Continuous records of the atmospheric greenhouse gases CO2, CH4, and N2O and their radiative forcing since the penultimate glacial maximum

Continuous records of the atmospheric greenhouse gases (GHGs) CO2, CH4, and N2O are necessary input data for transient climate simulations and their related radiative forcing important components in analyses of climate sensitivity and feedbacks. Since the available data from ice cores are discontinuous and partly ambiguous a well-documented decision process during data compilation followed by some interpolating post-processing are necessary to obtain those desired time series. Here we document our best-guess data compilation of published ice core records and recent measurements on firn air 5 and atmospheric samples covering the time window from the penultimate glacial maximum (∼156 kyr BP) to the beginning of year 2016 CE. A smoothing spline method is applied to translate the discrete and irregularly spaced data points into continuous time series. These splines are assumed to represent the evolution of the atmospheric mixing ratios for the three GHGs. Globalmean radiative forcing for each GHG is computed using well-established, simple formulations. Newest published age scales are used for the ice core data. While CO2 is representing an integrated global signal, we compile only a southern hemisphere record 10 of CH4 and identify how much larger a northern hemisphere or global CH4 record might have been due to its interhemispheric gradient. Data resolution and uncertainties are considered in the spline procedure and typical cutoff periods, defining the degree of smoothing, range from 5000 years for the less resolved older parts of the records to 4 years for the densely-sampled recent years. The data sets describe seamlessly the GHG evolution on orbital and millennial time scales for glacial and glacialinterglacial variations and on centennial and decadal time scales for anthropogenic times. Data connected with this paper, 15 including raw data and final splines, are available at https://doi.pangaea.de/10.1594/PANGAEA.871273.


Introduction
Our knowledge of changes in the atmospheric mixing ratios of the important greenhouse gases (GHGs) CO 2 , CH 4 and N 2 O beyond the instrumental record is mainly based on discrete data points derived from gas extractions in polar ice cores.While there are recent developments towards continuous CH 4 records using gas extraction and measurement systems coupled to continuous flow analysis systems (Schüpbach et al., 2009;Chappellaz et al., 2013;Rhodes et al., 2013Rhodes et al., , 2015)), none of this is within sight for CO 2 and N 2 O, the other two important GHGs, whose changes influence global climate via their radiative forcing.To Previous splines (similar to our approach here, but in detail not identical) have also been proposed to be used in interglacial experiments of the Holocene within PMIP4 (Otto-Bliesner et al., 2016).Within the most recent model intercomparison project, CMIP6, a slightly different compilation of GHGs for historical times or the Common Era has been presented (Meinshausen et al., 2016).While this alternative approach has its focus on the time since 1850 CE, its data compilation nevertheless extends back until the year 0 CE, based beyond instrumental times solely on the Law Dome ice core (MacFarling-Meure et al., 2006;Rubino et al., 2013).We will finally compare our splines with these forcing data sets proposed by Meinshausen et al. (2016) to be used within CMIP6.
As will be seen in detail in the next section the mathematical formulation of the spline smoothing method needs as input data some information on the uncertainties or errors of the data points supporting the spline.These data uncertainties taken here are the precisions of individual measurements (1σ errors) and are of the order of a few ppm for CO 2 or a few ppb for CH 4 or N 2 O.The uncertainty in the final spline, however, is larger, since the applied smoothing, which depends on the chosen cutoff periods, adds some additional uncertainty.Furthermore, the estimates of the radiative forcing based on these three GHGs are even more uncertain, since the calculations of the radiative forcing themselves are based on models (Myhre et al., 1998) with an embedded intrinsic uncertainty of about ∼ 10%.
Please note, that in the following we choose to state that the anthropogenic activities (or emissions) started at 1750 CE (or 200 BP).This is approximately the time, at which both CO 2 and CH 4 in our final splines started to rise rapidly.We keep to this distinction of anthropogenic and pre-anthropogenic changes in the GHGs at 1750 CE throughout the text, although the onset of the Anthropocene is still debated (e.g.Lewis and Maslin, 2015;Steffen et al., 2016;Williams et al., 2016).

Details on the spline smoothing method
The numerical code for spline smoothing is based on Enting (1987), but see also Bruno and Joos (1997); Enting et al. (2006) for further details, discussions and applications.It offers the possibility to select different cutoff periods for different time intervals/parts of the input data set, which is needed when data spacing is very different in different parts of the data set.
In a smoothing spline a cost function is minimised.This cost functions includes two terms: (i) the error-weighted deviation between the spline value and the actual data value, and (ii) the curvature of the spline/second derivative.A parameter λ defines how much weight is given to the curvature.The optimisation results in low curvature, i.e. a very smooth spline and relatively large deviations from the original data, for large values of λ.Similarly, increasing errors of the data results in a smoother spline for a given λ.In other words, the smooting of the spline depends on both the assumed errors of the data and the parameter λ.
According to Fourier, each data set can be represented by a sum of sine functions.Ideally, a smoothing spline removes all high frequencies sine functions and acts as a low pass filter.The period (or frequency) where half of the amplitude is attenuated is typically called cutoff period P cutoff (frequency).Thus, periods shorter (frequencies longer) than the cutoff are dampened in the spline.The parameter λ is linked to this cutoff period as described in detail below.
Let us assume input data is t j , y j , and v j corresponding e.g. to time, value, and error (1σ).For a given part of the input data an average error, v, and an average data spacing, ∆t, can be computed.The link between the cutoff period (P cutoff ), the data spacing (∆t), the 1σ error of the input data (v) is: (1) A positive aspect of Eq. 1 is that P cutoff depends only weakly on ∆t.
Let us now assume we have a data set where different parts or intervals have very different data spacing and/or for which we would like to apply a different smoothing.We can then modify λ to define an individual P cutoff for each part of the input data set.
Reference interval: λ is computed using Eq. 1 for the given cutoff period, average data spacing, and average error for this first interval.It follows: In the following, the reference interval is always the most recent time window (starting with the beginning of year 2016 CE) covered with instrumental measurements.
Other intervals: A modified λ = λ • s 2 with λ taken from the reference interval is used here, implying that for the reference interval s = 1 and λ = λ.The scaling factor s is chosen to gain the desired P cutoff after following that s is prescribed as: where ∆t and v are the mean data spacing and the mean error for the interval under consideration.2013) are all labeled as as "Law Dome" data for simplicity, although these two studies show data from the Law Dome deep ice core and from various shallow cores are combined with atmospheric data from Cape Grim, and South Pole.Please refer to the original publications for a precise characterisation of the sample origins.

Atmospheric CO 2
There are offsets in measured CO 2 of unidentified causes between records obtained from different ice cores (e.g.Ahn et al., 2012;Bereiter et al., 2012;Marcott et al., 2014;Bauska et al., 2015).These offsets may be related to inter-laboratory differences in the calibration, or potentially due to in situ artefact production of CO  (Bazin et al., 2013;Veres et al., 2013).The CO 2 record from WDC is used here on its more recent age scale WD2014 to have the timing of CO The CO 2 data contributing to this spline are the following (see also Table 2 for details): 1. We start our CO 2 data compilation at present day (end of year 2015 CE = beginning of year 2016 CE, or −66.0 BP).
From the NOAA network instrumental monthly CO 2 data are available online (Dlugokencky et al., 2016b).We here choose to take only the data of the original so-called "Keeling-Curve" started by C.  (Dlugokencky et al., 2016b).Hence, CO 2 data from Mauna Loa are only a (very good) approximation to a global value which might be used for calculating radiative forcing.However, in practise this interhemispheric difference does not matter, since no information on it exists beyond the instrumental period: CO 2 is only measured on ice cores from Antarctica, as the higher dust content can give rise to artefacts in any CO 2 measurement based on Greenlandic ice cores (e.g.Anklin et al., 1995;Stauffer et al., 1998).
Earth Syst. Sci. Data Discuss., doi:10.5194/essd-2017-6, 2017 Open Access ( Myhre et al., 1998) is 0.15 or 0.09 W m −2 for a reference concentration of 180 or 280 ppm, respectively.This is larger than the relative uncertainty (precision) of the order of 1 ppm attached to individual data points which is used to determine the smoothing spline through the data.
6. Further back in time all ice core records attached below have some data overlap with the previous record.There are some small offsets between these different records (for details see Bereiter et al., 2012).We treat them all alike so the spline averages over all cores and choose to select a rather larger cutoff period of 2000 years for the time between LGM and the end of the previous interglacial to account for those uncertainties.Rapid variations in CO 2 in glacial times (Fig. 2D-F) are best covered in the Siple Dome record between 21.9 and 48.7 kyr BP (Ahn and Brook, 2014), the Talos Dome record between 34.4-69.7 kyr BP (Bereiter et al., 2012), and the EDML record between 43.2-113.4kyr BP (Lüthi et al., 2010;Bereiter et al., 2012).Talos Dome CO 2 data include some outliers which disagree with CO 2 records from other ice cores by more than 10 ppm.Therefore, Talos Dome data are only considered for the times older than 38.0 kyr BP.
7. From 104.3-156.3kyr BP -covering the last glacial inception, the last interglacial, as well as Termination II and the penultimate glacial maximum (Fig. 2F) -CO 2 is best covered by data from EDC (Schneider et al., 2013;Lourantou et al., 2010a).
For every supporting data point j a 1σ uncertainty or error v j has to be assigned in order to be able to calculate the smoothing spline (see section 2 for details).Therefore, a nominal uncertainty of 0.3 ppm is assigned to the Mauna Loa data, following information given on the NOAA website.Uncertainties for the ice and firn data are taken either as reported or set to a minimal value of 0.5 ppm if the reported standard deviation is missing or less than 0.5 ppm.For Law Dome data published in MacFarling-Meure et al. (2006) we take a uniform uncertainty of 1.1 ppm as reported in their methods section.
The data selection as described above then leads to n = 2152 data points including 20 ages with duplicate entries.These duplicates are removed (reducing n to 2132) and their CO 2 values are calculated from the average of the single values with the assigned uncertainties based on this averaging.
To account for the given age spacing ∆t or data frequency of the data points (Fig. 1B) and the knowledge of abrupt changes in CO 2 during Termination I, the spline is divided into 12 intervals with different nominal cutoff periods, which varied between 4 years for instrumental times and 3000 years during most of the Holocene.A low P cutoff of 600 years could be assigned to the now well-covered interval of Termination I (11-18.5 kyr BP).For the glacial times older than 18.5 kyr BP P cutoff of 2000 years has been selected, whose usage is extended until 110 kyr BP.For the warm interglacial between 110 and 128 kyr BP we assign the same cutoff period of 3000 years as has been chosen before for the recent interglacial, the Holocene.Thereafter, we Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017 Open Access A summary of all details on the calculated spline is found in Table 3.
The final CO 2 record of the last 2 kyr to be used within CMIP6 (Meinshausen et al., 2016) is (i) within instrumental times nearly indistinguishable from our spline (Fig. 2A) and (ii) in the pre-anthropogenic time of the last 2 kyr partly larger by a few ppm than our spline (Fig. 2B).This difference is readily explained by the underlying data and the different filtering.While we here use a mixture of Law Dome and WDC data between 200 and 1210 BP, only Law Dome data are considered for CMIP6.
The CO 2 values chosen as boundary conditions for several time slice experiments within PMIP4 (Ivanovic et al., 2016;Otto-Bliesner et al., 2016) can be compared with snapshots from our splines (Table 4).However, one needs to be aware that some short-term fluctuations in our spline might offset the values from long-term averages and lead to differences between our final splines and the PMIP4 forcing data.For the Mid-Holocene (6 kyr experiment) both our approach and that of PMIP4 are based on the same EDC data and processed with the identical spline routines and cutoff frequencies, leading to identical values.Values differ by a few ppm for the experiments 1850 CE, 21 kyr and 127 kyr.
Since spline smoothing is a low-pass filter, abrupt changes in CO 2 are always smaller in the spline than in the original data sets.Therefore, if one wants to investigate in great detail the impact of very fast and abrupt rises in CO 2 on the climate system that have been identified during three points in time (around 11.6, 14.7 or 16.2 kyr BP) across Termination I (Monnin et al., 2001;Marcott et al., 2014) an alternative continuous CO 2 record needs to be constructed.One approach might be to further reduce the cutoff period to lower values so that the spline would include these pronounced jumps.In detail, one might want to obtain a rise in CO 2 of 12 and 13 ppm during one century at 16.2 kyr BP and 11.6 kyr BP, respectively, as identified in the WDC record (Marcott et al., 2014).For the abrupt rise in CO 2 around 14.7 kyr BP even an artificial rise of 15 ppm in 200 years, slightly larger than the 12 ppm of the WDC record, has been suggested to represent atmospheric changes in CO 2 potentially caused by permafrost thawing during the northern hemispheric warming into the Bølling/Allerød (Köhler et al., 2014(Köhler et al., , 2015)).Transient simulations forced by records containing these abrupt jumps in CO 2 might be able to investigate their impact in greater detail than simulations forced with our low-frequency spline.

Atmospheric CH 4
Our data compilation of CH 4 data and the consistently calculated CH 4 spline is restricted to the southern hemisphere (SH).
Northern hemispheric (NH) data are shown for comparison, but are not included in the spline, since for such efforts chronologies of ice cores from both hemispheres have to match perfectly during abrupt climate changes of the D/O events.However, as has been shown (Baumgartner et al., 2014)  was up to +4% (+14 ppb) and +10% (+60 ppb) larger than in the SH during glacial times and the Holocene, respectively.Using an approximation of the radiative forcing Our data compilation starts, identical to that for CO 2 , (end of year 2015 CE = beginning of year 2016 CE; −66.0 BP) and stops around 156 kyr BP, to cover the same time window as for CO 2 (Fig. 3A).The 11-points-running mean age distance between neighbouring data points, ∆t, is less than 100 year for most of the last 60 kyr, increasing to ∼700 years further back in time (Fig. 3B).Our strategy here is to select one (best) data set for each point in time, and use overlapping intervals only for confirmation of data consistency and the absence of any offsets in CH 4 .In detail, the following data sets are considered here: 1  Marcott et al., 2014;Buizert et al., 2015;Mitchell et al., 2013Mitchell et al., , 2011;;Sigl et al., 2016).Right now WDC CH 4 are the highest resolved data of the last glacial times and therefore our reference record (Fig. 4B-E).The data do not only contain the well known abrupt CH 4 changes at the onset and end of the millennial-scale D/O events in high resolution and accuracy, but also centennial-scale features, which are also understood to be of climatic origin (e.g.Mitchell et al., 2013). 5.The NH Greenland composite of CH 4 (Blunier and Brook, 2001;Dällenbach et al., 2000;Flückiger et al., 2004;Landais et al., 2004) is only plotted for comparison to the SH data (Fig. 4B-F).

Atmospheric N 2 O
For the data compilation of the third GHG, N 2 O, one has to be aware that during times of high dust input in situ production of N 2 O might occur leading to artefacts in the paleo record (Schilt et al., 2010a).Furthermore, the precise synchronisation of artefacts are reported in the pen-ultimate glacial maximum (Schilt et al., 2010a).The latter is also the case for EDML, whose data have not been taken to support the spline, since in large parts N 2 O of EDML and EDC agree, and the data from EDML have a lower temporal resolution than those of EDC (Schilt et al., 2010a).
In detail, the data sets contributing to the N 2 O stack are the following: 4. The highest resolved N 2 O record for large parts of Termination I are based on the horizontal ice core from Taylor Glacier (Schilt et al., 2014) which has been linked to the chronology of the WDC ice core (WD2014) via CH 4 (Baggenstos, 2015).In detail, the Taylor Glacier N 2 O record covers the time window 9.6 to 15.8 kyr BP (Fig. 6C) and is taken to support our spline.
5. The last glacial interval is well covered by data from the NGRIP record from Greenland (Flückiger et al., 2004;Schilt et al., 2010bSchilt et al., , 2013)).While the data cover the times between 11 kyr BP and 119.6 kyr BP, we only take those data older than 15 kyr BP due to the higher resolved Taylor Glacier N 2 O data during Termination I (Fig. 6C-F).Furthermore, five data points near the bedrock in the bottom part of the NGRIP records have apparent higher N 2 O values than found in ice cores from the southern hemisphere.These data points are rejected here, leading to the oldest NGRIP N 2 O data point at 118.6 kyr BP.We are aware that due to the imperfect north-south synchronization of gas records in AICC2012 (see N 2 O data, with Taylor Glacier data in-between both (Fig. 6C).We therefore believe, that a mixture of all three records, N 2 O from NGRIP, EDC and Talos Dome, most likely covers a reasonable mean global N 2 O value (Fig. 6C-F).The relatively large difference in N 2 O from different ice cores during the last glacial times indicates, that the uncertainty (accuracy) in N 2 O is probably higher than the reported measurement errors (precision) of about 10 ppb.
The general assigned 1σ uncertainty of each data point is 7 ppb for the Law Dome ice core (MacFarling-Meure et al., 2006).
Uncertainty of individual data points in other ice cores was in general less than 7 ppb (Flückiger et al., 2002;Schilt et al., 2014).For 58 times more than one data point for the same age exists.These duplicates are removed and the calculated mean and standard deviation of the averaging procedure (if larger than reported measurement error) are taken, reducing the amount of N 2 O data to n = 2344.For the instrumental measurements we take the reported uncertainties of around 1 ppb.Using an estimate of the radiative forcing of N 2 O, which neglects the interacting effects of CH 4 and N 2 O In MIS5.5, Termination II and the penultimate glacial maximum the mean age distance rises to ∼500 years (Fig. 5B).Based on this distribution of ∆t the prescribed cutoff periods for the spline varies for 7 different intervals between 4 years for the instrumental times to 5000 years for data older than 117 kyr BP.For a large part of the data (400 yr BP to 117 kyr BP) a P cutoff between 500 and 2000 years is prescribed.More details on the spline are found in Table 8.
If compared with the N 2 O compilation used within CMIP6 (Meinshausen et al., 2016) both approaches largely agree for instrumental times (Fig. 6A).Further back in time during the last 2 kyr, both approaches rely on the same data: the published Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017 Open Access  , 2006).Interestingly both time series differ by up to 6 ppb between 0.7 and 2.0 kyr BP (Fig. 6B).This difference is in the range of the ice core data uncertainty, and therefore still small, but we have no easy explanation at hand.Compared with our data compilation the records used in CMIP6 are higher than all data from Law Dome or other SH ice cores (Fig. 6B), for some apparent unknown reason.

Conclusions
Based on our best knowledge we have compiled available greenhouse gas records and by calculating a smoothing spline we were able to provide continuous records over the last glacial cycle, starting from the beginning of year 2016 CE and ranging back in time until 134 kyr BP for N 2 O and until about 156 kyr BP for CO 2 and CH 4 .These records should serve as radiative forcing in transient climate simulations as planned, for example, in the German project PALMOD, or might be used in the Last Deglaciation experiment within PMIP4 (Ivanovic et al., 2016) or in other future MIP.
The resulting radiative forcing of the three GHGs calculated here with the simplifed non-interacting Eqns.5-7 have relative uncertainties of ∼10% (Myhre et al., 1998).The results are very similar to recent calculations based on a complete and revised set of simplified equations, which consider also interacting effects between the three GHGs (Etminan et al., 2016), with differences between old and new expressions in ∆R [CO2] , ∆R [CH4] , and ∆R [N2O] of less than 0.01, 0.04, and 0.02 W m −2 , respectively.While the difference in ∆R [CO2] and ∆R [N2O] lie within their uncertainty bands, that of ∆R [CH4] is slightly higher, leading also to a revised relative uncertainty of 14% (Etminan et al., 2016).We nevertheless refrain from applying the new equations throughout the study, since the amplification in ∆R [CH4] by 40% through indirect effects of CH 4 (Hansen et al., 2005) is not yet considered therein, and we prefer to estimate the radiative forcing of one of the three GHGs individually without interacting effects.These forcing calculations nicely illustrate the dominant contribution of CO 2 , which is responsible for about two thirds of the total radiative forcing ∆R [GHG] during both the anthropogenic rise (Fig. 7A) and the reduction during the LGM (Fig. 7B).The higher resolved variability in CH 4 , resulting from smaller cutoff periods during spline calculations than for CO 2 , also imposes some fine-scale structure on the overall GHG radiative forcing (Fig. 7C,D), however, dominant features are still mainly provided by CO 2 .
Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017 Open Access led to the preparation of Table 1 (summary on data sites).This is a contribution to the German BMBF project PALMOD.
Long-term support of ice core research at the University of Bern by the Swiss National Science Foundation (SNF) is gratefully acknowledged.

Data Availability
Data connected with this paper are available in the scientific data base PANGAEA (https://doi.pangaea.de/10.1594/PANGAEA.871273).
In detail, for each of three GHGs the following data are available: -Finally compiled raw data (t j , y j , v j corresponding to time, value, assumed 1σ-error), including the data source, as described in the article.
-Pre-processed raw data (averaging of duplicate entries for similar times) -Calculated splines with time steps of ∆t = 1 year.
When using these data please consider citing the original publications from which the data underlying this compilation have been taken.

Author Contribution
PK initiated the work, compiled the data, calculated the spline and led the writing of the manuscript.CNA, JS, TFS, and HF contributed specific insights on the data selection and advised the spline smoothing.All co-authors commented on and improved the initial draft.

14
Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017 Open Access  Notes: 1: The data compilation of MacFarling-Meure et al. (2006); Rubino et al. (2013) on CO2, SH CH4, and N2O uses data from the Law Dome deep ice core and from various shallow ice and firn core in its vicinity, but also from atmospheric data from Cape Grim and firn core data from South Pole.While we here state all the relevant positions, the original source of the individual data points is not marked in Tables 2, 5 and 7, or in the data files uploaded to PANGAEA, but then only labeled with "Law Dome" as data source.Please see the original reference for further details.Table 3. Statistics of the CO2 spline.Interval iCO 2 ; s: scaling factor to fulfil the constrains given by the prescribed P cutoff ; P cutoff : average realised cutoff period.∆t: mean data spacing; v: mean 1σ error; exact time framing is given by the age of the first (tstart) and last (tstop) data point of the interval (in years BP); N : number of data points within interval.In the last column the underlying data source is briefly mentioned, see 1: In the discussion version of this paper a value of 284.6 ppm was given, which has been revised to the here stated value of 284.3 ppm.

22
Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017 Open 2 and the other two GHGs consistently on the same chronology.Using WD2014 instead of the original chronology WDC06A-7 leads to a shift of the timing of the deglacial CO 2 rise during Termination I by about 100 years towards younger ages.The related age difference for an update of the WDC chronology to WD2014 is for CO 2 during the last 1.2 kyr only about 10 years.Our CO 2 data compilation goes back in time until ∼156 kyr BP, at which point in time well-resolved CO 2 records stop.The full CO 2 spline covering the whole time window from 2016 CE to 156 kyr BP is plotted in Fig. 1, including the age distance, ∆t, of the compilation of data points in sub-panel B. The 11-points-running mean of ∆t is around 100 years in the Holocene, between 20 and 50 years during Termination I, and varies between 40 and 200 years between 20 and 70 kyr BP, rising to 500 to 1000 years later-only with a slightly higher resolution of 200 years only across Termination II.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.main focus on δ 13 C (Lourantou et al., 2010b; Schmitt et al., 2012), which have a somewhat lower precision in CO 2 concentrations.More details on this adjustment of the WDC data during Termination I are found in Figure A2.These offset corrections imply that the absolute CO 2 concentration is uncertain by about 5 ppm (accuracy).The corresponding radiative forcing following review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.continue with a 1000 years cutoff period for the CO 2 rise during Termination II (128-135 kyr BP), before again 2000 years are chosen during the penultimate glacial maximum.
Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6forjournal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.which neglects the interacting effects of CH 4 and N 2 O (Myhre et al., 1998), but which considers the approximately increase in ∆R [CH4] by 40% through indirect effects of CH 4 on stratospheric H 2 O and tropospheric O 3 (Hansen et al., 2005; Köhler et al., 2010), we estimate that this restriction of CH 4 to the SH only might underestimate the radiative forcing of CH 4 by less than 0.05 W m −2 .

4 .
We continue our SH CH 4 data compilation at the end of what is covered in WDC with data from EDC, covering 67 kyr BP -156 kyr BP(Loulergue et al., 2008) (Fig.4E-F).These EDC data would in principle go back in time over the whole 800 kyr covered by the EDC ice core, but since our focus here is on approximately the time since the penultimate glacial maximum (or approximately the last 156 kyr), those data further back in time are not taken into consideration.Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6forjournal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.
Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6forjournal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.northern and southern hemisphere records, as already explained for CH 4 , is crucial to get changes in N 2 O during millennialscale D/O events right.The compiled record, again, starts at the end of year 2015 CE ( = beginning of year 2016 CE; −66.0 BP), but extends back in time only until ∼134.5 kyr BP (Fig. 5A), because the ice cores on which the N 2 O compilation is based on in the older parts, Talos Dome, EDC and NGRIP, have no data points between 134.5 and 156 kyr BP, or only unreliable N 2 O data containing

1. 3 .
There are two contributions of N 2 O data based on instrumental measurements to the NOAA network or ESRL halocarbon program: (a) in situ data from 2016 CE back until 1999 CE, and (b) the RITS N 2 O data from 2000 CE back until 1988 CE, both representing global mean monthly values (Fig. 6A).Note that due to the long atmospheric lifetime of N 2 O, any interhemispheric gradient can be safely neglected.2. Law Dome and Cape Grim N 2 O data exist from 2004 CE back until 13 CE (= 1937 BP) (MacFarling-Meure et al., 2006),overlapping nicely with the instrumental data (Fig.6A,B).Here, the Law Dome data contribute to the spline only for those years not covered by the instrumental record, so from 1983 CE and earlier.In the Holocene N 2 O was measured at EDC(Flückiger et al., 2002) from 334 BP until 11.5 kyr BP.For the last two millennia the EDC N 2 O data points are sparser than the Law Dome data, therefore the EDC N 2 O data are only considered for times older than what is covered in the Law Dome N 2 O record, in detail before 1975 BP (Fig. 6B,D).

subsection 3 .
2 for details) the usage of N 2 O data from NGRIP might introduce erroneous phasing between N 2 O and the purely SH-based CH 4 spline during abrupt change connected to D/O events.However, N 2 O data coverage in the SH core Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6forjournal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License. is very sparse and a spline only based on SH data seems to be even less reliable.This potential synchronization problem is also addressed by large cutoff periods of the spline of 2000 to 5000 years beyond 16 kyr BP. 6.Additional N 2 O data going back to 134.4 kyr BP are obtained from the Talos Dome ice core (Schilt et al., 2010b) and from further data of the EDC ice core (compilation found in Schilt et al. (2010a); data source between 29.0-134.5 kyr BP: Stauffer et al. (2002); Spahni et al. (2005)).Since these are -besides EDML which very much agrees with EDC(Schilt et al., 2010a) -the only N 2 O records with reliable data going back into the penultimate glacial maximum we consider all data points from these ice cores here before 15 kyr BP.The data points of Talos Dome and EDC in general agree with the NGRIP data over the last glacial cycle, but NGRIP diverges from the SH records towards higher (probably biased) values in the warm previous interglacial around 115 kyr BP.As already explained above, these 5 NGRIP data points are rejected.However, across Termination I Talos Dome N 2 O data seems to be systematically lower than NGRIP

(
Myhre et al., 1998), we estimate that the 1σ error in N 2 O is related to an uncertainty in the radiative forcing of about 0.04 W m −2 , so slightly larger than the uncertainty in ∆R related to the CH 4 data.Comparing the different values of N 2 O in Talos Dome and NGRIP for same points in time reveals also differences on other order of about 10 ppb (e.g.Fig.6C-F), suggesting that the ice core specific values of N 2 O contain an intrinsic uncertainty of similar size than the measurement error.The mean age distance (11-points-running mean) of the underlying N 2 O data is around 50 years during large parts of the last glacial cycle (15-60 kyr BP), with slightly lower resolution of 100 years in the Holocene and between 60-115 kyr BP.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.Law Dome / Cape Grim N 2 O data (MacFarling-Meure et al. review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.

2 :
Data taken from Taylor Glacier are based on a "horizontal ice core", which has not a point location as all other sites.20 Earth Syst.Sci.Data Discuss., doi:10.5194/essd-2017-6forjournal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.
review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.

3
Greenhouse gas data compilations and final splinesOur GHG data compilations are based on various different data sets from thirteen global distributed locations.An overview of the locations, including latitude/longitude is obtained in Table1.Please note, that from some of those locations CH 4 data are taken only for comparison, but are not supporting the spline, since only a southern hemisphere spline is constructed.They are furthermore supported for the instrumental period by some global mean data from the NOAA observational network, including RITS Nitrous Oxide data from the NOAA/ESRL halocarbons program and Nitrous Oxide data from the NOAA/ESRL halocarbons in situ program, consisting of various globally distributed measurement sites.Furthermore, at the level of the individual point in the files uploaded to the data base PANGAEA, the entries from MacFarling-Meure et al. (2006) and Rubino et al. ( Bereiter et al. (2012)or a detailed discussion seeAhn et al. (2012)and the supplement toBereiter et al. (2012).Moreover, amplitudes of GHG variations may differ from one core to the other due to the site dependent bubble enclosure characteristics, which leads to more or less low-pass filtering.Offsets require the adjustment of individual records to avoid spurious CO 2 changes when linking different records from different laboratory and ice cores.Ice core data are considered here on the best (most recent) age model available, whose details are contained in Table2.AICC2012 refers to the most recent Antarctic Ice Cores Chronology providing age models for EDC, EDML, Talos Dome, Vostok, alongside the Greenlandic NGRIP record this is not the case for the most recent chronology AICC2012.NH CH 4 and consequently global CH 4 values should according to the well-known interhemispheric gradient be larger than our SH CH 4 values.Therefore, our SH CH 4 spline consists of a lower bound of CH 4 values.Baumgartner et al. (2012) found that NH CH 4 [CH4]given above these 1σ error are related to uncertainties in the radiative forcing of less than 0.01 W m −2 .Compiled data contain 2990 data points, from which for 37 ages duplicate entries exits.These duplicates are removed giving n = 2953 and for those ages new values (mean, uncertainties) are calculated.The whole data set is divided in nine intervals with different assigned cutoff periods.P cutoff ranges from 4 years for the interval covered by instrumental data to 60 years during Termination I and the LGM.Due to lower data coverage further back in time P cutoff is then increased to 200 (23-60 kyr BP), 500 (60-128 kyr BP), and 1000 years (128-156 kyr BP).More details are shown in Table 6.Bliesner et al., 2016) is not straight-forward, since we here only compile SH CH 4 data, while the PMIP4 experiments need and use global values.Taking the two records at face value one finds that our SH CH 4 is 13, 44, 25 ppb smaller than the global mean value used in PMIP4 for 1850 CE, 6 kyr, 127 kyr, respectively.Especially, the large SH-global difference of 44 ppb around 6 kyr seems to be rather large, but is readily explained by the centennial variability contained in the WDC CH 4 which leads to a local minimum in SH CH 4 around 6 ka.Furthermore, our SH CH 4 spline is 7 ppb higher than the global CH 4 value chosen within PMIP4 for the 21 kyr experiment, again explained by the centennial-scale variability contained in the WDC CH 4 record, with a local maximum just at 21 kyr BP.Hundred years later, our SH CH 4 spline has a local minimum which is 11 ppb smaller than the global CH 4 values taken for PMIP4 (Table (Ivanovic et al., 2016;Otto-d within CMIP6(Meinshausen et al., 2016)largely agrees with our SH spline (Fig.4A,B).However, during instrumental times the CMIP6 SH CH 4 record is consistently larger than our SH spline by about 10-15 ppb, probably caused by the consideration of various different stations in the calculation of the SH CH 4 record within CMIP6, while we here only rely on South Pole data.Prior to the instrumental CH 4 data around 1980 CE the difference between both approaches is with 30 ppb largest.This might be caused by the statistical routines within CMIP6 to account for missing stations.Further back in time (around 1150 BP, 1300 BP, 1900 BP) higher frequency variation contained in the WDC CH 4 record (used here, but ignored within CMIP6) leads to some CH 4 variabilities on the order 10-25 ppb within our SH spline, that are not contained within the CMIP6 SH CH 4 record.A comparison of our final spline with the GHG values chosen for the PMIP4 time slice experiments(Ivanovic et al., 2016;Otto-

Table 1 .
Locations of the different data sources, ordered north to south.Individual sites of the NOAA oberservational network are not explicitly mentioned here, when they only contribute to global mean calculations.SH CH4: Southern Hermisphere CH4.

Table 4 .
(Otto-Bliesner et al., 2016)ions.Comparison of our final spline data with values used for PMIP4 experiments for 21 kyr(Ivanovic et al., 2016)and 1850 CE, 6 kyr, 127 kyr(Otto-Bliesner et al., 2016).Please be aware that the PMIP4 values should be millenniual-scale mean numbers to serve as forcing values for time slice experiments, while the values depicted from this study are snapshots of the given points in time.Furthermore, we

Table 5 .
Data used to construct (or compare to) the Southern Hemisphere CH4 spline.Law Dome data are taken from various sources, see references for details.They exist for 2005 CE to 14 CE (or -55 BP to 1936 BP), but only the part bridging the gap between instrumental data and WDC are taken for the calculation of the spline(1982 CE to 1782 CE or -32 23EarthSyst.Sci.Data Discuss., doi:10.5194/essd-2017-6,2017

Table 6 .
Statistics of the CH4 spline.Interval iCH 4 ; s: scaling factor to fulfil the constrains given by the prescribed P cutoff ; P cutoff : average realised cutoff period.∆t:meandata spacing; v: mean 1σ error; exact time framing is given by the age of the first (tstart) and last (tstop) data point of the interval (in years BP); N : number of data points within interval.In the last column the underlying data source is briefly mentioned, see Table5for details and citations.

Discussions
Manuscript under review for journal Earth Syst.Sci.Data Discussion started: 20 February 2017 c Author(s) 2017.CC-BY 3.0 License.

Table 8 .
Statistics of N2O spline.Interval iN 2 O; s: scaling factor to fulfil the constrains given by the prescribed P cutoff ; P cutoff : average realised cutoff period.∆t:meandata spacing; v: mean 1σ error; exact time framing is given by the age of the first (tstart) and last (tstop) data point of the interval (in years BP); N : number of data points within interval.In the last column the underlying data source is briefly mentioned, see Table7for details and citations.