Sediment core MV1012-001KC was retrieved by research vessel Melville during cruise MV1012 in 2012 and subsequently stored and accessed at the Scripps Institution of Oceanography's sediment core and microfossil collection. We used a 30-cm-long section of this core, referred to as “KC1” for the remainder of this manuscript. Depths are expressed as relative centimeters, starting at the top of the sampled section. We additionally utilized data from three 5 cm replicated MSI measurements (approx. years 1913–1935) of boxcore SPR0901-05BC, originally published by Alfken et al. (2020), as an example of well-preserved varved deposition with known correlation to the instrumental record, named “SBX” in the remainder of the study.

The sediment samples were treated following the procedure developed by Alfken et al. (2019); an additional description of laboratory and processing workflows can be found in Zander et al. (2026). Briefly, after subsampling, X-ray photographs of the wet 30 cm sediment slabs (Fig. 1, step 1) were obtained using a Faxitron 43855A X-ray cabinet (Hewlett-Packard). 5 cm subsections were freeze-dried and embedded in an aqueous gelatine : CMC (4 % : 1 %) solution (gelatin from porcine skin, gel strength 300, type A and sodium carboxymethylcellulose, Sigma-Aldrich Chemie GmbH, Munich Germany, see Fig. 1, steps 1–2). After freezing, the embedded samples were cut into 100-µm slices using a cryomicrotome (Medite Cryostat M630), and the slices were placed on indium-tin-oxide coated (ITO) glass slides (Bruker Daltonik GmbH, Bremen, Germany, see Fig. 1, step 3). High-resolution RGB scans were performed as guidance for setting up FT-ICR MS measurements. Finally, organic compounds between m/z 554 and m/z 564 on each of the slices were measured using a 7T solariX XR FT-ICR MS coupled to an LDI source equipped with a Smartbeam II laser (Nd:YAG, UV-A Laser, 355 nm wavelength, Bruker Daltonik, Bremen). Details of the FT-ICR MS measurement settings can be found in Sect. S1 in the Supplement.

For each 5-cm depth interval of KC1 00–30 cm relative depth, the workflow was repeated on three subsequent slices, yielding triplicate measurements with the minimum achievable distance to the previous measurement area (100 µm cutting thickness).

MSI measurements were inspected and lockmass-calibrated using Bruker Compass DataAnalysis 5.0 SR1. Our main targets were the Na⁺ adducts of di- and tri-unsaturated alkenones (C37:2, C37:3). Each of the shots (orange dots in Fig. 1, step 4) resulted in a mass spectrum from which the intensities of the compounds, as well peak quality (“SNR_FT”) were extracted and used for further analyses, either per spectrum (spotwise-calculated U37K′, “swUk”) using the formula 1U37K′=C37:2(C37:2+C37:3) (Prahl and Wakeham, 1987) or aggregated per horizon (see Fig. 1, step 4–6 and Sect. 2.5 “Time-Series Analyses”). Peak intensities were only used when passing a SNR_FT quality threshold. Besides alkenones, we utilized pyropheophorbide α as a first-order estimation for primary production and calibrant during FT-ICR-MS lockmass calibration because of its ubiquity in SBB sediments (Liu et al., 2022) and stability over the timescales of this study (Szymczak-Żyła et al., 2011).

During the development of the MSI workflow by Wörmer et al. (2014) and in subsequent studies, the resulting U37K′ values were verified using GC-FID measurements (Alfken et al., 2020; Napier et al., 2022; Obreht et al., 2022a, b). Differences in U37K′ values obtained by different mass spectrometry techniques have been reported and attributed to varying ionization efficiencies of alkenones, the co-elution of other compounds, and additional factors (Chaler et al., 2000, 2003; Liao et al., 2023; Rama-Corredor et al., 2018). Notably, data derived from MSI in warm SST regions have shown a cold bias compared to GC-FID data, leading to the introduction of site-specific correction factors. However, in the temperate SST regime of SBB, MSI and GC-FID data did not exhibit significant differences and correlated well with CalCOFI buoy SST data at the site (Alfken et al., 2020; California State Department of Fish and Game; NOAA Fisheries; Scripps Institution of Oceanography, 2001).

The calculated U37K′ values were then translated to temperature using the calibration of Prahl and Wakeham (1987). Other calibrations are available (Tierney and Tingley, 2018); however, the experimental calibration of Prahl and Wakeham was confirmed by Müller et al. (1998) for 60° S–60° N and the applicability of this calibration for California specifically was supported by Herbert et al. (1998). As our results focus on replicability, they are not sensitive to the choice of calibration.

The Santa Barbara Basin is located in the Southern California Bight and part of the California Current System (Bograd et al., 2019). The basin covers ∼ 110 km² and is separated from the Pacific in the west and the Santa Monica Basin in the east by submarine sills, with the depocenter at ∼ 595 m (Soutar and Crill, 1977). The bathymetric layout, together with seasonal upwelling, high productivity in summer, and high sedimentation rates from winter runoff, can favor the preservation of laminated or varved sediments (Reimers et al., 1990; Thunell et al., 1995). The sediment sequence is interrupted by flood layers (“gray layers”) and mass movements from the upper shelf close to the shore (“massive olive layers”), and bioturbated intervals (Du et al., 2018; Hendy et al., 2013). Oxic time intervals allow colonization by macrofauna, for example, during the Macoma event 1835–1840 (Schimmelmann et al., 1992). During the Holocene and Late Pleistocene, varved, laminated, or bioturbated sections alternated in response to phases of varying oxygenation (Anderson et al., 1990). However, Bernhard et al. (2003) reported microscale bioturbation even under oxygen-limited or anoxic-sulfidic conditions.

We conducted our analyses on depth scale, as the focus was on replicability, which is independent of an exact age scale, and depth- or distance-based interpretations enhance the comparability of our results across different depths and sediment structures within the SBB. For general orientation, we anchored our sediment section KC1 to existing chronologies by identifying flood and event layers in the X-ray imagery (see Supplement Sect. S2). The topmost 0–5 cm relative depth piece contains a flood layer that aligns with the 1761 AD flood layer in SBB (Hendy et al., 2013). The presence of a massive olive-colored layer and partially mixed sections prevented the development of a detailed age model beyond linear interpolation between the 1761 and 1532 AD gray flood layers following approaches such as Zhao et al. (2000) and O'Mara et al. (2019). As a first-order age model to help interpret results of the spectral analyses, we use 1.4 mm yr⁻¹ based on average sedimentation rates and typical varve thickness for compacted sediments throughout SBB during the Holocene (Behl, 1995; Du et al., 2018; Ingram and Kennett, 1995; Schimmelmann et al., 1990; Schimmelmann and Lange, 1996).

We estimated the spatial correlation structure of the measured compounds and spotwise-calculated U37K′ (“swUk”) temperatures using variogram analysis (Cressie, 1993; Maroufpoor et al., 2020; Wikle et al., 2019). For each starting point, “shot” in our case, the semivariance γ between the starting point and all other points in specific distance bins was calculated. We compared differences in horizontal (within horizon, approximately parallel to lamina, see orange lines in Fig. 1, step 4) and vertical (downcore) semivariances as a metric for spatial (dis)similarity. Separation by direction enables the characterization of zonal and geometric anisotropy, which is the directional dependence of the differences or their increase over distance, which can be used to assess the influence of lamina preservation and the homogeneity of deposited horizons. Peak intensities of alkenones and pyropheophorbide α were log10 transformed, scaled, and centered to account for the strong non-normality of the data, whereas spotwise U37K′ (swUk) values were used without transformation. For the estimation of the sample variogram, we used the R::gstat implementation (Gräler et al., 2016; Pebesma, 2004; Posit team, 2025; R Core Team, 2024) choosing 100 µm bin widths, together with horizontal and vertical tolerances of 10° to ensure sufficient data coverage and account for deviations in sample or laminae orientation.

MSI data was mapped to the X-ray density maps using three tie points and an affine transformation to correct for potential sample deformation during embedding and cutting (Alfken et al., 2020) using the Python software msiAlign (Liu et al., 2025). The known flood layer in the upper part of the KC1 segment was identified using X-ray density scans and MSI data, and the respective subsection was excluded from the 0–5 cm interval replicates prior to statistical analyses. Horizons with very low numbers of MSI detections across all compounds (fewer than 10 simultaneous detections of C37:2 and C37:3) were removed. To avoid introducing gaps and ensure sufficient detections per horizon, the minimum horizon width was set to 200 µm. The time series were then formed by summing the intensities of each of the valid alkenone detections (simultaneous detection of both alkenones per shot, where each passes the SNR_FT threshold) within each horizon and calculating the U37K′. Short segments of missing horizons due to insufficient detections were linearly interpolated. On average, ∼ 4 % of horizons were missing: ∼ 1.4 % at the ends of series and ∼ 2.7 % within.

Following Laepple and Huybers (2014), power spectral density (PSD) estimates were calculated on linearly detrended, centered data using the multitaper method with three tapers and a time-bandwidth parameter of ω=2 (Thomson, 1982). The PSD estimates were smoothed using a Gaussian kernel with a constant width of 0.2 on the logarithmic timescale (base 10) (Kirchner, 2005), and the three lowest and highest frequencies were omitted.

We followed the framework developed by Münch and Laepple (2018), based on a partitioning of variance approach (Fisher et al., 1985), to separate the climatic signal and noise contributions in the spectral domain. We defined the climate signal as the common signal shared between the time series of a depth interval. The residual, individual variations, formed the noise component and allowed the quantification of the shared signal as signal-to-noise ratios.

Given a regional cluster of n proxy records with a similar climate between sites, the mean power spectrum, M, averaged across all individual records' spectra, will yield a precise estimate of the proxy spectrum. In contrast, the power spectrum S of the stacked record from averaging all datasets in the time domain will also contain the full common signal, but with the noise proportions reduced by a factor of n. By combining both quantities one can derive expressions for the climate and noise spectra (McPartland et al., 2024; Münch and Laepple, 2018) 2C=nn-1S-M/n3N=nn-1M-S-n-1n with the ratio of C:N yielding the frequency-resolved signal-to-noise ratio (SNR). We further compute the integrated SNR which yields the SNR of the time-series that one obtains if one measures at a specific sampling resolution 4iSNRfNyq=∫0fNyqCf∫0fNyqNf The relation of SNR to correlation can be found in the Supplement (Sect. S3), however as direct climatic interpretation is not the focus of this study, we report shared signal content between replicates expressed as SNR throughout the manuscript.

The measurement areas ranged from 14 907 to 26 937 laser shots for each 5-cm slice (20 957 ± 4134 = 1σ), and thus ∼ 40 shots per 100-µm horizon. On average, at 62 % of the shots, at least one of the three compounds was detected. Pyropheophorbide α was detected in 61 % of the spectra, while at least one alkenone was detected in 46. Both alkenones were detected simultaneously in 34 % of the spectra, which could then form swUk estimates (spotwise-calculated U37K′). When aggregating 200-µm horizons to time series, these detection rates amounted to ∼ 37 shots on average that yielded at least one alkenone compound and an average of ∼ 31 simultaneous detections per horizon (see Tables 1 and S5 in the Supplement).

We first investigated the replicability and signal content of individual MSI maps in terms of the similarity of their overall swUk (spotwise-calculated U37K′) values and distributions. The swUk maps showed strong spatial heterogeneity (Fig. 2B). Despite this, the average values were highly consistent between MSI replicates: the standard error of the mean swUk per triplicate was just 0.001, corresponding to ∼ 0.03 °C (Prahl and Wakeham, 1987). Similarly, the level of spatial variability was consistent across the replicates. The average swUk variance within maps (“field variance”) was 0.00595±0.00004, corresponding to 2.34 ± 0.19 °C. The swUk distributions were similar across depth intervals and were unaffected by sediment structure (Fig. 2C).

Values from the same horizon can be aggregated into a single data point, and thereby, a time series can be constructed. The average variance of swUk (spotwise U37K′) in these 200 µm horizons was similar to the observed total field variances: 0.006±0.00005 (2.35 ± 0.21 °C, Supplement Sect. S5). A visual comparison between replicate time-series showed good agreement in terms of mean values and slow variation (Fig. 2D), yet weak correlation of the fastest resolution of 200-µm steps, with an average Pearson's r of 0.12 and average RMSE of 0.03 (∼ 0.89 °C). Such weak correlations most likely result from a lack of fine-scale depth alignment.

After confirming consistency between replicates on a coarse whole-slice level, we examined the spatial proxy heterogeneity using maps of spotwise-calculated U37K′ values (swUk). These spatial patterns showed high variability both across depth and laterally, with average swUk variances across depths and horizons both approximating 0.00596 (vertical, downcore: 0.00597; horizontal: 0.00596).

To explore the structure of this heterogeneity in more detail, we focused on two representative depths, one laminated and one mixed, along with their respective replicates (Fig. 3). Notably, the clear laminations observed in the X-ray scans (e.g., at 5–7 cm; Fig. 2A, top row) are not reflected in the swUk patterns. Instead, both laminated and mixed intervals show swUk values forming small clusters of neighboring pixels with similar temperatures.

To describe local (dis) similarities in biomarker patterns more precisely, we obtained averaged variogram estimates of the compounds and swUk. This approach revealed that only a small portion of total variability is spatially structured (Table 2) in KC1. The total map-level variance, the “sill” in the variograms, was 0.00595 (∼ 2.34 °C), of which only 0.0002 (∼ 0.43 °C) showed small-scale spatial structure, that is, clustering or increased similarity among neighboring values. Most of the variability (∼ 93 %–99 %, avg. 96 %) is unstructured or unresolved, termed the nugget effect (C0). A high contribution of C0 can indicate variation, which is either spatially unresolved, intrinsic variability of the proxy, or measurement noise. In contrast, pyropheophorbide α showed a slower decay of similarity with distance and a lower C0 “nugget” component (80 %–96 %, avg. ∼ 86 %), indicating stronger spatial structure (Fig. 4A, C green lines).

We then separated vertical from horizontal similarity to distinguish between variability across successive depositional layers (in situ vertical, downcore) and within the same horizon (horizontal). It is important to note that the spatial plots in Figs. 2B and 3 are rotated by 90° to align with time-series convention; therefore, variograms estimated over the vertical downcore direction in the original sediment position have to be visualized left to right on these plots. As a proof of concept, we also applied this approach to a varved interval from a box core as the best-case reference without sediment mixing (Alfken et al., 2020, 2021). In both sample sets, this separation (solid lines: original sediment position vertical, downcore; dashed: horizontal, Fig. 4) reveals geometric anisotropy: similarity decreases more slowly in the horizontal than in the vertical direction. This can be interpreted as the within-horizon variability, parallel to the sediment surface, (dashed) which is lower and increases more gradually with distance than between layers (solid). This pattern is most clearly visible in the intensity data of pyrophophorbide α, but is also detectable in the C_37:3 alkenone and spotwise U37K′ data. As expected, the differences were clearer in the varved section of SBX than in the combined core KC1. Notably, in the varved section, pyropheophorbide α and, to a lesser extent, U37K′ show a sinusoidal pattern in γ with a distance between peaks and troughs that matches reported average annual sedimentation rates of ∼ 1.4 mm (Behl, 1995; Du et al., 2018; Ingram and Kennett, 1995; Schimmelmann et al., 1990). Variograms of individual sediment slices can be found in the Supplement (Sect. S6).

Replication within 5-cm depth intervals allowed us to quantify the reproducibility between individual MSI slices and estimate the SNR of the resulting time-series. The frequency-dependent separation of the shared signal and independent noise shows that the replicates contain increasing signal content with timescale and flat noise spectra. Signal spectra showed an average slope of ∼-1.7 estimated between f=1/10 mm⁻¹ and f=1/2 mm⁻¹, the interval of the spectral estimates not affected by bias and gaps. The noise components showed no frequency dependency (∼-0.13 between f=1/10 mm⁻¹ and f=1/2 mm⁻¹, see Fig. 5A).

As a reference, we added noise spectra derived from total field variance (black dotted line, Fig. 5A), downscaled by a factor of ∼ 37, the average number of swUk measurements per horizon at 200 µm resolution (see also Table 2), to illustrate how independent (measurement) noise scales with averaging. The actual noise level was ∼ 13 higher than this baseline, indicating a noise reduction weaker than 1/n when aggregating to horizonwise U37K′ data.

Integrating the spectra from high to low frequencies yields the integrated SNR (iSNR), which is an estimate of the expected SNR at a given sampling resolution. iSNR values reach from ∼ 0.24–∼ 3.1, considering sampling from submillimeter to centimeter steps (Fig. 5B).

We then assessed the influence of sediment structure on time-domain signal-to-noise variance ratios (Fisher et al., 1985) (that is SNR between the variance of the noise and signal components) of the different depth intervals in comparison to the “best case” varved 5 cm section published earlier (Alfken et al., 2020, 2021). We also repeated the analysis using different aggregation steps (Fig. 6) to assess the impact of data processing choices. The tested sampling widths ranged from 200 to 1400 µm, which, under ideal preservation and an average sedimentation rate of 1.4 mm yr⁻¹, would correspond to monthly to annual resolutions. For core segment KC1, these SNRs ranged from ∼ 0 in the most strongly mixed interval to ∼ 2 in an interval with partially preserved varves (Fig. 6). In comparison, the varved section of SBX reached SNRs of ∼ 2.8 on average.

Variogram analyses showed greater similarity across sedimentary horizons than downcore (Fig. 4, dashed and solid lines), consistent with the expectation of layered archives, where variation over depth includes additional variability from changing climate signals over time (Fisher et al., 1985; Münch and Laepple, 2018).

Biomarker maps revealed sub-millimeter-scale clustering and marked differences in spatial heterogeneity between compounds (Figs. 4, 5). For swUk, the variance that exhibited spatial structuring was minimal (∼ 4 %). The small-scale clustering (ranges ∼ 0.5 mm) observed in the biomarker maps implies that, for optimal sampling, MSI measurement windows should be in the range of a few millimeters to be wide enough to (1) have enough successful shots to overcome sparse detections and (2) avoid bias or correlated noise that could arise from only targeting single proxy clusters. Therefore, the typical measurement window widths of ∼ 4–8 mm used in this and earlier studies (Alfken et al., 2020, 2021; Liu et al., 2022; Napier et al., 2022; Obreht et al., 2022b; Wörmer et al., 2014, 2022b) should effectively reduce the noise in downcore records.

Under favorable sediment preservation conditions, varved sections manifest as oscillations in the variogram (Fig. 4A, green solid lines and B, purple solid lines). These can be interpreted as the results of measuring downcore, leaving areas of values that were deposited during a similar time interval and seasonal conditions, which increases average dissimilarity (γ, semivariance of pairwise comparisons as a function of distance), before dissimilarity decreases again when the values from these distances begin to again resemble the starting conditions. For example, when moving into sediments of the same preceding seasons. Therefore, peaks and throughs represent the differences between contrasting or identical seasons, respectively, and the typical distances between the same seasons match reported average sedimentation rates at SBB (Behl, 1995; Du et al., 2018; Ingram and Kennett, 1995; Schimmelmann and Lange, 1996).

The high proportion of unstructured variance in our biomarker and proxy maps suggests significant noise at the individual spot level. Aggregating data across horizons enhances the signal content because noise is largely uncorrelated between spots at the microscale. However, the data from this study does not allow us to determine whether this noise stems from measurement errors or the deposition of heterogeneous biomarker signals. Strong shot-to-shot variation is a common issue in MSI (Tobias and Hummon, 2020). On the other hand, classic bulk sediment U37K′ estimates reflect averaged or smoothed signals, and it is difficult to estimate potentially contained heterogeneity. Observed variability is influenced by the temporal and spatial scales that are aggregated within each sample. For instance, water column and sediment trap data display a broader range of U37K′ values around measured ocean temperatures than core tops (Conte and Eglinton, 1993; Freeman and Wakeham, 1992; Herbert et al., 1998). Our observed proxy spot-level variance is smaller than that of the monthly SST estimates for SBB, but larger than the interannual variation (Sect. S7), suggesting that the observed heterogeneity may indeed originate from SST and capture variability typically lost in bulk sediment samples. MSI-based U37K′ maps could help bridge processes along the proxy formation chain by resolving within-bulk-sample variability, and may provide access to high-frequency climate variability, including intraannual variations. This parallels developments in foraminiferal isotope proxies, where analyses of individual foraminifera complement bulk measurements (Glaubke et al., 2021) and have been argued to resolve seasonal to interannual variability.

Averaging across a diverse range of sediment preservation conditions in core segment KC1, we detected iSNRs around ∼ 1 at interannual resolution, increasing to 3 at subdecadal (ENSO) resolution. The extracted noise component shows no dependence on time scale, whereas the variance of the signal components increases with time scale. This suggests that temporal aggregation effectively reduces noise in individual time series and enhances the relative potential climate signal content. Additionally, stacking multiple replicates would further decrease the noise contribution by a factor of 1/n, boosting the relative signal content.

An iSNR ≫ 1 indicates that individual MSI series are representative of their core samples and demonstrate potential for high-resolution climate reconstructions at subdecadal resolution in the case of the SBB sedimentary archive containing laminated and mixed intervals. This aligns with findings of O'Mara et al. (2019), who estimated the effective temporal resolution of neighboring laminated SST records from San Lazaro Basin to approximately decadal without replication or stacking.

One limitation of our method is its sensitivity to depth misalignment between replicates, which could lead to a bias in SNR estimates at the fastest temporal and corresponding spatial scales (≪1 mm). To minimize these effects, we used sediment marker holes and mapped the MSI data onto X-ray imagery. More advanced image correction and alignment may further reduce these biases in the future. However, low SNRs at ≪1 mm also match expectations based on reports by Bernhard et al. (2003), who documented living microorganisms such as ciliates, foraminifera, and bacterial colonies growing under severe oxygen limitation, even within otherwise undisturbed laminae, pointing to an alternative cause for the low SNRs at the fastest resolutions.

The current SNR estimates are based on closely spaced samples (cryotome cuts within <1 mm horizontal distance) to provide as identical conditions as possible under natural conditions to estimate the baseline noise level for individual MSI-derived U37k′ time-series. This leads to the tradeoff that spatially correlated noise between adjacent cuts may also be interpreted as a shared signal, effectively barring the direct interpretation of the derived SNRs in terms of a pure climatic signal. However, the correlation of ∼ 0.6 between the MSI-derived U37k′ time series and in situ 0–30 m water column temperatures reported by Alfken et al. (2020) is close to the upper achievable limit of ∼ 0.77 as suggested by our SNR estimates for the exemplary varved pieces (Figs. 6A and S3). Together with the short-range spatial similarity found in the variogram analyses (Fig. 4) we expect correlated noise between replicates to play a minor role. Future studies using samples at larger lateral spatial scales (for example nearby sediment cores) combined with longer core segments will help to better characterize and separate the spatial structure of signal and noise components.

Variogram analyses demonstrated that the degree to which a spatial structure is preserved in the sedimentary archive strongly influences the spatial variation patterns of biomarkers. Under strong mixing conditions, such as olive turbidite 1D ∼ 1738 AD (Hendy et al., 2013), horizontal variation and variation over depth are similar. However, as lamina preservation increases, the anisotropy of the variograms becomes more pronounced, with longer ranges of similarity within horizons and generally greater variation over depth, suggesting potential additional climate variation (see also Sect. S6). Additionally, the variogram of a varved example section from the early 20th century captures the rhythmic seasonal variation of the biomarkers over depth (Fig. 4A green solid line), whereas this seasonal pattern is destroyed in the lower sediment sections.

Across a broad, yet plausible range of processing choices, the influence of sediment structure was the dominant factor affecting the high-frequency replicability (SNR) between individual MSI time-series. The achieved SNRs at sub-millimeter to interannual (∼ 1.4 mm) sampling closely follow the degree of preservation of lamina, as visible in the X-ray (Fig. 6C, D). Earlier studies have shown that the mixing and bioturbation intensities assessed in this study are representative of Holocene and Late Pleistocene SBB sediments (Anderson et al., 1990; Behl, 1995) and are closely linked to warm phases, which might limit the potential of high-resolution MSI variability estimates to snapshots during favorable oceanic regimes.