partitioning of crustal slip between linked, active faults in the eastern Qilian Shan, and evidence for a major seismic gap, the 'Tianzhu gap', on the western Haiyuan Fault, Gansu (China)

SUMMARY We have studied the Cenozoic and active tectonics of the north-eastern rim of Tibet west of the Yellow River (Gansu, China) where the western Haiyuan Fault enters the eastern Qilian Shan, a high mountainous region, which was the site of the 1927 May 23, M = 8-8.3, Gulang earthquake. Fieldwork, combined with analysis of aerial photographs and satellite images, reveals consistent cumulative left-lateral offsets of postglacial geomorphic features along the fault, but no recent rupture. West of the Tianzhu pull-apart basin, the levelling of offset-terrace risers implies Holocene horizontal and vertical slip rates on the steeply south-dipping, N11OE-striking fault of 11 f 4 and 1.3 f 0.3 mm yr-', respectively. The presence of subordinate, mostly normal, throws due to local changes in fault strike, and kinematic compatibility at the SW corner of the Tianzhu basin, constrains to about 120 km, consistent with the 95 f 15 km offset of the Yellow River ;ICI'OSS the Haiyuan Fault, but many times the offset (216 km) inferred on one rccent strand of that fault east of the river. Relative to the SE Gobi Desert, N E Tibet thus appears to have moved by a fair amount in the Late Cenozoic and is still moving fast. While some of this motion probably contributes to displace (towards the ESE) and rotate (CCW) the south-west edge of the Ordos block, much of it appears to be transmitted to the South China block, which leads, with the additional contribution of other large left-slip faults to the south and despite thrusting in the Lungmen Shan, to the extrusion (towards the ESE-SE) of that block relative to the Gobi, hcncc to north-eastern Asia. of of large its relatively high slip rate, one of the most critical sites for impending great earthquakes 7.5, probably M in the a here the 'Tianzhu gap', km north of Lanzhou and Xining, largest population of west-central instrumental of that fault particularly That the M 8, earthquake ruptured a complex thrust surface under high mountains in a restraining of the strike-slip fault suggests that the occurrence of comparable earthquakes in other areas with similar fault geometry, such as south of the big bend of the San Andreas Fault in California, not be

to about 120 km, consistent with the 95 f 15 km offset of the Yellow River ;ICI'OSS the Haiyuan Fault, but many times the offset (216 km) inferred on one rccent strand of that fault east of the river. Relative to the SE Gobi Desert, N E Tibet thus appears to have moved by a fair amount in the Late Cenozoic and is still moving fast. While some of this motion probably contributes to displace (towards the ESE) and rotate (CCW) the south-west edge of the Ordos block, much of it appears to be transmitted to the South China block, which leads, with the additional contribution of other large left-slip faults to the south and despite thrusting in the Lungmen Shan, to the extrusion (towards the ESE-SE) of that block relative to the Gobi, hcncc to north-eastern Asia.
The =260 km long western Haiyuan Fault links two faults that ruptured about 70 years ago during two great earthquakes only seven years apart. Despite spectacular evidence of Holocene movement, it bears no trace of a large earthquake in the past eight centuries, either in the field or in the historical record. Given its relatively high slip rate, it should therefore be singled out as one of the most critical sites for impending great earthquakes (at least M I 7.5, probably M 2 8) in the region. That such a seismic gap, called here the 'Tianzhu gap', lies only =lo0 km north of Lanzhou and Xining, largest population centres of west-central China, makes instrumental monitoring of that fault particularly urgent. That the M = 8, Gulang earthquake ruptured a complex thrust surface under high mountains in a restraining bend of the Haiyuan strike-slip fault suggests that the occurrence of comparable earthquakes in other areas with similar fault geometry, such as south of the big bend of the San Andreas Fault in California, should not be ruled out.

INTRODUCTION
The active, northern tectonic boundary of the Tibetan Plateau is composed of two parts. West of 94"E it is marked by the 1600 km long, -N70E-striking, left-lateral Altyn Tagh Fault, a narrow strike-slip zone that guides the north-eastward extrusion of the Plateau (Tapponnier & Molnar 1977;Armijo, Tapponnier & Han 1989). East of 94"E and north of the Qaidam Basin, that boundary is broader and more complex, with several splays of the Altyn Tagh Fault accommodating fractions of the movement of Tibet relative to the Gobi-Ala Shan Platform (Peltzer et al. 1988;Peltzer, Tapponnier & Armijo 1989) (Fig. 1). Those splay faults, which combine thrust with strike-slip movement, are responsible for the remarkable structure of the northeastern border of the Tibet-Qinghai highlands, characterized by many parallel, NW-SE-trending ranges locally higher than 5500 m, including the Qilian Shan. Such oblique thrusts may merge at mid-lower crustal depth into a shallow south-dipping dicollement, the north-eastern edge of Tibet thus forming a large crustal-stacking wedge (Mattauer 1986), with the detached underlying lithospheric mantle of the Gobi-Ala Shan Platform subducting 2500 km farther south, beneath the Kunlun (Tapponnier et al. 1990;Meyer 1991).
West of 101"E, widespread oblique thrusting changes to more localized crustal strain along the Haiyuan strike-slip zone (Fig. 1, inset), a predominantly left-lateral fault that runs NllO f 10E, roughly parallel to, and =70 km south of the frontal thrust ramps bounding the north-easternmost ranges of the Qilian Shan (Fig. 1). While the Tibetan highlands terminate around 103"E, the linear Haiyuan Fault zone continues east of the Huang He (=104"E), then veers to a N140-160E strike along the active Liupan Shan thrusts (e.g. Tapponnier Zhang et al. 1991), before merging with the Tianshui and Fengxian Faults near Baoji (=107"E) and resuming a N100-llOE strike along the northern boundary of the Qinling Shan, presently also a left-lateral fault, but with a normal component of slip (Tapponnier & Molnar 1977;Peltzer e l al. 1985;Bellier et al. 1988).
In order to understand the present-day deformation and recent tectonic evolution of north-eastern Tibet and of adjacent regions, it is important to determine quantitatively how the oblique movements that take place within the Qilian Shan relate to slip on the Haiyuan Fault zone and on faults farther east and north. This requires a precise analysis of the regional kinematics in terms of slip rates, slip-vector orientations, and cumulative displacements along the faults. As a step towards that goal, we studied the Haiyuan Fault zone between 102" and 104"E, west of the Yellow River (Huang He), as well as the region to the north, between Tianzhu and Wuwei, site of the great 1927 May 23, M = 8.3, Gulang earthquake (Fig. 1). In contrast to the wellstudied segment of the Haiyuan Fault that ruptured during the greater, 1920 December 16, M =8.7, Haiyuan earthquake, east of the Yellow River (e.g. Deng et al. 1984Deng et al. , 1986Zhang et al. 1987Zhang et al. , 1988aLiu & Zhou 1986;State Seismological Bureau 1990), little is known about active faulting in regions located west of the river. The field observations we gathered during short traverses of those regions were supplemented by structural and morphological studies of satellite images (Landsat MSS and SPOT) and air photographs, and by analysis of the topography and geology on maps at scales of 1/100,000 and 1/1,000,000. Using a digital theodolite distancemeter, we mapped horizontal displacements and levelled vertical offsets of postglacial morphological features at two sites along the western Haiyuan Fault. Although restricted by military interdiction and snowy weather, our work brings first-order information on Holocene and Cenozoic movement along the principal regional faults, on the mechanism of the Gulang earthquake, on fault architecture within the crust, and on potential seismic hazard in the area located north of Lanzhou and west of the Yellow River.

First-order segmentation of the fault
West of 104"E, the western Haiyuan Fault runs along ranges whose maximum elevations increase westwards from 23300 to =5200m (Fig. 1). The main strand of the fault is composed of two fairly linear segments, separated near the small city of Tianzhu by a = 7 km wide left step (Figs 1 and 2a). Owing to left-lateral motion on the fault, that step has resulted in the formation of a half-graben, bounded to the west by a prominent, mostly N30-80E-striking normal fault that connects the two linear segments. This half-graben, here named the Tianzhu pull-apart basin, interrupts the continuity of the ranges that follow the fault, thus providing fairly easy rail and road access to the Hexi corridor, into which it drains.
West of that pull-apart, the western segment, about 140 km long, first follows the north side of the Zhuanglang He Valley. Then, after merging with the Gulang Fault, it forms a narrow strike-slip rift across high, glaciated landforms along the Lenglong Ling (Dragon-of-Cold Range). East of the pull-apart, the eastern segment extends along the Maomao Shan and Lao Hu Shan (Old Tiger Mountain) to south of Jingtai, a distance of ~1 2 0 km ( Fig.   1). There, a = 5 km-wide left-step and pull-apart jog marks the separation between the western and the eastern Haiyuan Fault, which lies mostly east of the Yellow River. The =230 km long break of the 1920 earthquake on the latter fault stopped at that jog, about 20 km west of the river, near Santang Shan (Deng et ul. 1984(Deng et ul. , 1986State Seismological Bureau 1990;Liu & Zhou 1986).

Zhuanglang He-Lenglong Ling (western) segment
Near 101.7"E, as it enters the highest part of the Lenglong Ling, the western segment of the fault cuts three extinct, north and east-draining glacial valleys (Fig. 2b). The valley floors are fairly flat, several hundred metres wide, often with lateral moraines, and in part occupied by hummocky glacial till. Much of the area near and south of the fault, whose trace is straight and sharp, stands above 4000ma.s.l. The tips of three hanging glacier tongues lie only 1-1.5 km away from that trace. The bedrock topography peaks along thin crestlines (nunataks), implying thick ice-caps, during the last (Wiirm) glaciation, on much of the mountain relief visible on Fig. 2(b). The two most prominent mountain spurs between the glacial valleys are truncated by the fault, forming north-facing triangular facets. North of the fault, smooth, flat surfaces are preserved over a fairly large area between the glacial valleys. We interpret these surfaces, which have not been significantly incised by pluvial run off, to have been subject to efficient planation by periglacial erosion processes until a very recent epoch. The smooth texture of the surfaces also implies that they are mantled by loess. The edges of the glacial valleys and of the intervening ridges, which are in part outlined by lateral moraines, are clearly offset by the Haiyuan Fault, in a left-lateral sense (Fig. 2b). Such offsets range between 200 and 270m. The fault also offsets, by smaller amounts ( 4 5 and =loom), two stream channels that incise the floor of the westernmost valley (Fig. 2b).
The relationship between active faulting and glacial geomorphology derived from the interpretation of the SPOT image in the Lenglong Ling bears strong resemblance to that documented at several other Tibetan sites by Armijo et al. (1986Armijo et al. ( , 1989. The freshness and size of the glacial landforms imply that they date mostly from the last glacial maximum (about 20kaB.P.), a time at which glacier tongues fed by the ice capping the relief south of the fault may have extended several km northwards. The geomorphic offsets visible on the image thus probably postdate that maximum. More specifically, we interpret the 240 f 40 m offsets of the glacial valleys edges to postdate ice retreat above the fault trace, some time between 20 and 14 ka B.P., and the smaller stream offsets on the valley floor to postdate the pluvial period that marked the Early Holocene climatic optimum. Such inferences would be consistent with a left-slip rate of 15 f 5mmyr-I on the western Haiyuan Fault in the Lenglong Ling (Blusson, Tapponnier & Meyer 1989;Meyer 1991, p. 114, Fig. 10a).
Between 102"25'E and 102"45'E, west of the Tianzhu pull apart, the upper Zhuanglang river flows in an elongated basin with a flattish floor, at an elevation of -3100m (Fig.  2a). Bounded by steep relief on either side, that basin is filled by Plio-Quaternary clastics. To the north, the Lenglong Ling mountains (locally Leigong Shan) rise to ~4 3 0 0 m, .exposing mostly Silurian greenschists intruded by diorites (Gansu Geological Bureau 1975;Xu, He & Yan 1989). Well-preserved traces of the last glaciation are clear above 3700m on the south flank of the mountains, in the  form of fresh, hanging glacial cirques (Fig. 3a). The active Lenglong Ling. South of the fault, the fans coalesce to form Haiyuan Fault trace stands out along the north edge of the a rather uniform, gently SW-sloping piedmont. Rare, basin, near the 3500m elevation contour (Figs 2a, 3 and 4).
isolated hills or ridges of hummocky ground that stand The fault runs mostly across the apex of fans (Fig. 4b) and somewhat above the average piedmont surface may boulder-bearing fluvio-glacial terraces (Fig. 4b)   slices through bedrock, steeply south-dipping shear planes and cracks form a gouge zone several metres thick. The principal strand of the fault trace is outlined in most places by a south-facing, up to 30 metres high (Figs 4a and b), cumulative scarp with a rather steep slope (up to 35"), which makes it clear from afar in the landscape.
That scarp strikes mostly N110-120E but swings to a =N80E direction for =3 km near 102'40'E ( Fig. 2a). Bends in the Zhuanglang He course and in the Lenglong Ling crest, which forms a saddle at this longitude, mimic that swing south of trig point 4022 (Fig. 2a). The strike component of slip on the fault is documented by left-lateral  offsets of stream channels and terrace risers (Figs 4b and c), and by the presence of small dilational jogs, marked by penistent sag ponds or swampy peat bogs (Fig. 4a). The dip component of slip appears to result from the difference between the azimuths of the fault trace and of the regional fault-slip vector, as commonly observed along great strike-slip faults, and quantitatively documented on the Changma fault in the western Qilian Shan (Peltzer et al. 1988). Here in the Zhuanglang basin, fault segments that strike more northerly than ~N 1 1 5 E bear the clean-cut, steep morphology of normal fault scarps while rare segments striking more southerly than -N120E show the smoother relief and lobate geometry that typifies thrust-fault scarps (Figs 2a and 3).  (Fig. 4d). Big boulders within the streambed still outline a 1.7 f 0.5 m high step across which the water cascades (Fig. 4d). Using a digital theodolitedistancemeter (Wild T2000-D120), we levelled six profiles (S1 to S6) to document quantitatively the large-scale morphology of that site (Fig. 4e). S1 and S6 were levelled perpendicular to the fault in order to measure cumulative vertical offsets; S2 and S5, along the terrace-riser tops, to measure cumulative horizontal offsets. Two additional profiles (S3 and S4) were levelled along the bottom and top of the cumulative, N11OE-striking fault scarp. Several sources of uncertainty make measurement of offsets of morphological features such as terrace risers difficult. Given the maximum distance (typically of the order of a few km) between the station and measurement points, the high precision of the station instruments ( 2 5 X lop6 rad for angles, and 2 3 mm + 1 mm km-I for horizontal distance) makes instrumental errors (of the order of a few mm) negligible in view of the dimensions and offsets (typically of the order of several tens of metres) of the features studied (e.g. Peltzer et al. 1988). Another source of error arises from the fact that the top edge of a terrace riser may be locally difficult to identify because of degradation. At Sange Dun, however, the top edges of the risers are outlined by fairly sharp slope breaks, and we estimate in the field the error on the horizontal position of points measured along those edges to be of the order of *0.5 m.
By far the most serious source of error comes from difficulties in defining the large-scale shapes of the risers and from changes of those shapes with increasing fault offset. Fig. 5(a) illustrates how such difficulties may lead to fairly large uncertainties on horizontal offsets. In map view, the cumulative horizontal offset of a terrace riser is the distance, measured along the fault, that separates the intersections (piercing points) of the fault scarp with straight lines approximating the terrace-riser tops upstream and downstream from the scarp. If entrenchment of the stream has been rapid and if the riser top edge is sharp, with similar, fairly straight strikes above and below the fault scarp, defining the piercing points (open circles, Fig. 5a, left) and measuring the offset is straightforward. Fairly long profiles, extending far enough upstream and downstream from the fault scarp, are nevertheless required. If the length of the profiles does not suffice to define straight-and long-enough riser segments with matching strikes, the offset will be fraught with uncertainty. Given the sigmoidal bend that morphological markers transverse to strike-slip faults usually exhibit, matching the upstream and downstream segments of risers defined by short profiles will in general lead to apparent offsets smaller that the real offset (open triangles, Fig. 5a, middle). Finally, if the formation of a terrace riser results from slow, progressive entrenchment of the stream into previously deposited fanglomerates, erosion by the stream may make the riser edge recede away. Differential erosion and deposition may also change the riser shape in map view. At Sange Dun for instance, riser edges above the fault and west of the stream might have receded westwards, with carving and depositional smoothing by the stream leading to a greater sigmoidal deviation of the risers across the fault and to estimation of an apparent offset greater than the real offset (open squares, Fig. 5a, right). Clearly, the uncertainties on horizontal offsets resulting from the ill-defined large-scale geometry of the risers dwarf all others. They may reach tens of metres, amounting to a significant fraction of the offset itself (Fig. 5a).
Similar problems are encountered when determining the vertical throw of offset surfaces. The throw is the elevation difference between the two points where the fault plane intersects the regional slopes above and below the scarp (Fig. 5b). Neglecting millimetric instrumental errors, three factors contribute to the uncertainty on the throw. The first one comes from fitting regional slopes to straight lines. We levelled profiles as long as possible (several hundred metres), smoothing out obvious, small-scale surface irregularities, such as minor gullies or large boulders. For profiles S1 and S6 at Sange Dun and for five of the six profiles levelled west of Tianzhu, straight lines provided excellent fits to measured profile points for =200m above the scarps and for greater lengths below, thus constraining well the regional slopes. We estimate the uncertainty in those fits to be of the order of the scatter due to surface roughness, i.e. k0.5 m. The second factor is the uncertainty in the position of the fault within the scarp slope (Fig. 5b). In general (e.g. Bucknam & Anderson 1979;Hanks et al. 1984), scarps that do not integrate the effect of multiple, close-spaced faults exhibit an inflection point near mid-width that separates a region of erosion (above) from a region of deposition (colluvial wedge, below) (e.g. Wallace 1977;Armijo et al. 1986). Here, we assume that the fault plane is unique and intersects the scarp slope near that point, and allow the scarp-fault intersection to vary between the levelling points that bound the inflection (filled circles, Fig.  5b, top). The third factor is the uncertainty on the dip of the fault plane, shallower dips leading to greater throws (Fig.  5b). Since the fault is normal and strike slip, we take its dip to be greater than 45". Given such constraints, the combined uncertainty on the calculated values of vertical throws comes out to less than 20 per cent at Sange Dun and to generally less than 10 per cent at Tianzhu. At Sange Dun, the intermediate terrace riser has a dog-leg shape, with a linear upstream segment striking N18E for ~2 0 0 m (AB, Fig. 6a) and a slightly curved, downstream segment that strikes N30E on average from C to D (Fig. 6a). These two segments are linked, across the fault scarp, by a transition segment, BC, striking = N150E. Assuming the initial strike of the downstream riser segment to have been close to that now characterizing the upstream segment (N18E) would yield a cumulative left-lateral offset of 31.5 f 1.0m (open triangles, Fig. 6a, Table la). Taking the average strike (N30E) of the longest, best-defined downstream riser segment to be the original strike would lead to a larger offset: 37.5 f 1.0 m (open squares, Fig. 6a, Table la). The distance, along the NlOOE fault strike, between the points that define the top and base of the fault scarp on profile S2 (B and C, Figs 6a and c) is 33.5 f 1.0 m (Table la). Finally, projecting profiles S3 and S4 on a vertical plane striking NllOE, parallel to the fault, yields a left-lateral offset of 37.4 5 1.0 m between the riser tops at the base and top of the scarp (horizontal distance between P and M, Fig. 6b, Table la). We take the arithmetic mean of these four values (35.0m), and the greatest difference between that mean and any such value (3.5m), to be the most plausible value of, and uncertainty on, the horizontal offset of the intermediate terrace riser (Table la).
The more sinuous geometry of the outer terrace riser makes determination of its horizontal offset more difficult (Fig. 6a). Though curved and poorly defined, the segment of that riser that is farthest upstream, beyond E (Fig. 6a), strikes about N35E. Downwards, from E to F and below, the riser is more linear and takes a more northerly direction (NME), before veering counter-clockwise across the fault   (Table  la). As above, we assume the arithmetic mean of these three values (88.9 m), and the maximum difference between any of them and that mean (13.6m), to correspond to the most plausible value of, and uncertainty on, the horizontal offset of the outer terrace riser (Table la).
Projecting S1 on a vertical plane striking N20E, perpendicular to the fault, shows that the intermediate terrace surface slopes =lO"SW, both above and below the scarp (Fig. 6c). The scarp itself slopes 35"SW, a value close to the angle of repose of loosely consolidated alluvium (e.g. Hanks et al. 1984). The cumulative vertical throw on the fault is thus simply the vertical offset between the upper and lower segments of the profile (13.9 f 2.4 m, Fig. 6c, Table  la). Measured on the projection of the riser profile (S2) on a N2OE-striking vertical plane, that throw comes out to be __ smaller (11.7 f 1.7 m, Fig. 6c). This latter value is close to that deduced from projecting S3 and S4 on a vertical plane parallel to the fault (11.3 f 1.0 m, elevation difference between P and M , Fig. 6b, Table la).
On profile S6, the outer terrace has different, but well-constrained, slopes above (16"SW) and below ( 10°SW) the scarp. The corresponding cumulative vertical throw is 26.4 f 4.1 m (Fig. 6c, Table la). As for the intermediate terrace, that throw is smaller (17.8 f 4.3 m, Table la) when measured on the projection of the outer riser profile ( S 5 ) on a N20E-striking vertical plane (Fig. 6c). This latter value is poorly defined, however, because that riser has suffered significant slope and elevation changes since formation. The tell-tale sign that such changes have occurred is the less distinct, much smoother shape of the cumulative fault scarp on S5 than on S6 (Fig. 6c). Close inspection shows that a small stream has incised across the riser near the scarp top and that another one, flowing near the riser base above the fault, has degraded its slope, thus lowering the elevation of Downloaded from https://academic.oup.com/gji/article-abstract/120/3/599/776877 by guest on 29 July 2018 Assuming the age of a terrace riser to be the age of the terrace at its base (e.g. Weldon & Sieh 1985;Bull 1991;Avouac et al. 1993), the cumulative horizontal offset of the outer-terrace riser relates to the cumulative vertical throw of the intermediate terrace surface. For this terrace, tan p = 0.03 (Fig. 6b). With a cumulative left-lateral offset of 4 8 . 9 m, the vertical-offset correction due to strike-slip motion thus amounts to =2.7m, bringing the real vertical throw of that terrace on S1 to 11.2 * 2.8 m (Table la), much closer to the offset of its riser top (11.3-11.7m, Table la). We take the arithmetic mean of the three values (11.4m), and the average error (1.4 m) on the two values that do not depend on slope correction to be the most plausible value of, and uncertainty on, the vertical offset of the intermediate terrace (Table la).
For the outer terrace, tan / 3 == 0.06 (Fig. 6c), but the cumulative horizontal offset is unknown. It seems reasonable, however, to assume that the attitude of the slip vector has not changed significantly between the emplacement of the outer and intermediate terraces. Therefore, the ratio between vertical throw u and horizontal displacement of 18.0 f 4.5 m (Table la). This value is also much closer to the throw of the riser top on S5 (17.8f4.3m, Table la). The mean of these two values is 17.9~k4.4m (Table la). With r = 0.13 f 0.04, such a vertical throw corresponds to a left-lateral offset of 139.8 m, between extreme values of 79.9 and 227.6 m (-60/+90 m, Table la).
The values of the left-lateral offsets and vertical throws of the three inset terraces, five of which derive from independent measurements, are summarized on the plot of  Table la).
The values, listed in Table l(a), of the three horizontal offsets measured in the field at Sange Dun, are not unique to that site. Table l(b) shows, for comparison, the left-lateral offsets of streams and risers in five catchments that cross the Zhuanglang He segment of the fault east and west of Sange Dun, over a distance of ~4 . 5 km (Fig. 3). Such offsets, which are derived from inspection of an enlarged print of the aerial photo in Fig. 3(a), are only indicative, since they are not corrected for optical distortion. Nevertheless, most of them appear to cluster between 70 f 15 and 105 f 15 m, consistent with the 88.9f 13.6m offset at Sange Dun. The overall range of offsets (from 26* 15 to 140It 15m) is also consistent with that seen at Sange Dun (Table 1, (Table 1).
To our knowledge, none of the terraces a t S m g : h i ! I , 1, been directly dated. Given the high elevation of thlc (3400-3500 m) at 37"25'N, the fresh morphology ol th~: high-level terraces, and the steep slopes o f the proinin. i i i cumulative fault scarps that cut them, however. there call h, little doubt that both the terraces and scarps postih:c rhc last glacial maximum. Moreover, the rather large size 0 1 the terraces, and the fact that their material contains ahundiini reworked glacial boulders, suggests that the emplacement of' that material, possibly in part by debris flows. was related to deglaciation, starting with melting of the ice that filled the now ice-free glacial cirques hanging right above thc \ite. followed by increase in pluvial run off (e.g. Gasse t t ol 1991). We infer that such deglaciation occurred roughly in tune with the rapid global warming that marked the onset of the present interglacial, causing the rapid rise of sea-level between 14 and 11.5 kaB.P. (Fairbanks 1989;Bard rr ul. 1991), and culminating in the Early Holocene climatic optimum. Melting of the ice on Lenglong Ling, followed by greater rainfall, would have temporarily increased the water discharge, inducing fast and powerful flush by the streams and deposition of coarse, boulder-rich debris near the foot of the range, at the edge of the Zhuanglang He basin. Lowering of the base level in the Zhuanglang He basin, in part due to normal throw on the Haiyuan Fault, would have led to stream entrenchment and abandonment of the high-level terraces, particularly upstream from the fault. 'The emplacement of several prominent postglacial terraces and fans along other high mountain piedmonts of western China appears to have followed similar patterns, orchestrated by climate change (e.g. Peltzer et a/. 1989).
Assuming the oldest, outer terrace at Sange Dun to have been emplaced at 13.5 & 2 ka B.P. would imply average horizontal and vertical slip rates of 11 & 4 m m y r -' and 1.3 f 0.5 mm yr-', respectively (see Table 3 later), on the western Haiyuan Fault. If such rates had been constant, the age of the intermediate and inner terrace surfaces would be 8.4 f 4 ka B.P. and 3.0 * 1.5 ka B.P., respectively.
While the evidence we have at hand is insufficient to constrain with accuracy the parameters of individual seismic events on the Zhuanglang He segment of the western Haiyuan Fault, it is suggestive of rupture during infrequent, great earthquakes. That the 4 m high scarp across the inner terrace has been smoothed by erosion and cannot be traced within the high cumulative scarps of the intermediate and outer terraces (Figs 4c and d) suggests that the last rupture of that fault segment is at least several centuries old. This inference is corroborated by the lack of large ( M 2 7 ) recorded historical events in the region between the ancient cities of Lanzhou and Wuwei ( Fig. 1) since 1125 A.D., 869 years ago. If the fault at Sange Dun were to break only during similar great earthquakes with return periods of the order of 800 years, the cumulative horizontal offset and throw of the inner terrace could result from four such earthquakes, each with a left-lateral slip of the order of 8.8m. If the horizontal offset of the inner terrace were due to only three great earthquakes, then such earthquakes would be characterized by individual slip amounts of ~1 1 . 7 m and would recur every ~1 0 6 0 years. In either case, the slip amount per event would imply a magnitude of at least 8. While such slip values are common, it seems fan heads, causing prominent changes in scarp height along the fault (Figs 8a and b). Unfortunately, although the fault strikes between N30E and N70E across the site (Figs 8b and c), we found no clear evidence of left-lateral movement along the more easterly striking parts of the scarp. In contrast with the differentially uplifted surfaces on the footwall of the fault, the hanging-wall surface is mantled by a rather uniform fanglomerate cover, implying that deposition prevailed, smoothing out most terrain irregularities, except for a few risers bounding flood channels of present-day streams. We infer such widespread deposition to have hidden evidence for cumulative horizontal offsets. Besides, sustained accumulation of fanglomerates upon the hanging wall makes all cumulative vertical offsets measured at that site lower bounds of the actual offsets.
Six profiles (T1 to T6) were levelled roughly perpendicular to the fault scarp (Figs 8c and 9). The uniformity of the hanging-wall surface, which slopes 6"-7" SW for hundreds of metres on all profiles ( Fig. 9, Table la) is quite clear. But the footwall segments of the profiles, even on top of the highest bedrock spurs, also have fairly regular slopes, steeper by only 4-5" (up to 10-12", Fig. 9, Table la). This implies, in particular, that the bedrock surface on top of the truncated spurs has been abraded and flattened by efficient surface-shaping processes, probably of glacial or periglacial nature. Besides having similar slopes, the hanging-wall segments of profiles T2, T3 and T5 have absolute elevations within -2m of each other. On the other hand, the hanging-wall segment of T6 stands as much as ~1 4 . 5 m above that of T1. Thus, between those profiles, that are farthest apart ( 4 0 0 m, Fig. Sc), the hanging-wall surface slopes gently, by about 2"NE, towards the northern outlet of the Tianzhu Basin (Fig. 7).
Though separated by a small, hanging-stream valley, the two highest spurs to the east stand at similar elevations (Fig.  8a). The maximum scarp slopes on the corresponding profiles (T1 and T2), 37" and 30", are comparable to those at Sange Dun ( Fig. 9, Table 2). Here, however, bedrock, which maintains steeper slopes for longer times than alluvium does (e.g. Wallace 1978), forms the top of the cumulative scarp. The scarp heights on T1 and T2, 45.5 f 2.4 m and 41.2 f 2.5 m respectively, yield the maximum throw visible at the site (43.4*4.6m, Table 2). The footwall segment of profile T4, on the spur separating the two fans (Figs 8b and 9), lies on top of the uppermost terrace visible at the site. Although that terrace is chiefly made of fanglomerates containing big boulders, such fanglomerates have filled a previously more irregular bedrock topography. Bedrock asperites, such as that forming the scarp top on T4, thus subsist within the terrace. This accounts for the particularly steep slope of that scarp (41", Fig. 9, Table 2). The visible height of the cumulative scarp of the uppermost terrace on T4 is 28.2 f 1.8 m (Fig. 9, Table 2). The footwall segments of profiles T3, T5 and T6 all hug the top surface of the most prominent, intermediate terrace on either side of the T4 spur (Fig. 8b). The cumulative scarp heights on those profiles range from 17.0 f 2 m to 21.1 f 1.5 m, with a mean value of 19.5 f 3.0 m ( Fig. 9, Table 2). None of the six profiles reveals the distinctive scar of a recent seismic dislocation within the high cumulative scarps, whether in the form of a steeper freeface or of smoothed out, but visible, slope breaks. Thus, less likely that the inner-terrace offset could be the product of only two earthquakes, for such events would have caused =17[5m of horizontal slip each, an amount -25 per cent greater than the largest known in any strike-slip earthquake worldwide. The magnitude, slip amounts and recurrence times we infer to be most plausible for large seismic events on the Zhuanglang He segment of the western Haiyuan Fault are thus roughly comparable to those of a Haiyuan-type event on the eastern Haiyuan Fault (Zhang et af. 1987(Zhang et af. , 1988a.

Tianzhu pull-apart basin
Five kilometres east of Sange Dun, the main strand of the western Haiyuan Fault takes a sharp bend to the north-east, forming an outstanding, -10 km long, cumulative scarp bounding the eastern side of the Tianzhu Basin (Figs 1, 2a and 7). That scarp, whose strike swings between NlOE and N90E, truncates mountain spurs and marks the steep base of large triangular facets with hanging 'wine-glass' valleys, attesting to recent and fast norma! uplift (Figs 7 and 8). The height of the triangular facets increases towards the north-east and the visible throw of the cumulative scarp commonly reaches tens of metres. In addition, a few subsegments of the Tianzhu Fault that strike N60 to 90E offset streams within active fan-heads in a left-lateral sense, on aerial photographs (Fig. 7). Near 102"50'E, the Haiyuan Fault bends back to a N125E, then NllOE strike (Fig. 2a). In contrast with that west of Tianzhu, the recent vertical offset of the topography remains moderate for several kilometres east of that bend. Left-lateral offsets, on the other hand, are clear on SPOT images and aerial photographs. The southern side of the Tianzhu Basin appears to be bounded by a =NllOE-striking fault, in rough continuity with the Zhuanglang He segment of the Haiyuan Fault, but with much more subdued morphological evidence of Late Quaternary movement. Similarly, there is little evidence of prominent Late Quaternary faulting along the eastern side of the Tianzhu basin.
The visible fill of the basin is composed of Quaternary fanglomerates that have been fed mostly from the south and west, and that form three main sets of imbricated surfaces. In the east-central part of the basin, the highest, most ancient, fan surface has been incised tens of metres by streams that drain the basin northwards (Figs 7a and b). In the western part of the basin, small active fans built by streams incising the faulted rangefront lap onto the surface that caps much of the subsiding hanging wall of the Tianzhu Fault (Fig. 7). Save for these fans, and despite the fairly high average elevation (2900 m), the entire basin is cultivated, due to the presence of a top-soil layer of loess. Although geologic maps available to us here provide insufficient detail (Gansu Geological Bureau 1975), it is possible that the oldest fanglomerates visible in the basin postdate the last glacial maximum.
For a first-order quantitative assessment of recent movement on the Tianzhu Fault, we selected one site in the south-west comer of the basin (Figs 7 and 8). At that site, the fault cuts and offsets the heads of two alluvial fan complexes bounded by higher, truncated spurs carved into the Lower Palaeozoic bedrock. Deposition and incision by rivers in the footwall has produced inset terraces within the as at Sange Dun, the last earthquake to have ruptured the fluvioglacial terrace levels have been significantly uplifted by west Tianzhu normal fault must be at least several centuries movement on the fault. It is therefore likely that the old.
uppermost and intermediate terrace levels at the south-west As at Sange Dun, none of the main surfaces that marks corner of the Tianzhu Basin are coeval with the the footwall of the Tianzhu normal fault has been directly corresponding levels at Sange Dun, only 2 6 km to the west, dated. As observed at that first site, however, only two main and result from the same climatic forcing. In keeping with   this interpretation, we assume those two terraces to have ages of 13.5 f 2 and 8.4 f 4 ka B.P., respectively (Table 3).
We also assume the well-planed surface of the highest spurs, which despite its flatness and uniform slope is not covered by alluvium, to have been smoothed by cryoplanation and gelifluction (e.g. Bloom 1978;Chorley, Schumm & Sugden 1984;Peltzer et al. 1988) till the end of the last glacial maximum. Thus, we infer the age of that surface to be 18 f 2 ka B.P. (Table 3). Using such ages, the minimum throw rates derived from the visible offsets of the three surfaces are 2.4f0.5, 2.1 f 0.5 (mean value: 2.3 f 0.7 mm yr-') and 2.3 f lSmmyr-', respectively (Table 3). That such rates are within a few tenths of a millimetre per year of one another is consistent with the requirement that the slip rate on a segment of the Tianzhu normal fault only 0.5 km long be constant, and supports the inferences made on the ages of the surfaces. Though a lower bound, a throw rate of 2.3*0.7mmyr-', about 80 per cent faster than that at Sange Dun (Table 3), is expected from the more abrupt mountain front and more northerly direction of the fault west of Tianzhu.
Lacking information on the postglacial sediment thickness in the Tianzhu Basin, we can only hypothesize what the actual, total throw rate on the Tianzhu Fault might be. Plausible inferences can be made, however, by assuming the rate of deposition to have been roughly the same anywhere along the base of the scarp at any one time since the last glacial maximum (218 ka B.P.). Taking the total vertical throw-rate u, on the Tianzhu Fault to have been constant, and us,, us,, us3 to be constant deposition rates between 18 and 13.5 ka B.P. (from the glacial maximum to the onset of deglaciation), 13.5 and 6kaB.P. (from the onset of deglaciation to the end of early Holocene climatic optimum, Gasse et al. 1991), and between 6 ka B.P. and the present, respectively, the throws and ages of the surfaces require such rates to be related by the following linear system: 11.0*4.0mm/yr Fixing the value of usl determines uniquely the three other rates. If the sedimentation rate before the onset of deglaciation had been negligible (usl -0), then the total throw rate on the fault would be =3.4mmyr-', with us3 = 0.8 mm yr-' and us2 = 1.7 mrn yr-'. Assuming usl = 1 mm yr-', u, would become -4.4 mm yr-', and subsequent deposition rates uS2 = 2.7 and us3 = 1.8 mm yr-', respectively. In the latter case, the corresponding total cumulative throw and sediment thickness deposited since ~1 3 . 5 ka B.P.
High spur surface 2.4 f 0.5 m / y r 118 2 2 ka) Uppermost terrace 2.1 f 0.5 mm/yr (13.5 f 2ka) Intermediate terrace 2.3? 1.5 mm/yr (8.4 C 4 ka) would be -59.4 and 31.2 m. Hence, a little over half the total throw would be hidden by sediment.
Although such calculations, which are non-unique, are given as mere examples, they provide reasonable values. The second one, in particular, balances the visible relief of the scarp with the sediment thickness at its base and implies plausible ratios between deposition rates since the last glacial maximum while yielding a throw rate fast enough to be consistent with the spectacular morphology of the mountain front. That the throw rate on the Tianzhu normal fault should be faster than is typical of isolated normal faults (1-2 rnm yr-') is to be expected because that fault connects two relatively fast-slipping strike-slip fault segments. In southern Tibet for instance, Armijo et al. (1986) documented particularly fast throw rates ( 2 6 mm yr-') on normal faults linked with the Beng Co right-lateral strike-slip fault, whose horizontal slip rate is of the order of =15 mm yr-' (Armijo et al. 1989).

Kinematic compatibility between the Zhuanglang He and Tianzhu Faults
The prominent Holocene throw along the Tianzhu normal fault suggests that it takes up most of the movement on the Zhuanglang He Fault (Figs 7 and 10a). It is therefore worth exploring to what extent motion on those faults is compatible if one assumes rigid-block kinematics around the -45" bend between them, and what constraints our measurements at Sange Dun and near the SW corner of the Tianzhu Basin then place on fault dips and on the slip vector.
On average, the two faults strike NllOE and N65E, respectively, at the sites studied. Near the surface, the dip of the Tianzhu Fault must be greater than the maximum slope of its steepest cumulative scarp (41"). We allow it to vary between 45" and 90" (bold great circles, Fig. lob). In addition, the existence of left-lateral offsets on N60-90Estriking segments of the Tianzhu Fault east of our levelling site (Fig. 7) implies a left-lateral component of slip at that site too, even though we could not quantify it.
If strain within crustal blocks north and south of the faults were negligible, the slip vector on both fault planes would lie along their intersection (e.g. Armijo et al. 1986). Left-slip and a dipZ45" on the Tianzhu fault would thus constrain the azimuth of the slip vector common to both faults to range between NllOE and N155E, and the dip of the Zhuanglang He Fault to be at least 55" (light-shaded sector and bold great circle, respectively, on  Although both results are in qualitative agreement with the observation that the Zhuanglang He and Tianzhu Faults appear to be predominantly strike slip and prominently normal, respectively, they draw attention to a significant shortcoming of the rigid-kinematics assumption. From the diagram of Fig. 10(b), the plunge of the intersection between the fault planes would have to be steeper than =34", which cannot be reconciled with the steepest slip-vector plunge (-7") permitted by the u / h ratio of 0.125 at Sange Dun. That a common slip vector cannot be found along the intersection of the two planes is also clear from comparing the throw estimated from the height of the Tianzhu cumulative scarp to that predicted by rigid-block kinematics from the offsets measured at Sange Dun. Taking the movement on the Zhuanglang He Fault in the last 13.5 ka (=143m) to have been transferred rigidly to the Tianzhu fault (with S 2 45") would imply at least =lo6 m of (4) vertical throw on that fault, as much as =75 per cent more than inferred with uS1 = 1 mm yr-'.
A sizeable fraction of the horizontal slip on the Zhuanglang He Fault must therefore be taken up, past the bend, by the fault whose more subdued trace follows the southern limit of the Tianzhu pull-apart basin. In keeping with the hypotheses made above, 47 m of left-lateral slip on that fault since 13.5 ka B.P. would be required to bring the coeval offset and maximum throw at Sange Dun and Tianzhu to fit. The corresponding slip rate would be =3.5 mmyr-'. While that rate could be smaller if the sedimentation rate along the Tianzhu fault had been faster, the evidence we have at hand is insufficient to resolve the trade-off between these two parameters.
Although rigid-block kinematics does not hold at the south Tianzhu bend, the compatibility analysis points to three fairly reasonable inferences. First, the dip of the Zhuanglang He Fault is probably very steep (80-85"). while that of the Tianzhu Fault could not be much steeper than 45". Second, the slip-vector azimuth along the western Haiyuan Fault near Tianzhu probably lies in a narrow range, between NllOE and N115E. Third, the fault along the southern limit of the Tianzhu Basin must be active. At a more detailed level, we take the subdued morphological signature of this latter fault to imply that its postglacial left-slip rate is unlikely to have exceeded 3.5 mm yr-' (-30 Farther westwards, terrace tongues incised several metres deep by the streams that feed the prominent fan north of Song Shan (Fig. 12) have recorded clear cumulative left-lateral displacements. The risers that bound the entrenched stream channels are offset 20-40 m (Fig. 13b). While such evidence is too limited to warrant quantitative analysis, the observation that low-level terraces are offset a few tens of metres and that the last large seismic event to have ruptured the fault is at least several centuries old (the Ming dynasty lasted from 1368 till 1644A.D.) is in keeping with the conclusions reached at Sange Dun.
East of Song Shan, the western Haiyuan Fault slices the lower-Palaeozoic core of the Lao Hu Shan into two massifs of unequal width but similar elevations (3222 and 3316 m, respectively, Figs 11 and 14). From 103O45'E onwards, a 14 km long, 2400-2800 m high stretch of the fault cuts across the north slope of the southern massif, a fairly rugged area incised by streams with comparable spacing (a few hundred metres) and upstream lengths (1 to 2 km from massif crest to fault trace, Fig. 14b). That stretch displays outstanding cumulative offsets, whether of the stream channels or of the intervening crests, some of which are capped by high-level terraces (Figs 14a). By restoring the linearity, across the fault trace, of left-laterally displaced channels and crestlines clear on the SPOT image, the aerial photographs and the topographic map, we obtained 12 measurements of well-defined offsets (Fig. 14b). The offsets range between extremes of 75 f 20 and 200 f 20 m, with 60 per cent of them clustering between 100k20 and 150*20m, and an arithmetic mean of 120m. Short of detailed field sampling, which has not been done, one may only hypothesize what the ages of such offsets are. But it is at once apparent that the range of offset values is similar to that of the high-level terrace offsets near Sange Dun and in Lenglong Ling (Figs 6d and e, Table la). We interpret this to imply that, although the elevation of the fault trace and of the mountains south of it is 500 to lo00 m lower here than west of Tianzhu, the evolution of landforms was similarly influenced by the warmer, more humid climate that accompanied deglaciation. Hence we infer, as we have at Tianzhu and Sange Dun, that much of the entrenchment of small streams on the north slope of Lao Hu Shan occurred between the last glacial maximum and the end of the early Holocene climatic optimum (18-6 ka B.P.), and that increasing offset values result from increasing ages of entrenchment. Such a range of morphological ages would make postglacial horizontal slip rates on the Maomao-Laohu Shan and Zhuanglang He segments of the western Haiyuan Fault compatible (8.7-13.9 mm yr-' versus 11 f 4 mm yr-', Fig. 15a). If for instance, as near Tianzhu, the largest offsets (-200m) were 1 8 f 2 k a old, the slip rate would be 11.1 f 2.3 mm yr-'. Similarly, taking the ages of the 120-150 m and 75-100 m offsets to be 13.5 f 2 ka B.P. and 8.4 f 4 ka B.P., respectively, as at Sange Dun, would imply rates of 10 f 3 mm yr-' and 10.4 f 7 mm yrf'. Recall that the greatest merit of such reasoning, which is no substitute for direct dating of the offsets, is to test whether results that rest on the assumption that landforms over fairly vast regions are principally and coevally shaped by global changes of the climate are self-consistent, as argued in particular by Armijo et al. (1986Armijo et al. ( , 1989, Peltzer et d. (1988,1989) and Avouac et al. (1993). We think that such per cent of that at Sange Dun). Therefore, we find it plausible that sedimentation rates in the western part of the Tianzhu Basin could have exceeded 2.7 mm yr-', which suggests that this intramontane basin might have remained isolated from outside catchments until recently.
That the last earthquake on either of the two main faults that meet at the south Tianzhu bend is at least several hundred years old suggests that they may slip together during the same event. Given the slip partitioning we infer at that bend, events with 4 . 8 or =11.7m of left-lateral offset at Sange Dun might then produce at least ~3 . 5 or -4.7 m of normal throw on the west Tianzhu Fault north of the bend. During such events, the south Tianzhu Fault might slip left laterally by as much as -2.8 or =3.7 m, respectively. Whether this scenario, rather than slip of the three faults in a sequence of distinct earthquakes, possibly clustered in time, is more likely cannot be decided without trenching. But the fact that free faces are known to last at least -250 yr on normal fault scarps higher than -2m in west-central China ) supports the inference that if the west Tianzhu Fault does rupture during great earthquakes, it last did long ago.

Maomao Shan-Lao Hu Shan (eastern) segment
From 103"E to 104"E, over a distance of =lo0 km, most of the present movement along the Haiyuan Fault zone is taken up by a principal, continuous strand striking N100-105E on average (Fig. 11). On either side of that strand and less than 10 km north or south of it, subordinate splays or disconnected, active-fault segments strike mostly between N90E and N100E. Between 103"E and 103"20'E, the main fault strand hugs the northern flank of Maomao Shan, a bulky, 40 km long, 8 km wide, mountain ridge that culminates at 4074m. The core of the mountain exposes Silurian greenschists in the north and Ordovician volcanics in the south, separated by a pinched fault slice of Triassic sandstones parallel to the main active fault strand (Gansu Geological Bureau 1975;Xu et al. 1989). We could not reach that strand north of the mountain, but the satellite images and a few aerial photographs reveal high north-facing triangular facets in addition to prominent left-lateral offsets. As expected from the N110-115E slip vector found at Tianzhu, normal throw thus occurs on the fault, though in a sense opposite to that along the Lenglong Ling range, west of the Tianzhu jog (Figs 1 and 2). Such throw must have contributed to uplift the Maomao Shan.
East of Maomao Shan, the main fault strand crosses an area with more subdued relief (Fig. 11). mantled by loess. Whether on SPOT images, aerial photographs or in the field, its strike-slip trace is nonetheless outstanding. It becomes particularly spectacular east of its junction with a west-directed splay that may be linked, through faults along the south side of Maomao Shan, with the south Tianzhu fault (Figs 2, 11 and 12a). In the steppe north-east of Song Shan, a small fortified city whose wall, erected during the Ming Dynasty, stood intact until the M -5.8 shock of 1990 October 10 (Figs 1 and l l ) , a 2km long pull-apart basin signals a minor left step in the fault trace (Fig. 12b). Towards the west, left-stepping pressure ridges, several metres long and well smoothed by erosion, or a small north-facing scarp limit the floor of that basin (Fig. 13a). consistency, for a length of about 200 km, between =102" Shan, which have comparable elevations. In the direction and =104"E, along the Haiyuan Fault, may be taken to parallel to the fault (N108E), the separation of the massifs justify our use of that assumption as a first-order tool. eastern tips, as defined by the 2800, 3000, and 3200m Whether on satellite images or maps, two left-lateral elevation contours on the 1:100,000 topographic map (  1989). Despite kilometre-wide meanders, the clear, broadscale swing of the river course across the fault has a nearly symmetrical, sigmoidal shape (Fig. 1). Striking first =N-S for -15 km past the confluence with the Zuli He, that swing veers to N135-115E over the next -40 krn, then back to =N-S for -25 km (Fig. 1). The inflexion in the S-shape does not coincide with the intersection of the river with the present trace of the fault, but lies less than -8 km south of as it steps to the north and traverses desert sandy flats. It then resumes its sharpness, crosses the Huang He river along the base of a ZN100E-striking, =500m high, north-facing topographic step, and merges with the curved trace of the south-dipping, Tianjing Shan thrust ramp (Tapponnier & Molnar 1977;Deng et al. 1984;Zhang er al. 1990) (Fig. 1). Left slip on the Gulang Fault is attested by morphological offsets of decametric and kilometric scale. The most outstanding kilometric offset is that of the Gulang River, =8 km south of Gulang, which is deflected -5 km by a prominent shutter ridge that places uplifted Lower Palaeozoic rocks north of the fault against Cretaceous red beds south of it (Figs 1 and 16b). Clear decametric offsets exist at Tianxiao Kou in the upper valley of the Xiamenzu River (Figs 16 and 17), a -4km wide trough between -4400 m high ranges bearing strong, recent glacial imprint. On the north side of the valley, three strands of the fault cut the gently SSE-sloping surface of a terrace that caps abraded Triassic beds ( Fig. 17a and b) and is incised 50-90m by tributaries of the river. Two of the strands, which strike N90-100E, meet with a third, middle strand that bends to =N80E (Figs 16 and 17). While the first two strands show only minor dip slip, normal throw is prominent on the middle strand, possibly because of its more northerly azimuth. The northern and middle strands show inconclusive evidence for consistent horizontal offsets on the SPOT image. By contrast, a succession of four entrenched tributary channels are offset left laterally 60 to 70&20m by the southern strand, which we therefore infer to accommodate most of the strike component of slip along the Gulang Fault zone at this site (Figs 17a and b). Vertical throw on the middle strand has created a north-facing, cumulative scarp up to =40m high on the terrace top (Fig. 17c) and led to deeper entrenchment, hence greater width, of triburary channels south of that scarp (Fig. 17a).
As at Sange Dun, the freshness of glacial landforms around the mountain-locked, ~3 3 0 0 m high Tianxiao Kou site implies that incision of the terrace by the tributary streams postdates the last glacial maximum. The terrace forms the highest flat surface north of the Xiamenzu river and covers a broad part of the Tianxiao Kou trough. Moreover, it blends gently with the mountain slope. Hence, we infer much of the planation of the terrace surface to be due to periglacial smoothing, as inferred by Peltzer et al. (1988) in a comparable area of western Gansu. While such smoothing probably reached a climax during the last glacial maximum, it could have continued till the onset of rapid deglaciation on the mountains. Assuming entrenchment of the tributary streams to have started at most 18 ka and at least 14 ka B.P., their channel offsets (65 f 25 m) would be consistent with a constant left-slip rate ranging between 2.2 and 6.4 mm yr-' (4.3 f 2.1 mm yr-'), in a direction ranging between N85E and NlOOE (N92 f 7E), on the Gulang Fault. The 5 km offset of the Gulang River would thus result from slip during 1.5 f 0.7 Ma. Although the evidence from which such bounds derive is too limited to draw firmer conclusions, the values obtained are in keeping with the qualitative observation that the morphological signature of the Gulang Fault is less impressive, on satellite images and in the field, than that of the Haiyuan Fault, and imply coeval slip on those faults during much of the Quaternary. that trace, near another Late Cenozoic fault visible both on extant maps (Gansu Geological Bureau 1975;State Seismological Bureau 1990) and satellite images (Fig. 1). Since the swing links two fairly straight, 4 0 km long valley segments with similar average strikes north and south of the fault (N63 f 3E and N55 f 3 E , respectively, Fig. l), the offset between the two is well defined. Measuring that offset as we did for the terrace risers at Sange Dun, and allowing the fault strike in the area to range between NlOOE and N120E, yields a value of 95 f 15 km. Closer to the fault, the offset of the two NS-striking valley segments within the swing is 45 f 5 km. Fairly deep entrenchment (-500 m) of the Yellow River downstream from the fault probably reflects regional uplift of the upper crust above the active, south-dipping Tianjing thrust (Fig. 1). Even upstream, the river meanders in a flood channel (at most 6-7 km wide) that stands well below the summits of the mountainous landscape. The river course has therefore long been captive to the surrounding relief, unable to wander tens of kilometres away as great rivers can d o on occasion of millenal floods where flowing across flat plains. Besides, there is little evidence of capture, or of large abandoned river channels, on satellite images or on topographic and geologic maps. Therefore, we infer the 95& 15 km left-lateral offset of the Huang He Valley across the Haiyuan Fault zone to represent a lower bound of the total displacement on that zone (Gaudemer et al. 1989). At a constant rate of 11 f 4 mm yr-', 8.6 f 4.5 Ma would have been required to produce such an offset, implying slip along the fault zone since the upper Miocene, and at least since the onset of the Pliocene (Fig. 15b). The smaller offset within the Yellow River swing would be mostly of Plio-Quaternary age (4.1 f 1.9 Ma).

ACTIVE FAULTING A N D CRUSTAL SHORTENING NORTH OF T H E H A I Y U A N FAULT
Recent crustal deformation north of the Haiyuan Fault involves fairly complex faulting, principally along the Gulang left-lateral ramp and the en Cchelon, Shiyang-Dongqingding thrusts (Figs 1 and 16). There is convincing, though mostly indirect, evidence that the 1927 May 23, M = 8.3, Gulang earthquake ruptured the latter thrusts rather than the Gulang ramp. Our observations of surface breaks related to that earthquake, however, bear little resemblance to those previously reported (e.g. Jia 1982;Molnar & Deng 1984).

Gulang Fault
A clear, linear feature on the Landsat image (Tapponnier & Molnar 1977), the Gulang Fault zone splays eastwards from the Haiyuan Fault about 60 km west of Tianzhu (Figs 1 and  16). For -150 km, its closely spaced strands strike between N80E and N100E, oblique both to the Qilian Shan front and the Haiyuan Fault. Near 103"45'E, =60 km east of Gulang, one strand veers to NllOE at the base of the north-dipping limb of a =30 km long ridge of folded Palaeozoic rocks. East of that ridge, the =EW fault trace becomes less conspicuous Continuous line with open triangles at base of Carboniferous is inferred regional dCcollement. A-A' is line of section in Fig. 16(c). Epicentre of great 1927 earthquake (solid circle) is from Repetti (1928). Box is location of Fig. 17. (c) Top: interpretative cross-section of eastern Qilian ranges along line A-A' in Fig. 16(b). Topographic profile, with vertical exaggeration of =7, above geological section is from  Besides left-lateral slip, the Gulang Fault shows subordinate vertical throw over most of its length. In contrast with the Haiyuan Fault, along which normal slip components or pull-apart jogs are common, and save for the Tianxiao Kou scarp near that fault, however, such throw appears to result mainly from shortening. The Gulang River shutter ridge, for instance, and the larger ridge near 103O45'E (Fig. l), which reach elevations of 2800-2900 m, are best interpreted to represent growing ramp anticlines in compressional jogs along the fault. In addition to such features, the Gulang Fault follows the edge of an area of gently rising relief, made of uplifted pre-Cenozoic rocks, south of the Tengger Desert (Fig. 1). Most highs north of the fault are capped by north-sloping, thickly loess-mantled, Pleistocene alluvial aprons fed by the erosion of such relief. Moderate uplift south of the fault, the existence of push-up jogs along it, and the fact that it merges eastwards with the Tianjing thrust, are all consistent with a minor component of

Thrust Solution
north-directed thrusting. Hence, we interpret the Gulang Fault to represent the surface trace of a south-dipping lateral ramp, whose main function is to transfer some crustal shortening from the Qilian Shan to the Tianjing, Mibo and Yantong Shan thrusts. Given the strike difference of 38 f 5" between the Gulang Fault and the thrusts, a left-slip rate of 4.3 * 2.1 mm yr-' on that fault would be compatible with 2.8 * 1.6 mm yr-' of shortening across these thrusts. Such a value allows for the 1.5-2.7 mm yr-' shortening rate estimated by Burchfiel et al. (1989) and Zhang et al. (1990).
Where we could gain access to the Gulang Fault in the field, south of Gulang and at Tianxiao Kou, we found no clear evidence of recent seismic breaks on it. These two localities bracket the westernmost segment of the fault, which lies closest to the epicentre of the 1927 May 23 earthquake (e.g. Tapponnier (Fig. 16). The sharpness of the thrust traces, marked by outstanding slope breaks on satellite images and in the field, implies contemporaneous shortening and uplift along them (Fig. 16). West of Wuwei (Figs 16a and b), such uplift has in fact led to the formation of fresh, prominent colluvial-alluvial aprons north of the thrust trace, a feature characteristic of most active mountain fronts in Gansu and Qinghai (Tapponnier et al. 1990;Meyer 1991). In map view, the thrusts are arranged in a right-stepping array, with three principal segments (west Shiyang, east Shiyang and Dongqingding) offset from one another, along the base of three en Cchelon mountain ridges that rise to more than 3200m, -1700m above the Tengger Desert (Figs 1 and 16). The thrust traces thus step closer to the Haiyuan Fault towards the east (from =70 to =40 km, Figs 1 and 16). There is additional evidence for active shortening 10-30 km north of the topographic front. Such shortening is clearest west of the Shiyang River, where it is marked by local uplift of the river-fan surface along a narrow, =N110E-trending ridge incised by divergent fan channels (barren, light-grey stripe on the Landsat image, Figs 16a and b). That ridge, which is comparable to those that cross Holocene piedmont fans in the western Qilian Shan and Tien Shan (Meyer 1991;Avouac et al. 1993), appears to form the top of a growing anticline, which we interpret to signal the emergence, along a south-dipping ramp, of a shallow foreland dkcollement branching off the thrust beneath the mountain to the south (Fig. 16c). Slightly entrenched piedmont-bajada channels (Figs 16a and b) suggest that a similar foreland thrust ramp may exist near and east of Wuwei, but agriculture and construction in the populated city oasis obscure other surface evidence.
Constraints on the overall geometry and development of the active thrust system may be derived from the main structural traits of the eastern Qilian Shan. To a first order, the rocks belong to two principal units. The older unit is a basement of Early Palaeozoic rocks (mostly Cambro-Ordovician greenschists and slates, with some volcanics, quartzites and marbles) deformed, metamorphosed and intruded by large bodies of syn-to post-tectonic granodiorites and granites at the time of the Caledonian orogeny ( Fig. 16c) (Gansu Geological Bureau 1975;X u et al. 1989). The younger unit is a cover of platform sediments of Devonian to Quaternary age, =6000 m thick on average, resting unconformably on top of the basement. Locally thin Devonian red beds are overlain by Carboniferous black shales with interbeds of richly fossiliferous marine limestone, then by a massive sequence of Permo-Triassic and Jurassic continental sandstones with some conglomerates (Fig. 16c). There is little disconformity within this sequence, which reaches a thickness of ~4 5 0 0 m and forms impressive ~20"S-dipping monoclines north of Tianxiao Kou and on the Shiyang He-Jioutiaoling section west of Wuwei (Figs 16b and c). The most significant angular unconformity is at the base of the Jurassic sandstones, which contain coal horizons and whose thickness is variable (up to 1500m at places). Lower Cretaceous red beds (at most SO0 m thick) fill disconnected basins whose depocentres do not coincide with Permo-Triassic or Jurassic depocentres (Xu et al. 1989). Ochre-pink Neogene (particularly Pliocene) sandstones and siltstones, locally with metre-scale cross-bedding and at least 500m thick, are exposed in the two steps between the thrusts and in the Haxi syncline (Fig. 16b). At places, Neogene strata are capped by hundreds of metres of Quaternary conglomerates. The Plio-Quaternary beds, as well as other exhumed rocks, are commonly mantled by layers of loess up to several tens of metres thick The regional geology thus implies that a rather peaceful regime of shallow marine to terrestrial sedimentation presided over the Late Palaeozoic and Early Mesozoic epochs, with the Triassic-Jurassic boundary marking the onset of resumed tectonic warping and faulting, as in the Tien Shan (e.g. Bally et al. 1986;Peng & Zhang 1989Carroll ef al. 1990Graham et al. 1990;Hendrix et al. 1992;Enkin et al. 1992;Avouac et al. 1993). That the Lower Cretaceous red beds fill isolated basins implies that further faulting controlled their deposition, whether in pull-apart or synclinal troughs (Gansu Geological Bureau 197.5; Xu et al. 1989). The absence of Late Cretaceous and Early Tertiary deposits, except in the Xining basin (Gansu Geological Bureau 1975; Bureau of Geology and Mineral Resources of Qinghai Province, 1991), is suggestive of moderate elevation and relief at that time. Following this sedimentation lapse, the renewed, growing flux of Neogene-Quaternary clastics probably reflects the increasing proximity of rising relief, particularly since the Late Miocene.
In the Shiyang Valley, the cover is separated from the basement by a dCcollement (Fig. 16b). Intense strain, including tight disharmonic folding, flatish shear zones and boudinage, remains mostly localized within the Carboniferous shales, with the bulk of the more competent strata above showing only gentle warping and high-angle brittle faulting. The granite below the dicollement displays retrograde shear and cataclasite zones. Various shear indicators imply NNW-directed transport in the present geographic coordinates. Disharmonic folding is also prominent under the massive sandstone monocline on the Xiamenzu-Tianxiao Kou section. On the way up to Dongqingding mountain (Fig. 16b), a granite below Carboniferous rocks has been converted to gneiss, with a strong stretching lineation. From these three distant outcrops, we hypothesize decoupling between cover and basement to be of regional extent. The dkcollement, which dips towards the north at Dongqingding or in the northern part of the Shiyang He section and towards the south north of Tianxiao Kou, has been folded in harmony with the Carboniferous-Jurassic sedimentary cover, whose main patches on the map of Fig. 16(b) correspond to two large, asymmetric synclines with steep north-dipping south limbs. On the section of Fig. 16(c), both limbs are truncated by the south-dipping Gulang and Dongqingding thrust ramps, which bring the Ordovician metamorphic rocks onto the Permo-Triassic sandstones. A pinched syncline of steeply dipping Triassic beds separates the Gulang from the Haiyuan Fault at Tianxiao Kou.
The structure and morphology of the two mountain ridges south of the Dongqingding and Shiyang thrust traces (Fig.  16) imply that they form basement-ramp anticlines (e.g. Tapponnier et a/. 1990; Meyer 1991). That such anticlines are recent and still growing is clear from the canyons ('cluses') incised across them by rivers such as the Dongda He and Huangyang He, from the ponded floodplains upstream from such cluses, and from the flat tabular surfaces that top the ridges (Figs 1 and 16). Though hanging 1200-2400 m above the foreland and 500-1200 m above adjacent, entrenched valleys, those flats, which display complexly indented contours and slope gently (2-6") towards the Tengger Desert west of the Zamu He, still withstand vigorous erosion. We interpret them to represent vestigial patches of planation surfaces, heaved by relatively rapid movement on the underlying basement ramps, and dismembered by the resulting incision and headward retreat of streams. The age of planation is uncertain and may not be uniform, but the surfaces clearly postdate exhumation of the dicollement and apparently, at one locality, the Upper Pleistocene ( Q2-J basal unconformity (Gansu Geological Bureau 1975;Xu et al. 1989) (Fig. 16). Thus, planation of the surfaces appears to have taken place well after the onset of folding, possibly at most 700 ka B.P., and may be thought t o result chiefly from glacial abrasion and/or periglacial smoothing.
Though large-scale shortening characterizes the Late Cenozoic tectonics of the eastern Qilian Shan, several recent, -N115E-striking faults that cut the hanging-wall wedge of upper crust north of the Haiyuan-Gulang strike-slip zone appear to be normal. Those subordinate faults, which have particularly sharp and linear traces on satellite images (Fig. 16a), lie near and roughly parallel to the frontal thrusts (Fig. 16b). Three of them dip steeply, apparently towards the south, along the backs of the Shiyang ranges. Both the west Shiyang and east Shiyang rangeback faults are located north of the westward projections of the east Shiyang and east Dongqingding active thrusts, respectively (Figs 1 and 16). Given their orientation, the three faults may accommodate a component of left slip, in addition to down-to-the-south throw. Another normal fault dips north, only 2-4km south of the east Dongqingding thrust (Fig. 16). We surmise that the chief function of all four faults is to accommodate the hanging-wall strain that results from dip changes between flats and ramps on sole thrusts with weak frictional resistance (Meyer et al. 1990;Tapponnier 1993;Avouac, Meyer & Tapponnier 1992). More specifically, we interpret the Dongqingding normal fault as a simple compatibility fault (e.g. Avouac et al. 1992), possibly guided by the basement-cover dkcollement, and the Shiyang range back faults, as former thrust ramps, steepened and inverted while displaced past a flat-to-ramp kink on the basal, active Shiyang thrust (Fig. 16).
The interpretative, retrodeformable section of Fig. 16(c) (A-A', bottom) summarizes the overall, deep geometry of faulting that we infer from the regional structure (Figs 16a and b) and the topography (Fig. 16c, top). That three distinct, overlapping thrusts are active while more or less vigorous uplift affects the whole wedge of crust north of the Haiyuan Fault suggests that they merge downsection into a single, active, gently south-dipping sole thrust or dicollement and terminate to the east along lateral tip-lines or ramps (e.g. Boyer & Elliott 1982). We assume the dips of the Dongqingding and Shiyang thrust ramps to vary between  -topped anticline ridges would imply t h a t illc, I i l i i i j~\ reach depths of =12 and -18 km, i-cspecti\el\. 'I'hc hoIc, thrust, which probably steepens southward\ undcr. the hiyli hinterland relief. thus ought to lie 12 to 25 k m deep and dip only lo" to 20% consistent with the attitude\ ot tlic Tianxiao Kou and Jioutiaoling monoclines (Fig. I OL I. Similar reasoning implies that the shallou lorchiid dCcollement north of the east Shiyang range lies 1 to 2.5 k n i deep, probably near the base of the Pliocene (Fig. Ihc).
Beneath the Lenglong Ling range, wc infer the klai)u,iri Fault to extend steeply -25 kiii down to H Ixinchlinc M ill1 the sole thrust (Fig. 16c). That inference is in keeping \\it11 the presumed depth of the M = 8.7, Haiyuan eartliqu~ikc rupture (20 km, Molnar & Dcng 1984) and Iigioecntrc (17 km, Gongxu,Tinghuang & Zhenliang 19x0, Fig. 16c). which broke through thinner crust in a les4 mountainous region to the east. The observation that the right-stepping thrust ramps emerge farther from the Haiyuan I'xult westwards, as hinterland relief becomes higher and broadel-. implies that the branch line between that fault and the sole thrust plunges even deeper west o f the section o f Fig. 16(c). Conversely, as this branch line shallows eastwards. tho innermost parts of that section also shallow. resembling its outermost parts. That the inner thrust traces become less sharp and that hinterland-dipping normal faults come into being west of the lateral tip lines of the outer thrusts. suggests that a significant fraction o f the present-day shortening is transferred to the latter along section. as ought to occur within a leading, northwards growing, imbricate thrust fan (Boyer & Elliott 1982). Hence. though non-unique, the schematic section of Fig. 16(c) suffices to account fairly simply and completely for the large-wle structure and recent evolution of the eastern Qilian Shan. which is but a variant of a style similar t o that in othcr active, central Asian thrust belts (Tapponnier c' f (11. 1990: Meyer 1991Avouac et a/. 1993).
Conservation of surfaces on the cross-section of Fig. 16(c). and restoration of the folded and faulted cover/basenient dCcollement to a flat, slightly south-dipping unconformity. yields =2S km of horizontal shortening between the Gulang Fault and the foreland (=25 per cent of an initial length of =lo0 km). The greatest source of uncertainty on that number comes from the dips inferred for the thrust ramps and flats. The existence of the tight syncline south o f the Gulang ramp and thrust slip on that ramp probably add a few kilometres to the cumulative shortening across the mountains north of the western Haiyuan Fault. Although the basement/cover dicollement apparently postdates the Jurassic, our observations are insufficient to determine whether movement on it took place during, or prior to. the Cenozoic (Yenshan orogeny), and how large its contribution was to upper-crustal shortening. We simply conclude that 25 km is a lower bound on the amount o f Late Cenozoic shortening between the Haiyuan Fault and the Tengger foreland.
Rough estimates of uplift and shortening rates may also be derived from Section A-A' (Fig. Ihc). If most of the elevation difference (=1200 m) between the foreland and the planation surfaces topping the east Shiyang and east Dongqingding anticlines were due to tectonic uplift in the area, which extends between 101.6-103.5"E and 3.1-38.4"N, encompasses most of the Shiyang and Dongqingding ranges, as well as the southern part of the Tengger foreland, and includes the three inferred epicentral locations. Its long axis lies roughly along the right-stepping thrust traces that stretch north-west from Gulang past Wuwei (Fig. 1). The isoseismal of intensity IX (=110 km X 45 km) still surrounds completely the Dongqingding and east Shiyang thrust traces. Hence, although neither the epicentral locations nor the isoseismal contours should be considered accurate, the areal coincidence of all the above elements, together with the absence of surface ruptures along the Haiyuan and Gulang Faults near Tianzhu and Gulang, suggests that the 1927 earthquake resulted from slip on the Shiyang-Dongqingding thrust system. The body-waveform least-squares inversion method of Nabelek (1984) was used to examine the plausibility of that inference. The waveforms of the radial component P wave and the SH wave recorded on the horizontal component Galitzine instrument at the station UCCLE (Brussels, Belgium) rule out pure strike-slip faulting and are consistent with N24E-directed thrusting on a N136E-Striking, 50"s-dipping fault plane, with a minor component of right slip (Figs 1, 16b and d).
The most common, if imprecise, description of surface breaks attributed to that earthquake (e.g. Jia 1982;Deng et al. 1984;Molnar & Deng 1984) invokes two 'conjugate' fault segments intersecting at high angle in map view. Reportedly, 2 to 6 m of reverse throw with a component of right slip took place on a 33 km long segment oriented N160E, while 2 to 7.5 m of reverse throw with some left slip occurred on a 28 km long, south-dipping segment oriented NllOE, with slip maxima on either segment near their intersection. Deng et al. (1984, Landsat image of Fig. 2a in their paper) identify the first and second segments with the escarpments truncating the Haxi syncline north of Gulang, and marking the Dongqingding range front, respectively (Figs 1 and 16). Although we had only limited access to those escarpments, our field observations and conclusions are generally not consistent with the above numbers, and hence with earthquake parameters and other speculations derived from them (Molnar & Deng 1984).
The first N160E-striking escarpment, for instance, is only -17 km long on the Landsat image, about half the length alleged above, with an average strike of =N170E (Fig. 16). For most of its length, it forms the western riser of the Gulang river fan, resembling the other riser facing it east of the river. At a more detailed level, its slope is steeper and its height greater than are those of the eastern riser, and one hill remains isolated east of it (Fig. 16). From a distance of 1 to 2 km, however, and with binoculars (the utmost we could achieve in that strategic valley!), we saw little evidence for fresh, east-facing scarps several metres high along the base of that escarpment. Hence, though we cannot, for lack of access, rule out faulting along the north Gulang escarpment during the 1927 earthquake, nor question, for regional strain-compatibility reasons, the style of faulting, both the length of faulting and the amount of throw reported seem exaggerated, each probably by at least a factor of two.
The second escarpment is indeed =28 km long, with an average strike of N115E. But in the field, at two different sites about 10km apart (Fig. 16), again we found no evidence of large, recent, seismic throws. Near Chouniu last 700 ka, the vertical throw rate on the underlying ramps would be of the order of 1.7 mm yr-'. This number would represent a lower bound if the assessment of the Quaternary stratigraphy (Qz-3) on the geologic map were taken at face value. For the east Dongqingding ramp, the same rate (1.7mmyrF') would be obtained assuming that the Q34 fanglomerates that stand =200m above the piedmont and cap the gently ESE-dipping upper-Neogene beds of the north limb of the Haxi syncline (Gansu Geological Bureau 1975) were deposited by the Xiamenzu and adjacent river mostly during the Riss-Wiirm interglacial (~1 2 0 ka B.P.). That number would represent an upper bound if abundant fanglomerate deposition were restricted to warm and humid interglacials. Both estimates are in keeping with those derived farther west along the thrust belt from evidence of postglacial faulting and folding (0.9 to 1.9 mm yr-': Yumu Shan, Meyer 1991;Tapponnier et al. 1990, 1.5 f 0.3 mm yr-': Taxueh Shan, Meyer 1991). The highest planation surface, on top of the west Dongqingding range north of Xiyingerh Kou (Figs 1 and 16), might testify to more vigorous upheaval above the west Dongqingding ramp. If that surface, now up to 3900m high, had risen =2400m relative to the piedmont in 700kyr, the vertical throw rate on the ramp might have reached 3.4mmyr-'. Alternatively, and probably more likely in view of the distance between that range and the piedmont, planation might have taken place at a higher initial elevation and/or earlier than 700 ka B.P. We conclude that plausible shortening rates along the eastern Qilian Shan thrust system probably lie between 2 and 6 mm yr-' (4 f 2 mm yr-I), a range of values not unlike that inferred in the northern Tien Shan (3 f 1.5 mm yr-'), (Avouac et al. 1993). At 4 mm yr-', roughly 6 Myr would have been necessary to absorb 25 km of shortening.

Mechanism of the 1927 May 23 earthquake
As located by Gongxu et al. (1989), the hypocentre of the Wuwei-Gulang earthquake of 1927 May 23 lies at 37.7"N-102.2"E, under the west Dongqingding range front (Figs 1 and 16). Alternative locations place the epicentre at 37.6ON-102.6"E (Repetti 1928;Lee, Wu & Jacobson 1976), about 10km south of the east Dongqingding thrust or at 38.05"N-102.37"E (Molnar & Deng 1984), near the eastern extremity of the west Shiyang thrust trace in the foreland. Although the macroseismic depth of the earthquake hypocentre is reported to be =12 km (Gongxu et af. 1989, Fig. 16c), that value is probably a 'default depth'. While the magnitudes reported by various authors vary between 8 (Gongxu et al. 1989) and 8.3 (Molnar & Deng 1984), there is little doubt that the main shock, which had a M ~5 . 5 foreshock 49 minutes earlier, many large aftershocks (of which 2 with M = 6, and 6 with M 2 5.25) in the following 10 months, and a moment Mo-4.3 X 1OZ0Nm (Chen & Molnar 1977), ranks as a great earthquake. As defined by the isoseismal of intensity VIII (Gongxu et al. 1989, see also State Seismological Bureau 1993, and Ding 1985, its meizoseismal area is -200km long and -90 km wide, with a long axis oriented -N130E (Figs 1 and 16). Though the macroseismic evidence used in deriving the isoseismals is probably biased by the existence of rugged, high and uninhabited terrain towards the south-west, that Gou, north of the Dongqingding summit, we could see no clear, fresh scarp along the loess-smeared range-front thrust. At Wangjiaxin Zhuang, near Shuixia Kou, above the east bank of the Huangyang river, we examined a =NlIOEstriking, 6-7 m high scarp limiting a high-level, cultivated terrace behind adobe houses and reported to have formed in 1927 (State Seismological Bureau 1993). But neither the rather smoothed-out shape and slope of that scarp, nor ground disruption along it-all possibly altered by human action-testified unambiguously to a single rupture less than 70 years old. Having not been able to see a 1927 photograph, we thus infer this fairly high step, which stretches only = I km eastwards, to represent the cumulative record of several earthquakes on the high-level terrace. That inference is corroborated by the occurrence, about 1 km farther east along strike, of small (30-50cm high), well-preserved pressure ridges and thrust scarplets that fold and cut the turf (Fig. 18a). In aspect, these relatively fresh features are indistinguishable from those seen along the breaks of either the 1932, Changma (e.g. Peltzer et al. 1988;Meyer 1991) or 1920, Haiyuan earthquakes (e.g. Deng et al. 1984) and thus may be attributed to an earthquake in 1927. They lie along the base of the 20-30m high cumulative scarp that marks quite sharply here the eastern end of the Dongqingding range-front thrust by lifting into hanging position, relative to the Holocene piedmont bajada, loess-mantled troughs dug into gently south-dipping Neogene sandstones (Figs 16a, b and 18a). Our observations near Wangjiaxin Zhuang may thus be taken as consistent with sustained seismic thrusting on the Dongqingding range front, but the 1927 surface throw appears to have been modest, probably only about one tenth of the value used by Molnar & Deng (1984) to derive the seismic moment.
More impressive ground ruptures scar the landscape south of the range front. Their size and freshness testify to formation during the great 1927 earthquake, the only large regional event in the last few centuries. Many of these ruptures simply correspond to the free-faced head scarps of local, thin-skinned loess slides down steep valley slopes. Such slide-related scarps are widespread throughout the Haxi syncline and surrounding mountains, implying particularly strong ground shaking there, in keeping with the fact that the area lies within isoseismal X (Figs 1 and 16). But kilometres-long, free-faced scarps that slice uplifted Palaeozoic basement rocks and seem unrelated to superficial landsliding also exist within the Dongqingding range, mostly north of its crest. Several such scarps, oriented roughly parallel to the range crest and facing north or south, are clear when looking down from the television tower on top of the range (TT, Fig. 16b). They appear to belong to a discontinuous system, extending for at least 14 km between the Huangyang He in the east and the Zamu He in the west (State Seismological Bureau 1993, Fig. 5, p. 129). Upstream from Chouniu Gou, we could follow and crudely map one particularly prominent scarp zone across rugged mountainous terrain on either side of an E-W pass between two north-draining valleys (Figs 16,18b, c, d, 19a and b). Certain features along that zone resemble those depicted on photographs taken farther east, at Shenjiawo Pou (State Seismological Bureau 1993, photographs 34 and 35, p. 129).
Severe ground disruption starts transverse to the first, eastern valley (Taojung Gou), in the form of a narrow, N100-110E trending stretch of chaotically conioi-tcd iai-f. with exhumed boulders separated by water-setyinp lenws 01 black and grey-green gouge, which climlx up thc wc\t \itJt. of the valley. Movement appears to have raised thc rocks and ground surface up valley. south o f that Tone (Fig. 1%)). Farther up slope, antithetic scarps with well-prescr\cd frce faces up to -1 m high enclose a N120-13O~;-trc.nding pull-apart depression, about 10 rn wide and sever;\l 1cn5 01 metres long, with 1 to 2 m wide and deep collnp\c pits '11 either end (Fig. 19a). The zone then narrows into ;I narrower crack-like furrow oriented =NWE. and continue\ as a single north-facing, N100-1 10E-trending. free-liiccd scarp (Fig. 1Xb) all the way to the pass. uhosc distance from the valley floor we estimate to be of the order o f 600 ni (Fig.  19a). Consistently higher north-facing scarps all along the scarp zone east of the pass, as well as cumulative slope offsets of up to 3.4-3.9m opposite t o the generally south-directed slope gradient (Figs IXb and I%, inset AB). indicate down-to-the-north throw, which rules out 21 superficial landslide origin. At a more detailed level, the orientation, in map view, of the pull-apart depression relative to that of adjacent segments of the scarp zone suggests some component of right slip across that zone (Fig.  19a).
At the pass, the north-facing scarp is particularly outstanding, about S0m long and 1 to 1.7 m high, with a fresh free face and a modest debris wedge at its base (Fig.  18c). Sudden slippage has uprooted willow trees or cut and exposed their roots along that free face. Thc grass-covered morphological saddle of the pass is affected by open fissures roughly parallel to the free face, as well as by smaller, north-facing scarps, implying probably as much as 2 m of down-to-the-north seismic throw and significant extension in a -NNE azimuth, roughly parallel to the local divide (Fig.  19a, inset CD). West of the pass, the zone splays into two main, N-to NE-facing, free-faced scarps (Fig. 19a). The first one, which we did not follow, continues due west across willow thickets. The second one veers to NlSSE, then back to N130E, running obliquely down slope above the narrow gorge of a small stream, all the way to the confluence of that stream with the river west of the pass (Fig. 19a), a total distance of 4 0 0 m. That second scarp, which is 0.5 to 2 m high with down-to-the-north-east throw, opposite to the local slope gradient, limits a fairly linear and continuous furrow that cuts across basement rock and surface debris (Fig. 18d). One stretch of en Cchelon, paired scarps bounding metres-wide, right-stepping, N120E-trending grabens, indicates a component of left-lateral slip (Fig. 19a). Fresh landslides up slope from the scarp furrow funnel into it, implying that surface debris collapsed into ground fissures at the scarp base (Fig. 19a, inset EF). Short fir trees, with straight vertical trunks, at most several decades old, have grown in the furrow and associated slides.
In summary, the overall attitude of the ~1 . 5 km long scarp zone relative to the topographic relief, the freshness, amount and consistency of the deformation observed across the =300m high mountain spur between the two valleys-with south-side up and opposite components of lateral slip east and west of the pass-all concur to indicate formation, in 1927, by as much as 2 m of surface-slip on a =6O0NNE-dipping normal fault, and testify to local, N25 f 15E-directed extension (Fig. 19a). More credit should Gou, north of the Dongqingding summit, we could see no clear, fresh scarp along the loess-smeared range-front thrust. At Wangjiaxin Zhuang, near Shuixia Kou, above the east bank of the Huangyang river, we examined a ZNllOEstriking, 6-7 m high scarp limiting a high-level, cultivated terrace behind adobe houses and reported to have formed in 1927 (State Seismological Bureau 1993). But neither the rather smoothed-out shape and slope of that scarp, nor ground disruption along it-all possibly altered by human action-testified unambiguously to a single rupture less than 70 years old. Having not been able to see a 1927 photograph, we thus infer this fairly high step, which stretches only =1 km eastwards, to represent the cumulative record of several earthquakes on the high-level terrace. That inference is corroborated by the occurrence, about 1 km farther east along strike, of small (30-50 cm high), well-preserved pressure ridges and thrust scarplets that fold and cut the turf (Fig. 18a). In aspect, these relatively fresh features are indistinguishable from those seen along the breaks of either the 1932, Changma (e.g. Peltzer et a/. Meyer 1991) or 1920, Haiyuan earthquakes (e.g. Deng et al. 1984) and thus may be attributed to an earthquake in 1927. They lie along the base of the 20-30m high cumulative scarp that marks quite sharply here the eastern end of the Dongqingding range-front thrust by lifting into hanging position, relative to the Holocene piedmont bajada, loess-mantled troughs dug into gently south-dipping Neogene sandstones (Figs 16a, b and 18a). Our observations near Wangjiaxin Zhuang may thus be taken as consistent with sustained seismic thrusting on the Dongqingding range front, but the 1927 surface throw appears to have been modest, probably only about one tenth of the value used by Molnar & Deng (1984) to derive the seismic moment.
More impressive ground ruptures scar the landscape south of the range front. Their size and freshness testify to formation during the great 1927 earthquake, the only large regional event in the last few centuries. Many of these ruptures simply correspond to the free-faced head scarps of local, thin-skinned loess slides down steep valley slopes. Such slide-related scarps are widespread throughout the Haxi syncline and surrounding mountains, implying particularly strong ground shaking there, in keeping with the fact that the area lies within isoseismal X (Figs 1 and 16). But kilometres-long, free-faced scarps that slice uplifted Palaeozoic basement rocks and seem unrelated to superficial landsliding also exist within the Dongqingding range, mostly north of its crest. Several such scarps, oriented roughly parallel to the range crest and facing north or south, are clear when looking down from the television tower on top of the range (IT, Fig. 16b). They appear to belong to a discontinuous system, extending for at least 14 km between the Huangyang He in the east and the Zamu He in the west (State Seismological Bureau 1993, Fig. 5, p. 129). Upstream from Chouniu Gou, we could follow and crudely map one particularly prominent scarp zone across rugged mountainous terrain on either side of an E-W pass between two north-draining valleys (Figs 16, 18b, c, d, 19a and b). Certain features along that zone resemble those depicted on photographs taken farther east, at Shenjiawo Pou (State Seismological Bureau 1993, photographs 34 and 35, p. 129).
Severe ground disruption starts transverse to the first, eastern valley (Taojung Gou), in the form of a narrow, N100-110E trending stretch of chaotically contorted turf. with exhumed boulders separated by water-seeping lenses of black and grey-green gouge, which climbs up the west side of the valley. Movement appears to have raised thc rocks and ground surface up valley, south of that zone (Fig. 1%). Farther up slope, antithetic scarps with well-preserved free faces up to -1 m high enclose a N120-130E-trending pull-apart depression. about 10m wide and several tens of' metres long, with 1 to 2 m wide and deep collapse pits at either end (Fig. 1%). The zone then narrows into ;I narrower crack-like furrow oriented =NYOE. and continue\ as a single north-facing, NlOO-l IOE-trending. free-faced scarp (Fig. 1%) all the way to the pass, whose distance from the valley floor we estimate to be of the order of 600 m (Fig.   19a). Consistently higher north-facing scarps all along the scarp zone east of the pass, as well as cumulative slope offsets of up to 3.4-3.9 m opposite to the generally south-directed slope gradient (Figs 18b and 19a, inset AB). indicate down-to-the-north throw, which rules out a superficial landslide origin. At a more detailed level, the orientation, in map view, of the pull-apart depression relative to that of adjacent segments of the scarp zone suggests some component of right slip across that zone (Fig.  19a).
At the pass, the north-facing scarp is particularly outstanding, about 50m long and 1 to 1.7 m high, with a fresh free face and a modest debris wedge at its base (Fig.  18c). Sudden slippage has uprooted willow trees or cut and exposed their roots along that free face. The grass-covered morphological saddle of the pass is affected by open fissures roughly parallel to the free face, as well as by smaller, north-facing scarps, implying probably as much as 2 m of down-to-the-north seismic throw and significant extension in a =NNE azimuth, roughly parallel to the local divide (Fig.  19a, inset CD). West of the pass, the zone splays into two main, N-to NE-facing, free-faced scarps (Fig. 19a). The first one, which we did not follow, continues due west across willow thickets. The second one veers to N155E, then back to N130E, running obliquely down slope above the narrow gorge of a small stream, all the way to the confluence of that stream with the river west of the pass (Fig. 19a), a total distance of 4 0 0 m. That second scarp, which is 0.5 to 2 m high with down-to-the-north-east throw, opposite to the local slope gradient, limits a fairly linear and continuous furrow that cuts across basement rock and surface debris (Fig. 18d). One stretch of en Cchelon, paired scarps bounding metres-wide, right-stepping, N120E-trending grabens, indicates a component of left-lateral slip (Fig. 19a). Fresh landslides up slope from the scarp furrow funnel into it, implying that surface debris collapsed into ground fissures at the scarp base (Fig. 19a, inset EF). Short fir trees, with straight vertical trunks, at most several decades old, have grown in the furrow and associated slides.
In summary, the overall attitude of the ~1 . 5 km long scarp zone relative to the topographic relief, the freshness, amount and consistency of the deformation observed across the =300m high mountain spur between the two valleys-with south-side up and opposite components of lateral slip east and west of the pass-all concur to indicate formation, in 1927, by as much as 2 m of surface-slip on a -60"NNE-dipping normal fault, and testify to local, N25 f 15E-directed extension (Fig. 19a). More credit should thus be given to the map compiled by the State however, to accommodate shortening and left-lateral strike Seismological Bureau (Fig. 5, p. 129, 1993) than to other slip (State Seismological Bureau 1993, Fig. 5, p. 129). reports (e.g. Jia 1982;Deng el al. 1984; Rather, we interpret most of them to reflect, as that at 1984). Not only are the scarps that cut the rugged relief just Taojung Gou, extensional strain and normal slip on the north of the Dongqingding range crest of clear tectonic north-dipping Dongqingding fault and related splays (Fig.  origin, but, in the area we could visit, they are the largest 1%). surface ruptures due to the 1927 event. They do not seem, The picture that emerges is therefore one in which thrust estimated (Gongxu et al. 1989) and the moment inferred from spectral densities of mantle waves (Chen & Molnar 1977), should merely be viewed as illustrating the source parameters most consistent with our assessment of the regional tectonics. But an average slip value smaller than 1.5 m, as well as a magnitude significantly greater than 8 should probably be considered unlikely. At the other extreme, assuming a shortening rate of -4 mm yr-' and the recurrence time of a 1927-type event to be on the order of 1600yrs (Gongxu et al. 1989) would require larger average slip (-6 m) on a surface about one-quarter (35 X 60 km2) of the maximum surface above, in rough coincidence with the southernmost epicentral location and the contour of isoseismal X (Figs 1 and 16). Clearly, the limited information at hand does not permit derivation of more than such non-unique, first-order inferences.
slip on the Dongqingding ramp at depth caused incremental growth of the Dongqingding anticline and extension at its apex, with the Dongqingding normal fault accommodating down-to-the-north throw, probably due to flattening of the shallower part of the thrust past the anticlinal crest (Fig.  19b). While the scale here is greater, this interpretation is in line with well-documented occurrences of coseismic normal faulting in thrust earthquakes elsewhere, particularly in Algeria. The only clear surface break produced by the 1954 September 9, M = 6.7, El Asnam event, for instance, was along the south-dipping  (Avouac et al. 1993). If thrust offsets in 1927 had been no greater than those we saw near Wangjiaxin Zhuang, they could easily have gone undetected and for have disappeared in most places in only a few decades, particularly if they formed at the edges of, or within, cultivated fields. It is possible that such was the fate of most segments of the elusive N160"E-striking break reported by Jia (1982), whether it extended along the base of the north Gulang escarpment and farther north or, as shown in a more recent map (State Seismological Bureau 1993, Fig. 7-1, p. 176), 9-10 km west of it, in part across the irrigated foreland and along the western edge of the Huangyang river fan, all the way to the tip of the east Shiyang thrust (Fig. 16). While the actual lengths and throws of either the Dongqingding normal fault zone or the two 'conjugate' segments north of Dongqingding and Gulang (Molnar & Deng 1984) seem too small to have generated an earthquake of magnitude 8 to 8.3, the deformation mechanism in Fig.  19(b) and the fault geometry in Figs 1 and 16 suggest alternative ways to account for the moment and magnitude of the great 1927 event. Taking the shallower part of the Dongqingding thrust on Fig. 19b to dip = 15"S, the -60" dip of the Dongqingding normal fault and the distance between its trace and the range front (=3 km) would imply a ramp to flat bend about 1-1.5 km deep under the range. With =2 m of slip on the Dongqingding normal fault, kinematic compatibility at that 30" bend (e.g.  would require =4 m of thrust slip at depth (Fig. 19b, inset). Assuming an average slip Au = 4 m on the entire surface of the Wuwei-Dongqingding thrust, from =101.3 to =102.9"E and between the topographic front and the branch line with the Haiyuan-Gulang Fault (Lw = 8.5 X lo3 km2, Figs 1 and 161, with a shear modulus p = 3.3 X 10" N m-*, would yield a seismic moment M, = pLwAu = 11.2 X 10' ' N m, consistent with a magnitude M,,, -8.0 (e.g. Kanamori 1977). If less than the north-eastern half of that thrust surface (Lw = 3.3 X lo3 km'), west of the Shiyang river, had slipped by the same amount, M, and M , would be ~4 . 3 X lo2' N m and =7:, respectively. These numbers, which bring forth the discrepancy left between even the minimum magnitude

SUMMARY, DISCUSSION A N D IMPLICATIONS
Deep fault architecture, slip partitioning and regional kinematics.

Motion of NE Tibet relative to SE Gobi
Recent deformation near the eastern extremity of the Qilian Shan involves several fault zones with different strain styles. Our study helps understand how such strains are both partitioned and linked with one another, and combine into a coherent regional picture than can be related to crustal movements in adjacent regions, within the framework of the India-Asia collision. The two main components of deformation are ESE-striking, left-lateral faulting, and NNE-directed shortening. Qualitatively, this has long been known (e.g. Tapponnier 82 , and is in keeping with the situation found in the western Qilian Shan (e.g. Peltzer et al. 1988;Meyer 1991), or east of the Yellow River (e.g. Burchfiel et al. 1989;Zhang et al. 1990). But it is now possible to better assess the relative importance of either process.
At a large scale and regardless of the sense of horizontal slip, the tectonics of the area where the Haiyuan and Gulang Faults merge and veer 15-20' as they enter the Lenglong Ling mountains resembles that of the 'big bend' region east and north of Los Angeles, in California, where the San Andreas and San Jacinto Faults converge as they bend along the San Gabriel Mountains and central Transverse Ranges (e.g. Hill & Dibblee 1953;Allen 1981;Weldon, Meisling & Alexander 1993). It is the change of azimuth of the strike-slip faults that seems to be the chief cause of more prominent overthrusting and mountain building within either of those restraining bends (Figs 1 and 20), (e.g. Anderson 1971;Allen 1981;Jones el al. 1990).
At a more detailed level, each of the three main faults of the eastern Qilian Shan bears distinctive characters. The western Haiyuan Fault is the largest and fastest slipping. While predominantly left lateral, it shows consistent normal downthrow, either to the south or north, east of about 102.5"E (Figs 1, 2, 10, 11). Such secondary throw seems to be less in the Lenglong Ling range, west of the junction with the Gulang Fault and south of the Shiyang-Dongqingding thrust, where fully developed. East of 102SoE, the main active strand of the fault tends to step systematically northwards, resulting in the formation of asymmetric pull-apart basins of various sizes, with often prominent north-west boundary faults-the largest of which at Tianzhu-comparable to those segmenting the fault east of the Yellow River (e.g. Saltlake and Laohuyaoxian basins, Deng et al. 1986;Burchfiel et al. 1991;State Seismological Bureau 1990). The presently secondary strands of the western Haiyuan Fault show some degree of activity, however (Figs 2, 7 and 11). By contrast with those pull-apart steps, the highest, isolated mountains along the fault east of Laohu Shan (Hasi Shan, 3021 m and Nanhua Shan, 2969 m) seem to reflect contemporaneous shortening and thrusting within two clockwise restraining bends of the fault zone (Fig. 1).
The present-day slip rate on the Gulang Fault is less than half that on the western Haiyuan Fault. The secondary component of motion on the Culang Fault appears to be =N-directed thrusting, which has contributed to the formation of two distinct push-up mountains (Fig. 1). east of the north Gulang escarpment. That the Gulang Fault merges eastwards with the Tiangjing thrust further supports the inference that it acts as a lateral thrust ramp. Along either the Haiyuan or Gulang strike-slip faults, guidance by the early Palaeozoic fabric andlor by deep, large-scale boudinage such as that typifying the exhumed Red River-Ailao Shan shear one in Yunnan (Leloup er al. 1993) may account for the parallel surface strands separating lense-shaped crustal slivers (along Maomao Shan. near and east of Hasi Shan, and south of Haxi syncline. Figs 1, 2, 11, 16 and 17).
The en echelon, right-stepping Shiyang-Dongqingding thrusts provide a good example of fairly thick-skinned shortening across ancient basement, a widespread contemporary deformation process in the former peneplains of central Asia (Tapponnier et al. 1990: Meyer 1991: Avouac et al. 1993. The origin of those thrusts, and of the resulting Lenglong Ling crustal wedge or flake (Fig. 16) may be accounted for by the following scenario. Greater regional shortening west of 103"E probably led to the NNE propagation of a 10-20"s-dipping dicollement splaying from the Haiyuan Fault at a depth of =25 km. Ramping, probably in sequence, of two equally spaced, steeply south-dipping thrusts from that dkcollement to the surface subsequently led to warping of the upper crust into two large fault-propagation anticlines (Fig. 16c). Contemporaneous folding of Holocene deposits =10km north of the topographic front implies further north-directed propagation of the dicollement into the foreland. in keeping with the inference that the thrust system evolved mostly as a leading imbricate thrust fan. Our observations are insufficient to decide whether the three principal thrust segments between 102 and 103"E are bounded to the east by lateral tip lines or ramps. But the north Gulang escarpment, if active, might represent the emergence of a right-lateral thrust ramp, and another such ramp, with even more modest expression in the cumulative relief might account for the N160E-striking, 1927 scarps reported west of the Huangyang river, if real (Fig. 16b). Finally, in spite of the predominant NNE shortening regime west of 103"E, subordinate, ESE-striking normal faults apparently contribute to accommodate hanging-wall warping in anticlinal flanks (Figs 16,18 and 19). Whether of range-back or compatibility type (Figs 16c  and 19), we surmise that such normal faults, which often reactivate inherited discontinuities of the crust, assist localized mountain building by restricting it to narrower ridges (e.g. Tapponnier 1993).
We infer both the 'controversial' complexity of the surface breaks of the 1927 earthquake, due in part to prominent, hitherto poorly understood, hanging-wall extension, and its modest thrust surface signature to be typical characteristics of great thrust events rupturing relatively deep ramps and shallow-dipping dkcollements under and near high mountains, in Asia and elsewhere. Our study of active faulting in the source area of the M = 8, 1906 Manas earthquake in the Dzungarian Tien Shan (Avouac et al. 1993) points towards the same conclusion. While involving thicker layers of crust, the strain style of large thrust earthquakes would thus generally resemble that of smaller events (e.g. Stein & King 1984). We suspect such style and characteristics to account for the 'mystery' that still shrouds the great Himalayan earthquakes of 1897, 1905, 1934, and 1950, or the 1902 and 1911 earthquakes in the Tien Shan. for which suitable surface-rupture maps are not available. Though complexly segmented, the =150km long break of the 1932, M=7.6, Changma earthquake probably owes its clearer expression to oblique slip on generally steeper fault planes (Peltzer et al. 1988: Meyer 1991. The fact that the diverse faults of the eastern Qilian Shan all show evidence of coeval long-term slip. the detailed geometry of their traces, which bifurcate or step in map view, and the changes in strain styles. reflected in the cumulative relief, that coincide with such steps or bifurcations, imply that these faults connect at depth. The connections shown on the maps and section of Figs 1 and 16 are represented schematically in three dimensions on Figs 21(a) and (b). The architecture of the entire fault system resembles, in cross-section, that of a tree, with deep trunk and shallow branches. While quite complex at shallow crustal levels, it becomes simpler in the lower crust, with shear restricted to a narrower zone, a localization that probably continues in the upper mantle underneath. The fairly deep, active dkcollement provides an essential link between the most distant, emergent branches (see also Avouac et al. 1993).
At a grander scale to the west, the presence of an even deeper dkcollement in the mid-lower crust appears to account best for the contemporaneous growth of the parallel ranges of the central Qilian Shan (Tapponnier et al. 1990: Meyer 1991. To the east, it is probable that the Tiangjing, Mibo, and Yangtong Shan thrusts, which seem to reflect thin-skinned shortening (Zhang et al. 1990), mark the emergence of a relatively shallow dkcollement branching off the Haiyuan Fault one as shown in Fig. 21. We hypothesize deep connection between faults, or at least between localized, narrow shear zones in the Qilian Shan to be plausible because the crust, which is fairly old, and hence cold, thickens. Recent, relatively rapid thickening of cold crust ought to deepen further an already deep brittle-ductile transition, allowing for faulting to extend farther down, hence for greater earthquakes than is usual. Within a system like that depicted in Fig. 21, fault junctions are not stable. Even with steady-state boundary conditions at the regional scale, such a system will thus behave as a very dynamic feature, breeding short-lived jogs and splays, block rotation, and eventually slip inversion on steep ramps or high-angle strands (e.g. Peltzer et af. 1988). Such unstability probably suffices to account for the existence of the transient push-up mountains, pull-apart basins (with typical life times on the order of a few 10' yrs, given the centrimetric slip rate: e.g. Burchfiel et al. 1991), and shortening perpendicular normal faults observed.
A corollary of the geometry outlined in Fig. 21 and of the underlying inference that the whole 'fault tree' behaves mechanically as one system is that deep, oblique slip on the 'trunk fault' ought to be partitioned into dominantly strike-perpendicular and strike-parallel slip among the shallower branches (Armijo et af. 1986). The first-order rates obtained on the three principal faults may be used to test quantitatively the reality of such partitioning and to constrain the regional kinematics (Fig. 22). Combining the 11 f 4 mm yr-' slip, in the N112.5 f 2.5E direction, on the Haiyuan Fault with the 4.3k2.1 mmyr-' slip, in the N92 f 7E direction, on the Gulang Fault yields a total of 15.1 f 6.0 mm yr-' of left-lateral motion, in the N107 -f 4E direction (Fig. 22, top), between the north-easternmost part of Tibet (region of Lanzhou and Xining) and the southern part of the Gobi-Ala Shan platform (Tengger desert). Projecting the latter vector on the more southerly average strike (N119f5E) of the Haiyuan Fault in the Lenglong Ling, between 101.2 and 102"E, west of the Gulang-Haiyuan junction, yields 14.8 f 5.8 mm yr-' of strikeparallel, left-lateral slip on that high mountain segment of the fault and 3.3 f 1.7 mm yr-' of strike-perpendicular (N29 f 5E) shortening (Fig. 22, bottom). The first number is consistent with the rate (15 f 5 mm yr-') implied by the offsets of the glacial valley edges visible on the SPOT image of Fig. 2(b). The second vector is comparable to the shortening rate (4 f 2 mm yr-l, N25 f 10E) inferred from the structure and Pleistocene uplift (Fig. 22, bottom). The error ellipses are too large to allow a discussion of the misfit (=0.7 mm yr-', zN07E) between the two shortening rates but, if real, it might reflect a minor component of roughly strike-perpendicular shortening along the Gulang Fault, in keeping with the interpretation that this fault is a lateral thrust ramp. We conclude that, to the first order, slip partitioning accounts well for the regional strain observed and that the postglacial and current rate of overall left-lateral motion between north-eastern Tibet and the south-western Gobi is probably about 15 f 6 mm yr-' in the N105 f 6E direction (Fig. 22).
The results of our study of Cenozoic movements in the eastern Qilian Shan differ from those of similar studies in regions farther east. The horizontal slip rates reported by others on the Haiyuan Fault east of Jingtai (Fig. 1) are generally smaller than the one we obtain at Sange Dun \I/ \ . Strike-perpendicular rib on western (11 f 4 mmyr-'). Liu & Zhou (1986), for instance, infer from offset Holocene ridges and creeks an average long-term slip rate of -5 mm yr-', less than half the Sange Dun value. Their basis for estimating the ages of such offsets, however, is unclear. From more precise fieldwork on the Xihua-Nanhua Shan segment of the fault west of Haiyuan (Fig. l), Zhang et al. (1988b) derive a left-slip rate of 8 f 2 mm yr-' (Fig. 15a). Given the uncertainties, the Sange Dun value might be viewed as more consistent with this latter rate, but it still exceeds this rate by about 35 per cent. This discrepancy may be accounted for in various ways. First, as clearly stated by Zhang et al. (1988b), the work near Haiyuan, which is based on offsets of streams that incise C,,-dated terraces, only provides lower bounds on the slip rate and, while =7 mm yr-I is the bound deemed most plausible from measurements at two preferred sites (Shaomayin and Yehupo), two other bounds may be taken to imply rates in excess of 10mmyr-' (Fig. 15a).
Secondly, the Xihua-Nanhua segment of the Haiyuan Fault strikes =N125E on average, 215" more southerly than at Sange Dun, and lies about 250 km to the east (Fig. 1). Hence, both the geographical separation and the change of strike might account for a fraction of the difference in rates. Thirdly and probably more important, fault slip south of the Xihua-Nanhua segment, on strands that reportedly did not rupture in the M -8.7, 1920 Haiyuan earthquake, could add a few millimetres to the long-term rate documented at Shaomayin and Yehupo. Though less clear where they cross loess-blanketed landforms, the traces of such fault strands, which are discernable on Landsat images (e.g. Tapponnier & Molnar 1977, Fig. 5) and quite sharp on panchromatic SPOT scenes KJ 259-276, 260-277, 261 -277, stretch along two different zones (Fig. 1). Some, mostly outside the area mapped by Burchfiel ef ai. (1991) or Deng et al. (1986) in the field, bound both Nanhua and Xihua Shan to the south-west. Others hug the SW flanks of Quwu and Yueliang Shan. In view of their orientation relative to N105 f 6E, most of these presently subordinate fault strands ought to move obliquely, with components of both SW-directed thrust (Burchfield et al. 1991;State Seismological Bureau 1990) and left-lateral slip, possibly contributing to the rise of the four mountains, which form distinct highs, as push-up folds along the Haiyuan system. Last, and even though the high mountain site on the single-stranded fault at Sange Dun is remarkable, as are the Wallace and Cajon creeks sites on the San Andreas Fault (Sieh & Jahns 1984;Weldon & Sieh 1985), one ought to recall the strengths and weaknesses of the studies at stake here, and ponder the different results in perspective. Our approach relies on a novel, comprehensive scrutiny of the entire Qilian-Haiyuan Fault system and on analysis of the self-consistency of offsets over great distances, a goal unattainable without the help of SPOT images, but on debatable inferences concerning the relationship between regional morphology and globally dated climate changes. In contrast, Zhang et a1.k approach (1988b), while limited by locally focused, if detailed, geological mapping, rests firmly on C,, dates. Hence, at this stage, it is probably wiser to emphasize the first-order consistency of the numbers derived at two distant places with such different approaches than to speculate on the difference between them.
The conclusions we reach regarding the date of onset of Cenozoic deformation, the kinematic regime between that date and the present, and the amounts of finite Cenozoic movement in the eastern Qilian Shan, on the other hand, are much more difficult to reconcile with those derived from fieldwork in Ningxia Hui Autonomous Region, east of the Yellow River. There, Burchfiel et al. (1989, see also Deng et al. 1986State Seismological Bureau 1990;Burchfiel et a/. 1991;Zhang et at. 1991) infer that Cenozoic regional deformation started only in the Late Pliocene, and that a =90° clockwise swing of the regional shortening direction (from =NE to =SE) in the Early Pleistocene triggered a radical shift in tectonic style, from shortening along NW trending thrusts and folds, to left slip on the Haiyuan Fault. That fault is thus interpreted to have originated only 1.6-1.8MaB.P., and to cut the thrusts and folds that it strictly postdates.
Not only do we find that the evidence discussed by the above authors does not require the interpretation they propose but there is little sign, if any, for such drastic change and late onset of deformation west of the Yellow River. Instead, in the eastern Qilian Shan, contemporaneous thrusting and strike-slip faulting take place in adjacent areas and we find no suggestion in the geological record that such a situation has not been going on unabated for at least several million years, with slip-partitioning accounting for current rates as well as for most of the finite movements. The simple interpretation that the Shiyang-Dongqingding thrusts result from compatibility shortening in a large restraining bend of the Haiyuan Fault (Figs 20-22) implies that both structures moved in tune during much of their recent history. We also find it difficult to account for the finite displacements on either the thrusts or the strike-slip faults if motion on them lasted less than 4-6 Ma. Assuming a constant, maximum shortening rate of 6mmyr-', for instance, more than 4Ma would have been necessary to absorb a minimum finite shortening of 25 km across the mountains north of the Haiyuan Fault. Similarly, with a constant left-slip rate of 11 mm yr-', more than 3.7 Ma or 7.3 Ma, depending on the minimum offset value chosen (40 or 80 km), would be required to form the great bend of the Yellow River across that fault (Fig. 15b). If the latter rate had only been 8mmyr-' (Zhang et al. 1988b), those time-spans would increase to =5 and -10 Ma, respectively. Our results are thus best interpreted to imply that shortening in the eastern Qilian Shan and left slip on the Haiyuan Fault started sometime in the late Miocene, consistent with the regional stratigraphy, rather than in the late Pliocene. Assuming partitioning of slip at constant rates on the Haiyuan Fault and eastern Qilian Shun thrust> ( and 11 mm yr-', Figs 21 and 22) since 7-X Ma B.P. NOUILI yield cumulative sinistral offset and shor:ening of 7 i --88 k m . and 28-32 km, respectively. Adding cumulative slip on t he Gulang Fault during that same period to that on thc Haiyuan Fault would bring the total east-south-eastu;\rJ\ displacement of NE Tibet relative to the SE Gohi to ;it Iea\t 100-120 km.
That number contrasts with the small finite left-lateral \lip inferred east of the Yellow River. By matching occurrence\ of rocks of different ages on either side of a 60 k m 4tretch 01 the 600-700km long fault, Burchtiel c' / nl. ( I Y N . I Y Y I ) .
Zhang ef a/. (1990,1991), and Deng et crl. (1YX6) h a k c concluded that the total displacement on the Haiyuan Fault could not exceed 16 km (10.5-15.5 km). quite a small valuc for a fault of that length. While wt' do not question rhc quality of their detailed mapping, we note that thc. otf\ct.s they document concern only one segment o f the presentl! active strand of the Haiyuan Fault zone. Besides, hecause visible outcrops north of that segment are small. close t o it and extensively surrounded by a blanket o f late Pleistocene ( Q 3 ) loess, the offsets of rocks or contacts older than liite Pleistocene might be underestimated. The length of tault they studied is in fact too short to rule out the exktcnce of offsets greater than 60 km, particularly in view of the observation that Precambrian and Palaeozoic rocks similar to those used to constrain the larger offsets extend at least t o Hasi Shan, about 100 kmNE or' Nanhua Shan. outside the Ningxia Hui Autonomous Region (Gansu Geological Bureau 1975). Thus, until a different, albeit convincing. explanation of the outstanding left-lateral deviation of the greatest river of north China across the Haiyuan Fault zone is found, and until mapping of offsets along that zone is completed at the appropriate scale. we consider it more likely that NE Tibet moved 100-150 rather than =16 km relative to the SE Gobi, and choose to interpret the 'total' offset in Ningxia as possibly indicative of the initiation of slip on the active Xihua-Nanhua fault strand (Fig. 1%).
Exactly how such a relatively large displacement might have been transferred to blocks farther east and south remains a target for future studies, but it could not have been entirely absorbed by crustal thickening in the modest, 52900 m high, Madong, Liupan and Xiaoguan mountains. where less than 18 km of shortening is reported (Zhang el ul. 1991). Hence, while some of that displacement probably rifted and rotated the Ordos block away from the Gobi-Ala Shan platform, the greater part of it ought to have contributed to push the SE China craton towards the ESE or SE, tearing it away from the rest of north-eastern Asia (Tapponnier et a/. 1982;Peltzer et al. 1985;Armijo et ril. 1986Armijo et ril. , 1989Peltzer & Tapponnier 1988;Zhang, Vergeiy & Mercier 1994).

Seismic hazard in NE Gansu and comparable regions
While several aspects of the mechanism of the great, 1927 Gulang earthquake remain obscure, one robust conclusion of our study is that this earthquake ruptured the Shiyang-Dongqingding thrusts, not the Gulang or Haiyuan Faults. Moreover, neither the Gulang nor the Haiyuan Fault bears field evidence of seismic dislocation in the last few centuries at this longitude. This is particularly clear for the Haiyuan Fault, whose trace we inspected in detail at several localities. Whether near Sange Dun, Tianzhu or Song Shan. the smoothed-out scarps or pressure ridges that mark that trace are suggestive of infrequent, large earthquakes. That the walled city of Song Shan, located only 2 km south of the fault and built during the Ming Dynasty, remained intact until 1990 AD implies that no such earthquake has occurred since at least 1644 AD. a period of 350 years.
Additional inferences on the seismic behaviour of the western Haiyuan Fault may be derived from the Catalogue of Chinese Earthquakes (Gongxu et al. 1989). Since the great 1920 and 1927 earthquakes resulted in intensities of VII, V1, VI and VI, VIII, V, in Lanzhou, Wuwei, and Xining, respectively, comparable intensities would probably characterize a Iarge event on the western Haiyuan Fault.
Relatively few seismic events assigned peak intensities ZVI are recorded in the catalogue within the region encompassing those three ancient cities. Before assessing the value of catalogue entries, whether in terms of location or estimated intensities, it must be recalled that very little is known of the seismicity of Tibet and Qinghai, immediately south-west of Gansu province, prior to the advent of instrumental data (~1 9 0 0 AD), implying that this province always formed a frontier between nomad and sedentary ways of life. Thus, the catalogue is probably quite incomplete, and its entries must depend strongly on critical stages that have marked the history of the Chinese civilisation and empire, of its westwards conquests, and of population growth in settlements along and west of the upper reaches of the Yellow River, within lands of ancestral nomadic culture. Foremost among those stages are the establishment of the Qin dynasty (231 BC), the onset of sustained commercial exchange along the silk road (=50BC), the successful and rapid expansion of Buddhism (=400 AD, Northern Wei dynasty), then Islam (~7 0 0 AD, Tang dynasty), and concurrent establishment of permanent centres of culture with durable architecture-none of the Mingoi (Buddhist cave monasteries) in Gansu and Ningxia, for instance, predates 400BC, and much of the presently visible remains of the great wall in NE Gansu date from about 600AD (Sui dynasty). In what has how become Gansu province, as in most other parts of China, there is in fact no earthquake record before the foundation of the Qin dynasty in 231 BC, 2225 years ago (Gongxu et al. 1989).
The cities of Lanzhou (formerly Jingcheng), Wuwei (formerly Liangzhou) and Xining are mentioned for the first time in the catalogue in 138 A D and 406 AD, 143 AD, and 318 AD, respectively, with the two oldest of these four events (peak intensities IX), as well as four more ancient ones, hitting more strongly towns located south-east of the Yellow River (Lintao, Gangu, Longxi, Wudu). Hence, we regard it likely that many events in NE Gansu, including large or very large earthquakes in the mountains roughly midway between those three cities, were not recorded prior to about 1600-1700 years ago. Since then, periods of war or important dynastic changes could have resulted in significant loss of information, and we doubt that location or intensity assessment in that part of China, even for such large events, becomes acceptable prior to the Mongol conquest (1206 AD) and its subsequent consolidation into the Yuan (1271 AD) and Ming dynasties (1368 AD). It must be noted, for example, that the first peak intensity of XI ( M = 8 ) , reported in the catalogue is that of the 1303 Linfen earthquake, in the much more central and populated province of Shanxi, to the east.
Hence, we infer that a large earthquake on the western Haiyuan Fault would probably have been located and identified as such in the last ~8 0 0 years, might have been felt in certain cities but assigned incorrect location and size in the preceding 400-900 years, and would probably have been missed altogether prior to about 300 AD. With this in mind, shocks assigned intensities of VJ to VII felt repeatedly in Wuwei in the fourth, sixth and eleventh centuries AD (362, 366, 374; 506, 575; 1092) (Gongxu et al. 1989;State Seismological Bureau 1991), closest to either the Shiyang-Dongqingding thrusts, or the Gulang and western Haiyuan Faults, might have been great events rupturing either strike-slip fault or, since apparently none of these shocks was felt in Lanzhou, 'smaller' ones on the nearer thrusts. The major historical event felt in and around Lanzhou in 1125AD (Z=IX, M=7.2?, Fig. l), whose location is uncertain (Gongxu et al. 1989), might have activated the Haiyuan Fault, but the fact that it was not felt in Wuwei may be taken to imply that it activated a smaller fault south of Lanzhou (Fig. 1). Similarly, the two shocks felt in 1440 at and south of Yongden (VIII, VII), but neither in Lanzhou nor Wuwei, more likely activated a fault within 30-40 km of that city, probably to the south-west. The I=VIII, ( M = 6.5?) 1888 February 1 earthquake near Jingtai, on the other hand, which is felt nowhere else, may be safely inferred to be a moderate-size event on one strand of the Haiyuan Fault. New field evidence (Liu Baichi, personnal communication, 1992), in fact suggests that this event ruptured a 20-30 km long stretch of the main fault strand in the Lao Hu Shan, =20km east of Song Shan, with 1.5-2m of maximum left slip. While this discussion mostly emphasizes how little we know of the pre-instrumental seismicity of NE Gansu, it does suggest that, short of gross omissions in the catalogue, no large earthquake occurred along most of the lengths of the Gulang and western Haiyuan Faults during the last 800 years at least, in keeping with the qualitative impression gained in the field. Barring the occurrence of fault creep, for which there is no evidence at this time, both faults thus ought to be singled out as likely sites for such earthquakes in the future.
The western Haiyuan Fault, in particular, whose left-slip rate is fairly fast, looms as the foremost seismic-hazard source to the region. Such hazard is all the more serious that great earthquakes, rupturing several discrete segments, with millenal return intervals, are not rare on comparable fault zones in north-eastern China. Peltzer ef al. (1988), for instance, estimated a recurrence interval between 900 and 1900 years for events comparable to the M=7.6, 1932 earthquake, which ruptured at least five distinct segments of the Changma Fault zone, in the north-western Qilian Shan (Meyer 1991). Zhang et al. (1987) derived a recurrence interval of 800 to 1600 years for events similar to the M = 8.7, 1920 earthquake, which ruptured eight segments, separated by several kilometres wide pull-aparts, of the eastern Haiyuan Fault (Deng et al. 1986). It would thus be no surprise if the western stretch of that fault behaved in a similar way.
At a more detailed level, a rate of =ll mm yr-' would imply more than 9 m of cumulative left slip along the western Haiyuan Fault in over 800 years. The main geometric features that may be inferred to segment the fault are the Gulang-Haiyuan junction, the Tianzhu pull-apart basin, and the Song Shan junction and pull-apart basin. Independent rupture of the three western segments, whose lengths are respectively =90, -50 and -60 km, to depths of =20km, each with =6m of average slip, would generate three earthquakes with M,,, 2 7.5. If the entire length of the fault (~2 2 0 km) west of the site of the 1888 event ruptured to =20 km, with -10 m of slip in a single earthquake, then a M , 2 8 event (M, = 1.4 X lo2' N m), comparable to the 1920, Haiyuan earthquake would ensue. Although alternative scenarios are possible, we find the latter more plausible in view of the usually accepted scaling law between average slip and rupture length for intraplate earthquakes (e.g. Scholz, Aviles & Wesnousky 1986), and of the examples of the 1920, 1927 and 1932 earthquakes.
That the Qilian-Haiyuan Fault system, seen as a whole (Fig. 21), generated the 1888, 1920 and 1927 earthquakes, the last two only seven years apart, after a long period of quiescence, implies that it re-entered an active phase of its cycle about 100 yrs ago. That two events with M > 5.5, the strongest since 1928, have taken place since 1985 near either extremity of the apparently quiescent, 220 km long, western stretch of fault-which we call the Tianzhu seismic gap- (Figs 1 and 21a), raises even greater concern. The first one (1986 August 26; M =5.9), which struck the region north of Menyuan, is located either =20 km north of the trace of the Haiyuan Fault (centroid-moment-tensor inversion, Dziewonski et al. 1987), or a few kilometres south of it (broad-band body-wave phases, Ekstom 1987) (Figs 1  and 21a). Its focal mechanism is compatible with =NE-directed thrusting on =SE-striking planes ( Fig. 1) (Dziewonski et al. 1987;; see also, Molnar & Lyon-Caen 1989Molnar 1990). The second one (1990 October 2 0 M = 5.8) occurred in the Lao Hu Shan, =20 km east of Song Shan (Figs 1 and 21a), where it caused significant damage; segments of the Ming walls, for instance, collapsed. Its focal mechanism is consistent with pure left-lateral slip on a vertical, =NlOOE-striking plane (Dziewonski et al. 1991), corresponding either to the main strand of the Haiyuan Fault or to that north of it (Fig. 1). Both the 1986 and the 1990 events, therefore, apparently ruptured faults belonging to the Qilian-Haiyuan system as defined on Fig. 21. We conclude that the likelihood of impending large or great earthquakes on the western Haiyuan Fault, between Song Shan and Menyuan, only =100km north of the most densely populated region of west-central China, is such that extensive instrumental monitoring of that part of the fault should be implemented at once.
Finally, to those of us who study great earthquakes and active faults in China, the urge for a comparison with southern California comes to mind once more. The geometrical resemblance brought out by Fig. 20 may be taken to imply simple kinematic and mechanical similarities.
For instance, given the combined slip rate (=42 mm yr-') on the San Andreas and San Jacinto Faults, south of the San Bernardino Mountains (Weldon & Sieh 1985;Weldon & Humphreys 1986;Weldon et al. 1993), and the angle that they make with the Mojave segment of the San Andreas Fault, between Cajon and Tejon Passes (=lSo), a vector diagram similar to that in Fig. 22 would imply as much as 13 mm yrC' of strike-perpendicular (SSW) shortening along that segment, much of which on the thrusts south of the San Gabriel Mountains. This would be consistent with the recent revision, based on GPS measurements, of shortening rates south of the fault (11 f 3 mm yr-': e.g. Donnellan. Hager & King 1995;.
The similarities should not take precedence over thc differences, however, which are significant. While the Haiyuan and Gulang Faults straddle the north-eastern edge of the Tibet plateau, the San Andreas and San Jacinto Faults extend south-eastwards into the extensional, Gulf of California, plate boundary. Boundary-movement rates in California (Minster & Jordan 1984;DeMets rf ul. 1990) are more than three times greater than in Gansu. More deformation occurs adjacent to the San Andreas Fault zone (e.g. Feigl ef al. 1993). Horizontal slip at rates of 7-8mmyr-' takes place north-east of the San Andreas Fault, along the Garlock Fault (e.g. McGill & Sieh 1993) and the east Mojave shear zone, recently activated by the 1992 June 28, M , = 7.3, Landers earthquake (e.g. Sieh ef a / . 1993). South-west of the San Andreas Fault, more dextral motion of crustal slivers along the Elsinore. Palos Verdes and other offshore faults contributes to make the central and western transverse Range thrusts more complex and to load them differently (e.g. Jones et al. 1990;Hauksson & Jones 1991) with increasing amount of shortening westwards (e.g. Yeats 1988). Many of those thrusts remain oblique to the Mojave segment of the San Andreas, which accounts for the sinistral component of slip on them (Fig. 20, top) and further kinematic complication arises from the clockwise block rotations thus induced (Jackson & Molnar 1990). The crust in eastern Gansu is thicker than in southern California. and has a different structure. Hence, although the relief in both regions is comparable, the elevations are not. Last, the thickness of recent, poorly consolidated sediments appears to be much greater in southern California, which probably implies more widespread occurrence of flat dkcollements at various depths (e.g. Yeats 1983: Davis, Nanson & Yerkes 1989. But the resemblance illustrated in Fig. 20, and its implications regarding 3-D geometry and kinematics (Figs 21 and 22), adds weight to simple inferences. For instance. even at the hundred kilometre scale, in California or China, earthquake coupling and fault interaction (e.g. Stein, King & Lin 1992) should probably be considered to coinvolve, in three dimensions, all active faults, whether thrust, normal or strike slip. Also, as argued elsewhere (e.g. Hauksson & Jones 1991;Avouac et al. 1993), nucleation of the greatest thrust earthquake in the Los Angeles region deep under the highest and steepest relief, with propagation to a number of apparently discrete, shallower ramp segments and tear faults, as illustrated by the examples of the Changma (e.g. Peltzer et al. 1988;Meyer 1991), Manas (Avouac et af. 1993), and Gulang earthquakes (Fig. 20, bottom), ought not to be discarded as a remote possibility.

ACKNOWLEDGMENTS
This study is part of a long-term cooperation program between the Institut National des Sciences de I'Univers and