A complexity measure is introduced which is the integral of the coarse-grained, scale dependent entropies. As an application we analyze a Gaussian distribution model for noise signals. We find that among all possible distributions the one having the 1/f power spectrum maximizes the complexity measure