Here's a very superficial answer to your challenging question but one that affects practical systems. If you start with a so-called 'hard' biometric such as fingerprint, palm vein, iris or face, you have a well characterized confidence, which can be interpreted using a DET curve or similar analyses. When you integrate a so-called 'soft' biometric to the hard biometric, for example, fusing a face score with a behavioral score, one question to ask is whether you gain any additional confidence in your binary decision -- is this person who they claim to be? Simple stated, you only get additional confidence if the new information contributes a comparable amount of information to your fusion model. A face system that gives 1 false non-match for every 100,000 matches does not benefit much from additional behavioral data that gives a false non-match for ever 100 matches. But that much is obvious. So when and why would you add a soft biometric to a hard one? One opportunity might be in anti-spoofing. For example, a presentation attack on a face system might benefit from a behavioral system that detects glitches in spoken answers to question due to underlying stress of the attacker. Another example might be in establishing additional characteristics of an enrollee who might blink at a given rate through the matching process. Faster blinking might indicate stress that would sound an alarm related to those entering the system adjacent to the enrollee. But simple addition of a hard and soft biometric doesn't typically result in more confidence -- the opportunities might come from broader questions.