Summer sea-ice variability on the Antarctic margin during the last glacial period reconstructed from snow petrel (Pagodroma nivea) stomach-oil deposits

Antarctic sea ice is a critical component of the climate system, affecting a range of physical and biogeochemical 15 feedbacks, and supporting unique ecosystems. During the last glacial stage, Antarctic sea ice was more extensive than today, but uncertainties in geological (marine sediments), glaciological (ice core), and climate model reconstructions of past sea-ice extent continue to limit our understanding of its role in the Earth system. Here, we present a novel archive of past sea-ice environments from regurgitated stomach oils of snow petrels (Pagodroma nivea), preserved at nesting sites in Dronning Maud Land, Antarctica. We show that by combining information from fatty acid distributions and their stable carbon isotope 20 ratios with measurements of bulk carbon and nitrogen stable isotopes and trace metal data, it is possible to reconstruct changing snow petrel diet within Marine Isotope Stage 2 (ca. 22.6-28.8 cal. kyr BP). We show that, as today, a mixed diet of krill and fish characterises much of the record. However, between 25.7-26.8 cal. kyr BP signals of krill almost disappear. By linking dietary signals in the stomach-oil deposits to modern feeding habits and foraging ranges, we infer the use by snow petrels of open water habitats (‘polynyas’) in the sea ice during our interval of study. The periods when consumption of krill 25 was reduced are interpreted to correspond to the opening of polynyas over the continental shelf, which became the preferred foraging habitat. Our results challenge hypotheses that the development of extensive, thick, multi-year sea-ice close to the continent was a key driver of positive sea ice-climate feedbacks during glacial stages, and highlight the potential of stomachoil deposits as a palaeo-environmental archive of Southern Ocean conditions. https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c © Author(s) 2021. CC BY 4.0 License.


Introduction 30
Antarctic sea ice is globally important: large seasonal fluctuations in its spatial extent influence planetary albedo, oceanatmosphere exchanges of heat and climatically-active gases including CO2, and the formation of intermediate and deep water masses which create the world's largest sink of heat and carbon (e.g. Ackley et al., 2015;Delille et al., 2014;Arrigo et al., 2008;Frölicher et al., 2011). Antarctic sea ice extent is projected to decline by 16-67% by 2100 depending on greenhouse gas emission scenarios, but with low confidence due to a wide range of model responses (Collins et al., 2013). 35 The geological record offers an opportunity to set the relatively short (~50 yr) instrumental observations of sea ice into a longer-term context. During the Last Glacial Maximum (LGM, ~19-23 ka) (Tierney et al., 2020), the development of thick, multi-year and more extensive sea-ice is inferred from the disappearance of, or diagnostic changes in, assemblages of marine microfossils from many Southern Ocean sediment cores (e.g. Hillenbrand and Cortese, 2006;Grobe and Mackensen, 1993;40 Bonn et al., 1998;Lucchi et al., 2002;Gersonde et al., 2005;Benz et al., 2016;Collins et al., 2012;Allen et al., 2011). An expanded 'sea-ice cap' is proposed to have increased deep-ocean storage of CO2, by limiting the air-sea gas exchange and by enhancing CO2 export through increased Antarctic Bottom Water production (e.g. Stephens and Keeling, 2000;Ferrari et al., 2014).

45
Antarctic sea-ice environments today are dynamic and complex at a range of temporal and spatial scales (Parkinson, 2019;Turner et al., 2020), and include open waters within the sea-ice pack ('polynyas') which span a large size range (1000-400,000 km2 (Arrigo and van Dijken, 2003), are poorly represented in models (Mohrmann et al., 2021), and yet impact ocean circulation, sea ice formation and air-sea gas exchange (Mohrmann et al., 2021;Morales Maqueda et al., 2004). A similarly complex picture has emerged for the Last Glacial period: in the Atlantic sector of the Southern Ocean, spanning the 50 Weddell, Lazarev and Scotia Seas ( Fig. 1) there is fragmentary empirical evidence for polynyas during and before the LGM, as detailed by intervals of high productivity in marine sediment cores, and occupation of nesting sites by seabirds that require open water within their foraging range (e.g. Smith et al., 2010;Sprenk et al., 2014;Mackensen et al., 1989;Thatje et al., 2008;Berg et al., 2019). Millennial-scale variability in the extent of the seasonal sea-ice zone has also been described (Collins et al., 2012;Gersonde et al., 2003;Rae et al., 2018), but large (10-20%) uncertainties remain in both models and 55 geological datasets (Collins et al., 2012;Roche et al., 2012;Gersonde et al., 2003). Maximum summer sea-ice extent was likely reached at 30-22 ka (Collins et al., 2012;Gersonde et al., 2003;Allen et al., 2011;Xiao et al., 2016), raising the possibility that Southern Ocean sea-ice / climate feedbacks were more important before the global LGM (Allen et al., 2011;Xiao et al., 2016).

70
Here, we investigate changes in the sea-ice environment during the time of proposed maximum sea-ice extent (30-22 ka) by analysing a sequence of preserved stomach oils of snow petrels (Pagodroma nivea). Although sometimes referred to as 'Antarctic mumiyo' (e.g. Berg et al., 2019;Thor and Low, 2011), this is a misnomer because its biological origin is so different from the original use of 'mumiyo' which was used to describe an organic, tar-like substance of unknown origin, 75 found in high altitude rocks and caves especially in Asia (Hiller et al., 1988 and references therein; Aiello et al., 2011). We therefore refer to Antarctic 'stomach-oil deposits' here.
Here, we undertake elemental scans, organic geochemistry and stable isotope analysis to investigate snow petrel diet during the Last Glacial stage from a stomach-oil deposit collected at Lake Untersee in central Dronning Maud Land (DML). Since snow petrels have a restricted foraging range during the breeding season (Delord et al., 2016), our analyses enable us to 95 reconstruct changes in their diet and foraging habitat relatively close to the Antarctic margin ( Fig. 1). In turn, we hypothesise that the biochemistry of the stomach-oil deposits provides diagnostic signatures of snow petrels foraging in sea-ice over or beyond the continental shelf, and in polynyas, offering novel insights into the evolution of sea-ice environments during the last glacial stage.

Untersee Oasis sequence WMM7 and its regional context
The stomach-oil deposit WMM7 was collected during the GeoMaud expedition (1995/1996), from the Untersee Oasis (71° 21.6' S, 13° 18.96' E) in DML ( Fig. 1) (Wand and Hermichen, 2005). The deposit was retrieved from under a boulder on a steep slope west of Lake Untersee, c. 320 m above the lake surface, at an elevation of 880 m above sea level. WMM7 was kept in dark, cold storage (4°C, 60% humidity) at the Alfred Wegener Institute, Germany, until analysis. The sample had a 105 waxy consistency, so to preserve its internal structure it was frozen immediately before slicing and sub-sampling at Durham University. A central slab of WMM7 was sectioned for non-destructive, high-resolution X-ray fluorescence (XRF) analysis (40 mm thick, 50 mm wide and 150 mm long, marked on Fig. 2a). Samples for stable isotope and lipid analysis were taken from adjacent to the XRF slab (Fig. 2a).

110
Before sub-sampling, WMM7 weighed c. 1 kg, and was 155-194 mm thick, 144 mm wide, and 120 mm deep (Fig. 2a). It had an irregular, mammillated outer surface, but was characterised internally by mm-scale laminae that were traced through the deposit. The laminae visibly slope away from the centre of the deposit, showing that the deposit progressively draped https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License. over the underlying rock. Three distinct units were observed in the sequence during sampling (Fig. 2b): the oldest part of the sequence, (160-110 mm depth, Unit III) was characterised by laminated, yellow-brown deposits, overlain by a zone of 115 relatively dark, brown-black deposits with more poorly defined laminae (Unit II). The Unit II/I transition was gradual (80-70 mm), and Unit I contained yellow-brown deposits with sub-mm scale laminae (Unit I). No hiatuses were visible in the stratigraphy. 120 Figure 2: Stomach-oil deposit WMM7 stratigraphy and age-depth model. (a.) cross-section of the section used for sub-sampling, scaled with panels (b) and (c). Note the clear laminae which slope to the right away from the main slab. A zone of darker brown and less well defined laminae is found in the central section (Unit "II" in panel (b)). A 5 cm wide slab was cut for XRF analysis. Data from the centre-line scan is shown in Fig. 3 and used for statistical analysis. The results of additional scans to the left and right of centre, and from two new cuts (3cm left, 3 cm right; see arrows) and shown in Figure B1. Biomarker and bulk stable 125 isotope samples were taken immediately to the left of the XRF slab. (b.) three-unit stratigraphy and age-depth model constrained by 6 bulk radiocarbon dates (see Table 1 for 14 C calibration). (c.) accumulation rate between age control points. Linear interpolation was applied to generate an age-depth model for all sampling points. https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License.

Radiocarbon analysis
Previous 14 C analysis of the top of WMM7 provided a age constraint of 21,551 ± 110 yr . To establish an 130 age-depth model, six additional bulk samples were dated (Table 1), and linear interpolation was applied between dating points. Each sample was digested in 2M HCl (80°C, 8 hours), washed free from mineral acid with deionised water, then dried. Samples were graphitized using an automated graphitization system (Rethermeyer et al., 2019) and analysed for 14 C by accelerator mass spectrometry (AMS) at CologneAMS, Germany. Radiocarbon ages were converted to calendar ages (Table 1) using the MARINE13 radiocarbon age calibration (Reimer et al., 2013) corrected for the Southern Ocean marine 135 reservoir effect by applying a R of 880 yr ± 100 yr from pre-bomb 14 C ages at Hope Bay in the western Weddell Sea (Björck et al., 1991;Sterken et al., 2012). We apply MARINE13 here since MARINE20 is not recommended for polar regions with variable sea-ice extent (Heaton et al., 2020). A comparison of the two approaches for WMM7 yields calibrated ages which are within error (Table A11) and thus do not affect our interpretations or conclusions. However, we acknowledge that the application of a constant, modern R to a region affected by greater sea-ice extent during MIS 2 leads to larger 140 uncertainty and potential bias in our calendar ages than indicated by the results in Table 1 (Heaton et al., 2020).

XRF analysis
XRF analysis was performed using the ITRAX core scanner at the National Oceanography Centre, Southampton, U.K. (Croudace et al., 2006). A molybdenum X-ray source was used (45 kV, 40 mA), at a step size of 200 µm and exposure time of 400 secs per increment. To test for potential contamination during sub-sampling, and to explore internal consistency of the recovered signals, a further 5 profiles were measured on the central slab (Fig. 2a, Figure B1). Multiple XRF scans confirmed 150 that the major patterns outlined below are consistently recorded in WMM7, with differences accounted for by changes in the orientation and spatial continuity of the laminae ( Figure B1).

Bulk organic matter elemental composition and stable isotope analysis
Directly adjacent to the left margin of the XRF slab, 15 contiguous 10 mm sub-samples were taken by scalpel for bulk stable isotope analysis (Fig. 2a). Carbon and nitrogen stable isotope analyses were performed using a Costech Elemental Analyser 155 (ECS 4010) connected to a Thermo Scientific Delta V Advantage isotope ratio mass spectrometer. Carbon isotope ratios are corrected for 17O and reported in standard delta (δ) notation in per mil (‰) relative to Vienna Pee Dee Belemnite (VPDB).

Biomarker distributions and stable carbon isotope analysis
A 2mm stainless steel dermal punch was used to extract samples for biomarker analysis, from the same line as the bulk 165 stable isotope samples (Fig. 2a), including triplicates at 1 cm depth and 8 cm depth. Lipids were extracted from 0.03-0.3 g of each sample using repeated ultra-sonication (3 x 10 mins) in 10 ml dichloromethane/methanol (3:1), following addition of two internal standards of known concentration (androstanol, hexatriacontane). Ultrasonic extraction yielded between 16-43 mg total lipid extract.

170
An aliquot of the total lipid extract was dissolved in acetone, and analysed by UV-Vis spectrophotometry using a Dionex HPLC Quaternary pump and photo-diode array detector (McClymont et al., 2007). A grass standard and blank acetone were injected regularly to monitor potential instrument drift. The relative absorbance at wavelengths characteristic of chlorophyll derivatives ('chlorins') and potential carotenoid derivatives, at 410, 435 and 665 nm (Jeffrey et al., 1997) was calculated for each sample: 175 Where P = relative magnitude of absorbance for a given wavelength (abs. g-1); A = integrated area for that wavelength, averaged over three repeat measurements; DF = dilution factor i.e. the aliquot of the total sample injected; M = mass of material which was extracted. To take into account the potential influence of variable organic matter deposition or preservation, we also normalised all photosynthetic pigment absorbance to TOC content (generated from section 2.4). There 180 may be multiple sources of our target pigments (Jeffrey et al., 1997), so we refer to them as P410, P435 and P665 respectively, to identify the trends in absorbance for each specific wavelength. The remaining extracts were saponified using 2 ml KOH (8%) in methanol (95%) and heated for 2 hours at 70°C. Neutral lipids were extracted using hexane; the remaining extracts were acidified using 2 M HCl, and the fatty acids were extracted 185 using hexane. An internal standard of known concentration (heptadecanoic acid, 0.4 mg ml-1) was added to the fatty acid fractions, before generating fatty acid methyl esters (FAMEs) by methylating with 3 ml methanol: HCl (95: 5) for 12 hours at 70°C, then allowing to cool to room to temperature. After adding 4 ml of DCM-rinsed H2O to each sample, FAMEs were recovered sequentially using hexane followed by hexane: dichloromethane (4:1) and pooled. FAME fractions were then taken to dryness in a stream of N2. The isotopic value of the methanol was determined through methylation of a phthalic acid 190 with known isotopic value (Lee et al., 2017).
The FAMEs were identified and quantified with a Thermo Trace 1310 gas chromatograph linked to an ISQ LT single quadrupole mass spectrometer (GC-MS). Chromatographic separation was performed with a Restek Famewax 30m×0.25mm×0.25µm column. Sample extracts were injected (0.8µL) in CT split mode (80:1 ratio) into a PTV injector at 195 230°C with a constant helium carrier gas flowrate of 1.5ml/min. The oven temperature was initially set to 100°C for 3.0 min, then ramped at 2°C/min to 230°C, and held at 230°C for 10 min. The transfer line temperature was set to 230°C and the ion volume temperature to 220°C. A mass range of 38 to 600 m/z was scanned every 0.5 seconds giving at least 20 data points per compound peak. Samples were identified and quantified by comparison to a Supelco 37 component ME mix (CRM47885, Sigma-Aldrich) with peak area ratios calculated with reference to the peak area of the internal standard. 200 Spectral confirmation was performed using a NIST EI reference library.
The carbon isotopic compositions ( 13 C) of individual saturated FAMEs were determined on 8 samples (22-132 mm depth) using a Thermo GC-C-IRMS system at the Department of Geography, Durham University. All samples were run in duplicate. The system was composed of a Trace 1310 GC coupled to a Thermo Delta V Plus via a GC IsoLink II and a 205 Restek Famewax 30m×0.25mm×0.25µm column. Samples were injected (2µL) in splitless mode into a S/SL injector set to 240°C with a helium carrier gas flowrate of 1.5ml/min. The oven temperature was initially set to 50°C for 1 min, then ramped to 100°C at 10°C/min, then ramped to 240° at 3°C/min, and finally held for 10 min at 240°C. The alumina (with CuO,NiO and Pt wires) combustion reactor was operated at 1000°C and conditioned with oxygen each day immediately before use. CO2 reference gas pulses were introduced at the start and end of each chromatogram to provide an isotope ratio 210 reference point and to check the system stability during the run. All the quantified FAME peaks in the selected carbon range were baseline resolved apart from the cis and trans isomers of C18:1. A FAME standard (CRM47885 -Supelco 37 component mix, Sigma-aldrich) was run with each batch of isotope reference standards to confirm the retention time of the FAME peaks. Individual FAME isotope ratio values were corrected using a multipoint linear normalization of a C14-C20 FAME reference material (F8-3 standard provided by A. Schimmelmann, Indiana University, Bloomington). Reference standard 215 FAMEs from C14-C20 were used to generate the normalization curve, covering  13 C values from −23.24 to −30.92 ‰.
gave an r2 value of at least 0.995 for the normalization plot. The concentration of the F8-3 FAME standard used for the linear normalization was adjusted to obtain reference amplifier intensities within this range (1000 to 6000 mV). Each sample was then diluted and pre-run to determine the optimum solvent volume required to fit within the amplifier signal range of the 220 reference standards. The long term pooled standard deviations of the ME F8-3 reference compounds were all <0.35‰ (ranging from 0.21 to 0.34‰, n=21). Fatty acid  13 C was calculated through mass balance corrections of the measured FAME  13 C and known  13 C of the added methyl group (Lee et al., 2017) and reported relative to Vienna PeeDee Belemnite (V-PDB).

Statistical analysis 225
Principal component analysis (PCA) was performed to investigate the potential for co-evolving relationships in the stomach oil geochemistry, using the PAST3 software (Hammer et al., 2001). We focussed on the XRF data for PCA since it was analysed at the highest resolution and with the highest number of variables, and used the variance-covariance matrix since all variables were measured in the same units (i.e. counts). Elements which consistently recorded counts below 500 were excluded from the analysis. 230 Cluster analysis was performed to identify units of similar geochemical composition using the rioja package in R 3.6.0 (Juggins, 2020), whereby a hierarchical clustering is performed, constrained by sample order, and to a broken-stick model of a random distribution of zones within a sequence (Bennett, 1996). Due to the different sampling resolutions of the geochemical methods employed here, we re-sampled the XRF data by averaging element counts across the same depth-235 window sampled by each of the discrete geochemical measurements (10 mm diameter). This re-sampling approach does not alter the main signals of the principal components, although additional clusters are identified using the original XRF data ( Figure C1). We present the re-sampled XRF PCA here as an independent measure of geochemical change that is comparable to the resolution of the lipid distribution and stable isotope signatures.

Elemental composition (XRF) 245
Deposit WMM7 was dominated by Fe, Ca, Cu and Ti. Elements likely to be indicative of minerogenic inputs (e.g. Ti, Al, K, Ca), for example from windblown particulates, had low signals overall ( Figure D1). There were no horizons of elevated minerogenic inputs which might be expected during a hiatus in stomach oil accumulation. Elements with both a biogenic and minerogenic (e.g. Cu, Fe) or principally biogenic source (e.g. As, Zn) (Huang et al., 2009;Liu et al., 2013) varied both at the mm-scale (likely reflecting individual laminae) and with depth in the deposit ( Figure D1). To account for the influence of 250 minerogenic contributions to the elemental composition of WMM7, all data were normalised to Ti (Fig. 3, Figure E1). There were no clear down-core trends in Fe/Ti and Si/Ti (Fig. 3a,b). Cu/Ti was high but showed a long-term decrease from the base of the deposit (28.8 ka) to 25.7 ka (Units III and II). High Cu/Ti was recorded between 25.7-24.0 ka (Unit I), then values decreased to the top of the deposit (Fig. 3c). In contrast, Br/Ti and S/Ti were low from 28.8 ka to 26.8 ka (Unit III), then elevated between 26.8-25.7 ka (Unit II; Fig. 3d,e). A gradual increase in Br/Ti and S/Ti was then recorded until a 255 second maximum at 23.9 ka (Unit I); both Br and S then decreased to 22.9 ka. A dominant biogenic source for the trends in Si, Cu, Br and S is inferred for WMM7, noting that the patterns in Br/Ti and S/Ti were different to those of Cu/Ti, and may include a sea-salt origin given some similarities to Cl/Ti (not shown; r2=0.36). 265 The PCA of the XRF data confirms that there were two main patterns in the elemental composition of WMM7 (Fig. 3f,g; Table 2). PC1, which accounted for 93% of the variance, and driven by positive loading from Fe, Cu and Ca, which could have both minerogenic and biogenic origins. PC2 accounted for 6% of the variance, driven by Cu (positive loading), Ca and Fe (negative loading), but PC2 was not a statistically significant factor using the broken-stick threshold in PAST3. Peaks in 270 PC1 occurred before 27.0 ka (Unit III), and between at 25.5 and 24.75 ka (lower Unit I). PC1 minima occurred in Unit II (26.75-26.0 ka) when high Br/Ti and S/Ti confirmed enhanced organic matter inputs. In contrast, PC2 broadly followed the Cu/Ti trend, recording declining values from 28.25-26.0 ka (Units III and II), and peak values at 25.25 ka (lower Unit Il, Fig.   3). Three clusters were identified in the re-sampled XRF data (Fig. 3i), broadly aligned to Units I-III.
The major pigments in WMM7 were P410 and P435 (Fig. 4h), consistent with chlorophyll derivatives ("chlorins"; (Harris 305 and Maxwell, 1995)) and some carotenoid sources (Jeffrey et al., 1997). P665 nm, recording exclusively chlorins, was very low (3 orders of magnitude below P410, Figure F1). P410 showed a sustained maximum between 28.7-25.5 (extending across the Unit II/I boundary, Fig. 4h). P435 remained low from the base of the sequence until a small peak at 25.5 ka. A rapid decrease in P410 and P435 resulted in pigment minima at 25.1 ka. From 24.3-22.6 ka, P410 oscillated at intermediate values and there is a slight increase in P435 (Fig. 4h). 310 Table 3 Examples of snow petrel prey biochemistry, and oceanographic conditions related to snow petrel diet and prey distributions. Snow petrels forage in close association with the sea ice, either at the sea-ice edge or in leads or polynyas within the sea ice pack. The prey biochemistry information is used as a framework to interpret the chemical signatures recorded in stomachoil deposit WMM7, including our use of fatty acid ratios to assess relative contributions of prey (Figs 4 and 5).

Discussion 345
Stomach-oil deposit WMM7 is characterised by biochemical variability which can be attributed to changing biogenic composition through time. In this section, we first evaluate the likely sources of this variation, infer changes in snow petrel diet through time, and investigate how the sea-ice environment may have varied. By doing this we aim to provide greater insight both into the biochemistry of the stomach-oil deposit, and into critically evaluating its use as a palaeo-environmental archive. 350

Contributions to stomach-oil biochemistry
The stomach oil regurgitated by snow petrels at nesting sites provide an integrated archive of diet during the most recent foraging trip in the summer breeding season. These trips may be for several days (Barbraud et al., 1999). During this time, the snow petrel concentrates the lipid components of their prey into stomach oil; this provides an energy-rich food source for their chicks (Warham, 1977;Watts and Warham, 1976), and can be spat in defence against predators, or against other snow 355 petrels in disputes over access to suitable crevices for nesting. Unlike muscle, feather or adipose tissues, which involve biosynthesis (e.g. Rau et al., 1992), the basic biochemical composition of the prey is not altered during stomach oil formation . Thus, the stomach-oil deposits provide a window into the biochemistry of the prey consumed by snow petrels in open waters at the margins of, or within, the sea-ice zone.

360
A contribution of snow petrel guano to stomach-oil deposits has been indicated previously Hiller et al., 1988). However, WMM7 does not show the negative relationship between C/N and other elements (Cl, P, S) which would have been expected if there was a strong residual signature of guano contributions to the deposits ). The C/N ratios are consistent with particulate organic carbon (~10) in modern sea ice (Henley et al., 2012) and fish collected in the Weddell and Lazarev Seas (~3-11) (Rau et al., 1992;Friedrich and Hagen, 1994). We suggest that any contribution of 365 guano to WMM7 is lower than for other published sequences derived from DML stomach-oil deposits , perhaps due to local topography (including aspect and microclimate), the distance from the nest that the sample was taken, or the history of occupation.
We did not observe formation of K-and C-bearing phosphates by weathering ; rather, the highest P counts 370 were recorded in Unit II where Ca counts were lower and there was no change in K ( Figure B1). Progressive oxidation of https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License. organic matter was not evident: no systematic trends in labile pigment abundances (Fig. 4h) nor fatty acid  13 C were observed (Fig. 5). Furthermore, bulk  13 C are negatively correlated to C/N, as observed today in Antarctic fish (and other predators), reflecting the variable contributions of 13 C-depleted and low N lipids to the bulk signal (Rau et al., 1992;Friedrich and Hagen, 1994;Cherel et al., 2011). Again, this difference in weathering signal between deposits may reflect 375 local factors at the nesting sites (discussed above).
As observed in other glacial-stage stomach-oil deposits , WMM7 differs from some modern snow petrel stomach oils (e.g. Warham et al., 1976) by substantially lower contributions of C18:1 relative to C16:0 and C14:0. The lower contributions of polyunsaturated (<1%) and monounsaturated fatty acids (e.g. C18:1 9 ± 2%, C16:1 2.6 ± 0.3%) than in fresh 380 stomach oil of other procellariiform seabirds (Connan et al., 2007) including snow petrels Watts and Warham, 1976) could indicate post-depositional oxidation . However, similar fatty acid contributions to WMM7 have been recorded in a late Holocene DML stomach-oil deposit (Aiello et al., 2011), in prey of snow petrels (Cripps et al., 1999) and in stomach oils from other Procellariiformes (Wang et al., 2007). As previously noted , the samples of fresh stomach oils Watts and Warham, 1976) were from snow petrels foraging 385 in the Ross Sea, where prey availability, and hence stomach oil biochemistry, may also be different. We therefore suggest that the fatty acid signatures in WMM7 primarily signal a dietary intake, rather than variable preservation.
Our interpretation of C14:0 as an indicator of krill inputs to snow petrel diet is broadly supported by elevated Cu (measured as Cu/Ti) (Rainbow, 1989;Palmer Locarnini and Presley, 1995;Liu et al., 2013)  C16:0/C14:0 and C18:0/C14:0) (Fig. 4). Thus, we here identify two independent measures of the relative contributions of krill which can be applied to WMM7 to investigate broad-scale changes in snow petrel diet through time.

Snow petrel diet 28.8-22.6 ka
The variation in biogenic contributions to stomach-oil deposit WMM7 reveal changing snow petrel diet at decadal (XRF data), centennial and millennial timescales, for the early part of Marine Isotope Stage (MIS) 2, which includes the start of the 410 LGM (28.8-22.6 ka). We note that although each stomach oil regurgitation provides a snapshot of snow petrel diet during the summer breeding season (November to February), prey biochemistry (and thus stomach oil composition) may also provide information over varying timescales, reflecting lifecycle and tissue turnover rates of the fish, krill and squid consumed by the snow petrel.

415
Our multi-proxy analysis confirmed that three zones could be identified. Between 28.8-26.8 ka (~Unit III) elevated Cu/Ti and C14:0 contributions (low C16:0/C14:0 and C18:0/C14:0) identified krill as an important component of snow petrel diet, but likely decreasing through time. Although offset by ~2.0‰, supporting different sources, both  13 C14:0 and  13 C16:0 decreased across this time interval, suggesting either a common environmental (prey habitat) signal or that krill contributed to both C14:0 and C16:0. In contrast, the much higher (and increasing)  13 C18:0 and  13 C18:1 confirmed a different prey source and/or 420 habitat signal contributing to the C18 fatty acids. Thus, the prey other than krill, which were most likely to be fish, occupied a different foraging habitat (see Sect. 4.3).
At 26.8 ka (Unit II) there was a shift in the dominant contributions to the snow petrel stomach oils. Low Cu/Ti and increasing C16:0/C14:0 and C18:0/C14:0 indicate a prolonged (~1100 yr) interval where krill was not a major component of snow 425 petrel diet (Figs 3,4). Reduced minerogenic inputs (Fig. 3f) and increased marine organic matter (Figs 4h,i) characterise Unit II: low PC1, elevated S, Br (e.g. Leri et al., 2010), chlorins, and C18:1 (Jónasdóttir, 2019;Cripps et al., 1999). Although this biochemical signature suggests a dominant phytoplankton fingerprint in snow petrel diet between 26.8-24.7 ka, phytoplankton are too small to be consumed directly by seabirds. The elevated chlorins (Fig. 4h) seem more likely to indicate the snow petrels consumed prey with a phytoplankton-dominated diet. Decreasing  15 Nbulk in Unit II may also reflect 430 a shorter food-chain, although the shift is small (<1‰, Fig. 5). Preservation of intact phytoplankton in stomach oils is also feasible through secondary ingestion (i.e. in the stomachs of snow petrel prey), since undigested algal cells have been reported in penguin guano (Mychra and Tatur, 1991) and diatoms have been isolated in P. antarctica larvae (Vallet et al., 2011) and other stomach-oil deposits .

435
The similarity in the trends between  13 C16:0 and  13 C18:1 (Fig. 5c)  through their consumption of copepods, squid or fish, which can occur with minimal alteration (e.g. Lee et al., 1971). Elevated (~20%) C18:1 through Units I and II supports an increased contribution of fish to the snow petrel diet between 26.8-25.7 ka, but identifying the particular fish species is more challenging. Both P. antarctica (Mayzaud et al., 2011) and several 440 species of myctophid fish including E. antarctica (Imber, 1976;Raclot et al., 1998) contain C18:1, and their  13 C and  15 N values suggest that P. antarctica and E. antarctica have overlapping trophic niches (Rau et al., 1992). High abundances of C16:0 in myctophids (Mayzaud et al., 2011) may account for the similarity in  13 C16:0 and  13 C18:1 in Unit II. Although the main myctophid prey today are euphausiids (including krill, (Saunders et al., 2019)), this diet is not consistent with the low Cu/Ti and low C14:0 contributions to WMM7 in Unit II, nor with the >4‰ offset between  13 C14:0 and  13 C18:1. Alternatively, 445 an increased contribution from nototheniid fish (e.g. P. antarctica), the diet of which includes copepods as well as krill, might explain the apparently elevated phytoplankton signature and the reduced contribution of krill to snow petrel diet throughout Unit II.
From 25.7-24.2 ka (lower Unit I), krill returned as an important component (high Cu/Ti, higher C14:0 contributions), although 450 the positive relationship between  13 C16:0 and  13 C18:1 suggests that fish was also important in the diet. Reduced input of marine organic matter was apparent in Br, S and chlorins (Figs 3 and 4). Fatty acids  13 C showed higher variability than in Units II and III. From 24.2 ka to the top of the deposit (22.6 ka; Upper Unit I), increasing minerogenic contributions may explain the simultaneous decrease in Cu/Ti, Br/Ti and S/Ti, which contrasts with the anti-phase behaviour between Cu and Br or S discussed previously, and for the Cu/Ti minima whilst C14:0 inputs increase slightly. Fatty acids  13 C also fluctuate in 455 parallel from 24.2-23.5 ka, suggesting there may be a strong baseline (i.e. primary producer) control over the stable isotope values.
We note that the changes in relative contributions of krill and fish in the diet that we inferred for WMM7 are not reflected in the range of  15 Nbulk (~1.2‰), which lies below modern trophic level offsets of >2 ‰ (e.g. Seyboth et al., 2018) including 460 some krill-fish offsets of ~5-7 ‰ in the Southern Ocean (e.g. Polito et al., 2011), and suggests minimal changes to snow petrel trophic level between 28.8-22.6 ka. However,  15 Nbulk reflects  15 N in prey tissues which can themselves vary over space and time (Quillfeldt and Masello, 2020), so that trophic values can only be calculated if baseline (primary producer)  15 N is known (Post, 2002). Since baseline  15 N is poorly constrained for our interval of study, and since snow petrel prey integrate  15 N across different temporal and spatial scales, it seems likely that the small range of  15 Nbulk in WMM7 reflects 465 either baseline  15 Nbulk or mixing of multiple signals over time, which we are unable to disentangle with our analyses.
In summary, the stomach-oil deposit WMM7 indicates several changes to snow petrel diet between 28. phytoplankton fingerprint during this time also suggests that the fish themselves had a reduced krill contribution to their diet, highlighting a substantial change to the food chain in the foraging habitat of the snow petrels. In the following section, we consider the likely palaeo-environmental drivers and implications of these signals.

Evolution of sea-ice environments (28.8-22.6 ka)
Seabird diets vary over time and space, depending on the availability of prey species, in turn reflecting environmental 475 conditions within their foraging range (Mills et al., 2020;Mills et al., 2021). For example, it has been proposed that myctophid fish become more important to seabird diet when krill availability is low, during winter when sea ice is more extensive and/or in response to temporal and spatial changes in summer production (Nicol, 2006;Watanuki and Thiebot, 2018). Prey biochemistry may also reveal environmental information:  13 C decreases with increasing latitude (Francois et al., 1993;Trull and Armand, 2001) or between coastal and offshore waters (Trull and Armand, 2001), whereas  13 C 480 increases within the sea ice during spring melt (Dunbar and Leventer, 1992). These differences are reflected in the  13 C of crustaceans, fish and squid in these habitats, and in turn in their predators including seabirds (Cherel et al., 2011;Jaeger and Cherel, 2011;Delord et al., 2016;Phillips et al., 2009).
The presence of multiple stomach-oil deposits in DML during MIS 3 and the early LGM  confirms that 485 there must have been accessible prey within range of the breeding snow petrels at Lake Untersee during the spring and summer. WMM7 has a very stable accumulation rate (Fig. 2) and we found no evidence for hiatuses, suggesting that the site was not affected by long periods of nest-site abandonment related to unfavourable climatic conditions (e.g. snow cover over the site, perennial sea ice preventing access to prey within foraging range) or local factors (e.g. burial by rock) (Olivier and Wotherspoon, 2006;Olivier et al., 2005). We consider it unlikely that the snow petrels were foraging at an oscillating 490 spring/summer sea-ice margin, because WMM7 deposition coincides with an expanding then maximum summer sea ice extent in the Atlantic sector of the Southern Ocean between 29-22 ka ( Fig. 6) (Collins et al., 2012;Gersonde et al., 2005;Gersonde et al., 2003;Allen et al., 2011;Fischer et al., 2007;Xiao et al., 2016), which places the summer sea-ice margin at 55-60°S, >1500-2000 km away from WMM7 and far beyond the typical maximum foraging range for this species (Fig. 1).

495
Today, snow petrel foraging in high (>80%) sea-ice concentrations occurs in the margins of polynyas (Ainley et al., 1984), regions of open water within the winter-or multi-year sea-ice pack, which are critical for enabling high levels of primary productivity, carbon cycling and air-sea gas exchange (Arrigo et al., 2008;Sherrell et al., 2015). Several marine sites in the Weddell and Lazarev Seas have indicated sporadic polynya formation during MIS 2 and 3 ( Fig. 1) (Sprenk et al., 2014;Smith et al., 2010;Thatje et al., 2008). Deposit WMM7 confirms previous proposals that polynyas were important for 500 https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License. supporting snow petrel colonies through the last glacial (Thatje et al., 2008;Berg et al., 2019). We interpret our new records 510 of changes in snow petrel diet in MIS 2 to have occurred as a result of changes to conditions within, and/or the locations of, polynyas offshore of DML. Unlike today, where polynyas are largely winter/spring phenomena which impact spring-summer productivity during or after sea-ice melt as "post polynyas" (Arrigo and van Dijken, 2003), summer sea surface temperatures below freezing in the Weddell Sea (Gersonde et al., 2005) would have supported more extensive sea ice (and polynya) formation in summer during the last glacial stage (Fischer et al., 2007). 515 Diet studies in recent decades indicate that snow petrels consume a mixed diet of krill, squid and myctophid fish when feeding beyond the continental shelf, whereas fish are most important in shelf waters, at least in the Ross Sea Fijn et al., 2012;e.g. Ridoux and Offredo, 1989;Ainley et al., 1984;Falla, 1937). A similar pattern in Antarctic petrels (Thalassoica antarctica) has been attributed to a lack of Antarctic krill inshore during the breeding season (Nicol, 1993). 520 Although adult krill and larvae over-winter within the sea ice, post-larval krill are mostly oceanic (Atkinson et al., 2008).
During the austral spring and summer, corresponding to the breeding season of the snow petrels, adult krill move to deeper waters for egg development (Nicol, 2006), which may account for the observed increase in krill in snow petrel diets when they forage beyond the continental shelf. Given the importance of deep waters for the krill lifecycle, we infer that the mixed diet of krill and fish in Units I and III represent snow petrels foraging in polynyas located beyond the continental shelf. In 525 contrast, the observed shift in fatty acid and element profiles in Unit II suggests that fish became more important to snow petrel diet, suggesting that polynyas had opened up over the continental shelf between 26.8-25.7 ka. We hypothesise that these shifts in foraging habitat reflect changes in sea ice conditions, by either influencing prey distributions or access to surface waters for feeding.

530
We discount a latitudinal control over  13 C in WMM7, because the seasonal and spatial ranges of  13 C within sea ice can exceed the Southern Ocean latitudinal  13 C gradient (Dunbar and Leventer, 1992;Kennedy et al., 2002) due to changing CO2(aq) supply and productivity (Henley et al., 2012;Kennedy et al., 2002). Organic matter in the sea-ice zone tends to have elevated  13 C compared to open waters beyond the ice edge (Henley et al., 2012;Kennedy et al., 2002), but  13 C14:0 production in open waters for WMM7 is not consistent with the expanded sea ice extent of MIS 3-2 outlined above. Rather, 535 low  13 C14:0 is consistent with (krill) production in the winter/spring sea ice (Dunbar and Leventer, 1992;Henley et al., 2012) and/or more oceanic (rather than coastal) habitats (Cherel et al., 2011), with 13 C-enriched C18 fatty acids in WMM7 reflecting primary production and fish tissues from summer and/or more coastal habitats. No long-term trends are recorded in  13 C18:0 or  13 C18:1, and all four of the major fatty acids have highly variable  13 C (ranges >2‰) after 25.3 ka (Unit I). In contrast, a long-term decrease in  13 C14:0 between 27.9-26.3 ka (Units III and II) occurs alongside declining krill 540 contributions (Cu/Ti and C18:0/C14:0). The cause of the long-term decline in  13 C14:0 is unclear, but could reflect prolonged https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License. winter (low  13 C) sea ice, which leads to 13 C-depletion (Dunbar and Leventer, 1992) and also to restricted access to krill (Watanabe et al., 2020), even though increased winter sea ice is important for Antarctic krill recruitment (Atkinson et al., 2004;Jaeger and Cherel, 2011). Alternatively, declining  13 C14:0 could reflect enhanced upwelling of CO2(aq) rich waters during the season of krill production. The decreasing  13 C14:0 and krill contributions from 28.8 ka culminate in the shift to a 545 diet rich in fish by 26.8 ka (Unit II) which we infer to represent coastal polynya development, reducing krill recruitment in shallower waters and/or reducing snow petrel access to krill with more widespread spring sea ice. Elevated phytoplankton contributions between 25.5-24.7 ka in WMM7 are also consistent with the highest primary productivity and/or low grazing observed today in shallow (continental shelf) waters, including coastal polynyas (Delmont et al., 2014;Karnovsky et al., 2007;Arrigo and van Dijken, 2003). 550 The cause of the millennial-scale switching between foraging in coastal and open-ocean polynyas is not clear, but could reflect ocean temperatures, wind speeds and directions (Morales Maqueda et al., 2004) and availability of continental shelf waters. Colder Southern Ocean waters during glacial stages (Gersonde et al., 2005) are expected to limit formation of those open-ocean polynyas which rely on ocean-driven sea ice melt (Ferrari et al., 2014) and reduced vertical mixing may have 555 resulted from sea-ice driven stratification in the upper water column (Crosta and Shemesh, 2002). However, upwelling driven by wind and/or terraces on the continental slope may explain LGM polynya formation at site PS1506 (Thatje et al., 2008) and upstream of site PS1795 (Sprenk et al., 2014), within the foraging range of snow petrels breeding at Lake Untersee (Fig. 1). Enhanced wind speeds during glacial stages are also expected to have encouraged sea-ice break up and coastal (ice sheet-or ice shelf-proximal) polynya formation (Morales Maqueda and Rahmstorf, 2002;Sprenk et al., 2014;560 Smith et al., 2010), as observed today (Morales Maqueda et al., 2004).
Unlike other parts of the Antarctic ice sheet (Bentley et al., 2014), access of snow petrels to continental shelf waters during MIS 2 may also have been facilitated by relatively minor changes to vertical and lateral ice-sheet extent in DML over the last 100,000 years (Mackintosh et al., 2014;Hillenbrand et al., 2014). The exact position of the LGM ice-sheet limit on the DML 565 continental shelf is not well constrained (Fig. 1) (Mackintosh et al., 2014), but our data from WMM7 suggests that at least part of the continental shelf was free of glacial ice, enabling the development of coastal polynyas within MIS 2. Times of ice-sheet advance may then have reduced the area of available continental shelf and increased wind-driven polynya formation over the continental slope, as suggested by Sprenk et al. (2014), which would account for the increased contribution of krill to snow petrel diet. The relatively restricted extent of the continental shelf at DML (<100 km) may make 570 coastal polynya formation in this region particularly sensitive to ice-sheet advance.
The sustained presence of polynyas through MIS 3-2, as recorded in WMM7 and in other DML sequences (Fig. 6)   may also offer an explanation for the small amplitude of millennial-scale oscillations in sea-salt Na to the EPICA ice core, and the weak relationship to Antarctic temperatures (Fischer et al., 2007). Although we show that polynya properties changed through time (Fig. 6), the open waters within the sea ice pack would have provided a continuous supply of sea-salt Na to the ice core site. Polynyas may also have affected the strength of the sea ice/climate feedbacks during MIS 2: introducing only 2-8% open waters into the LGM sea-ice pack (compared to 10-20% for winter today) reduces the Southern Ocean contribution to the LGM CO2 draw-down from ~80% to 15-50% via enhanced ocean-atmosphere CO2 transfer (Morales Maqueda and Rahmstorf, 2002). In contrast, increasing brine formation, either over the continental shelves 580 or at the ice-sheet margin, would have been conducive to formation of dense glacial AABW and the associated deep-ocean storage of CO2 (Paillard and Parrenin, 2004;Adkins, 2013;Adkins et al., 2002). It is currently difficult to evaluate the relationship between the proposed variability in polynya positions and the millennial-scale oscillations in atmospheric CO2 ( Fig. 6), in part because it is unclear whether the variations in surface ocean productivity observed in the stomach-oil deposits are related to changes in the efficiency of the biological pump and CO2 drawdown (e.g for Unit II with high chlorin 585 inputs). Our age model uncertainties also limit confident correlation between WMM7 and the ice core CO2 record, so that further testing is required to explore whether polynya development along the DML coastline impacted atmospheric CO2.

Conclusions
Here, we present a multi-proxy analysis of stomach-oil deposits of snow petrels from Dronning Maud Land, Antarctica. We show that variation in trace metals and fatty acid distributions strongly suggest changes in snow petrel diet between 28.8-590 22.6 ka. We show that, as today, a mixed diet of krill and fish characterises much of the record. However, between 26.8-25.7 cal. kyr BP signals of krill in the diet almost disappear. By linking dietary signals in the deposits to modern feeding habits and foraging ranges, we highlight the presence of open water ('polynyas') within more extensive summer sea-ice cover during MIS 2. The reduced contribution of krill to the snow petrel diet between 26.8-25.7 ka suggests restriction of polynyas to the continental shelf, limiting krill recruitment or access to waters where krill were present. Our results show that 595 extensive, thick and multi-year sea ice was not always present close to the continent during MIS 2. These results challenge existing hypotheses which emphasise multi-year sea ice as a key driver of positive sea ice-climate feedbacks during glacial stages, whilst also highlighting the potential of stomach-oil deposits as a palaeoenvironmental archive of Southern Ocean conditions. 600 https://doi.org/10.5194/cp-2021-134 Preprint. Discussion started: 4 October 2021 c Author(s) 2021. CC BY 4.0 License.  Figure B1: consistency of XRF scanning signals in deposit WMM7. The XRF data presented in Fig. 3 was recovered from a central slab of WMM7 (Fig. f), where we identified the longest sequence of horizontal laminae (Fig. 3). Here we show additional scans to investigate the internal consistency of the signals. Panels (a.-e.) correspond to lines (a.-e.) on panel f. Scans were undertaken to the left (b.) and right (d.) of the centre-line (c.), then two further cuts were used 610 to investigate signals perpendicular to the cut surface (a. and e.). Away from the central line the signals are less strong and sometimes intermittent, but the laminae also dip sharply away from the centre (f.), suggesting more heterogenous accumulation, perhaps due to distance from the nest and/or nest morphology.  Figure C1: comparison of cluster analysis results for original XRF data, re-sampled XRF data (shown in Fig. 3) and organic indicator data (shown in Fig. 4). XRF data between 0-10 mm and 155-160 mm were removed before analysis. The left panel shows the original Cu/Ti data (grey line) with ~100 yr smoothing (orange line, as in Fig. 3), alongside 620 re-sampled Cu/Ti data (orange squares) used for cluster analysis. The right panel shows the P410 pigment absorbance signal, which may account for the offset in the O1/O2 cluster boundary compared to 1/2 boundary in other analyses.