Empirical orthogonal functions (EOFs) and the corresponding eigenvalues are an important ingredient in formulating the linear detection algorithm developed by North et al. <1995>. Naturally, reliability of this detection method is an important issue when only inaccurate EOFs and eigenvalues are available for constructing a detection filter. To address this issue, this study examines how sensitive the signal-to-noise ratio (SNR) is, which is an important property of a detection filter, to the accuracy of EOFs and eigenvalues. A relationship is derived between an error bound of the signal-to-noise ratio squared (SNR2) and EOF and eigenvalue errors <Kim, this issue>. The derived error bound estimate for SNR2 rather accurately delineates the actual SNR2 error due to an inaccurate filter providing that the EOFs and eigenvalues constructing a detection filter are ''reasonably'' accurate. Whether or not a set of EOFs and eigenvalues is reasonably accurate depends upon the magnitude of sampling error and is addressed in terms of the sampling insufficiency in space and time, and the cutoff eigenvalue which is the minimum eigenvalue retained for formulating a detection filter. If the cutoff eigenvalue is very low, an error bound tends to seriously underestimate an actual SNR2 error. The utility of the derived error bound is demonstrated by examining the relative uncertainty in the SNR of a hypothetical greenhouse warming signal when using two different sets of EOFs and eigenvalues. ¿ American Geophysical Union 1996