Statistics Glossary → Stationary point and canonical analysis (response surface)
Stationary point and canonical analysis (response surface)concept
The point on a second-order response surface where all slopes are zero, and the analysis that uses eigenvalues to determine whether that point is a maximum, a minimum or a saddle point.
The second-order model is written ŷ = b0 + x'b + x'Bx, where b holds the linear coefficients and B is the symmetric matrix of quadratic and interaction coefficients. Setting the derivatives to zero gives the stationary point x_s = -½ B⁻¹ b. Canonical analysis computes the eigenvalues and eigenvectors of B and rewrites the surface on axes rotated about the stationary point: if all eigenvalues are negative the point is a maximum, if all are positive a minimum, and if their signs are mixed a saddle point.
The size of each eigenvalue shows how sharply the surface curves in that direction. If an eigenvalue is very close to zero, the surface is almost flat in that direction (a ridge); the stationary point is then estimated unstably and can land far away, but since the response hardly changes along the ridge there are in practice many equivalent operating conditions. If the stationary point falls outside the experimental region, extrapolating to it is unreliable; ridge analysis is used instead to find the best direction within the region.
An everyday example
A bakery studies how the softness of a cake depends on oven temperature and baking time. The stationary point of the fitted surface is at 175 °C and 32 minutes, and both eigenvalues are negative, so this is the setting with the greatest softness. Had one eigenvalue been very close to zero, many settings that raise the temperature slightly and shorten the time slightly would have given almost the same result.
In MerQur
'Canonical analysis' (automatic output) in Response Surface Regression: it gives the stationary point, eigenvalues, eigenvectors and the type of surface (minimum, maximum, saddle), with a warning for small eigenvalues.
Analyses where this concept appears
Response Surface Regression (RSREG)