Differing local‐scale responses of Bolivian Amazon forest ecotones to middle Holocene drought based upon multiproxy soil data

Uncertainty remains over local‐scale responses of ecotonal Amazonian forests to middle Holocene drying due to the scarcity, and coarse spatial resolution, of lacustrine pollen records. This paper examines the palaeoecological potential of soil phytoliths, stable carbon isotopes and charcoal for capturing local‐scale ecotonal responses of different types of Bolivian Amazonian forest to middle Holocene climate change. Soil pits 1 m deep were dug at ecotones between rainforest, dry forest, Chaco woodland and savannah, and sampled at 5–10 cm resolution. Both phytolith and stable carbon isotope records indicate stability of dry forest–savannah ecotones over the last ca. 6000 years, despite middle Holocene drought, revealing the dominance of edaphic factors over climate in controlling this type of ecotone. In contrast, δ13C data reveal that rainforest–savannah ecotones were more responsive to climate change, with rainforest likely replaced by drought‐tolerant dry forest or savannah vegetation during the mid‐Holocene, consistent with regional‐scale lacustrine pollen records. However, such shifts are not apparent in most of our phytolith records due to insufficient taxonomic resolution in differentiating rainforest from dry forest. Charcoal data show that ecotonal dry forests experienced greater fire activity than rainforests and that recent high fire activity at all forest sites is unprecedented since at least the middle Holocene.


INTRODUCTION
Ecotones are of significant scientific interest for several reasons. They exhibit high biodiversity (beta-diversity) due to their mosaic of different ecosystems (i.e. habitat heterogeneity), they play a key role in rainforest speciation via allopatric and sympatric mechanisms (Smith et al., 1997), and they will likely become important migration corridors, facilitating species range shifts in response to climate change (Hannah et al., 2002). Southern Amazonian ecotones face the twin threats of deforestation, driven by agricultural expansion (Miles et al., 2006;Nobre et al., 2016;Prieto-Torres et al., 2021) and climate change, i.e. increasing drought frequency and fire activity (Brando et al., 2014;Brienen et al., 2015;Cochrane et al., 1999;Feldpausch et al., 2016).
Palaeoecology can be a useful tool for examining ecotonal dynamics by providing empirical palaeo-evidence of ecotonal responses to past intervals of drier climatic conditions. Such palaeodata are useful for testing the ability of Earth System models to predict the ecological impacts of future climate change . Palaeoclimate data from a variety of proxies across a broad network of sites (e.g. Baker et al., 2001;Bird et al., 2011;Reese et al., 2013;Seltzer et al., 2000;Thompson et al., 1998) show that southwestern Amazonia was markedly drier than at present during the middle Holocene (9000-3000 cal a BP), due to a weakening of the South American summer monsoon (Cruz et al., 2009;Prado et al., 2013).
Fossil pollen records from several large lakes in northern Bolivia (Lagunas Bella Vista, Chaplin, Oricoré, Rogaguado) have revealed expansion of dry forest and savannah at the expense of rainforest in response to middle Holocene drought (Brugger et al., 2016;Burbridge et al., 2004;Carson et al., 2014;Mayle et al., 2000;Smith and Mayle, 2018). In the Chiquitano dry forest of easternmost Bolivia, pollen data from two other large ecotonal lakes -Lagunas La Gaiba (Whitney et al., 2011) and Mandioré (Plumpton et al., 2019)point to resilience of the dry forest biome to both drought and fire (i.e. Power et al., 2016), albeit with floristic turnover favouring more drought-tolerant taxa. (Note: in this paper the term 'resilience' is used according to Holling's (1973) definition; i.e. the amount of disturbance an ecosystem can withstand without changing self-organised processes and structures).
However, while these lacustrine fossil pollen records provide clear evidence for climate-driven, middle to late Holocene ecotonal responses in southwestern Amazonia, considerable uncertainty remains over the precise nature of these ecotonal responses, due to: a) the paucity of suitable lakes for pollen analysis, b) the mismatch in spatial resolution between the coarse-scale pollen catchments of these large lakes versus the fine-scale spatial heterogeneity of forest-savannah ecotones, and c) uncertainty in palynological differentiation of different types of savannah (upland versus seasonally flooded) and Chaco woodland versus dry forest (Gosling et al., 2005(Gosling et al., , 2009Jones et al., 2011). The aforementioned lakes are over several kilometres in diameter, with regional-scale pollen catchments >100 km 2 , encompassing and amalgamating spatially complex mosaics of different types of tropical forest and savannah, reflecting local-scale differences in topography, hydrology (i.e. flooded savannahs) and edaphic factors (i.e. upland savannahs). These lacustrine pollen records are therefore poorly suited to differentiating the responses of different types of ecotone to middle Holocene drought over fine spatial scales. Furthermore, the paucity of sufficiently old lakes limits the ability of lacustrine pollen data to capture the full range of ecotonal responses across Amazonian Bolivia, a region which encompasses a diverse range of vegetation types due to the north-to-south gradient in precipitation. Middle Holocene age lakes are scarce in the fluvially dominated rainforest-savannah ecotones of northernmost Bolivia, and even more so in the Chiquitano dry forest in eastern Bolivia and Chaco woodland further south.
Stable carbon isotope (δ 13 C) records from soil profiles can be used to indicate the ratio of C3/C4 plants that persisted in past environments, allowing for the differentiation of C3 forest versus C4 savannah ecosystems (Pessenda et al., , 2004. Unlike lacustrine pollen records, these soil records reflect vegetation patterns at ultra-local spatial scales, from plants that have decomposed in situ, meaning that they have the spatial resolution to detect subtle ecotonal shifts and determine whether middle Holocene drought exerted a greater control over ecotonal dynamics compared with local edaphic or hydrological factors. Stable carbon isotope analyses from soil pit transects across ecotones in Rondonia state, northwestern Brazil (de Freitas et al., 2001;Pessenda et al., 1998Pessenda et al., , 2001 and central Brazil (Wright et al., 2020) corroborate the lake pollen records by evidencing ecotonal savannah expansion in response to middle Holocene drought. However, no soil-based stable carbon isotope studies have yet been conducted in lowland Amazonian Bolivia, where forest-savannah ecotones are best preserved.
Soil phytolith analysis is an alternative technique to reconstruct vegetation histories. It has long been used as an 'on-site' archaeobotanical tool to explore pre-Columbian crop cultivation and land use (e.g. Hilbert et al., 2017;Watling et al., 2015), but is still in its infancy as a palaeoecological tool for exploring past land use beyond archaeological settings (i.e. 'off-site') (e.g. Iriarte et al., 2020;McMichael et al., 2015;Watling et al., 2017), and its potential for determining the impact of Holocene climate change upon ecotonal forest dynamics has yet to be explored in Amazonia.
The overall aim of this paper is to examine the effectiveness of soil phytolith, bulk stable carbon isotope and charcoal analyses in capturing local-scale responses of different types of ecotonal tropical forest to drier than present middle Holocene climatic conditions across Amazonian Bolivia.
Specific questions: 1. How do soil-based phytolith and stable carbon isotope palaeovegetation records compare with each other and with previously published lacustrine fossil pollen records, and how can any differences be accounted for? 2. What is the relative importance of climate versus edaphic factors in controlling different types of forest-savannah ecotone? 3. What is the relationship between fire activity, climate change and ecotonal dynamics since the middle Holocene?

STUDY AREAS AND APPROACH
These aims are addressed by examining phytoliths, stable carbon isotopes and macroscopic charcoal from the same soil samples in 1 m deep soil pits, to enable direct interproxy comparisons. Soil pits were excavated from different types of ecotonal forest across Bolivian Amazonia: interfluvial and fluvial evergreen rainforest, 'Chiquitano' semideciduous dry forest, cerrado (non-flooded) savannah, and 'Chaco' woodland (Table 1; Fig. 1). The soils at each site are classified as 'ferrasols', which are strongly weathered, acidic, and often have a sand/clay dominant texture (Table 1; Quesada et al., 2011). Figure 1 shows the distribution of the four major vegetation types in lowland Amazonian Bolivia, which are examined in this study. Humid evergreen tropical forest (hereafter referred to as 'rainforest') grows in northern and northeastern Bolivia, where mean annual precipitation exceeds 1500 mm and the dry season is less than 3-4 months (Ibisch et al., 2004). Due to the dense tree canopy, there is only a sparse herbaceous understorey. Further south, in central and eastern Bolivia, where mean annual precipitation is <1500 mm, and the dry season is at least 5-6 months, rainforest grades into a complex mosaic of 'Chiquitano' semideciduous dry forest and 'cerrado' savannah in non-flooded areas. This Chiquitano dry forest has a more open canopy, and hence a denser herbaceous understorey, compared with rainforest, and also supports Cactaceae and Bambusoideae (Pennington et al., 2009). Cerrado savannah comprises scattered, fire-adapted trees, with a ground layer of grasses and other light-demanding herbs. The dry forest, in turn, is ecotonal with Chaco woodland in southern Bolivia, which, unlike dry forest, is entirely deciduous and consists primarily of short thorny trees and scrub, alongside columnar Cereus cacti. Although the seasonally flooded 'Llanos de Moxos' covers a large portion of Amazonian Bolivia, our study sites are located in terra firme landscapes beyond this sedimentary basin.

METHODS
At most forest ecotone sites (Table 1), single, 1 m deep soil pits were dug at the perimeter of 1 ha (20 × 500 m) forest inventory plots established by the RAINFOR network (Lopez-Gonzalez et al., 2009, between 100 m and 10 km of the ecotone. However, at the Florida site (adjacent to Noel Kempff Mercado National Park (NKMNP)), two neighbouring soil pits were dug, 150 m apart, to enable the spatial representativeness of soil pits to be assessed. This pair of rainforest soil pits lies 2 km away from a savannah island, and ca. 50 km north of the ecotone with Chiquitano dry forest. Furthermore, this pair of Florida soil pits are located only 20 km from the Laguna Chaplin fossil pollen record, providing the opportunity to directly compare local-scale, soil-based, phytolith, charcoal and δ 13 C reconstructions with those of regional-scale lacustrine pollen records.
At each soil pit, soil samples were collected from the cleaned profile face in contiguous 5 cm increments. Soil-surface samples (0-5 cm depth) were also collected from the savannah (n = 3) and Chaco (n = 1) to create modern phytolith and δ 13 C analogues (Table 2; Fig. 2). Leaf litter samples were collected from rainforest (n = 3) and dry forest (n = 3) to provide modern δ 13 C analogues (Table 2; Fig. 2).

Phytoliths
Phytolith analysis was undertaken at 5-10 cm resolution using the wet oxidation method (Piperno, 2006). Samples were divided into 'A' silt (<50 µm) and 'C' sand (>50 µm) fractions. Silt-fraction phytoliths were examined at 500× magnification. A minimum of 200 phytoliths were counted, and the rest of the slide was scanned to identify other diagnostic types. For the sand fraction, the entire slide was scanned at 200× magnification and all diagnostic taxa were counted. Identifications were made using published atlases (Dickau et al., 2013;Iriarte and Paz, 2009;Morcote-Ríos et al., 2016;Piperno, 2006;Piperno and Pearsall, 1998;Watling et al., 2016) as well as the phytolith reference collection at the University of Reading Tropical Palaeoecology Laboratory, which is based on herbarium material from the 'Noel Kempff Mercado' Natural History Museum in Santa Cruz, Bolivia.

Charcoal
Macroscopic charcoal was analysed for: a) the reconstruction of long-term trends in fire activity, via changes in charcoal particle abundance through the profile, and b) δ 13 C analysis at El Tigre and Tumichucua sites, to assist the interpretation of bulk-soil carbon isotopic records; i.e. the extent to which bulksoil δ 13 C records are influenced by microbial decomposition (Wynn, 2007). Charcoal concentration was based on 3 cm 3 of soil, sampled every 5 cm, using a modified macroscopic sieving method with >250 µm and 125-250 µm size classes to distinguish between local versus extra-local charcoal sources, respectively (Clark, 1988;Watling et al., 2017;Whitlock and Larsen, 2001). Macroscopic charcoal pieces >0.5 cm in size, were also collected for radiocarbon dating (Table 3). Four AMS dates per soil pit were obtained and calibrated to 2σ accuracy using the IntCal 20 calibration curve using OxCal 4.4 (Bronk Ramsey, 2009;McCormac et al., 2004;Reimer et al., 2020).

Stable carbon isotopes
Stable carbon isotopes (δ 13 C) serve as an independent palaeovegetation proxy which can be compared with the phytolith record, to provide corroboratory data on vegetation history. The δ 13 C from bulk-soil organic matter (SOM) was analysed from each profile at 5 cm increments using a Thermo Fisher Delta V Plus Isotope Ratio Mass Spectrometer (IRMS) fitted with an Elemental Analyser and expressed using the delta (δ) notation as per mille (‰) deviations relative to the Vienna Pee Dee Belemnite standard (VPDB). Analytical precision on circa 100 μg C sample of international standards was better than 0.1‰. Samples were run alongside international standards and were drift and stretch corrected. Each sample was replicated three times and an average value was used. Surfacesoil δ 13 C signatures incorporate changes resulting from fossil fuel emissions over the last century, which can potentially skew soil δ 13 C values by up to −1.5‰ (Hare et al., 2018). The latter was corrected for using the calculation found in Bostrom et al. (2007).
SOM 13 C values can increase by up to 6‰ due to microbial decomposition (Wynn, 2007), potentially obscuring downprofile δ 13 C changes associated with changing proportions of C4 savannah versus C3 forest. However, charcoal is highly recalcitrant in soils, is not subject to microbial decomposition, preserves the original δ 13 C of the parent plant and can be used as a proxy for past climatic changes, since δ 13 C values are known to be influenced by mean annual precipitation (Hare et al., 2018;Kohn, 2010). Therefore, the δ 13 C from the charcoal fractions at El Tigre and Tumichucua (two sites that feature increasing bulk δ 13 C trends <6‰) were also analysed, alongside the δ 13 C values associated with each charcoal-based AMS date (Table 3) for each of the other sites, to gain δ 13 C records not impacted by microbial decomposition.

Soil analysis
Available soil nutrients play a role in governing the modernday distributions of some tropical forests and savannah, like dry forests which occur on more fertile soils, with higher calcium and magnesium levels compared with neighbouring    (Dubs, 1992;Ratter, 1992;Furley and Ratter, 1988). Therefore, carbon, nitrogen and nutrient exchangeable cations (Ca, Mg, K, and Na) were extracted to measure soil nutrient availability. By measuring cations alongside soil palaeovegetation proxies, the role that soil nutrients played in governing ecotones at precise spatial scales over the middle to late Holocene can be examined. Exchangeable cations were determined using the ammonium acetate leaching procedure (Rowell, 1994) at 10 cm stratigraphic resolution and expressed as cmolc/kg soil.
Physical and chemical soil properties were measured, as some studies show they may influence phytolith stratigraphy, e.g. soil pH affecting phytolith preservation (Alexandre et al., 1997;Fraysse et al., 2006) and soil particle size influencing phytolith translocation (Fishkis et al., 2010;Fraysse et al., 2006). The pH was measured using a calibrated pH meter on samples taken at 10 cm stratigraphic intervals. Soil particle size was measured at 5 cm intervals using a Mastersizer 3000 laser diffraction analyser on the <2 mm fraction after pretreatment with a mortar and Calgon. The  Bulk charcoal n/a n/a n/a n/a volume-based percentages produced by laser diffraction were converted to mass-based percentages using a calibration model (Yang et al., 2015) as laser diffraction underestimates the proportion of clay particles (Campbell, 2003).

Multivariate analysis
A constrained redundancy analysis (RDA) was conducted to help visualise the dataset and to examine the extent to which the phytolith assemblages are negatively influenced by environmental variables (i.e. soil texture, chemistry, charcoal). Constrained ordination was selected over unconstrained ordination (i.e. principal component analysis (PCA)) as it can highlight which environmental variables, if any, best explain species composition, while unconstrained ordinations provide a more general analysis, searching for any variable which best explains the dataset (Legendre and Birks, 2012;McCune and Grace 2002). RDA was selected over canonical correspondence analysis by performing detrended correspondence analysis, which revealed relatively short gradients in the dataset appropriate for RDA (Braak and Prentice 2004).

Modern plant-phytolith and plant-isotope relationships
Analysis of soil-surface phytolith assemblages and δ 13 C values were used to draw palaeoecological inferences from our soil profiles. Phytolith surface-sample assemblages from this study ( Fig. 2; Table 2), as well as from Dickau et al. (2013) and Watling et al. (2016Watling et al. ( , 2020, demonstrate that tropical forests can be differentiated from savannah (both open and wooded), based on different ratios of arboreal versus Poaceae phytolith morphotypes (Table 2). Phytolith assemblages from forest ecosystems are dominated by arboreal and palm taxa (>40%) and have low abundances of Poaceae (<20%) and other herbs, while those of savannah ecosystems are dominated by nonbamboo grass (>50%) and have lower percentages of palms (<20%) (see Fig. 2 and Dickau et al. (2013)).
On the other hand, rainforest and dry forest soil phytolith assemblages are generally indistinguishable from each other due to considerable overlap in bamboo phytolith percentages (Figs. 3-4;Dickau et al., 2013;Watling et al., 2016Watling et al., , 2017 and an inability to differentiate most dicotyledonous arboreal taxa beyond the rugulose-sphere morphotype (e.g. Iriarte et al., 2020;McMichael et al., 2012;Robinson et al., 2020;Watling et al., 2017Watling et al., , 2018. Consequently, any turnover between arboreal rainforest versus dry forest taxa is likely to be masked by low phytolith taxonomic resolution. Furthermore, the extent to which phytolith assemblages of Chaco woodland can be differentiated from those of tropical forests is also uncertain, given that our Chaco soil-surface phytolith sample, which differs from dry forest with respect to Asteraceae and Chloridoideae grass percentages, is the only such sample to have been collected, not only in the Bolivian Chaco, but the Gran Chaco ecoregion as a whole. Bulk δ 13 C surface-soil samples collected from NKMNP (Table 2) by Dickau et al. (2013), as well as leaf litter samples taken from rainforest (−31 to −32‰) and dry forest (−28 to −29‰) collected in this study (Table 2) and previous studies (e.g. Ometto et al., 2006;Sobrado and Ehleringer, 1997;Mooney et al., 1989) demonstrate that C3-dominated closedcanopy tropical forests (−22 to −32‰) can be distinguished from C4-dominated open savannahs (−19 to −22‰). δ 13 C values of C3 taxa from a range of temperate and tropical forest ecosystems vary widely between −22 and −36‰ (Hare et al., 2018;Kohn, 2010). The latter correlate with mean annual precipitation (MAP), with δ 13 C values > −25‰ in arid regions (MAP <500 mm/yr) and δ 13 C < −31‰ in closedcanopy tropical rainforests where MAP is >2000 mm/yr (Kohn, 2010). This means that δ 13 C values within the C3 range can potentially be used as a proxy for past climatic changes (e.g. middle Holocene drought), although they cannot reliably differentiate the different forest types in our study region. Soil-surface and leaf-litter data demonstrate this overlap in δ 13 C values between rainforest, dry forest, wooded savannah and Chaco woodland ecosystems (Table 2; Dickau et al., 2013;Ometto et al., 2006;Sobrado and Ehleringer, 1997;Mooney et al., 1989;Gatica et al., 2017). As with the phytoliths, soil-surface δ 13 C data are lacking from the Chaco woodland environment, so the representativeness of the single δ 13 C soil sample (Fig. 2) for this ecosystem is uncertain.

Chronology and age-depth relationships
Soils experience considerable vertical mixing, of both minerogenic and organic matter, due to a number of factorsdeep burrowing of both vertebrates and insects; deep root penetration from large trees; and up-rooting of large trees through wind-throw (Hart, 2003). Consequently, age-depth relationships are less secure for most soil profiles than they are for most lake sediments, and the soil profiles in this study are no exception (Table 3). The El Tigre profile shows the most robust age-depth relationship, with no age reversals and a middle Holocene basal date of 7760 cal a BP. Age reversals, or anomalously young or old ages, occur at the other sites to varying extents. However, even for those sites where the radiocarbon ages show little age-depth relationship at all (i.e. the Ottavio and Santa Cruz dry forest sites), one can reasonably infer a minimum basal age of the profile based upon the age of the oldest charcoal sample, irrespective of where it lies in the profile. For example, the oldest dates for Ottavio (5450 cal a BP), Santa Cruz (5880 cal a BP) and Florida 1 (7670 cal a BP) are all middle Holocene, conforming with the basal (or near-basal) ages for the rainforest sites El Tigre (7760 cal a BP), Los Tajibos (7060 cal a BP) and Tumichucua (6340 cal a BP). Although the oldest date at the Florida 2 site is only 2000 cal a BP, the fact that this pit is located only 150 m from the Florida 1 site leads us to infer that this 1 m profile also likely dates to the middle Holocene. At all sites, the dated charcoal particles are over several millimetres in size and therefore originate from woody plants growing in close proximity to the soil pit. Because none of our sites are flood-prone and the soil pits were not dug in hollows or valley bottoms, input of anomalously old charcoal from beyond the study site, via flooding or erosion, can be discounted. Considering all these radiocarbon ages in the round, together with middle Holocene basal ages for soil pits of similar depth elsewhere in Amazonia (e.g. Balesdent et al., 2018;McMichael et al., 2015;Pessenda et al., 1998Pessenda et al., , 2001Pessenda et al., , 2004Watling et al., 2017), we can cautiously infer that all our 1 m deep soil profiles likely span the middle Holocene to the present day. Despite Bolivia's Chiquitano dry forest being a global conservation priority, as the largest and most threatened intact block of dry forest remaining in South America (Miles et al., 2006), its palaeoenvironmental history remains poorly understood, based on only two neighbouring fossil pollen records from its eastern ecotone (Whitney et al. 2011;Plumpton et al. 2020)). The Ottavio and Santa Cruz data presented here are the first soil profile palaeodata from this ecosystem. Therefore, despite their weak chronological control, these two records are nevertheless informative and present an important step forward in

Soil profile results and palaeovegetation interpretations Florida 1 and 2 interfluvial rainforest
The dominance of arboreal phytoliths (ca. 70-90%; Figs. 3 and 6), low percentages of palm phytoliths (ca. 10%, except for the ca. 30% peak at the surface of Florida 1), and negligible values of Poaceae and bamboo phytoliths (<2%) throughout the two profiles are consistent with continuous rainforest cover at the Florida site throughout the middle to late Holocene. At first glance, these data appear to contradict fossil pollen records from the two large lakes in NKMNP -Lagunas Chaplin and Bella Vistawhich instead show a >130 km expansion of dry forest and savannah during the middle Holocene, relative to the present (Mayle et al., 2000;Burbridge et al., 2004). We attribute this apparent contradiction to the aforementioned low taxonomic resolution of arboreal dicotyledonous phytolith taxa masking a climate-driven replacement of rainforest by dry forest.
Both soil pits show progressively heavier bulk-sediment δ 13 C values down-profile (−28 to −22‰ at Florida 1 and −28 to −23‰ at Florida 2; Figs. 3 and 9), which implies ecotonal replacement of humid rainforest by wooded savannah and/or dry forest under drier middle Holocene conditions, consistent with the Chaplin and Bella Vista lacustrine fossil pollen data (Table 2; Fig. 3). Although wooded savannah and dry forest have strongly overlapping δ 13 C values (Dickau et al., 2013), the absence of any savannah signal from the phytolith records (i.e. >40% Poaceae; Dickau et al., 2013;Fig. 2) suggests that the δ 13 C records from the pair of Florida soil pits reflect a droughtinduced ecotonal expansion of the Chiquitano dry forest at  the expense of rainforest during the middle Holocene (by ca. 50 km relative to the modern ecotone).
Tumichucua pre-Columbian ring-ditch within fluvial rainforest Abundance of arboreal phytoliths (40-60%) and palms (peaking at 60% in the upper half of the profile), together with low percentages of Poaceae and bamboo (5-20% and 5-10%, respectively) (Figs. 3-6), signify forested conditions over the last ca. 6000 yrs. As with Florida, the low taxonomic resolution of arboreal phytoliths could be masking the expansion of dry forest during the middle Holocene. Although the nearest modern-day dry forest ecotone is ca. 700 km south in northeastern Bolivia, it is known that dry forest expanded at least several hundred kilometres into the Beni basin during the middle Holocene (Lombardo et al., 2019). Dry forest could have migrated into the Riberalta region via gallery forest up the River Mamoré, as well as through long-distance wind-dispersal of seeds, outcompeting the less drought-tolerant rainforest taxa.
Small peaks in Poaceae (20%) and Bambusoideae (10%), a large peak in discoloured phytoliths (50%), and the presence of Heliconia at ca. 40-45 cm, indicate small-scale forest disturbance and also coincide with a late Holocene increase in palms (10-50%) in the upper part of the profile.
The δ 13 C values from the bulk-soil fraction (Figs. 3 and 9) become progressively more negative (−25 to −30‰) from the base of the profile (i.e. middle Holocene) to the soil surface (i.e. late Holocene), signifying a transition from drought-adapted dry forest taxa in the middle Holocene (MAP <500 mm/yr), to drought-intolerant rainforest taxa in the late Holocene (MAP >2000 mm/yr) (Kohn, 2010). Although there is large overlap in δ 13 C values between wooded savannah versus dry forest (Table 2), the absence of any savannah signal from the phytolith record implies that the middle Holocene δ 13 C signatures originate from dry forest. However, as the amplitude of δ 13 C shift through the profile is only 5‰, one cannot exclude the possibility that it reflects microbial decomposition of SOM, rather than palaeovegetation change, since soil decomposition through microbial activity has been shown to increase δ 13 C values downprofile by up to 6‰ (Wynn, 2007). This decomposition effect may at least partially explain why phytolith and δ 13 C records may not always align as expected, not only in this study but elsewhere (e.g. Watling et al., 2017;Robinson et al., 2020).
The δ 13 C data from the >250 μm charcoal fraction (Fig. 3) exhibit a weaker trend through the soil profile than that of the bulk fraction, and also feature less-negative values throughout (−23 to −26‰), but nevertheless broadly show a similar trend toward lighter δ 13 C values up-profile. Interestingly, however, the δ 13 C signal from the 125-250 μm charcoal fraction shows no such trend and remains constant between −27 and −26‰ throughout the profile, for reasons which are unclear.
Charcoal is present throughout, with peaks between 15 cm and the surface and at 50-55 cm (Figs. 3 and 8), indicating that fire activity was relatively higher during the late Holocene (upper part of the profile) and was highest over the past century (upper 15 cm).
Soil texture changes from clay/silt between the profile base and 35-40 cm depth to silty/sand above (Fig. 3). Acidity peaks from 100-60 cm (pH 4), decreasing above 60 cm. The soil profile is enriched with available nutrients, particularly with C (40% dry weight) and Ca (13 cmolc/kg) at the soil surface.

Los Tajibos fluvial rainforest
Tajibos is ca. 1 km beyond the Tumichucua ring-ditch profile ( Fig. 1) and exhibits similar stratigraphic trends in phytolith abundance of the dominant taxa (Figs. 3 and 6).
The SOM δ 13 C record of Tajibos shows a similar trend to that of Tumichucua, but of greater amplitude (ca. −20 to −29‰ up-profile) (Figs. 3 and 9), providing robust evidence for a forest-savannah ecotonal shift. Therefore, an apparent contradiction exists between the phytolith and δ 13 C records at Tajibos which we cannot explain, with the former indicating continuous forest cover through the middle to late Holocene and the latter instead indicating forest-savannah biome turnover. Both proxy records are consistent with lacustrine fossil pollen records, which show evidence for both dry forest and savannah expansion in northeastern Bolivia (e.g. Carson et al., 2014;Mayle et al., 2000).
The four δ 13 C values from the >250 μm charcoal fraction (Fig. 3) show a similar trend through the profile, but are unsurprisingly more negative (−26 to −34‰, 80-30 cm) than those in the bulk SOM fraction, since the latter contains carbon, not only from C3 woody plants, but also isotopically heavier C4 savannah grasses. These charcoal isotope data indicate a transition from more drought-tolerant C3 woody taxa (either savannah or dry forest taxa, or both) in the middle Holocene to humid evergreen rainforest taxa in the late Holocene.
Macroscopic charcoal is present throughout the profile, with peaks between 15 cm and the surface (Fig. 8), demonstrating that fire activity was present throughout the last ca. 7000 yrs but was greatest over the past century. Soil texture remains uniform silt/ sand down the profile, apart from a peak in clay (20%) at 90-95 cm (Fig. 3). The soil is acidic with a uniform pH 4 throughout. Available nutrients are lower than in the Tumichucua profile. The surface-sample peak in Na (2.5 cmolc/kg) is highest among the exchangeable cations for the site.

El Tigre interfluvial rainforest
The phytolith record is similar to that of the other rainforest sites, dominated by arboreal and palm taxa, with the latter peaking at ca. 60% at 25-30 cm depth. As at the other rainforest sites, phytolith data alone cannot discriminate between changes in rainforest versus dry forest taxa through the middle to late Holocene.
The bulk SOM δ 13 C data show a similar up-profile trend toward more negative values as found at the previous sites, but the low amplitude of this shift (−27 to −32‰; Figs. 4 and 9) cannot discriminate between shifts in dry forest versus wooded savannah versus rainforest (Dickau et al., 2013), or even whether it signifies real vegetation change or merely microbial decomposition (Wynn, 2007). However, microbial decomposition is an unlikely explanation because the δ 13 C values from the recalcitrant charcoal fractions show a trend similar to that of the bulk soil, albeit with less-negative values throughout (125-250 μm: −24 to −28‰; >250 μm: −24 to −31‰; Fig. 4). This indicates a shift from drier to wetter conditions over the middle to late Holocene, based upon a comparison with values from modern C3 taxa across a precipitation gradient (Kohn, 2010).
Macroscopic charcoal is present throughout, with small peaks between 55-60 cm and 0-10 cm (Figs. 4 and 8) indicating that fires have occurred throughout the middle to late Holocene. Since fire return intervals of less than 90 years can lead to canopy opening (Cochrane et al., 1999, Balch et al., 2015, it is likely that fires at El Tigre were rare, with only centennial-to millennial-scale frequency (Saldarriaga and West 1986;Sanford et al., 1985, Bush et al., 2008. Soil texture changes from predominantly clay/silt in the lower profile (65-100 cm; Fig. 4) to silt/sand in the upper profile. The soil profile is acidic throughout (ca. pH 3-4) and soil nutrient availability is low, with 10-17% C, <2% N, and exchangeable cations ranging from 0.9 cmolc/kg (Ca) to 0.01 cmolc/kg (K).

Ottavio interfluvial dry forest
Macroscopic charcoal concentration is three times higher than that of the rainforest profiles (Figs. 4 and 7). As with the rainforest sites, arboreal and palm phytoliths dominate the assemblages, with a similar pattern of peak arboreal percentages (ca. 50-80%) in the lower half of the profile and peak palm percentages (ca. 30-60%) in the upper half (Figs. 4  and 7). These abundances are comparable to those of modern surface-soil phytolith assemblages from other dry forest plots (e.g. Dickau et al., 2013). The 30% Poaceae peak at the base of the profile correlates with a peak in discoloured arboreal phytoliths (ca. 20%), the presence of Heliconia, and a peak in charcoal (Figs. 4 and 8), likely indicative of a period of forest disturbance.
The bulk-soil δ 13 C values are constant throughout the profile, fluctuating around −27 to −28‰ (Figs. 4 and 9), and are similar to those of modern dry forest surface soils ( Table 2). The four δ 13 C values from the >250 µm charcoal fraction are broadly similar, albeit slightly less negative (−26 to −27‰). These isotope results, considered in isolation, could indicate dry forest stability throughout the last ca. 6000 yrs or instead a drought-induced replacement of dry forest by wooded savannah in the middle Holocene, followed by dry forest expansion in the late Holocene as precipitation increases, since δ 13 C values between these two ecosystems overlap (Table 2; Dickau et al., 2013). However, when both the phytolith and carbon isotope data are considered together, the most parsimonious interpretation is continuous dry forest cover at Ottavio throughout the sequence; i.e. stability of the dry forest-savannah ecotone since the middle Holocene and thus dry forest resilience to middle Holocene drought.
Soil texture comprises a sand/silt mix throughout the profile, except for a 60% spike in sand at 60-65 cm (Fig. 4). The reason for this peak in sand percentages is unclear, especially given the lack of any commensurate fluctuations in either the phytolith or δ 13 C records. However, we can rule out a fluvial sand origin via flooding, given the absence of any nearby streams or rivers, as well as sand input via hillslope erosion/ runoff, given the relatively flat topography. The soil is slightly less acidic (pH 6) and more enriched in nutrients (e.g. 10-30% C; 5-10 cmolc/kg) than the rainforest soils.

Santa Cruz interfluvial dry forest
The phytolith record of this site is quite distinct from those of the other five sites, with markedly lower percentages of palms (10-20%) and arboreal (10-20%) types, and markedly higher percentages of grass (30-50%) and bamboo (10-20%) through most of the profile (0-80 cm depth) (Figs. 4 and 7). The surface-sample Poaceae phytolith percentage of 50% is unusual, as this value is significantly higher than that of other dry forest soils (e.g. Ottavio) and is more typical of savannah soils (Dickau et al., 2013). Nevertheless, considered alone, the phytolith record for this site is consistent with dry forest stability throughout the middle to late Holocene, albeit with changes in canopy density implied by reductions in grass percentages at 30-35 cm, 60-65 cm and 90-95 cm. However, one cannot exclude the possibility that past ecotonal shifts between Chaco woodland and dry forest occurred, but are not apparent due to limitations in phytolith taxonomic resolution and a lack of surface-sample phytolith data from Chaco ecosystems. Unfortunately, no pollen data exist at this ecotone for comparison (either fossil or modern), in part due to the scarcity of lakes/swamps required for pollen preservation.
Furthermore, the stable carbon isotope record calls into question this phytolith-based vegetation reconstruction, as the bulk δ 13 C values between ca. −24 and −18‰ through the lower two thirds of the profile (Figs. 4 and 9) are similar to those of open savannah soil-surface samples (−18 to −22‰; Dickau et al., 2013). This suggests that the dry forest-Chaco ecotone at Santa Cruz may have given way to more drought-tolerant savannah under middle Holocene drought, despite the absence of savannah from the immediate vicinity today. Alternatively, these heavier isotopic values could signify a floristic change toward increased prevalence of drought-tolerant CAM taxa (e.g. Cereus cacti) which grow in both dry forest and Chaco and have heavy δ 13 C signatures, ranging from −11 to −24‰ (Llano and Ugan, 2014). The three δ 13 C values from the >250 μm charcoal fraction show similar results to those from bulk-soil samples, albeit slightly more negative (−25 to −24‰, 85-20 cm). Despite uncertainty over the palaeoecological significance of the phytolith and stable carbon isotope profiles at Santa Cruz, the high-amplitude (10‰) shift in the bulk δ 13 C values through the profile clearly reflects real floristic or structural vegetation changes, since it exceeds the 6‰ threshold associated with microbial decomposition (Wynn, 2007).
Charcoal concentrations are generally similar to those of Ottavio through most of the profile (Figs. 4 and 8), but decrease by over 50% between 10 and 25 cm. Whether this late Holocene dip in charcoal concentrations reflects a climate-induced change in fire regime, or instead a change in human land use, is unclear. However, the upturn in charcoal concentrations in the uppermost 10 cm is consistent with the charcoal records of all the other sites in this study and points to a marked increase in anthropogenic burning in recent decades/centuries. The Santa Cruz soil profile has a silty/sandy texture with low clay content, except for a spike in clay (ca. 60%) between 75 and 80 cm (Fig. 4). The soil pH and available nutrients are broadly similar to those of Ottavio.

Potential impacts of soil pH and texture upon phytolith records
Differential chemical dissolution and translocation of phytoliths through the soil profile, due to variations in soil texture, have been shown to negatively impact the preservation and stratigraphic integrity of phytolith records in some soils (Piperno, 2006;Fishkis et al., 2010), but are unlikely to have distorted the phytolith records in our study because: a) all the soils are acidic (pH 3-6), favouring the preservation of phytoliths, and b) stratigraphic changes in soil texture do not correlate with any stratigraphic changes in the phytolith records. Although age inversions occur in some profiles, excessive bioturbation does not appear to have unduly influenced the stratigraphic integrity of either phytolith or δ 13 C profiles, because the marked biostratigraphic changes which we observe would have been smoothed out if there had been significant vertical mixing.
Redundancy analysis results (Fig. 10) All phytolith results (RDA 1) Arboreal phytoliths and palm phytoliths are negatively correlated with each other across sites, reflecting the shift from arboreal to palm-dominated phytolith assemblages that occurs over the middle to late Holocene. Arboreal taxa negatively correlate with herbs (i.e. Poaceae, Heliconia), demonstrating that reductions in arboreal cover result in the expansion of shade-intolerant herbs. Assemblages from the Santa Cruz dry forest site during the middle and late Holocene are closely correlated with Poaceae phytoliths, demonstrating the importance of these types at this site. The Florida rainforest assemblages correlate with arboreal phytoliths, highlighting the dominance of arboreal phytoliths throughout these profiles. In contrast, however, the El Tigre rainforest samples correlate with palm, particularly in the middle to late Holocene, demonstrating the increasing importance of palm through the Holocene at this site. The Ottavio dry forest site has a similar transition, with the early and middle Holocene assemblages correlating more closely with burnt and arboreal phytoliths, respectively, while samples from the late Holocene correlate more with palms.

Inclusion of other components (RDA 2)
Bulk δ 13 C values negatively correlate with palm phytolith percentages and correlate positively with Poaceae phytolith percentages, demonstrating the sensitivity of the δ 13 C to turnover between these taxa. Charcoal (>250 μm) concentrations correlate positively with Poaceae phytolith concentrations, and negatively with arboreal phytolith taxa, demonstrating the greater prevalence of fire in grassy ecosystems. Charcoal concentrations also correlate well with soil carbon weight, indicating that charcoal is a primary source of SOM carbon. Clay correlates closely with herbaceous taxa (i.e. Bamboo, Heliconia), cation and δ 13 C data, particularly at Santa Cruz, Tumichucua (early to middle Holocene) and Ottavio (early Holocene) but is negatively correlated with palm and arboreal phytoliths. This highlights a potential role for clay in skewing δ 13 C and herbaceous phytolith records but also demonstrates that palm and arboreal phytoliths are not negatively impacted by soil texture.

Comparison of δ 13 C records (RDA 3)
A site-specific RDA at El Tigre, where bulk-soil and charcoal (>250 µm and 125-250 µm) δ 13 C records are included, reveals a correlation between the bulk SOM and the 125-250 µm charcoal δ 13 C fraction, as well as with arboreal phytoliths. The latter indicates that the SOM δ 13 C record at El Tigre is capturing an arboreal-dominant signal and has not been influenced by either C4 taxa or microbial decomposition. Although this may be true of El Tigre, where the bulk δ 13 C trend is consistent with the charcoal δ 13 C fractions, it is unlikely to apply to other sites, e.g. Tumichucua, where there is no consistent trend among the different δ 13 C fractions, perhaps due to the differences in isotopic composition between different C3 taxa.

Drivers of palaeovegetation shifts and fire activity
Although there are discrepancies between the phytolith and δ 13 C signals from the rainforest sites (Tumichucua, Florida and El Tigre), it is still likely that these ecotones were sensitive to the reduced precipitation of the middle Holocene (e.g. Baker et al., 2001;Bird et al., 2011). The δ 13 C records indicate the expansion of either drought-tolerant dry forest or wooded savannah during the middle Holocene which then shifted to drought-sensitive rainforest in the late Holocene. Phytoliths demonstrate that no wooded savannah expansion occurred, since Poaceae phytolith abundances are below those from modern savannahs (i.e. <50%) (Dickau et al., 2013). However, if the low taxonomic resolution of arboreal phytoliths is masking a shift from rainforest to dry forest during the middle Holocene, as is likely the case, then dry forest may have expanded at the expense of rainforest at all the rainforest sites.
It is unsurprising that these rainforest ecotones underwent a drought-induced contraction during the middle Holocene, since rainforest lacks the drought adaptations common to dry forest and savannah tree species (e.g. thick bark, semideciduousness). Furthermore, rainforest ecotonal resilience would contradict regional lake pollen data which show rainforest contractions during the middle Holocene; i.e. Lago Rogaguado (Brugger et al., 2016), Laguna Oricoré (Carson et al., 2014), Lagunas Bella Vista and Chaplin (Mayle et al., 2000).
The Tajibos record differs from those of the other rainforest sites with its contradictory phytolith and δ 13 C signaturesthe former pointing to continuous forest cover and the latter implying middle Holocene expansion of open savannah. It is unlikely that this signal discrepancy is due to taxonomic limitations of phytolith analysis, since forest and savannah can be clearly differentiated from one another via phytoliths (i.e. Figs. 2-4;Dickau et al., 2013), Poaceae are prolific phytolith producers (Piperno, 2006), and they preserve well in acidic soils (Cabanes and Shahack-Gross, 2015). However, overlapping δ 13 C values between modern savannah and dry forest (Table 2) could potentially hamper interpretations of the record. Edaphic factors might explain why, at the rainforest sites Tumichucua, El Tigre and Florida, there is no middle Holocene ecotonal savannah expansion, even though the pits are in close proximity to the modern rainforest-savannah boundary. The savannah islands near Riberalta and Florida are surrounded by rainforest and clearly co-occur under a climate regime suitable for rainforest, indicating the overriding importance of edaphic controls, rather than climate, at these fine spatial scales. However, this does not appear to be the case for Tajibos, where the highamplitude middle Holocene δ 13 C shift is best explained by ecotonal savannah expansion at the expense of rainforest. The greater ecotonal sensitivity seen at Tajibos could also be due to differences in edaphic conditions, such as soil texture (i.e. lower clay) and/or lower available nutrients (i.e. Ca, Mg, C), which can influence drought tolerance and ecotone position at fine spatial scales (Arruda et al., 2017).
Fire activity at the rainforest sites was relatively low during the middle Holocene drought (i.e. lower half of the profile) and only peaked within the upper 15 cm of the profile, indicating that climate was not a driver of fire activity at these ecotones. Instead, anthropogenic burning associated with land use likely drove changes in fire activity, since: a) there is a spatial gradient of increasing fire activity from the sparsely occupied interfluvial site (El Tigre) to the indigenous settlement/pre-Columbian riverine occupation site (Tumichucua), including a peak in fire activity between 40 and 60 cm during the late Holocene when higher precipitation levels would have likely suppressed natural fires; and b) fire activity at all sites increases in the uppermost 15 cm of soil profiles, corresponding to anthropogenic burning associated with recent land use over the past century.
At Tumichucua and Tajibos, the floristic changes in the top halves of the phytolith profiles (with respect to grass, bamboo, Heliconia and palms) are best explained by increasing pre-Columbian land use during the late Holocene (Arroyo-Kalin and Riris, 2020), rather than middle Holocene drought. Palms are especially useful to indigenous Amazonians and are a core element of agro-forestry in the vicinity of occupation sites (e.g. Clement et al., 2015;Schroth et al., 2003). Over millennia, this pattern of land use would manifest itself in the phytolith record as an enrichment in palm types at the expense of dicotyledonous arboreal types, as seen among the pre-Columbian earthworks in eastern Acre, Brazil (Watling et al., 2017).
The stability of the dry forest-savannah ecotone at the Ottavio site, despite middle Holocene drought, implies that edaphic factors were more important than climate in controlling this type of ecotone over at least the past ca. 6000 years, an inference corroborated by the edaphic patterns associated with present-day dry forest-savannah mosaics, whereby dry forest is confined to calcium-rich soils, and savannahs to nutrient-poor soils (Furley and Ratter, 1988;Ratter, 1992).
In contrast, however, the δ 13 C record at the Santa Cruz site points to a significant middle Holocene drought response at the dry forest-Chaco woodland ecotoneeither dry forest replacement by savannah or Chaco woodland, or instead floristic turnover with expansion of CAM (Crassulacean acid metabolism) taxa (e.g. Cereus cacti). Dry forest-savannah turnover at the Santa Cruz site would be consistent with the fossil pollen record from Laguna Mandioré at the eastern limit of the Chiquitano dry forest, where middle Holocene drought led to localised savannah expansion (Plumpton et al., 2019).
The finding that charcoal frequencies throughout the Ottavio and Santa Cruz profiles are several times higher than those of the rainforest sites demonstrates the greater flammability of dry forests compared with rainforests and challenges the assumption (e.g. Pennington et al., 2009) that dry forests are not fire-adapted ecosystems. We find no evidence that pre-Columbian humans were a dominant control upon fire activity at our dry forest sites, as there is no clear charcoal peak correlative with late Holocene human population expansion (Arroyo-Kalin and Riris, 2020). Despite this, Holocene anthropogenic fire activity in the broader Chiquitano dry forest region cannot be ruled out, as trends in the Laguna La Gaiba charcoal record (Power et al., 2016) from easternmost Bolivia correlate poorly with Holocene climate change, raising the prospect of Holocene fire regimes strongly influenced by pre-Columbian land use.
At Ottavio, the peak in fire activity at 80-85 cm, seen in the 125-250 µm charcoal fraction, likely thinned the canopy, allowing the light-demanding herbs (i.e. Heliconia and Poaceae) to colonise the understorey. However, these fires were not of a sufficient magnitude or frequency to lead to savannah expansion, since Poaceae remains well below the >50% abundance seen in modern savannah assemblages (Dickau et al., 2013). Therefore, the Ottavio soil profile reveals that even those dry forest plant communities within only 200 m of a sharp savannah ecotone were resilient to fires.
At the Santa Cruz dry forest-Chaco ecotone, fire activity was high throughout the middle Holocene dry period and then dipped during the wetter late Holocene, consistent with a climatic control on fire activity at this site. However, it remains unclear whether climate had a corresponding impact on vegetation, given the uncertainty of differentiating dry forest from Chaco woodlands via phytoliths and δ 13 C. If savannah did in fact expand at the expense of dry forest at Santa Cruz, as potentially indicated by the δ 13 C record, then increased fire activity might not reflect a change in dry forest fire regime, but instead a biome shift toward flammable savannah.

Implications for conservation
The remarkable degree of resilience of the Ottavio dry forest ecotone to both drought and fire over the last 6000 yrs, despite its proximity to savannah (only 200 m away), is an encouraging finding from a conservation perspective. It implies that at least some dry forest ecotones will remain resilient to future drought, due to edaphic factors, consistent with pollen data from Laguna Gaiba (Whitney et al., 2011). However, other dry forest ecotones will likely be sensitive to climate change, since potential savannah incursions occurred at the Santa Cruz dry forest-Chaco woodland ecotone, and definitely occurred at the Laguna Mandioré dry forest-savannah ecotone (Plumpton et al., 2019). These palaeoecological insights into the differing responses of dry forest ecotones to past climate change can help conservationists target their limited resources to those dry forest ecotones most at risk from future drought, e.g. by establishing ecosystem corridors (Hannah et al., 2002;Mayle et al., 2007).
The results from rainforest-savannah ecotones in northern and northeastern Bolivia highlight the possibility that future climate change may not just result in the expansion of savannah at the expense of rainforest, as one might predict, but may also lead to ecotonal expansion of dry forest. This would likely reduce the biodiversity in these ecotonal regions as well as the size of the carbon store, since dry forest holds less carbon and is less species-rich than rainforest (Fekete et al., 2017;Pennington et al., 2009). In this scenario, a dynamic conservation approach that allows rainforest taxa to migrate into more suitable climatic zones via protected ecosystem corridors would be most appropriate (Hannah et al., 2002).
Recent fire activity driven by modern land use far exceeds the fires seen throughout most of the Holocene in both dry forest and rainforests, with peak charcoal concentrations occurring in the top 15 cm of soil across all the sites. Interestingly, despite this increase, both rainforest and dry forest sites currently remain intact, suggesting that the background levels of fire that these forests were exposed to throughout the Holocene may have pre-conditioned them to better withstand the elevated fire regimes they are experiencing today, as proposed by Gosling et al. (2021). However, this does not imply that these forest ecotones will remain resilient to fire over the coming century, given that the synergistic impacts of modern land useforest fragmentation and burning will likely interact with predicted enhanced drought and rising temperatures to promote forest die-back (Brando et al., 2012;Prieto-Torres et al., 2021).

Methodological recommendations
The multiproxy analysis used in this study demonstrates differences in the sensitivity of soil phytoliths and δ 13 C to ecological changes through time at forest-savannah ecotones, highlighting the risks that come from interpreting these proxies in isolation. There is a clear need to further research both soil phytolith and δ 13 C records, in terms of taxonomic resolution, surface-sample assemblages and post-depositional biases, to improve their robustness as palaeovegetation proxies.
Improved taxonomic resolution of phytoliths from dicotyledonous arboreal taxa is needed to better resolve ecotonal shifts between rainforest and dry forest, a goal which can only be achieved by expanding modern phytolith reference collections (e.g. Morcote-Ríos et al., 2016;Watling et al., 2020b;Piperno, McMichael, 2020). More phytolith and δ 13 C soil-surface samples are needed from dry forest and Chaco woodlands to improve understanding of modern vegetation proxy relationships and thereby enhance palaeoecological reconstructions of these ecosystems (e.g. Dickau et al., 2013;Watling et al., 2016Watling et al., , 2020.
Radiocarbon dating phytoliths alongside charcoal and SOM from the same stratigraphic horizons may help to improve chronological control and better quantify the extent to which post-depositional factors distort phytolith records. Compound-specific δ 13 C analysis, used alongside bulk δ 13 C analysis, could help quantify the extent to which microbial decomposition distorts bulk fraction values. Macro-charcoal δ 13 C analysis could potentially be used as a novel palaeoclimate proxy free of the influence of microbial decomposition. However, the mixed δ 13 C signals found among our charcoal fractions demonstrate the need for anthracological identifications of both forest and savanna woody taxa.

CONCLUSION
Soil phytolith records from rainforest-savannah ecotones in northern and northeastern lowland Bolivia imply ecotonal stability through the Holocene, whereas δ 13 C data from the same soil samples instead capture a shift to more droughttolerant dry forest or savannah taxa under the drier climatic conditions of the middle Holocene. We attribute these apparently conflicting proxy signals at four of the five sites to the taxonomic limitations of phytolith analysis which are likely masking climate-driven turnover between rainforest and dry forest ecosystemsan inference corroborated by lake fossil pollen records (e.g. Carson et al., 2014;Mayle et al., 2000) which show drought-adapted dry forest and savannah expanding during the middle Holocene at the expense of drought-intolerant rainforest.
The spatial precision afforded by soil profiles captures a higher degree of stability at the dry forest-savannah ecotone 'Ottavio' (located only 200 m from a sharp ecotonal boundary) than could ever be captured with coarse-scale lake pollen records (e.g. Whitney et al., 2011). This remarkable degree of resilience to both middle Holocene drought and fires indicates that edaphic factors play an important role in stabilising these ecotones.
The profiles also demonstrate a persistent history of fire in dry forests over the last 7000 years, a finding supported by the lake charcoal records from eastern Bolivia (Plumpton et al., 2011;Power et al., 2016;Gosling et al., 2021). The finding challenges the assumption made by some ecologists (Pennington et al., 2009) that dry forests are not adapted to fire. The drivers of Holocene fire activity at our sites are unclear. At the Santa Cruz dry forest-Chaco woodland ecotone, climate may have indirectly increased fire activity by forcing a biome shift to pyrophilous savannah. Low charcoal concentrations at the rainforest sites demonstrate that fire has not been a persistent feature in these ecosystems over the Holocene. Increased fire activity during the late Holocene was likely driven by pre-Columbian land use, with charcoal concentrations correlating with proximity to indigenous occupation. Among both rainforest and dry forest sites, recent fire activity due to modern land use far exceeds that observed during the Holocene.
The comparative, multiproxy approach taken in this study demonstrates differing sensitivities of phytolith and δ 13 C proxies to middle to late Holocene forest-savannah ecotonal dynamics, due to limitations in taxonomic resolution and post-depositional factors, highlighting the need for further research into these methods to improve their effectiveness.

Acknowledgments
This research was funded by a NERC 'Scenario' DTP PhD awarded to JH (grant NE/L002566/1). Radiocarbon dates were funded by the UK NERC Radiocarbon Facility (allocation number 2179.0319 to FM). The University of Reading contributed funds toward the fieldwork.
Author contributions-FM, JH and SB conceived and designed the overall project; FM led the fieldwork; JH, DS, EC and VV assisted in collecting the soil samples; DS and EC undertook qualitative floristic surveys; JH undertook the laboratory analyses; SB directed the physical and geochemical analyses of soil properties; JH wrote the initial draft of the paper; JH, FM and SB contributed to the interpretation of the data and subsequent drafts. DS, EC, VV and the Noel Kempff Mercado Natural History Museum provided logistical support during fieldwork. Heather Plumpton, Richard Smith and John Carson assisted with fieldwork. We thank the two reviewers, whose comments improved the manuscript.
Conflict of interest statement-We declare we have no competing interests.

Data availability statement
The data that support the findings of this study are available from the corresponding author upon reasonable request. The raw phytolith data have also been uploaded to Neotoma.