⚡ electrical engineering

Segmented Continuous Optimization

This paper introduces Segmented Continuous Optimization (SCO), a novel framework that enhances the analysis of non-stationary time-series data by performing piecewise continuous curve fitting on diverse non-linear models with $C^1$ continuity to capture both local and global signal patterns.

Original authors: Teymur Aghayev

Published 2026-07-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Teymur Aghayev

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to describe a wild, winding rollercoaster ride to a friend who has never seen it. You could try to draw the whole thing with one giant, smooth curve, but that would be a lie; the ride has sharp drops, sudden loops, and gentle hills that no single line could ever capture. Or, you could try to describe it by breaking it into tiny, straight sticks, but that would look like a jagged, robotic staircase, missing all the smooth, flowing motion of the real thing. This is the classic struggle scientists face when analyzing "signals"—those endless streams of data that represent everything from a heartbeat and brainwaves to the price of Bitcoin or the speed of a car.

For a long time, researchers have had two main tools for this job. The first is like using a giant, flexible ruler that can bend a little but mostly tries to fit a simple straight line or a basic curve (like a polynomial) across the whole picture. The second is "segmented regression," which chops the data into chunks and fits a simple line to each chunk. The problem is that real life is messy. Signals often wiggle, oscillate, or spike in ways that simple lines can't catch. They need something that can be wiggly and complex, but also smooth enough to make sense as a continuous story. This is where the new idea comes in: a way to stitch together complex, custom-shaped curves so they fit perfectly together without any bumps or jumps at the seams.

Enter Segmented Continuous Optimization (SCO), a new framework proposed by Teymur Aghayev. Think of this algorithm as a master tailor who doesn't just sew straight lines; they sew together custom-tailored, wiggly fabric patches (like sine waves, exponential curves, or polynomials) to fit a very specific, complex body of data. The magic of SCO isn't just that it uses these fancy shapes; it's that it ensures the patches are sewn together so perfectly that not only do they touch at the edges (C0 continuity), but they also flow into each other with the exact same slope, like a seamless transition from one hill to the next (C1 continuity).

The paper suggests that this method is a significant step forward because it allows scientists to use "user-defined" models—meaning you can tell the computer, "I think this part of the data looks like a sine wave," or "this part looks like a decay curve"—and the algorithm will figure out the perfect numbers to make it fit, while keeping the whole picture smooth. The authors tested this on some tricky, real-world data: the chaotic daily price of Bitcoin, the speed of a car driving on a secondary road, and the electrical signals of a human brain (EEG) while someone was doing mental math.

The results were promising. The algorithm managed to fit these complex signals with high accuracy, often keeping the error lower than the natural "noise" or volatility of the data itself. For instance, on the Bitcoin data, it achieved a local error score (SRMSE) of about 0.553, and on the brainwave data, it was even lower at 0.552. It also did this surprisingly fast, taking just a couple of seconds to process the data on a standard computer. However, the paper also notes that the method isn't a magic bullet for everything; it can get a bit unstable if the data is too flat or if the initial guesses for the shapes are way off, which is why the authors built in some "safety nets" like processing the data in overlapping batches to prevent errors from snowballing.

In short, this paper doesn't claim to have solved every problem in signal analysis, but it offers a powerful new way to look at data that is too wiggly for simple lines and too complex for rigid models. By stitching together custom curves that flow seamlessly into one another, SCO gives researchers a sharper, more flexible lens to see the hidden patterns in the noise, whether they are tracking a car's acceleration or decoding the electrical chatter of a thinking brain.

Technical Summary: Segmented Continuous Optimization (SCO)

Problem Statement

Traditional regression algorithms for time-series analysis predominantly rely on linear or polynomial functions. While effective for certain data types, these methods often fail to capture the complex, oscillatory, or transcendental behaviors inherent in many real-world non-stationary signals (e.g., biological signals, financial markets, and physical dynamics). Existing segmented approaches, such as Multivariate Adaptive Regression Splines (MARS) or High-Order Polynomial Segmenters (HOPS), often lack the flexibility to incorporate user-defined non-linear models while ensuring strict continuity between segments. There is a need for a generalized framework capable of performing piecewise continuous curve fitting on diverse non-linear models (trigonometric, exponential, etc.) while maintaining $C^1$ continuity (continuity of both value and first derivative) to accurately analyze local and global trends.

Methodology

The paper proposes Segmented Continuous Optimization (SCO), a framework designed to fit user-defined, differentiable functions to signal segments while enforcing algebraic continuity constraints.

1. Signal Preprocessing and Normalization

To enhance optimization stability and parameter interpretability, the raw signal ( $x, y$ ) undergoes standard scaling (Z-score normalization). Furthermore, each segment is locally translated. This prevents parameters from being skewed by the global magnitude of the $x$ -values, which is critical for higher-order models.

2. Optimization Core: Levenberg-Marquardt (LM)

The core optimization engine utilizes the Levenberg-Marquardt algorithm, implemented via the JAX library for high-performance computing (XLA compilation).

Parameter Bounds: To handle user-defined bounds $[L_k, U_k]$ (e.g., limiting frequency to prevent overfitting), the algorithm employs an arctanh transformation. This maps constrained parameters into an unconstrained space for the optimizer, with the Jacobian matrix adjusted via the chain rule to maintain gradient accuracy.
Objective: The algorithm minimizes the sum of squared residuals between the observed data and the user-defined function $f(x, \beta)$ within each segment.

3. Continuity Enforcement ( $C^0$ and $C^1$ )

A defining feature of SCO is the enforcement of continuity at segment boundaries.

Mechanism: For each segment $k > 1$ $k > 1$ , two specific parameters ( $\gamma_0$ $γ_{0}$ for value, $\gamma_1$ $γ_{1}$ for slope) are excluded from the direct optimization loop. Instead, they are algebraically derived during the residual calculation to satisfy:
- $C^0$ : $f(0, \beta_k) = f(\text{boundary}, \beta_{k-1})$
- $C^1$ : $f'(0, \beta_k) = f'(\text{boundary}, \beta_{k-1})$
Constraint: $C^1$ continuity requires the model to possess a linear term; otherwise, only $C^0$ continuity is enforced.

4. Stability Techniques

To address the non-convexity and error propagation caused by strict continuity constraints, SCO introduces two stability mechanisms:

Batched Optimization: The algorithm optimizes batches of $q$ segments (default $q=5$ ) simultaneously. This allows the optimizer to adjust the initial segment of a batch to find a compromised solution that accommodates continuity constraints from the previous batch.
Forward Fit: Within each batch, $q+1$ segments are optimized, but the last segment is discarded from the final fit. Instead, its optimized parameters serve as the initial guess for the first segment of the next batch. This ensures that the continuity constraints for the upcoming batch are relevant and mitigates catastrophic error propagation.

5. Parameter Unscaling

Post-optimization, parameters are mathematically "unscaled" by substituting the normalization equations back into the general function. This restores physical units and meaning to the coefficients (e.g., converting normalized amplitude back to the original signal scale).

Key Contributions

Generalized Framework: SCO enables piecewise fitting for a wide variety of user-defined non-linear models (sinusoidal, Gaussian, logistic, rational, etc.), moving beyond the limitations of polynomial-only approaches.
Algebraic Continuity: The framework enforces strict $C^1$ continuity by deriving specific parameters algebraically rather than relying solely on penalty terms in the loss function.
Stability Mechanisms: The introduction of batched optimization combined with a "Forward Fit" strategy effectively solves instability issues common in sequential non-linear least squares problems with continuity constraints.
Open-Source Implementation: The authors provide a complete, open-source implementation on GitHub, utilizing JAX for efficient compilation and execution.

Experimental Results

The framework was evaluated on diverse datasets, including Bitcoin price data, GPS velocity (UAH-DriveSet), and Electroencephalogram (EEG) signals.

Accuracy: Across 30 datasets and multiple model types, SCO achieved a local Standardized Root Mean Square Error (SRMSE) generally below 1.0, indicating fit errors lower than the data's inherent volatility.
- Best Performers: Sinusoidal (5-parameter), Gaussian, and Logistic models demonstrated the highest accuracy due to their flexibility.
- Challenges: The Fourier model showed higher instability on short, flat segments but excelled at capturing global patterns.
Efficiency: Tests on an Intel i7-12700 CPU and NVIDIA RTX 3050 GPU showed that subsequent runs (after JAX JIT compilation) are highly efficient. For example, fitting 100,000 data points with 10,000 segments took approximately 3–6 seconds depending on model complexity.
Case Studies:
- Velocity Data: SCO successfully modeled GPS velocity with a cubic model, allowing for the analytical derivation of acceleration ( $C^0$ continuous) from the velocity fit.
- EEG Data: Using a 5-parameter sine model, SCO analyzed mental arithmetic tasks, extracting local frequency, amplitude, and trend, and calculating the integral of the curve to represent accumulated voltage.

Significance and Claims

The paper claims that SCO provides a robust solution for analyzing non-stationary signals where traditional regression fails. By combining user-defined model flexibility with strict continuity enforcement and stability techniques, the framework allows researchers to:

Extract meaningful physical parameters (frequency, volatility, trend) from complex signals.
Perform analytical operations (differentiation and integration) on the fitted curves with guaranteed continuity properties.
Adapt to various scales and volatilities without manual re-tuning of the core algorithm.

The authors position SCO as a practical tool for fields ranging from medicine to econometrics, offering a balance between the interpretability of analytical functions and the adaptability required for real-world signal analysis. Future work is modestly proposed to focus on real-time implementation, higher-order continuity derivation, and expanding the library of preset models.