Rebonato–Jäckel correlation repair
Constructs valid correlation matrices from inconsistent inputs using spectral or hyperspherical methods.
How it works
In the spectral method, decompose the symmetric input along its eigenvectors, drop modes with negative eigenvalues and form a matrix from the remaining modes. Adjust the rows and columns so all diagonal entries are one. The same source also describes a separate hyperspherical optimisation over unit vectors.
Worked example
Take correlations of 0.8, 0.8 and −0.8 for the three pairs in a three-asset example. The assembled matrix has eigenvalues −0.6, 1.8 and 1.8. Clipping and rescaling produces 0.5, 0.5 and −0.5 and a zero eigenvalue. Pairwise estimates on different dates can create this pattern.
Validation and limitations
Verify symmetry and spectral tolerance and examine how much important pairwise inputs moved. Rescaling after spectral clipping does not solve the nearest-correlation objective in general. The comparison with Higham and a fixed-entry repair shows the different outputs and their hedging implications.