On Efficient LHS-Based Yield Analysis of Analog Circuits

The Latin hypercube sampling (LHS) has been used as a variance-reduction estimation tool for an efficient sampling-based variability analysis of analog circuits. For a certain estimation confidence interval, a lower number of LHS samples is needed than that of Monte Carlo due to the estimation variance reduction. In this paper, an analysis of variance decomposition of the indicator function, the yield function, reveals strong contribution of interactive terms in the variance of the yield function, leading to limited performance gain of the traditional LHS sampling. In order to improve its efficiency, two correlation-controlled LHS methods are developed to reduce the required number of LHS samples for analog circuit yield estimation.