Reproduction on OS X 10.10.4 and Numpy 1.10.2: https://gist.github.com/ihaque/882dc3fbc1731f88abb2

np.linalg.eigh computes incorrect eigenvalues/eigenvectors for matrices with at least 560 rows/columns, when built against Accelerate in OS X 10.10.4 (maybe other versions?). Everything is fine if matrices are 559 rows/cols or below, and the results are OK if using Numpy built against a different BLAS such as MKL.

np.linalg.eig and scipy.linalg.eigh are not affected by this problem.

This may be a bug in OS X 10.10.4's Accelerate; @rmcgibbo could not repro on 10.10.3: https://gist.github.com/rmcgibbo/945b6a8cc0a2f9ca21ae