Non-Monotone Characteristic of Spectral Statistics in the Transition between Poisson and Gauss

We have computed the spectral number variances of an extended random matrix ensemble predicted by Guhr's supersymmetry formula, showing a non-monotone increase of the curves that arises from an "overshoot" of the two-level correlation function above unity. On the basis of the most general form of $N$-level joint distribution that meets sound probabilistic conditions on matrix spaces, the above characteristic may be attributed to the {\it attractiveness} of the pair potential in long range($E >$ Thouless energy) of the underlying level gas. The approach of level dynamics indicates that the result is "anti-screening" of the level repulsion in short-range statistics of the usual random matrix prediction until the joint level distribution undergoes a phase transition (the Anderson transition).