Representational Similarity Analysis (RSA)
An analysis framework that characterizes a neural signal not by which stimulus or class it decodes but by the geometry of its representation: compute a representational dissimilarity matrix (RDM) whose entries are the distance (for example 1 − correlation, or a cross-validated Mahalanobis / crossnobis distance) between the response patterns to every pair of conditions. Two systems — brain regions, subjects, time points, or a model and the brain — are compared by correlating their RDMs, abstracting away the incommensurable individual channels or units.
RSA is second-order and model-comparison oriented: rather than asking whether a class can be decoded, it asks which candidate representational geometry (a physical-feature model, a semantic model, a deep-network layer) best predicts the measured dissimilarity structure. Cross-validated distance estimators (crossnobis) give an unbiased dissimilarity that is zero under the null, making RSA a rigorous alternative to accuracy when the question is about representational structure rather than control.
To ask whether motor cortex encodes reach direction geometrically, build the RDM of population responses to eight directions; a circular RDM (adjacent directions similar, opposite directions distant) matches a cosine-tuning model better than a categorical one.
RSA compares representational geometry, not decodability.
RSA and decoding are complementary. Decodability shows information is linearly available; RSA characterizes its format. High decoding accuracy with a mismatched RDM tells you the information is there but not in the geometry your model predicted.