uncorrelated with the subduction velocity (R = 0.17; Figure 9c), confirming the observation of Pacheco et al. . This proves that, on the 1 century time scale, the seismic potential is far from being fully expressed in most subduction zones.
 Relations with vs appeared to be nonlinear. Correlation coefficients were improved when second-order polynomial regressions were applied. Though only a few segments reached such velocities, from approximately 80 mm yr−1, the trend was inverted and Mmax/MMRR decreased with vs. As a result, during the 1900–2007 time window explored in the present study, the largest earthquake magnitudes were not associated with the fastest subduction zones, but rather with zones moving at speeds of 20 < vs < 80 mm yr−1. Moreover, none of the vs > 80 mm yr−1 subduction zones (e.g., northern Tonga; vs = 220 mm yr−1; Table 1) had experienced Mw > 8.1 events. This yielded direct consequences on the χ distribution (Figure 9c). The largest χ values were found to be associated with vs < 80 mm yr−1 whereas none of the vs > 80 mm yr−1 subduction zones had χ values > 0.2. Such behavior of the subduction interface seismogenic zone at the fastest subduction velocities is inconsistent with the theory that Mmax is constant over subduction zones in the long term [e.g., McCaffrey, 2008]. Indeed, as the fastest subduction zones are associated with the shortest recurrence times [e.g., Gutscher and Westbrook, 2009], they are theoretically more prone to have experienced Mw ≥ 8.5 earthquakes and to express large χ values during a limited time span. The variabilities in the potential size of earthquakes and in the long-term seismic coupling are thus likely to be real features.
 In summary, our results demonstrate that the subduction velocity cannot be the only parameter that controls the subduction interface variability. The statistical portrait depicted in Figure 14 was not verified for all subduction interface parameters, especially for the genesis of Mw ≥ 8.5 events. The correlation between the subduction interface geometry parameters vs, ϕ, τ and Mmax was generally low, and in most of the bivariate relations, the segments where Mw ≥ 8.5 events occurred were generally out of the main trend (Figures 9, 10, and 12). Among other differences and in opposition to the results of Ruff and Kanamori , the parameters Mmax, MMRR and χ appeared uncorrelated with ϕ. The relationship between seismic moment-related parameters and the nature of the upper plate (Figure 11) likely represents an important ingredient to completing the picture. Subduction zones associated with continental upper plates were on average slower (the mean vs was 53 ± 28 mm yr−1) than intraoceanic subduction zones (the mean vs was 76 ± 55 mm yr−1). As a result, this can represent a possible consequence of the [Mmax; vs] relation for vs < 80 mm yr−1 subduction zones. However, the underrepresentation of intraoceanic subduction zones can bias this observation.
4.3. Conditions for Mw ≥ 8.5 Earthquake Genesis
 The seismic moment released during earthquake failures increases with the event rupture area [e.g., Hanks and Kanamori, 1979]. The potential earthquake magnitude of a given trench segment is thus conditioned by both the maximum possible rupture width and the maximum along-trench length along which the rupture is able to propagate.
 The maximum rupture width, which corresponds to the downdip extent W of the subduction plate interface seismogenic zone, was found to be uncorrelated with Mmax (Figure 12). Megaevents (Mw ≥ 8.5) were associated with the entire range of possible W values, from the largest (e.g., 250 km in the Andaman; Table 1) to the smallest (e.g., 75 km in the central Aleutians; Table 1). The same trend was observed for the plate interface dip (possible values ranged from 9° in the Andaman to 35° in the central Aleutians; Table 1). The fact that megaevents were not associated with any particular value of W highlights the dominant role played by the along-trench rupture propagation [Ruff, 1989; McCaffrey, 2007, 2008; Hayes and Conrad, 2007]. This means that subduction zones characterized by a limited lateral extent (<500 km; e.g., Yap, Palau, Puysegur and Cotobato) [see Lallemand et al., 2005; Heuret, 2005] are unable to generate megaevents, while longer subduction zones may potentially be able to extend the rupture along their whole length (e.g., approximately 6000 km of potential lateral rupture in Sunda, South America or NE Japan-Kuril). The observed lateral ruptures of megaevents were, however, restricted to only limited portions of the trench system (500–2000 km) [e.g., McCann et al., 1979; Kostoglodov, 1988].
 To speculate on the reason for the natural cutoff that occurs in the lateral rupture of interplate subduction faults, it is necessary to recall the mechanism that produces interface events. The subducting plate complexity (i.e., variable sediment thickness, horsts and grabens, fracture zones, aseismic ridges and seamounts) and upper plate strength (see Bilek  for a comprehensive review) are known to produce local frictional variations along the fault, focusing the high-slip location during an earthquake [e.g., Rice, 1993]. This concept is described by the asperity model, where asperity is seismically defined as the area dominated by stick-slip behavior and the high-slip location during an earthquake [e.g., Lay et al., 1982; Kanamori, 1986]. In this view, interface earthquake ruptures are segmented, being separated by weak zones (i.e., barriers [e.g., Seno, 2003]) that inhibit rupture propagation. We found that more than 75% of global subduction zones have Mmax ≥ 7.5 (Figure 8b), regardless of the subduction conditions, which implies that the inherited segmentation is always able to generate moderately large-sized earthquakes. Conversely, the occurrence of megaearthquakes was limited to only a few subduction zones (Figures 1 and 8b), requiring the progressive failure of adjacent asperities (i.e., multisegment ruptures [e.g., Thatcher, 1990; Ruff, 1996]). For example, at the Ecuador-Columbia margin, the area ruptured in 1906 as a single and wide segment (Mw = 8.6 event). This segment was composed of three adjacent seismic segments that ruptured independently into smaller-size events in 1942 (Mw = 7.8), 1958 (Mw = 7.7) and 1979 (Mw = 8.2) [Marcaillou et al., 2008; Bilek, 2009]. The genesis of megaevents should thus be promoted if the rupture is able to easily propagate laterally into adjacent segment(s) through possible barriers or weak zones (e.g., subducting ridges or fracture zones).
 The statistical results presented here show that such behavior seems to be preferentially enhanced in specific areas that are characterized by a neutral back-arc regime or continental upper plate (Figures 12a–12d), as well as those that are in the vicinity of slab borders (Figure 13).
4.3.1. Upper Plate Strain, Plate Coupling, and Megaearthquake Genesis
 Although compressive UPS values were underrepresented in current subduction zones (only 15%, whereas neutral UPS values represent approximately 60% of the subduction segments analyzed here), it is noteworthy that 85% of the recorded Mmax ≥ 8.5 events occurred at neutral UPS subductions. This observation yields to speculation concerning alternative mechanism from those usually though. Since the report by Uyeda and Kanamori , megaearthquakes have indeed been generally associated with compressive back-arc deformations. In this view, where the back-arc area is used as a strain tensor to determine the efficiency of the plate coupling [Lallemand et al., 2008], large compressive stresses recorded in the back-arc area are thought to reflect larger and stronger stress accumulations of the subduction fault and, in turn, larger asperities [Ruff and Kanamori, 1980]. However, as first highlighted by Hayes and Conrad , this mechanism does not take into account that, even if this tectonic framework favors the release of a large seismic moment during the occurrence of the initial asperity rupture, the large critical stresses associated with adjacent asperities may inhibit the rupture from propagating laterally and, in turn, restrain the possibility for the generation of megaevents. The opposite scenario occurs with extensional UPS, where asperities are expected to be smaller. Low critical stresses should theoretically facilitate the lateral propagation of the rupture to adjacent asperities. However, a limited initial seismic moment release may not be sufficient to trigger megaevents. The statistical association of megaearthquakes with neutral subduction zones can thus be explained as being related to the most favorable interplay between a significantly large initial released seismic moment and a low critical stress for the lateral rupture propagation [Hayes and Conrad, 2007; Rosenau and Oncken, 2010; Corbi et al., submitted manuscript, 2010].
 This mechanism could also explain why any association between megaevents and single subduction parameters is missing. The UPS and plate coupling magnitude reflect the balance between driving and resistive forces (e.g., slab pull, bending and mantle resistance) [Lallemand et al., 2008] acting at a given subduction. This balance is supposed to be influenced by a wide range of subduction parameters and should be kinematically mirrored by combinations of the subduction velocity and absolute plate motion [e.g., Uyeda and Kanamori, 1979; Ruff and Kanamori, 1980; Conrad et al., 2004]. Lallemand et al.  found that neutral subductions verify a [Vsub; Vup] relation that is satisfied by a wide range of kinematic combinations and subduction parameters. Thus, the condition that likely enhances the occurrence of megaearthquakes is not associated with any single tested subduction parameter. This is true even for those parameters that are commonly assumed to exert a primary control on the plate interface mechanical behavior. Moreover, neutral UPS may be indifferently associated with the advancing or retreating upper plates. This is in contrast with the arguments of several previous studies by researchers [Uyeda and Kanamori, 1979; Peterson and Seno, 1984; Conrad et al., 2004] who recognized that the occurrence of megaearthquakes correlates with the advancement of the upper plate.
 By assuming UPS as an indicator of the ability of the rupture to propagate laterally (that is, extensional, compressive and neutral UPS determine the increasing rupture potential length), it is possible to estimate the potential rupture area of a given segment in a [UPS; W] diagram (Figure 15). The smallest rupture areas should settled in the lower left corner of this diagram (short downdip and lateral rupture potential) and the largest ones would be in the upper right corner (large downdip and lateral rupture potential). The expected trend is verified in Figure 15; the extensional UPS and narrow seismogenic zones are associated with low Mmax, whereas the neutral UPS and large seismogenic zones are usually associated with the largest events. However, the largest magnitudes observed in several subduction zones were found to be smaller than what is suggested by their W and UPS values. Owing to the century over which these instrumental observations were made, we hypothesize that these subduction zones have not yet experienced the largest earthquake magnitude that they should be able to produce. The Mmax observed for subduction zones with similar W-UPS pairs could yield an estimate of their actual magnitude potential. For example, the Sandwich and southern Kermadec (recorded Mmax < 7.5) could produce at most Mw = 8.1 events. The Cocos Ridge subduction (recorded Mmax < 7.5) could produce Mw ≥ 8.0 events, and megaevents are likely to occur in the Antilles, Costa Rica and Manila segments (7.5 ≤ recorded Mmax < 8.0). A magnitude potential Mw = 9.0 could characterize Cascadia. A similar event did occur in the year 1700, and this region is currently on high alert for future earthquakes and tsunamis that are larger than those known from written and instrumental records (7.5 ≤ Mmax < 8.0 [Satake et al., 2003]).
4.3.2. Three-Dimensional Control of Megaearthquake Genesis
 Finally, it is appealing to speculate that a spatial preferential setting controls the occurrence of megaearthquakes (Mw ≥ 8.5). Theoretically, a differential trench migration is expected along the strike length; trench motion rates are inversely related with the slab width and decrease near the center of the slabs [e.g., Morra et al., 2006; Schellart et al., 2007; Guillaume et al., 2010]. As a first-order consequence, the local upper plate compressive strain is expected to progressively diminish toward the slab edges. This prediction is reasonably respected in natural cases where back-arc compressions are preferentially located at slab centers (e.g., Japan and South America) and extension is often reached at slab edges [Heuret and Lallemand, 2005; Heuret, 2005; Schellart, 2008]. In our view, such a scenario should tune the plate coupling and normal stress acting on the subduction fault, enhancing the occurrence of megaevents at slab edges. A similar background relationship may thus explain the observed frequency of large earthquakes near the borders of long trenches, such as in South America, Aleutian-Alaska, Sunda and NE Japan-Kuriles (Figure 13). Largest τ were instead observed toward slab centers, in agreement with Scholz and Small  observation of an inverse relationship between small earthquake seismicity and large earthquakes [Scholz and Small, 1997]. However, this inverse relation is only verified for segments where megaevents have occurred. The general law is rather a negligible but positive correlation between τ and Mmax (R = 0.25; Figure 3). In any case τ is of questionable use in predicting extreme events because megaevents are not part of the Gutenberg and Richter  distribution of the small earthquakes [Scholz, 2002].
 The three-dimensional control of megaevents genesis can be described for an idealized single wide (L > 5000 km) trench system where large normal compressive stresses associated with the compressive UPS of the slab center diminish toward the lateral slab edges, reaching neutral strain at one border and extensional strain at the other border. At the neutral edge, the moderate compressive stresses enable the nucleation and lateral propagation of large ruptures over a large number of asperities. The genesis of megaevents is possible. Such events release almost the entire available seismic moment. As a consequence, the plate interface has a large seismic coupling (vs ≈ vss), low seismic moment remains for additional events and seismicity is characterized by a relatively small number of large-sized events (e.g., southern Chile and eastern Alaska). Toward the slab center, normal compressive stresses at the plate interface increase, and relatively large ruptures are less able to propagate toward adjacent asperities. A large earthquake magnitude potential remains (Mmax 8.0–8.5) but the proportion of uncoupled areas increases. The seismicity is characterized by a relatively large number of moderately sized events and intermediate seismic coupling (e.g., Japan). Toward the extensional edge, normal compressive stresses become small. Asperities become increasingly smaller and more isolated. The proportion of uncoupled areas is at a maximum, but the large subduction velocities associated with back-arc spreading enables the few asperities to fail frequently. Large moment rates can occur, but the magnitude potential (Mmax ≤ 8.1) and seismic coupling (χ ≤ 0.2) values are low. The two latter features best characterize the seismicity of such regions (e.g., northern Tonga, New Hebrides and the Marianas).
 Statistics performed over worldwide subduction seismogenic zones reveal some of the major features of the seismogenesis of subduction plate interfaces. We confirmed that most seismogenic zones end in the fore-arc mantle rather than at the upper plate Moho depth. The subduction velocity was found to likely be the first-order controlling parameter of the variability in physical characteristics of the plate interfaces, determining both the geometry and the mechanical behavior. As such, fast subduction zones and cold slabs were found to be associated with large and steep plate interfaces, which, in turn, had large seismic rates. The earthquake magnitude potential and occurrence of Mw ≥ 8.5 events, in particular, determined both the level of the seismic moment that was released and whether the subduction was seismically coupled, regardless of the seismic rate. The subduction velocity was not able to account for the entire earthquake magnitude potential diversity observed along trenches. In fact, Mw ≥ 8.5 events preferentially occurred in the vicinity of slab edges, where the upper plate was continental and the back-arc strain was neutral. This observation was interpreted in terms of the level of compressive normal stresses along the plate interface. Large lateral ruptures associated with Mw ≥ 8.5 events should be promoted in neutral subduction zones due to moderate compressive stresses along the plate interface, which enable a rupture to easily propagate laterally.
along-strike length of seismic activity distribution (km).
downdip width of seismic activity distribution (km).
number of plate interface thrusts events.
analyzed time period (yr).
highest earthquake magnitude observed along the section.
moment released rate (N m per century and per 103 km of trench length).
seismic moment released by an individual earthquake (N m).
horizontal extent of the seismogenic zone width (km).
vertical extent of the seismogenic zone width (km).
mean arc-trench distance (km).
thermal parameter of the subducting plate measured at trench (km).
age of the subducting plate measured at trench (Myr).
fore-arc Moho depth (km).
depth difference between the fore-arc Moho and the downdip limit of the seismogenic zone (km).
 We thank Kelin Wang, an anonymous reviewer, and the editor for constructive suggestions which improved the quality of the manuscript. This research was supported as part of the Eurohorcs/ESF-European Young Investigators Awards Scheme (F.F.), by funds from the National Research Council of Italy and other National Funding Agencies participating in the 3rd Memorandum of Understanding, as well as from the EC Sixth Framework Programme. Discussion with Clint Conrad has been fundamental to strengthen ideas related to the importance of neutral regime in enhancing the triggering of Mw ≥ 8.5 earthquakes.