Vrhel, Gershon and Iwan [Vrhel et al. 1994] have recently re-measured the chromaticity coordinates of 64 Munsell chips, 120 Du Pont paint chips and 170 reflectance spectra from various man--made and natural objects using a high precision spectrometer. This work is primarily targeted towards image processing engineers, but it provides an interesting perspective on the problem of measurement and representation of distributions photon quanta -- the very problem which the visual system must solve.
Vrhel's data suggest that at least four principal components are necessary to reduce and then reproduce complex spectral data from real world objects with reasonable accuracy (see Figure 14). This result is indicative that the perceptual color space needs to have at least 4 orthogonal dimensions in order to have a reasonably stable metric. In fact, Vrhel's analysis notes that substantial accuracy is once again gained when a fifth principal component is added as a basis vector for these reflectance spectra.
Figure 14. CIE chromaticity diagrams showing chromaticity errors on reproducing 354 naturally occurring reflectance spectra from (A) four principal components, and from (B) three principal components [from Vrhel et al, 1994, page 521].
In general there are only three cone types sensitive to differences in wavelength of visible light. We will choose to ignore for the moment the recent work of Neitz and colleagues [Neitz 1990, Neitz 1993] who have shown conclusive evidence of four cone types in a few human retina. Then does this imply that there can only be three orthogonal dimensions of color information available to the system? That would only be the case if the contrast sensitivity functions of the individual cone types could be linearized with respect to each other, if the resultant color was a linear combination of the cone outputs, and if positional information was not shared during the construction of the composite color output. It is likely that all three of these predicates are false. This gives the visual system the potential to construct at least four and possibly as many as seven orthogonal dimensions from the interactions between the three cone types.