The samples that we analyzed for major, trace, and rare earth element concentrations were originally collected for moisture and density (MAD) measurements onboard the JOIDES Resolution during IODP Expedition 353. Each sample was collected with a 2 cm diameter plastic syringe that fits into the top of a 10 cm3 glass vial, allowing for the vial to be completely filled with sediment (Clemens et al., 2016). The samples were dried in a convective oven at 105∘C±5∘C for 24 h (Clemens et al., 2016). The remaining sample preparation, digestions, and analyses were conducted at Boston University, and a detailed description of the analytical geochemical procedures are presented in Dunlea et al. (2015). In summary, sediment samples were hand-powdered with an agate mortar and pestle. For major elements, sample powders were digested by flux fusion (Murray et al., 2000) and analyzed by inductively coupled plasma-emission spectrometry (ICP-ES). For analysis of trace and rare earth elements, sample powders were dissolved in a heated acid cocktail (HNO3, HCl, and HF, with later additions of HNO3 and H2O2 after samples were dried down) under clean-lab conditions and analyzed by inductively couple plasma-mass spectrometry (ICP-MS). Three separate digestions of a matrix-matched in-house sediment standard were analyzed with each batch and determined precision [(standard deviation)/(average)x100] was ∼2 % of the measured value for each element. The international Standard Reference Material BHVO-2 was analyzed as an unknown with each batch, and the results were consistently found to be accurate within precision for each element.

Analyses of the abundance of total carbon (TC), total inorganic carbon (TIC), total organic carbon (TOC), nitrogen (N), and the δ13C of the TOC component were performed at Woods Hole Oceanographic Institution, and the methods are described in Whiteside et al. (2011). In brief, samples for TOC were weighed into tared silver boats and then acidified to remove carbonates in a closed desiccator for 3 d at 60–65 ∘C over concentrated HCl. All samples were flash-combusted in a Costech 4010 elemental analyzer coupled via a Finnigan-MAT Conflo II interface to a Thermo DELTA V Plus isotope ratio mass spectrometer. Data were recorded and integrated using the Isodat software package. Post-run calculations were performed for blank corrections, quantifications, and final calibrations.

The analyses of compound-specific biomarkers were performed at Brown University (e.g., Daniels et al., 2017). Samples were freeze-dried and lipids were extracted from 3.5 to 4.5 g of sediment using a Dionex 350 accelerated solvent extractor (ASE) with dichloromethane : methanol (9:1 v/v). The fatty acids in the total lipid extract were separated from the neutral lipids using aminopropyl silica gel chromatography, eluting with a dichloromethane : isopropanol solution followed by ether with 5 % acetic acid.

The fatty acids were methylated to form fatty acid methyl ester (FAME) by dissolving the dried down acid fraction in ∼0.3 mL of toluene and ∼1 mL of 5:95 acetyl chloride : methanol with known isotopic composition. Nitrogen replaced the headspace in the vial before they were capped tightly and heated at 60 ∘C for 12 h. Once the reaction was complete, the FAMEs were separated from the water by-products formed during the methylation reaction. Samples received ∼1 mL of synthetic saline solution (50 g NaCl L-1 of double-distilled water) and ∼1 mL of hexane, were vigorously shaken, and were then allowed to rest until the hexane separated from the water. The hexane fraction was pipetted into a new vial, avoiding the water. Another ∼1 mL of hexane was added to the sample, shaken, and pipetted into the new vial. To clean the solution and isolate the fatty acids, samples were run through a second silica gel column, eluting with hexane to remove unwanted acids and then dichloromethane to acquire the clean FAME fraction.

The FAME fraction was analyzed on an Agilent 6890 gas chromatograph with a flame ionization detector (GC-FID). Sample blanks were analyzed with every batch. The isotope ratios of the FAME fraction (δDn-acid and δ13Cn-acid) were measured on a Thermo Finnigan Deltaplus XL isotope ratio mass spectrometer with a HP 6890 gas chromatograph and a high-temperature pyrolysis reactor for sample introduction. For δD, three injections of each sample were analyzed, and for δ13C, two injections of each sample were analyzed. Between every six injections, an in-house lab standard mixture containing known amounts of various n-acids was analyzed to monitor instrument accuracy and precision. The δD and δ13C were corrected for the methyl groups added during derivatization using the following equations: δDcorrected=(2n+2)⋅δDmeasured+123.7⋅3/(2n-1) and δ13Ccorrected=(n+1)⋅δ13Cmeasured+36.52/n, where n is the carbon chain length of the compound, and the methanol added has δD =-123.7 ‰ and δ13C =-36.52 ‰. Analytical uncertainty was calculated by [standard deviation/average] of the injections and is typically less than 3 % for δD and less than 1 % for δ13C. The standard deviations are reported in Table S3 in the Supplement. For every instrument run, samples were analyzed in random order.

We corrected the δD for differences in C3 versus C4 plant physiology (Fig. E1; Smith and Freeman, 2006; Chikaraishi and Naraoka, 2007). First, we calculated the fraction of C3 versus C4 vegetation with δ13CFA for each sample, estimating that C3 and C4 vegetation have a δ13CFA of -37.7±1.8 ‰ and -21.1±1.4 ‰, respectively (Chikaraishi et al., 2004; Ponton et al., 2012). Then, we approximate that C3 plants have a δD that is 30 ‰ lighter than C4 plants (Smith and Freeman, 2006; Chikaraishi and Naraoka, 2007; Ghosh et al., 2017) and use the fraction of C3 versus C4 to correct the δD of each sample for the differences in plant physiology (Fig. E1). The corrected data show the same overall trend as the uncorrected data, except the values are shifted to be more negative and the overall change in δD is steeper (Fig. E1).

The δ13CFA of terrestrial plants is primarily a function of the photosynthetic pathway and isotopic composition of atmospheric CO2 (e.g., Farquhar et al., 1989). In this study, the 5 ‰ increase in δ13CFA is greater than the reconstructed δ13C of atmospheric CO2 (≤1 ‰; Tipple et al., 2010), suggesting that a correction for δ13CCO2 would only slightly adjust our results. Thus, we interpret δ13CFA as reflecting the amount of C4 relative to C3 vegetation produced in the CMZ.

From ∼6 Ma until ∼3.5 Ma, approximately 51 % to 81 % (mean 69 % ±9 % SD) of the vegetation in the CMZ utilized a C4 photosynthetic pathway (Fig. E1). Thus, the environmental threshold for the C4 photosynthetic pathway had already been crossed before the end of the late Miocene. Later in the mid-Pliocene, the reconstruction shows another distinct expansion reaching 64 % to 92 % (mean 81 % ±7 % SD) of C4 vegetation in the early Pleistocene (Fig. 2b). The change in vegetation from ∼3.5 to ∼1.5 Ma suggests multiple steps of C4 expansion in the CMZ, rather than a singular late-Miocene expansion. From ∼1.5 Ma to the present, the average proportion of C4 vegetation decreased and became more variable (58 % to 92 %, mean 76 % ±10 % SD; Fig. 2b), which may reflect the sensitivity of the region to glacial–interglacial variations observed in shorter records from this region (e.g., Zorzi et al., 2015, and references therein).

After correcting for the effects of plant physiology (Sect. 2.4), the amount of precipitation and mixing of different air masses can each vary the hydrogen isotopic composition of leaf wax fatty acids (δDFA; e.g., Eglinton and Eglinton, 2008). The mixing of two air masses with unique δD values was recently observed to drive δD of rainfall in New Delhi, India, but, similar to the amount of precipitation, the relatively depleted δD corresponded to wetter conditions (Hein et al., 2017). Thus, we interpreted the δDFA in this study as a qualitative proxy for aridity or the relative amount of precipitation. The δDFA results suggest an overall drying of the CMZ on the Indian Peninsula over the past 6 Myr. The shorter-term scatter in the δDFA record may reflect higher frequency variations in aridity or rainfall.

In this section, we compare our record of C4 expansion with other compound-specific biomarker records of C4 expansion at sites in the Indian Ocean or the adjacent land and seas. Multiple proxy records document a late-Miocene C4 expansion in the Ganges or Brahmaputra river basins such as the Siwalik Group in northern Pakistan or BoB sites receiving outflow sediment (Fig. 3c, France-Lanord and Derry, 1994; Quade and Cerling, 1995; Cerling et al., 1997; Freeman and Colarusso, 2001; Sanyal et al., 2004; Behrensmeyer et al., 2007; Ghosh et al., 2017). Collectively, the reported timing of C4 expansions in the Himalaya region ranges from 9 to 5 Myr, most commonly 8 to 6 Myr (Behrensmeyer et al., 2007). Rather than a uniform timing, detailed sampling of various deposits around the Siwalik regions shows that C4 vegetation expansion was staggered amongst nearby sub-environments with different local conditions (Ghosh et al., 2017, and references therein). Another biomarker record documents a late-Miocene expansion in a wide continental region north and west of the Arabian Sea (Site 722; Figs. 1a, 3d; Huang et al., 2007). Once C4 vegetation expanded at each of these sites, the records suggest there is generally little or no systematic change in the amount of C4 vegetation after the late Miocene.

In contrast, the CMZ of the Indian Peninsula and a few other records around the Indian Ocean document an expansion of C4 vegetation during the Pliocene (Fig. 3). Marine deposits in the Gulf of Aden originate from northeast Africa and record a late-Miocene C4 expansion, followed by a relapse to predominantly C3 vegetation ∼4.3 Ma and a re-expansion of C4 plants in the Pliocene (Site 231; Figs. 1a, 3e; Feakins et al., 2005, 2013; Liddy et al., 2016). The Pliocene re-expansion is consistent with other records from tropical East Africa (e.g., Levin et al., 2004; Cerling et al., 2011). A C4 expansion in the Pliocene is also observed in northwest Australia (Site 763A; Fig. 3g; Andrae et al., 2018). There is little evidence of significant C4 vegetation prior to the Pliocene, suggesting a relatively late onset of C4 vegetation expansion in northwest Australia (Fig. 3). Additionally, there is evidence that East, South, and Central Asia experienced multiple steps of C4 expansion through the Pliocene (e.g., An et al., 2005; Passey et al., 2009; Zhou et al., 2014; Miao et al., 2017; Koutsodendris et al., 2019). Collectively, a significant regional expansion in the Pliocene, distinctly after the first late-Miocene expansion, is common at least amongst tropical East Africa, northwest Australia, Asia, and the Indian Peninsula.

The adaptations of the C4 photosynthetic pathway provides a competitive advantage over C3 vegetation in environmental conditions with low pCO2, high temperature, high aridity, or extreme seasonality that can lead to frequent floods, droughts, and fires (e.g., Edwards et al., 2010, and references therein). Many studies hypothesize that the global expansion of C4 vegetation in the late Miocene was triggered by pCO2 decreasing below a temperature-dependent threshold (e.g. Polissar et al., 2019; Herbert et al., 2016, and references therein). During the Pliocene, decreasing pCO2 may have also contributed to the expansion of C4 vegetation (Fig. 3a; Pagani et al., 2009; Tripati et al., 2009; Seki et al., 2010; Bartoli et al., 2011). However, the heterogenous regional response of the Pliocene C4 expansion suggests pCO2 cannot be the only driver. Given that the Pliocene C4 expansion is observed in multiple regions surrounding the Indian Ocean that have highly seasonal rainfall, regional hydrodynamics likely played a role in the Pliocene C4 expansion.

In the modern era, a complex interplay of multiple modes of variability dictate the unique seasonal and multi-decadal precipitation patterns in the regions surrounding the Indian Ocean. For example, rainfall on the Indian Peninsula follows quintessential monsoon behavior, but it is also tied to the intertropical convergence zone (ITCZ), the Walker circulation, and the Indian Ocean dipole (IOD; Gadgil, 2003; Wang et al., 2017). The biannual rains of the (semi)arid tropical East Africa are related to the ITCZ, monsoon winds, the sea surface temperature (SST), and the Walker circulation (e.g., Williams and Funk, 2011; Tierney et al., 2015; Yang et al., 2015). Monsoon rains annually quench northern Australia, but the El Niño–Southern Oscillation (ENSO), the Walker circulation, and the amount of Indonesian Throughflow (ITF) better explain the precipitation in other parts of Australia (Ummenhofer et al., 2009, 2011a, b). Previous studies reason that changes in monsoon precipitation or the frequency of fires likely triggered local expansion of C4 vegetation (An et al., 2005; Passey et al., 2009; Zhou et al., 2014; Miao et al., 2017; Andrae et al., 2018). However, the sampling resolution of our million-year study cannot resolve these seasonal, decadal, or centennial rhythms in precipitation, so we examine the underlying processes driving variations in each of these modes of variability.

The physical mechanism of each of these modes of variability can be related back to atmospheric pressure gradients. Monsoons are classically defined as a seasonal wind reversal induced by the pressure gradient caused by differential heating of land and sea (Wang et al., 2017). Broadly, the ITCZ is formed by the convergence and uplift of air masses near the Equator due to differential heating between low and high latitudes and the formation of Hadley cells. The IOD and ENSO are cyclic changes in the zonal gradient across the Indian and Pacific oceans, respectively, and affect the Walker circulation. Variable Walker circulation affects ocean upwelling and thermocline depth, which in turn will influence sea surface temperature. Sea surface temperatures can interact with the atmosphere and reinforce or dampen the atmospheric pressure gradients affiliated with the monsoon, ITCZ, IOD, or ENSO dynamics. Changes in any combination of these atmospheric gradients may have altered the seasonal precipitation in the regions surrounding the Indian Ocean during the Pliocene (Fedorov et al., 2013; Wang et al., 2017).

What could have changed atmospheric pressure gradients in the Pliocene? Indonesian Throughflow was being restricted from 5 to 3 Ma and there is evidence for strengthening SST gradients during the Pliocene (Cane and Molnar, 2001; Wara et al., 2005; Karas et al., 2009; Ford et al., 2012; Zhang et al., 2014; Burls and Fedorov, 2017; Christensen et al., 2017; White and Ravelo, 2020). Changes in the zonal and meridional SST gradients would have affected atmospheric pressure gradients in and surrounding the Indian Ocean and modified seasonal precipitation patterns. At the same time, the onset of Northern Hemisphere glaciation likely also helped reorganize the atmospheric pressure gradients surrounding the Indian Ocean (e.g., Koutsodendris et al., 2019). Before the intensification of glaciation at 2.7 Ma, there is evidence of four early local glaciation events at 4.9–4.8, 4.0, 3.6, and 3.3 Ma (De Schepper et al., 2014). The latter two events are within the age model uncertainty of 3.5 Myr when C4 vegetation expands in various regions around the Indian Ocean. Thus, it is possible that the early local growth of ice, not the intensification of glaciation, altered the Siberian high or other atmospheric pressure gradients near the Indian Ocean, changing cyclic precipitation patterns that lead to a regional expansion in C4 vegetation.

On the Indian Peninsula, C4 vegetation expands gradually from 3.5 Ma until the mid-Pleistocene when the range of variability in C4 vegetation increases. The timing of the change coincides within uncertainty with the mid-Pleistocene transition (∼1.3 Ma) where the glacial periods transitioned from 41 to 100 kyr cycles. Although the trigger for the switch in cyclicity is unknown, the processes responsible for the change or the ice sheets themselves may have affected atmospheric gradients and the modes of climate variability that altered the dynamics of C4 expansion on the Indian Peninsula. The sampling resolution from 1.5 to 0 Ma in our study is too low to resolve the change in cyclicity. However, previous studies documenting variations in C4 vegetation relative to C3 since the last glacial maximum demonstrate the sensitivity of plant communities to recent glacial–interglacial cycles (Galy et al., 2008; Phillips et al., 2014; Zorzi et al., 2015, and references therein).