How to build low dimensional representations of high-dimensional data is a challenging problem in image analysis and texture modeling. We review scattering representation which is built from a non-linear cascade of multiscale complex wavelet transform. Low dimensional structures are found in the scattering representation through tensor factorization on the Turbulent fluid images simulated from Magnetohydrodynamics. To capture dependencies of the phase in the wavelet coefficients, we introduce phase harmonics. Statistical dependencies can be extracted through covariance of the wavelet coefficients after non-linear phase harmonic transformations. These representations can be applied to build low-dimensional models for textures.