This thesis concerns Integrated Information Theory (IIT), a branch of information theory aimed at providing a fundamental theory of consciousness. At its core, lie two powerful intuitions:

• That a system that is somehow more than the sum of its parts has non-zero integrated information, Φ; and

• That a system with non-zero integrated information is conscious.

The audacity of IIT’s claims about consciousness has (understandably) sparked vigorous criticism, and experimental evidence for IIT as a theory of consciousness remains scarce and indirect. Nevertheless, I argue that IIT still has merits as a theory of informational complexity within complexity science, leaving aside all claims about consciousness. In my work I follow this broad line of reasoning: showcasing applications where IIT yields rich analyses of complex systems, while critically examining its merits and limitations as a theory of consciousness.

This thesis is divided in three parts. First, I describe three example applications of IIT to complex systems from the computational neuroscience literature (coupled oscillators, spiking neurons, and cellular automata), and develop novel Φ estimators to extend IIT’s range of applicability. Second, I show two important limitations of current IIT: that its axiomatic foundation is not specific enough to determine a unique measure of integrated information; and that available measures do not behave as predicted by the theory when applied to neurophysiological data.

Finally, I present new theoretical developments aimed at alleviating some of IIT’s flaws. These are based on the concepts of partial information decomposition and lead to a unification of both theories, Integrated Information Decomposition, or ΦID. The thesis concludes with two experimental studies on M/EEG data, showing that a much simpler informational theory of consciousness – the entropic brain hypothesis – can yield valuable insight without the mathematical challenges brought by IIT.