Abstract: The temperature correlation matrix C and random correlation matrix R are constructed based on the NCEP/NCAR temperature reanalysis data and random series with the characteristics of these two matrixes analyzed in this paper. The results show that there are genuine correlations as well as the correlation ‘noise’ existing in the temperature correlation matrixes. These genuine correlations can be divided to two parts, one is the correlation between the nearest and nextnearest points, called the short distance correlation (SC); the other is the long distance correlation (LC), i.e. the correlation between the El Nino area and other remote areas such as the warm pools. For the different scales, these two kinds of correlations show different features. On the 1-10 d scale, the SC is more important than the LC, while on the 15 d and larger scales, the SC and LC both play an important role in the temperature correlation matrixes. Most correlation information is contained in several eigenvectors with larger eigenvalues, and the projection of the global temperature series on these eigenvectors will show, in some cases, the whole characteristics of global temperature changes. Besides, the temperature correlations have significant temporal and spatial variabilities: for example, the correlation is better during 1950-1956, 1972-1977 and 1996-2000 than during 1978-1982 and 1991-1996. On the 1 d scale, the correlation shows however a good latitudinally symmetric spatial distribution, but it is relatively worse on the 15 d scale owing to the oceanland difference between the Northern and Southern Hemisphere. In contrast, on the 15 d scale, the correlation shows a longitudinally symmetric distribution.