Introduction
Geophysical and astrophysical time series, such as isotope concentrations in ice cores or historical sunspot records, are fundamental to understanding Earth's climate and solar dynamics. However, these records are frequently plagued by gaps and irregular sampling. This poses a significant barrier to modern non-linear analysis techniques—such as calculating Lyapunov exponents or attractor dimensions—which strictly require equidistant data points. Traditional linear interpolation often fails to capture the chaotic, non-stationary nature of these natural systems, effectively "smoothing out" the very dynamics scientists wish to study.
The Manifold Approach: Beyond Linear Interpolation
The researchers propose a "neuromathematical" solution based on the premise that complex physical data often resides on a low-dimensional manifold. Even if a dataset appears high-dimensional or noisy, its underlying physics may be governed by a few active degrees of freedom. By modeling the data as a geometric shape (a manifold) rather than a simple sequence of numbers, missing values can be reconstructed based on the system's inherent geometry.
As shown in Figure 1, the method identifies the manifold (line $M$ or curve $f(t)$) that best fits the available data. When a point is missing (represented by the line $L_x$), its value is "recovered" by finding the intersection of the manifold and the missing coordinate's axis.
Methodology: Embedding and Neural Networks
The implementation utilizes a neural network system called "Famaster-2" to execute three modeling strategies:
- Linear Models: A generalization of Principal Component Analysis (PCA) capable of handling missing coordinates.
- Quasi-linear Models: Utilizing cubic splines or polynomials to create a flexible "skeleton" for the data.
- Essentially Non-linear Models (Self-Organizing Curves - SOC): Based on Kohonen self-organizing maps, these curves iteratively learn the complex distribution of the data.
A critical component of this approach is Takens’ Embedding Theorem. Instead of treating missing points in isolation, the system creates "vectors" of consecutive measurements. This "embedding" allows the neural network to capture the dynamics of the system—the trajectory of how it moves from one state to another. By understanding the path, the network can accurately predict the missing segment of that path.
Experimental Results: Sunspots and Isotopes
The method was rigorously tested on several real-world datasets. In one experiment, the researchers took the Annual Wolf Index (a record of sunspot activity) and intentionally deleted 50% of the data to see if the model could recover the signal.
As illustrated in Figure 2, the Self-Organizing Curve (SOC) model successfully recovered the peaks and troughs of the 11-year solar cycles. The recovered data (thick lines) aligns closely with the real data (thin lines), demonstrating that the model preserved the underlying physical "memory" of the solar system. Similar success was achieved with cosmogenic isotopes $^{14}C$ (radiocarbon) and $^{10}Be$ (beryllium), where quasi-linear models filled gaps while maintaining the statistical and fractal characteristics of the original records.
Conclusion: Unlocking Advanced Analysis
The ultimate value of this gap-filling technique is that it enables advanced scientific inquiries that were previously impossible. For instance, after repairing a radiocarbon record spanning 1510–1954 AD, the authors successfully performed a multi-fractal analysis. They calculated a stable $f(\alpha)$ spectrum, revealing a wide range of multi-fractal properties (from 1.0 to 2.2). This provides a new, high-resolution benchmark for studying solar-terrestrial relationships, proving that manifold-based reconstruction is a powerful tool for salvaging and enriching incomplete historical data.
Summary updated