“…This question has been well investigated in a line of work [17,38,29,12,3,35,36] based on common dimensionality reduction methods like CountSketch [9], sparse 0-1 matrices [14], Gaussian and sub-Gaussian matrices [39,28], and FFT-based fast constructions [1,44]. In many applications, the coefficient data A, b admit multi-linear (tensor) structures, for example in: spatio-temporal data analysis [4,19], higher-order tensor decompositions [15,24], approximating polynomial kernels [22,34], linearized PDE inverse problems [10,32] and so on. In the sketching setting, we can utilize the tensor structure in the original objective function to speed up forming the sketched problem by making the sketching matrix have a corresponding tensor structure.…”