The control of line-edge roughness (LER) and line-width roughness (LWR) is a key issue in addressing the growing challenge of device variability in large-scale integrations. The accurate characterization of LER and LWR forms a basis for this effort and mostly hinges on reducing the effects of noise inherent in experimental results. This article reports how a power spectral density (PSD) is affected by a statistical noise that originates from the finiteness of the number N-L of available samples. To achieve this, the authors numerically generated line-width data using the Monte Carlo (MC) method and assuming an exponential autocorrelation function (ACF). By analyzing the pseudoexperimental PSDs obtained using the MC data, they found that the standard deviation eta of normalized analysis errors was determined by the total number N-ALL of width data used in each analysis, regardless of N-L and the number N of width data in each line segment. The authors found that eta decreased with N-ALL approximately in inverse proportion to N-ALL(3/4). It is noteworthy that they could obtain accurate results even in the case of N-L=1 as long as N-ALL was sufficiently large, although the distribution of PSDs was large due to a large statistical noise. This resulted from the fact that the PSD distribution was not completely irregular, but centered at the true value and that the best-fitted PSD accordingly approached the true one with an increasing N. On the other hand, eta at a fixed N-ALL decreased with the ratio Delta y/xi of an interval Delta y of width data to a correlation length xi, approximately in inverse proportion to (Delta y/xi)(3/8). As a result, N-ALL at a specified eta decreased with Delta y/xi in inverse proportion to the square root of Delta y/xi in the case when Delta y/xi was 0.3 or smaller. Beyond this threshold of Delta y/xi, the authors needed to increase N-ALL markedly to achieve the same accuracy of analyses. This comes from a decrease in the range of the PSD with an increasing Delta y/xi and a subsequent loss of sensitivity of the PSD to the change of xi. Based on these results, they established guidelines for accurate analyses as follows: Delta y/xi <= 0.3 and N-ALL >= A eta(-4/3)(Delta y/xi)(-1/2), where A is 1.8 x 10(2) for xi and 7.2 x 10(1) for the variance of widths, respectively. Equivalently in terms of the total measurement length L-ALL, instead of N-ALL, the guidelines are given in Delta y/xi <= 0.3 and L-ALL/xi >= A eta(-4/3)(Delta y/xi)(1/2) using the same A's as those of N-ALL. Being expressed in universal forms like these, the guidelines of this study can be applied to many practical problems beyond LER and LWR to accurately analyze PSDs, as long as the stochastic processes have exponential ACFs. (C) 2010 American Vacuum Society. [DOI: 10.1116/1.3499647]