Recent developments in ice melter systems and continuous flow analysis (CFA) techniques now allow higher-resolution ice core analysis. Here, we present a new method to aid interpretation of high-resolution ice core stable water isotope records. Using a set of simple isotopic recording and postdepositional assumptions, the European Centre for Medium-Range Weather Forecasts' 40 year reanalysis time series of temperature and precipitation are converted to “virtual core” depth series across the Antarctic Peninsula, helping us to understand what information can be gleaned from the CFA high-resolution observations. Virtual core temperatures are transferred onto time using three different depth-age transfer assumptions: (1) a perfect depth-age model, (2) a depth-age model constructed from single or dual annual photochemical tie points, and (3) a cross-dated depth-age model. Comparing the sampled temperatures on the various depth-age models with the original time series allows quantification of the effect of ice core sample resolution and dating. We show that accurate annual layer count depth-age models should allow some subseasonal temperature anomalies to be recovered using a sample resolution of around 40 mm, or 10–20 samples per year. Seasonal temperature anomalies may be recovered using sample lengths closer to 60 mm, or about 7–14 samples per year. These results tend to confirm the value of current CFA ice core sampling strategies and indicate that it should be possible to recover about a third of subannual (but not synoptic) temperature anomaly information from annually “layer-counted” peninsula ice cores.
If you can't find a tool you're looking for, please click the link at the top of the page to "Go to old article view". Alternatively, view our Knowledge Base articles for additional help. Your feedback is important to us, so please let us know if you have comments or ideas for improvement.
 The Antarctic Peninsula has experienced significant climate change in recent decades. Since continuous meteorological monitoring began in the late 1950s, the mean annual temperature on the west coast has risen by more than 2°C [Vaughan et al., 2001, 2003; Turner et al., 2005], faster than anywhere else in the southern hemisphere. However, our understanding of rapid and ongoing regional changes in this remote region are hindered by a lack of long-term meteorological observations [Schneider et al., 2004; Steig et al., 2009; Thomas et al., 2009].
 The stable water isotope composition of precipitation (for brevity referred to hereafter as δ to indicate either δ18O or δD) can provide a valuable record of past surface temperatures. High-resolution δ records in ice cores can extend the climate history beyond the observational era [Thompson et al., 1994; Mulvaney et al., 2002; Schneider et al., 2006; Vinther et al., 2010; Thomas et al., 2009]. Where air parcels travel in isolation, δ in precipitation is controlled by temperature differences between the evaporation and condensation sites [Dansgaard, 1964; Jouzel et al., 1997]. If evaporation site temperature remains approximately constant, which is considered likely for centennial length studies in the Antarctic vicinity [Schneider et al., 2006], δ in precipitation is largely dependent on the condensation temperature [Dansgaard, 1964]. Since condensation or inversion temperature and site temperature are closely related [e.g., Connolley, 1996; Helsen et al., 2007] the site surface temperature can be used to represent temperature during condensation. It is therefore likely that the δ climate record is largely representative of site temperature when precipitation falls. Across the Antarctic Peninsula, records of δ from high-accumulation ice cores sites have provided an excellent resource for reconstructing past climate [e.g., Swithinbank, 1977].
 Traditionally, δ in cores has been discretely sampled by manually cutting the cores into short sections. Once the samples have been cut, they are melted and passed through an isotope ratio mass spectrometer. Typical discreet sample lengths have been between 10 and 500 mm of the core depth [Littot et al., 2002]. Recent developments using ice core melter systems [Kaufmann et al., 2008; McConnell et al., 2002; Röthlisberger et al., 2000; Sigg et al., 1994] allow measurements using continuous melt flow at higher resolution [Kaufmann et al., 2008]. Continuous flow analysis (CFA) techniques can allow cheaper high-resolution analysis of physical and chemical constituents in ice cores at millimeter scales [Federer et al., 2008; Kaufmann et al., 2008]. These technical advances in CFA techniques suggest a new need to understand what information can be gleaned from high-resolution δ observations. In particular, to address the question: What resolution of δ measurement is required to achieve good reconstructions of past site temperature?
 Here, to help address that question, we develop a new method for constructing virtual cores and apply it to ECMWF (European Centre for Medium-Range Weather Forecasts) ERA-40 reanalysis data [Uppala et al., 2005]. The method is applied to the Antarctic Peninsula region, where ice core sites have been drilled by the British Antarctic Survey between 1986 and 2008 (Figure 1). Results allow diagnosis of how sampling resolution, and depth-age model uncertainties, may affect temperature information recovery from ice core stable water isotope records.
2. The Virtual Core Method
 Using simple assumptions, the relationship between depth and time is mapped out for “virtual cores.” Knowledge of the depth-age virtual core structure allows testing of the potential for temperature information recovery from δ in ice cores, for example due to sampling or depth-age model specifications.
 Neglecting any influence on δ other than site temperature during precipitation, the temperature recorded in precipitation (TP) and δ in ice cores are directly equivalent [Sime et al., 2008, 2009a]. Using this assumption, the impacts of ice core location and δ sampling resolution can be investigated by analyzing time series of temperature and precipitation, such as those from a weather station or climate model.
 Accumulation depth, Z is the integral of precipitation P over time t, from time start ts to time end te, divided by the snow density ρ (in g per cm3),
This yields the relationship between time (t) and depth (Z) for any ice core site, assuming no postdepositional losses or mixing. The relationship is often referred to as the depth-age model.
 To examine how the accumulation pattern at a particular site influences the preserved temperature history, the precipitation weighted temperature can be calculated over specific discrete depth intervals (dz) as:
 If this calculation is applied over a consecutive set of depth intervals, we can obtain a depth profile. Note that P and T are the precipitation and temperature that occurred over the site, over the specified time interval now represented by the ice core depth interval zb (bottom depth) to zt (top depth). We refer to these profiles of Tp (equation (2)) against depth (equation (1)) herein as virtual cores. Once the impact of diffusion is taken into account (see section 3.2.3), if Tp is considered equivalent to δ, virtual core profiles are equivalent to sampling ice core δ using the sample length resolution dz. Note site location (x, y) is omitted from equations for clarity, however all calculations are carried out on a location by location basis.
 A brief description of the ERA-40 data used here, and a description of the peninsula ice core characteristics which feed into the analysis, is provided below. The peninsula ice core characteristics largely determine the shape of the analysis.
3.1. ECMWF ERA-40 Reanalysis Data
 The virtual core method above can be applied to any given time series of temperature and precipitation, such as from a climate model or a weather station. For this initial application, 22 years (1980–2002) of surface air temperature (T) and precipitation (P) data (Figure 2) from the ERA-40 reanalysis are used [Marshall, 2003; Bromwich and Fogt, 2004]. The ERA-40 reanalysis is thought to be reliable at the high southern latitudes after 1979, when satellite observations were assimilated into the model. It is considered more accurate than National Center for Atmospheric Research–National Centers for Environmental Prediction (NCAR-NCEP) reanalysis for this region [e.g., Bromwich and Fogt, 2004; Miles et al., 2008].
 A comparison between the observed and reanalysis values in Table 1 suggests that ERA-40 tends to underestimate total precipitation over the Antarctic Peninsula, with the average underestimation around 26%. Detailed topographic features of the peninsula, such as James Ross Island, and the height of the peninsula are not well represented: the ERA-40 grid size (geographical resolution) leads to orographic deficiencies which cause the ERA-40 Antarctic Peninsula to be too dry [Orr et al., 2008].
Table 1. Mean Annual Site Characteristics for Antarctic Peninsula Ice Core Site Observations and 22 Years of ERA-40 Reanalysisa
 A study using the virtual core method requires application of suitable sampling and dating characteristics.
3.2.1. Range of Ice Core Sample Lengths
 The recent developments in ice core CFA, generating the ability to sample cheaply at high resolution, motivate this study. CFA instrumental limits typically allow for 1–10 mm lengths of sample resolutions [Federer et al., 2008; Kaufmann et al., 2008]. Past Antarctic Peninsula studies sampling by discrete ice cutting have generally used sample lengths between about 10 and 250 mm.
 By varying the length of the sample (dz), the influence of these different sampling resolution decisions on the temperature reconstructions from ice core δ records can be investigated. The range of sample lengths used here generally covers from 1 to 250 mm.
3.2.2. Snow Density Values
 Snow densification for the Antarctic Peninsula means that, while fresh snow can have a density of 0.3 g per cm3, average surface snow ρ values are about 0.5 g per cm3. Once firnification is complete, maximum ice ρ values are about 0.91 g per cm3. However, since we are working with ideal data over a regional scale, and are only here comparing virtual cores with original temperature time series, for clarity and brevity the simple ρ = 1 g per cm3 density value is used throughout this study, i.e., using a pure water equivalent (w.e.) depth. Note, when comparing virtual cores to actual cores, sets of measured density values can be used to either convert the ice core observations to a w.e. depth scale, or alternatively to convert the virtual core depth scale to a realistic depth scale.
 To account for diffusion effects, the virtual core temperatures are smoothed using a running mean 161 mm in length. This means that water molecules (which carry the δ signal) are moved an average of about 81 mm from their initial positions. This diffusion length of 81 mm is chosen to fit with the lengths given by Johnsen et al. . Note, the impact of the running mean filtering very closely approximates the impact of filtering the depth series using a symmetric Gaussian low-pass filter, with a standard deviation of about 60, on a 1 mm resolution. (In both cases the median movement of the water molecule from its initial point, as a result of the filtering, is about 81 mm.) The running mean filter is for us computationally quicker, therefore although the Gaussian filter is more representative of the process of molecular diffusion, given the close numerical equivalence of the results, here we simulate diffusion using the running mean approach. This simulation of “diffusion” smoothing is always applied to Tp interpolated to a 1 mm resolution. The interpolation and diffusion simulation is always applied after Tp is calculated at a specified sample resolution (dz). Note, in future it would be a potentially useful extension of the virtual core framework to fully implement Johnsen et al.  diffusion, but this is outwith the scope of this paper.
 An 81 mm w.e. diffusion length is roughly equivalent to about 90 mm of peninsula ice core, using a more realistic ice core density (of about 0.91 g per cm3), or about 270 mm of very fresh snow with a density of 0.3 g per cm3. While the effect of small-scale postdepositional changes across the Antarctic Peninsula, such as wind redistribution of surface snow, are thought to be mostly small [van Lipzig et al., 2004], this simple simulation of diffusion does therefore also help account for windblown snow mixing effects, e.g., windblown dunes or ripples of heights of less than about 270 mm. Any deeper mixing of the near-surface firn layer, from internal snow pack hoar frost redistribution or melt lenses, may be less well accounted for using this approach.
3.2.4. Dating Peninsula Ice Cores
 The past rates of precipitation over ice core sites are generally unknown. This leads to uncertainty about the depth-age relationship. In these cases the depth-age model could be written as:
where Pest is the estimated precipitation which is accumulating, and e is the resultant unknown depth-age error associated with uncertainties on the precipitation estimate. The virtual core approach is very useful in enabling investigation of the influence of the depth-age errors associated with unknowns in past precipitation rates.
 When past accumulated precipitation rates are not known, depth-age models are usually constructed using some depth-age tie points, i.e., regular recurrent (e.g., annual sea ice) or irregular (e.g., volcanic) markers of known ages are used. Accumulated precipitation estimates are used to fill in the gaps between age markers.
 High-resolution ice core records from the Antarctic Peninsula have been successfully dated using annual tie points from the seasonal cycle of chemical species contained within the core, or sometimes using the δ signal itself. Seasonal variations exist for many species such as dust, ammonia, calcium, and nitrates [e.g., McConnell et al., 2002]. This approach is usually known as annual layer counting. A particularly useful marker is hydrogen peroxide (H2O2); a photochemical species that exhibits a clear maximum (minimum) corresponding to the summer (winter) solstice [e.g., Anklin et al., 1998; Sigg and Neftel, 1988]. In ice core records from cold, high-accumulation sites H2O2 tends to be well preserved, including in cores from the Antarctic Peninsula region [Thomas et al., 2008]. Dating using H2O2 can be carried out using tie points corresponding to precise annual solar markers [e.g., Frey et al., 2006]. An alternative method of dating is cross-dating records against one another. This is effectively a method of attempting to simultaneously use many, somewhat uncertain, tie points.
 Both the annual layer and the cross-correlation tie point approaches are applied in the following work. The impact of nonzero e terms (depth-age errors) on the recovery of temperature information is assessed.
4. Description of Analysis
 By using the daily ERA-40 reanalysis data, alongside the assumptions of no precipitation loss and a constant and uniform ice density, we have complete knowledge of the relationship between depth and time for ERA-40 virtual cores. This perfect depth-age model allows us to comprehensively test the potential for temperature information recovery from δ in ice cores.
4.1. Dating Methods
 Based on the dating methods discussed above, three cases of automated virtual core dating are analyzed and discussed: (1) the simplest case of perfectly dated virtual cores, i.e., no errors on each individual ERA-40 virtual core depth-age model; (2) independently dated virtual cores, using “annual layer counting,” achieved by automatically aligning the record of each virtual core to recurrent age markers and then using an assumption of constant accumulation to fill in the depth-age model between each of these known annual (or semiannual) age markers; and (3) cross-dated records, where it is assumed that one virtual core record can be perfectly dated, and the other records are automatically cross-dated to this target depth-age model using the cross-dating algorithm MATCH [Lisiecki and Lisiecki, 2002].
 The MATCH algorithm provides a largely automated cross-dating method which cross-correlates paleoclimate records [Lisiecki and Lisiecki, 2002]. The MATCH algorithm divides each record (the “signal” and “target”) into intervals and calculates a score for each possible alignment of the intervals (in sequence) from the square of the differences between the points in each matched interval, and the sum of any applied penalties. The lowest score has the optimal alignment. The penalties, which can be auto-calculated, aim to prevent unphysical alignments involving unrealistically rapid changes in the relative accumulation rate between the signal and target cores. MATCH has typically been used for correlating marine records [e.g., Lisiecki and Raymo, 2005; Lang and Wolff, 2010] but has diverse applications and is capable of aligning records with many cycles [e.g., Milkovich and Head, 2005]. MATCH is applied here by assigning the longest (Table 1) target Beethoven ERA-40 virtual core a perfect depth-age model. Other undated ERA-40 virtual core signals are then matched to this one target Beethoven depth-age model, using the auto-calculated MATCH penalties [Lisiecki and Lisiecki, 2002].
4.2. Correlation Analysis
 In each dating case, the analysis presented consists of correlating the ERA-40 virtual core reconstructed temperature record Tp against the ERA-40 original time series of temperature T. The complete knowledge of the relationship between depth z and time t, in each of the three depth-age dating cases, allows transfer of Tp(z) into time coordinates Tp(t) and the correlation analysis to be performed using the original time series T(t) and the reconstructed Tp(t) values. As noted, varying the sample resolution (dz), allows an investigation of the impact of this ice core cutting choice on the recovery of the original temperature signal. The correlation coefficients show how strongly the original time series T(t) and the reconstructed virtual core Tp(z) depth series are related, given a specified depth-age model (as described above). We report the correlation coefficients as R values, and note that the square of R yields the explained variance.
 Results are organized so that the three depth-age model cases are considered in order (see section 4.1). In each case, the depth-age models are generated automatically using solely the assumptions given above and the ERA-40 temperature and precipitation time series. The simulated virtual core diffusion length is always 81 mm (except for Figure 5, where three different diffusion lengths are explored). See section 6 for a general overview of the results.
5.1. Case 1: Using a Perfect Depth-Age Model
 In the perfect dating case, we are asking the question “What is the maximum amount of temperature information that can be recovered from an ice core?,” i.e., what is the maximum recovery potential at different sample resolutions across the Antarctic Peninsula? In the course of examining this, it is useful to consider the effect of the seasonal cycle on results, and generally to address the question of what frequency of anomalies we wish to recover from the ice cores.
 Maps of correlation results at three representative sample resolutions, between 10 and 250 mm, are shown in Figure 3. From Figure 3a to Figure 3c the sample resolution decreases, from an average of around sixty samples per year at dz = 10 mm, to about two samples per year for dz = 250 mm. The correlations reduce with the precipitation gradient, from the northwest to south. Thus the basic geographical pattern of temperature information recovery is dependent on the sample resolution and the precipitation amount (Table 2).
Table 2. Correlations Between Virtual Core T(t) and Tp(t) Using Daily Data on a Perfect Depth-Age Modela
 While the initial analysis above (Figure 3) allows a feel for the sample resolution problem, the results are strongly affected by the mean climatological (i.e., seasonal) temperature cycle. This seasonal cycle is the most easily identifiable mode of temperature variability (Figure 4a, red line). While seasonal cycles allow for annual layer counting dating methods, in examining the question of temperature information recovery, we are actually more interested in temperature anomalies, rather than mean climatological results. Removing the mean seasonal cycle allows a more accurate assessment of ability to recover climatically interesting temperature anomalies from ice cores.
 The mean seasonal cycle can be removed by first constructing a climatological mean by averaging across the 22 year time series, and then secondly subtracting this climatological cycle for each individual year. The use of 30 day moving averages during construction of our mean climatological seasonal cycles also helps counteract noise due to the short 22 year sample size. An example of the mean seasonal cycle removal is shown by the blue line in Figure 4a.
 The strong impact of the seasonal cycle removal can be seen in the difference between the solid lines versus dashed lines in Figure 5a and in comparing Figures 3a–3c against Figures 6a–6c. Correlation coefficients tend to drop from R = 0.7 ± 0.12 to R = 0.2 ± 0.2 (at dz between 20 and 100 mm). Thus, the high R values in Figures 3 and 5a (solid lines) are effectively an artefact of the seasonal cycle.
5.1.2. What Type of Temperature Anomalies Can We Recover?
 After removing the mean seasonal cycle, diffusion effects not withstanding, it is still theoretically possible to pick out temperature excursions for sites without precipitation hiatuses at any frequency. By examining the results from Figure 5a (dashed lines) it can be seen that (without diffusion) synoptic-scale temperature recovery, at R = 0.6, would be potentially possible. R values of 0.6 occur for the highest-resolution dz = 3 mm samples at Dyer, JRI, and Dolleman. ERA-40 precipitation falls frequently at these sites, which enables about one-third of synoptic temperature excursions to be recorded (see also Table 1). This low-precipitation intermittency is reflected in the high correlations.
 However, once realistic diffusion is considered (Figures 5c and 5e, dashed lines), the maximum synoptic correlations drop, indicating that even given sample lengths of 3 mm, synoptic temperature anomalies are not generally recoverable. This is because molecular diffusion of ice cores is effectively equivalent to a form of low-pass filtering. Thus, even for a perfectly dated (virtual) ice core, because of the effects of diffusion, longer-term temperature anomalies can be more reliably recovered from peninsula cores. Low-pass filtering eliminates synoptic time scale temperature signals, leaving only the lower-frequency temperature anomalies, which are more diffusion resilient. (See, in Figure 4, the synoptically affected (thin blue lines) compared to low-pass filtered (thicker gray and black lines) for an illustration.) This lead to the result that correlations between Tp (t) and T(t) are generally higher once both signals have been low-pass filtered.
 The results from Figure 6 (which use the standard 81 mm diffusion length) show how sample resolution, temperature anomaly frequency, and correlation vary together. The pattern between and within the panels depicts how the potential to recover information depends together on precipitation amount, anomaly frequency, and sample resolution. With this same realistic diffusion length, Figure 5f indicates that 60 day low-pass temperature anomalies can be recovered using sample lengths of around dz = 45 mm or shorter (about R = 0.55). For 180 day low-pass temperatures, dz ≈ 50 mm tends to give the highest correlations (about R = 0.75), using the perfect depth-age model.
5.2. Case 2: Using a Photochemical “Layer Counted” Depth-Age Model
 In section 5.1, full use was made of perfect depth-age models. We now consider the impact of likely errors in depth-age models, due to unknown past accumulation rates (equation (3)). This case is based on using annual layer counting tie point information, from independent seasonal photochemical markers. An assumption of constant precipitation is used to fill in the depth-age model between these markers.
 Depth-age models are tied on 1 January, for the once a year tie point approach, and additionally also on 1 July for a twice a year annual solar marker tie point approach. These dates are used to simulate approximate solstice dates [Frey et al., 2006]. Dating error (e), in years, reduces to zero at these depth-age tie points, and increases where the estimated accumulation Pest(t) (using the constant precipitation between markers) deviates from actual P(t). Figure 7 and Table 3 show the impacts on dating accuracy of using this single (Figures 7a, 7c, 7e, 7g, and 7i) or dual (Figures 7b, 7d, 7f, and 7j) depth-age tie point approach.
Table 3. Root-Mean-Square Errors (RMSEs) in Years For Virtual Core Depth-Age Dating Models, For Data at Sample Resolutions as Specified
 The root-mean-square errors (RMSEs) in dating (Table 3), alongside the correlation results (Figure 8) show that, for dz less than around 50 mm, depth-age models are generally more sensitive to whether there are one or two tie points specified than they are to the sample resolution. For example, Table 3 indicates RMSE values of 0.072 ± 0.010 yr (and 0.078 ± 0.014 yr) for the single tie point approach for dz ≈ 10 mm (and dz ≈ 50 mm), but significantly lower 0.051 ± 0.014 yr (and 0.062 ± 0.012 yr), values for the dual tie point approach for dz ≈ 10 mm (and dz ≈ 50 mm). At lower 100 mm sample resolutions RMSE values are larger, again indicating significant sensitivity of the depth-age model to sample resolutions dz > 70 mm for the single tie point, and about dz > 40–50 mm for the dual tie point approach.
 These dating errors explain the following: (1) The noticeably uniform R values obtained for sample resolution dz less than ≈50–70 mm (Figure 8). In essence, temperature information recovery shows limited sensitivity to dz, provided it is less than 70 mm, for the single tie points dating approach, and less than about 50 mm for the dual tie point approach. (2) Synoptic-scale temperature information cannot be reliably recovered for any dz for single or dual seasonal dating, shown by R values less than 0.3 (Figures 8a and 8c, dashed lines). (3) R values less than 0.5 for the single tie point, and less than 0.6 for the dual tie point, at all dz, indicate a relatively low recovery potential for 60 day low-pass temperature information (Figures 8b and 8d, solid lines). (4) Lower frequency, 180 day low-pass temperature can be moderately well recovered (R > 0.6) for two sites using single tie point dating at dz up to about 10 mm (Figure 8b, dashed lines). Using dual tie point dating, these low-pass temperatures can be recovered quite well (R > 0.6) for most core sites at dz < 50 mm (Figure 8d, dashed lines).
 In summary, dating errors using a photochemical annual-layer counting depth-age model suggest that temperature anomalies with periods longer than 180 days (half a year) can be potentially recovered from most peninsula ice cores, provided that: (1) the cores are sampled at resolutions better than about 50 mm; and (2) dual annual dating tie points are used. Shorter time scale temperature anomalies cannot be reliably recovered using this type of dating approach, with one or two tie points, using even very high resolution ice core sampling.
5.3. Case 3: Using a Cross-Dating Depth-Age Model
 The final depth-age model considered is the cross-dating case. For this case we use the paleoclimate MATCH program, which implements a form of time series pattern (or wiggle) matching [Lisiecki and Lisiecki, 2002]. The automatic program parameter selection used is given by using the graphical user interface supplied with MATCH, and no other predefined depth-age model tie points. This eliminates any user subjectivity. The highest accumulation Beethoven virtual core is used as the cross-dating target, alongside a simple assumption that this one record is perfectly dated. Beethoven is marked as the target record (Figure 9b), and the other virtual cores on their original depth scales (Figure 9a) are then cross-dated to this target core depth scale (Figures 9c and 9d). In each case the signal and target virtual cores are sampled at the same resolution. If the cross-dating procedure were perfect then, through this approach, perfect age models could be recovered.
 Errors in the recovered depth-age models are shown in Figure 10 and Table 3 (“Cross-Matched to Beethoven” entries), i.e., the errors generated by imperfections in the cross-dating. Dating RMSE values are 0.094 ± 0.040 yr for dz = 10 mm; 0.103 ± 0.030 yr for dz = 50 mm; and rising to 0.113 ± 0.048 yr for dz = 100 mm. These errors are larger than those from the photochemical dating case above.
 Unlike the previous case, where the dating assumptions could be easily applied to the time series generated at any dz between 1 and 250 mm, use of the MATCH graphical user interface would make undertaking the cross-dating procedure for every dz between 1 and 250 a time-consuming task (allowing around five minutes per cross-dating procedure, for five virtual cores, implies about one month devoted solely to this task). For the purposes of this exploratory case, we therefore restrict ourselves to undertaking the cross-dating at the illustrative intervals of dz at 10, 50, and 100 mm.
 Correlation results for the 10, 50, and 100 mm cases are presented in Table 4. Potential temperature recovery is significantly affected by sample resolution for all frequencies of temperature anomalies (Table 4, “No Seasonal” and “Low-Pass” entries). For the highest sample resolution (dz = 10 mm) correlation results, R never exceeds 0.52. Virtual cores from the highest accumulation, and closer to Beethoven, Gomez and Dyer sites have the highest R values, and also tend to have the lowest dating errors. In general, as for section 5.2, the lowest depth-age model errors tend to yield the best correlation results. However, since all cross-dated age model errors are relatively large (Table 3), this suggests that cross-dating ice cores, using one well-dated target core without any further age marker constraints, does not tend to enable good ice core temperature information recovery. Consideration of the result in conjunction with sections 5.1 and 5.2 suggests that R is unlikely to exceed about 0.55 for our peninsula virtual cores at any sample resolution.
Table 4. Correlations Between Virtual Core T(t) and Tp (t) Using Cross-Dated Depth-Age Modelsa
 In summary, depth-age model errors from using unconstrained MATCH cross-dating prevents the recovery of subannual temperature anomalies, even at high-resolution ice core sampling. Additional depth-age information, e.g., combining photochemical tie points and cross-dating methods together, could reduce the depth-age model uncertainties. This would make it likely that lower-frequency (e.g., 180 day low-pass filtered) recovery potential could be improved.
6. What Causes These Results?
 The simple virtual core method presented means that the results obtained above are primarily dependent on precipitation amount and intermittency effects. Using R threshold values of 0.5, 0.6, and 0.7 to characterize the integrity of the temperature reconstruction means that 25%, 36%, and 49%, respectively, of the original temperature time series variance is explained by the recovered Tp virtual core values.
Figures 11a, 11b, 12a, and 12b depict the sample resolution lengths at which particular correlation coefficient thresholds are passed using a perfect depth-age model, and Figures 11c, 11d, 12c, and 12d use annually dated photochemical depth-age models. The points shown are for all ERA-40 grid points on or close to the Antarctic Peninsula (as shown in Figure 1). The lines in Figure 11 are not best fits, rather they show the approximate Nyquist sample rates. These approximate rates are also indicated by the shading in Figure 12. The approximate Nyquist rates are calculated assuming that the 180 day low-pass time series have two cycles per year; and the 60 day low-pass time series have six cycles per year.
6.1. The Accumulation Amount Effect
 Consideration of the amount effect is straightforward: the amount and sample resolution are directly proportional (e.g., see Figure 11). Given identical intermittency characteristics and neglecting diffusion, an ice core sampled at a resolution dz = 10 mm, obtained from a site accumulating 1000 mm w.e. yr−1, is directly equivalent to a sample resolution of dz = 1 mm applied to a core accumulating 100 mm w.e. yr−1. Conversion of the sample resolution length to a number of samples per year is useful in enabling effects other than this amount effect to be considered.
6.2. The Precipitation Intermittency Effect
 Accumulation intermittency is less straightforward. Figure 12 shows the dz sample resolution length results, from Figure 11, converted to average numbers of samples per year. The converted resolution length results are presented against a measure of precipitation intermittency; the coefficient of variation (COV), which is the standard deviation divided by the mean.
 In terms of understanding the effect of precipitation intermittency, notwithstanding diffusion, a core constructed under conditions of constant precipitation would potentially allow any daily temperature time series to be perfectly recovered, provided sample resolution size dz is kept small enough to provide a daily Tp(z) value. Constant precipitation would also ensure no errors on a depth-age model, providing it had one or more accurate depth-age tie points. (Note that a constant accumulation site would have a zero precipitation intermittency COV value.) For cores constructed from variable precipitation time series, the temperature signal is recorded only when precipitation falls. Progressively more of the temperature signal is lost as precipitation intermittency increases, and the difficulty associated with dating the core increases. Thus, hiatuses in accumulation both prevent information recording, resulting in unavoidable temperature information loss, and also make accurate dating difficult. This is why there are no sample resolutions which provide jointly high R and COV values in Figures 12b–12d.
Figure 12 quantifies intermittency effects. For the perfectly dated cores, there is a log linear relationship between precipitation intermittency (COV) and the number of samples per year for each of the R correlation thresholds, both for the 180 and 60 day low-pass results (Figures 12a and 12b). Most of the peninsula sites have a COV value of about 1.7 (Table 1). This suggests that getting 180 day signals from perfectly dated cores requires about 6 samples per year for R = 0.5; around 8 samples per year for R = 0.6; and around 11 samples per year for R = 0.7 (Figure 12a). Shorter 60 day signals from perfectly dated cores seem to require about 10 samples per year for R = 0.5; and around 16 sample per year for R = 0.6. While extrapolation implies that around 25 samples per year for R = 0.7 would fit the log linear trend, the precipitation recording at COV = 1.7 is too incomplete to reliably allow temperature information recovery with R greater than 0.6 (Figure 12b).
 The dating errors which affect the annually dated Figures 12c and 12d results produce more scatter in the information recovery versus precipitation intermittency relationship. Nevertheless, the same sort of log linear relationship still occurs. For the low-pass filtered data, the number of samples required per year is higher than the approximate Nyquist rate, as shown by the shaded panels, for these imperfectly dated cores. For ice core sites with stronger precipitation intermittency effects (COV of around 1.7), it is not usually possible to recover temperature information at the R = 0.7 threshold, for 180 day filtered results. For the 60 day results, even the R = 0.5 threshold is generally not exceeded. Thus the maximum percentage of information which can be recovered at annually dated, or layer counted, peninsula ice core sites (COV = 1.7) is about 36% for 180 day low-pass filtered series, and is less than 25% for 60 day low-pass filtered temperature series. For the perfectly dated cores, Figures 12a and 12b indicate that 50% or more of the information can usually be recovered.
7. A Brief Discussion of Other Short Core Considerations
 While the results above are dependent on precipitation amount and intermittency effects, other phenomena may also affect what temperature information can be recovered from δ measured in ice cores.
7.1. Mixing in CFA Instruments
 The recent developments in ice core CFA and melter techniques that have led to higher-resolution data sets being achieved motivate this study [Federer et al., 2008; Kaufmann et al., 2008; McConnell et al., 2002; Röthlisberger et al., 2000; Sigg et al., 1994]. Assuming horizontal homogeneity of the ice, the effective CFA sample resolution tends to be limited by mixing that occurs during the measurement process. Intersample mixing is possible in: the melting unit; during degassing processes; in tubing; in reagent mixing apparatus and in detector cells [Olsen et al., 2006; Rasmussen et al., 2005; Sigg et al., 1994]. The volume (cross-section area) of ice at any particular depth set aside for analysis, the speed of analysis; the isotope ratio mass spectrometer precision, and CFA system mixing characteristics will set the eventual instrument δ sample resolution recovery limit [e.g., Olsen et al., 2006]. Since, in most CFA systems, sample length resolution will be shorter than 10 mm, this implies that CFA system limitations on δ measurements are unlikely to adversely affect temperature recovery for short Antarctic Peninsula cores.
7.2. Impacts of Other Isotopic Effects
 The individual air parcel model, which suggests that Tp and δ are equivalent, is a simplification. More complex changes in the mixing of different air parcels affect δ [e.g., Noone, 2008]. For the Antarctic Peninsula, climate changes such as sea-ice anomalies [Abram et al., 2010], local circulation and precipitation anomalies [Sime et al., 2009a], and large-scale atmospheric circulation changes [Marshall and Connolley, 2006], occur alongside temperature changes and are likely to induce additional complexity to the δ versus temperature relationship [e.g., Noone and Simmonds, 2002; Helsen et al., 2007; Noone, 2008; Sime et al., 2009b]. Although it would be possible to add noise to δ versus temperature relationship, using the virtual core approach, it is not clear how one might accurately specify the noise characteristics. Therefore, the most powerful means to diagnose these type of effects seems likely to be to perform a similar analysis on δ values obtained from a water isotope enabled climate simulation for the Antarctic Peninsula. This also has the potential added benefit of allowing observed regional isotopic signals to be used to deduce atmospheric circulation and sea ice changes.
 Although many previous isotope enabled climate simulations have been run at resolutions which are too coarse to examine local core site variations across the Antarctic Peninsula [Sime et al., 2008, 2009b; Tindall et al., 2009], significant progress in undertaking higher-resolution isotopic simulations [e.g., Sturm et al., 2010; Noone and Sturm, 2010] suggests that regional or high-resolution global modeling of stable water isotopes across the Antarctic Peninsula would be very useful. The virtual core techniques introduced here could be of value in making best use of regional isotopic model output.
8. Summary and Conclusions
 This study presents a new method which is helpful in interpreting stable water isotope signals in ice cores. A set of simple isotopic recording and postdepositional assumptions allow European Center for Medium-Range Weather Forecasts 40 year reanalysis time series of temperature and precipitation to be converted to virtual core depth series. In the context of recent developments in high-resolution ice melter systems and continuous flow analysis techniques, this is useful in helping to understand what temperature information can be gleaned from high-resolution isotopic measurements.
 Three types of depth-age models were analyzed: perfectly dated cores; independent “annual layer counted” photochemically dated cores; and cross-dated records where other cores are matched to one target core. The analysis, correlating virtual core reconstructed temperature records against the ERA-40 original time series of temperature, quantifies how strongly the original temperature time series and the reconstructed virtual core temperature variables are related, given a specified sample resolution and depth-age model.
 The first result we show is that Antarctic Peninsula virtual core correlation results are strongly affected by the mean climatological (i.e., the seasonal) temperature cycle. Since interest in past climate tends to be focused on recovering temperature anomalies, rather than mean climatological signals, the mean seasonal cycle needs to be removed to allow an accurate assessment of ability to recover temperature anomalies.
 For the perfect depth-age model case, and once the seasonal mean cycle has been removed, at some sites 36% of the 60 day low-pass filtered temperature information can be recovered using around 10–15 samples per year (around 40–80 mm resolution). For 180 day low-pass temperature anomalies, about 5 to 9 samples per year (60–120 mm) tends to allow about 36% recovery. It is unlikely that higher-frequency synoptic time-scale temperature anomalies can be recovered from Antarctic Peninsula cores, due to the joint impact of precipitation intermittency and diffusion effects.
 The independent seasonal photochemical marker depth-age model case, has relatively small depth-age model errors. For sample resolution, dz, less than about 50–70 mm, the magnitude of the depth-age model errors are more sensitive to whether there are one or two annual tie points specified than they are to the sample resolution. The dating errors affect the potential for temperature information recovery. The errors ensure that, while 36% of temperature anomaly information with frequencies lower than 180 days (half a year) can be recovered from ice cores, about 7 to 14 samples per year (40–80 mm) are required, i.e., higher sampling resolution than for perfectly dated cores. Moderate, 25%, temperature anomaly information recovery at frequencies lower than 60 days may require up to 30 samples per year (20 mm). However this level of information recovery is possible only from sites with levels of precipitation intermittency which are lower than average Antarctic Peninsula intermittency. For the cross-dating depth-age models explored here, subannual temperature anomalies cannot be reliably recovered. This is due to the magnitude of the depth-age model errors induced by the unconstrained cross-dating approach explored here.
 In general, for most Antarctic Peninsula ice core sites, the maximum percentage of information which can be recovered at annually dated, or layer counted, peninsula ice core sites is about 36% for 180 day low-pass filtered series, and usually less than 25% for 60 day low-pass filtered temperature series.
 There are limits to the virtual core method as presented here. If the method were to be applied to deeper ice cores, ice deformation and/or thinning would also have to be considered. A further useful development would also be to implement the full diffusion model of Johnsen et al. . The use of an individual air parcel model, which assumes that site temperature during precipitation and δ in ice core are equivalent is also a simplification. Changes in the mixing of different air parcels and potential vapor source changes caused by: sea-ice anomalies; local circulation and precipitation anomalies, and large-scale atmospheric circulation changes, are all likely to occur alongside temperature anomalies, and may also have independent affects on δ and temperature at peninsula ice core sites [e.g., Noone and Simmonds, 2002; Noone, 2008; Sime et al., 2009a]. High-resolution isotopic modeling, perhaps together with the virtual core techniques introduced here, would provide a means to allow significant future progress in this area of research.
 In conclusion, the new virtual core method presented is useful in gaining insight into the recovery of temperature information and in investigating dating methods for ice core records. The method, and these study results, can aid decision making about ice core dating and at what resolution to sample δ. The analysis indicates that δ samples of lengths around 40–50 mm are likely to provide valuable information about subannual peninsula past site temperature anomalies.
 Thanks to Kevin Oliver for advice on methods, to Peter Fretwell for producing the Figure 1 map, to all members of the British Antarctic Survey Ice Coring group for interesting discussions on past British Antarctic Survey ice coring and sampling strategy, and to Hans Christian Steen-Larsen and an anonymous reviewer for insightful reviews. This study is part of the British Antarctic Survey Polar Science for Planet Earth Programme. All authors gratefully acknowledge funding from United Kingdom Natural Environment Research Council.