Frequently Asked Questions

This section addresses common questions and issues when using pydelt.

Multivariate Derivatives

Q: Why do I get non-zero gradients at points where the gradient should be zero?

A: This is due to numerical smoothing in interpolation methods. For example, with the function f(x,y) = (x-y)², the gradient should be zero along the line x=y, but numerical methods will show small non-zero values due to smoothing effects.

Solution: - Use higher resolution sampling near critical points - Reduce smoothing parameters (but beware of overfitting) - Validate against analytical solutions when possible - Consider neural network methods with automatic differentiation

Q: How accurate are the multivariate derivatives?

A: Accuracy depends on several factors: - Function smoothness: Smooth functions give better results - Sampling density: More data points improve accuracy - Interpolation method: LLA/GLLA typically most accurate, LOWESS/LOESS moderate - Distance from boundaries: Accuracy decreases near domain edges

Q: When should I use different interpolation methods for multivariate derivatives?

A: Choose based on your needs: - SplineInterpolator: Good balance of speed and accuracy - LlaInterpolator/GllaInterpolator: Best accuracy for smooth functions - LowessInterpolator: Robust to outliers, moderate accuracy - NeuralNetworkInterpolator: Best for exact derivatives and complex functions

Q: Can I compute mixed partial derivatives?

A: Traditional interpolation methods approximate mixed partials as zero. For exact mixed partials, use neural network methods with automatic differentiation:

from pydelt.multivariate import NeuralNetworkMultivariateDerivatives

# This will provide exact mixed partial derivatives
nn_mv = NeuralNetworkMultivariateDerivatives()
nn_mv.fit(input_data, output_data)

Numerical Accuracy

Q: Why do my derivatives look “smoothed out” compared to the analytical solution?

A: All numerical interpolation methods apply some degree of smoothing to handle noise and ensure stability. This smoothing: - Reduces sharp features and discontinuities - Makes critical points appear “rounded” - Can mask important mathematical properties

Q: How can I improve derivative accuracy?

A: Try these strategies: 1. Increase sampling density near regions of interest 2. Reduce smoothing parameters (e.g., lower smoothing in SplineInterpolator) 3. Use analytical derivatives when available for validation 4. Choose appropriate interpolation method for your function type 5. Consider neural networks for complex multivariate functions

Q: What causes boundary effects in derivatives?

A: Near the edges of your data domain: - Interpolation has fewer neighboring points to work with - Extrapolation may be required, reducing accuracy - Edge effects from smoothing become more pronounced

Solution: Extend your sampling domain beyond the region where you need accurate derivatives.

Performance and Memory

Q: My multivariate derivative computation is slow. How can I speed it up?

A: Performance optimization strategies: 1. Reduce data size if possible while maintaining accuracy 2. Use SplineInterpolator for fastest computation 3. Pre-compute derivatives at fixed evaluation points 4. Consider neural networks for batch processing large datasets

Q: I’m running out of memory with large datasets. What should I do?

A: Memory management approaches: 1. Process data in chunks rather than all at once 2. Use lower-resolution sampling for initial exploration 3. Consider neural network methods which can handle larger datasets more efficiently

Error Messages and Troubleshooting

Q: I get “Interpolation failed” errors. What’s wrong?

A: Common causes and solutions: - Duplicate input points: Remove or average duplicate x-values - Insufficient data: Ensure you have enough points for the chosen method - Extreme smoothing: Try reducing smoothing parameters - Data scaling issues: Consider normalizing your input/output data

Q: My Hessian computation returns unexpected results. Why?

A: The Hessian implementation computes only diagonal elements (pure second derivatives). Mixed partial derivatives are approximated as zero for traditional methods. For full Hessian matrices, use neural network methods.

Q: Neural network derivatives are inconsistent between runs. Is this normal?

A: Yes, this is expected due to: - Random weight initialization - Stochastic training process - Local minima in optimization

For more consistent results, set random seeds or use multiple training runs with averaging.

Integration with Other Libraries

Q: Can I use pydelt with pandas DataFrames?

A: Yes, convert to numpy arrays:

import pandas as pd
import numpy as np

# Convert DataFrame to numpy arrays
input_data = df[['x', 'y']].values
output_data = df['z'].values

Q: How do I integrate with scipy optimization routines?

A: Use pydelt derivatives as objective function gradients:

from scipy.optimize import minimize
from pydelt.multivariate import MultivariateDerivatives

# Fit derivatives
mv = MultivariateDerivatives(SplineInterpolator)
mv.fit(input_data, output_data)
gradient_func = mv.gradient()

# Use in optimization
result = minimize(objective_func, x0, jac=gradient_func)

Q: Can I use pydelt with JAX or other autodiff libraries?

A: pydelt focuses on interpolation-based derivatives. For automatic differentiation, use JAX, PyTorch, or TensorFlow directly. However, you can use pydelt to validate autodiff results or handle cases where analytical functions aren’t available.

Getting Help

Q: I found a bug or have a feature request. Where should I report it?

A: Please report issues on the GitHub repository: https://github.com/yourusername/pydelt

Q: How can I contribute to pydelt?

A: Contributions are welcome! See the contributing guidelines in the repository for details on: - Code style and testing requirements - Documentation standards - Pull request process

Q: Where can I find more examples?

A: Check out: - The examples section in this documentation - The /local/tests/ directory for visual test examples - The GitHub repository for additional demos and notebooks