Figure 3
Effect of noise on the effective number of components (x) deduced from the decrease of eigenvalues. (a, b) 100 replications of a random mixture of seven pure components (r = 7, m = 100), and scree plot showing a clear separation between the 7 significant and 93 non-significant components. (c, d) Same theoretical set after addition of Gaussian noise with a s/n ratio of 80 and a standard error of 20. The scree plot shows an `elbow' at j = 8, instead of nominally 7, and Δ[log(λ)] shows two maxima at j = 5 and 7. For clarity, jmax = 30. |