A Lake Record of Geomagnetic Secular Variations for the Last 23 ka From Lake Chala: Toward a Composite Directional Lake Record of the Earth's Magnetic Field for Equatorial East Africa

The documentation and understanding of variations in the Earth's magnetic field through time is fundamental for several disciplines, but current geomagnetic models rely on datasets heavily biased toward the mid‐ and high northern latitudes. The African continent and surrounding islands and oceans are particularly underrepresented. Here, we present a new record of paleo‐secular variation (PSV) of the inclinations over the last 23 ka from Lake Chala, situated at 3°S near Mt Kilimanjaro in eastern equatorial Africa. This groundwater‐fed crater lake is characterized by a high sedimentation rate (ca. 1 cm/10 years) and a particularly well‐constrained age model based on 210Pb and 14C dating. The magnetic mineralogy of the sediments is tested with rock magnetic analyses. The Lake Chala inclination record shows four highs and lows over 20 ka and compares well with that of Lake Malawi (10°S) between 20 and 16.2 ka, and from 9.8 to 2.6 ka. This record is linked to PSV records at Lakes Victoria and Malawi using a sequence slotting technique to generate a composite PSV model for east Africa. Analyzed at best‐possible resolutions up to 200 years, the Lake Chala PSV record not only represents an important contribution to improve our understanding of local and global features of the Earth's magnetic field. It also expands the utility of paleomagnetism as a key tool for dating and correlation both for archeological sites throughout East Africa and the many volcanoes, active or dormant, of the East African Rift System.


Introduction
Direct measurements of the Earth's magnetic field from satellites are limited to the last 50 years and indirect observations to the last 400 years (from navigation records, Jackson et al., 2000).To understand the characteristics and behavior of the geomagnetic field in the past, we rely on paleomagnetic (direction and intensity) data recorded in volcanic rocks, sediments and heated archeological artifacts.
Global magnetic datasets (e.g., MagIC, www.earthref.org;GEOMAGIA, Brown et al., 2015) store published data and provide an invaluable resource for building geomagnetic models to understand the characteristics, features and evolution of the Earth's magnetic field.The global and regional geomagnetic models describe secular to millennial changes and their spatial and temporal resolution, the reliability of which has improved with the increasing data availability in the last few decades.However, the current dataset is still limited, leaving large areas across the Southern Hemisphere and the African continent poorly represented.Indeed, of the entire GEOMAGIA dataset (v.3;Brown et al., 2015), only 6% of data derived from sediment cores come from the Southern Hemisphere (Korte et al., 2019).In terms of paleomagnetic directions, only the 3% are from the African continent (Di Chiara, 2019).
Gathering new data from under-represented areas is important for several research disciplines.A robust dataset would refine the temporal and spatial resolution of the global geomagnetic dataset and expand the use of paleomagnetism as a dating and correlating tool.For instance, in volcanology and stratigraphy, paleomagnetism is largely employed to date and correlate volcanic or sedimentary sequences (e.g., Di Chiara et al., 2014;Korte et al., 2011Korte et al., , 2019)).In archeology and ancient and modern history, paleomagnetism helps to constrain the ages of archeological sites and paintings (Chiari & Lanza, 1997) and in paleoclimatology and paleoecology, magnetic and paleomagnetic methods can provide key paleoclimate proxies and date climatic events (e.g., Evans & Heller, 2003;Maher & Thompson, 1999;Sandgren & Snowball, 2001).
On late-Quaternary timescales, paleomagnetic dating is based on the statistical comparison of a paleo-secular variation (PSV) reference curve and the paleomagnetic directions recorded by undated targets (lava or sedimentary sequence/archeological object, etc.).Despite the Earth's magnetic field being predominantly dipolar, the non-dipolar components impart a significant local to regional variability to the geomagnetic field, such that PSV curves do not have global validity (Merrill et al., 1996) but local records can be re-located within regions using the virtual geomagnetic pole (VGP) conversion method (Noel & Batt, 1990).The relocation can induce an error of up to 7°in declination or inclination if sites are 1,700 km distant (Casas & Incoronato, 2007).Global models should overcome the pole relocation error, but their behavior typically smooths datasets (tending to average out the secular variations, e.g., Snowball et al., 2007), limiting the dating resolution that can be achieved.
In this study, we focus on eastern equatorial Africa, whose importance extends beyond paleomagnetism.It is a "hotspot" of hominid evolution and trans-migration (e.g., Kaboth-Bahr et al., 2021), hosting multiple major archeological sites-constraining the age of which can be challenging.Moreover, eastern equatorial Africa is part of the East African Rift System, a 3,000-km long rift system from the Afar depression in the north to Lake Malawi in the south (Chorowicz, 2005) with an abundance of active, dormant, or extinct volcanoes (Fontijn et al., 2012;Macdonald, 2002;Sengör & Burke, 1978), some posing significant potential hazard and risk.The East African Rift System also hosts about 180 lakes (Tiercelin & Lezzar, 2002).However, the scarcity of reliable archaeomagnetic or volcanic paleomagnetic data limits the application of paleomagnetism as a dating tool in this important region.Indeed, for the entire African region, only 50 studies are available (Di Chiara & Pavón-Carrasco, 2022), the majority of which are discussed in a recent review by Di Chiara (2019), with five were published more recently (Kapper et al., 2020;Lund et al., 2021Lund et al., , 2022;;Madingou et al., 2020;Nami et al., 2020).Overall, of late Quaternary/Holocene data from Africa, 52% (26 publications) are obtained from archaeomagnetic artefacts, 22% from volcanic rocks (9 publications from the Canary Islands, 1 from Kenya, Skinner et al., 1975;and 1 in Mt. Cameroon, Herrero-Bervera et al., 2004) and 26% from lacustrine and marine sediments (13 studies).52% of the data are from northern Africa (from 35°N to 12°N), 32% from central Africa (12°N to 10°S) and 8% from southern Africa (10°S to 34°S).Finally, only six studies from lacustrine sediments (Barton & Torgersen, 1988;Lund et al., 2016Lund et al., , 2021Lund et al., , 2022;;Mothersill et al., 1996;Williamson et al., 1991) have been available until now for eastern Equatorial Africa (blue circles in Figure 1).
Here, we present new and reliable paleomagnetic data from the high-resolution sedimentary record from Lake Chala (in central-eastern subequatorial Africa, on the Kenya-Tanzania border; red star in Figure 1).The new data are combined with other suitable lake PSV records to generate a robust composite PSV record for east Africa for the last 23 ka.The new composite record will enhance the applicability of paleomagnetism as an independent dating tool for other lakes, archeological sites and volcanic deposits in this key region.

Comparable PSV Records for Other East African Lakes
The paleomagnetic record from Lake Tanganyika (6.30°S, 29.50°E, ∼1,000 km WSW of Lake Chala; Williamson et al., 1991; Figure 1) comes from two piston cores.Its age model is based on magnetostratigraphic correlation with Lake Barombi Mbo, in SW Cameroon (Thouveny & Williamson, 1988), ∼3,200 km to the ENE of Lake Tanganyika, and six 14 C dates (for MPU13 core) from 25 to 12 ka and two (for core MPU3) from 38 to 12 kyrs.Because of the absence of age control for the Holocene part of the record, any correlation with other regional PSV records can only be tentative.Moreover, the declination record from Lake Tanganyika displays extraordinary directional swings (between 50 and 60°), which are neither expected at such low latitudes nor replicated by other records elsewhere.
The PSV record from Lake Turkana (3.35°N, 36.7°E),∼800 km to the NNW of Lake Chala (Barton & Torgersen, 1988; Figure 1) is based on two cores (TA and TD).Unfortunately, the age model is poorly constrained, based on low-resolution 210 Pb dating, and the sedimentation rate cannot be used as it has been modeled by aligning its paleomagnetic data to Lake Barombi Mbo, more than 5,000 km to the west (Maley et al., 1990;Thouveny & Williamson, 1988).Recently, a new record by Lund et al. (2022) was published from Lake Turkana, with three piston cores, covering the last 4 ka (with a sedimentation rate of ∼200 cm/kyrs).The chronology of the cores was determined from four radiocarbon dates on one core (core 4P), together with the comparison with Lake Malawi (Lund et al., 2016) and Victoria (Lund et al., 2021) paleomagnetic data.The comparison suggests that the Turkana radiocarbon dates are ∼500 years older than the paleomagnetic dates and a correction for that offset restored the consistency of three of the four radiocarbon ages with the PSV age estimates.
Finally, the PSV records from Lake Malawi (10.0°S, 34.3°E; Figure 1), ∼1,000 km to the SSW of Lake Chala, are based on two sediment cores (M98-6P and M98 3-P, 100 m apart) covering the last 25 ka (Lund et al., 2016).The age models are obtained from 22 14 C dates from these and two other cores correlated with the cores using magnetic susceptibility and rock magnetic data.The old carbon correction was 450 14 C yrs.Hence, until now the studies of Lund et al. (2016) from Lake Malawi and Lund et al. (2021) from Lake Victoria are the only reliable datasets due to the reasonable inter-core reproducibility, their better age control compared to the other published lake records, and the comparable time interval covered.These records can therefore be compared with our Lake Chala record, characterized by a high sedimentation rate and a particularly wellconstrained age model.

The Lake Chala Record
Lake Chala (3.317°S, 37.699°E) is a 4.2-km 2 , 94-m-deep crater lake situated at 840 m altitude on the Kenya-Tanzania border in the southeastern foothills of Mt.Kilimanjaro (Figure 1).The lake is currently the object of extensive multidisciplinary studies for reconstructing the climate and environmental history of East Africa (e.g., Barker et al., 2011;Baxter et al., 2023;Maitituerdi, 2022;Meyer et al., 2020;Verschuren et al., 2009;Wolff et al., 2011) and to reconstruct the volcanic history of surrounding scoriae cone fields (Martin-Jones et al., 2020).In 2016, in the framework of the ICDP DeepCHALLA Project, a 214.5 m long record of lacustrine sediments was sampled.The sediments are finely laminated siliceous muds (Baxter et al., 2023;Maitituerdi, 2022) with a sedimentation rate of ∼100 cm/kyrs, comprising ∼80% water, a dry mass dominantly comprising (∼60%) freshwater diatom frustules, and a clastic mineral fraction that is almost entirely limited to the input of aeolian dust (representing 20%-30% of dry mass).Sediment compaction is inhibited due to the high silica content, thereby minimizing any potential inclination shallowing bias (Tauxe, 2005).The age model is based on the transfer of an extremely detailed 210 Pb/ 14 C-based age model (i.e., 168 14 C dates for the 22 m-long core) from the adjacent CHALLACEA core site (Blaauw et al., 2011) using cross-correlation of the fine laminations and varves shared by both sequences (Baxter et al., 2023;Maitituerdi, 2022;Wolff et al., 2011).The composite Chala depth record was created by visual cross-correlation of fine sediment lamination at millimeter scale in overlapping core sections drilled in two adjacent holes A and B (a few meters apart, Baxter et al., 2023).In this study, we consider the uppermost 21.74 m of the composite DeepCHALLA sediment core, in which the lithology varies from mm-scale laminated to cm-scale banded intervals (Baxter et al., 2023).All composite sediment depths were adjusted by removing all tephras and turbidites ≥0.5 cm from the stratigraphy to produce event-free depths (efd) (Baxter et al., 2023).The record studied here thus reaches a depth of 20.66 mefd.

Methods and Sampling
The studied section was sampled along the coring-axis at a total of 113 depth intervals spaced at 2.5, 5 or 25 cm on a preliminary event-free depth scale with standard paleomagnetic Japanese pots (non-magnetic polycarbonate boxes with volume of 7 cm 3 ).
The natural remanent magnetization (NRM) was measured for all specimens.NRM was demagnetized using 10-15 incremental steps of up to 100 mT alternating-field (AF) peak amplitude (2.5, 5, 7.5, 10, 15, 20, 25, 30, 40, 50, 60, 70, 80, 90, 100 mT) using the static AF demagnetization system of the RAPID 2G cryogenic superconducting quantum interference device (SQUID) magnetometer.For inter-comparison, 12 specimens were demagnetized using a reversing-tumbling demagnetizer and the NRM was measured using GM400 SQUID and AGICO-JR6 magnetometers.The magnetic components were calculated using a DosBox version of LINEFIND (Kent et al., 1983; software details in Hounslow ( 2023)).LINEFIND is more sophisticated than conventional principal component analysis in that it uses the measurement variance in a model of the directional uncertainty and determines α 95 for line-fits rather than maximum angular deviation values (Heslop & Roberts, 2016, 2020).
The rock magnetic properties of 66 specimens were analyzed to identify the magnetic carriers, by first measuring the bulk low frequency susceptibility (for the whole 7 cm 3 pot, in units of 10 9 m 3 ), then inducing an anhysteretic remanent magnetization (ARM as a proxy of nanoscale SP/SD-superparamagnetic/single-domain magnetite ;Maher, 1988) in two or three DC field steps (60, 80, 100 μT) and a peak AC field of 80 mT.The ARM was subsequently static AF demagnetized (at steps of 5, 10, 15, 20, 25, 40 mT).Isothermal remanent magnetization (IRM) was acquired in 6 or 7 pulsed DC field steps (10, 20, 50, 100, 300 mT) and 500 and 1,000 mT (saturation, SIRM) non-pulsed DC field.The SIRM was then static AF demagnetized (5, 10, 15, 20 mT), and finally, one step of pulsed back-field IRM acquisition at 100 mT was measured.Specimen wet mass was measured for each specimen and used to normalize the ARM and IRM.
The median destructive field, MDF ARM , defined as the field required to decrease the ARM by 50%, is indicative of the ferrimagnetic grain size.In the Maher plot (Maher, 1988), MDF ARM values from a compilation of synthetic magnetite samples from the literature are plotted against the ratio of χ ARM and the SIRM (saturation IRM at 1 T) to indicate magnetite grain sizes.In this study (Figure 2a), we modified the Maher plot and plot MDF ARM from Lake Chala with the ratio of the χ ARM /IRM measured at 100 mT, to exclude any potential IRM contribution from higher-coercivity minerals and compare only the magnetite content with the reference datasets.The presence of other magnetic minerals (e.g., titanomagnetite, hematite, pyrrhotite, and greigite) was assessed via a range of diagnostic magnetic measurement ratios (Peters & Thompson, 1998; Figure 2b).

Carriers of the Magnetization and Paleomagnetic Behavior
The Maher plot (Figure 2a) displays data from the Chala sediments (black circles) compared with grain-size specific synthetic and ground magnetite powders (Dankers, 1981;Maher, 1988;Özdemir & Banerjee, 1982).The magnetite present in the Chala sediments displays variable grain sizes, including sub-micron particles, but are predominantly <10 μm.(Maher, 1988), which provides a proxy for magnetite grain size, with black symbols indicating the data from Chala and the numbers refer to the sample numbers (see the SI table, https://doi.org/10.6084/m9.figshare.24175503).Gray symbols show the published data, the numbers indicating particle sizes of the magnetite (both synthetic and natural/ground) in micrometers (μm symbol omitted in the plot).(b) Peters and Thompson (1998) plot showing the backfield IRM at 100 mT normalized to the SIRM versus the ARM at 80 mT demagnetized at 40 mT field normalized by the ARM at 100 mT (static fields).Hardness refers to the coercivity increasing toward higher values of the ratios.
The Peters and Thompson plot (Figure 2b) indicates that the magnetic carriers in the Chala sediments are magnetite and/or titanomagnetite, and importantly that greigite is absent (while these data do not exclude the presence of pyrrhotite, there is no evidence of sulphide diagenesis in these sub-oxic sediments).Some data points fall outside the envelope around the reference samples (probably due to other effects, such as interactions, non-stoichiometry etc.).From the average HIRM (calculated as the IRM 300mT -IRM 1000mT ) values of 2.4 *10 6 A.m. 2 /kg, we infer the presence of a significant amount of hematite more common above 8 m (Figure 3h).
The magnetic susceptibility (Figure 3c) displays a decrease from 21 to 12 m, higher values between 9.4 and 8 m and constant low values until a marked increase from 2 to 0 m (Figure 3b).The NRM and ARM show constant values from 21 to 14 m, a decreasing trend from 14 to 11 m and from 9 to 3.5 m (Figure 3), interrupted by higher values from around 9 m and a rising trend from 2.5 m to the core top.The NRM/SIRM ratio (Figure 3f) displays similar trends as ARM.The IRM 10 /SIRM (Figure 3g) ratios show noticeably less variability, while the %HIRM (calculated as 100-the percentage of IRM 300mT /IRM 1000mT ; Figure 3h) shows opposite behavior, with higher values from about 8 to 3 m.Interestingly, the thicknesses of the lamina (Figure 3a) from mm-scale laminations to cm-cm scale sedimentary bands do not consistently relate to any of the rock magnetic data.
Notwithstanding the high water and silica content of the sediment, the NRM intensity is readily measurable, averaging 1.019 mA/m (minimum of 0.049 and maximum of 10.63 mA/m; Figure 3d).Upon AF demagnetization of the NRM, an inferred viscous remanent magnetization (VRM) was usually isolated between 2 and 6 mT, and the characteristic remanent magnetization (ChRM) between 10 and 30 to 40 mT (Figure S1 in Supporting Information S1).ChRM was isolated in 90 of 113 samples; in the other specimens, either no ChRM was isolated or the α 95 was too large.The average α 95 calculated from LINEFIND (Kent et al., 1983;Figure S2 in Supporting Information S1) for the 91 successful samples was 3.4°.
Overall, four types of AF-demagnetization behavior can be recognized (Figure S1 in Supporting Information S1).The majority of specimens (∼59%) are Type 1, with single component magnetite, demagnetizing beyond 10 mT straight to the origin (Figure S1a in Supporting Information S1).Type 2 specimens (∼29%) display less stable but magnetite-dominated behavior (decaying up to and beyond 40 mT, Figure S1b in Supporting Information S1).Type 3 specimens (∼1% of specimens) display a mix of stable magnetite and high-coercivity (hematite-like) behavior (with little demagnetization after 40 mT, Figure S1c in Supporting Information S1).Type 4 specimens (∼11%) display erratic behavior (Figure S1d in Supporting Information S1).Typically, for Type 1 and 2 specimens, a stable ChRM could be isolated.
Generally, the inclinations of the ChRM for the specimens younger than 1.5 ka display significant swings ( 49°t o + 32°, Figure 3b), with an average value of 11.3°; that is, similar to the expected geocentric axial dipole (GAD) magnetic field for Lake Chala of 6.6°.This similarity suggests that the ChRM is unaffected by inclination shallowing (Zhu et al., 1999).Because the core was not azimuthally oriented while sampling, the declination record is not deemed reliable, and it is not considered further in this study (see data on MagIC, Di Chiara et al., 2023, and on figshare; Hounslow et al., 2023).

Quantitative Comparison of the Lake Chala to PSV Records of Lakes Victoria and Malawi
The new inclination record from Lake Chala, covering the last 23 ka, is now compared with the PSV records from Lakes Victoria and Malawi, both lying within its 2,000 km radius.PSV records from locations farther than 2,000 km radius are not compared due to significant relocation error (Casas & Incoronato, 2007).
Uncertainty in PSV datasets is often related to the recording process in the sediment archive (e.g., inclination flattening, lock-in depths, undetected sediment disturbance), or to the uncertainty in the declination and inclination measurements, or the sampling resolution.Substantial uncertainty can also often attach to the age model, which places the measured PSV on the calendar scale.Chronological disparity may also relate to differences in magnetization lock-in depths (Bliedtner et al., 2023;Mellström et al., 2015;Sakuramoto et al., 2017) as well as analytical and calibration uncertainty and old-carbon corrections to radiocarbon dates.Simple quantitative comparisons of PSV records are often challenging because of these issues.Ignoring uncertainties in the age models, for example, and assuming that the lock-in depth is consistent, is a simple approach, but clearly flawed.One solution is to relax the age models to obtain better coherence between PSV records (Nilsson et al., 2018;Reilly et al., 2018).Other approaches to compare and correlate PSV records involve classic VGP relocations (Casas & Incoronato, 2007;Noel & Batt, 1990) of the records to a common locality (e.g., Turner & Corkill, 2023), or performing linear regression between subsequent pairs of selected tie-points (e.g., using StratFit software Sagnotti & Caricchi, 2018;Caricchi et al., 2022) to estimate the equivalent stratigraphic depth of the correlated curve in the depth scale of a selected master curve, but this approach is qualitative.Recently, Reilly et al. (2023) first aligned PSV irrespective of the ages (Geomagnetic Network Analysis) by rearranging the available radiocarbon dates, combining and stacked into one age-depth profile.
Here, to compare PSV datasets, we use sequence slotting, an objective method for relating different sets of stratigraphic sample data (Benito & Birks, 2020;Gary et al., 2005;Thompson et al., 2012;Wakefield et al., 2022).Sequence slotting parameterizes the relationships between datasets by evaluation of a distance metric (in our case a Euclidean distance) and finds an optimum slotting by utilizing this metric when the PSV records are combined.We use the CPLSlot version of sequence slotting (Hounslow & Clark, 2023) based on Clark (1985).Sequence slotting does not utilize time or age models in its evaluation, but only the relative stratigraphic order of sample datasets and the coherent structure in the compared data (Thompson et al., 2012).In order to use time as a constraint in sequence slotting, we introduce a time variable for the PSV age model (scaled from 0 to 1.0, with the 1.0 corresponding to the maximum age in the comparison datasets).In the case of full declination/inclination, directions are converted to X, Y, Z direction cosines [X = cos(inc)*cos(dec); Y = sin(dec)*cos(inc); Z = sin(inc)] and used in the slotting.When the declination is not available, only inclination (i.e.Z) is used.So for full directional sets, X, Y, and Z are used in the slotting and for comparison of inclination-only sets, only Z is used.To evaluate the relationships between PSV sets, we progressively relax the age control by changing the relative weight used with the X, Y, Z components, symbolized as wt XYZ .Default slotting uses weights for all variables set to 1.0, so that each variable in the dataset has an equal contribution to the distance metric (assuming the scaling of each is similar; Thompson & Clark, 1989).When wt XYZ ∼0, the contribution to the slotting relationship is almost entirely dominated by the age model, while progressively larger values of wt XYZ include more structure from the directional data.

Geochemistry, Geophysics, Geosystems
Hence, by increasing the values of wt XYZ , the age model is progressively relaxed and the structure in the PSV is increasingly used to define the slotting relationship between the records.At some point during the relaxation of the age model, the fitted PSV data should show better coherence, but without much increase in the age-mismatch of the data.This introduced additional age mismatch is appropriately judged by the radiocarbon age uncertainty or magnitude of the old-carbon corrections.
To demonstrate this approach, we use two test cases, the details of which are outlined in Supporting Information S1 (Figures S3-S6).The first uses only inclination data, and the second uses both declination and inclination.Slotting uses a reference (or A-seq) dataset in which the B-seq sample-set are slotted into the A-seq sample order, so the final slotted model is in the age or depth model used by the A-seq.Various statistics assess the short-scale and long-scale coherence between the two datasets (Table 1).Note.The values at wt XYZ = 0.001 are correlations effectively without slotting-age adjustments, and some of those in the optimum columns are from the slottings in Figure 6, and Figures S5 and S7 in Supporting Information S1.V = using VGP latitude, longitude with a rotation scaling of vertical to longitude, latitude = 000°,45°, to remove the fold-around of VGP longitude across the 000,90°pole.Note the increase in R p and decrease in Z-MaD and Y-MaD between none and optimum slotting.See Table 1 for parameter meaning.
We apply this approach in (a) assessing the coherence in PSV datasets for east African lake cores and their similarity to the record from Chala (Table 2), and (b) producing a regional PSV composite using the Lake Chala record and the previously published Victoria and Malawi records.
First, we combine the dual core records at Lake Victoria and at Lake Malawi into a single record at each of these lakes (Figures 4 and 5) using this same slotting approach.The two core records from Lake Victoria show a good degree of similarity (Table 2), although angular divergences are larger around 1.6-2.5 ka and 5.5-6.3ka (Figure 4c; Figures S5 and S6 in Supporting Information S1).The two Lake Malawi cores are also very similar in  2021).The upper part of (a) shows the 14 C dates, mapped via the depth scale onto the PSV record (from Table 1 in Lund et al., 2021).In each case, the black curve is the Fisher mean of these data, with a smoothing window of 9 data points, which approximately matches the width of the convolution window on the pass-through magnetometer used by Lund et al. (2021).(c) The angular standard deviation of the Fisher mean, ASD = cos 1 (R/n), where n = number of points in the smoothing window.The extreme edges of this Fisher mean smoother use a minimum of a 3-data point window.Prior to 1.9 ka, the mean and ASD use a declination for LV95-7P based on interpolation of the declination from LV95-1P.A comparison of the slotting and original inter-core correlations in depthscale is shown in Figure S6 in Supporting Information S1.
PSV structure, with angular divergences largest near some of the major swings in declination (Figure 5; Figure S7 in Supporting Information S1).

Comparison of Slotting Coherence Between Three East African Lake PSV Records
Overall, the slotting approach results in a significant improvement in PSV coherence (larger Pearson correlation coefficients, R p , and lower mean absolute deviation, MaD; Table 2) between the three East African lakes at the optimum wt XYZ .For example, the Victoria/Malawi comparisons to Chala display improvements in R p of ∼0.5.For the comparison of Lake Victoria to Lake Malawi PSV, the initially negative R p (using the age models alone) changes to moderate correlation coefficients of 0.4-0.47 for the X and Y components (Figure 6a).In the case of the Malawi-Victoria comparison, the optimum slotting interval is marked by a stabilization of the R p and MaD values after a jump in the age-σ (age model standard deviation), which also show the PSV data age-aligned on the core M98-6P (i.e., seq-A) age model, with the peaks labeled as in Lund et al. (2016).The upper part of (a) shows the 14 C dates mapped via the depth scale onto the PSV record (from Table 1 in Lund et al., 2016).In each case, the black curve is the Fisher mean of these data, with a smoothing window of 9 data points, which approximately matches the width of the convolution window on the pass-through magnetometer used by Lund et al. (2016).(c) is the angular standard deviation of the Fisher mean, ASD = cos 1 (R/n), where n = number of points in smoothing window.The extreme edges of the mean smoother use a minimum of a three-point data window.The age uncertainty intervals in (a) and (b) are for the declination and inclination peaks indicated in both cores (gray = M98-6p, blue = M98-3P) using the data in Table 2 of Lund et al. (2016).These indicate that the slotting composite largely conforms with the uncertainty in the original inter-core correlations.
nears plateaus in the same region of wt XYZ (Figures 6a and 6b).This approach also highlights those regions of poorer age and PSV fits (see the comments column in Table 2), where either the age models are inappropriate and/or the structure in the PSV data is divergent (Figure 6c).Such divergence is certainly responsible for the larger age-MaD for the Malawi-Chala and Malawi-Victoria comparisons (Figure 6b), since outside these regions the age-MaD is smaller (e.g., 168 years compared to 399 years for Malawi-Chala, Table 2).A visual expression of this is in the VGP comparison of the slotted correlation between Lake Malawi and Lake Victoria (Figure 7), and the inter-lake slotting of Lake Victoria and Lake Chala to the reference age-scale at Lake Malawi (Figure 8).
The slotting comparison approach used here is not appropriate for assessing the similarity to field models, since the age structure embedded in these models is more complex than the age/depth-based models used in the sediment core records.However, some simple model/data comparison of these is shown in the Figures S12-S14 in Supporting Information S1.

A Composite East African Lake PSV: E.AFRL.24k
Using the same slotting approach we construct a composite east-African Lake PSV (termed E.AFRL.24k) using the optimum slotting wt XYZ in Table 2.This is a composite initially based on in the age-depth model of the Lake Malawi composite (core M98-6P; Figure S10 in Supporting Information S1).For E.AFRL.24k,however, we have modified the reference core M98-6P age model.We have re-modeled the original dating control points Figure 6.The data for slotting-runs for various wt XYZ for the correlation of the PSV composites between Lake Malawi and Lake Victoria (this is using VGP space-i.e. Figure 7).The optimum weight window is marked in pink in (a) and (b).The slotting analysis suggests a significant apparent age divergence at >8.9 ka, which is either a problem with the age models or the PSV data.The latter seems the more likely possibility.Age control points are marked in (c), with the age displacements of the Lake Victoria dates (introduced by the slotting) shown by black "error bars." (mostly radiocarbon dates) available from all these cores using a new Bayesian age model (using Bchron, Parnell, 2016), and re-calibrating all the radiocarbon dates to IntCal20 (Figures 9 and 10).The modified age model highlights some divergent age control points between the three lakes' PSV records.These differences are most severe for Lake Malawi between 17 and 13 ka, where BChron flags these as potential outliers (Figure 9b).Some significant divergences are also apparent for the Lake Victoria control dates compared with the numerous 14 C dates for Lake Chala.Some of this divergence may relate to a poorer slotting over these intervals or could be a consequence of problems with the fixed-old carbon corrections applied to the Lake Malawi and Victoria 14 C dates.
For our new African composite record, E.AFRL.24k, the PSV data were relocated to Chala and uncertainty expressed in the Fisher mean direction (angular standard deviation, ASD) and MaD of age (Figure 10).The Fisher means use a window width of 9 data points (average 115 years width, maximum 610 years width), and have an average age-MaD of 200 years.The ages ranges (in original age scales) used for this are for the Malawi composite The Lake Malawi and Lake Victoria composite PSV sets with the slotting analysis applied at the optimum wt XYZ of 0.3.This can be compared to the original age models in Figures S8c and S8d in Supporting Information S1.The age discordance at >8.9 ka (see Figure 6) is the reason for the clumped data levels at ca.9 ka.The age model is that of Lake Malawi core M98-6P.from 0.450 to 26.628 ka, for the Victoria composite from 1.303 to 11.436 ka and for the Chala record from 0.396 to 23.166 ka.From the Malawi and Victoria composites the full directional (declination and inclination) set could be used, while from the Chala record only the inclination data.Thus, the declination record is entirely that of Lake Malawi and Victoria, whereas the inclination uses all three (Chala, Victoria and Malawi) records.
The E.AFRL.24krecord lacks detail in the last 900 years, where field models agree well anyway (Figure S12b in Supporting Information S1); this part could be improved for E. Africa by adding the recent data from Lake Turkana data (Lund et al., 2022).However, this is not done here since Lake Chala data are sparse in this interval; hence, making use of the age control would be unviable.The strength of the Lake Chala age control for the composite PSV is of most valuable value for the older intervals.Specifically, the three PSV records display of low angular deviation at 7.4 ka, 5.5 ka, 3.5-3.1 ka, and 2.0-1.5 ka.Outside of these common intervals, the E. AFRL.24k composite is most similar to CALS10k.2 in the intervals 5.5-4 ka and 2.4-2 ka, and most similar to SHA.DIF.14k in the intervals 13.5-13.2ka, 11-10.2ka, and 9.5-8 ka (Figure S14 in Supporting Information S1).
It is apparent that development of a PSV reference curve for central E. Africa, for use both as a dating tool and to constrain the geomagnetic field behaviour in the last 20 ka, required additional robust and reproducible Figure 8.The summary of the data used for the east Africa Lake composite, with the optimum slottings used as shown in Table S1 in Supporting Information S1.These PSV records are all in Lake Chala coordinates and are limited to the age overlap of the three datasets.The age model used here is that at Lake Malawi, since this is the reference proxy-scale used for age in the slotting analysis.The Bayesian age model shown in Figure 9 applied to these data is shown in Figure S10 in Supporting Information S1.
records.Such records may be available within some of the 35 large-and medium-sized lakes and more than 150 crater lakes (Tiercelin & Lezzar, 2002) located in the East African Rift System.The approach used here, especially if combined with more comprehensive age modeling of lake sediment successions, could improve the resultant composite PSV records.Ideally, future analytical developments will better combine the structure, correlations and uncertainty in lake PSV datasets and the uncertainty in the age models to build improved composite PSV age models.Future paleomagnetic studies could include volcanic deposits from the East African Rift System, with its an abundance of active, dormant, or extinct Quaternary volcanoes (Fontijn et al., 2012;Macdonald, 2002;Sengör & Burke, 1978).For instance, in the central-south section of the East The age control points used for the BChron age model.The small histograms are the probability densities for each date (color coded by Lake, as labeled in the key), with the histogram base placed at the position of the age in the Lake Malawi-age model composite (i.e., X-scale in Figure 8).The X-scale shows intcal20 calibrated ages, with prior old carbon corrections applied.These were blanket 500 and 450 years corrections for Lake Malawi and Lake Victoria but more complex old carbon corrections for Challa (Blaauw et al., 2011).The three lines through the probability densities are the median and 95% confidence interval from BChron.(b) The outlier probability of the dates with color and symbols according to lake.
African Rift System, 17 of the 25 volcanoes in Kenya, 9 of 21 in Tanzania, 7 of 8 in Uganda and all 3 in Rwanda have erupted at least once during the Holocene.Given independent age control, these volcanic records can contribute both to refining the PSV curve for central East Africa and could providing a robust basis for local volcanic risk assessment.

Conclusions
This study presents a new, high-quality inclination record from Lake Chala in equatorial East Africa, covering the last 23 ka.A key strength of this record is the very well constrained age model available for the Chala PSV data, of much greater detail than exists for the PSV datasets from Lake Victoria and Malawi.Our results show that the finely laminated sediments carry a stable magnetic signal with magnetite and titanomagnetite as the main carriers.
We compiled the Chala inclinations record with PSV curves from Lake Victoria and Lake Malawi to produce a PSV composite for east African Lakes, adopting a new technique which improves the coherence between the records, whilst still constrained by the sediment age models.This composite is located at the Chala coordinates.The Chala PSV record can be hence included in an increasingly improving dataset for construction of a robust PSV master curve for the central eastern African region.Suaster curve will provide a valuable and independent tool for dating not only Pleistocene and Holocene volcanic deposits from the African Rift but also the abundant archeological sites of this area, key for understanding hominid evolution.Meeren and all DeepCHALLA parties are thanked for their diverse assistance.We thank the DeepCHALLA sampling party and Heather, Erin and Els for their help with the sampling at the UGent core repository and Tomasz Gonet with the magnetic measurements.Javier Pavòn Carrasco provided useful insight and discussion regarding the relocation of PSV curves.

Figure 2 .
Figure 2. (a).Modified Maher Plot(Maher, 1988), which provides a proxy for magnetite grain size, with black symbols indicating the data from Chala and the numbers refer to the sample numbers (see the SI table, https://doi.org/10.6084/m9.figshare.24175503).Gray symbols show the published data, the numbers indicating particle sizes of the magnetite (both synthetic and natural/ground) in micrometers (μm symbol omitted in the plot).(b)Peters and Thompson (1998) plot showing the backfield IRM at 100 mT normalized to the SIRM versus the ARM at 80 mT demagnetized at 40 mT field normalized by the ARM at 100 mT (static fields).Hardness refers to the coercivity increasing toward higher values of the ratios.

Figure 4 .
Figure 4.The data for the Lake Victoria composite model (at wt xyz = 0.8) based on the slotting analysis.(a) and (b) show the PSV data age-aligned on the core LV95-1P age model, with the peaks labeled as inLund et al. (2021).The upper part of (a) shows the 14 C dates, mapped via the depth scale onto the PSV record (from Table1inLund et al., 2021).In each case, the black curve is the Fisher mean of these data, with a smoothing window of 9 data points, which approximately matches the width of the convolution window on the pass-through magnetometer used byLund et al. (2021).(c) The angular standard deviation of the Fisher mean, ASD = cos 1 (R/n), where n = number of points in the smoothing window.The extreme edges of this Fisher mean smoother use a minimum of a 3-data point window.Prior to 1.9 ka, the mean and ASD use a declination for LV95-7P based on interpolation of the declination from LV95-1P.A comparison of the slotting and original inter-core correlations in depthscale is shown in FigureS6in Supporting Information S1.

Figure 5 .
Figure5.The Lake Malawi composite model (at wt xyz = 0.6) based on the slotting analysis.(a) and (b) show the PSV data age-aligned on the core M98-6P (i.e., seq-A) age model, with the peaks labeled as inLund et al. (2016).The upper part of (a) shows the 14 C dates mapped via the depth scale onto the PSV record (from Table1inLund et al., 2016).In each case, the black curve is the Fisher mean of these data, with a smoothing window of 9 data points, which approximately matches the width of the convolution window on the pass-through magnetometer used byLund et al. (2016).(c) is the angular standard deviation of the Fisher mean, ASD = cos 1 (R/n), where n = number of points in smoothing window.The extreme edges of the mean smoother use a minimum of a three-point data window.The age uncertainty intervals in (a) and (b) are for the declination and inclination peaks indicated in both cores (gray = M98-6p, blue = M98-3P) using the data in Table2ofLund et al. (2016).These indicate that the slotting composite largely conforms with the uncertainty in the original inter-core correlations.

Figure 7 .
Figure7.The Lake Malawi and Lake Victoria composite PSV sets with the slotting analysis applied at the optimum wt XYZ of 0.3.This can be compared to the original age models in FiguresS8c and S8din Supporting Information S1.The age discordance at >8.9 ka (see Figure6) is the reason for the clumped data levels at ca.9 ka.The age model is that of Lake Malawi core M98-6P.

Figure 9 .
Figure 9. (a)The age control points used for the BChron age model.The small histograms are the probability densities for each date (color coded by Lake, as labeled in the key), with the histogram base placed at the position of the age in the Lake Malawi-age model composite (i.e., X-scale in Figure8).The X-scale shows intcal20 calibrated ages, with prior old carbon corrections applied.These were blanket 500 and 450 years corrections for Lake Malawi and Lake Victoria but more complex old carbon corrections for Challa(Blaauw et al., 2011).The three lines through the probability densities are the median and 95% confidence interval from BChron.(b) The outlier probability of the dates with color and symbols according to lake.

Figure 10 .
Figure10.The East Africa Lake composite PSV with the BChron age model from Figure9.Directional mean determined using the Fisher mean (9 point data window) of the data in Figure8(Lake Chala declination used is an interpolated average of lake Victoria and Malawi adjacent to the Chala sample in the slotting).In (a) and (b) the gray bars represent the ±ASD values of the mean.

Table 1
Statistics Generated From Slotting MaD The mean absolute deviation of the difference between the values of age (in years), Z 1 or Y 1 in A-seq, and the B-seq value at the slotted age a Linearly interpolated value of Y or Z from the A-seq value at the slotted age for B-seq used in this calculation with the corresponding B-seq value.

Table 2
Summary Statistics of the Comparison of PSV Datasets Using Sequence Slotting for East Africa Lakes This study is part of the International Continental Scientific Drilling Program through the DeepCHALLA project (https://www.icdp-online.org/projects/world/africa/lake-challa/) and was partially supported by the NERC fund (NE/P011969/1).The recovery of the Lake Chala sediment record was facilitated by the government of Kenya (permit P/16/ 7890/10400 from the National Commission for Science, Technology and Innovation), license (EIA/PSL/3851 from the National Environmental Management Authority), and research passes for foreign nationals issued by the Department of Immigration; and by the government of Tanzania through permits NA-2016-67 (270-285) and NA-2016-201 (277-292) from the Tanzania Commission for Science and Technology (COSTECH), permit EIA/10/0143/V.I/04 from the National Environmental Management Council, and resident permits issued by the Immigration Department.Thijs Van der