[motivation] ## Motivation Many differential equations arising in physics and engineering contain a small dimensionless parameter $\epsilon$ — the ratio of two disparate scales. It might be the ratio of viscous to inertial forces in a fluid, the ratio of a reaction rate to a diffusion rate, or the ratio of a membrane thickness to its lateral extent. When $\epsilon$ is small, we seek an *asymptotic approximation*: an expression that becomes increasingly accurate as $\epsilon \to 0$, even though we cannot solve the equation exactly. The most natural first attempt is to expand the solution in a [power series](/page/Power%20Series) in $\epsilon$ and substitute into the equation. This is called a **regular perturbation expansion**, and it works when the character of the equation does not change as $\epsilon \to 0$. But it can fail catastrophically when $\epsilon$ multiplies the highest-order derivative. Setting $\epsilon = 0$ then *reduces the order* of the equation, and the resulting lower-order equation has fewer free constants than [boundary](/page/Boundary) or initial conditions to satisfy. No single expansion in $\epsilon$ is valid across the entire domain. The resolution is that the solution develops structure on two different scales: a "slow" outer scale where the regular expansion works, and a "fast" inner scale — a thin region of width $\delta(\epsilon) \to 0$ — where the solution varies rapidly. The method of **matched asymptotic expansions** constructs separate Poincaré expansions on each scale and stitches them together by demanding consistency in the overlap region where both are valid. This page develops the method from first principles: the formal setup, the rescaling procedure, dominant balance, matching, and composite approximations. [/motivation]

## The Problem

We consider a differential equation of the form

\begin{align*} F\!\left(x, y, y', y'', \ldots; \epsilon\right) = 0, \qquad x \in \Omega \subseteq \mathbb{R}, \end{align*}

subject to boundary or initial conditions, in the [limit](/page/Limit) $\epsilon \searrow 0$. Here $y = y(x;\epsilon)$ depends on the independent variable $x$ and parametrically on $\epsilon$. Our goal is to find a **Poincaré asymptotic expansion** of $y$ — that is, an expression of the form

\begin{align*} y(x;\epsilon) \sim \sum_{k=0}^{N} \mu_k(\epsilon) \, y_k(x) \quad \text{as } \epsilon \to 0, \end{align*}

where $\{\mu_k(\epsilon)\}$ is an asymptotic [sequence](/page/Sequence) (i.e. $\mu_{k+1}(\epsilon) = o(\mu_k(\epsilon))$ as $\epsilon \to 0$) called the **gauge [functions](/page/Function)**, and each $y_k(x)$ is independent of $\epsilon$. The gauge functions are not known in advance — they are determined by the structure of the equation itself.

In a singular perturbation problem, no single such expansion is valid for all $x \in \Omega$. Instead, we will construct *two* Poincaré expansions — one valid in an "outer" region and one in an "inner" region — and match them in their common domain of validity.

## Failure of Regular Perturbation Theory

[example: Failure of Regular Perturbation — A Singularly Perturbed IVP] Consider the initial-value problem \begin{align*} \epsilon \frac{d^2 y}{dt^2} + \frac{dy}{dt} + y = 0, \qquad y(0) = 0, \quad y'(0) = 1, \end{align*} in the limit $\epsilon \searrow 0$. We attempt a regular perturbation expansion \begin{align*} y(t;\epsilon) \sim y_0(t) + \epsilon \, y_1(t) + \cdots \quad \text{as } \epsilon \to 0, \end{align*} with $t$ fixed. Substituting $y \sim y_0 + \epsilon\,y_1 + \cdots$ into $\epsilon y'' + y' + y = 0$: \begin{align*} \epsilon(y_0'' + \epsilon\,y_1'' + \cdots) + (y_0' + \epsilon\,y_1' + \cdots) + (y_0 + \epsilon\,y_1 + \cdots) = 0. \end{align*} Expanding and grouping by powers of $\epsilon$: \begin{align*} \underbrace{(y_0' + y_0)}_{O(1)} + \epsilon\underbrace{(y_0'' + y_1' + y_1)}_{O(\epsilon)} + \cdots = 0. \end{align*} At $O(1)$, the $\epsilon y_0''$ term is absorbed into the $O(\epsilon)$ group, leaving: \begin{align*} \frac{dy_0}{dt} + y_0 = 0, \end{align*} a first-order equation with general solution $y_0 = A e^{-t}$. There is one free constant and two initial conditions — $y(0) = 0$ forces $A = 0$, while $y'(0) = 1$ forces $y_0'(0) = -A = 1$, so $A = -1$. Both cannot hold simultaneously. The cause is clear: $\epsilon$ multiplies the highest derivative $y''$. Setting $\epsilon = 0$ reduces the equation from second order (two-parameter family of solutions) to first order (one-parameter family). The expansion cannot hold uniformly on any interval containing $t = 0$. [/example]

[remark: Signals of Singular Perturbation] The most common signal of a singular perturbation is that the small parameter multiplies the highest-order derivative. But more generally, any situation where setting $\epsilon = 0$ changes the *qualitative character* of the equation — its order, its type (elliptic to parabolic), or the nature of its singularities — should be suspected as singular. The failure of the regular expansion is the diagnostic, not the cause. [/remark]

## The Inner Region: Rescaling and Dominant Balance

The failure of the regular expansion tells us that the solution must vary rapidly in some narrow region — a **boundary layer** or **initial layer** — where the highest-order derivative cannot be neglected. The goal of this section is to determine the width of this layer rigorously.

### The rescaling

We make the following **structural assumption**: there exists a point $\bar{x} \in \overline{\Omega}$ and a function $\delta(\epsilon) \to 0$ as $\epsilon \to 0$, such that the solution varies on the scale $\delta$ near $\bar{x}$. We introduce the rescaled variable and function

\begin{align*} \xi := \frac{x - \bar{x}}{\delta(\epsilon)}, \qquad Y(\xi;\epsilon) := y\!\left(\bar{x} + \delta(\epsilon)\,\xi;\,\epsilon\right). \end{align*}

This is an exact change of variables. By the chain rule:

\begin{align*} \frac{dy}{dx} = \frac{1}{\delta}\frac{dY}{d\xi}, \qquad \frac{d^2 y}{dx^2} = \frac{1}{\delta^2}\frac{d^2 Y}{d\xi^2}. \end{align*}

We then assume that $Y(\xi;\epsilon)$ admits a Poincaré expansion

\begin{align*} Y(\xi;\epsilon) \sim \mu_0(\epsilon)\,Y_0(\xi) + \mu_1(\epsilon)\,Y_1(\xi) + \cdots \quad \text{as } \epsilon \to 0, \; \xi \text{ fixed}, \end{align*}