1: \begin{abstract} 

2: 

3: %{\it Sketched gradient methods} \cite{tang2016gradient} have been previously introduced for efficiently solving the large-scale constrained Least-squares regressions. Those sketched gradient methods utilizes iterative sketching techniques to reduce the sample size w.r.t the intrinsic low-dimensional structure of the constrained set where the solution lives in, and is proven to be able to exploit this structure to accelerate computation. In this paper we reveal the trade-off between the sketch size and step size of the sketched gradients and computational cost w.r.t the (sparsity) constraint, via theoretical analysis and numerical experiments on large data sets and image reconstruction applications.

4: 

5: {\it Sketched gradient algorithms} \cite{tang2016gradient} have been recently introduced for efficiently solving the large-scale constrained Least-squares regressions. In this paper we provide novel convergence analysis for the basic method {\it Gradient Projection Classical Sketch} (GPCS) to reveal the fast linear convergence rate of GPCS thanks to the intrinsic low-dimensional geometric structure of the solution prompted by constraint set. Similar to our analysis we observe computational and sketch size trade-offs in numerical experiments. Hence we justify that the combination of gradient methods and the sketching technique is a way of designing efficient algorithms which can actively exploit the low-dimensional structure to accelerate computation in large scale data regression and signal processing applications. 

6: 

7: \end{abstract}