In compressive sampling (CS) MRI, non-adaptive variable density sampling (VDS) offers many advantages over adaptive sampling techniques. However achieving a better k-space sampling distribution, especially at lower sampling rates, is a challenging problem using conventional VDS method. As a result, the quality of CS reconstruction is affected. To overcome this problem, an improved VDS method is proposed in this paper. A probability density function (PDF) is used to generate variable density sampling patterns. The PDF is divided into number of smaller segments to control the sampling distribution in a better way. The samples are then distributed to these segments according to their local density values. The method is implemented on fully acquired k-space data of phantom and multi-slice brain imaging respectively. Images are reconstructed using /1-norm minimization technique. Data reduction of upto 75% is achieved in phantom data and upto 60% is achieved in brain MRI data respectively. The reconstruction results are compared with the results of conventional VDS method. It is found that the quality of reconstruction is significantly improved over conventional method. Thus the CS reconstruction quality is improved using the proposed method by achieving better k-space sampling distribution. Copyright © 2015 American Scientific Publishers All rights reserved.