A random process is wide-sense stationary, or WSS, when its first two moments do not change with absolute time.

Its mean is constant:

$$\mathbb E[X(t)]=\mu_X,$$

and for a real-valued process its autocorrelation depends only on the time separation $\tau$:

$$R_X(\tau)=\mathbb E[X(t)X(t+\tau)].$$

Autocorrelation measures how strongly values of the process at two separated times resemble one another on average. At zero lag,

$$R_X(0)=\mathbb E[X(t)^2],$$

so it gives the process's mean squared value.

A process can have zero mean without being uncorrelated across time, and it can be uncorrelated at nonzero lags without individual samples being statistically independent.

Stationarity makes long-running noise and signal models tractable because their second-order statistics can be described by lag rather than by two independent time coordinates.