Nonlocal Models

For centuries, our understanding of the physical world has been guided by the principle of local action—the idea that what happens at a point is determined solely by its immediate surroundings. This concept is the foundation of classical continuum mechanics and has served us magnificently in countless engineering applications. However, this elegant picture breaks down dramatically when we confront phenomena at the edge of its applicability, such as the infinite stress predicted at a crack tip or the unphysical behavior seen in simulations of material failure. These failures reveal a critical knowledge gap, signaling that the assumption of locality is not a universal truth but a powerful approximation with clear limits. This article delves into the world of nonlocal models, a more profound framework that rectifies these shortcomings. It begins by exploring the core tenets and mathematical formulations in "Principles and Mechanisms," explaining how introducing an intrinsic length scale heals the pathologies of local theories. Following this, "Applications and Interdisciplinary Connections" will showcase the remarkable power and breadth of the nonlocal perspective, demonstrating its essential role in fields ranging from fracture mechanics and nanotechnology to tumor modeling and atmospheric science.

In the grand tapestry of physics, some of the most beautiful threads are the principles we once held as absolute, only to discover they are brilliant approximations of a deeper, more subtle reality. One such thread is the idea of ​​locality​​. For centuries, from Newton to the engineers of the 20th century, continuum physics has been built on a beautifully simple foundation: the state of a material at a point in space—its stress, its temperature, its velocity—is determined solely by what is happening at that exact point. This is the ​​principle of local action​​. It tells us that to know the force within a tiny piece of steel, you only need to know how that infinitesimal piece is being stretched or squeezed. You don't need to ask its neighbors.

This assumption is the bedrock of classical theories like linear elasticity and the Navier-Stokes equations for fluid flow. It imagines a material as a collection of independent points, each responding to its own infinitesimal environment. And for an enormous range of problems, from designing bridges to predicting weather patterns, this local picture works magnificently. The reason it works so well is an implicit assumption of ​​scale separation​​. We assume that the microscopic world of atoms, crystal grains, or molecules is so much smaller than the macroscopic world of our interest that we can safely average it all away into smooth, point-wise properties. But what happens when the universe decides not to be so neat? What happens when the clear separation between the micro and macro worlds begins to blur?

Nature has a way of creating situations where the local approximation breaks down, sometimes with catastrophic consequences. The most dramatic example is the tip of a crack. If you take the equations of classical, local elasticity and apply them to a body with a sharp crack, the theory predicts that the stress right at the tip is infinite. This is, of course, physically impossible. An infinite stress would require infinite energy, and it signals that our theory has been pushed beyond its limits. The equations are screaming at us that something is wrong with the idea of a "point" at the crack tip.

A more subtle but equally profound failure occurs in materials that soften, a common behavior in soils, concrete, and some polymers. Imagine pulling on a bar of such a material. At first, it resists, but after reaching a peak strength, a small zone begins to weaken and stretch more than the rest. All subsequent deformation concentrates, or localizes, into this narrow band, which gets progressively weaker until the bar snaps. When we try to simulate this with a computer model based on local physics, a disaster occurs. The localization band pathologically shrinks to the width of a single computational element. If we refine our mesh to get a more accurate answer, the band just gets even narrower, and the total energy dissipated to break the bar spuriously drops to zero. This isn't just a numerical glitch; it's a symptom of a deep malady in the underlying equations. The problem is said to have lost ​​ellipticity​​, becoming ill-posed and losing its predictive power.

This breakdown is not confined to solids. In a dilute gas flowing through a microchannel, if the channel's width $L$ is not much larger than the average distance a gas molecule travels between collisions (the ​​mean free path​​ $\lambda$), the gas stops behaving like a simple fluid. The ratio $Kn = \lambda / L$, known as the ​​Knudsen number​​, tells us when we are in trouble. When $Kn$ is not vanishingly small, the classical fluid dynamics equations fail, predicting things like zero fluid velocity at the walls, when in reality the gas slips past.

In all these cases—the crack tip, the softening band, the micro-channel flow—the common theme is the loss of scale separation. The "macroscopic" fields of stress or velocity are changing dramatically over distances comparable to the material's own internal, microscopic length scale. The neighborhood matters.

The remedy is as elegant as it is profound: we must abandon the tyranny of the point. A ​​nonlocal model​​ proposes that the state of a material at a point depends not just on that point itself, but on a weighted average of the state within a surrounding neighborhood. We replace the idea of a material point with that of a material neighborhood.

This simple shift introduces a new, fundamental character into our physical description: an ​​intrinsic length scale​​, often denoted by $\ell$. This isn't the size of the object or a computational parameter; it is a true material property, like density or stiffness, that characterizes the range of interaction within the material. This length is physically rooted in the microstructure: it might be related to the crystal grain size in a metal, the aggregate size in concrete, or the mean free path in a gas. By embedding this length scale directly into the constitutive law, the model now has a built-in ruler to measure itself against, allowing it to capture the real physical phenomena, like size effects, that local models miss.

How do we translate this philosophical shift into working mathematics? Two main approaches have emerged, revealing a beautiful duality in the way we can describe nonlocality.

The Integral Approach: The Honest Average

The most direct and physically intuitive way to formalize nonlocality is to write it down as an explicit average. In this view, the stress $\boldsymbol{\sigma}$ at a point $\mathbf{x}$ is given by an integral of the local stress response over the entire body.

Here, $\mathbf{C}:\boldsymbol{\epsilon}(\mathbf{x}')$ is the stress that a local theory would predict at point $\mathbf{x}'$. The ​​kernel function​​ $\alpha(\|\mathbf{x}-\mathbf{x}'\|)$ is a weighting factor that depends on the distance between the "cause" point $\mathbf{x}'$ and the "effect" point $\mathbf{x}$. It is typically chosen to be large for nearby points and decay to zero for distant points, with its effective range determined by the intrinsic length $\ell$. For this to be a proper weighted average, the kernel is usually normalized such that its integral over all space is one.

The effect of this integral is to "smear" or smooth out the physical fields over the length scale $\ell$. The infinite stress at a crack tip is regularized into a large but finite value. The pathological localization band is forced to have a finite width comparable to $\ell$, and the dissipated energy becomes a stable, mesh-independent quantity. The model is healed. The price, however, is computational complexity. In a simulation, each point now talks to many other points, creating a dense web of interactions that can be costly to solve.

The Gradient Approach: The Clever Shortcut

Is there a way to capture the essence of nonlocality without the computational burden of integration? For fields that are relatively smooth, a Taylor series expansion gives us a clue. The average of a function in a small neighborhood is related to the value at the center plus a correction involving its curvature, or second derivative. This inspires the family of ​​gradient-enhanced models​​.

Instead of an integral, we add new terms involving spatial gradients to the constitutive law. A famous example is the differential form of Eringen's model:

Here, $\mathbb{I}$ is the identity operator, and $\nabla^2$ is the Laplacian. The equation can be rearranged to $\boldsymbol{\sigma} = (\mathbf{C} : \boldsymbol{\varepsilon}) + \ell^2 \nabla^2 \boldsymbol{\sigma}$. The new term, $\ell^2 \nabla^2 \boldsymbol{\sigma}$, is the gradient correction. It acts to penalize sharp changes or high curvature in the stress field, effectively smoothing it out, just as the integral model did. This approach turns the algebraic constitutive law of local theory into a partial differential equation.

This is often much friendlier for computation. But this mathematical shortcut comes with a subtle and important consequence. By increasing the order of the derivatives in our system of equations (from second to fourth or higher order for the overall problem), we mathematically require additional ​​boundary conditions​​ to obtain a unique solution—a requirement that is absent in the more fundamental integral formulation.

Are the integral and gradient models merely two different ways to be nonlocal, or are they related? The connection, revealed through the powerful lens of Fourier analysis, is a cornerstone of the theory. A spatial convolution (the integral) becomes simple multiplication in Fourier space. The differential operator also becomes a simple multiplier. For the two models to be truly identical, their Fourier space multipliers must match for all wavelengths.

This leads to a remarkable conclusion: the differential gradient model is exactly equivalent to the integral model only if the integral kernel has one very specific mathematical form—the ​​Yukawa potential​​ (or screened-Coulomb potential), $\alpha(r) \propto \exp(-r/\ell)/r$ in three dimensions. For any other kernel, such as a Gaussian, the gradient model is only an approximation of the integral model, valid for long wavelengths (i.e., slowly varying fields).

This approximation has a weakness. A classic "paradox" arises when applying the strain-driven differential model to a simple cantilever beam with a point load at its tip. According to the laws of equilibrium, the bending moment inside the beam must be a linear function, meaning its second derivative is zero. The gradient term $\ell^2 \nabla^2 M$ in the constitutive law vanishes identically! The model paradoxically collapses back to the purely local theory, predicting no nonlocal effects whatsoever, and the boundary value problem becomes ill-posed. The more fundamental integral model does not suffer from this pathology. It's a beautiful, cautionary tale about the limits of mathematical approximations.

The journey into nonlocality reveals a richer, more interconnected physical world. It teaches us that at small scales, or during rapid changes, points are not isolated entities but are in constant communication with their neighbors. This conversation can happen across space, as we've seen, but it can also happen across time. Models that incorporate ​​temporal nonlocality​​, or memory, are essential when a material's internal relaxation time is comparable to the timescale of an event, a condition captured by the ​​Deborah number​​. By embracing this richer, nonlocal description, we not only fix the failures of older theories but gain a more profound and unified understanding of the material world.

Having journeyed through the principles of nonlocality, we now arrive at the most exciting part of our exploration: seeing these ideas at work in the real world. You might be tempted to think of nonlocality as an exotic correction, a minor detail for specialists. Nothing could be further from the truth. The shift from a local to a nonlocal perspective is a recurring revolution across countless scientific disciplines. It represents a deeper understanding, a recognition that the world is more interconnected than our simplest models assume. It is where our idealized, point-like descriptions of reality confront the rich, structured fabric of the underlying micro-world. Let us now tour this fascinating landscape, from the materials that build our world to the very code of life.

Imagine a solid object, perhaps a concrete beam in a bridge or a bone in your own body. How does it break? A simple, "local" way of thinking would treat each point in the material as an independent actor. It experiences a certain stress, and if that stress exceeds a certain threshold, it "fails." This is the heart of classical continuum mechanics. But if you try to build a computer simulation on this principle to predict how a crack grows, you run into a terrible, unphysical problem: the result depends entirely on the size of the computational grid you use! A finer grid leads to a sharper, more brittle failure, and in the limit, the energy required to break the material goes to zero. This is a mathematical disaster, telling us our model has missed something essential.

What it has missed is that fracture is not a point-like event. When a material like bone or soil begins to fail, a "process zone" of micro-cracks and damage forms ahead of the main crack tip. The behavior at one point is influenced by the damage in a finite neighborhood around it. The material has a memory of its own texture—the size of its grains, the spacing of its fibers, the scale of its micro-structure. Nonlocal models rescue us from the numerical nightmare by introducing an intrinsic length scale, a parameter that tells the model how far its "action at a distance" extends. By making the state of one point dependent on the average state in a surrounding region, the model becomes insensitive to the computational grid and correctly predicts that fracture consumes a finite amount of energy. This same principle that saves our predictions for bone fatigue also applies to understanding the stability of foundations built on soil, where strain can soften the material and lead to catastrophic shear bands.

This idea of nonlocal interactions regularizing sharp interfaces extends even to complex fluids. Consider certain long-chain polymer solutions, like those found in some shampoos or industrial fluids. Under shear, they can do a remarkable thing: spontaneously separate into bands flowing at different speeds, a phenomenon called shear banding. A local model predicts that the interface between these bands should be infinitely thin—a mathematical discontinuity that is physically impossible. By adding a simple nonlocal term to the equations—a "stress diffusion" term that says stress at one point can diffuse to its neighbors—the model is beautifully regularized. Not only does this smooth the interface to a finite, physical width, but it also uniquely selects the one-and-only stress at which the two bands can coexist. Nonlocality, it turns out, is the arbiter that chooses the physically correct reality from a multitude of mathematical possibilities.

The power of the nonlocal view becomes even more apparent as we zoom into the microscopic world, where the "particles" we averaged over to get our continuum theories come back to the fore. Think about heat flow. Fourier's law of heat conduction, a cornerstone of thermodynamics, is quintessentially local: the heat flux at a point is directly proportional to the temperature gradient at that very point. This works wonderfully for everyday objects.

But in the world of nanotechnology, this law breaks down spectacularly. In a tiny silicon membrane, for instance, heat is carried by quantized vibrations called phonons. If the membrane is small enough—comparable to the average distance a phonon travels before scattering (its "mean free path")—then heat transport becomes "ballistic." Phonons shoot straight across the device like bullets. The heat flux at one point no longer depends on the local gradient, but on the temperature of the surface far away where the phonons were launched. To describe this, one must abandon local models, even more sophisticated ones, and turn to a fully nonlocal description like the Boltzmann Transport Equation (BTE), which keeps track of the trajectories of the heat carriers.

A similar story unfolds in the quantum realm. Modern electronics are pushing the boundaries of what is possible, with devices like the Tunnel Field-Effect Transistor (TFET) that operate on the principle of quantum tunneling. To predict the current in such a device, one might use a simple, local model that assumes the electron is tunneling through a barrier created by a constant electric field. But in a real, aggressively scaled transistor, the electric field changes dramatically over the tiny tunneling distance. A more accurate, nonlocal model based on the WKB approximation calculates the tunneling probability by integrating along the entire path. This reveals a beautiful subtlety: the effective electric field that governs tunneling is a kind of harmonic mean of the field along the path. This means that regions of low field have a disproportionately large effect, dramatically reducing the tunneling current compared to the naive local estimate. The electron, in its quantum journey, feels out the entire path, and its decision to cross is a nonlocal one.

Even the way we model life's machinery is imbued with this local-to-nonlocal transition. To simulate a protein, we must account for the sea of water molecules surrounding it. A common simplification is to replace the explicit water molecules with a continuous medium that has a position-dependent dielectric constant, $\varepsilon(\mathbf{r})$. This is a local model. But we know that water molecules are dipoles that interact with each other; their orientations are correlated over a finite distance. A true description of the solvent's response to an electric field is nonlocal. The polarization at a point $\mathbf{r}$ depends on the electric field in a whole neighborhood around it. The local model, $\varepsilon(\mathbf{r})$, is only a good approximation when the electric field varies slowly compared to the solvent's correlation length. Understanding this—knowing when the separation of scales allows for a local approximation and when it breaks down—is crucial for accurate biomolecular simulations.

The nonlocal perspective finds its most profound and modern applications in the realm of complex systems, where heterogeneity and emergent behavior reign. Consider the daunting challenge of modeling a solid tumor. These are not uniform blobs of tissue; they are wildly heterogeneous, with blood vessels, cell densities, and permeability varying chaotically from place to place. The very concept of a "Representative Elementary Volume" (REV)—a small volume that is nonetheless large enough to represent the average properties of the medium, the bedrock of local continuum mechanics—can completely break down. Imaging data from tumors often reveals correlations in tissue structure that persist over very long distances.

When the REV is lost, so is locality. Transport of drugs or nutrients through such a medium can no longer be described by classical diffusion. Instead, it resembles a "Lévy flight," characterized by many small steps interspersed with rare, sudden long jumps through preferential pathways. The mathematics to describe this is found in fractional calculus, where the familiar integer-order derivatives are replaced by fractional ones. The fractional Laplacian, $(-\Delta)^{\mu/2}$, is the quintessential nonlocal operator, perfectly capturing this anomalous, super-diffusive transport that arises from the long-range structural correlations of the tumor.

We see a similar failure of local models when we look up at the sky. Air quality models that predict the dispersion of pollutants often use a local "eddy-diffusivity" theory, which assumes that turbulent eddies mix pollutants down the concentration gradient, from high to low. On a sunny day, however, the atmosphere is dominated by large, powerful thermals—coherent updrafts that can transport a puff of pollution from the surface high into the atmosphere in one swift motion. In certain layers of the atmosphere, this can lead to the bizarre phenomenon of "counter-gradient" transport, where the net flux of pollutants is up the mean concentration gradient. This is impossible in a local diffusion model. To capture it, meteorological models must incorporate nonlocal closures, such as schemes that explicitly account for the mass flux in these large, coherent eddies or theories that define mixing via a "transilient" matrix that allows for direct exchange between non-adjacent layers of the atmosphere.

Finally, the concept of nonlocality can even be lifted from the continuous space of physics into the abstract world of networks. Imagine modeling how immune cells migrate through the body's network of lymph nodes. A local model would assume cells only move between connected nodes. But we know cells can enter the bloodstream and travel to a distant, non-adjacent node. How can we model this? The answer lies in the "fractional graph Laplacian." By taking a fractional power of the standard graph Laplacian operator, we transform a sparse matrix describing local connections into a dense matrix that allows for long-range jumps. This powerful mathematical tool, derived directly from the spectral properties of the network, provides a principled way to model anomalous, nonlocal transport on complex networks, a concept with applications from immunology to social science and beyond.

From fracture mechanics to quantum electronics, from protein folding to cancer therapy, the nonlocal viewpoint offers a more profound and accurate description of our world. It teaches us that to understand the whole, we must often look beyond the point and appreciate the connections that span a finite, and sometimes vast, distance. This is not a complication, but an enrichment of our understanding, revealing the beautiful and intricate unity of nature across all scales.