Multimodal biometric authentication (BA) has shown perennial successes both in research and applications. This paper casts a light on why BA systems can be improved by fusing opinions of different experts, principally due to diversity of biometric modalities, features, classifiers and samples. These techniques are collectively called variance reduction (VR) techniques. A thorough survey was carried out and showed that these techniques have been employed in one way or another in the literature, but there was no systematic comparison of these techniques, as done here. Despite the architectural diversity, we show that the improved classification result is due to reduced (class-dependent) variance. The analysis does not assume that scores to be fused are uncorrelated. It does however assume that the class-dependent scores have Gaussian distributions. As many as 180 independent experiments from different sources show that such assumption is acceptable in practice. The theoretical explanation has its root in regression problems. Our contribution is to relate the reduced variance to a reduced classification error commonly used in BA, called Equal Error Rate. In addition to the theoretical evidence, we carried out as many as 104 fusion experiments using commonly used classifiers on the XM2VTS multimodal database to measure the gain due to fusion. This investigation leads to the conclusion that different ways of exploiting diversity incur different hardware and computation cost. In particular, higher diversity incurs higher computation and sometimes hardware cost and vice-versa. Therefore, this study can serve as an engineering guide to choosing a VR technique that will provide a good trade-off between the level of accuracy required and its associated cost.