Logo of pnasPNASInfo for AuthorsSubscriptionsAboutThis Article
Proc Natl Acad Sci U S A. 2002 Apr 2; 99(7): 4167–4171.
Published online 2002 Mar 19. doi:  10.1073/pnas.022055499
PMCID: PMC123620

Atmospheric carbon dioxide levels for the last 500 million years


The last 500 million years of the strontium-isotope record are shown to correlate significantly with the concurrent record of isotopic fractionation between inorganic and organic carbon after the effects of recycled sediment are removed from the strontium signal. The correlation is shown to result from the common dependence of both signals on weathering and magmatic processes. Because the long-term evolution of carbon dioxide levels depends similarly on weathering and magmatism, the relative fluctuations of CO2 levels are inferred from the shared fluctuations of the isotopic records. The resulting CO2 signal exhibits no systematic correspondence with the geologic record of climatic variations at tectonic time scales.

The long-term carbon cycle is controlled by chemical weathering, volcanic and metamorphic degassing, and the burial of organic carbon (1, 2). Ancient atmospheric carbon dioxide levels are reflected in the isotopic content of organic carbon (3) and, less directly, strontium (4) in marine sedimentary rocks; the former because photosynthetic carbon isotope fractionation is sensitive to CO2 levels, and the latter because weathering and degassing are associated with extreme values of the abundance ratio 87Sr/86Sr. However, attempts to use these geochemical signals to estimate past CO2 levels (58) are hindered by the signals' additional relationships to various tectonic (9, 10) and biological (11) effects. Moreover, the strontium signal has proven especially difficult to parse (1215).

Here, I attempt to resolve these ambiguities in the isotopic signals of carbon and strontium. First, it is shown that the last 500 million years of the strontium signal, after transformation to remove the effects of recycled sediment (16, 17), correlate significantly with the concurrent record of isotopic fractionation between inorganic and organic carbon (3). This empirical result is supplemented by the theoretical deduction that the two records are linked by their common dependence on rates of continental weathering and magmatic activity. The assumption that CO2 levels fall with the former and rise with the latter then indicates that an appropriate average of the two records should reflect the long-term fluctuations of the partial pressure of atmospheric CO2. The CO2 signal derived from this analysis represents fluctuations at time scales greater than about 10 million years (My). Comparison with the geologic record of climatic variations (18) reveals no obvious correspondence.

Strontium and Carbon Isotopic Signals

Fig. Fig.11 shows the strontium and carbon isotopic signals for the last 500 My. The data for the strontium isotope ratios 87Sr/86Sr were compiled by Veizer et al. (4) and Walter et al. (19); the former source accounts for all 87Sr/86Sr data younger than 520 My, whereas the latter was used for data extending back to 608 My (not shown). Both sets of 87Sr/86Sr data were first averaged in nonoverlapping time windows of 10 My. The resulting unevenly spaced record was then transformed (20) to an evenly spaced record with time increments of approximately 10 My and no contributions to the power spectral density at periods less than 21 My.

Figure 1
Data for epsilontoc (red filled circles) (3) and 87Sr/86Sr (blue open squares) (4). The time scale for epsilontoc has been revised from the original to match the scheme (32) used for the strontium data. The capital letters correspond to the ...

The second record in Fig. Fig.1,1, compiled by Hayes et al. (3) derives from the isotopic composition of marine organic carbon and carbonate carbon. From isotopic abundance ratios Rx = (13C/12C)x for carbon in sample x, the isotopic fractionation between sample x and a standard (STD) sample, δx = 1,000[(RxRSTD)/RSTD], is obtained for carbonate (δa) and organic (δo) carbon. The isotopic fractionation epsilontoc between total organic carbon and sedimentary carbonates is then given approximately by epsilontoc = δa − δo, which is the second signal plotted in Fig. Fig.11.

Fig. Fig.11 shows a surprising similarity between the two records for fluctuations with periods less than about 100 My. However, the correlation between the two time series is not statistically significant [Spearman rank correlation coefficient (21) Rs = −0.40, P = 0.17, N = 46], because the longer-period fluctuations are not in phase. It is therefore interesting to ask why the two records appear similar at shorter time scales and dissimilar at longer time scales.

Recent work has shown that the measurements of epsilontoc approximately fit the empirical relation (3, 22)

equation M1

The parameter epsilon0 represents the isotopic effects of photosynthesis and secondary biological processes along with the isotopic depletion of dissolved CO2 in surface waters relative to sedimentary carbonate. Because epsilon0 is approximately constant throughout the period plotted in Fig. Fig.1,1, the main source of epsilontoc's fluctuations is contained in isotopic effects due to changing algal physiology—the permeability, surface-to-volume ratio, and growth rate of algal cells—represented by Γ, and the concentration Φ of dissolved carbon dioxide in surface waters (3, 8).

Although the mechanisms responsible for the fluctuations of s = 87Sr/86Sr are subject to much debate (4, 9, 10, 1215), some aspects of the signal's evolution are nevertheless clear. Because 87Rb decays to 87Sr with a half-life of ~48 billion years, the supply of 87Sr may be taken to be approximately constant over the last 500 My. However, it is not uniformly distributed: the fluvial input to the oceans is derived in part from rocks—both silicates and metamorphosed carbonates (1215)—that are relatively enriched in radiogenic Sr (s [similar, equals] 0.712 or greater) compared to the nonradiogenic Sr of mantle origin supplied at hydrothermal vents (s [similar, equals] 0.7035) (23). The value of s(t) at some particular time t represents, to first approximation, the relative fluxes of these two extreme values.

However, this first approximation ignores the recycling of rocks in the sedimentary cycle (16). As pointed out by Brass (17), about 75% of the strontium input to the oceans should come from the weathering of exposed carbonates of marine origin. Because these carbonates retain a memory of s at the time of deposition, their contribution to sedimentary Sr significantly damps the signal coming from hydrothermal vents (low s) and rocks containing radiogenic Sr (high s) (9).

Denoting the fluvial flux of radiogenic Sr by r, the input from hydrothermal vents by vh, and the memory effect by m, a simple expression for the strontium isotope ratio s is

equation M2

where 0 ≤ μ ≤ 1 is the fraction of sedimentary Sr deriving from the memory flux, and g is a function that depends on both r and vh. The memory term can be approximated by assuming that sedimentary strontium is weathered at a rate proportional to its mass (16). Then m(t) is simply a weighted (exponentially decaying) average of s(τ), τ < t (17), which we express by the convolution

equation M3

The parameter λ is related to the half-life t1/2 of sedimentary Sr; i.e., λ = −(ln 2)/t1/2. Brass estimates 57 My < t1/2 < 102 My (17). Therefore, the memory effect should strongly influence the fluctuations of s at long time scales (greater than about 100 My), whereas the short time scale fluctuations should be relatively unaffected.

The short-time correlations may be explained in part by noting the common dependence of epsilontoc and g on the global weathering rate w (i.e., the flux of all weathered products from the continents to the oceans) and the rate of magmatic activity v associated not only with the hydrothermal flux vh but also with volcanic degassing. Assuming that r = r(w) and vh = vh(v), g's dependence on w and v may be written§

equation M4

With respect to epsilontoc, the physiological term Γ may likewise depend on w because of changing nutrient concentrations in the oceans. Γ could also depend on CO2 levels, which in turn depend on degassing, weathering rates—because of the net uptake rate u = u(w) of atmospheric CO2 associated with silicate weathering (1, 2, 24)—and other processes, such as organic carbon burial, which we collectively designate by b. Aggregating these dependencies then yields

equation M5

Comparison of Eqs. 4 and 5 directly reveals the joint dependence of g and epsilontoc on v and w.

Removal of the Effects of Sedimentary Recycling

Because the strontium signal s contains a memory effect, whereas epsilontoc does not, removal of the memory effect should reveal correlations between s and epsilontoc at both short and long time scales, thereby confirming the joint dependence on v and w. To test this hypothesis, I first assume it is true and use it to estimate g. For a given λ and μ, m is calculated by discretizing Eq. 3 for τ > −608 My. Eq. 2 is then solved for g:

equation M6

Here the subscripts make explicit the dependencies on λ and μ.

Let Rλμ(epsilontoc, gλμ) be the Pearson product-moment correlation coefficient (21) that quantifies the similarity of epsilontoc and gλμ. From the foregoing argument, the best estimate of λ and μ should be that which minimizes Rλμ (because the expected correlation is negative). Fig. Fig.22 shows contours of R as a function of μ and t1/2 = −(ln 2)/λ. The minimum Rλμ = −0.81 occurs at t1/2 = 54 My and μ = 0.99. However, the corresponding estimate of gλμ includes values below the hydrothermal minimum of 0.7035. Such nonphysical ratios probably derive from the amplification of any measurement noise resulting from division by the small quantity 1 − μ in Eq. 6. Indeed, the dark gray region in Fig. Fig.22 indicates that all such results occur only for μ close to unity. I therefore define the best estimates λ*, μ* of λ, μ by minimizing Rλμ subject to the constraint that gλμ > 0.7035. One then finds t*1/2 = −ln 2/λ* = 41 My and μ* = 0.83, corresponding to Rλ*μ* = −0.80 (within 1% of the unconstrained minimum). These results are in reasonable accord with Brass's estimate of the half-life of sedimentary Sr (57 − 102 My) (17) and his conclusion that “strontium leached from limestones is about 75% of the total input” (17).

Figure 2
Contours of Rλμ as a function of μ, the fraction of sedimentary Sr deriving from the memory flux, and t1/2 = −(ln 2)/λ, the half-life of sedimentary Sr. R was calculated for μ = 0, 0.01,  , ...

Fig. Fig.33 shows gλ*μ* compared to epsilontoc. Compared with Fig. Fig.1,1, one sees that both the long and short period fluctuations are now not only approximately equally correlated, but the correlation is also significant (Rs = −0.74, P < 10−3, N = 46). The hypothesis that recycled sediment partially obscured an inherent correlation because of a shared dependence on weathering and volcanic processes is therefore confirmed.

Figure 3
The function g (blue squares) obtained by removing the memory flux from the 87Sr/86Sr data of Fig. Fig.1,1, along with epsilontoc (red circles). Note that the range of the 87Sr/86Sr curve is approximately five times greater than ...

Relationship to CO2 Levels

To understand further the correlation, I make the following assumptions concerning the functional dependencies contained in Eqs. 4 and 5:

equation M7

The first two state that the flux of radiogenic Sr to the oceans and the uptake of atmospheric CO2 because of silicate weathering increase as the global weathering rate (of all minerals in all terranes) increases. The third formalizes the natural assumption that the hydrothermal flux of Sr varies with the same sign as all magmatic processes.

Because the Sr isotope ratios increase with increasing r (i.e., [partial differential]g/[partial differential]r > 0) and decrease with increasing vh (i.e., [partial differential]g/[partial differential]vh < 0), one has

equation M8

Because g and epsilontoc are negatively correlated, one expects that epsilontoc depends on v and w in the opposite sense:

equation M9

The inequalities 8 and 9 have been obtained without making any explicit assumption concerning any dependence of r on u nor Γ on v and w. Because they show that v and w each influence g and epsilontoc with opposite signs, the relative fluctuations of any other quantity that also depends on v and w with opposite signs may be inferred from the shared fluctuations of g and epsilontoc. Specifically, the concentration Φ of oceanic CO2 responds positively to v while its response to weathering is negative:

equation M10

Thus the shared fluctuations of g and epsilontoc indicate the relative fluctuations of ancient CO2 levels.

Estimation of the CO2 Signal

I proceed to estimate the CO2 signal explicitly. First, the highest measurement of epsilontoc, at −175 My, is dropped because of its statistical insignificance (3). I then transform g(t) → epsilong(t), a strontium-derived estimate of epsilontoc, by mapping the left axis of Fig. Fig.33 to the corresponding value on the right axis. The quantities epsilontoc and epsilong are then averaged to obtain

equation M11

The times ti are given by the points where epsilontoc is specified and epsilong(ti) is obtained by linear interpolation. Here it is implicitly assumed that epsilontoc and epsilong contain a common signal that is enhanced by averaging. Statistical arguments suggest that ζ's rms signal-to-noise ratio lies between about 2 and 3.**

Assume now that Γ and epsilon0 are constant and define the dimensionless CO2 concentration [var phi] = Φepsilon0/Γ (8). In terms of [var phi] and ζ, Eq. 1 then yields

equation M12

At long time scales such that the oceans are in equilibrium with the atmosphere, the partial pressure pCO2 is proportional to [var phi] (1). The relative fluctuations of pCO2 are therefore given by [var phi](t)/[var phi](0), which is plotted in Fig. Fig.44 for epsilon0 = 36 per mil (‰). The gray area surrounding the pCO2 curve in Fig. Fig.44 brackets this result for epsilon0 = 35‰ and 38‰, a range consistent with previous estimates (3).

Figure 4
Fluctuations of pCO2 for the last 500 My, normalized by the estimate of pCO2 obtained from the most recent value of ζ. The solid line is obtained from Eq. 12 by using epsilon0 = 36‰. The lower and upper limits of the gray area surrounding ...

Fig. Fig.44 reveals that CO2 levels have mostly decreased for the last 175 My. Prior to that point they appear to have fluctuated from about two to four times modern levels with a dominant period of about 100 My. The decline for the last 175 My is also present in several previous pCO2 reconstructions (7, 8, 26, 27), and the entire curve displays some similarity to a previous estimate derived from the geologic record of carbonate formation (26). Although the period before −175 My differs substantially from previous geochemical model calculations (7), an approximate error estimate lends considerable credence to the pCO2 curve of Fig. Fig.4.4. Specifically, [var phi] should inherit ζ's signal-to-noise ratio of 2–3. This correspondence would be exact if [var phi](ζ) were linear. Because [var phi](ζ) is monotonic for the observed range of ζ, its nonlinearity does not affect the timing of the maxima and minima of the pCO2 curve. Thus the linear error estimate remains pertinent.

Comparison with the Climate Record

Using a variety of sedimentological criteria, Frakes et al. (18) have concluded that Earth's climate has cycled several times between warm and cool modes for roughly the last 600 My. Recent work by Veizer et al. (28), based on measurements of oxygen isotopes in calcite and aragonite shells, appears to confirm the existence of these long-period (~135 My) climatic fluctuations. Changes in CO2 levels are usually assumed to be among the dominant mechanisms driving such long-term climate change (29).

It is therefore interesting to ask what, if any, correspondence exists between ancient climate and the estimate of pCO2 in Fig. Fig.4.4. The gray bars at the top of Fig. Fig.44 correspond to the periods when the global climate was cool; the intervening white space corresponds to the warm modes (18). The most recent cool period corresponds to relatively low CO2 levels, as is widely expected (30). However, no correspondence between pCO2 and climate is evident in the remainder of the record, in part because the apparent 100 My cycle of the pCO2 record does not match the longer climatic cycle. The lack of correlation remains if one calculates the change in average global surface temperature resulting from changes in pCO2 and the solar constant using energy-balance arguments (7, 26).

Superficially, this observation would seem to imply that pCO2 does not exert dominant control on Earth's climate at time scales greater than about 10 My. A wealth of evidence, however, suggests that pCO2 exerts at least some control [see Crowley and Berner (30) for a recent review]. Fig. Fig.44 cannot by itself refute this assumption. Instead, it simply shows that the “null hypothesis” that pCO2 and climate are unrelated cannot be rejected on the basis of this evidence alone.

Discussion and Conclusion

One of the principal contributions of this study is methodological. From observations of a weak correlation between strontium and carbon isotopic signals (Fig. (Fig.1)1) and their shared dependence on global weathering rates and magmatic activity, weathering and magmatism are deduced to be the main processes driving the signals' fluctuations. Correction for the effects of sedimentary recycling enhances the correlation (Fig. (Fig.3),3), indicates a strong shared signal, and strengthens this conclusion.

A second, crucial step is to note that any quantity with a similar joint dependence on weathering and magmatic processes may be expected to display similar fluctuations. Here attention has been focused on CO2 levels; as for the strontium and carbon isotopic signals, CO2 levels depend on weathering and magmatism with opposite signs and should therefore fluctuate roughly in sync with the isotopic signals. Because the reasoning is general, it need not be limited to CO2. Among the many possible applications, the case of oceanic phosphate concentrations is particularly interesting. Phosphate concentrations should increase with weathering and decrease with hydrothermal activity (31); thus the methodology in this paper may be applicable to their reconstruction. Moreover, because phosphorus is a limiting nutrient, oceanic productivity may be expected to covary positively with its concentration in seawater, suggesting that CO2 levels and productivity covary negatively at geologic time scales (8).

Such reasoning naturally raises the issue of cause and effect. This study indicates that degassing and silicate weathering were the primary controls on the carbon cycle for the last 500 My. But the results do not themselves indicate whether either of these mechanisms dominated, or whether weathering was driven by the diversification of land plants (8), continental collisions (9), or a complex combination of tectonic, biological, and geochemical processes (7). They do, however, offer a new view of the long-term fluctuations of pCO2 that will hopefully stimulate novel approaches to the study of biogeochemical cycles at evolutionary time scales.


I thank O. Aharonson, L. Derry, J. Hayes, P. Hoffman, A. Knoll, L. Kump, J. Sachs, R. Summons, and the late John Edmond for helpful remarks. This work was supported in part by National Science Foundation Grant DEB-0083983.


million years


Correlations between epsilontoc(t) and s(t) or g(t) are computed from the N equal-time pairs obtained after linearly interpolating the Sr signal so that it is sampled at the same times as epsilontoc. The statistical significance P is one-sided and was estimated by using the Monte Carlo technique described in ref. 8.

§Although w and v could conceivably depend on one another, here such dependencies are assumed to be insignificant compared to any independent variations.

One also requires that if changes in v or w dominate one process then they dominate all processes.

Because Φ can also depend on other processes b such as organic carbon burial, the shared fluctuations of g and epsilontoc do not necessarily reflect the full evolution of Φ but rather the “partial” contribution of v and w alone. However, the good correlation of Fig. Fig.33 indicates that b's influence on Φ's fluctuations has been small, except possibly in the Carboniferous. Thus one may conclude that the fluctuations of g and epsilontoc follow to good approximation the fluctuations of Φ with opposite signs.

**To estimate the signal-to-noise ratio of the time series ζ(t) defined by Eq. 11, assume that epsilontoc and epsilong are each composed of a shared but unknown signal z in the presence of zero-mean noise ηtoc and ηg, respectively: epsilontoc = z + ηtoc and epsilong = z + ηg. The quantity ζ is an estimate of z. Its signal-to-noise ratio may be computed under the assumption that ηtoc and ηg each have variance σequation M13, z has variance σequation M14, and ηtoc, ηg, and z are each uncorrelated to the other. The expectation of the correlation coefficient R is then σequation M15/(σequation M16 + σequation M17) (25). The observed correlation |R| = 0.80 then yields σequation M18equation M19 = 4, or that the rms signal fluctuation is twice that of the noise. The rms signal-to-noise ratio of ζ can be as large as a factor of equation M20 greater, i.e., it should range between about 2 and 3.


1. Walker J C G. Evolution of the Atmosphere. New York: Macmillan; 1977.
2. Holland H D. The Chemistry of the Atmosphere and Oceans. New York: Wiley; 1978.
3. Hayes J M, Strauss H, Kaufman A J. Chem Geol. 1999;161:103–125.
4. Veizer J, Ala D, Azmy D, Bruckschen P, Buhl D, Bruhn F, Carden G, Diener A, Ebneth S, Godderis Y, et al. Chem Geol. 1999;161:59–88.
5. Freeman K, Hayes J M. Global Biogeochem Cycles. 1992;6:185–198. [PubMed]
6. Francois L M, Walker J C G. Am J Sci. 1992;292:81–135. [PubMed]
7. Berner R A. Am J Sci. 1994;294:56–91.
8. Rothman D H. Proc Natl Acad Sci USA. 2001;98:4305–4310. [PMC free article] [PubMed]
9. Edmond J. Science. 1992;258:1594–1597. [PubMed]
10. Richter F M, Rowley D B, DePaolo D J. Earth Planet Sci Lett. 1992;109:11–23.
11. Hayes J M. Marine Geol. 1993;113:111–125.
12. Derry L A, France-Lanord C. Earth Planet Sci Lett. 1996;142:59–74.
13. Quade J, Roe L, DeCelles P G, Ojha T P. Science. 1997;276:1828–1831.
14. Sharma M, Wasserburg G J, Hofmann A, Chakrapani G J. Geochim Cosmochim Acta. 1999;63:4005–4012.
15. Basu A R, Jacobsen S B, Poreda R J, Dowling C B, Aggarwal P K. Science. 2001;293:1470–1473. [PubMed]
16. Garrels R M, Mackenzie F T. Evolution of Sedimentary Rocks. New York: Norton; 1971.
17. Brass G W. Geochim Cosmochim Acta. 1976;40:721–730.
18. Frakes L A, Francis J E, Syktus J I. Climate Modes of the Phanerozoic. Cambridge, U.K.: Cambridge Univ. Press; 1992.
19. Walter M R, Veevers J J, Calver C R, Gorjan P, Hill A C. Precambrian Res. 2000;100:371–433.
20. Vio R, Strohmer T, Wamsteker W. Publ Astrono Soc Pac. 2000;112:74–90.
21. Press W H, Flannery B P, Teukolsky S A, Vetterling W T. Numerical Recipes in C: The Art of Scientific Computing. Cambridge, U.K.: Cambridge Univ. Press; 1995.
22. Popp B N, Laws E A, Bidigare R R, Dore J E, Hanson K L, Wakeham S G. Geochim Cosmochim Acta. 1998;62:69–77.
23. Palmer M R, Edmond J. Earth Planet Sci Lett. 1989;92:11–26.
24. Urey H C. The Planets. New Haven, CT: Yale Univ. Press; 1952.
25. Feller W. An Introduction to Probability Theory and Its Applications. New York: Wiley; 1968.
26. Budyko M I, Ronov A B, Yanshin A L. History of the Earth's Atmosphere. Berlin: Springer; 1987.
27. Ekart D D, Cerling T E, Montanez I P, Tabor N J. Am J Sci. 1999;299:805–827.
28. Veizer J, Godderis Y, Francois L M. Nature (London) 2000;408:698–701. [PubMed]
29. Kump L R. Nature (London) 2000;408:651–652. [PubMed]
30. Crowley T J, Berner R A. Science. 2001;292:870–872. [PubMed]
31. Wheat C G, Feely R A, Mottl M J. Geochim Cosmochim Acta. 1996;60:3593–3608.
32. Harland W B, Armstrong R, Cox A, Craig L, Smith A G, Smith D G. A Geologic Time Scale 1989. Cambridge, U.K.: Cambridge Univ. Press; 1990.

Articles from Proceedings of the National Academy of Sciences of the United States of America are provided here courtesy of National Academy of Sciences
PubReader format: click here to try


Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...


  • PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...