On the Time Scale Structure of Climate Variability and Response

Abstract

This thesis analyzes several aspects pertaining to the time scale structure of climate variability and the temperature response to external radiative forcing. The analysis begins with an examination of the structure of temperature variability on synoptic time scales. Non-normal characteristics of temperature distributions have important implications for extreme climatic events. Mechanisms giving rise to the observed non-normality have been posited across a range of time scales. The structure of non-normality is examined in radiosonde time series of winter temperature using linear filters. These linear filters are shown to suppress non-normal variability. Observed non-normality on longer time is shown to likely be introduced at the highest resolved frequency and propagated through autocorrelation. Conversely, observations of normal distributions on synoptic time scales are shown to be an artifact of filtering. In the process, a Monte-Carlo based test for non-normality is developed for use in time series presenting autocorrelation. On longer time scales the single most important climate quantity is, arguably, a measure of the long-term warming that results from increased greenhouse gas concentrations, termed Equilibrium Climate Sensitivity or ECS. Estimates of ECS are drawn from numerical simulations, historical data, and paleoclimate proxies. The historical estimates have a range of 1.5-3 C compared to the 2-4.5 C range drawn from simulations and proxies. These historical estimates, however, are suspected to be biased low, as they rely on an assumption of a single radiative response to warming across all timescales. This bias is quantified using a Bayesian methodology to parse fast and slow modes in the evolution of Earth's temperature and radiation within an ensemble of 24 climate GCMs. Centennial-scale modes with stronger amplifying feedbacks ultimately contribute 44% of the long-term warming, but account for only 3 % of current warming. Thus, although the GCM ensemble has a median equilibrium sensitivity of 3.4 C, historical forcing would yield a biased estimate of only 2.5 C, consistent with the observational range. The equilibrium response is unlikely to be realized in upcoming centuries, as the radiative forcing is expected to vary on the same time scales. The temporal structure of atmospheric CO2 following present and future carbon emissions is simulated using a response function consisting of three eigenmodes fitted to 16 models of the carbon cycle. Convolving these response functions with those of the CMIP5 models provides a comprehensive probabilistic distribution for the evolution of temperature in response to any given emission scenario. The framework allows for calculation of emission budget associated with a given peak temperature threshold and a given probability of exceeding that threshold. As expected, the uncertainty range is sensitive to the fast decadal modes of the physical system and the carbon cycle response for narrow emission profiles. However, for more realistic broad emission profiles the uncertainty range is most sensitive to the slow centennial modes of the physical system. Due to the low degree of representation in modern observations, constraining the evolution of slow climatic modes will likely require use of paleoclimate proxies. The dominant modes of variability over the Pleistocene is associated with periodic changes in the earth's orbital configuration termed Milankovitch cycles. Such cycles have been described in pre-Pleistocene sediment records spanning warmer climates that may provide better analogues for future warming. Most of these studies, however, employ climatic records set on orbitally tuned chronologies. The bias introduced in spectral power estimates from the background continuum in the presence of tuning is quantified using Monte-Carlo methods, and appropriate hypothesis test for orbital forcing is developed. The test is applied to two marine sediment d18 O records spanning the Oligo-Miocene, from ODP cores 1090 and 1218. Orbital tuning is found to increase the statistical significance of a precession peak, whereas the obliquity and eccentricity peaks are no longer significant when compared to a null hypothesis of tuned background noise.