Exact Interpolation under Noise: A Reproducible Comparison of Clough-Tocher and Multiquadric RBF Surfaces

This paper presents a reproducible comparison of Clough-Tocher and Multiquadric RBF interpolants, demonstrating that while both methods overfit noisy data, the cubic interpolant offers greater stability and that noisy thermodynamic measurements can be effectively structured to recover meaningful process behavior.

The Gist

Here is an explanation of the paper using simple language, everyday analogies, and creative metaphors.

The Big Picture: Connecting the Dots in a Messy World

Imagine you are an environmental engineer trying to understand how a chemical plant works. You have a bunch of data points: temperature, pressure, and flow rate. But the data is scattered (like stars in the sky) and noisy (like a radio with static).

Your goal is to draw a smooth, continuous map (a 3D surface) that connects these dots so you can predict what happens in the empty spaces between them. This is called interpolation.

The paper compares two different "map-makers" (algorithms) to see which one does a better job:

1. The Cubic Interpolator (Clough-Tocher): Think of this as a flexible, heavy rubber sheet. If you pin it down at your data points, it stretches smoothly between them, trying to keep the overall curve gentle.
2. The Multiquadric RBF Interpolator: Think of this as a bouncy trampoline. When you pin it down at data points, it creates sharp, bouncy peaks and valleys right around those pins to match them perfectly.

The Experiment: The "Fair Fight"

The authors wanted to know: Which map-maker is better?

In the past, people compared these tools using different data or different rules, which made the results unfair. This paper sets up a strictly fair arena:

• The Arena: They created a fake, perfect world (a synthetic dataset) where they know the "true" answer.
• The Rules: They split the data into a "training set" (to build the map) and a "test set" (to check the map). They did this 40 times with different random splits to ensure the results weren't just luck.
• The Twist: They ran the test twice.
  1. The "Clean Room": The data points are perfect, with no errors.
  2. The "Dirty Room": The data points have "noise" (random errors, like measurement mistakes or sensor glitches).

The Results: What Happened?

1. In the Clean Room (No Noise)

Both map-makers did an amazing job.

• The Rubber Sheet (Cubic) was smooth and accurate.
• The Trampoline (RBF) was also accurate, sometimes even better at capturing tiny, sharp details.
• Verdict: If your data is perfect, you can use either one. They are both excellent.

2. In the Dirty Room (With Noise)

This is where things got interesting. Real-world data is rarely perfect; sensors glitch, and measurements vary.

• The Trampoline (RBF) went crazy. Because it is programmed to hit every single data point exactly, when it hit a "noisy" point (a mistake), it didn't ignore it. Instead, it created a massive, sharp spike or a deep crater right at that mistake. It tried to fit the error, which made the whole map look jagged and unreliable. It was overfitting—memorizing the mistakes instead of learning the pattern.
• The Rubber Sheet (Cubic) stayed calm. It still tried to connect the dots, but because it's a "stiffer" sheet, it didn't bend wildly just because one point was slightly off. It smoothed over the mistakes, keeping the general shape of the map intact.
• Verdict: When data is messy, the Cubic (Rubber Sheet) method is much more stable and trustworthy. The Trampoline method became chaotic and produced wild, wrong predictions.

The "Aha!" Moment: Don't Throw Away Bad Data

The most important takeaway for engineers and scientists is this:

Don't throw away "messy" data points just because they look weird.

In the past, if a sensor gave a weird reading, engineers might have deleted it as an "outlier." This paper argues that you shouldn't do that. Instead, you should treat those messy points as part of the system. By using the right interpolation method (like the stable Cubic one), you can turn that "messy" data into a useful map that reveals the true behavior of the system.

The Analogy of the "Noisy Party"

Imagine you are trying to guess the temperature of a room based on 10 thermometers.

• 9 thermometers say it's 70°F.

• 1 thermometer is broken and says 200°F.

• The Trampoline (RBF) would try to connect the dots perfectly. It would create a giant, impossible mountain of heat right where the broken thermometer is, making the whole room look like a volcano.

• The Rubber Sheet (Cubic) would see that one point is weird. It would gently stretch over it, realizing, "Okay, that one is a bit high, but the rest of the sheet says 70°F." It would give you a map that says the room is roughly 70°F, ignoring the glitch.

Conclusion

This paper is a guide for scientists on how to draw maps from messy data.

• If your data is perfect: You can use the fancy, bouncy Trampoline method.
• If your data is messy (which it usually is): Use the sturdy Rubber Sheet method. It ignores the glitches and gives you a reliable picture of reality.

The authors also provided a free, open-source computer script so anyone can repeat this experiment and see the results for themselves, ensuring that science remains transparent and reproducible.

Technical Summary

Here is a detailed technical summary of the paper "Exact Interpolation under Noise: A Reproducible Comparison of Clough-Tocher and Multiquadric RBF Surfaces" by Mirkan Emir Sancak.

1. Problem Statement

The paper addresses the challenge of visualizing and analyzing multivariate datasets (specifically 3D input spaces mapping to scalar outputs) in fields like environmental engineering, materials science, and machine learning.

• The Core Issue: Traditional 2D plotting fails to capture complex non-linear dependencies in high-dimensional data. While 3D surface plots are a solution, generating them from sparse, noisy, or irregularly sampled data is difficult.
• The Specific Gap: Existing methods often lack a unified, reproducible framework for comparing interpolation techniques. Furthermore, there is a practical dilemma: should noisy or "inconsistent" measurements in thermodynamic systems be discarded as outliers, or can they be interpolated to recover meaningful process behavior?
• The Objective: To rigorously compare two common interpolation strategies—Cubic Interpolation (Clough-Tocher) and Multiquadric Radial Basis Function (RBF) interpolation—under controlled conditions to determine their robustness against measurement noise.

2. Methodology

The study employs a rigorous, reproducible computational workflow implemented in Python (using NumPy, SciPy, Pandas, and Matplotlib) to eliminate evaluation bias.

A. Experimental Design

• Synthetic Dataset: A full factorial design was generated with 48 unique data points derived from three input variables ($X_1, X_2, X_3$) and three output variables ($Y_1, Y_2, Y_3$).
• Ground Truth: Outputs were generated using non-linear mathematical expressions (combining polynomials and trigonometric functions).
• Noise Regimes: Two regimes were tested:
  1. Noise-Free: Exact analytical values.
  2. Noisy: Gaussian noise ($\epsilon \sim N(0, \sigma^2)$) was added with varying standard deviations ($\sigma = 0.1, 1.0, 2.0$) to simulate real-world measurement uncertainty.
• Slice-wise Strategy: To visualize the 3D input space, the data was sliced by fixing one variable, reducing the problem to 2D interpolation tasks.

B. Interpolation Models

1. Cubic Interpolation: Implemented via SciPy's CloughTocher2DInterpolator. This constructs a piecewise polynomial surface ensuring derivative continuity.
2. Multiquadric RBF: Implemented via SciPy's RBFInterpolator with a multiquadric kernel ($\phi(r) = \sqrt{1 + (\epsilon r)^2}$) and a smoothing parameter set to zero (exact interpolation).

C. Evaluation Protocol (The Key Innovation)

To ensure a fair comparison, the study introduced a unified evaluation protocol:

• Identical Splits: Both methods were trained and tested on the exact same random train/test splits (70% training, 30% testing) across 40 repetitions per slice.
• Metrics: Performance was measured using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and the Coefficient of Determination ($R^2$).
• Uncertainty Quantification: Non-parametric bootstrap resampling (1,000 iterations) was used to generate confidence intervals and assess split-to-split reliability.
• Reproducibility: All experiments used a fixed random seed (42) and are fully reproducible via a single script.

3. Key Contributions

1. Unified Evaluation Framework: The paper moves beyond "visualization showcases" by establishing a protocol where algorithmic differences are isolated from data-handling artifacts. Both methods face identical geometric and sampling conditions.
2. Quantitative Evidence on Noise Sensitivity: It provides empirical proof that exact interpolation (forcing the surface to pass through every noisy node) leads to severe overfitting and out-of-sample degradation, particularly for RBF methods.
3. Diagnostic Visualization: The study introduces a multi-layered diagnostic approach combining aggregate metrics, distributional analysis (boxplots of RMSE), surface reconstruction visuals, and predicted-vs-true scatter plots to explain why failures occur.
4. Practical Implication for Engineering: It argues against discarding noisy data in thermodynamic systems. Instead, it suggests that with appropriate (potentially regularized) interpolation, these "inconsistent" points can be structured to recover physically meaningful process behaviors.

4. Key Results

A. Noise-Free Regime

• Both methods achieved high accuracy ($R^2 > 0.98$ for most cases).
• Performance Variance: The "better" method depended on the specific output function. Cubic interpolation performed slightly better for Output 1 and 3, while RBF was superior for Output 2. This indicates that in clean data, method selection is problem-dependent.

B. Noisy Regime (Critical Findings)

• General Degradation: Exact interpolation for both methods suffered significantly when noise was introduced, as the models attempted to fit the noise rather than the underlying trend.
• Cubic vs. RBF Stability:
  • Cubic Interpolation proved comparatively more stable. While errors increased, the degradation was moderate.
  • Multiquadric RBF exhibited severe instability. In difficult cases (Output 2 and 3), the $R^2$ values collapsed to highly negative numbers (e.g., $-10.8$ and $-52.2$), indicating the model performed worse than a simple mean predictor.
• Distributional Analysis: Boxplots of RMSE across repeated splits showed that RBF had much wider interquartile ranges and heavier "tails" (risk of catastrophic failure) compared to Cubic.
• Geometric Evidence: Visualizations revealed that RBF surfaces developed localized warping, amplified peaks/troughs, and ripple-like fluctuations in response to noise, whereas Cubic surfaces preserved the global topology more effectively.

5. Significance and Future Directions

• Methodological Shift: The paper demonstrates that in noisy environments, exact interpolation is often unreliable. The "best" method is not just about mathematical elegance but about robustness to measurement error.
• Call for Regularization: The results strongly motivate the use of regularized interpolation (e.g., RBF with non-zero smoothing parameters or spline smoothing) for real-world engineering data. This trades a small amount of bias for a massive reduction in variance.
• Environmental Engineering Application: The study validates a workflow where noisy sensor data from thermodynamic processes (reactors, treatment units) should not be discarded. Instead, they should be processed through robust interpolation pipelines to aid in process diagnosis and optimization.
• Future Work: The authors propose extending this framework to include:
  • Non-zero smoothing parameters (regularization).
  • Full 3D scattered interpolation (not just slices).
  • Real-world datasets with non-Gaussian noise.
  • Physics-constrained regularization (e.g., enforcing thermodynamic continuity).

In summary, this paper provides a reproducible, statistically rigorous benchmark proving that while Cubic and RBF interpolants are comparable in clean data, Cubic interpolation offers superior stability in noisy regimes, and that exact interpolation is generally unsuitable for noisy data without regularization.