Abstract
The morbidity problem of the GM(1,1) power model in parameter identification is discussed by using multiple and rotation transformation of vectors. Firstly we consider the morbidity problem of the special matrix and prove that the condition number of the coefficient matrix is determined by the ratio of lengths and the included angle of the column vector, which could be adjusted by multiple and rotation transformation to turn the matrix to a well-conditioned one. Then partition the corresponding matrix of the GM(1,1) power model in accordance with the column vector and regulate the matrix to a well-conditioned one by multiple and rotation transformation of vectors, which completely solve the instability problem of the GM(1,1) power model. Numerical results show that vector transformation is a new method in studying the stability problem of the GM(1,1) power model.
Introduction
In Grey System Theory, the GM(1,1) power model is a robust nonlinear tool for forecasting with small datasets. However, it frequently suffers from morbidity—a state of numerical instability where the coefficient matrix used for parameter identification is "ill-conditioned." In such cases, minor measurement or rounding errors in input data can lead to wildly inaccurate results.
Mathematically, this instability is quantified by the condition number. A high condition number (e.g., >1,000) indicates an unstable model. While linear grey models have been stabilized in previous research, the nonlinear GM(1,1) power model remained challenging due to its complex exponent and time-response formulas. As shown in Figure 1, the condition number can skyrocket toward infinity as certain parameters, such as the power exponent $\alpha$, approach critical values.
Methodology: Taming the Coefficient Matrix
The authors propose a two-step geometric approach to stabilize the coefficient matrix $B$ by treating its columns as vectors in space.
1. Multiple Transformation (Scaling)
The paper proves that the condition number of a two-column matrix is primarily determined by two factors: the ratio of the lengths of its column vectors and the angle between them. By scaling these vectors so their lengths are equal (Multiple Transformation), the condition number reaches a local minimum.
2. Rotation Transformation
Scaling alone is often insufficient if the angle between the vectors is very small or very large (near $0$ or $\pi$). The authors introduce a Rotation Transformation—a specific linear combination of the vectors—to adjust the included angle $\theta$. By forcing $\theta$ toward $\pi/2$ (perpendicularity), the matrix enters a "well-conditioned" state. Table 1 (Figure 6) details the specific angular ranges required to achieve target condition numbers.
3. Parameter Recovery
Because these transformations are mathematically rigorous, the authors provide inverse formulas to "map back" the parameters calculated from the stabilized matrix to the original model, ensuring the final forecast remains valid.
Key Insights and Results
The study reveals that model stability is inextricably linked to the power exponent $\alpha$. When $\alpha$ is near 1, the vectors in the coefficient matrix become nearly parallel, causing the condition number to explode. In these scenarios, simple data scaling (Multiple Transformation) fails because it does not address the underlying geometric alignment of the vectors.
Numerical validation using socioeconomic data (Chinese high school graduate proportions, 1999–2008) produced striking results for a "difficult" exponent of $\alpha = 0.99$:
- Raw Data: Condition number = 378,570 (extreme instability).
- After Multiple Transformation (MT): Condition number = 377,160 (negligible improvement).
- After MT + Rotation Transformation (RT): Condition number = 1.0033 (near-perfect stability).
Conclusion
By reducing the condition number from nearly 400,000 to approximately 1, the vector transformation method effectively eliminates numerical instability in the GM(1,1) power model. While currently optimized for two-column matrices (covering the standard power and Verhulst models), future work will focus on extending this geometric framework to higher-order grey models with more complex multi-column coefficient matrices.
Summary updated