Gerhard Tintner’s mathematical research article brings canonical correlation, principal components, weighted regression, and discriminant analysis into a common framework of constrained optimization. Its contribution is a systematic account of their formal relations: different statistical objectives can produce closely related matrix equations when their coefficient vectors, quadratic forms, and normalization conditions are identified appropriately.
We will only deal with estimation problems, and not endeavour to derive distributions.
This restriction defines the argument’s scope. Tintner works with sample relationships as estimates of population relationships, without claiming that algebraic similarities establish identical sampling distributions. After introducing matrix notation, he develops a general stationary-value problem, distinguishes two principal special cases, and then interweaves further specializations with the four statistical applications.
The general theory starts from bilinear and quadratic forms involving several vectors and two sets of matrices. Combining these forms with scalar coefficients, subsequently interpreted as Lagrange multipliers, yields homogeneous linear equations for stationary values. Nontrivial solutions require the system’s determinant to vanish. Case A identifies the two matrix sets; Case B gives all vectors the same dimension and reduces the matrix sets to two fixed matrices. These restrictions supply the article’s organizing scheme.
Further specialization is necessary to give the required solutions.
The point is methodological: the general equations become statistically useful only after an application specifies what is optimized and what is held constant. Canonical correlation enters through the two-vector specialization of Case A. Tintner maximizes the correlation between linear combinations of two groups of variables while fixing both variances at unity. Eliminating one coefficient vector produces an equation equivalent to the single-vector version of Case B. This reduction is the central bridge between canonical correlation and the other methods.
Principal components enter Case B through a covariance matrix paired with the identity matrix. Maximizing component contributions while reproducing the original variances and covariances leads to a characteristic-root equation. Tintner distinguishes his covariance-based presentation from Hotelling’s use of standardized variables and a correlation matrix. Girshick’s related formulations broaden the connection: minimizing error variance subject to unit total variance, and maximizing a component’s squared correlations with the original variables, also fall within the single-vector framework.
Weighted regression gives the two matrices a different statistical meaning. Observations consist of systematic parts plus errors, and the aim is to estimate linear relations among the systematic parts when the error covariance matrix is given. Tintner minimizes a least-squares criterion weighted by the inverse error covariance matrix. After eliminating the unknown systematic deviations and imposing orthogonality and normalization conditions, the remaining optimization belongs to Case B, with sample covariance and error covariance as its two matrices.
The formal equivalence of the method of principal components and the method of weighted regression has been noted by Geary (1948).
This acknowledgment locates Tintner’s contribution in the integration of existing connections. Principal components and weighted regression retain distinct objectives and assumptions, but their estimating equations share a structure. The earlier reduction of canonical correlation extends that network of relations; Tintner also recalls weighted regression’s connection with Bartlett’s multivariate analysis of variance.
The final application makes discriminant analysis a further specialization. For two groups, Tintner maximizes the squared linear combination of their mean differences subject to unit variance. The outer product of the mean-difference vector supplies a singular matrix, placing the problem in subcase B₂.
Hence the formal analogy with subcase $A_1$ (canonical correlation), noted above, holds also for discriminant analysis.
The article closes by completing this chain of reductions rather than by proposing a new empirical procedure. Its relevance lies in showing how distinct estimation problems can be organized around shared stationary equations and determinant conditions. Tintner’s opening hope that formal relations might suggest new methods remains a prospect; the demonstrated achievement is a compact algebraic map of four established approaches.
This work was divided into 9 sections when it entered the library's research corpus—an apparatus for search and citation, not necessarily the author's own table of contents. Each title opens its summary.
Put a question to this work; the Librarian answers from its 9 sections and cites the passage.
Ask the Librarian