In this correspondence, we focus on the performance analysis of the widely-used minimum description length (MDL) source enumeration technique in array processing. Unfortunately, available theoretical analysis exhibit deviation from the simulation results. We present an accurate and insightful performance analysis for the probability of missed detection. We also show that the statistical performance of the MDL is approximately the same under both deterministic and stochastic signal models. Simulation results show the superiority of the proposed analysis over available results. Index Terms-Minimum description length (MDL), source enumeration, performance analysis, deterministic signal. EDICS Category: SAM-PERF, SAM-SDET I. INTRODUCTION AND PRELIMINARIES MDL [1], is one of the most successful methods for determining the number of present signals in array processing and channel order detection [2]. MDL is a low complexity information theoretic criteria which does not need any subjective threshold setting usual in detection theoretic criteria. Other statistical properties, specially its asymptotic consistency [1], makes it a favorable choice for source enumeration. Unfortunately, only few approximate finite-sample performance analysis are available on the MDL method [3]-[8]. In [3], a simple asymptotic statistical model for the eigenvalues of the sample correlation matrix was used. Unfortunately, the theoretical results showed persistent bias from the simulation results [4].The next work [5], gives a computational approach for calculation of the probability of false alarm p f a . In calculating the probability of missed detection pm, the same inaccurate statistical model is used as in [3]. In [6], instead of exact performance estimation, theoretical bounds for performance were presented. A qualitative performance evaluation in terms of gap between noise and signal eigenvalues and also the dispersion of each group is given in [7]. In a recent work [8], a significantly different approach was used. Our simulation results show improved results of [8] in comparison with [3]. The performance analysis was generalized to the non-Gaussian signals while it was shown that the results reduce to the results of [5], [6] in Gaussian signals. We will show that the same modelling errors have degraded the analysis in [8] as in [3]- [6].In this correspondence, we use an approach very similar to [3]-[5] to estimate pm, including in the analysis the finite sample O(n −1 ) biases of the eigenvalues. The noise subspace eigenvalue spread is taken into account which prevents the signal subspace eigenvalues to approach σ 2 , the noise variance. The bias of the noise power estimator in MDL is calculated to get excellent match between theoretical and simulation results. We will not calculate p f a which is negligible.In the previous works, only the case of stochastic signal has been considered. Here, we use a perturbation analysis to calculate biases and variances of the eigenvalues under deterministic signal, too. Using these results, we show t...