We found eight similar earthquakes of M4.8 regularly occurring from 1957 to 1995 on the plate boundary off Sanriku, northeastern Japan. The mean recurrence interval of the events is 5.35 years and its standard deviation is only 0.53 years. These events have the same focal mechanisms of low-angle thrust fault type and their source sizes are estimated to be around 1 km. The slip of each event is comparable to the cumulative amount of relative plate motion in the inter-seismic period. Thus, these events are probably caused by repeating slip at a small but strong asperity surrounded by creeping zone on the plate boundary. The next event was expected to occur by the end of November 2001 with 99% probability and actually M4.7 event occurred on November 13, 2001.
 Characteristic earthquake model with a nearly constant recurrence interval time has been used in the long-term earthquake prediction [e.g., Shimazaki and Nakata, 1980]. However, the nature of the characteristic events is still vague and the model is still debatable [e.g., Kagan, 1993]. Numerical simulations based on the recent rate- and state-dependent friction laws predict that the whole coupling area surrounded by steady-slip region on the plate boundary will slip at the same time and the recurrence interval is almost constant [Tse and Rice, 1986; Kato and Hirasawa, 1997]. The simulations, however, are carried out assuming homogeneously coupled area. On the real plate boundary, the coupling must be very heterogeneous spatially and the recurrence interval will be affected by slips on the nearby segments. However, if the coupling is strong only on the small asperities sparsely distributed on the plate boundary, the interaction among the asperities would be negligible. This would be the case in a creep segment where the plate boundary is mostly decoupled and only small earthquakes can occur on small asperities. Actually, characteristic small earthquake repeaters have been found in Stone Canyon [Ellsworth, 1995] and Parkfield [Nadeau and McEvilly, 1997], California. If the observed repeaters in these areas can be explained by the model mentioned above, such repeaters must be found in other creep segments in the world.
 Far off Sanriku, northeastern Japan, is one of the most seismically active regions in the world. There occurred many large interpalte earthquakes close to the Japan trench. However, there are no historical records of earthquakes with magnitudes six or larger along the plate boundary area close to the seashore between 39°N and 40°N, although microearthquake activity is very high there (Figure 1). From the GPS data analyses, Nishimura et al.  pointed out that the area may be a velocity-strengthening region. Thus, the plate boundary in the area is probably creeping just the same as the creep segment in the San Andreas fault.
 If the boundary is mostly creeping in this region, it is expected that clusters consisting of characteristic repeaters occur there just the same as Parkfield. In this paper, we searched hypocenter catalogues for the characteristic events, and investigated the nature of the events. Note that when we refer the magnitude of some event explicitly, we show the magnitude determined by JMA (Japan Meteorological Agency) through this paper .
2. Characteristic Earthquake Sequence
 We systematically investigated magnitude-time (M-T) distribution for events in the earthquake clusters off Sanriku using microearthquake catalogue by the Tohoku University (TU) seismological network, which contains the hypocenter data for the period from the middle of April 1975. As a result, we found one cluster where four M4.8 events have occurred with almost constant intervals. The location of the cluster is about 10 km away from the seashore as indicated by an arrow in Figures 1b and 1c and its focal depth is about 50 km. Note that the cluster is isolated from other large clusters indicating that the interaction among the clusters will be very small.
 In order to investigate the activity before 1975, we searched JMA catalogue, which contains the hypocenter data for the period from 1926. By comparing the hypocenters determined by TU and JMA, we found that the discrepancies in the hypocenter locations between the two catalogues are sometimes larger than 10 km for the period before 1980. Thus, we relocated hypocenters of older events with M ≥ 4.5 using the same velocity structure as that utilized in the routine hypocenter determination at TU [Hasegawa et al., 1978].
 In order to avoid the hypocenter location bias caused by different station set, it is desirable to use the homogeneous station method [Ansell and Smith, 1975]. However, JMA stations commonly used in the hypocenter determination for the older events are too few to apply the homogeneous station method. Thus, we adopted a simple technique using only the six stations closest to each event. The comparison of the original and relocated hypocenters is shown in Figure 2. The estimation errors (standard deviations) of relocated events are less than 8 km. Four events are concentrated to the location of the cluster shown in Figure 1c and there are no other ambiguous earthquakes near the cluster. Thus, we judged that these four events located in the rectangle shown in Figure 2 also belong to the cluster.
Figure 3 shows (a) M-T diagram and (b) cumulative moment for the events located in the rectangles shown in Figure 1c and Figure 2. The M-T diagram indicates that M4.8–4.9 events have regularly occurred in the cluster. Although the 1962 and 1968 events have magnitudes of 4.9, we will call these M4.8–4.9 events simply as ‘M4.8 events’ hereafter. The mean recurrence interval calculated from the eight M4.8 events is 5.35 years with standard deviation of 0.53 years. Thus, fluctuation of the interval is only 10% of the mean value. In the calculation of seismic moment (Mo; dyne.cm) shown in Figure 3b, we estimated the moment from the magnitude using the relation [Aki, 1972] as:
where M denotes the magnitude of an event. Figure 3b indicates that the moment release has been controlled by M4.8 events and other small events are negligible. Although we did not relocate old events smaller than M4.5, such small events will not largely change the pattern shown in Figure 3b even if they exist. We could not identify M4.8 events before 1957 in JMA catalogue. However, the completeness of the JMA catalogue for earthquakes smaller than M5 is questionable for the period before the middle of 1960s [Ishikawa, 1987]. Thus, it is uncertain when this ‘characteristic’ activity actually started.
 We estimated focal mechanisms for the latest four characteristic events using P-wave initial motion data of TU and JMA; it was impossible to determine the mechanisms of older events because they had very few polarity data. All the recent four events have similar focal mechanisms of low-angle thrust fault type indicating that these events occurred on the plate boundary. The estimated mechanisms have fault planes with strike of about 205 degrees and dip angle of about 30 degrees.
 We also checked the waveform similarity for the events using both JMA and TU data. For JMA data, we checked smoke paper records for these eight events recorded at JMA Sendai station (about 160 km southwest away from the cluster) and found that all the events show almost the same waveforms and their amplitude fluctuations are within 20%, which corresponds to the magnitude differecne of 0.1. However, since the dominant frequencies of these waveforms are several hertz, the paper speed was too slow (0.5 or 1.0 mm/s) to analyze the events further. On the other hand, TU digital waveform data have been stored since 1984. We compared the waveforms for the 1985, 1990 and 1995 events observed at several stations. As a result, we found that the waveform similarities are quite good up to about 3 Hz at least.
 The characteristics in the seismic activity, focal mechanisms and waveforms mentioned above indicate that these M4.8 events are caused by repeating slip at the same asperity on the plate boundary. In order to verify this hypothesis, we relocate the recent four events using the arrival-time difference (ATD) method (master event method) [Spence, 1980]. The relocated hypocenters are concentrated within a region as small as 1 km and it is possible to interpret that all the events occurred at the same location taking the estimation errors (several hundred meters) into account. Thus we inferred from these results that these events occurred on the same asperity on the plate boundary.
3. Size of the Asperity
 We estimated the source sizes of the characteristic events using acceleration seismograms [The Mining and Materials Processing Institute of Japan, 1998] recorded at Kamaishi station installed by JNC (Japan Nuclear Cycle Development Institute). The station location is indicated by a cross in Figure 1b. Waveform data for the 1990 and 1995 events were available. We estimated the mean QS value between the source and station assuming that the omega-square model [Aki, 1967] was correct. As a result, QS = 800 was found to be the most appropriate; this QS value was consistent with previous works [Hasegawa et al., 1979; Umino and Hasegawa, 1984]. From the spectra corrected with the QS, corner frequencies were estimated to be about 3.5 Hz for both the 1990 and 1995 events.
 Size of the rupture area estimated from corner frequency depends on the source model used in the analysis. In those source models, Brune's model [Brune, 1970, 1971] is the simplest and represented by the fewest parameters. In Brune's model, source radius (R) and corner frequency (fc) are connected by the equation as R = 0.372VS / fc, where VS denotes S-wave velocity. Substituting fc = 3.5 Hz and VS = 4.4 km/s which is the typical value at a depth of 50 km, we obtain R = 0.47 km corresponding to the source area of 0.69 km2. We also estimated the size of source area using a circular crack model [Madariaga, 1976] and unilateral Haskell's model [Haskell, 1964] taking the ray direction to Kamaishi station into account. The circular crack model predicts the area of 1.01 km2 while the Haskell's model shows 0.51 km2 if we assume the ratio of length to width of the fault is 2.
 All the models mentioned above give almost the same source lengths of about 1 km and source areas of 0.5 to 1.0 km2. The amount of slip (d) is expressed as
where Mo, μ and A are the seismic scalar moment, rigidity and source area, respectively [Aki, 1966]. Here, Mo is estimated to be 1.6 × 1023 dyne.cm by substituting M = 4.8 into the equation (1). Substituting Mo and μ = 7×1011 dyne/cm2 into equation (2), d is then estimated to be 46 cm for the model with A = 0.5 km2 and 23 cm for A = 1.0 km2. The stress drop is estimated to be 380–1100 bar.
 Since the descending rate of the Pacific plate is about 8 cm/year [DeMets et al., 1990], the slip deficit for 5.35 years corresponds to 43 cm if seismic coupling is 100%. This value is almost the same as the slip amount by the M4.8 event estimated in the previous section (23–46 cm). This result indicates that the asperity is almost 100% coupled if the source area is about 0.5 km2.
Nadeau and Johnson  proposed a new scaling relationship between the slip (d) and seismic moment (Mo) for small earthquakes based on the data of the repeating earthquakes as
where the unit of d is cm (equation 13 in their paper). Substituting Mo for M4.8 event in their scaling relations, the slip is calculated to be 38 cm, which is comparable to ours. This is quite remarkable because their relation is based on the results obtained from very shallow earthquakes occurring along the transform plate boundary, while our results are from rather deep earthquakes whose depths are about 50 km in a subduction zone.
Ellsworth  and Nadeau and McEvilly  compared the recurrence intervals with surface laser ranging data and concluded that episodic creeps control the recurrence intervals of the repeating earthquakes. In the case of our study, the recurrence interval is almost constant: standard deviation is only 10% of the mean. This suggests that the region surrounding the cluster off Sanriku is almost stably creeping. The coupling on the plate boundary in the northeastern Japan arc are thought to persist down to a depth of around 50 km [Umino and Hasegawa, 1982]; the boundary below 50 km depth is considered to be stably creeping. Since there are no low-angle thrust fault type events to the west of the cluster we analyzed here [Igarashi et al., 2001], the cluster probably corresponds to the deepest coupling area in this region.
 It is hard to find other characteristic small events using only the hypocenter catalogue because the location errors are quite large for events far off shore. Recently, Igarashi et al.  revealed that the characteristic repeating small earthquakes are widely distributed on the plate boundary off Sanriku using the waveform similarity analysis. They also shows that such repeaters are not located in the regions where the plate boundary is inferred to be ‘locked’ from GPS data analysis [Nishimura, 2000].
 From the above consideration, we propose a following model for the region between 39°N and 40°N. Asperities are widely distributed on the plate boundary off Sanriku but disappear below the depth of around 50 km. There are large asperities close to the trench while there are only small asperities distributed sparsely close to the seashore. The region surrounding the cluster analyzed in this study is mostly creeping just the same as the region deeper than 50 km.
 Since the recurrence interval was so stable, Matsuzawa et al.  predicted that the next event would occur by the end of January 2001 with 68% probabilty and by the end of November 2001 with 99% probability assuming that the recurrence interval would follow the normal distribution. The magnitude would be 4.8 ± 0.1 taking very small fluctuation in the waveform amplitudes into consideration.
 The expected event actually occurred on November 13, 2001, and its magnitude was preliminary reported to be 4.7 by JMA. The hypocenter location was almost the same as the previous events (within several hundred meters). The interval after the 1995 event (6.68 years before the 2001 event) is the longest in the eight intervals while the one before the 1995 event (4.65 years after the 1990 event) is the shortest. Nishimura et al.  shows that the after-slip area of the 1994 M7.5 Far-off Sanriku earthquake was extended to the region we analyzed here. Thus we infer that the after-slip region surrounding the asperity accelerated the occurrence of the 1995 M4.8 event.
 Since the averaged interval and standard deviation have been renewed to be 5.52 and 0.68 years, the next event will occurr in May 2007 ± 21 month with 99% probability based upon the time since the last event assuming that the interval follows the normal distribution.
 We thank T. Hirasawa and K. Maeda for their valuable suggestions, Japan Nuclear Cycle Development Institute for providing us the digital waveform data recorded at Kamaishi station. This work was partly supported by Grant-in-Aid for Scientific Research (C) no.11680465, The Ministry of Education, Science, Sports and Culture.