Reconstruction of multi-millennial summer climate variations in central Japan by integrating tree-ring cellulose oxygen and hydrogen isotope ratios

Oxygen isotope ratios (18O) of tree-ring cellulose are a novel proxy of summer hydroclimate in monsoonal Asia. In central Japan, we collected 67 conifer wood samples, mainly Chamaecyparis obtusa, with ages encompassing the past 2,600 yr. The samples were taken from living old trees, excavated archeological wood, old architectural wood, and naturally buried 25 logs. We analyzed stable isotope ratios of oxygen (18O) and hydrogen (2H) in tree-ring cellulose in these samples without using a pooling method, and constructed a statistically reliable tree-ring cellulose 18O time-series for the past 2,500 yr. However, there were distinct age trends and level offsets in the 18O record, and cellulose 18O values showed a gradual decrease as an individual tree matures. This suggested it is difficult to establish a cellulose 18O chronology for low-frequency signals by simple averaging of all the 18O time-series data. However, there were opposite age trends in the cellulose 2H, and 30 2H gradually increased with tree age. There were clear positive correlations in the short periodicity variations between 18O and 2H, probably indicating a common climate signal. A comparison of the 18O and 2H time-series in individual trees with tree-ring width suggested that the opposite age trends of 18O and 2H are caused by temporal changes in the degree of postphotosynthetic isotope exchange with xylem water, accompanied by changes in stem growth rate (growth effect) that are influenced by human activity in the forests of central Japan. Based on the assumptions that cellulose 18O and 2H vary 35 https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c © Author(s) 2020. CC BY 4.0 License.


Measurement of tree-ring cellulose  18 O and  2 H
Cellulose is extracted from tree-rings using the method of Kagawa et al. (2015). Firstly, 1-mm-thick cross-sectional laths perpendicular to the cellulose fibers were sliced from all wood samples and directly subjected to chemical treatment to remove 100 components other than cellulose. A 100-300 g fragment of cellulose representing the whole annual layer was then cut from the cellulose lath year-by-year using a fine blade under a microscope. Each sample was wrapped in silver foil for  18 O and  2 H measurements using a mass spectrometer combined with a pyrolysis elemental analyzer (TCEA/Delta V Advantage; Thermofisher Scientific, Bremen, Germany). After measurements of every eight samples, a cellulose standard (cellulose reagent; Merck) with known  18 O and  2 H was analyzed to calibrate analytical drift and calculate the  18 O and  2 H of samples 105 against the international standard of VSMOW, following Eqs (1) and (2) Given that cellulose contains 30% of OH-group hydrogen that is exchangeable with experimental water during chemical 110 treatment, it is necessary to remove the OH-group by nitration of the cellulose (synthesis of nitrocellulose) before the isotope measurements (Epstein et al., 1976;DeNiro, 1981). It was not possible to nitrate all the cellulose samples analyzed in this study. However, we can expect that all OH-group hydrogen in a cellulose lath will have a unique  2 H value, as it would have been replaced by the hydrogen of homogeneous water in the test tube during the process of cellulose extraction, such that the measured tree-ring cellulose  2 H time-series will retain the original isotopic variations, although the amplitude of these 115 variations will decrease by up to 70% and the absolute values will have been modified (Filot et al., 2006). We confirmed that there are consistent correlations in  2 H variations between cellulose and nitrocellulose produced using the traditional nitration method (Epstein et al., 1976;DeNiro, 1981) for two tree samples (Fig. 2). This demonstrates that it is possible to reconstruct original variations in tree-ring cellulose  2 H data without nitrating the cellulose. However, our  2 H measurements are significantly influenced by a memory effect from the previous few measurements, probably due to the absorption of H 2 120 molecules in the pathway of the pyrolysis elemental analyzer (TCEA); this reduces the statistical precision of the  2 H analyses.

Results and Discussion
3.1 Variations in the tree-ring cellulose  18 O and  2 H data https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c Author(s) 2020. CC BY 4.0 License. juvenile period (Totman, 1989). In addition, this tree experienced a sudden increase in growth rate (i.e., tree-ring width) at ca. 155 1950 CE, probably due to logging of neighboring trees after the end of World War II in 1945 CE (Fig. 4b). Tree No. 49 germinated in the 12 th century, and survived the intense logging activity in the 17-18 th century CE because it was left as a seed tree. By the time of logging in the 17-18 th centuries, the growth environment of the tree had improved step by step. We can recognize three episodes of drastic growth rate increase of this tree at ca. 1600, 1710, and 1720 CE (Fig. 5b). After 1720 CE, the growth rate decreased gradually along with that of neighboring (younger) trees. 160 Figures 4 and 5 show that  18 O increased and  2 H decreased, regardless of the tree age, corresponding to an increase in tree-ring width (i.e., growth rate), in 1600, 1700, and 1720 CE for the No. 49 tree and 1950 CE for the No. 65 tree. Therefore, the long-term opposite trends in the tree-ring cellulose  18 O and  2 H data are not a simple age trend. (Note that the long-term  18 O and  2 H variations are not perfect mirror images because they are also affected by long-term climate variations in which  18 O and 2 H are positively correlated.) Given the growth rates of trees that germinated in open spaces after logging or fires 165 gradually decline as the trees mature, the  18 O and  2 H values decrease and increase with time, respectively. However, the growth rate not only decreases due to canopy closure, but also increases randomly due to human activities in the forest. As such, it is impossible to apply the RCS or negative exponential curve proposed by Esper et al. (2010) for removing the apparent age trend of tree-ring cellulose  18 O in central Japan, where intense logging activity has been undertaken in forest environments for a long time. 170 To date, the age trend in tree-ring cellulose  18 O data has been explained by the relatively shallow root system of juvenile trees and the evaporative enrichment of 18 O in soil water at the near-surface. However, this cannot explain the long-term variations in tree-ring cellulose  18 O data in this study , because the root depth cannot become shallower when the growth rate increases and soil water evaporation must increase both  18 O and  2 H. Figures 4b and 5b clearly show that the long-term variations in tree-ring cellulose  18 O and  2 H are influenced by a physiological effect corresponding to long-term 175 where cel and rain indicate the tree-ring cellulose and precipitation, respectively; h is the relative humidity during photosynthesis; f is the proportion of oxygen atoms in carbohydrate exchanged with xylem water during post-photosynthetic processes before cellulose synthesis, which is assumed to be equal to that of the hydrogen atoms;  eo ( eh ) and  ko ( kh ) are 185 isotopic fractionation factors for oxygen (hydrogen) during equilibrium evaporation between water and water vapor and kinetic water vapor diffusion through leaf stomata, respectively; and  ao ( ah ) and  ho ( hh ) are isotopic fractionation factors for oxygen (hydrogen) between water and carbohydrate during photosynthesis in the leaf and the post-photosynthetic processes before cellulose synthesis, respectively. Values of  eo ,  ko ,  ao , and  ho are +9, +29, +27, and +27, respectively, while  eh and  kh are +80 and +25, respectively, and  ah and  hh are approximately -150 and +150, respectively . 190 Given that (Dansgaard, 1964) and any change of h produces changes in  18 O cel and  2 H cel in Eqs (3) and (4)  However, if f increases over the long-term due to some physiological reason, it can be inferred from Eqs (3) and (4) that  18 O cel decreases while  2 H cel increases. This is because, for oxygen, the magnitude of isotopic fractionation between water and carbohydrate is the same (+27) for photosynthesis ( ao ) and post-photosynthesis processes ( ho ), and the increase in f 200 reduces the effect of leaf water 18 O enrichment, ( Oe +  Ok )(1 -h) in Eq. (3), resulting in lower  18 O cel . However, for hydrogen, the magnitude of isotopic fractionation between water and carbohydrate during post-photosynthesis processes ( hh = +150) is significantly larger than that during photosynthesis ( ah = -150), such that the increase in f results in higher  2 H cel . This causes long-term opposite trends in  18 O cel and  2 H cel data (Figs 3, 4b, and 5b), which can be interpreted as an increase in the rate of post-photosynthetic isotopic exchange between carbohydrate and xylem water (f). Possible physiological mechanisms for this 205 include an increase in the rate of utilization of stored carbohydrates for stem cellulose synthesis (Nabeshima et al., 2018), rather than using photosynthetic products directly for rapid tree growth during the juvenile period. This could also occur in the period following an abrupt improvement in the growth environment, due to logging of neighboring trees.
Although the long-periodicity variations in  18 O cel and  2 H cel are influenced by predominant physiological effects (Figs 4b and 5b), this does not mean that the long-periodicity variations in  18 O cel and  2 H cel do not contain climatological components. 210 In fact, climate varies at all time-scales, such that long-term variations in  18 O cel and  2 H cel inevitably include climatological components. It is therefore challenging to resolve the climatic signals from the physiological effects. In dendrochronological studies based on tree-ring width, RCS is used to separate and estimate climatological components in tree-ring width time-series (Esper et al., 2002;Grudd et al., 2002;Büntgen et al., 2005). However, it is difficult to create a regional standardized  18 O cel curve like the RCS for the samples analyzed in this study, because the physiological effects on  18 O cel are not solely an age 215 trend (Figs 4b and 5b), and also reflect past logging activity.

Classification of  18 O and  2 H variations into climatological and physiological components
To extract the climatological component of the variations in the tree-ring cellulose  18 O data, we modified the model of cellulose  18 O and  2 H in Eqs (3) and (4) into climatological and physiological components. Given there are four variables ( 18 O rain ,  2 H rain , h, and f) in Eqs (3) and (4), we first define their variations as follows: 220 where  18 O rain(0) ,  2 H rain(0) , h (0) , and f (0) are  18 O rain ,  2 H rain , h, and f in a fixed year (0), respectively, and  18 O rain ,  2 H rain , h, 225 and f are deviations in  18 O rain ,  2 H rain , h, and f from the fixed year (0) to an arbitrary year, respectively. By substituting Eqs (5)-(8) into Eqs (3)-(4) and neglecting the second-order minor terms (fh, f 18 O rain , and f 2 H rain ) Eqs (3) and (4) can be rewritten as follows: We can now introduce new equations for the climatological and physiological components in the tree-ring cellulose  18 O 235 and  2 H time-series, as follows: where  18 O cel(climate) ,  18 O cel(physiol) ,  2 H cel(climate) , and  2 H cel(physiol) are variations in the tree-ring cellulose  18 O and  2 H from the fixed year (0) with respect to an arbitrary year due to climatological and physiological factors, respectively. We can then reformulate  18 O cel and  2 H cel as the sum of climatological and physiological components, as follows: where  18 O cel(0) and  2 H cel(0) are  18 O cel and  2 H cel at the fixed year (0), respectively.

A method to extract the climatological component from cellulose  18 O
Here we propose a new method to calculate the climatological component in variations of the tree-ring cellulose  18 O( 18 O cel(climate) ) by solving simultaneous equations consisting of Eqs (15)-(16), with two additional equations based on the relationship between  18 O cel(climate) and  2 H cel(climate) and between  18 O cel(physiol) and  2 H cel(physiol) (i.e., Eqs (17)-(18)). 250 This is based on the theoretical and observational understanding that tree-ring cellulose  18 O and  2 H data correlate positively and negatively due to climatological and physiological factors, respectively. Although the assumption that A and B are constant might not be valid, this assumption is needed to explicitly calculate  18 O cel(climate) using Eq. (21). Hence, we tentatively assumed that A and B are constant, and calculated  18 O cel(climate) over the past 2,600 yr using Eq. (21). We then verify this by comparison with numerous local, regional, and global meteorological, historical, and paleoclimatological records of past summer climate over various time-scales.

Method to determine the proportional coefficients A and B 270
To utilize Eq. (21), we need to determine the proportional coefficients A and B in Eqs (17) We first consider the feasibility of the theoretical approach. It is not easy to determine A using Eq. (19), because there are three variables ( 18 O rain ,  2 H rain , and h). If the rainwater isotope ratios do not change (both  18 O rain and  2 H rain = 0), then 275 A is equal to ( eh +  kh )/( eo +  ko ), which is (80 + 25)/(9 + 29) = 2.76 . However, if relative humidity does not change (h = 0), A =  2 H rain  18 O rain , which is equal to 8 if the meteoric water line is followed (Dansgaard, 1964). Given the wide range of potential A values from 2.76 to 8, we cannot easily theoretically define A. On the other hand, B may be easier to determine theoretically, because there are no variables in Eq. (20). For example, if the relative humidity h (0) is 0.5, B = 13 because  eo ,  ko ,  ao ,  ho ,  eh ,  kh ,  ah , and  hh are +9, +29, +27, +27, +80, +25, ca. -150, and +150, respectively 280 . However, in fact, it is not easy to fix h (0) because the relative humidity typically shows large diurnal variations, and the timing of  18 O incorporation into leaf carbohydrate is unknown. Moreover, the  ah and  hh values of -150 and +150 are just approximations. For example, while Yakir and DeNiro (1990) obtained values of -171 and +158 for  ah and  hh , respectively, Estep and Hoering (1981) obtained values of -100 to -120 for  ah and Luo and Sternberg (1992) reported values of +144 to +166 for  hh . Therefore, we used the empirical approach for estimating A and B. The A value in Eq. (17)  in Fig. 3a, in order to find the B value that produced the best match between the two time-series. This procedure assumes that, although individual trees possess different age trends and level offsets, utilizing all data from all trees cancels out these effects and allows climate variations to be discerned, at least in terms of the overall trend.

Method to combine individual tree-ring time-series with large level offsets
To reconstruct multi-millennial variations in the climatological component of the tree-ring cellulose  18 O data 300 ( 18 O cel(climate) ), we must combine individual time-series from different trees with variable level offsets due to different sample locations. As such, we cannot simply average all data for individual trees, because this produces steps in the composite record at both ends of the time-series of individual trees. Numerous procedures have been proposed to combine tree-ring time-series (e.g., Hangartner et al., 2012); here we propose a new iterative calculation method (Fig. 6).
Firstly, we simply average all the individual time-series to make a preliminary combined time-series. Secondly, we offset 305 each individual time-series up or down, retaining their original patterns of temporal variations, to the position where the average of the individual time-series becomes equal to that of the preliminary combined time-series for the corresponding period of the individual tree. Thirdly, we average all the offsetted individual time-series to make a refined combined timeseries. We iterate this procedure until the average of the individual time-series becomes equal to that of the combined timeseries for the corresponding period, without any further offsetting of the individual time-series. This method assumes that the 310 tree-ring absolute isotope ratios of individual trees are not important, because they depend on the sample locations, but that the temporal variations are well correlated amongst different trees due to a common regional climate signal.  . 3d), we reintegrate the long-periodicity variations in  18 O cel(climate) and short-periodicity variations in  18 O cel for all individual trees at an adequate threshold periodicity and combine all individual data again, because the physiological factors 320

Procedure
do not appear to affect the short-periodicity  18 O cel variation. In this procedure (Fig. 7a), the iterative calculation used to combine many tree-ring time-series in the third step is the most time-consuming because various B values are tested.
Although A can be determined independently for each tree, we can only obtain one B value for all the trees. In a practical sense, it is not meaningful to determine A separately for each tree, because all the trees were collected within central Japan ( Fig. 1). If we assume that A and B are unique for all trees in this study, we can simplify the procedure as shown in Fig. 7b to 325 reduce the time required for the iterative calculation. Considering that all the calculations used to combine and integrate the time-series in Fig. 7 are linear, we can change the order of calculation between combination and integration. In fact, if we use common A and B values, the resultant combined time-series for  18 O cel(climate) does not change when using the two procedures ( Fig. 7a and b). In Fig. 7b, we first combine the  18 O cel and  2 H cel time-series of all trees in Fig. 3

Determination of the climatological and physiological proportional coefficients (A and B)
We combined all the time-series for  18 O cel and  2 H cel shown in Fig. 3 using the method illustrated in Fig. 6. The final combined  18 O cel and  2 H cel time-series after 1000 iterations are shown in Fig. 8. The long-term variations in the combined  18 O cel and  2 H cel time-series obviously reflect the accumulated age trends, in which the  18 O cel and  2 H cel tend to decrease and increase over a long time-scale, respectively. 340 We found that the climatological proportional coefficient A in Eq. (17) can be set to 5, because there are positive correlations in the short-periodicity variations of the combined  2 H cel and  18 O cel records with a slope of ca. 5 (Fig. 9), irrespective of the threshold year for extracting the short-periodicity variations (i.e., 5, 11, or 21 yr). A value of 5 is within the theoretical range of A of 2.76 to 8 obtained from Eq. (19). We used the  2 H cel data directly for the calculation of A, although the amplitude of variations in  2 H cel is reduced to 70% of that of the original cellulose, such that a value of 5 is equivalent to To determine the most appropriate value for the physiological proportional coefficient B in Eq. (18), we integrated the combined  18 O cel and  2 H cel time-series in Fig. 10 using Eq. (21), in order to calculate  18 O cel(climate) with B values of 3, 5, 7, and 9, and an A value 5. We then examined the overall trend in  18 O cel(climate) and found that it became equal to the overall trend of the average  18 O cel (Fig. 3a), when B = 5.4 (Fig. 10). Given the 70% amplitude reduction of  2 H cel , 5.4 is equivalent 350 to 7.7 from the theoretical estimation. If we assume that  eo ,  ko ,  ao ,  ho ,  eh ,  kh ,  ah , and  hh are +9, +29, +27, +27, +80, +25, -150, and +150, respectively, then B = 7.7 means that h (0) is 0.25 in Eq. (20). Given that 0.25 is too low for the relative humidity in central Japan, the values of -150 and +150 for  ah and  hh , respectively, may be overestimated. However, because the overall trend of the integrated  18 O cel(climate) using B = 5.4 is equal to that of the average of raw  18 O cel values, we use this value of B hereafter. 355

Calculation of temporal variations in the climatological component ( 18 Ocel(climate))
We calculated the temporal variation in the climatological component ( 18 O cel(climate) ) of tree-ring cellulose  18 O (Fig. 11a) by using A = 5 and B = 5.4, and the smoothly combined  18 O cel and  2 H cel time-series in Fig. 8 and Eq. (21). We used the temporal average of ( 2 H cel + B  18 O cel )/(A + B) during the 30 yr from 1961 to 1990 as ( 2 H cel(0) + B  18 O cel(0) )/(A + B) in Eq.
(21). 360 The  18 O cel(climate) variations shown in Fig. 11a must be influenced by the low quality (low R-bar and EPS values) of the original  2 H cel time-series (Fig. 3d) (Fig. 11b). The purpose of introducing the  2 H cel signal into the  18 O cel time-series was to remove the physiological effects from the  18 O cel time-series, but the short-periodicity 365 variations in the  18 O cel time-series do not originally contain physiological effects, so it is not necessary to integrate the  2 H cel signals into the short-periodicity components of the  18 O cel time-series. Therefore, at two threshold periodicities (21 and 51 yr), we tentatively separated the long-periodicity component (21-and 51-yr running means) from the integrated  18 O cel(climate) time-series in Fig. 11a and short-periodicity component (deviations from 21-and 51-yr running means) from the original combined  18 O cel in Fig. 8, and reintegrated these into a new time-series for  18 O cel(climate) to remove the influence of the low 370 quality  2 H cel data from the short-periodicity component. We also applied this reintegration procedure between the long ( 18 O cel(climate) ) and short ( 18 O cel ) periodicity components to the data for individual trees, and calculated the EPS values for the reintegrated  18 O cel(climate) datasets (Fig. 11b). The resultant EPS values were >0.85, and much higher than the original  18 O cel(climate) for almost all periods during last 2,500 yr when either 21 or 51 yr were used as the thresholds. This suggests that the reintegration procedure ensures the reliability of the datasets without the influence of the low-quality  2 H cel data. 375 However, this reintegration procedure may recover potential physiological effects with intermediate periodicities of less than 21 and 51 yr. Hence, we compared the long-term variations (periodicity > 11 yr) of the two reintegrated  18 O cel(climate) https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c Author(s) 2020. CC BY 4.0 License. time-series using 21 and 51 yr as the thresholds and the original  18 O cel(climate) time-series (Fig. 11a) in Fig. 11c to investigate whether there are significant differences. Each of the three time-series shown in Fig. 11c almost coincide over all time-scales, indicating there are no significant physiological effects with a periodicity between 11 and 51 yr. However, we used 21 yr as 380 the threshold for the reintegration of the long-and short-periodicity components, in order to robustly remove the influence of physiological effects. It does not result in a lower quality reintegrated  18 O cel(climate) record, given the nearly equal EPS values using the 21-and 51-yr thresholds in Fig. 11b. We utilized the reintegrated  18 O cel(climate) time-series between the longperiodicity (>21 yr) domain of  18 O cel(climate) in Fig. 11a and short-periodicity (<21 yr) domain of  18 O cel in Fig. 8 as the final time-series of the climatological component in the tree-ring cellulose  18 O data (Fig. 12). 385 In contrast to the combined  2 H cel record (Fig. 8), the multi-centennial variation in the combined  18 O cel record (Fig. 8) does not appear to be very similar to that of the climatological component  18 O cel(climate) (Fig. 12). This is partly because there is an apparent age trend in the combined  18 O cel record, which overlaps the multi-millennial climatological decrease in Fig. 10.
This means the multi-centennial variations are ambiguous in the combined  18 O cel record (Fig. 8). However, there may be an anthropogenic explanation for this, whereby in wetter and cooler periods, the number of trees in Japanese forests might have 390 decreased due to logging for fuel. Rapid tree growth in the resultant more open and lighter forest would have increased  18 O cel and decreased  2 H cel values due to physiological effects. The wetter and cooler climate may have also lowered  18 O and  2 H in leaf water. The combined effects of climate variations and anthropogenic factors on Japanese forests might have reduced and enhanced the multi-centennial variations in  18 O cel and  2 H cel , respectively (Fig. 8). However, we can robustly extract the climatological component  18 O cel(climate) independently of the local forest history by integrating  18 O cel and  2 H cel data. This 395 is the most important paleoclimatological innovation of our study.

Comparison of  18 Ocel(climate)with other summer climate records
In dendroclimatological studies focused on inter-annual variability, statistical methods to calibrate and verify the relationship between tree-ring data and instrumental meteorological observations have been well established. However, there is no commonly accepted statistical method to validate the reliability of long (i.e., centennial or millennial) periodicity climate 400 reconstructions. In the case of low-frequency data recovered from speleothems, ice cores, and sediments, climate reconstructions are typically not based on correlations with meteorological observations. These reconstructions are verified by different methods, such as mechanical models based on the relationship between oxygen isotope ratios and environmental factors, empirical knowledge of the relationship between pollen assemblages and climate, and experimental studies between biomarker compositions and water temperature. The reconstruction of  18 O cel(climate) variations is principally based on the 405 mechanical model developed in Eqs (3)-(21), but it is necessary to validate our results by comparison with other past summer climate records. Figure 13 shows the sensitivity of  18 O cel(climate) to local monthly mean temperature, mean relative humidity, and precipitation during the period from 1901-2005 at Kyoto, Nagoya, and Iida in central Japan.  18 O cel(climate) shows significant https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c Author(s) 2020. CC BY 4.0 License. negative correlations with precipitation and relative humidity, and a positive correlation with temperature during summer (Fig.  410 13d-f), when annual precipitation is at its maximum (Fig. 13a-c), as demonstrated by previous studies of monsoonal Asia (Li et al., 2015;Liu et al., 2017;Pumijumnong et al., 2019;Sano et al., 2012Sano et al., , 2013Sano et al., , 2017Seo et al., 2019;Xu et al., 2013Xu et al., , 2018Xu et al., , 2019. Spatial correlations of  18 O cel(climate) with June-July precipitation in East Asia (Fig. 14) indicate that  18 O cel(climate) in central Japan reflects precipitation in an extended region from the lower reaches of the Yangtze River in China to southern Honshu in Japan, corresponding to the Baiu/Meiyu front, which is an early summer stagnant rain belt characteristic of the East 415 Asian summer monsoon (Fig. 14a).  18 O cel(climate) has a significant positive correlation with June-July mean temperature across a wide region of Japan, Korea, and China, suggesting that  18 O cel(climate) may be a good proxy for the East Asia summer monsoon (Fig. 14c).
The negative correlations between  18 O cel(climate) and relative humidity and precipitation can be explained by the direct negative relationship between  18 O cel(climate) and relative humidity in Eq. (11) and the amount effect, whereby  18 O rain becomes 420 lower when precipitation increases (Dansgaard, 1964;Araguás-Araguás et al., 1998). The highest correlation (<0.6) area of June-July precipitation is located just to the south of the sample sites (Fig. 14b). This is because, in the summer season, water vapor usually comes from the south (Pacific Ocean) to the sample sites and  18 O rain becomes lower when heavy rainfall occurs just before arrival of the air mass carrying the water vapor. The positive correlation between  18 O cel(climate) and temperature must be caused by the meteorologically reverse relationship between summer precipitation and temperature in humid 425 monsoonal Asia, including Japan. In fact, the center of the highest correlation area of June-July mean temperature is located slightly to the west of the sample site (Fig. 14c). As such, when the temperature in western Japan is high and it is characterized by high pressure, the wind blows from the north to the sample site resulting in dry conditions and low rainfall.
Given the relationship between  18 O cel(climate) and modern meteorological observations evident in Figs 13 and 14, we compared  18 O cel(climate) in central Japan with long-term historical and paleoclimatological records of summer climate before 430 the 19 th century. During the Edo era (1603-1868 CE), people wrote numerous diaries throughout Japan in which daily weather conditions were routinely described. Mizukoshi (1993) compiled many diary weather descriptions for central Japan and reconstructed inter-annual variations in early summer precipitation for Osaka since 1692 CE (Fig. 15). The diary-based (1692-1882) and instrumentally observed (1883-1990) precipitation reconstructions for Osaka are negatively correlated with  18 O cel(climate) in central Japan, not only at an inter-annual time-scale, but also at a multi-decadal time-scale, indicating that 435  18 O cel(climate) records long-term variations in summer climate.
During the Medieval Period from the 11 th to 16 th centuries in Japan, there were numerous extreme meteorological events (Fujiki, 2007). We used flood-and drought-related disaster records during the summer season (June-August) to construct an index of the "flood disaster ratio" that is the proportion of flood-related document numbers to the total flood-and droughtrelated document numbers. Given the total document numbers are scarcer in the older period, we calculated an 11-yr running 440 mean of the "flood disaster ratio" and compared it with  18 O cel(climate) (Fig. 16). In the 10 th and 11 th centuries during the Medieval Climate Anomaly, there were numerous droughts in Japan, corresponding to the highest values of  18 O cel(climate) . https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c Author(s) 2020. CC BY 4.0 License.
After the 11 th century, both the documentary records and  18 O cel(climate) values demonstrate that climate became progressively wetter towards the Edo era in the 17 th century.
Although we could not find historical records of extreme climate events prior to the 10 th century CE in Japan, Sakaguchi 445 (1983Sakaguchi 445 ( , 1989) reconstructed long-term summer temperature variations using the pollen percentage of a cold region pine (Pinus pumila) in the Ozegahara peatland of east Japan (Fig. 17). Although the pollen data after the 3 rd century CE are not very reliable due to the lower sedimentation rate and human disturbance, the high sedimentation rate before the 2 nd century CE enabled us to compare it with  18 O cel(climate) . Both the pollen and  18 O cel(climate) records show similar variations from the 6 th century BCE to 2 nd century CE, indicating a warmer and drier climate from the 5 th to 2 nd century BCE and a cooler and wetter climate after 450 the 1 st century BCE (Fig. 17). After the 3 rd century CE, both datasets show similar millennial variations, although the temporal resolution of the pollen data is not high.
The climatological component of the variations in tree-ring cellulose  18 O ( 18 O cel(climate) ) correlates well with meteorological, historic, and vegetation data over various time-scales in central Japan, and also shows similar long-term patterns as paleoclimatological global temperature and East Asian precipitation data. Figure 18a shows that  18 O cel(climate) 455 exhibits almost identical variations as air temperature reconstructions for land areas in the Northern Hemisphere , except for the period of the Medieval Climate Anomaly. Variations in summer precipitation reconstructed from diatom assemblages in lake sediments in Taiwan  are also similar to the variations in  18 O cel(climate) in central Japan (Fig. 18b). Two time-series of carbonate  18 O values in speleothems from northwest China Tan et al., 2010) also match the variations in  18 O cel(climate) (Fig. 18c). Note that the directions of the y-axis are reversed between the 460 speleothem and  18 O cel(climate) data, reflecting the meridional disparity of precipitation patterns in East Asia (Fig. 14a), as demonstrated by Liu et al. (2014) and Chen et al. (2015).
The climatological component of the variations in the tree-ring cellulose  18 O data ( 18 O cel(climate) ) in central Japan (Fig.   12) correspond well over various time-scales with summer precipitation and temperature records in central Japan, which have been derived from various meteorological, historical, and paleo-vegetation archives (Figs 13-17). This indicates that 465  18 O cel(climate) is a reliable proxy of summer climate, such as the activity of the East Asian summer monsoon. Multi-centennial and millennial  18 O cel(climate) variations are similar to those of paleoclimatological reconstructions of global temperatures and East Asian precipitation (Fig. 18), indicating a drier climate during the Medieval Climate Anomaly and wetter climate during the Little Ice Age in central Japan. As such,  18 O cel(climate) can be utilized in climatological, historical, and archeological studies.

Conclusions 470
We constructed a statistically reliable multi-millennial tree-ring dataset of cellulose  18 O in central Japan by analysing tree-rings of 67 trees without using a pooling method. We found that there are distinct age trends in the  18 O time-series. By comparison with the  2 H time-series, we showed that the age trend in  18 O is caused by an increase in the degree of post-photosynthesis isotopic exchange with xylem water before cellulose synthesis as the trees mature. Because the physiological conditions of the post-photosynthesis isotopic exchange are not simply controlled by tree age, but also related to the tree growth 475 environment randomly influenced by human activity, it was not possible to remove the age trend by application of the negative exponential curve or RCS.
Given that tree-ring cellulose  18 O and  2 H are correlated positively and negatively due to climatological and physiological factors, respectively, we formulated simultaneous equations for the climatological and physiological components of the treering cellulose  18 O and  2 H data. We solved these equations to cancel out the physiological effects and established a multi-480 millennial record of the climatological component of tree-ring cellulose  18 O ( 18 O cel(climate) ). The  18 O cel(climate) time-series is well correlated with local, regional, and global variations in summer climate reconstructed by instrumental, historical, and paleoclimatological methods. This suggests that  18 O cel(climate) records summer climate variations in central Japan during the past 2,600 yr on annual to millennial time-scales.
However, further research is needed to make  18 O cel(climate) a more reliable index of summer climate. Firstly, the analytical 485 precision of the tree-ring cellulose  2 H measurements needs to be improved. In order to minimize the negative influence of the exchangeable OH-group hydrogen, it is necessary to fix the  2 H in the OH-group (Filot et al., 2006). The memory effect of H 2 molecules in the pyrolysis elemental analyzer also needs to be reduced. Secondly, more global tree-ring cellulose  2 H data need to be acquired to expand the  18 O and  2 H datasets (and  18 O cel(climate) Esper et al. (2010), hardwoods may be intrinsically free from long-495 term age trends (Duffy et al., 2017). Therefore, in most isotopic dendrochronological studies, the cellulose  2 H data will be a supplementary index to ensure there are no significant age trends (An et al., 2014). However, when such studies are based on a small number of conifer woods collected from archeological artifacts and/or architectural material, and where their growth environments may have been disturbed by human activities, the simultaneous measurement of  2 H with  18 O allows of tree-ring cellulose; TN, MS, YS analyzed isotopic data mathematically; All authors discussed the results and provided input to the manuscript. 505 Competing interests. The authors declare that they have no conflict of interest.    were calculated for a total period of 51 yr, including 25 yr before and after the year shown. https://doi.org/10.5194/cp-2020-6 Preprint. Discussion started: 7 February 2020 c Author(s) 2020. CC BY 4.0 License.