The central issue in global change studies is distinguishing anthropogenic signals from natural climate variations. Evidence is accumulating that much of the changes in greenhouse gases and other radiatively active trace species will be transmitted globally through changes in the mean tropical SST and its variability. A small positive trend over the last 50 years in the dominant pattern of tropical Indo-pacific SST variability is indeed already evident in observations. The question is: What is the likelihood of such a trend arising from purely natural variations? Many modeling groups are attempting to answer this question by examining the natural variability in global coupled climate models without secular changes in radiative forcing. The existence of substantial mean tropical biases in the models, as well as misrepresentations of ENSO variability, however prevents them from reaching a definite conclusion at present.

Here we take an alternative approach, using a linear inverse model (LIM) of tropical Indo-Pacific SST variability derived from the observed zero-lag and 4-month lag covariances of detrended SSTs over the last half century. The model can be integrated forward in time, using a linear evolution operator determined as above, and a stochastic forcing whose amplitude and spatial structure are also estimated from those same observed covariance statistics. These characteristics of the stochastic forcing can be approximated as SST-independent (“additive forcing”) or weakly SST-dependent (“multiplicative forcing”). We have made three 50,000 year integrations of this LIM with additive forcing and two different types of multiplicative forcing, and split the simulated time series of the dominant pattern of SST variability into 1000 separate 50-year segments. The sample mean, standard deviation, and trend over 50 years are calculated for each of these 50-year segments, and their probability distribution is estimated from histograms plotted using results from the 1000 segments. By construction, the population mean statistics of the mean and standard deviation should be zero and the observed standard deviation, respectively, and they are. There is no such constraint on the distribution of the 50-year trends, since the observed trend is not used to train the LIM. And indeed, the observed 50-year trend not only does not match the mean trend in the 1000 segments, but lies at the extreme positive end of the distribution of those trends, This result is obtained regardless of whether the specified stochastic forcing is assumed to be additive or multiplicative. Our analysis thus leads us to conclude that the observed 50-year trend in the dominant pattern of tropical Indo-Pacific SST variability is not consistent with naturally occurring 50-year trends expected from the observed statistics of interannual SST variability over the last half century. We have likewise also assessed the significance of the SST trends in the West Pacific warm pool and the Indian ocean.