Philippa Kaur

We primarily relied on the “tidyverse”61, “metafor”30, “lmerTest”62, “effects”63, “maptools”64, “sf”65, and “raster”66 packages of the R software67 for organizing and analyzing data, and visualizing the results.
Data assembly
To support the use of meta-analysis to quantify the effects of species richness on plant litter decomposition (i.e., species-mixing effects), we searched the literature using the ISI Web of Science (WoS) database (https://clarivate.com/products/web-of-science/; up to and including December 2018). We used a combination of “decomposition” AND “litter”. These keywords matched 12,278 publications. We then reduced this list of the literature with the keywords of “mix* litter,” OR “litter diversity,” OR “litter mix*,” OR “litter species diversity,” OR “litter species richness,” OR “species mix*,” OR “mix* plant litter,” OR “multi* species litter,” OR “mixing effect*,” resulting in 416 publications. We also searched the literature using Scopus (https://www.scopus.com/home.uri) and Google Scholar (https://scholar.google.ca/) and the latter set of keywords, resulting in 765 publications. An anonymous expert identified 15 additional publications. We decided not to include reports in the grey literature among the papers we found, as we could not verify whether, as in all WoS-listed papers, the papers had undergone independent peer review as a quality check. We read through the papers carefully to select those that focused on quantifying the effects of litter diversity on decomposition rates based on a litter-bag experiment using both mixed- and mono-species litter. We focused on the publications that reported values of either mass loss (or mass remaining) or a decomposition rate constant k that quantified the decomposition rate. Specifically, some papers for litter-mixing effects reported only for the additive effects of mixture and had no reports on mass loss or the constant k. In some cases, no sample size nor standard deviations/errors were reported. Mass loss or standard deviations/errors values were reported only for mono-species or multi-species litter. These publications were not included in the present analyses. In cases that had no report on means but instead showed median values, we estimated means and standard errors based on individual data points or percentile values of boxplots. Note that, standard deviations/errors were often invisible because they were too small to read and thus hidden by their data points; in this case, a conservative approach was adopted by using possible maximum values of deviations/errors (i.e., the size of these symbols). Based on these criteria, we extracted information required for our meta-analysis (see below) from a total of 151 studies (Supplementary Data 1). Out of these 151 studies, 131 and 45 studies measured mass loss and the decomposition rate constant k, respectively. Also see the PRISMA work flow diagram (Supplementary Fig. 6).
For the selected studies, we identified the sets of comparisons between mixed- and mono-species litter bags in each publication. That is, a single study could have multiple treatments that focused on different litter substrates, using different mesh sizes for litter bags, having multiple experiments in different locations, changing abiotic conditions, and so on. The decomposition rate should be comparable in a given treatment under the same set of environmental conditions except for differences caused by a different number of plant species. Even for a given treatment, decomposition rates were often reported multiple times in the literature; in such cases, we only included comparisons for litter retrieved at the same time (i.e., after the same incubation period). Note that in a given treatment, the same mono-species litter could be used for multiple comparisons; for instance, mass loss from a litter bag that contained a three-species mixture (species A, B, and C simultaneously) can be compared with the mass loss in three different mono-species litter bags (only species A, B, or C). This could lead to issues related to pseudo replication, which we considered based on a multilevel random effects meta-analysis (see Data analyses). As a result, we identified 6535 comparisons (across 1949 treatments from the 131 studies) for the values of mass loss and 1423 comparisons (across 504 treatments from the 45 studies) for values of the decomposition rate constant k.
After identifying the sets of comparisons, we extracted the sample size (number of litter-bag replicates, n), and the mean and standard deviation (SD) of the decomposition rate from the main text, from any tables and figures, and from the supplemental materials of the selected studies. If standard errors or 95% confidence intervals (CI) were given, we transformed them into SD values. If only figures were given, we used version 3.4 of the Webplot-digitizer software (https://automeris.io/WebPlotDigitizer/) to extract these parameters from the graphs. We also recorded the study location (longitude, latitude, elevation), climatic region (subarctic, boreal, temperate, subtropical, tropical, or other), ecosystem type (forest, grassland, shrubland, desert, wetland, stream, seagrass, lake, or ex situ microcosm), and litter substrate (leaf, root, stem, branch, straw, or other). Microcosm studies were further divided into two categories: terrestrial or aquatic. The former and latter, respectively, placed litter bags on the soil and in water. For litter substrate, most studies used terrestrial plants, but two studies used macrophytes for their decomposition experiment; these data were included in the main analysis and excluded from the subsequent analyses that focused on leaf litter. Note that authors reported the decomposition rate for mixed-species litter bags in different ways, as the values for all species together or for individual species in the same bag. Removing data based on the latter classification had little effect on our results, so we retained that data. Because of the limited data availability, we did not consider the potential influences of species richness (here, the number of litter species; more than 74% of the comparisons used two- or three-species mixtures) and instead considered species richness as a random term (see Data analyses). Based on the geographic locations of the studies, we estimated their present bioclimatic conditions based on the WorldClim database (www.worldclim.org).
Species-mixing effects on litter decomposition
We calculated the unbiased standardized mean difference (Hedges’ d)30 of the decomposition rate between the mean values for the mixed- and mono-species litter. Hedges’ d is a bias-corrected, unit-free index that expresses the magnitude of the deviation from no difference in the response variable between comparisons. Note that many studies performed multiple comparisons for the decomposition rate between mixed- and mono-species litter, potentially causing issues of pseudo-replication and non-independence in the dataset, as is often the case for ecological syntheses68. We thus applied a multilevel random effects meta-analysis30,69 to account for this problem. In a random effects model, the effect sizes for individual comparisons are weighted by the sum of two values: the inverse of the within-study variance and the between-study variance. The multilevel model can also account for a nested structure in the dataset (in which different treatments and multiple comparisons were nested within each study) and is thus appropriate for dealing with non-independence within a dataset. To calculate the effect sizes, we defined their values to be positive for comparisons in which decomposition was faster in mixed-species litter than in mono-species litter and negative when mono-species litter decomposed faster. This is based on a species gain perspective. Note that some have mono- and two-species mixtures, and others could have for instance from 1 to 16 species in their experiments. Considering mono-species litter as a control is therefore the only way to be consistent across all studies. That is, quantification based on a species loss perspective requires us to obtain the effect sizes using mixed-species litter as a control. In this case, different studies have different levels of richness for a control, making it impossible to have a quantification in a standardized way. We first calculated the effect sizes based on mass loss and the decomposition rate constant k (Fig. 1b, c); because of the limited data availability for the constant k, we only calculated the effect sizes for the entire dataset and the subset of data for leaf litter. For mass loss, we further calculated the effect sizes for the different ecosystem types in different climatic regions; subsets of the data that originated from at least three different studies were used. We conducted a multilevel mixed effects meta-regression by considering random effects due to non-independence among some data points, resulting from a nested structure of the dataset that individual data points were nested within a treatment and treatments were also nested within a study. We used the Q statistic for the test of significance.
Note that, because publication bias is a problem in meta-analysis, we visually evaluated the possibility of such a potential bias by plotting the values of the effect sizes and their variances against sample size. If there is no publication bias, studies with small sample sizes should have an increased sampling error relative to those with larger sample sizes, and the variance should decrease with increasing sample size. In addition, the effect size should be independent of the sample size. Also, there should be large variation in effect sizes at the smallest sample sizes. Prior to the meta-analysis, we confirmed that these conditions existed (Supplementary Fig. 1), and found no publication bias large enough to invalidate our analysis.
We also visually confirmed that the datasets were normally distributed based on normal quantile plots. Furthermore, following the method of Gibson et al.70, we randomly selected only one comparison per treatment and then calculated the effect sizes for the decomposition rate (based on mass loss); we repeated this procedure 10,000 times (with replacement) and found that the species-mixing effects on decomposition were significantly positive (0.247 ± 0.045 for the mean ± 95% CI; Supplementary Fig. 2). We thus confirmed that the overall results were not affected by a publication bias or by non-independence of the dataset. We additionally focused on a subset of the entire dataset, which reported mass loss data for each component species from mixtures and thus had comparisons between mass loss of mono-species litter and that of the same species in a mixture (e.g., mass loss of mono-species litter for species A, B, and C was only compared with that for species A, B, and C within a three-species mixture, respectively). We found the results for this subset of the data (Supplementary Fig. 3) almost identical to the main results (Fig. 2a, b). Note that because of the limited data availability, we did not consider the effects of using different equation forms to estimate the rate constant k and instead considered these among-study differences based on a random term in our multilevel meta-analysis. We also calculated the influence of incubation period (days) on the effect size for the decomposition rate (based on mass loss of leaf litter) for the subsets of the dataset using a mixed effects meta-regression30 (Supplementary Fig. 4). We limited this analysis to studies that retrieved litter bags at least two different time points. To account for non-independence of data points, the study identity was included as a random term.
Impacts of global change drivers on litter decomposition
Litter decomposition is primarily controlled by three factors: the environment (e.g., climate), the community of decomposers, and litter traits46,47,48. By relying on the dataset of Makkonen et al.50, who conducted a full reciprocal litter transplant experiment with 16 plant species that varied in their traits and origins (four forest sites from subarctic to tropics), we aimed to obtain a standardized equation to estimate how decomposition rate changed along a climatic gradient (hereafter, the “standardized climate–decomposition relationship”). First, by considering litter decomposition rates at the coldest location (i.e., the subarctic site) as a reference (control), we calculated the effect sizes (i.e., standardized mean difference based on Hedges’ d) for their decomposition rate constant k for all possible comparisons (i.e., between data from the coldest site and the value, one at a time, for the three other warmer sites). As explained above, the effect sizes were calculated only for the comparisons between decomposition rates of litter that were obtained using the same protocol (litter species, origin, and decomposers). The effect sizes therefore represent the climatic effects on litter decomposition after removing the effects of the decomposer community and litter traits.
We obtained the bioclimatic variables for these four study sites from the WorldClim database, and then modelled the standardized climate–decomposition relationships using a multilevel mixed effects meta-regression30,69. Annual mean temperature, the mean temperature of the wettest quarter, or precipitation of the driest quarter were used as an explanatory variable, with the protocol as a random term. This allowed us to evaluate how decomposition rate can be altered by climate along a latitudinal gradient from tropics to subarctic. We then calculated the decomposition rate constant k for our dataset in exactly the same manner as Makkonen et al.50 and calculated the effect sizes; because of limited datasets for most biomes, we performed this analysis only for the subset of our dataset that we obtained for forest biomes, and we used only leaf litter (57 studies) so the results would be comparable with those of Makkonen et al.50. We used these effect sizes and the standardized climate–decomposition relationships to convert the species-mixing effects into climatic effects; that is, based on the temperature or precipitation required to alter the effect size to the same magnitude as the litter diversity effect, we estimated the climate-equivalency of the diversity effects. This means that we relied on the slope of the relationships to quantify the diversity effects, as has been done in previous analyses38,71. We also compared these estimates to the future projections of changes in decomposition rates in response to climate change (with estimates for the 2070s for two of the representative concentration pathway scenarios of CO2 used in the 5th Climate Model Intercomparison Project; CMIP5 RCP 2.6 and 8.5) for the locations of the original studies (the coordinates of individual studies were collected using the Google Map; www.google.com). This comparison was aimed at quantifying the impacts of different global change drivers (biodiversity and climate change) on decomposition processes. Note that we additionally conducted the above analysis after excluding data with no species identity information for mono and/or mixed-species litter bags, those that measured ash-free dry mass loss (this exclusion was to be consistent with the procedure used in Makkonen et al.50), or both of them (Table S1). As an additional confirmation, we also relied on an alternative method used by Hooper et al.45 to convert the species-mixing effects into climatic effects (mean annual temperature), and compared the impacts of different global change drivers on the decomposition rate (Supplementary Fig. 5).
We further analyzed the potential changes in decomposition resulting from litter diversity and climate change. For the dataset of Makkonen et al.50, we first modelled the relationship between climate (annual mean temperature) and the decomposition rate constant k using an LMM with the protocol as a random term. The constant k was log-transformed to improve homoscedasticity. With this equation and the values of annual mean temperature at the 57 study locations in the forest biomes included in our dataset, we then estimated the expected values of k at these study locations (the present k values). Note that because the experiment of Makkonen et al.50 had no mixed-species litter, the expected k using the above LMM was used primarily to estimate the decomposition rate of mono-species litter at a given annual mean temperature. By combining this estimate with the aforementioned estimate of the climate-equivalency of the litter diversity effect (for annual mean temperature; Fig. 3a), it was possible to estimate the potential changes in the k that resulted from increasing litter diversity from mono- to mixed-species litter at the 57 forest study locations. Specifically, the diversity effect converted into the temperature effect (Fig. 3a) was added to present estimates of annual mean temperature at these study locations, and these values of temperature change were used as an explanatory variable to project changes in k based on the above LMM (the projected k values). In other words, we quantified the projected k values under a scenario in which we brought mono-species litter to areas with a warmer annual mean temperature equivalent to the temperature increase in the climate-equivalency effect. We then converted the present and projected values of the decomposition rate constant k to a mass loss per unit time, making it possible to calculate percentage changes in the decomposition rate due to litter diversification. Furthermore, using the above LMM for the temperature–decomposition relationship with the projected changes in mean annual temperature (based on the CMIP5 RCP 2.6 and 8.5 scenarios) as an explanatory variable, we projected future changes in the decomposition rate constant k at these study locations.
Again, we converted these estimates to a litter mass loss to obtain the percentage changes in the decomposition rate that would result from climate change. The rationale for using the temperature–decomposition relationship was as follows: A similar magnitude of increase in k can, in absolute terms, have different influences on decomposition rates at different locations with a different k. For instance, a change in the value of k from 0.1 to 0.2 and a change from 3.0 to 3.1 are substantially different when considered based on litter mass loss or litter amount remaining after decomposition. We thus included annual mean temperature in our model for quantifying the effects of litter diversity on decomposition. More

1.
Katz, B. J. Controlling factors on source rock development—A review of productivity, preservation, and sedimentation rate. Depos. Org. Sediments Model. Mech. Consequences SEPM Spec. Publ. 82, 7–16 (2005).
CAS  Google Scholar 
2.
Tyson, R. V. The, “productivity versus preservation” controversy: Cause, flaws, and resolution. Depos. Org. Sediments Model. Mech. Consequences SEPM Spec. Publ. 82, 17–33 (2005).
CAS  Google Scholar 

3.
Harris, N. B., Freeman, K. H., Pancost, R. D., White, T. S. & Mitchell, G. D. The character and origin of lacustrine source rocks in the Lower Cretaceous synrift section, Congo Basin, west Africa. Am. Assoc. Pet. Geol. Bull. 88, 1163–1184 (2004).
CAS  Google Scholar 

4.
Talbot, M. R., Filippi, M. L., Jensen, N. B. & Tiercelin, J.-J. An abrupt change in the African monsoon at the end of the Younger Dryas. Geochem. Geophys. Geosyst. 8(3), Q03005. https://doi.org/10.1029/2006GC001465 (2007).
ADS  Article  Google Scholar 

5.
Scholz, C. A. et al. Scientific drilling in the Great Rift Valley: The 2005 Lake Malawi Scientific Drilling Project—An overview of the past 145,000 years of climate variability in Southern Hemisphere East Africa. Palaeogeogr. Palaeoclimatol. Palaeoecol. 303, 3–19 (2011).
Google Scholar 

6.
Ellis, G. S., Katz, B. J., Scholz, C. A. & Swart, P. K. Organic sedimentation in modern lacustrine systems: A case study from Lake Malawi, East Africa. In Paying Attention to Mudrocks: Priceless! (eds. Larsen, D. et al.) 515, (Geological Society of America, 2015).

7.
Ivory, S. J. et al. East African weathering dynamics controlled by vegetation-climate feedbacks. Geology 45, 823–826 (2017).
ADS  Google Scholar 

8.
Lambiase, J. & Morley, C. Hydrocarbons in rift basins: The role of stratigraphy. Philos. Trans. R. Soc. London. Ser. A Math. Phys. Eng. Sci. 357, 877–900 (1999).
ADS  Google Scholar 

9.
Huc, A. Y., Le Fournier, J., Vandenbroucke, M. & Bessereau, G. Northern Lake Tanganyika—An example of organic sedimentation in an anoxic rift lake. In Lacustrine Basin Exploration: Case Studies and Modern Analogs (ed. Katz B. J.) 50, 169–185 (American Association of Petroleum Geologists, 1990).

10.
Katz, B. J. A survey of rift basin source rocks. Geol. Soc. Lond. Spec. Publ. 80, 213 (1995).
ADS  Google Scholar 

11.
Lyons, R. P., Scholz, C. A., Buoniconti, M. R. & Martin, M. R. Late Quaternary stratigraphic analysis of the Lake Malawi Rift, East Africa: An integration of drill-core and seismic-reflection data. Palaeogeogr. Palaeoclimatol. Palaeoecol. 303, 20–37 (2011).
Google Scholar 

12.
Harris, N. B. & Tucker, G. E. Soils, slopes and source rocks: Application of a soil chemistry model to nutrient delivery to rift lakes. Sediment. Geol. 323, 31–42 (2015).
ADS  CAS  Google Scholar 

13.
Talbot, M. R. & Johannessen, T. A high resolution palaeoclimatic record for the last 27,500 years in tropical West Africa from the carbon and nitrogen isotopic composition of lacustrine organic matter. Earth Planet. Sci. Lett. 110, 23–37 (1992).
ADS  CAS  Google Scholar 

14.
Talbot, M. R. The origins of lacustrine oil source rocks: Evidence from the lakes of tropical Africa. Geol. Soc. Lond. Spec. Publ. 40, 29–43 (1988).
ADS  Google Scholar 

15.
Petersen, H. I. et al. World-class Paleogene oil-prone source rocks from a cored lacustrine syn-rift succession, Bach Long Vi Island, Song Hong Basin, offshore northern Vietnam. J. Pet. Geol. 37, 373–389 (2014).
CAS  Google Scholar 

16.
Nytoft, H. P. et al. Novel saturated hexacyclic C34 and C35 hopanes in lacustrine oils and source rocks. Org. Geochem. 87, 107–118 (2015).
CAS  Google Scholar 

17.
Hovikoski, J. et al. Density-flow deposition in a fresh-water Lacustrine rift basin, Paleogene Bach Long Vi Graben, Vietnam. J. Sediment. Res. 86, 982–1007 (2016).
ADS  CAS  Google Scholar 

18.
Fyhn, M. B. W. et al. Linking paleogene rifting and inversion in the Northern Song Hong and Beibuwan Basins, Vietnam, with left-lateral motion on the ailao shan-red river shear zone. Tectonics 37, 2559–2585 (2018).
Google Scholar 

19.
Tuyen, N. T. Biostratigraphic report of the ENRECA‐3 well, Bach Long Vi Island prepared for the VPI‐ODA Project, Ho Chi Minh City, Vietnam Oil and Gas Group, Vietnam Petroleum Institute (VPI), Analysis Laboratory Center. (2013).

20.
Rizzi, M. et al. Hinterland setting and composition of an Oligocene deep rift-lake sequence, Gulf of Tonkin, Vietnam: Implications for petroleum source rock deposition. Mar. Pet. Geol. 111, 496–509 (2020).
CAS  Google Scholar 

21.
Dymond, J., Suess, E. & Lyle, M. Barium in deep-sea sediment: A geochemical proxy for paleoproductivity. Paleoceanography 7, 163–181 (1992).
ADS  Google Scholar 

22.
Föllmi, K. B. The phosphorus cycle, phosphogenesis and marine phosphate-rich deposits. Earth-Sci. Rev. 40, 55–124 (1996).
ADS  Google Scholar 

23.
Kump, L. R. & Arthur, M. A. Interpreting carbon-isotope excursions: Carbonates and organic matter. Chem. Geol. 161, 181–198 (1999).
ADS  CAS  Google Scholar 

24.
Tribovillard, N., Algeo, T. J., Lyons, T. & Riboulleau, A. Trace metals as paleoredox and paleoproductivity proxies: An update. Chem. Geol. 232, 12–32 (2006).
ADS  CAS  Google Scholar 

25.
Boyle, J. F. Inorganic geochemical methods in palaeolimnology. In Tracking Environmental Change Using Lake Sediments (eds Last, W. M. & Smol, J. P.) 83–141 (Springer, Berlin, 2005).
Google Scholar 

26.
Calvert, S. & Pedersen, T. F. Elemental proxies for palaeoclimatic and palaeoceanographic variability in marine sediments: Interpretation and application. Dev. Mar. Geol. 1, 567–644 (2007).
Google Scholar 

27.
Wersin, P., Höhener, P., Giovanoli, R. & Stumm, W. Early diagenetic influences on iron transformations in a freshwater lake sediment. Chem. Geol. 90, 233–252 (1991).
ADS  CAS  Google Scholar 

28.
Melles, M. et al. 2.8 million years of arctic climate change from lake El’gygytgyn, NE Russia. Science 337, 315 (2012).
ADS  CAS  PubMed  Google Scholar 

29.
Naeher, S., Gilli, A., North, R. P., Hamann, Y. & Schubert, C. J. Tracing bottom water oxygenation with sedimentary Mn/Fe ratios in Lake Zurich, Switzerland. Chem. Geol. 352, 125–133 (2013).
ADS  CAS  Google Scholar 

30.
Hatch, J. R. & Leventhal, J. S. Relationship between inferred redox potential of the depositional environment and geochemistry of the Upper Pennsylvanian (Missourian) Stark Shale Member of the Dennis Limestone, Wabaunsee County, Kansas, U.S.A.. Chem. Geol. 99, 65–82 (1992).
ADS  CAS  Google Scholar 

31.
Rimmer, S. M. Geochemical paleoredox indicators in Devonian-Mississippian black shales, Central Appalachian Basin (USA). Chem. Geol. 206, 373–391 (2004).
ADS  CAS  Google Scholar 

32.
Cuven, S., Francus, P. & Lamoureux, S. Mid to Late Holocene hydroclimatic and geochemical records from the varved sediments of East Lake, Cape Bounty, Canadian High Arctic. Quat. Sci. Rev. 30, 2651–2665 (2011).
ADS  Google Scholar 

33.
Johnson, T. C., Brown, E. T. & Shi, J. Biogenic silica deposition in Lake Malawi, East Africa over the past 150,000 years. Palaeogeogr. Palaeoclimatol. Palaeoecol. 303, 103–109 (2011).
Google Scholar 

34.
Lerman, A. Lakes: Chemistry, Geology, Physics (Springer, Berlin , 1978). https://doi.org/10.1007/978-1-4757-1152-3.
Google Scholar 

35.
Jia, J., Liu, Z., Bechtel, A., Strobl, S. A. I. & Sun, P. Tectonic and climate control of oil shale deposition in the Upper Cretaceous Qingshankou Formation (Songliao Basin, NE China). Int. J. Earth Sci. 102, 1717–1734 (2013).
CAS  Google Scholar 

36.
Meng, Q., Liu, Z., Bruch, A. A., Liu, R. & Hu, F. Palaeoclimatic evolution during Eocene and its influence on oil shale mineralisation, Fushun basin, China. J. Asian Earth Sci. 45, 95–105 (2012).
ADS  Google Scholar 

37.
Didyk, B. M., Simoneit, B. R. T., Brassell, S. C. & Eglinton, G. Organic geochemical indicators of palaeoenvironmental conditions of sedimentation. Nature 272, 216–222 (1978).
ADS  CAS  Google Scholar 

38.
Sinninghe Damsté, J. S., Ten Haven, H. L., de Leeuw, J. W. & Rullkotter, J. Pristane/phytane ratio as environmental indicator-Reply. Nature 333, 604 (1988).
Google Scholar 

39.
Sinninghe Damsté, J. S. et al. Evidence for gammacerane as an indicator of water column stratification. Geochim. Cosmochim. Acta 59, 1895–1900 (1995).
ADS  PubMed  Google Scholar 

40.
Irwin, H., Curtis, C. & Coleman, M. Isotopic evidence for source of diagenetic carbonates formed during burial of organic-rich sediments. Nature 269, 209–213 (1977).
ADS  CAS  Google Scholar 

41.
Curtis, C. Mineralogical consequences of organic matter degradation in sediments: Inorganic/organic diagenesis. In Marine Clastic Sedimentology (eds Leggett, J. K. & Zuffa, G. G.) 108–123 (Springer, Berlin, 1987). https://doi.org/10.1007/978-94-009-3241-8_6.
Google Scholar 

42.
Mozley, P. S. & Wersin, P. Isotopic composition of siderite as an indicator of depositional environment. Geology 20, 817–820 (1992).
ADS  CAS  Google Scholar 

43.
Schieber, J., Southard, J. & Thaisen, K. Accretion of mudstone beds from migrating floccule ripples. Science 318, 1760 (2007).
ADS  CAS  PubMed  Google Scholar 

44.
Macquaker, J. H. S. & Bohacs, K. M. On the accumulation of mud. Science 318, 1734 (2007).
CAS  PubMed  Google Scholar 

45.
Föllmi, K. B. & Grimm, K. A. Doomed pioneers: Gravity-flow deposition and bioturbation in marine oxygen-deficient environments. Geology 18, 1069–1072 (1990).
ADS  Google Scholar 

46.
Ingall, E. & Jahnke, R. Evidence for enhanced phosphorus regeneration from marine sediments overlain by oxygen depleted waters. Geochim. Cosmochim. Acta 58, 2571–2575 (1994).
ADS  CAS  Google Scholar 

47.
Fyhn, M. B. W., Hoang, B. H., Anh, N. T., Hovikoski, J. & Cuong, T. D. Eocene-Oligocene syn-rift deposition in the northern Gulf of Tonkin, Vietnam. Mar. Pet. Geol. 111, 390–413 (2020).
Google Scholar 

48.
Talling, P. J. On the triggers, resulting flow types and frequencies of subaqueous sediment density flows in different settings. Mar. Geol. 352, 155–182 (2014).
ADS  Google Scholar 

49.
Morley, R. A review of the Cenozoic palaeoclimate history of Southeast Asia. In Biotic Evolution and Environmental Change in Southeast Asia (ed. Rosen, B.) 79–114 (Cambridge University Press, Cambridge, 2012).
Google Scholar 

50.
Bohacs, K. M., Carroll, A. R., Neal, J. E. & Mankiewicz, P. J. Lake-Basin Type, Source Potentional, and Hydrocarbon Character:An Integrated Sequence-Stratigraphic-Geochemichal Framework. Lake basins through space and time: AAPG Studies in Geology 46, (American Association of Petroleum Geologists, 2000).

51.
Demaison, G. J. & Moore, G. T. Anoxic environments and oil source bed genesis. Org. Geochem. 2, 9–31 (1980).
CAS  Google Scholar 

52.
Cohen, A. S. Facies relationships and sedimentation in large rift lakes and implications for hydrocarbon exploration: Examples from lakes Turkana and Tanganyika. Palaeogeogr. Palaeoclimatol. Palaeoecol. 70, 65–80 (1989).
Google Scholar 

53.
Calvert, S. E., Bustin, R. M. & Pedersen, T. F. Lack of evidence for enhanced preservation of sedimentary organic matter in the oxygen minimum of the Gulf of California. Geology 20, 757–760 (1992).
ADS  CAS  Google Scholar 

54.
Katz, B. J. Lacustrine basin hydrocarbon exploration -current thoughts. J. Palaeolimnol. 26, 161–179 (2001).
ADS  Google Scholar 

55.
Bohacs, K. M. et al. Production, destruction, and dilution—The many paths to source-rock development. Spec. Publ. SEPM 82, 61–101 (2005).
CAS  Google Scholar 

56.
Lehmann, M. F., Bernasconi, S. M., Barbieri, A. & McKenzie, J. A. Preservation of organic matter and alteration of its carbon and nitrogen isotope composition during simulated and in situ early sedimentary diagenesis. Geochim. Cosmochim. Acta 66, 3573–3584 (2002).
ADS  CAS  Google Scholar 

57.
Harris, N. B. et al. Patterns of organic-carbon enrichment in a lacustrine source rock in relation to paleo-lake level, Congo Basin, West Africa. Depos. Org. Sediments Model. Mech. Consequences SEPM Spec. Publ. 82, 103–123 (2005).
CAS  Google Scholar 

58.
Bojanowski, M. J. & Clarkson, E. N. K. Origin of siderite concretions in microenvironments of methanogenesis developed in a sulfate reduction zone: An exception or a rule?. J. Sediment. Res. 82, 585–598 (2012).
ADS  CAS  Google Scholar 

59.
Berner, R. A. Early Diagenesis—A Theoretical Approach. Princeton Series in Geochemistry (Princeton University Press, Princeton, 1980). https://doi.org/10.1016/0037-0738(81)90046-4.
Google Scholar 

60.
Raiswell, R. Chemical model for the origin of minor limestone-shale cycles by anaerobic methane oxidation. Geology 16, 641–644 (1988).
ADS  CAS  Google Scholar 

61.
Canfield, D. E. Reactive iron in marine sediments. Geochim. Cosmochim. Acta 53, 619–632 (1989).
ADS  CAS  PubMed  Google Scholar 

62.
Katsev, S., Sundby, B. & Mucci, A. Modeling vertical excursions of the redox boundary in sediments: Application to deep basins of the Arctic Ocean. Limnol. Oceanogr. 51, 1581–1593 (2006).
ADS  CAS  Google Scholar 

63.
Esbensen, K. H., Guyot, D., Westad, F. & Houmoller, L. P. Multivariate Data Analysis—In Practice: An Introduction to Multivariate Data Analysis and Experimental Design. (Camo, 2002).

64.
Bish, D. L. & Post, J. E. Quantitative mineralogical analysis using the Rietveld full-pattern fitting method. Am. Mineral. 78, 932–940 (1993).
CAS  Google Scholar 

65.
Hutton, A. C. Petrographic classification of oil shales. Int. J. Coal Geol. 8, 203–231 (1987).
Google Scholar 

66.
Taylor, G. H. et al. Organic Petrology (Gebrüder Borntraeger, Berlin, 1998).
Google Scholar 

67.
International, Committee for Coal and Organic Petrology. The new inertinite classification (ICCP System 1994). Fuel 80, 459–471 (2001).
Google Scholar 

68.
Sýkorová, I. et al. Classification of huminite—ICCP System 1994. Int. J. Coal Geol. 62, 85–106 (2005).
Google Scholar 

69.
Nytoft, H. P. et al. Biomarkers of Oligocene lacustrine source rocks, Beibuwan-Song Hong basin junction, offshore northern Vietnam. Mar. Pet. Geol. 114, 104196 (2020).
CAS  Google Scholar 

70.
Radke, M., Willsch, H. & Welte, D. H. Preparative hydrocarbon group type determination by automated medium pressure liquid chromatography. Anal. Chem. 52, 406–411 (1980).
CAS  Google Scholar 

71.
Craig, H. Isotopic standards for carbon and oxygen and correction factors for mass-spectrometric analysis of carbon dioxide. Geochim. Cosmochim. Acta 12, 133–149 (1957).
ADS  CAS  Google Scholar 

72.
Friedman, I. & O’Neil, J. R. Compilation of stable isotope fractionation factors of geochemical interest. US Gov. Print. Off. 440, 1–11 (1977).
Google Scholar  More

1.
Karl D, Michaels A, Bergman B, Capone D, Carpenter E, Letelier R, et al. Dinitrogen fixation in the world’s oceans. Biogeochemistry. 2002;57/58:47–98.
CAS  Article  Google Scholar 
2.
Bombar D, Paerl RW, Riemann L. Marine non-cyanobacterial diazotrophs: moving beyond molecular detection. Trends Microbiol. 2016;24:916–27.
CAS  Article  Google Scholar 

3.
Bird C, Martinez JM, O’Donnell AG, Wyman M. Spatial distribution and transcriptional activity of an uncultured clade of planktonic diazotrophic γ-proteobacteria in the Arabian Sea. Appl Environ Microbiol. 2005;71:2079–85.
CAS  Article  Google Scholar 

4.
Moisander PH, Beinart RA, Voss M, Zehr JP. Diversity and abundance of diazotrophic microorganisms in the South China Sea during intermonsoon. ISME J. 2008;2:954–67.
CAS  Article  Google Scholar 

5.
Moisander PH, Serros T, Paerl RW, Beinart RA, Zehr JP. Gammaproteobacterial diazotrophs and nifH gene expression in surface waters of the South Pacific Ocean. The. ISME J. 2014;8:1962–73.
CAS  Article  Google Scholar 

6.
Langlois R, Großkopf T, Mills M, Takeda S, LaRoche J. Widespread distribution and expression of Gamma A (UMB), an uncultured, diazotrophic, γ-proteobacterial nifH phylotype. PLoS ONE. 2015;10:e0128912.
Article  Google Scholar 

7.
Zehr JP, Turner PJ. Nitrogen fixation: nitrogenase genes and gene expression. Methods Microbiol. 2001;30:271–86.
CAS  Article  Google Scholar 

8.
Church MJ, Short CM, Jenkins BD, Karl DM, Zehr JP. Temporal patterns of nitrogenase gene (nifH) expression in the oligotrophic North Pacific Ocean. Appl Environ Microbiol. 2005;71:5362–70.
CAS  Article  Google Scholar 

9.
Langlois RJ, Hümmer D, LaRoche J. Abundances and distributions of the dominant nifH phylotypes in the Northern Atlantic Ocean. Appl Environ Microbiol. 2008;74:1922–31.
CAS  Article  Google Scholar 

10.
Bonnet S, Rodier M, Turk-Kubo KA, Germineaud C, Menkes C, Ganachaud A, et al. Contrasted geographical distribution of N2 fixation rates and nifH phylotypes in the Coral and Solomon Seas (southwestern Pacific) during austral winter conditions. Glob Biogeochem Cycles. 2015;29:1874–92.
CAS  Article  Google Scholar 

11.
Gradoville MR, Bombar D, Crump BC, Letelier RM, Zehr JP, White AE. Diversity and activity of nitrogen-fixing communities across ocean basins. Limnol Oceanogr. 2017;62:1895–909.
Article  Google Scholar 

12.
Chen TY, Chen YL, Sheu DS, Chen HY, Lin YH, Shiozaki T. Community and abundance of heterotrophic diazotrophs in the northern South China Sea: revealing the potential importance of a new alphaproteobacterium in N2 fixation. Deep Sea Res Part I. 2019;143:104–14.
CAS  Article  Google Scholar 

13.
Moisander PH, Benavides M, Bonnet S, Berman-Frank I, White AE, Riemann L. Chasing after non-cyanobacterial nitrogen fixation in marine pelagic environments. Front Microbiol. 2017;8:1736.
Article  Google Scholar 

14.
Farnelid H, Turk-Kubo K, Ploug H, Ossolinski JE, Collins JR, Van Mooy BA, et al. Diverse diazotrophs are present on sinking particles in the North Pacific Subtropical Gyre. ISME J. 2019;13:170–82.
Article  Google Scholar 

15.
Cornejo-Castillo FM. Diversity, ecology and evolution of marine diazotrophic microorganisms. Barcelona: Universitat Politècnica de Catalunya; 2017.

16.
Delmont TO, Quince C, Shaiber A, Esen ÖC, Lee ST, Rappé MS, et al. Nitrogen-fixing populations of planctomycetes and proteobacteria are abundant in surface ocean metagenomes. Nat Microbiol. 2018;3:804–13.
CAS  Article  Google Scholar 

17.
Benavides M, Moisander PH, Daley MC, Bode A, Aristegui J. Longitudinal variability of diazotroph abundances in the subtropical North Atlantic Ocean. J Plankton Res. 2016;38:662–72.
CAS  Article  Google Scholar 

18.
Gradoville MR, Farnelid H, White AE, Turk-Kubo KA, Stewart B, Ribalet F, et al. Latitudinal constraints on the abundance and activity of the cyanobacterium UCYN-A and other marine diazotrophs in the North Pacific. Limnol Oceanogr. 2020. https://doi.org/10.1002/lno.11423.

19.
Scavotto RE, Dziallas C, Bentzon-Tilia M, Riemann L, Moisander PH. Nitrogen-fixing bacteria associated with copepods in coastal waters of the North Atlantic Ocean. Environ Microbiol. 2015;17:3754–65.
CAS  Article  Google Scholar 

20.
Carradec Q, Pelletier E, Da Silva C, Alberti A, Seeleuthner Y, Blanc-Mathieu R, et al. A global ocean atlas of eukaryotic genes. Nat Commun. 2018;9:1–3.
CAS  Article  Google Scholar 

21.
Karl DM, Church MJ, Dore JE, Letelier RM, Mahaffey C. Predictable and efficient carbon sequestration in the North Pacific Ocean supported by symbiotic nitrogen fixation. Proc Natl Acad Sci. 2012;109:1842–9.
CAS  Article  Google Scholar 

22.
Pedersen JN, Bombar D, Paerl RW, Riemann L. Diazotrophs and N2-fixation associated with particles in coastal estuarine waters. Front Microbiol. 2018;9:2759.
Article  Google Scholar 

23.
Ploug H, Kühl M, Buchholz-Cleven B, Jørgensen BB. Anoxic aggregates-an ephemeral phenomenon in the pelagic environment? Aquat Microb Ecol. 1997;13:285–94.
Article  Google Scholar 

24.
Ploug H. Small-scale oxygen fluxes and remineralization in sinking aggregates. Limnol Oceanogr. 2001;46:1624–31.
CAS  Article  Google Scholar 

25.
Bentzon-Tilia M, Severin I, Hansen LH, Riemann L. Genomics and ecophysiology of heterotrophic nitrogen-fixing bacteria isolated from estuarine surface water. MBio. 2015;6:e00929–15.
CAS  Article  Google Scholar 

26.
Bergman B, Sandh G, Lin S, Larsson J, Carpenter EJ. Trichodesmium—a widespread marine cyanobacterium with unusual nitrogen fixation properties. FEMS Microbiol Rev. 2013;37:286–302.
CAS  Article  Google Scholar 

27.
Moraru C, Lam P, Fuchs BM, Kuypers MM, Amann R. GeneFISH—an in situ technique for linking gene presence and cell identity in environmental microorganisms. Environ Microbiol. 2010;12:3057–73.
CAS  Article  Google Scholar 

28.
Batani G, Bayer K, Böge J, Hentschel U, Thomas T. Fluorescence in situ hybridization (FISH) and cell sorting of living bacteria. Sci Rep. 2019;9:1–3.
Article  Google Scholar  More

Demographic dynamics of urban systems
We now derive the population size distribution of cities in the most general way, set by their demographic dynamics. The change of the population Ni in city i during unit time Δt necessarily results from the balance of births, deaths, and migration as

$${N}_{i}(t+Delta t)={N}_{i}(t)+Delta tleft[{upsilon }_{i}{N}_{i}(t)+sum_{j = 1,jne i}^{{N}_{{rm{c}}}}({J}_{ji}-{J}_{ij})right],$$
(2)

where Nc is the total number of cities, υi = bi − di is the city’s vital rate, the difference between its per capita average birth rate, bi, and the death rate, di. The current Jij is the number of people moving from city i to city j over the time interval Δt, and vice-versa for Jji. For simplicity of notation we will work in units where Δt = 1. Eq. (2) only accounts explicitly for migration between different specific cities, which can cross political borders. Migration to city i from non-specific places outside the system, for example from non-urban regions or other nations, can be incorporated into the vital rates υi. Note that this model is very general, and naturally accommodates the inclusion of new cities over time when their sizes become non-zero, and indeed the disappearance of others, if their size vanishes.
To proceed we need to explore the dependence of the currents Jij on population size. We start by writing these currents as population rates, Jij = mijNi(t), where mij is the probability that someone in city i moves to city j over the time period. This allows us to write Eq. (2) in matrix form

$${bf{N}}(t)={bf{A}}(t){bf{N}}(t-1),$$
(3)

where ({bf{N}}(t)={[{N}_{1}(t),ldots ,{N}_{i}(t),ldots ,{N}_{{N}_{{rm{c}}}}(t)]}^{T}) is the vector of populations in each city at time t and the matrix A projects the population at time t − 1 to time t. This matrix is known in population dynamics as the environment39, with elements

$${A}_{ij}=left{begin{array}{ll}1+{upsilon }_{i}-{m}_{i}^{{rm{out}}},&{rm{if}},i=j\ {m}_{ji},&{rm{if}},ine jend{array}right.,$$
(4)

where ({m}_{i}^{{rm{out}}}=mathop{sum }nolimits_{k = 1,kne i}^{{N}_{{rm{c}}}}{m}_{ik}) is the total probability that a person migrates out of city i per unit time. We also define the population structure vector x = [xi], with xi = Ni/NT, which is the probability of finding a person in city i out of the total population NT. Note that (mathop{sum }nolimits_{i = 1}^{{N}_{{rm{c}}}}{x}_{i}=1), so that the structure vector has Nc − 1 degrees of freedom (the magnitude of x is fixed).
Equation (3) can now be solved by repeated iteration

$${bf{N}}(t)={bf{A}}(t){bf{A}}(t-1)ldots {bf{A}}(1){bf{N}}(0).$$
(5)

City size distributions in different environments
A number of important ergodic theorems40 in population dynamics tell us about the properties of the solutions under different conditions on the environments A(t). These results rely on some constraints on the properties of A; specifically, that the product of matrices in Eq. (5) is positive for sufficiently long times39,41.
First, the weak ergodic theorem guarantees that for a dynamical sequence of environments, the difference between two different initial population structure vectors decays to zero over time. This means that there is typically an asymptotic city size “distribution”, which is a function of environmental dynamics only, independent of initial conditions. When the environment is stochastic but otherwise time independent, the strong stochastic ergodic theorem states that the structure vector becomes a random variable whose probability distribution converges to a fixed stationary distribution, regardless of initial conditions. This is the sense in which most derivations of Zipf’s law apply17,33. For stochastic environments, only probability distributions of structure vectors, not vectors themselves, can be predicted. Finally, in situations where the environment is time dependent and stochastic, the weak stochastic ergodic theorem states that the difference between the probability distributions for the structure vector resulting from any two initial populations, exposed to independent sample paths, decays to zero. Again, in cases where the environment is explicitly dynamic, besides being stochastic in a stationary sense, we cannot say much about the actual probability distribution of city sizes, only that the importance of initial conditions vanishes for sufficiently long times.
To get more intuition on these results we will now show some explicit solutions. The city size distribution can be calculated exactly when the environments A are arbitrarily complicated, but static. This is the result of the strong ergodic theorem, which says that in a constant environment A, any initial population vector converges to a fixed stable structure given by the leading eigenvector of the environment39. This is guaranteed to exist if A is a strongly connected aperiodic graph, meaning that any city can be reached from any other city in a finite and diverse number of intermediate steps along the non-zero migration flows between them. This is a reasonable assumption to make about an urban system, which may even serve as a definition. The leading eigenvector of A is the eigenvector centrality of the urban system. It describes the probabilistic location of a random walker over the graph of migration flows42. This is equal to the stationary solution of numerically integrating a network described by environment A. Figure 1 shows two numerical solutions with different initial conditions, but the same environment A. The cities in the two runs converge to similar sizes, as they experience a common environmental matrix with all entries set to be the same plus a little noise. The insets show how the population structure converges from both initial conditions to the one defined by the leading eigenvector e0. Because the structure vector is the probability of finding a person out of the total population in the urban system in a specific city, we use the Kullback-Leibler (KL) divergence43({D}_{{rm{KL}}}(P| {P}_{{{bf{e}}}_{0}})={sum }_{x}P[x]{mathrm{log}},P[x]/{P}_{{{bf{e}}}_{0}}[x]) as a measure of the ‘distance’ between the initial distribution, P, and the final, ({P}_{{{bf{e}}}_{0}}). The KL divergence is a foundational quantity in information theory. It measures the information gain in describing a probabilistic state using one distribution, ({P}_{{{bf{e}}}_{0}}), versus another, P, in units of information (bits). Note that (Pto {P}_{{{bf{e}}}_{0}}) and thus ({D}_{{rm{KL}}}(P| {P}_{{{bf{e}}}_{0}})to 0) over time, see insets in Fig. 1.
Fig. 1: Temporal evolution of the relative population structure of cities in the same environment for different initial conditions.

Lines shows the trajectory of each city in terms of the fraction its population, Ni to the total NT. a shows an initial situation where all cities start out with similar sizes, whereas in b they are initiated following Zipf’s law (shown in log-scale). In both cases, the relative city size distribution eventually becomes stationary at long times (vertical red line) with the same population structure. This is given by the eigenvector e0, corresponding to the leading eigenvalue of the environment. The insets show how the system converges in both cases to this common structure set by the environment, because the Kullback–Leibler divergence ({D}_{{rm{KL}}}(P| {P}_{{{bf{e}}}_{0}})) approaches zero at late times.

Full size image

The convergence rate of the city size distribution to the leading eigenvector is given by the ratio of secondary eigenvalues to the dominant one. Explicitly, we can now write this solution as

$${bf{N}}(t)={{bf{A}}}^{t}{bf{N}}(0)to {bf{N}}(t)=sum_{i = 1}^{{N}_{{rm{c}}}}{({lambda }_{i})}^{t}{c}_{i}{{bf{e}}}_{i}.$$
(6)

where Aei = λiei, so that ei are the eigenvectors of the matrix A and λi the corresponding eigenvalues, so that λ0  > Re(λ1)  > Re(λ2)  > . . . , where Re(λi) is the real part of the eigenvalue. The projection coefficients ci are such that ({bf{N}}(0)=mathop{sum }nolimits_{i = 1}^{{N}_{{rm{c}}}}{c}_{i}{{bf{e}}}_{i}), which can be written as c = E−1N(0), where E is a matrix whose jth column vector is ej. The strong ergodic theorem states that the largest eigenvalue λ0 is positive and real and that the associated eigenvector e0 is also positive in terms of all its entries. This means that, over time, the dynamics of the urban system approaches that of the growth of its dominant eigenvector, because the projections on all other eigenvectors decay exponentially in relative terms, as

$${left(frac{{lambda }_{i}}{{lambda }_{0}}right)}^{t}={e}^{-mathrm{ln},frac{{lambda }_{0}}{{lambda }_{i}}t},{mathop{-hskip -4.3pt-hskip -4.3ptlongrightarrow}limits_{tto infty}},0$$
(7)

and exemplified in the two insets in Fig. 1. The longest characteristic time is associated with the difference in magnitude between the two largest eigenvalues ({t}_{*}=1/{mathrm{ln}},frac{{lambda }_{0}}{| {lambda }_{1}| }). Consequently, for t ≫ ({t}_{*}) we observe an extreme dimensional reduction from an initial condition characterized in general by Nc degrees of freedom to a final one with just one, set by the environment’s leading eigenvector. Using the values of migration flows in the US urban system over the last decade, we can compute the value of ({t}_{*}). It is rather long—of the order of several centuries—when measured using census data44 and a little shorter using data on tax returns45.
These results allow us to consider very general classes of dynamics and initial conditions. We therefore conclude that in general, given a specific set of demographic conditions, the total population can grow exponentially with a common growth rate across cities. In addition, the relative population distribution at late times does not resemble anything like Zipf’s law. As expected from the statements of the ergodic theorems, the population structure is set by the nature of the intercity flows (and vital rates), or equivalently by the environment A.
Stochastic demographic dynamics and symmetry breaking
The step towards a statistical solution for the city size distribution—and obtaining Zipf’s law in particular—requires the additional consideration of statistical fluctuations, which must arise in the vital and migration rates. Just as in models of statistical physics, the introduction of stochasticity implies not just fluctuating quantities but potentially also the restoration of broken symmetries46,47. In the context of urban systems, restoring broken symmetries means removing preferences for population growth in specific cities over others, associated with imbalances in vital rates and migration flows. To make these expectations explicit we write the migration rates mij as

$${J}_{ij}={m}_{ij}{N}_{i}(t)=left(frac{{s}_{ij}+{delta }_{ij}}{2}right)frac{{N}_{j}(t)}{{N}_{{rm{T}}}(t)}{N}_{i}(t).$$
(8)

Eq. (8) splits the migration flows into two parts: sij describes the symmetric flow between two cities (sij = sji), and δij the anti-symmetric part (δij = −δji). Note that any correlations between Ni and Nj are preserved in these two quantities. Making the population size of the destination city explicit will allow to diagonalize Eq. (2) and therefore to find self-consistent solutions for the structure vector. This form of the migration currents also makes contact with the gravity law of geography, which is typically written as

$${J}_{ij}={G}_{g}frac{{N}_{i}{N}_{j}}{{d}_{ij}^{gamma }}={J}_{ji}.$$
(9)

In this form, the gravity law is symmetric, corresponding to Eq. (8) when δij = 0 and sij = GgNT(t)f(dij), where Gg is the ‘gravitational’ constant, (f({d}_{ij})=1/{d}_{ij}^{gamma }) is a decaying function of distance dij and γ is the distance exponent. Only the anti-symmetric part contributes to the dynamics of population size change, since we can write Eq. (2) as

$${N}_{i}(t+1)=left[1+{upsilon }_{i}-{overline{delta }}_{i}right]{N}_{i}(t).$$
(10)

with ({overline{delta }}_{i}=mathop{sum }nolimits_{j = 1}^{{N}_{{rm{c}}}}{delta }_{ij}{x}_{j}). We can now appreciate the importance of the bi-linearity of the migration flows on both the origin and destination population as a means to diagonalize the demographic dynamics. We have achieved this however at the cost of introducing a slow non-linearity in each city’s growth rate, via their dependence on the population structure vector x. We can establish the existence of a self-consistent solution x* such that for long times

$${upsilon }_{i}-{overline{upsilon }}^{* }=sum_{j = 1}^{{N}_{{rm{c}}}}{delta }_{ij}{x}_{j}^{* },$$
(11)

where ({overline{upsilon }}^{* }=mathop{sum }nolimits_{i}^{{N}_{{rm{c}}}}{upsilon }_{i}{x}_{i}^{* }), since (mathop{sum }nolimits_{i,j = 1}^{{N}_{{rm{c}}}}{delta }_{ij}{x}_{i}{x}_{j}=0), by the conservation of the migration currents. Note that we did not assume any explicit form for the migration flows as a function of distance between places, so that our discussion includes many specific models, including different version of the gravity law.
This equation has a clear geometric meaning: The matrix δ = [δij] is anti-symmetric and real, so that it behaves as an infinitesimal generator of rotations: R(x) = x + δx. We can write the self-consistent solution in terms of these rotation as

$$R({{bf{x}}}^{* })={{bf{x}}}^{* }+({boldsymbol{upsilon }}-overline{upsilon }),$$
(12)

where (overline{upsilon }) is still a function of x and υ = [υi]. Note that these rotations are associated with very small angles, and that the structure vector is subject to constraints, such as staying positive. The first two terms of Eq. (12) ask for a general solution for x that is invariant under rotations. However, the last term, which introduces differences in relative growth rates for different cities, breaks this symmetry explicitly. This specificity of the relative growth rates leads to particular solutions that do not coincide with Zipf’s law. Fig. 2a shows the example of a numerical solution in a non-linear, non-stochastic environment, defined by Eq. (8). The final population structure in this environment is still well defined, as it reaches a steady state. However, compared to the time independent case (see Fig. 1), the final structure takes much longer to unfold and has a stronger urban hierarchy.
Fig. 2: City size distribution emerging from stochastic demographic dynamics differ significantly from those in time independent environments.

a depicts the temporal trajectory for a set of 100 cities in a non-stochastic, non-linear environment, described by Eq. (8). Despite generalizing the demographic dynamics in Fig. 1, this results in a fixed urban hierarchy at late times (vertical red line), regardless of initial conditions. However, when migration and vital rates become stochastic, the demographic dynamics no longer has a static solution at long times, b. Instead, cities constantly change their relative sizes (rank) over time. After an initial transient, the population structure fluctuates close to Zipf’s law. The insets show the population structure (red) at the point in time when the non-stochastic evolution (a) becomes approximately stationary, with Zipf’s law (dashed blue line) shown for reference.

Full size image

Symmetry restoration and the emergence of Zipf’s law
The stage is now set to consider what it takes to counter the symmetry breaking resulting from the selective structure of the environment. To do that, we must consider the statistics of the demographic rates. We start by rewriting Eq. (10) in terms of the dynamical equations for the structure vectors x as

$${x}_{i}(t+1)=left[1+{upsilon }_{i}-overline{upsilon }-{overline{delta }}_{i}right]{x}_{i}(t)=(1+{epsilon }_{i}){x}_{i}(t).$$
(13)

We now specify the relative growth rate ({epsilon }_{i}={upsilon }_{i}-overline{upsilon }-{overline{delta }}_{i}) for each xi as a statistical quantity. It is clear by construction that its average over cities vanishes, ({overline{epsilon }}_{i}=0). The only remaining question is how ϵi varies over time. There is strong empirical evidence that on the time scale of years to decades some cities can maintain larger growth rates than others. For example, presently in the US, most cities of the Southwest and South are growing fast, while other cities show slower growth or relative decay, such as many urban areas in the Midwest and Appalachia48. Moreover, it has been true in recent decades that mid-sized cities in the US have been growing faster than either the largest cities or small micropolitan areas, so that even population aggregated growth rates over certain size ranges differ. Going further back in time one could argue that the same cities of the Midwest were once booming as they became the manufacturing centers of the US, especially between the Civil War and the decades post WWII49. Similar arguments may be made for cities in other urban systems, such as in Europe50,51. Thus, it is possible that on very long time scales, T—on the order of a century or longer—the growth rates of cities equalize to very similar numbers and therefore the temporal average (langle {epsilon }_{i}rangle =frac{1}{T}mathop{sum }nolimits_{t}^{T}{epsilon }_{i}(t)to 0). This requires that the temporal average of Eq. (11) remains zero, which means that the structure vector is both rotated and dilated ’randomly’ in ways that over sufficiently long times lead to 〈ϵi〉 = 0, for each city. Numerical solutions of the demographic equations under these conditions show a population structure that closely resembles Zipf’s law for long times, see Fig. 2b.
In the following, we show that Zipf’s law is a probability density for this derived dynamics in the long run. To do this, let us characterize the variance in the temporal growth rate fluctuations as ({sigma }_{i}^{2}=langle {epsilon }_{i}^{2}rangle ={sigma }_{{upsilon }_{i}-upsilon }^{2}+{sigma }_{{overline{delta }}_{i}}^{2}+2{rm{COV}}({upsilon }_{i}-upsilon ,{overline{delta }}_{i}) , > , 0). Over the same long time scales, statistical fluctuations in the growth rate may obey a limit theorem, so that they become approximately Gaussian, with variance, ({sigma }_{i}^{2}). Then, Eq. (13) becomes simpler, and acquires a direct correspondence to familiar stochastic growth processes. It can be written in analogy to geometric Brownian motion without drift as

$$d{x}_{i}={epsilon }_{i}{x}_{i}={x}_{i}{sigma }_{i}dW$$
(14)

where W is a standard Brownian motion. We now see how demographic dynamics can be simplified through a chain of assumptions into a form consistent with typical stochastic processes leading to Zipf’s law as proposed by Gabaix20,52.
Equation (14) can be expressed as an equation for the probability density P ≡ P[xi, t∣xi(0), 0] of observing xi at time t, having started with an initial condition where xi(0) was observed at time zero,

$$frac{d}{dt}P=frac{{d}^{2}}{d{x}_{i}^{2}}{sigma }_{i}^{2}{x}_{i}^{2}P=frac{d{J}_{{x}_{i}}}{d{x}_{i}}.$$
(15)

The last term describes a probability current, ({J}_{{x}_{i}}=frac{d}{d{x}_{i}}{sigma }_{i}^{2}{x}_{i}^{2}P({x}_{i})). This equation is exactly solvable, see Methods for technical details. To find the stationary solutions, we ask that the right hand side of Eq. (15) vanishes. The first of the two solutions corresponds to a vanishing current, ({J}_{{x}_{i}}=frac{d}{d{x}_{i}}{sigma }_{i}^{2}{x}_{i}^{2}P({x}_{i})=0) and is (P=frac{c^{prime} }{{sigma }_{i}^{2}{x}_{i}^{2}}), where (c^{prime} ) is a normalization constant. If ({sigma }_{i}^{2}={sigma }^{2}) independent of x, as proposed by Gibrat’s law, we can drop the index so this becomes Zipf’s law, (P(N)={P}_{{rm{z}}}({N}_{i}) sim frac{c}{{N}^{2}}). Note for example that if ({sigma }_{i}^{2} sim {N}^{-alpha }) (α may be negative), which violates Gibrat’s law in terms of fluctuations of the growth rate, then P ~ 1/N2−α. Thus, the city size dependence of growth rate fluctuations will change the exponent in Zipf’s law and may destroy scale-invariance altogether if it is not a power law. The second solution applies for constant current, i. e. Jx = J, which leads to (P=frac{c^{primeprime} }{{sigma }^{2}x}), where c″ is a normalization constant. This represents a flow of probability across the urban hierarchy. Both solutions are attractors of the stochastic dynamics that emerge for long times, t ≫ 1/σ2.
We find that in the limit where the relative difference in vital rates vanishes due to fluctuations, the structure vector becomes rotationally invariant on average and the angular symmetry of x is effectively restored. This leads to an effective symmetry of the relative city size distribution on the time scale tr = 1/2σ2, when growth rate fluctuations vanish. The time tr can be estimated using US data to be typically very long, on the range of many centuries or even millennia. Under these conditions all cities share statistically identical dynamics and can be interchanged, resulting in Zipf’s law as the time average of population size distributions, see Fig. 3. In the same limit, the anti-symmetric components of intercity flows will average out and the gravity law will emerge in its conventional form. However, the observations of city sizes at each time are samples of this distribution and may reflect decade-long observable positive and negative preferences for certain cities. Figure 4 shows this effect for the US urban system using decennial census data from 1790 to 1990. For time periods of a few decades, the temporal average does not visibly converge to Zipf’s law: The blue line in the inset of Fig. 4 depicts this effect. However, the cumulative temporal average of the size distributions starts to approximate Zipf’s law after about five decades. The red line (inset) shows the KL divergence between the cumulative temporal average and Zipf’s law, starting from 1790 to subsequent census.
Fig. 3: Stochastic migration rates restore the symmetry of city growth rates over time, leading to Zipf’s law.

The figures show the temporal solution of the demographic dynamics (Eq. (13)) when rotational symmetry is restored through stochasticity averaged over long times. a Shows the population structure at different times (dashed lines) and the distribution with smallest divergence from Zipf’s law over a long run (red), measured by the smallest observed DKL(P∣Pz). b shows the trajectory of the Kullback-Leibler divergence away from Zipf’s law over time. After a long time period, the population structure approximates Zipf’s law and fluctuates around it. In this regime, at any single time only samples showing some deviations from Zipf’s distribution are observable: it is only on the average over many structure vectors at long times that Zipf’s law emerges as a good characterization of the city size distribution. Fig. 4 shows that this was indeed the case for the US urban system until recently.

Full size image

Fig. 4: The dynamics of the rank-size distribution for cities in the US between 1790 and 1990.

Population structure distributions for the largest 100 cities in the US53 are shown as dashed lines, while the average over all years is shown in red. The inset depicts the DKL(P∣Pz) to Zipf’s law of the population structure for each available year (blue) and of the cumulative average over all structure vectors up the specific year (red). We see that the temporal average steadily approaches Zipf’s law as more years are added. After an initial phase, the average is always closer to Zipf’s law than the distribution in any single year. Note that in recent decades, the population structure vector began consistently deviating from Zipf’s law, leading also to a small divergence of the cumulative temporal averages. This effect is to a large extent the consequence of mid-sized and large metropolitan areas growing faster over this period than the largest cities in the US urban system.

Full size image More

Protected areas
We determined the structural connectivity of the global network of PAs by quantifying intact continuous pathways (areas largely devoid of high anthropogenic pressures) between PAs. Data on PA location and boundary were obtained from the May 2019 World Database of Protected Areas (WDPA)70. We only considered PAs that had a land area of at least 10 km2. As China removed most of its PAs from the public May 2019 WDPA version, we used the April 2018 WDPA for China only, which contained the full set of Chinese PAs at the time. It is important to note that our statistics may differ from those reported by countries and territories due to methodologies and dataset differences used to measure terrestrial area of a country or territory.
Measure of human pressure
We used the latest global terrestrial human footprint (HFP) maps—a cumulative index of eight variables measuring human pressure on the global environment—to calculate the average human pressure between PAs35. While there are other human pressure maps74,75,76, the HFP is a well-accepted dataset that provided a validation analysis using scored pressures from 3114 × 1 km2 random sample plots. The root mean squared error for the 3114 validation plots was 0.125 on the normalized 0–1 scale, indicating an average error of approximately 13%. The Kappa statistic was 0.737, also indicating high concurrence between the HFP and the validation dataset. The HFP 2013 map uses the following variables: (1) the extent of built human environments, (2) population density, (3) electric infrastructure, (4) crop lands, (5) pasture lands, (6) roads, (7) railways, and (8) navigable waterways.
Navigable waterways such as rivers and lakes are included within HFP as they can act as conduits for people to access nature35. In the latest HFP (2013), rivers and lakes are included based on size and visually identified shipping traffic and shore side settlements. Venter et al. treated the great lakes of North America, Lake Nicaragua, Lake Titicaca, Lake Onega, Lake Peipus, Lake Balkash, Lake Issyk Kul, Lake Victoria, Lake Tanganyika and Lake Malawi, as they did navigable marine coasts (i.e. only considered coasts as navigable for 80 km either direction of signs of a human settlement, which were mapped as a night lights signal with a Digital Number (DN)  > 6 within 4 km of the coast35). Rivers were included if their depth was >2 m and there were night-time lights (DN  > = 6) within 4 km of their banks, or if contiguous with a navigable coast or large inland lake, and then for a distance of 80 km or until stream depth is too shallow for boats. To map rivers and their depth, Venter et al. used the hydrosheds (hydrological data and maps based on shuttle elevation derivatives at multiple scales) dataset on stream discharge, and the following formulae: stream width = 8.1× (discharge[m3/s])0.58; and velocity = 4.0 × (discharge[m3/s])0.6/(width[m]); and cross-sectional area = discharge/velocity; and depth = 1:5× area/width35.
Each human pressure was scaled from 0–10, then weighted within that range according to estimates of their relative levels of human pressure following Sanderson et al.77. The resulting standardized pressures were then summed together to create the HFP maps for all non-Antarctic land areas35.
Within the main manuscript, we defined intact land as any 1 km2 pixel with a HFP value not higher than or equal to 4. Within this threshold, all areas with a HFP score higher than 4 are defined as nonintact. While previous analyses showed that a >4 score is a key threshold above which species extinction risk greatly increases39, we recognized that there is no one true threshold, which impacts all species equally. Some species may require no human pressure to successfully disperse, while others might successfully navigate through more intensively modified landscapes. Therefore, we conducted our analyses for two additional HFP thresholds. The first used a HFP score More

1.
Mitchard, E. T. A. The tropical forest carbon cycle and climate change. Nature 559, 527–534 (2018).
ADS  CAS  PubMed  Google Scholar 
2.
Saatchi, S. S. et al. Benchmark map of forest carbon stocks in tropical regions across three continents. Proc. Natl Acad. Sci. USA 108, 9899–9904 (2011).
ADS  CAS  PubMed  Google Scholar 

3.
Baccini, A. et al. Estimated carbon dioxide emissions from tropical deforestation improved by carbon-density maps. Nat. Clim. Change 2, 182–185 (2012).
ADS  CAS  Google Scholar 

4.
Liu, Y. Y. et al. Recent reversal in loss of global terrestrial biomass. Nat. Clim. Change 5, 470–474 (2015).
ADS  Google Scholar 

5.
Harris, N. L. et al. Baseline map of carbon emissions from deforestation in tropical regions. Science 336, 1573–1576 (2012).
ADS  CAS  PubMed  Google Scholar 

6.
Marco, M. D., Watson, J. E. M., Currie, D. J., Possingham, H. P. & Venter, O. The extent and predictability of the biodiversity–carbon correlation. Ecol. Lett. 21, 365–375 (2018).
PubMed  Google Scholar 

7.
Giardina, F. et al. Tall Amazonian forests are less sensitive to precipitation variability. Nat. Geosci. 11, 405–409 (2018).
ADS  CAS  Google Scholar 

8.
Erb, K.-H. et al. Unexpectedly large impact of forest management and grazing on global vegetation biomass. Nature 553, 73–76 (2018).
ADS  CAS  PubMed  Google Scholar 

9.
Zarin, D. J. et al. Can carbon emissions from tropical deforestation drop by 50% in 5 years? Glob. Change Biol. 22, 1336–1347 (2016).
ADS  Google Scholar 

10.
Chaplin-Kramer, R. et al. Degradation in carbon stocks near tropical forest edges. Nat. Commun. 6, 10158 (2015).
ADS  CAS  PubMed  PubMed Central  Google Scholar 

11.
Brandt, M. et al. Satellite passive microwaves reveal recent climate-induced carbon losses in African drylands. Nat. Ecol. Evol. 2, 827 (2018).
PubMed  Google Scholar 

12.
Fan, L. et al. Satellite-observed pantropical carbon dynamics. Nat. Plants 5, 944–951 (2019).
CAS  PubMed  Google Scholar 

13.
Mitchard, E. T. et al. Markedly divergent estimates of Amazon forest carbon density from ground plots and satellites. Glob. Ecol. Biogeogr. 23, 935–946 (2014).
PubMed  PubMed Central  Google Scholar 

14.
Mitchard, E. T. et al. Uncertainty in the spatial distribution of tropical forest biomass: a comparison of pan-tropical maps. Carbon Balance Manag. 8, 10 (2013).
PubMed  PubMed Central  Google Scholar 

15.
Avitabile, V. et al. An integrated pan-tropical biomass map using multiple reference datasets. Glob. Change Biol. 22, 1406–1420 (2016).
ADS  Google Scholar 

16.
Réjou-Méchain, M. et al. Upscaling forest biomass from field to satellite measurements: Sources of errors and ways to reduce them. Surv. Geophys. 40, 881–911 (2019).
ADS  Google Scholar 

17.
Saatchi, S. Mapping tropical forest biomass: synthesis of ground and remote sensing inventory. Consult. Rep. 2 High Carbon Stock Sci. Study (2015).

18.
Ploton, P. et al. A map of African humid tropical forest aboveground biomass derived from management inventories. Sci. Data 7, 221 (2020).
PubMed  PubMed Central  Google Scholar 

19.
Philippon, N. et al. The light-deficient climates of Western Central African evergreen forests. Environ. Res. Lett. 14, 034007 (2018).

20.
Saatchi, S. et al. Seeing the forest beyond the trees. Glob. Ecol. Biogeogr. 24, 606–610 (2015).
Google Scholar 

21.
Mermoz, S., Le Toan, T., Villard, L., Réjou-Méchain, M. & Seifert-Granzin, J. Biomass assessment in the Cameroon savanna using ALOS PALSAR data. Remote Sens. Environ. 155, 109–119 (2014).
ADS  Google Scholar 

22.
Lewis, S. L. et al. Above-ground biomass and structure of 260 African tropical forests. Philos. Trans. R. Soc. B Biol. Sci. 368, 20120295 (2013).
Google Scholar 

23.
Hansen, M. C., Potapov, P. & Tyukavina, A. Comment on “Tropical forests are a net carbon source based on aboveground measurements of gain and loss”. Science 363, eaar3629 (2019).
CAS  PubMed  Google Scholar 

24.
Baccini, A. et al. Tropical forests are a net carbon source based on aboveground measurements of gain and loss. Science 358, 230–234 (2017).
ADS  MathSciNet  CAS  PubMed  PubMed Central  MATH  Google Scholar 

25.
Breiman, L. Random forests. Mach. Learn 45, 5–32 (2001).
MATH  Google Scholar 

26.
Lyapustin, A., Wang, Y., Korkin, S. & Huang, D. MODIS Collection 6 MAIAC algorithm. Atmos. Meas. Tech. 11, 5741–5765 (2018).

27.
Fick, S. E. & Hijmans, R. J. WorldClim 2: new 1‐km spatial resolution climate surfaces for global land areas. Int. J. Climatol. 37, 4302–4315 (2017).
Google Scholar 

28.
Kühn, I. Incorporating spatial autocorrelation may invert observed patterns. Divers. Distrib. 13, 66–69 (2007).
Google Scholar 

29.
Dormann, C. F. Effects of incorporating spatial autocorrelation into the analysis of species distribution data. Glob. Ecol. Biogeogr. 16, 129–138 (2007).
Google Scholar 

30.
Roberts, D. R. et al. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40, 913–929 (2017).
Google Scholar 

31.
Valavi, R., Elith, J., Lahoz‐Monfort, J. J. & Guillera‐Arroita, G. blockCV: an r package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models. Methods Ecol. Evol. 10, 225–232 (2019).
Google Scholar 

32.
Parmentier, I. et al. Predicting alpha diversity of African rain forests: models based on climate and satellite-derived data do not perform better than a purely spatial model. J. Biogeogr. 38, 1164–1176 (2011).
Google Scholar 

33.
Baccini, A., Walker, W., Carvalho, L., Farina, M. & Houghton, R. A. Response to Comment on “Tropical forests are a net carbon source based on aboveground measurements of gain and loss”. Science 363, eaat1205 (2019).
CAS  PubMed  MATH  Google Scholar 

34.
Xu, L., Saatchi, S. S., Yang, Y., Yu, Y. & White, L. Performance of non-parametric algorithms for spatial mapping of tropical forest structure. Carbon Balance Manag. 11, 18 (2016).
PubMed  PubMed Central  Google Scholar 

35.
Dormann, C. F. et al. Methods to account for spatial autocorrelation in the analysis of species distributional data: a review. Ecography 30, 609–628 (2007).
Google Scholar 

36.
Irwin, A. The ecologist who wants to map everything. Nature 573, 478–481 (2019).
ADS  PubMed  Google Scholar 

37.
Legendre, P. Spatial autocorrelation: trouble or new paradigm? Ecology 74, 1659–1673 (1993).
Google Scholar 

38.
Meyer, H., Reudenbach, C., Wöllauer, S. & Nauss, T. Importance of spatial predictor variable selection in machine learning applications–Moving from data reproduction to spatial prediction. Ecol. Model. 411, 108815 (2019).
Google Scholar 

39.
Strobl, C., Boulesteix, A.-L., Kneib, T., Augustin, T. & Zeileis, A. Conditional variable importance for random forests. BMC Bioinformatics 9, 307 (2008).
PubMed  PubMed Central  Google Scholar 

40.
Lefsky, M. A. et al. Estimates of forest canopy height and aboveground biomass using ICESat. Geophys. Res. Lett. 32, L22S02 (2005).

41.
Réjou-Méchain, M. et al. Upscaling Forest biomass from field to satellite measurements: sources of errors and ways to reduce them. Surv. Geophys. 40, 881–911 (2019).

42.
Mitchard, E. T. A. et al. Comment on ‘A first map of tropical Africa’s above-ground biomass derived from satellite imagery’. Environ. Res. Lett. 6, 049001 (2011).
ADS  Google Scholar 

43.
Asner, G. P. et al. High-resolution carbon mapping on the million-hectare Island of Hawaii. Front. Ecol. Environ. 9, 434–439 (2011).
Google Scholar 

44.
Asner, G. P. et al. Human and environmental controls over aboveground carbon storage in Madagascar. Carbon Balance Manag. 7, 2 (2012).
PubMed  PubMed Central  Google Scholar 

45.
Asner, G. P. et al. High-resolution mapping of forest carbon stocks in the Colombian Amazon. Biogeosciences 9, 2683 (2012).
ADS  CAS  Google Scholar 

46.
Asner, G. P. et al. Mapped aboveground carbon stocks to advance forest conservation and recovery in Malaysian Borneo. Biol. Conserv. 217, 289–310 (2018).
Google Scholar 

47.
Xu, L. et al. Spatial distribution of carbon stored in forests of the Democratic Republic of Congo. Sci. Rep. 7, 15030 (2017).
ADS  PubMed  PubMed Central  Google Scholar 

48.
Schepaschenko, D. et al. The Forest Observation System, building a global reference dataset for remote sensing of forest biomass. Sci. Data 6, 1–11 (2019).
Google Scholar 

49.
Chave, J. et al. Ground data are essential for biomass remote sensing missions. Surv. Geophys. 40, 863–880 (2019).
ADS  Google Scholar 

50.
van den Hoogen, J. et al. Soil nematode abundance and functional group composition at a global scale. Nature 572, 194–198 (2019).
ADS  CAS  PubMed  Google Scholar 

51.
Steidinger, B. S. et al. Climatic controls of decomposition drive the global biogeography of forest-tree symbioses. Nature 569, 404–408 (2019).
ADS  CAS  PubMed  Google Scholar 

52.
Bastin, J.-F. et al. The global tree restoration potential. Science 365, 76–79 (2019).
ADS  CAS  PubMed  Google Scholar 

53.
Trabucco, A. & Zomer, R. J. Global aridity index (global-aridity) and global potential evapo-transpiration (global-PET) geospatial database. CGIAR Consort Spat Information (2009).

54.
Wilson, A. M. & Jetz, W. Remotely sensed high-resolution global cloud dynamics for predicting ecosystem and biodiversity distributions. PLoS Biol. 14, e1002415 (2016).
PubMed  PubMed Central  Google Scholar 

55.
Bowman, D. M., Williamson, G. J., Keenan, R. J. & Prior, L. D. A warmer world will reduce tree growth in evergreen broadleaf forests: evidence from A ustralian temperate and subtropical eucalypt forests. Glob. Ecol. Biogeogr. 23, 925–934 (2014).
Google Scholar 

56.
Malhi, Y. et al. Exploring the likelihood and mechanism of a climate-change-induced dieback of the Amazon rainforest. Proc. Natl Acad. Sci. USA 106, 20610–20615 (2009).
ADS  CAS  PubMed  Google Scholar 

57.
Rennó, C. D. et al. HAND, a new terrain descriptor using SRTM-DEM: Mapping terra-firme rainforest environments in Amazonia. Remote Sens. Environ. 112, 3469–3481 (2008).
ADS  Google Scholar 

58.
Nachtergaele, F., Velthuizen, H. V., Verelst, L. & Wiberg, D. Harmonized World Soil Database (HWSD) (Food and Agriculture Organization, U. N. Rome, 2009).

59.
Defourny, P. et al. Algorithm Theoretical Basis Document for Land Cover Climate Change Initiative. Technical report (European Space Agency, 2014).

60.
Segal, M. & Xiao, Y. Multivariate random forests. WIREs Data Min. Knowl. Discov. 1, 80–87 (2011).
Google Scholar 

61.
CCI, ESA. New Release of 300 m Global Land Cover and 150 m Water Products (v.1.6.1) and new version of the User Tool (3.10) for Download (ESA CCI Land cover website, 2016). More

Global P losses from soils and soil P balances
All continents result in negative P balances (e.g., net P losses from agricultural systems, Table 1; Fig. 2) except Asia, Oceania and Australia, with Asia having a slightly positive but near zero P balance. This is in spite of high to very high chemical fertilizer inputs (with a range of 1.7 to 13 kg ha−1 yr−1 between the different continents, with national values reaching up to 14 and 19 kg ha−1 yr−1 for the European Union 15 old Member States (EU15, member states joining before 2004 mainly in western and northern Europe) and China, respectively). Most negative P balances are indicated for Africa due to very low chemical fertilizer input of 1.7 kg ha−1 yr−1 paired with high losses due to soil erosion of 9.6 kg ha−1 yr−1. South America as well as Central and Eastern Europe (NEU11; which are the new member states joining the EU after 2004 with the exception of Cyprus and Malta) also exhibit high P losses but for different reasons. South America has a very high chemical fertilizer input but also high losses due to soil erosion paired with high P exports due to organic P management (calculated as the sum of manure and residue input minus plant uptake). In contrast, the eastern European Union New Member States (NEU11) have rather low erosional losses but also very low chemical fertilizer input. With the hypothetical assumption of no replenishment due to chemical fertilizer (e.g., due to economic or technical constraints), calculation of soil P balances results in negative balances globally, as well as for all the continents and regions considered (depletion between 4 and 20 kg P ha−1 yr−1; Figs. 3 and 4; Table 1). The latter demonstrates the vulnerability of today’s global land management system and its strong dependency on chemical P fertilizers from non-renewable mineable P deposits.
Table 1 Main phosphorus (P) statistics in kg P ha−1yr−1 for all continents and selected countries.
Full size table

Fig. 2: Global average phosphorus (P) losses due to soil erosion in kg ha−1 yr−1.

The chromatic scale represents the P losses estimates, while the gray color indicates the cropland areas that were excluded from the modeling due to data unavailability. Note that classes are not regularly scale ranked but are divided into six classes using the quantile classification method. Only plant available fractions were considered. For the more residual P fractions please refer to Table 1 or Figs. 3 and 4).

Full size image

Fig. 3: Global P soil pools and depletion due to erosion.

Arrows indicate fluxes (positive: net input to soils, negative: depletion of soils). *Organic P management = sum of manure and residue input minus plant uptake. Non-plant P = non-plant available P. Inorganic and organic P give plant available fractions. Total soil P: sum of P fractions lost from soil via erosion with relative errors. No/with chemical = P balance with and without chemical fertilizer.

Full size image

Fig. 4: Soil P pools and depletion due to erosion in Africa, Europe and North America.

AD = Atmospheric Deposition. CF = Chemical Fertilizer. OM = Organic P management = sum of manure and residue input minus plant uptake. Arrows indicate fluxes (positive: net input to soils, negative: depletion of soils). Non-plant P = non-plant available P. Inorganic and organic P give plant available fractions. Soil Plost: sum of P fractions lost from soil via erosion with relative errors. No/with chemical = P balance with and without chemical fertilizer.

Full size image

Our area related calculations result in an average P loss for arable soils, due to erosion by water, of approximately (5.9_{ – 0.79% }^{ + 1.17% }) kg ha−1 yr−1 globally (Fig. 3, Table 1). This is around 60% of the rates given by Smil16 who estimated 10 kg P ha−1 yr−1 from arable fields due to soil erosion by water. Our total P losses due to soil erosion by water from arable soils globally result in 6.3 Tg yr−1 with 1.5 Tg yr−1 for organic and 4.8 Tg yr−1 for inorganic P. With these values the results are at the lower end of the range discussed in the literature (between 1–19 Tg yr−1, Table 2).
Table 2 Global fluxes from soils/arable systems to waters as discussed in recent literature. For comparability, all values were normalized to 1 billion ha of arable land (with original values given in brackets).
Full size table

Liu34 in estimating net input to, and output from, cropland systems indicated a net loss of P from the world’s croplands of about 12.8 Tg  yr−1 (calculating input from atmosphere, weathering and chemical fertilizer versus output from organic P management, soil erosion and runoff), which would be, according to their calculations, the same order of magnitude as synthetic fertilizer input (13.8 Tg yr−1 for statistical year 2003/2004). Our calculations result in an approximate net soil loss due to erosion of 6.3 Tg yr−1 with a global average chemical fertilizer input of 9.2 Tg yr−1. The fluxes clearly show the critical dependence on chemical fertilizer globally, with a hypothetical net-average-area related depletion of 10.7 kg ha−1 yr−1 globally without the compensation due to chemical fertilizer, from which a loss of 5.2 kg ha−1 yr−1 stems from organic P management (sum of manure and residue input minus plant uptake) and 5.9 kg ha−1 yr−1 from soil erosion (Table 1). Continental and national erosional P losses are between 40 and 85% of total P losses from agricultural systems with the exception of Europe and Australia (16 and 19%, respectively). Globally, as well as for Africa, South America and Asia, P soil losses due to erosion are higher than losses due to organic P management.
A recent quantification of atmospheric P dust input, based on dust measurements in the Sierra Nevada, concluded that measured dust fluxes are greater than, or equal to, modern erosional outputs and a large fractional contribution relative to bedrock35. However, even though their measured and modeled maximum atmospheric P fluxes were in the same order of magnitude as the atmospheric flux data of Wang et al.36 used in this study (0.1 kg ha−1 yr−1 for North America), erosion rates from sediment trap measurements in their investigated forest ecosystems were considerably lower (maximum of 0.06 kg ha−1 yr−135) than might be expected in arable lands worldwide. As such, we can clearly not agree with the above conclusion.
Mitigation of the soils P status in the long term by decreasing the deficits of the current organic P management seems difficult and rather unlikely in many regions of the world (see discussion of continental balances below). As such, and considering the expected shortage of P supply from industrial fertilizers in the future, the evaluation of P fluxes clearly shows that soil erosion has to be limited to the feasible absolute minimum in the future.
Only a small fraction of the total soil P is plant available, because a large fraction is either bound in, adsorbed to, or made unavailable by occlusion in minerals (apatite and occluded P)37. Our modeling approach follows the approach of Yang et al.37 that builds on existing knowledge of soil P processes and data bases to provide spatially explicit estimates of different forms of naturally occurring soil P on the global scale. Yang et al.37 uses data acquired with the Hedley fractionation method38 which splits soil P into different fractions that are extracted sequentially with successively stronger reagents and which are merged into various inorganic and organic pools. There is great uncertainty to associate these fractions with functional plant uptake39,40 but some recent work aims at quantifying residence times of the different fractions17,41. We present the total P loss as well as the plant-available pool (Figs. 3 and 4, Table 1) as the most labile P which can participate to plant nutrition in a time scale up to months17 (i.e., the so-called labile inorganic P, inorganic P bound to secondary minerals, labile and stable organic P33,37). We contrast these more labile, short lived fractions with fractions considered very stable which would not be plant available at short time scales (i.e., inorganic P associated with minerals such as apatite or occluded P17, corresponding to 52% of total soil P in our estimates at the global scale).
Verification of soil P loss with a comparison to riverine P exports
Verification of our proposed P soil losses might be done indirectly via a comparison to riverine P loads. However, to verify our on-site P soil loss approach with off-site P river export data requires two prerequisites: (i) an estimation of the agricultural erosion and runoff contribution to total P loads in the rivers (as the latter will be the total sum of agricultural, urban and industrial runoff) and (ii) an assumption on sediment delivery rates (e.g., the percent of sediments reaching the rivers from the total eroded sediments) to estimate P loads to rivers from on-site P soil loss. Regarding the first prerequisite, we searched for published river P export data with a separation of the agricultural erosion and runoff contribution from total P loads in the rivers. Regarding the second prerequisite, there is a lack of models describing integrated sediment-delivery to rivers on a continental or global scale, but it has been discussed that sediment delivery ratios generally decrease as drainage area increases, ranging from roughly 30–100% in small catchments (≤0.1 km2) to 2–20% at large spatial scales (e.g., ≥1000 km2)42. For our comparison, we used a range of average sediment delivery rates between 11–30% as used in recent large scale studies from continental to global scale43,44. In doing so, we would like to point out, that even if 70–89% of the P lost from agricultural soils might be re-deposited within catchments, potential threats of P loss from soils are not reduced. Erosion (and thus P loss) occurs predominantly on agricultural soils while re-deposition will mostly occur in depositional hollows, wetlands, riparian zones or buffer strips. Thus, P is lost as a nutrient on food and feed production sites but re-deposited as a potential ecological threat to biodiversity and ecosystem health due to its eutrophication effect in less intensively or unmanaged ecosystems. Last but not least we would like to point out that RUSLE only considers soil displacement due to rill and inter-rill erosion neither considering tillage and gully erosion nor land sliding.
A comparison of calculated potential P export to rivers from our on-site soil P losses is well within the range of published riverine P exports (Table 3). Beusen et al.45 used total suspended sediment measurements from the GEMS-GLORI database to extrapolate spatially distributed sediment rates for the world’s largest rivers with land use, topography, lithology and precipitation as factors in a multiple linear regression approach accounting for soil erosion as well as sediment trapping. The associated nutrient exports for all continents, as well as global assessment were made by calibrating nutrient export to sediment rates with the model Global News45 using established correlations between sediment and nutrient concentrations. In comparing their P export with suspended sediments in rivers45 to our assessments, we underestimated P export globally as well as for all continents, with the exception of Africa (Table 3), which is, however, (i) strongly related to assumed sediment delivery rates and (ii) our RUSLE application only considering rill and inter-rill erosional processes with an unknown contribution of gullies, landslides and tillage erosion.
Table 3 On-site P loss from gross soil erosion (this study), calculated potential riverine loads with sediment delivery ratios between 11–30%43,44 and comparison to global and regional riverine P export studies. All values in kg P ha−1yr−1.
Full size table

An analysis of 17 large scale European catchments (250–11,000 km2) quantifying the loss of P to surface waters from an off-site perspective (P flux measurement in waters) resulted in 0.05–1.5 kg ha−1 yr−1 of P loss due to soil erosion and runoff from agricultural lands into streams and rivers (excluding a Greek catchment with a very high P export of 6 kg ha−1 yr−1)46. Recalculating our on-site soil P losses with sediment delivery ratios between 11–30% results in rates for geographic Europe between 0.1–0.4 kg ha−1 yr−1 which are at the lower end of the range of Kronvang, et al.46 (Table 3). Potential P losses as riverine export due to agricultural runoff and erosion of 143 watersheds across the U.S.47 are in the same range as the calculated P loss assessment of our study, while P loads to Lake Erie assessed in a regional study48 seem slightly higher (Table 3). However, no partitioning between agricultural, urban or industrial fluxes was possible from the latter study. An assessment of the Yangtze River Basin (with 1.8 Million km2 near 20% of the whole Chinese territory) gives modeled soil P losses (on-site perspective) between 0–196 kg ha−1 yr−1 demonstrating the huge spatial heterogeneity of on-site soil erosion rates49. The model output was calibrated with total measured nutrient loads in rivers, while the partitioning differed between dissolved point and non-point as well as adsorbed non-point pollution which can be mostly attributed to soil erosion of agricultural fields. The average P loads due to soil erosion from agricultural fields in the Yangtze River Basin compares well with the range assessed in our study for China (2.7 versus a range of 1.4 to 3.7 kg ha−1 yr−1, respectively). The same holds true for assessment comparisons of Africa50 and South America51 (Table 3), even though it should be considered that especially for these latter studies scale differs considerably from our approach.
Regional P losses and balances
Parallel to the distribution pattern and dynamics of global soil erosion by water32, P losses from soils due to water erosion are most dramatic in countries and regions with intensive agriculture and/or extreme climates (e.g., droughts followed by significant rain events or high frequencies of heavy rain storms) due to high erosivity effects52. As such, our calculations result in extremely high P losses due to erosion ( >20 kg ha−1 yr−1) in regions such as eastern China, many regions in Indonesia, parts of east and south-eastern Africa (Ethiopia, Eritrea, Mozambique), Central America and parts of South America (South-Eastern Brazil; Southern Chile, Peru (Fig. 2)). A very high P loss (10 to 20 kg ha−1 yr−1) is estimated for parts of Southern Africa (South Africa, Madagascar, Tanzania) and South America (Bolivia) and a high loss (5–10 kg ha−1 yr−1) for most of India, as well as regions in Southern Africa (Angola, Zambia) and South America (Uruguay) (Fig. 2). Even though the underlying erosion model algorithm does not calculate the net catchment output but rather the on-site displacement of soil sediments which might then be re-located to other parts of the fields or even buried at depositional places, the considered on-site field management will clearly be confronted with substantial P losses due to soil erosion by water. Only considering agronomic P inputs and outputs without including P losses due to erosion by MacDonald et al.22 resulted in a very different global P pattern: most widespread large deficits were in South America (North-Eastern countries, e.g., Argentina and Paraguay), the northern United States and Eastern Europe while the largest surpluses covered most of East Asia, Western and Southern Europe, the coastal United States, South-Eastern Brazil and Uruguay.
With average soil depletion due to erosion of 9.6 kg ha−1 yr−1 in Africa, the overall P balance is already negative by 9.7 kg ha−1 yr−1 today (Table 1, Fig. 4). As the average P depletion in Africa due to negative fluxes in organic P management equals the input fluxes from the atmosphere plus chemical fertilizer, African farmers could decrease P losses to near zero with effective soil erosion mitigation. Even though the system’s P depletion due to organic P management is relatively low in Africa (−2.2 kg ha−1 yr−1) compared to a global average (−5.2 kg ha−1 yr−1), the high overall P losses are unlikely to be covered neither from a mitigated and more sustainable organic P management nor from increased chemical fertilizer input. P fluxes due to organic P management are calculated here as the sum of manure and residue input minus plant uptake (which results in biomass export in arable systems with the exception of residues left on the field). The overall sum of plant uptake is likely to increase with increased need for food and feed parallel to a predicted population and livestock growth in Africa in the future. Many soils in sub-Saharan Africa have already been characterized as deficient for levels of plant-available P for the last decades53. Manure and residue input is simultaneously in demand in Africa today (shortage of biomass in general, low animal production and even if there is manure available, there are no means to transport it to where it is needed), which results in the recommendations of an integrated farm management with combinations of organic and inorganic fertilizers54,55,56. With the inorganic P fertilizers becoming increasingly scarce, the depletion due to organic P management can be expected to increase in Africa in the future. Simultaneously, today’s prices for chemical fertilizer can already be 2–6 times more expensive for a farmer in Africa than in Europe due to higher transport and storage costs3, even though Africa itself has the highest geological P deposits in the world (according to today’s estimates 80% of the global geological P deposits are located in Morocco and the Western Sarah8). As such, and if the political situation does not change dramatically (e.g., that the P supplies are marketed within Africa instead of being exported to US, Europe and China), the only realistic means of reducing P depletion of African soils today, and in the future, is to drastically reduce soil erosion.
We recognize that the values calculated in Table 1 and Fig. 4 are gross estimates over large scales and that spatial context and scale, especially on the African continent is important. P deficiency is a country, district, farm and soil specific issue in Africa, for example parts of east Africa and the Sahel have substantial deficiencies57. In sub Saharan Africa ~40% of soils are considered to have low nutrient reserves ( More

Regionally-downscaled climate change projections
We used a management strategy evaluation (MSE) applied to ensemble projections of a climate-enhanced multispecies stock assessment within the integrated modeling framework of the Alaska Climate Change Integrated Modeling project (ACLIM)19. For this, six high resolution downscaled projections of oceanographic and lower trophic level conditions in the Bering Sea (using the Regional Ocean Modeling System49,50) were coupled to the BESTNPZ nutrient-phytoplankton-zooplankton model51; we refer to this model complex throughout this paper as the Bering10K ROMSNPZ, or just ROMSNPZ, model. Boundary conditions were driven by three global general circulation models (GFDL-ESM2M52, CESM153, and MIROC-ESM54) projected (2006–2099) under the high-baseline emission scenario Representative Concentration Pathway 8.5 (RCP 8.5) and midrange global carbon mitigation (RCP 4.5; note, that for CESM under RCP 4.5, projections from 2080–2100 were unavailable so conditions from 2080–2099 were held constant at 2080 conditions for that scenario only) future scenarios from the Coupled Model Intercomparison Project phase 5 (CMIP5)29,55. Hermann et al.30,56 also report on downscaled hindcasts of oceanographic and lower trophic conditions in the EBS from 1970–2012 (see refs. 30,56,57 for detailed descriptions of model evaluation and performance). For each downscaled model simulation, we replicated the National Marine Fisheries Service Alaska Fisheries Science Center annual summer bottom-trawl survey in time and space in the ROMSNPZ model (using historical mean survey date at each latitude and longitude of each gridded survey station) to derive estimates of sea surface and bottom temperatures (Fig. 1). We additionally used a polygon mask of the survey area to estimate the average zooplankton abundance in the system during spring, summer, winter, and fall months. These indices were derived for each climate projection scenario, as well as a persistence scenario where conditions were held constant at the average of those for 2006–2017 from a hindcast simulation. All index projections were bias corrected to the 2006–2017 hindcast period using the delta method assuming unequal variance in the GCM projections and hindcast58 such that:

$$T_{mathop {{{mathrm{fut}}}}limits^prime ,y} = bar T_{{mathrm{hind}},overrightarrow {scriptstyle{mathrm{ref}}} } + frac{{sigma _{{mathrm{hind}},overrightarrow {scriptstyle{mathrm{ref}}} }}}{{sigma _{{mathrm{fut}}, overrightarrow{scriptstyle{mathrm{ ref}}} }}}left( {T_{{mathrm{fut}},y} – bar T_{{mathrm{fut}},overrightarrow {scriptstyle{mathrm{ref}}} }} right)$$
(1)

where (T_{mathop {{{mathrm{fut}}}}limits^prime ,y}) is the bias-corrected projected timeseries, (T_{{mathrm{fut}},y}) is the raw projected timeseries, (bar T_{{mathrm{hind}},overrightarrow {scriptstyle{mathrm{ref}}} }) is the mean of the hindcast during the reference years (overrightarrow {{mathrm{ref}}}) (2006–2017), (bar T_{{mathrm{fut}},overrightarrow {scriptstyle{mathrm{ref}}} }) is the mean of the raw projected timeseries during the reference years (overrightarrow {{mathrm{ref}}}), (sigma _{{mathrm{hind}},overrightarrow {scriptstyle{mathrm{ref}}} }) is the standard deviation of the hindcast during the reference years (overrightarrow {{mathrm{ref}}}), (sigma _{{mathrm{fut}},overrightarrow {scriptstyle{mathrm{ref}}} }) is the standard deviation of the raw projection timeseries during the reference years (overrightarrow {{mathrm{ref}}}).
Climate-enhanced multispecies stock assessment model
Bias-corrected indices were then used as covariates in the climate-enhanced multispecies stock assessment model for the Bering Sea (hereafter CEATTLE)59 to evaluate the performance of alternative management approaches on future fish biomass and catch. CEATTLE is a climate-enhanced multispecies statistical age-structured assessment model with parameters for growth that are functions of temperature (i.e., temperature-specific average weight-at-age) and predation that are functions of temperature (via a bioenergetics-based predation sub-model)59,60,61. Since 2016, the model has been used operationally in the Bering sea as a supplement to the annual BSAI pollock stock assessment61. Various configurations of CEATTLE are possible; for this study we chose one where temperature-specific predator and prey interactions influenced natural mortality, temperature influenced weight-at-age, and the spawner-recruit relationship was a function of physical and biological future conditions as well as random variability (i.e., a climate-informed multispecies model). We fit the model using penalized maximum likelihood to survey biomass, diet, and fishery harvest data for three groundfish species pollock, Pacific cod, and arrowtooth flounder from the EBS in the EBS over the period 1979–2017. We also used the Bering10K ROMSNPZ model to produce detailed hindcasts of temperature for the period 1970–2017. We used hindcast-extracted timeseries from the ROMSNPZ model and CEATTLE model estimates of recruitment ((R_{i,y,l})) and spawning biomass ((B_{i,y – 1})) in hindcast year y for each species i to fit a climate-enhanced logistic recruitment per spawner model36, such that:

$$ln left( {hat R_{i,y}} right) = alpha _i – beta _{0,i}B_{i,y – 1} + ln left( {B_{i,y – 1}} right) + {{{bf{B}}}}_i{{{bf{X}}}} + varepsilon _{i,y}$$
(2)

where ({{{bf{B}}}}_i{{{bf{X}}_l}}) is the summed product of each covariate parameter (beta _{ij}) and the corresponding environmental covariate (X_{j,y}) for each bias-corrected environmental index (j = ( {1,2…n_j} )). We selected indices representative of ecological conditions important for groundfish recruitment in the Bering sea39; spring and fall large zooplankton abundances, survey replicated bottom temperature, and extent of the residual cold pool of extremely dense and cold sea water that persists across the EBS shelf following spring sea ice retreat. We assumed normally distributed (in log space) residual errors for each species ((varepsilon _{i,y} sim Nleft( {0,sigma _i^2} right))). The CEATTLE model was then projected forward where ROMSNPZ indices from individual projections drove growth, predation, and recruitment in each future simulation year36,62.
Evaluation of harvest management approaches
Previous authors have defined EM (i.e., the incorporation of ecosystem information into marine resource management) as a continuum between two paradigms of management and focus18. On one end is within-sector single-species management that considers ecosystem information (EAFM) and on the other is cross-sectoral whole of ecosystem management (i.e., EBM). EBFM is intermediate between these and is defined by quantitative incorporation of ecosystem interactions into assessment models and target setting (EBFM). Most fisheries management in the Bering Sea can be characterized as EBFM or EAFM, with increasing trends toward cross-sectoral coordination at the scale of EBM. Here we focus on one aspect on this scale of potential management options, operational EBFM and EAFM as captured through the CEATTLE multispecies stock assessment model and harvest policies decisions made annually under the constraint of the 2 MT cap (modeled via the ATTACH model).
MSE is a process of “assessing the consequences of a range of management strategies or options and presenting the results in a way which lays bare the tradeoffs in performance across a range of management objectives”63. MSE has been frequently used to evaluate alternative management strategies based on single-species estimation methods64. It is increasingly used to evaluate ecosystem management performance, although these evaluations are far less commonplace due to the complexity of modeling and assessing the performance of ecosystem level metrics64. Importantly, MSE “does not seek to proscribe an optimal strategy or decision”63, rather it aims to describe the uncertainty and tradeoffs inherent in alternative strategies and scenarios. In this case, through a series of workshops, we worked with managers and stakeholders to identify priority scenarios and outputs19. From this, risk, sensitivity, and uncertainty under contrasting climate scenarios were requested outputs of the analysis, as was the performance of current climate-naive EBFM policies.
A key component of MSE is identifying and quantifying uncertainty (i.e., process, observation, estimation, model, and implementation error) and representing it using an operating model. In the case of this MSE, the focus was on process error uncertainty due to variation in recruitment about the fitted stock-recruitment relationship, one major source of model error in the form of alternative climate scenarios, and implementation error. The MSE does not account for estimation error (uncertainty in the parameters of the operating model) nor observation error. This is because the estimates of recruitment and spawning biomass from CEATTLE for the BSAI are very precise (see Fig. 10 in ref. 60) and the estimation and operating models are therefore very similar. Thus, CEATTLE is the operating model for this MSE and implicitly the estimation method. In this approach we assume that while allowing for observation error would have increased overall error, the effect would have been minor compared to the investigated uncertainties. Future analyses using a full MSE (i.e., separate operating and estimation models) could evaluate the effect of observation error, but perhaps more importantly, the potential for model error, whereby the population dynamics model (on which the estimation method is based) differs from that of the operating model such that the estimates on which management decisions are made are biased relative to the true values in the operating model.
Given this we summarized the relative change in catch and biomass for the three species in the model under the following fishing scenarios (Fig. 1): (a) projections without harvest ((F_{i,y} = 0)) in each year y of scenario l for each species i, (b) projections under target harvest rate (Supplementary Fig. 7 left) and with a sloping harvest control rule (HCR) (Supplementary Fig. 7 right), (c) as in 2 but with the constraint of a 2 MT cap applied dynamically to the three focal species only.
Under the North Pacific Fishery Management Council (NPFMC) constraint of the 2 MT cap on cumulative total annual catch, realized harvest (i.e., catch) and specification of individual species harvest limits known as Total Allowable Catch (TAC; metric tons) are a function of the acceptable biological catch (ABC) for the given species, as well as ABC of other valuable species in the aggregate complex19,34 (https://github.com/amandafaig/catchfunction). TAC must be set at or below ABC for each species, therefore TAC of individual species are traded-off with one another to avoid exceeding the 2 MT cap. From 1981 to 1983, the TAC of pollock was reduced significantly below the ABC and in 1984 the 2 MT cap became part of the BSAI fishery management plan21,34,65. Pacific cod regulations have changed markedly over recent decades and it was only in the 1990s that in many years the catch and TAC approached its ABC. Thus, we used the socioeconomic ATTACH model (the R package ATTACHv1.6.0 is available with permission at https://github.com/amandafaig/catchfunction34) to model realized catch in each simulation year as a function of CEATTLE assessment estimates of ABC (tons) for pollock, Pacific cod, and arrowtooth flounder under future projections (2018–2100). This entailed three steps for each future simulation year (y):
1.
project the population forward from (y – 1) to (y) using estimated parameters from the multispecies mode of the CEATTLE model fit to data from 1979 to 2017 and recruitment based on biomass in simulation year (y) and future environmental covariates from the ROMSNPZ model downscaled projections (see “Methods” above) to determine ({mathrm{ABC}}_{i,y,l}) for each species (i) under each scenario (l) given the sloping harvest control rule for pollock, Pacific cod, and arrowtooth flounder in each simulation year (y);

2.
use ({mathrm{ABC}}_{i,y,l}) of each species from step 1 as inputs to the ATTACH model in order to determine the North Pacific Marine Fishery Council Total Allowable Catch (TACi,y,l) for the given simulation l year y;

3.
use TACi,y,l from step 2 to estimate catch (tons) in the simulation year (Fig. 1); remove catch from the population and advance the simulation forward 1 year.

Determine the annual ABC
We used end-of-century projections (2095–2099) to derive a maximum sustainable yield (MSY) proxy for future harvest recommendations (ABCi,y,l) for each scenario l. To replicate current management, we used a climate-specific harvest control rule that uses climate-naive unfished and target spawning biomass reference points ((B_{0,i}) and (B_{{mathrm{target},i}}), respectively) and corresponding harvest rates ((F_{i,y} = 0) and (F_{i,y} = F_{{mathrm{target}}}) and (B_{i,y,l}) in each simulation l year (y) for each species (i)

$${mathrm{ABC}}_{i,y,l} = mathop {sum }limits_a^{A_i} left( {frac{{S_{i,a}F_{{mathrm{ABC}},i,y,l}}}{{Z_{i,a,y,l}}}left( {1 – e^{ – Z_{i,a,y,l}}} right)N_{i,a,y,l}W_{i,a,y,l}} right)$$
(3)

where (W_{i,a,y,l}), (N_{i,a,y}), and (Z_{i,a,y,l}) is the climate-simulation specific annual weight, number, and mortality (i.e., influenced through temperature effects on recruitment, predation, and growth) at age (a) for (A_i) ages in the model, and (S_{i,a}) is the average fishery age selectivity from the estimation period 1979–201759,60. (F_{{mathrm{ABC}},i,y,l}) and was determined in each simulation timestep using an iterative approach66 whereby we: (i) first determined average (B_{0,i}) values in years 2095–2099 by projecting the model forward without harvest (i.e., (F_{i,y} = 0)) for each species under the persistence scenario. We then (ii) iteratively solved for the harvest rate that results in an average spawning biomass ((B_{i,y})) during 2095–2099, that is, 40% of (B_{0,i}) (i.e., (F_{{mathrm{target}},i})) for pollock and Pacific cod simultaneously, with arrowtooth flounder (F_{i,y}) set to the historical average (as historical F for arrowtooth flounder ≪(F_{40% })); once (F_{{mathrm{target}},i}) for pollock and Pacific cod were found, we then iteratively solved for (F_{{mathrm{target}},i}) for arrowtooth flounder (Supplementary Fig. 7 left panel)59,60. Last, (iii) to derive a climate-informed ({mathrm{ABC}}_{i,y,l}) in each simulation year, the North Pacific Marine Fisheries Council (hereafter, “Council”) Tier 3 sloping harvest control rule with an ecosystem cutoff at (B_{20% }) was applied to adjust (F_{{mathrm{ABC}},i,y,l}) lower than (F_{{mathrm{target}},i}) if the simulation specific (climate-informed) (B_{i,y,l}) was lower than 40% of the climate-naive (B_{0,i}) at the start of a given year or set to 0 if (B_{i,y,l} , More