No abstract uploaded!

The onset of thermal convection in a rapidly rotating spherical shell is studied by linear stability analysis based on the fully compressible Navier-Stokes equations. Compressibility is quantified by the number of density scale heights N-rho, which measures the intensity of density stratification of the motionless, polytropic base state. The nearly adiabatic flow with polytropic index n = 1.499 < na = 1.5 is considered, where na is the adiabatic polytropic index. By investigating the stability of the base state with respect to the disturbance of specified wavenumber, the instability process is found to be sensitive to the Prandtl number Pr and to N-rho. For large Pr and small N-rho, the quasi-geostrophic columnar mode loses stability first; while for relatively small Pr a new quasi-geostrophic compressible mode is identified, which becomes unstable first under strong density stratification. The inertial mode can also occur first for relatively small Pr and a certain intensity of density stratification in the parameter range considered. Although the Rayleigh numbers Ra for the onsets of the quasi-geostrophic compressible mode and columnar mode are different by several orders of magnitude, we find that they follow very similar scaling laws with the Taylor number. The critical Ra for convection onset is found to be always positive, in contrast with previous results based on the widely used anelastic model that convection can occur at negative Ra. By evaluating the relative magnitude of the time derivative of density perturbation in the continuity equation, we show that the anelastic approximation in the present system cannot be applied in the small-Ra and large-N-rho regime.

It is well acknowledged in image processing domain that the information can be decomposed into different frequency parts and each part has its own merits. However, existing neural networks always ignore the distinctions and straightforwardly feed all the information into neural networks together, treating them equally. In this paper, we propose a novel neural networks framework named J-Net that decomposes images into different frequency bands and then processes them sequentially. Concretely, the images have been decomposed by wavelet transformation and then the wavelet coefficients are fed into neural networks gradually in different depth according to their decomposition levels. An attention module is utilized to facilitate the fusion of neural network features and injected information, yielding significant performance gain. Furthermore, we show how does the information with different frequency impact the accuracy of neural networks. Experiments show that 5.91%, 5.32% and 2.00% accuracy improvements on Caltech 101, Caltech256 and ImageNet, respectively.

The spectral radius of a graph is the largest eigenvalue of the adjacency matrix of the graph. Let T(n,D
l) be the tree which minimizes the spectral radius of all trees of order n with exactly l vertices of maximum degree D
. In this paper, T(n,D
l) is determined for D=3, and for l<= 3 and n large enough. It is proven that for sufficiently large n, T(n, 3, l) is a caterpillar with (almost) uniformly distributed legs, T(n,D
2) is a dumbbell, and T(n,
D,3) is a tree consisting of three distinct stars of order D
connected by three disjoint paths of (almost) equal length from their centers to a common vertex. The unique tree with the largest spectral radius among all such trees is also determined. These extend earlier results of Lov´asz and Pelik´an, Simi´c and To˘si´c, Wu, Yuan and Xiao, and Xu, Lin and Shu.

No abstract uploaded!