Testing for nonlinearity in European climatic time series by the method of surrogate data

Using temperature and pressure records from Czech meteorological stations and NCEP/NCAR reanalysis series, we tested for the presence of detectable nonlinearity in univariate and multivariate climatic time series. The method of surrogate data was utilized for nonlinearity detection – results of nonlinear prediction for the original series were compared to the results for series whose nonlinear structure was randomized. The prediction was done by means of local linear models in the reconstructed phase space. None or very weak nonlinearity was found in the single (univariate) series, and pressure series generally exhibited stronger nonlinearity than series of temperature (daily mean, minimum or maximum). Distinct nonlinearity was found in all tested multivariate systems, especially when both temperatures and pressures were used simultaneously to form the phase space. Nonlinearity tests were carried out for 30-year and 10-year-long datasets and nonlinear behavior was generally more apparent in the longer versions. In addition, the tested systems showed more substantial nonlinearity when the success of short-range prediction was used as the discriminating statistic; with an increase of the prediction time, detectable nonlinearity became weaker and it disappeared completely for long-term prediction.