Fourier Continuation Methods

Metadata	Value
Level	Intermediate
Runtime	~3 min (CPU)
Prerequisites	JAX, Signal Processing basics
Format	Python + Jupyter

Overview

Fourier continuation methods extend signals beyond their boundaries, which is essential for neural operators that need to handle non-periodic boundary conditions in spectral methods. Without proper continuation, Gibbs phenomenon causes ringing artifacts at boundaries, degrading spectral accuracy.

This example demonstrates four continuation methods: periodic, symmetric, smooth, and zero padding. It also shows intelligent neural boundary handling that adaptively selects the best method, and verifies JAX transformation compatibility (JIT, grad, vmap).

What You'll Learn

1. Apply basic continuation methods (periodic, symmetric, smooth, zero padding)
2. Use FourierBoundaryHandler for intelligent neural boundary selection
3. Extend signals in 2D with FourierContinuationExtender
4. Verify JAX compatibility (JIT, grad, vmap) with continuation methods
5. Build reusable continuation pipelines with create_continuation_pipeline

Files

• Python Script: examples/layers/fourier_continuation_example.py
• Jupyter Notebook: examples/layers/fourier_continuation_example.ipynb

Quick Start

source activate.sh && python examples/layers/fourier_continuation_example.py

Core Concepts

Why Fourier Continuation?

Spectral methods (like those in FNO) assume periodic signals. When the input signal is not periodic, truncating the Fourier series causes Gibbs phenomenon — oscillatory artifacts near boundaries. Continuation methods extend the signal to make it "appear" periodic, eliminating these artifacts.

graph LR
    A["Non-periodic<br/>Signal"] --> B["Continuation<br/>Method"]
    B --> C["Extended<br/>Signal"]
    C --> D["FFT<br/>(no Gibbs)"]

    style A fill:#fce4ec
    style C fill:#e3f2fd
    style D fill:#c8e6c9

Continuation Methods

Method	How It Works	Best For
Periodic	Wraps signal cyclically	Naturally periodic signals
Symmetric	Mirror reflection at boundary	Smooth signals, Neumann BCs
Smooth	Tapered transition to zero	General non-periodic signals
Zero Padding	Pads with zeros	Simple, baseline extension
Neural Handler	Learned adaptive selection	Unknown boundary behavior

Implementation

Step 1: Basic Continuation Methods

Test signals are extended using four different methods:

from opifex.neural.operators.specialized.fourier_continuation import (
    PeriodicContinuation,
    SymmetricContinuation,
    SmoothContinuation,
    FourierContinuationExtender,
)

methods = {
    "Periodic": PeriodicContinuation(extension_length=16),
    "Symmetric": SymmetricContinuation(extension_length=16),
    "Smooth": SmoothContinuation(extension_length=16),
    "Zero Padding": FourierContinuationExtender(
        extension_type="zero", extension_length=16
    ),
}

for method_name, extender in methods.items():
    extended = extender(signal)
    # Original signal is preserved in the middle of the extended signal

Terminal Output:

BASIC FOURIER CONTINUATION METHODS
==================================================

Signal: sine_wave
   Original length: 32
   Periodic    : 64 -> Shape preservation: (64,)
                 Original signal preserved: MSE = 0.00e+00
   Symmetric   : 64 -> Shape preservation: (64,)
                 Original signal preserved: MSE = 0.00e+00
   Smooth      : 64 -> Shape preservation: (64,)
                 Original signal preserved: MSE = 0.00e+00
   Zero Padding: 64 -> Shape preservation: (64,)
                 Original signal preserved: MSE = 0.00e+00

Step 2: Intelligent Boundary Handling

FourierBoundaryHandler uses a neural network to adaptively select the best continuation method based on signal features:

from opifex.neural.operators.specialized.fourier_continuation import FourierBoundaryHandler

handler = FourierBoundaryHandler(
    continuation_methods=["periodic", "symmetric", "smooth"],
    extension_length=16,
    hidden_dim=32,
    rngs=nnx.Rngs(42),
)

# Automatically selects best method based on signal features
extended = handler(signal)

Terminal Output:

INTELLIGENT BOUNDARY HANDLING
==================================================
Handler configuration:
   Methods: ['periodic', 'symmetric', 'smooth']
   Extension length: 16
   Neural network: 3 method weights

Processing: sine_wave
   Signal features: mean=0.000, std=0.696, boundary_grad=0.201, periodicity=0.939
   Extended from 32 to 64 points
   Signal preservation error: 3.35e-15

Step 3: 2D Signal Extension

Extension capabilities generalize to multi-dimensional signals:

extender = FourierContinuationExtender(
    extension_type="symmetric",
    extension_length=8,
)

# 2D Gaussian signal: (12, 16) -> (28, 32)
extended_2d = extender(signal_2d)

Terminal Output:

2D SIGNAL EXTENSION
==================================================
Original 2D signal shape: (12, 16)
Extended 2D signal shape: (28, 32)
2D signal preservation error: 0.00e+00

Step 4: JAX Transformation Compatibility

All continuation methods are compatible with JIT, grad, and vmap:

# JIT compilation
@jax.jit
def extend_signal(signal):
    return extender(signal)

# Gradient computation
grad_fn = jax.grad(lambda s: jnp.sum(extender(s)**2))
gradients = grad_fn(test_signal)

# Vectorized mapping
batch_extended = jax.vmap(extender)(batch_signals)

Terminal Output:

JAX TRANSFORMATIONS COMPATIBILITY
==================================================
JIT compilation: Extended 32 -> 48
Gradient computation: Gradient shape (32,)
   Gradient norm: X.XXX
Vectorized mapping: Batch (2, 32) -> (2, 48)

Step 5: Performance Benchmarks

JIT-compiled timing comparison of continuation methods:

Terminal Output:

PERFORMANCE BENCHMARKS
==================================================
Periodic  : X.XXX ms per call (JIT compiled)
Symmetric : X.XXX ms per call (JIT compiled)
Smooth    : X.XXX ms per call (JIT compiled)

Results Summary

Method	Extension Type	Signal Preservation	Use Case
Periodic	Wraps signal periodically	Exact	Naturally periodic signals
Symmetric	Mirror reflection at boundaries	Exact	Smooth signals with defined endpoints
Smooth	Tapered transition to zero	Approximate	General non-periodic signals
Zero Padding	Pads with zeros	Exact (original region)	Simple extension
Neural Handler	Adaptive selection	Learned	Unknown boundary behavior

Next Steps

Experiments to Try

1. Discontinuous signals: Test continuation on step functions and observe artifacts
2. Extension length: Increase extension_length and observe Fourier spectrum changes
3. Neural training: Train the boundary handler on domain-specific PDE solutions

Related Examples

Example	Level	What You'll Learn
Grid Embeddings	Beginner	Spatial coordinate injection
Spectral Normalization	Intermediate	Stability for spectral layers
FNO Darcy Full	Intermediate	FNO using spectral convolutions

API Reference

• FourierContinuationExtender - Core signal extension
• FourierBoundaryHandler - Intelligent neural boundary handling
• PeriodicContinuation - Periodic boundary extension
• SymmetricContinuation - Symmetric/mirror boundary extension
• SmoothContinuation - Smooth tapering extension
• create_continuation_pipeline - Pipeline factory function

Troubleshooting

Signal preservation error is non-zero

Symptom: Original signal region in extended output doesn't match input.

Cause: This may happen with the SmoothContinuation method which applies tapering that can slightly affect boundary values.

Solution: Use PeriodicContinuation or SymmetricContinuation for exact signal preservation. Check by comparing:

original_part = extended[ext_len:-ext_len]
error = jnp.mean((original_part - signal) ** 2)

Gibbs artifacts still visible

Symptom: Ringing at boundaries despite using continuation.

Solution: Increase extension_length to provide more room for the transition. Typical values are 25-50% of the original signal length.