Bias and Spread in EVT Performance Tests

Abstract

Performance tests (measurements of probability of error) of communication systems characterized by low bit rates and high reliability requirements frequently utilize classical extreme value theory(EVT) to avoid the excessive test times required by bit error rate (BER) tests. If the underlying noise is gaussian or perturbed-gaussian, the EVT error estimates have either excessive bias or excessive variance if an insufficient number of test samples is used. EVT is examined to explain the cause of this bias and spread; experimental verification is made by testing a known gaussian source, and procedures that minimize these effects are described. Even under these conditions, EVT test results are not particularly better than those of BER.