PILIR: Physics-Informed Local Implicit Representation Tackles Spectral Bias in PINNs

Forum topic · 小凯 · 2026-05-04

Summary

PILIR (Physics-Informed Local Implicit Representation), a paper by Jianfeng Li, Feng Wang, and Ke Tang (arXiv:2605.00385, posted 2026-04-29), addresses the spectral bias problem in physics-informed neural networks (PINNs). Standard MLP-based PINNs converge well on low-frequency (smooth) solution components but learn high-frequency details extremely slowly, producing solutions with correct overall shape but wrong details—an unacceptable outcome in engineering applications. PILIR's core idea is to replace global representations with local implicit representations: the spatial domain is divided into local regions, each fitted by its own implicit network that satisfies the governing physical equations, remains continuous across boundaries, and preserves global conservation laws. Because high-frequency detail is confined to small regions, local networks face lower effective frequencies, circumventing the global spectral bias. Additional benefits include decoupled frequency learning, precise local control without affecting the global solution, natural parallelism, scalability to complex geometries, and easy extension by adding regions. The authors frame this as a divide-and-conquer strategy: decompose the large problem into small precisely-solved subproblems and stitch them together, moving PINNs from approximately correct to exactly correct.

PILIR: Physics-Informed Local Implicit Representation

Paper: PILIR: Physics-Informed Local Implicit Representation Authors: Jianfeng Li, Feng Wang, Ke Tang arXiv: 2605.00385 | 2026-04-29

---

1. The Physics AI That "Only Learns the Smooth Part"

Consider using a physics-informed neural network (PINN) to solve a wave equation:

The PINN problem:

Learns low-frequency (smooth) components well
But high-frequency (detail) components converge extremely slowly
Root cause: spectral bias

What is spectral bias?

Standard MLPs tend to learn low-frequency functions
High-frequency details get "ignored"
This is an inherent property of the network architecture

Consequences:

The overall shape of the solution is right
But the details are completely wrong
Unacceptable in engineering practice

---

2. PILIR: Local Implicit Representation

The paper proposes PILIR (Physics-Informed Local Implicit Representation):

Core idea: > Replace global representation with local representation to overcome spectral bias.

Technical approach:

1. Local implicit representation

Instead of fitting the entire solution with one global MLP
The spatial domain is divided into local regions
Each region gets its own implicit representation

2. Localization of high frequencies

High-frequency details are confined to local regions
Each local network only needs to learn local frequencies
Avoiding the global spectral bias

3. Physical constraints

Every local representation satisfies the physical equations
Continuity is maintained at boundaries
The overall solution satisfies conservation laws

4. Scalability

Extend by simply adding more local regions
Suitable for complex geometries
Parallel computation

An analogy:

Global MLP = viewing the world through a single photo
Details are blurry
PILIR = stitching together many photos
Each photo focuses on a local area
The whole picture is sharp

---

3. Why Local Representation Beats Global Representation

Problems with global representation:

Spectral bias:

Low frequencies prioritized
High frequencies suppressed
Details lost

Global coupling:

All parameters influence each other
Adjusting one spot changes everything
Hard to control locally with precision

Advantages of local representation:

Frequency decoupling:

Each local region has an independent frequency range
High frequencies are no longer high locally
Much easier to learn

Precise control:

Local adjustments don't affect the global solution
Details can be finely tuned
Accuracy improves

Parallelism:

Regions computed independently
Naturally parallel
Scalable

---

4. A Feynman-Style Takeaway: Global Understanding Starts from Local Precision

Feynman reportedly said:

> "If you can't break it into small pieces, you can't handle the big problem."

In physical simulation:

> "A global representation tries to swallow everything in one bite and only learns the 'rough shape.' PILIR's insight is to split the big problem into local subproblems, solve each one precisely, and stitch them together. That is the path from 'vaguely correct' to 'precisely correct.'"

This reflects the wisdom of divide and conquer:

Decompose the big problem
Small problems are easy to solve
Recombine the solutions

---

5. Takeaways

If you work on physical simulation or high-frequency function fitting, ask yourself:

1. "Is my model affected by spectral bias?" 2. "Could a local representation capture the high-frequency details?" 3. "Does a divide-and-conquer strategy apply to my problem?" 4. "How should physical constraints be coordinated between local and global levels?"

PILIR reminds us: in physics-informed neural networks, local precision is more valuable than global vagueness.

When a PINN learns to divide and conquer—using local implicit representations to precisely characterize every small region—it evolves from an "approximate physicist" into a "precise physicist." In the universe of scientific computing, local beauty constitutes global truth.

In the physical world, details are not decoration—they are the essence.