Sarah De Ceustera,
Dimitra Machairab and
Patrick Degryse*ab
aKU Leuven, Earth and Environmental Sciences, Celestijnenlaan 200E, 3001 Leuven, Belgium. E-mail: patrick.degryse@kuleuven.be
bLeiden University, Faculty of Archaeology, Einsteinweg 2, 2333 CC Leiden, The Netherlands
First published on 28th June 2023
Lead isotope analysis has been used to determine the provenance of metals such as lead, silver and bronze for many decades. Nevertheless, different approaches to interpret lead isotopic ratios have been proposed. In this study, three methods to couple the lead isotopic signature of archaeological artefacts to their possible mineral resources will be compared: the conventional assessment of biplots, a clustering method combined with calculating model ages (as applied by F. Albarède et al., J. Archaeol. Sci., 2020, 121, 105194), and relative probability calculations using kernel density estimates (as proposed by De Ceuster and Degryse, Archaeometry, 2020, 62(1), 107–116). The three different approaches will be applied to a dataset of lead isotopic analyses of 99 Roman Republican silver coins previously analyzed, pointing to a primary origin of the silver in the mining regions of Spain, NW-Europe and the Aegean, but showing signs of mixing and/or recycling. The interpretations made through the different approaches are compared, indicating the strengths and weaknesses for each one. This study argues that, although the conventional biplot method gives valid visual information, it is no longer feasible due to ever growing datasets. Calculating the relative probabilities via kernel density estimation provides a more transparent and statistically correct approach that generates an overview of plausible provenance candidates per artefact. The geological perspective introduced in the cluster and model age method by F. Albarède et al., J. Archaeol. Sci., 2020, 121, 105194 broadens the analytical spectrum with geologically informed parameters and improved visualization. However, the results when applying their method as a stand-alone approach are of low resolution and may lose archaeological relevance. Their approach regarding clustering should be revised.
After the introduction of lead isotope analysis (LIA) to provenancing in archaeological science (e.g. Brill and Wampler4), the method gained increasing prominence when Gale and Stos-Gale5 studied the provenance of Late Bronze Age (LBA) copper alloys. Traditionally, bivariate plots are used to assess the data, visually matching lead isotopic compositions of artefacts and ore deposits. Much debate has always been associated with the interpretation of LIA (e.g. Pollard6). It has been pointed out that most geographically and/or geologically distinct regions show overlap in their lead isotope ratios and that the lead isotope signature in a single region can be very broad. Furthermore, lead isotope ratios of an ore body or mining region are rarely normally distributed. Baxter et al.7,8 argued that assuming normality of lead isotope data leads to false graphical presentations, though the usual biplot approach was defended by Scaife et al.9 for provenancing purposes, again the cause of much debate.
Efforts to deal with the limitations of LIA are frequently made. The use of kernel density estimates (KDEs) was proposed to address the non-parametric distribution. Choosing the optimal parameters would be challenging, though, and the need for larger sample sizes for the ore sources was conveyed.7,8 De Ceuster & Degryse2 demonstrated kernel density estimation (KDE) as a method for provenancing purposes. The parameters chosen for calculating the KDEs were provisionally defined, but need further evaluation. Clearly, the main advantages of this method are that it tackles the assumption of normality and that it provides a mathematical, statistical basis for further calculations.
Another criticism of the use of LIA in archaeology has been the tendency to use isotope ratios as simple numbers that characterize an ore source, while these ratios evolve according to laws of radioactive decay and geochemistry.10 Albarède et al.1 use hierarchical cluster analysis to group samples, then examine the geochemical parameters per group, to assign each group to a Pb model age, thus trying to identify a potential ore source. They tested their method on a data set of 99 Roman Republican silver coins.11–13
This paper makes an in-depth comparison of three common methods used in assessing lead isotope data for archaeological provenancing purposes: the conventional bivariate plots, the cluster and model age method advocated by Albarède et al.1 and the KDE method proposed by De Ceuster and Degryse.2 Each method will be applied on the data set of 99 silver coins as used in Albarède et al.11 The paper also aims to define for each approach the advantages and disadvantages and to determine to what extent these methods contrast or complement one another.
As conventional biplots the 207Pb/204Pb vs. 206Pb/204Pb and the 207Pb/204Pb vs. 208Pb/204Pb ratios were displayed. These were then visually assessed per coin and per mining region. Often, in literature, confidence ellipses are added to these scatterplots as “ore fields”, assuming normality instead of the non-parametric distribution that is given,6–8 to then appraise which artefacts fall inside the ellipses on both plots. This assessment was performed here by checking if the measured values would lie within each confidence interval.
For the cluster and model age method, the same process was followed as proposed by Albarède et al.1 Before clustering the artefacts, their isotopic ratio data was whitened, i.e. decorrelated and standardized. In this paper, the ZCA transformation matrix was chosen in order to keep the whitened data as close as possible to the original measurements. For the hierarchical clustering Ward's linkage algorithm was applied with Euclidean distances. The resulting groups were then displayed in scatterplots, thus representing the relations between the geological parameters model age (Tm), κ (Th/U) and μ (U/Pb), and between different lead isotope ratios. These plots are then visually assessed.
For the KDE approach, the method advocated by De Ceuster and Degryse2 (Appendix 1†) was further developed.
The kernel density is estimated per region and per ratio in the following way. The estimator for a probability density function f(x) can be motivated as follows:
For every coin in the assemblage (cf. Albarède et al.11) the measured lead isotope ratios are used to calculate the relative probability that this coin originates from each mining region. These results are then illustrated with a bar chart per region and per ratio for each coin, and assessed visually. Although this is a multidimensional setting, the choice was made to visually assess the calculated results per ratio, because adding dimensions enlarges the error in the estimated density function, particularly in this case of small sample sizes. This phenomenon is referred to in the literature as the curse of dimensionality.15
All calculations and plots are generated using RStudio, a free and open-source integrated development environment for R, a programming language for statistical computing and graphics. The R code used in this paper can be consulted in the Appendix 2.†
Fig. 2 The confidence 95 ellipses of the mining regions that give a ‘match’ and the scatterplots of the coins (grey dots). The coins with a match are highlighted. |
Albarède et al.1 state that the practice of employing LIA to link every metal artefact separately with their ore source does not exploit all possibilities of the LIA dataset of the artefacts. In their paper they apply their method twice, once with 12 groups for which they argue it is too many and once with four groups for which they argue it is not enough. The groups are created for description purposes. The hierarchical cluster analysis, using Euclidean distances in the Ward linkage algorithm, of the coins dataset after whitening with the ZCA method results in six significant clusters. In a next step better ways to visualize these clusters are proposed. In the top left scatterplot of Fig. 3 a ‘mixing triangle’ can be identified where the vertices – clusters 2, 4, and 1 and 5 – represent the unmixed groups and the other clusters are probably the result of mixing and/or recycling. Lead isotope fractionation can be disregarded in ancient metallurgical processes,18 meaning the isotopic signature of the ore used is passed unchanged to metal coins. Care has to be taken, however, when ores are mixed (e.g. the separate addition of another metal such as lead), giving a mixed isotopic signature, or when different batches of silver are recycled together. Model ages, and μ and κ are added as valuable extra dimensions to the dataset (Fig. 3). Group 2 corresponds to young model ages associated with the Baetic orogeny. The vertex represented by group 4, with Mesozoic ages, probably relates to the hydrothermal activity that led to Pb–Zn deposits throughout Western Europe (e.g. Muchez et al.19 and Burisch et al.20). Groups 1 and 5 have Paleozoic model ages pointing to the Hercynian orogeny. Groups 3 and 6 plot within the ‘mixing triangle’ and might therefor represent mixed and/or recycled raw materials, which can be the result of a geological or an anthropological process. These groups and allocations can be consulted in the table in Appendix 5.† Albarède et al.1 conclude their case study by confirming the attained groups with histograms of the minting ages per group and suggesting that the Cenozoic source of silver mined throughout the Second Punic War was replaced by Mesozoic sources. Although this can be useful complementary information, this does not answer the provenance questions asked nor does it meet the desired archaeologically relevant resolution. A map of Pb model ages in Europe and the Mediterranean region compiled by Blichert-Toft et al.,21 using mainly lead ores, demonstrates that all model ages appearing in this dataset of coins occur on several locations throughout the area that was once the Roman Empire and beyond, though their abundance varies regionally. The conclusion thereby does not allocate possible ore sources, nor does it shed a light on human behavior, or the processes of production, consumption and possible exchange and/or trade.
Fig. 3 Lead isotope ratios and parameters Tm, μ and κ cfr. Albarède et al. 2020.1 |
KDEs are a valid alternative to tackle the problem of non-parametric datasets. Per coin a bar chart is generated that shows the relative probability that the raw materials for that coin originate from a specific region per isotopic ratio. The regions where all three ratios are highest are potential source regions for the ores. These charts were visually divided in groups with a similar signature and further confirmed through inspection of the calculated probability tables (Appendix 6†). The regions for which the minimum probability is 5% for each ratio were highlighted together with the two highest minima (example Table 1). The names or descriptions of the groups are the regions with the highest relative probabilities, they are probable ore sources, not immediate answers. The first group of coins is consistent with a Spanish origin. One coin, sample R118, scores high on several Spanish regions in the southern half of Spain. Furthermore, there are two subgroups: one that points at an origin on the Balearic Islands for 10 samples, and a second one in Gipuzkoa in northern Spain for 9 samples with relatively high probabilities and another 7 samples with lower probabilities. The second group of 14 coins reveals a Northwestern European origin, though for most samples it is difficult to distinguish between the different regions in France, Britain and Germany. Three coins match with several of these Northwestern European regions. One coin, sample MS056, seems to find its origin in France: most likely in Armorica though the Pyrenees are also a likely candidate. The 10 other coins in this group show a high relative probability to have a Portuguese origin besides their high Northwestern European values. The third group of 19 coins indicates a Greek source. Seven coins match with multiple Greek regions: Chalkidiki, Laurion and Thasos. One coin, 22-Denarius-44-5, points to an origin in Chalkidiki, a second subgroup contains 6 coins that have a high relative probability for Chalkidiki and Thasos. The third subgroup for Greece includes 5 coins that seem to originate from the mines of Laurion. The fourth group consists of 4 Bulgarian subgroups: 3 coins imply a provenance in the Burgas region, 2 in the Central Rhodope region, one in the East Rhodope region and 2 in the West Rhodope region. A fifth group holds 13 coins that have a very similar signature and thus, very likely, have the same provenance. Their bar charts show higher relative probabilities for Bulgaria, Siphnos, Almeria in Spain and Türkiye. The sixth group contains 18 coins that have no clear signature. Two coins with very comparable signatures appear to originate in Iran, though they also have higher relative probabilities for France. These may be the result of the existence of the Silk Road, trading with the Parthian Empire. Five coins show quite low relative probabilities for Bulgaria and France, and 6 coins show equally low relative probabilities for Bulgaria and Türkiye. The last subgroup holds 5 coins, again with very similar bar charts that point to no specific region, they have a very low match almost everywhere. The reason these low probabilities occur might be the absence of the actual mining region in the database, or mixing and/or recycling.
Theoretically, every mining region with even the smallest relative probabilities for each ratio is a possible ore source, but a minimum probability of 5% was chosen for significance. For each coin the regions for which the minimum probability is 5% for each ratio were listed in the table in Appendix 5† as “matches”. Notice that this is often a subset of the visual biplot results. The reason why there are several matching regions is of course the overlap between them. The region with the highest probability for the 3 ratios was listed in the next column, though this is not necessarily the ore source. In the Greek group (3), for example, each coin – except for 22-Denarius-44-5 – matches with a combination of Chalkidiki, Thasos, Laurion and western Türkiye. The highest probabilities suggest a Greek provenance. The overlap plots for Greece (Fig. 4) show that it is impossible to differentiate between Chalkidiki and Thasos, which results in group 3b. Both ore sources have high probabilities for these coins and it is impossible to differentiate between them. However, to distinguish Laurion from them in group 3, it is interesting to inspect the orange 206Pb/204Pb probability bars on the bar charts, because that density plots next to the ones for Chalkidiki and Thasos (Fig. 4). This means that samples MS002 and MS006 can be added to group 3c, for which the ores were probably mined in Laurion. The utility of these overlap plots is further demonstrated for the Bulgarian group 4. As can be deduced from the overlap plot for the Bulgarian group (Fig. 5) the difference between Burgas and the others is easier made than between the eastern and central Rhodope region. One should also stay aware that any of these Bulgarian ores has a small chance to originate in the western Rhodope region, because of the broad variance of the ratios for this region. The samples in group 5 may not be easily allocated to a specific region, due to the overlap between the KDEs of the Bulgarian regions, the Turkish regions, Siphnos, and Almeria in Spain (Fig. 6). However, the 3-Victoriatus-44-1 and the R82 coins have a high probability to have their source in Siphnos, because of the high relative probability for 207Pb/204Pb.
Fig. 4 Visualization of the overlap between the lead isotopic KDEs for the three isotopic ratios for Chalkidiki, Thasos and Laurion in Greece. |
Fig. 5 Visualization of the overlap between the lead isotopic KDEs for the three isotopic ratios for the Bulgarian mining regions. |
Fig. 6 Visualization of the overlap between the lead isotopic KDEs for the three isotopic ratios for KDE group 5. |
Albarède et al.1 advocate the use of all geological information contained in the samples analyzed and propose better visualization methods. Their results, however, have low resolution, linking the coins to an orogenic event, which makes their stand-alone usefulness questionable. The KDE-approach offers a higher-resolution image of the dataset, rendering ‘ore fields’ that match the isotopic ratios. The KDE overlap plots greatly contribute to the interpretation, though one should be mindful that low probability still is probability. However, since datasets grow it might not stay effective to assess every sample separately and to group them visually. A clustering method as suggested by Albarède et al.1 might be preferable to describe a dataset more efficiently, though Ward's linkage algorithm with Euclidean distances might not be the most suitable, even after whitening. To compare both methods further, the groups formed after visual inspection of the KDEs were represented as suggested by Albarède et al.,1 displaying a ‘mixing triangle’ and geological parameters Tm, μ and κ (Fig. 7). When compared to the same plots with the clusters formed cfr. Albarède et al.1 (Fig. 3), it is clear the groups formed by the cluster analysis of Albarède et al.1 are different from the ones composed through visually examining the KDE bar charts. Group Albarède-1 largely coincides with the Northwestern European KDE-group and group Albarède-2 partly corresponds with the Greek KDE-group. The other groups are dissimilar. Although this dataset contains 99 coins, which is a large dataset to assess manually and visually, it only provides a very small sample when taking the represented time span and geographic area into account. It is not unreasonable to assume that several of these coins do not belong to a cluster, but this is not supported by the applied clustering algorithm. A density based clustering method might then give a better outcome.
Fig. 7 Lead isotope ratios and parameters Tm, μ and κ, visualized cfr. Albarède et al. (2020)1 with groups cfr. KDE method. |
Besides the comparison it might be interesting to explore whether the geological information promoted by Albarède et al.1 can be used to further discriminate between the results of the KDE approach or to confirm them. The ‘mixing triangle’ (Fig. 7a) implies that groups KDE-6b, 6c and 6d and the subset of group KDE-1b that plots together with group KDE-6, might be the result of mixing and/or recycling. This conclusion is supported by the KDE giving no matches for most of these coins. The other subset of KDE-1b has lower μ and κ values (μκ ≈ 38) and has younger model ages (Tm ≈ 100 Ma), so, combining the KDE bar charts, the ‘mixing triangle’, and maps of Pb model ages, μ and μκ in Europe and the Mediterranean region,20 Gipuzkoa region is a valuable candidate for their provenance. Group KDE-1a splits into 2 groups, a larger one with higher κ values and slightly older Tm, and 4 samples with lower κ values and younger Tm (Fig. 7c). Both groups might not originate on the Balearic Islands, because μκ ≥ 38. The larger group plots together with group KDE-4b-4c-4d and their KDE bar charts also show elevated relative probabilities for the Bulgarian Rhodope regions. Looking at the overlap between their lead isotopic KDEs (Fig. 8) it might be impossible to differentiate between both the Balearic Islands and the Rhodope regions, though the Tm and μκ support a Bulgarian origin. Group KDE-2b could be subdivided in a group with lower μκ and Tm and one with slightly higher μκ and Tm. However, both might originate in Portugal, France, Britain or Germany. In addition, this group isn't necessarily homogenous. In group KDE-2, 10-Denarius-53-2 and MS042 show low μκ and Mesozoic model ages, confirming the results of the KDEs. The third sample, 19-Denarius-112-2a, shows Paleozoic Tm and elevated μκ, and it has a high relative probability to find its origin in Germany (Siegerland) or in Britain (Northern Pennines or Alderley Edge). Looking at a map showing the geographic distribution of μκ,20 a British origin is more probable. For groups KDE-3-3b-3c the young model ages and 9.7 < μ < 9.8 are consistent with the Greek origins in Chalkidiki, Thasos and Laurion for those samples. In group KDE-4, subgroup KDE-4a has the lowest μκ values and young model ages, consistent with the Burgas region, the other subgroups have relatively high μκ values and Mesozoic Tm, supporting their origin in the Rhodope Mountains. The coins in group KDE-5 (Tm ≈ 100 Ma, μ > 9.77) seem to fall apart in several subgroups: one that forms the vertex in the ‘mixing triangle’ and has κ > 4, one with 3 samples measuring 3.97 < κ < 4, and 4 samples with κ ≈ 3.95, of which 2 samples with slightly older model ages and higher μ values. The high μ values and young model ages for most of the group KDE-5 coins seem to be present in the Central Taurus region in Türkiye, in western Türkiye, in southern Greece and in southern Spain.20 So, for this group it is not possible to narrow down the options obtained by the KDE method.
Fig. 8 Visualization of the overlap between the lead isotopic KDEs for the three isotopic ratios for the Bulgarian Rhodope regions and the Balearic Islands. |
By combining the KDE approach, inspection of the overlap plots and the extra geological parameters advocated by Albarède et al.,1 extra information concerning the provenance can be derived from this dataset of coins. These results were also added to the table in Appendix 5.†
Footnote |
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d3ra02763e |
This journal is © The Royal Society of Chemistry 2023 |