The Wundt Curve in Music: Compressibility in a Learned Natural Basis Predicts Aesthetic Preference

The Wundt curve---an inverted-U relating stimulus complexity to listener preference---has been a theoretical staple of musical aesthetics since Berlyne, but empirical tests have been bedeviled by sign reversals: within-genre Pearson correlations between complexity measures and preference vary in sign across genres even inside a single corpus. We develop a geometric account in which cross-genre sign reversals are consistent with---indeed, are predicted under---a common inverted-U once preference is measured as compressibility in the perceiver's learned natural (principal-component) basis of CLAP audio embeddings. On the Free Music Archive corpus (n=24,801 tracks across nine top-level genres), within-corpus quadratic regression of (1+ ) on compressibility shows negative curvature at Bonferroni-corrected z=11.77, 8.38, 5.97 for 2-, 3-, and 4-bit scalar quantization (random-basis controls z=2.25, 1.61, 0.50). A between-genre meta-regression of within-genre OLS slope on within-genre mean compressibility is directionally consistent with a common inverted-U at all three bit widths: the 2-bit cell clears the family-wise Bonferroni threshold (z=6.50, z_ =5.74 across six pre-registered cells); the 3-bit cell (z=5.44) clears the uncorrected threshold but not the family-wise correction; the 4-bit cell (z=5.25) clears neither. The theoretical prediction is basis-specific: random orthogonal rotations of the embedding space yield no structure within or across genres. We treat the music-specific scope as a feature rather than a limitation: the same mathematical framework should generalize to other modalities, but we demonstrate it here on a single well-annotated audio corpus with a single fixed encoder, so that the empirical content is sharp. Author preprint deposited for archival and citation. Draft — pending author review.