Measuring What We Mean: Neural Properties and the Tools That Stand In for Them

We rarely observe the neural properties we care about. Instead, we compute a proxy (a spectral component, a connectivity estimate, an entropy, a dimensionality) and treat it as the property itself. This talk is about the tension that opens up in that substitution: between the target property we intend to characterize and the statistical or mathematical tool we use to stand in for it. I will draw examples from across methods work in electrophysiology and functional connectivity. The aperiodic (1/f) component is one case: amplitude modulation of an oscillation reproduces the same spectral signature attributed to a separate “aperiodic” process, so the spectrum underdetermines which mechanism is present. Connectivity and causality measures raise the same issue in a different guise, where the estimated interaction is a property of the estimator as much as of the system. My aim is not to catalogue artifacts but to take the tension seriously as a conceptual problem: a given mathematical tool carves the signal in a particular way, and that choice quietly fixes what can and cannot be concluded. I will argue that generative modeling and forward simulation are the discipline that keeps the proxy honest — that lets us state which claims about neural mechanism an observable can actually support.