Highly Problematic Correlation Matrices
(indefinite, singular, fast-collinear)
Diagnostics, numerical therapy, relation fidelity, and analyses of laws and regularities
by Rudolf Sponsel, Erlangen
prepared in part with the support of artificial intelligence (AI).
AI transparency notice (EU AI Act)
Correlation matrices are the foundation of many multivariate methods, including factor analysis, regression analysis, structural equation modelling, and eigenvalue analysis. If a matrix is mathematically deranged, all subsequent results may be false, unstable, or impossible to interpret responsibly.
The book therefore examines highly problematic correlation matrices systematically. It focuses on indefinite matrices, positive semidefinite and nearly singular matrices, and matrices with small eigenvalues in the fast-collinearity range. The investigation follows a necessary sequence. The matrix is diagnosed first. An indefinite matrix must then be treated successfully by numerical therapy. Relation fidelity is examined after treatment. Only a treatment matrix that is at least positive semidefinite may be used for a subsequent analysis of possible laws and regularities.
The main volume explains the questions, methods, principal results, and scientific meaning. The appendix volumes document the complete tables and calculations. Short explanations and reading examples make the large numerical tables understandable without constant reference to the main volume.
8 November 1939 - 25 July 2025
I dedicate this work to the mathematician Dr Bernhard Hain, the pioneer of fast-collinearity research in correlation matrices. After eleven years of unsuccessful searching, he discovered the cause of my pathological partial correlation coefficient of 1.388. This discovery led to our book published in 1994:
Numerically Unstable Matrices and Collinearity in Psychology. Diagnosis, Relevance & Utility, Frequency, Aetiology, Therapy. Ill-Conditioned Matrices and Collinearity in Psychology. German-English.
A regular correlation matrix must be symmetric, have ones on its main diagonal, contain coefficients within the admissible range, and be at least positive semidefinite. An indefinite matrix violates the last condition: it has at least one real eigenvalue below -0.001. Such a matrix is not a mathematically admissible correlation matrix and must not be processed as if it were unproblematic.
A positive semidefinite matrix has no eigenvalue below -0.001 and at least one eigenvalue in the numerical-zero range from -0.001 to +0.001. It is mathematically admissible but may contain linear dependencies or practical null directions. A positive definite matrix has only eigenvalues above +0.001.
Three disjoint classes are used for small eigenvalues: F0 for -0.001 ≤ EV ≤ 0.001, G0 for 0.001 < EV < 0.02, and FR for 0.02 ≤ EV < 0.20. The binding FKOL count is F0 + G0 + FR. These limits are project-wide diagnostic conventions.
| Volume | Title and purpose | Language |
|---|---|---|
| Volume 1 | Main volume containing Part I Diagnostics, Part II Numerical Therapy, Part III Relation Fidelity, and Part IV Laws and Regularities. | German |
| Volume 2 | English version of the main volume. | English |
| Volume 3 | Appendix to Part I Diagnostics: the recalculated 1994 report and the evaluation of 3,569 matrices by 43 diagnostic criteria. | German and English |
| Volume 4 | Appendix to Part II Numerical Therapy: results of 20 therapy methods for 268 indefinite source matrices. | German and English |
| Volume 5 | Appendix to Part III Relation Fidelity: comparison of preserved, altered, and lost relations. | German and English |
| Volume 6 | Appendix to Part IV Laws and Regularities: GRA cases, source analyses, and result documentation. | German and English |
The appendices are not intended as repositories of unexplained numbers. Each appendix explains its abbreviations and uses concrete reading examples. Detailed methodological explanations remain in the corresponding part of the main volume.
Every matrix is examined by 43 criteria. They cover formal matrix properties, eigenvalues, definiteness, rank, inadmissible partial and multiple correlations, communalities, sums of pathological derangements, and diagnostic checks. The individual evaluations show what kind of problem is present in each matrix.
| Definiteness class | Number | Meaning |
|---|---|---|
| positive definite | 1,899 | All eigenvalues exceed +0.001. |
| positive semidefinite | 1,402 | No eigenvalue is below -0.001; at least one is in the numerical-zero range. |
| indefinite | 268 | At least one real eigenvalue is below -0.001; numerical therapy is required. |
| Total | 3,569 | Complete current diagnostic set. |
| matrices with FKOL | 2,934 | At least one eigenvalue belongs to F0, G0, or FR. |
Volume 3 combines the title matter, documentation of the generating scripts, the recalculated 1994 report with recalculations highlighted in white, and the 43-criterion evaluation of all 3,569 matrices.
Numerical therapy comprises computational methods designed to turn a mathematically inadmissible or highly problematic matrix into a matrix that is at least positive semidefinite. A therapy is not assessed merely by the disappearance of negative eigenvalues. The unit diagonal, symmetry, admissible coefficient range, and the extent of changes to the original relations must also be considered.
The book set contains 268 indefinite source matrices. Methods Th01 to Th20 were applied to every matrix. A selection score combines definiteness, the case-specific FKOL share, unit diagonal, coefficient range, and symmetry. The score is a transparent selection rule, not proof of empirical truth.
| Measure | Number | Explanation |
|---|---|---|
| source matrices | 268 | Indefinite correlation matrices in the current book set. |
| therapy methods | 20 | Th01 to Th20 for every source matrix. |
| therapy results | 5,360 | 268 source matrices multiplied by 20 methods. |
| positive definite | 1,232 | All eigenvalues exceed +0.001 after treatment. |
| positive semidefinite | 3,246 | Admissible matrices with one or more practical null directions. |
| still indefinite | 882 | The therapy did not remove the indefiniteness. |
| regularly assessable for FKOL | 5,350 | The three FKOL classes can be counted regularly. |
| not assessable for FKOL | 10 | No regular count because of complex-imaginary eigenvalues. |
| therapy matrices with FKOL | 5,104 | At least one eigenvalue belongs to F0, G0, or FR. |
| FKOL eigenvalues | 20,009 | Total number of F0, G0, and FR eigenvalues in assessable therapy matrices. |
Matrix counts and eigenvalue counts must be distinguished. The number 5,104 refers to therapy matrices with at least one FKOL eigenvalue, whereas 20,009 is the total number of FKOL eigenvalues.
A mathematically successful therapy may alter the source matrix to very different degrees. Relation fidelity therefore examines which original relations are preserved. The comparison includes rank, definiteness, eigenvalue relations, FKOL, F0, G0, and FR.
| NumTher comparisons | Number | Meaning |
|---|---|---|
| total | 5,360 | One comparison for every source-matrix and therapy combination. |
| fully assessable | 5,318 | All required comparison features are available. |
| not assessable | 42 | At least one required feature is unavailable because of complex source or therapy matrices or complex-imaginary eigenvalues. |
The appendix to Part III is divided into six areas: RD_CE_CN_KOR, Scaling, NumTher, Standard, LINA, and Other. The sixth area, Other, is currently empty.
GRA examines whether small eigenvalues and their relations may indicate laws or regularities. Such an interpretation is permissible only if the matrix used is mathematically suitable. An originally indefinite matrix must first be treated successfully and the treatment matrix must be at least positive semidefinite.
Wright's 32 × 32 tetrachoric correlation matrix is an instructive NumTher case. The original matrix has seven distinctly negative eigenvalues and is therefore indefinite. Twenty therapies were compared by the defined selection procedure. Th18 is the only method to achieve a score of 5.00 and rank 1.
After Th18 the matrix is positive semidefinite. The unit diagonal, admissible coefficient range, and symmetry are preserved. The original matrix has five FKOL eigenvalues; Th18 has 19. The necessary condition for a later independent GRA is therefore fulfilled.
Wright is discussed in detail in Part II of the main volume. In Volume 4 Wright is used only as a reading example showing how to interpret the 20 therapy rows, FKOL, selection score, and rank. The full case study is not duplicated in the appendix.
Publication of the six volumes is planned for late autumn 2026.
| Area | Status on 10 October 2026 |
|---|---|
| Part I and Volume 3: Diagnostics | Completed. Volume 3 contains 3,569 matrices, 43 diagnostic criteria, and the recalculated 1994 report. |
| Part II and Volume 4: Numerical therapy | In final preparation. The 268 × 20 evaluation set, FKOL evaluation, selection scores, and rankings are available. |
| Part III and Volume 5: Relation fidelity | Evaluation set and HTML appendix available. The six areas are organised. |
| Part IV: Laws and regularities | In preparation. The GRA cases, source analyses, and appendix are not yet complete. |
Artificial intelligence was used in part for work preparation, structural review, formulation support, and technical checks. Rudolf Sponsel retains responsibility for the scientific review, selection, evaluation, and conclusions. Inventory counts and evaluation results come from the documented project records and are not replaced by AI estimates.
| Search the IP-GIPT with Google: <search term> site: www.sgipt.org |
Changes are made occasionally; comments and criticism are welcome.
10 October 2026: English parallel version created from the confirmed book records, including the current diagnostic, numerical-therapy and relation-fidelity figures, the dedication to Bernhard Hain, the Wright example, and the status of the four parts.