Traditional methods for nonlinear dynamic analysis, such as correlation dimension, Lyapunov exponent, approximate entropy, detrended fluctuation analysis, using a single parameter, cannot fully describe the extremely sophisticated behavior of electroencephalogram (EEG). The multifractal formalism reveals more "hidden" information of EEG by using singularity spectrum to characterize its nonlinear dynamics. In this paper, the zero-crossing time intervals of sleep EEG were studied using multifractal analysis. A new multifractal measure Δ as α was proposed to describe the asymmetry of singularity spectrum, and compared with the singularity strength range Δα that was normally used as a degree indicator of multifractality. One-way analysis of variance and multiple comparison tests showed that the new measure we proposed gave better discrimination of sleep stages, especially in the discrimination between sleep and awake, and between sleep stages 3 and 4.The research of nonlinear dynamics in electroencephalogram (EEG) has made much headway in recent years. Nonlinear analysis methods have been successfully applied to the studies of brain functions and pathological changes in EEG [1][2][3] . These studies also proved that EEG exhibited at least partly chaotic characteristics. The detrended fluctuation analysis (DFA) [4] , which has been widely used recently, revealed the longrange power-law correlation in EEG, indicating time scale invariant and fractal structure [5,6] . However, EEG is also rather noisy, displaying short-term decorrelation like white noise, and consequently, the EEG has been traditionally considered as a linear stationary random process. The paradoxical combination of short-term decorrelation and long-range correlation, stochastic and deterministic suggests that a single nonlinear parameter, such as largest Lyapunov exponent, correlation dimension, fractal dimension, scaling exponent, etc., may not be able to fully characterize the "stochastic chaos" (as named by Freeman [7] ) nature of EEG.The long time behavior of chaotic, nonlinear dynamic systems can often be characterized by (mono) fractal or multifractal measures. Monofractals are homogeneous in the sense that they have the same scaling properties, characterized by a single singularity exponent throughout the entire signal. In contrast, multifractals can be decomposed into many (possibly infinite) sub-sets characterized by different exponents. Multifractal signals are intrinsically more complex and inhomogeneous than monofractals. Multifractal models have been used to account for scale invariance properties of various objects in very different domains ranging from the energy dissipation or the velocity field in turbulent flows [8] to underlying hierarchical structure in proteins [9] . Physiologic signals generated by complex self-regulating systems, such as heartbeat interval, electrocardiogram, gait etc. have been proven to be multifractal, and the degree of multifractality often relates to pathological state or natural aging process [10][11][12] ....