The probability distribution of the differential group delay (DGD) at any fiber length is determined by use of a physically reasonable model of the fiber birefringence. We show that if the fiber correlation length is of the same order as or larger than the beat length, the DGD distribution approaches a Maxwellian in roughly 30 fiber correlation lengths, corresponding to a couple of kilometers in realistic cases. We also find that the probability distribution function of the polarization dispersion vector at the output of the fiber depends on the angle between it and the local birefringence vector on the Poincaré sphere, showing that the DGD remains correlated with the orientation of the local birefringence axes over arbitrarily long distances. Polarization mode dispersion (PMD) is caused by the random birefringence that is present in optical f ibers. It can lead to pulse spreading and depolarization, and it is detrimental to system performance. As transmission rates continue to increase, the PMD has become an increasingly important fiber impairment, thus motivating extensive experimental and theoretical study over the past few years. PMD is characterized by a three-component dispersion vector, V. Its magnitude, jVj, gives the differential group delay (DGD) between the principal states, and its direction gives the orientation of the slow principal state of polarization at the output on the Poincaré sphere. 1 At short distances, PMD is deterministic, and the DGD probability distribution is a d function. At long distances, however, previous work in which a weak random birefringence model was used showed that the three components of V are independent and Gaussian distributed, so the DGD distribution is Maxwellian.1 Similar results are also obtained if one assumes that the fiber birefringence completely randomizes the polarization state over the Poincaré sphere.2 Both analysis and numerous numerical and experimental studies have lead to the generally accepted wisdom that the asymptotic (long-length limit) distribution function of the DGD resulting from PMD is Maxwellian. The transient behavior of the distribution, however, is not so well elucidated. In this Letter we study the fundamental question of determining the probability distribution of the DGD that is due to PMD at any fiber length with general fiber correlation and beat-length parameters. The results of our analytical work indicate that current systems approach a Maxwellian distribution in just a few kilometers, confirming the experimental work of Gisin et al. 3 and placing on a f irm theoretical foundation the widely used approach of calculating the penalties on transmission that are due to PMD by assuming that the distribution of the DGD is Maxwellian. Moreover, our results indicate when this assumption breaks down, which may be of use in future systems. In addition, we f ind, for what is to our knowledge the first time, that the probability distribution function of the polarization dispersion vector at the output of the fiber depends on the angle between ...