Dansgaard-Oeschger events: tipping points in the climate system

Dansgaard-Oeschger events are a prominent mode of variability in the records of the last glacial cycle. Various prototype models have been proposed to explain these rapid climate fluctuations, and no agreement has emerged on which may be the more correct for describing the paleoclimatic signal. In this work, we assess the bimodality of the system reconstructing the topology of the multi–dimensional attractor over which the climate system evolves. We use high-resolution ice core isotope data to investigate the statistical properties of the climate fluctuations in the period before the onset of the abrupt change. We show that Dansgaard-Oeschger events have weak early warning signals if the ensemble of events is considered. We find that the statistics are consistent with the switches between two different climate equilibrium states in response to a changing external forcing (e.g. solar, ice sheets…), either forcing directly the transition or pacing it through stochastic resonance. These findings are most consistent with a model that associates Dansgaard-Oeschger with changing boundary conditions, and with the presence of a bifurcation point.

💡 Research Summary

The paper investigates the nature of Dansgaard‑Oeschger (DO) events—abrupt climate fluctuations that occurred repeatedly during the last glacial period—by analysing high‑resolution δ¹⁸O isotope records from the NGRIP Greenland ice core. The dataset spans 59 k to 15 k years before present with an average temporal resolution of 2.7 years, covering DO events numbered 2 through 16.

First, the authors address the long‑standing claim that the DO time series is bimodal, i.e., that the climate system switches between two distinct equilibrium states. To avoid artefacts caused by projecting a high‑dimensional climate system onto a single scalar record, they employ Takens’ embedding theorem to reconstruct the system’s phase space. After interpolating the irregular series to a regular 4‑year grid, detrending, and selecting an optimal lag of 20 years (based on the first minimum of the average mutual information), they embed the series in a four‑dimensional space—the minimum dimension required to suppress global false neighbours. Probability density functions (PDFs) for each coordinate are estimated using Gaussian kernel density estimation and compared against a multivariate normal null hypothesis.

The results show a clear deviation from normality in the pre‑22 k yr segment (59 k–22 k yr). All four dimensions exhibit PDFs that lie outside the confidence intervals of the normal distribution, especially the angular coordinates, which display two distinct peaks—direct evidence of genuine bimodality. After 22 k yr, the PDFs converge toward a normal distribution, indicating a transition to a unimodal regime. This transition is interpreted as the climate system moving away from a region of multiple equilibria toward a more stable, single‑equilibrium configuration.

Having established the dynamical structure, the authors turn to early‑warning signals (EWS) that might precede each abrupt transition. They compute three statistical indicators—lag‑1 autocorrelation (c), variance (σ²), and the Detrended Fluctuation Analysis exponent (α)—for each DO event, then average these quantities across the ensemble of events to improve the signal‑to‑noise ratio. The ensemble‑averaged trends reveal modest but systematic increases in c, σ², and α as the onset of a DO approaches (approximately 18 k yr ago). The α values approach 1.5, consistent with “critical slowing down” as the system’s restoring force weakens near a bifurcation. The modest magnitude of the signals explains why they are not detectable in single‑event analyses.

To interpret these statistical signatures, the paper compares three prototype dynamical models, all based on the stochastic differential equation

  ẋ = −x³ + x + q + σ η(t),

where x represents the climate state, q is a control parameter, σ η(t) is additive white noise, and the cubic term creates a double‑well potential.

1. Forced Bifurcation Model – q varies slowly in time (representing external forcing such as solar irradiance or ice‑sheet changes). When q crosses one of the two fold bifurcation points (q₀ = ±2√3/9), one equilibrium disappears, forcing the system to jump to the other well. Simulations of this scenario reproduce the observed increases in c, σ², and α before transitions.

2. Noise‑Induced Transition Model – q is held constant, and a sufficiently large σ triggers random jumps between the wells. In this autonomous case, statistical indicators remain flat; no EWS appear, contradicting the empirical findings.

3. Stochastic Resonance Model – q oscillates periodically (q = q₀ sin(2πt/τ)), mimicking a weak periodic external driver (e.g., Milankovitch‑scale variations). The system is most susceptible to noise‑driven jumps when q is near a bifurcation, leading to intermittent EWS: some events show clear precursory increases, while others do not. This pattern matches the observed heterogeneity of EWS across the DO ensemble.

The comparison indicates that the data are most consistent with mechanisms involving a slowly varying external forcing that brings the system close to a bifurcation, or with stochastic resonance where periodic forcing modulates the distance to the critical point. Purely autonomous noise‑induced transitions are ruled out.

The phase‑space analysis also reveals that the post‑22 k yr distribution is not perfectly spherical, suggesting that the underlying climate dynamics are higher‑dimensional and involve coupled ocean‑atmosphere‑ice processes rather than a simple one‑dimensional potential. The authors caution that the detection of EWS is limited by the resolution of the ice‑core record, uncertainties in dating, and assumptions about the noise (additive, white, and uncorrelated). They advocate for higher‑resolution, multi‑proxy records and for incorporating colored noise in future modelling efforts.

In summary, the paper provides robust evidence that DO events are manifestations of a climate system with two competing equilibria, driven toward abrupt transitions by external forcing that moves the system close to a bifurcation point, possibly enhanced by stochastic resonance. Early‑warning signals are weak but detectable when averaged across many events, supporting the view that DOs are not purely random noise‑induced switches but are rooted in the system’s underlying nonlinear stability landscape.