The generalized likelihood ratio (GLR) test is a widely used method fo
r detecting abrupt changes in linear systems and signals, In this pape
r the marginalized likelihood ratio (MLR) test is introduced for elimi
nating three shortcomings of GLR while preserving its applicability an
d generality, First, the need for a user-chosen threshold is eliminate
d in MLR. Second, the noise levels need not be known exactly and may e
ven change over time, which means that MLR is robust. Finally, a very
efficient exact implementation with linear in time complexity for batc
h-wise data processing is developed, This should be compared to the qu
adratic in time complexity of the exact GLR.