A system, method, and computer program product for computing device mismatch variation contributions to circuit performance variation. Embodiments estimate which individual devices in a simulated circuit design have the largest impact on circuit performance, while requiring far fewer simulations than traditional multivariate linear regressions. When the samples exceed the mismatch parameters, a linear model is solved by least squares. Otherwise, a linear model is solved by orthogonal matching pursuit (OMP), and if that solution is too inaccurate then a new mixed method builds a better linear model. If the linear solution is too inaccurate, a full linear and quadratic model is made using OMP to select the most important variables, and the full model is fitted using OMP with selected cross terms. The embodiments summarize the output variance in each device, and rank the mismatch contributions based on the summarized contributions.