Method of Matched Asymptotic Expansions

A regular perturbation expansion in a small parameter $\varepsilon$ quietly assumes the solution depends smoothly on $\varepsilon$ and that each correction stays uniformly small everywhere. Both assumptions fail the moment $\varepsilon$ multiplies the highest derivative of a differential equation. Setting $\varepsilon=0$ lowers the order of the equation, and the reduced (outer) problem no longer has enough free constants to satisfy all the boundary conditions. The discarded term was small in magnitude — but it was the only one that could resist a steep gradient, and where the solution must turn sharply it comes roaring back to leading order. That thin zone of rapid adjustment is a boundary layer, the signature of a singular perturbation problem. The method of matched asymptotic expansions is the systematic cure. Rather than one series valid everywhere, one builds two: an outer expansion in the original variable $x$ , good away from the layer, and an inner expansion in a stretched variable $X=x/\delta(\varepsilon)$ , good inside it. The thickness $\delta(\varepsilon)$ is not guessed but read off by dominant balance — chosen so the rescaled equation keeps its highest derivative at leading order. Each expansion keeps undetermined constants: the outer carries the far boundary condition, the inner the wall condition, and the remaining freedom is fixed by demanding the two agree in an intermediate overlap region. Prandtl saw this physically in 1904; Van Dyke turned it into an operational rule; the pieces then combine into a single composite approximation, uniformly valid to the order worked. The same discipline governs high-Reynolds-number flow, WKB turning points, enzyme kinetics, and semiconductor junctions — one grammar for every problem with a hidden thin scale.

Key takeaways

When a small parameter $\varepsilon$ multiplies the highest derivative, the naive $\varepsilon\to0$ limit is singular: the equation loses an order and a single power series cannot stay valid across the whole domain.
Split the domain into an outer region (the reduced equation holds) and a thin inner layer (rapid variation), expanding in each with its own gauge functions and, in the layer, a stretched coordinate.
The inner scale $\delta(\varepsilon)$ is fixed by dominant balance so the highest derivative survives: $\delta\sim\varepsilon$ for a convection–diffusion layer, $\delta\sim\varepsilon^{2/3}$ for a WKB turning point.
Join the two expansions in the overlap region by Prandtl's matching of limits or Van Dyke's rule, fixing the free constants.
Add the pieces and subtract their common part for a composite $y_{\mathrm{comp}}=y_{\mathrm{out}}+y_{\mathrm{in}}-y_{\mathrm{match}}$ , uniformly valid to the worked order.
Key idea: when $\varepsilon$ multiplies the highest derivative the limit is singular — build separate outer and stretched-inner expansions, set the inner scale by dominant balance, match them in the overlap region to fix the free constants, then add and subtract the common part for one uniformly valid composite.

The core insight

When $\varepsilon$ multiplies the highest derivative the limit is singular: build separate expansions in the outer variable and a stretched inner variable (thickness set by dominant balance), match them in the overlap region to fix the free constants, then add them and subtract the common part for a uniformly valid composite.

Where this lives today

Matched asymptotics is a daily tool wherever a problem hides a thin scale. In aerodynamics the outer potential-flow solution is still matched to a Prandtl boundary layer to predict skin-friction drag and flow separation on wings and turbine blades, cross-checked against direct numerical simulation. In systems biology the Michaelis–Menten law $v=V_{\max}[S]/(K_M+[S])$ is derived — and its error bounded — as the quasi-steady outer limit of an enzyme reaction whose inner layer is the fast binding transient. In semiconductor physics the small Debye length plays the role of $\varepsilon$ , producing the depletion layers at $p$ – $n$ junctions that make diodes and transistors work, and in gas dynamics a small viscosity smooths a shock into an inner layer of thickness $\sim\varepsilon$ . And in quantum mechanics the turning-point matching of Derivation 3 is the content of every WKB estimate of tunnelling rates and bound-state energies — from alpha decay to molecular spectra — with the $\pi/4$ phase supplying the zero-point $\tfrac12$ in Bohr–Sommerfeld quantisation. One grammar, many dialects.

The physical picture

Picture a fast-decaying transient glued onto a slow background. Far from the wall the solution tracks the smooth outer behaviour of the reduced equation; near the wall a thin sliver of rapid adjustment lets it snap down to its boundary value. The width of that sliver is the only length the small parameter can build, and dominant balance reads it off. The first figure is the case worked throughout this article — $\varepsilon y''+y'+y=0$ on the unit interval — drawn at a real, small $\varepsilon$ so you see the outer curve and the layer at once.

Figure

Boundary-layer solution of ε y″ + y′ + y = 0 with y(0) = 0 and y(1) = 1 at ε = 0.05. The smooth outer part e^(1−x) would sit at e ≈ 2.718 at the wall, but a layer of thickness δ = ε ≈ 0.05 plunges the composite y = e^(1−x) − e^(1−x/ε) down to obey y(0) = 0. Away from the wall the composite rejoins the outer curve.

Matching is conservation of information across scales. The outer solution knows everything except the fine structure of the layer; the inner knows the layer but not the level it should climb to. Their overlap — large compared with the layer yet small compared with the domain — is where each transmits the one number it is missing. The second figure makes the handshake explicit: the inner profile rises to a plateau, the outer curve falls to that same plateau, and their agreement in the middle is the matching condition.

Figure

Matching in the overlap region. The inner profile ỹ = e(1 − e^(−X)) climbs from the wall value 0 to a plateau of e ≈ 2.718 as X → ∞ — exactly the value the outer solution e^(1−x) approaches as x → 0. The shared plateau is the common part, and equating the two limits fixes the last free constant d₁ = e.

Why the highest derivative is the whole story: dropping $\varepsilon y''$ deletes precisely the term that penalises sharp curvature. The full equation is then free to turn over a distance $\sim\varepsilon$ , a turn invisible to the outer problem — and stretching $x\to x/\varepsilon$ magnifies it until the second derivative is order one again and the lost physics reappears.

The core equations

The starting point is a linear second-order boundary-value problem with $\varepsilon$ on the highest derivative — it holds every feature of the method: an order that collapses at $\varepsilon=0$ , two boundary conditions, and a sign that decides where the layer lives.

\varepsilon\,y''+a(x)\,y'+b(x)\,y=0,\qquad y(0)=A,\ \ y(1)=B,\ \ 0<\varepsilon\ll1 \qquad (1)

where: $\varepsilon$ — small parameter on the highest derivative; $a(x),b(x)$ — smooth coefficients, the sign of $a$ locating the layer; $A,B$ — boundary values at the two ends. Reading it: at $\varepsilon=0$ the $y''$ term vanishes and the second-order problem becomes first-order — it cannot honour both $A$ and $B$ . Away from the layer the solution is smooth, so we expand it as a power series in the original variable. The leading term obeys the reduced equation — a first-order ODE with one integration constant.

y_{\mathrm{out}}(x)=\sum_{n=0}^{\infty}\varepsilon^{n}\,Y_n(x),\qquad a(x)\,Y_0'+b(x)\,Y_0=0 \qquad (2)

where: $Y_n(x)$ — outer coefficients in $x$ ; the order drop — with $y''$ gone, only one boundary condition can be imposed, so the outer solution keeps the one away from the layer and abandons the other. Inside the layer nothing is smooth, so we magnify it with a stretched coordinate and expand in gauge functions $\nu_n(\varepsilon)$ set by the equation.

X=\frac{x}{\delta(\varepsilon)},\qquad y_{\mathrm{in}}(X)=\sum_{n=0}^{\infty}\nu_n(\varepsilon)\,\tilde Y_n(X) \qquad (3)

where: $X$ — the inner (stretched) coordinate; $\delta(\varepsilon)$ — the layer thickness, still to be fixed; $\nu_n(\varepsilon)$ — inner gauge functions, generally not the integer powers $\varepsilon^n$ . Reading it: an $O(\varepsilon)$ layer under $X=x/\varepsilon$ becomes an $O(1)$ region, where derivatives are large and the discarded term revives. The thickness $\delta$ is chosen — never guessed — by dominant balance: substitute $X=x/\delta$ and demand the highest derivative return at leading order beside another.

\frac{\varepsilon}{\delta^{2}}\,\frac{\mathrm{d}^2\tilde Y}{\mathrm{d}X^2}+\frac{a}{\delta}\,\frac{\mathrm{d}\tilde Y}{\mathrm{d}X}+b\,\tilde Y=0\ \ \Longrightarrow\ \ \delta\sim\varepsilon \qquad (4)

where: $\varepsilon/\delta^2$ — coefficient of the restored second derivative; $a/\delta$ — coefficient of the first derivative. Reading it: balancing $\varepsilon/\delta^2\sim a/\delta$ gives $\delta\sim\varepsilon$ and leaves $b\tilde Y$ subdominant; the rival $\delta\sim\sqrt\varepsilon$ captures no layer. The missing constants are fixed by insisting the two pieces agree in the overlap. Prandtl's statement: the outer limit of the inner equals the inner limit of the outer.

\lim_{X\to\infty} y_{\mathrm{in}}(X)\;=\;\lim_{x\to 0^{+}} y_{\mathrm{out}}(x) \qquad (5)

where: $X\to\infty$ — leaving the layer on the inner side; $x\to0^{+}$ — approaching it from the outer side. Reading it: as you march out of the layer the inner solution flattens to a constant, which must equal the value the outer solution reaches marching in. Van Dyke packaged the same requirement as a term-by-term rule on truncated expansions — what one computes to higher order.

\big[E_m\,H_n\,y\big]=\big[H_n\,E_m\,y\big] \qquad (6)

where: $H_n$ — take the inner expansion to $n$ terms, re-express in the outer variable, expand to $n$ terms; $E_m$ — the same with outer and inner swapped. Reading it: the inner expansion of the outer solution must equal the outer expansion of the inner, order by order — each order supplying one matching equation. Once matched, add the two solutions and subtract the shared overlap value they both reproduce.

y_{\mathrm{comp}}(x)=y_{\mathrm{out}}(x)+y_{\mathrm{in}}(X)-y_{\mathrm{match}} \qquad (7)

where: $y_{\mathrm{match}}$ — the common part, the shared overlap limit; $y_{\mathrm{comp}}$ — one expression valid across the whole domain. Reading it: far from the layer $y_{\mathrm{in}}\to y_{\mathrm{match}}$ and the composite collapses to $y_{\mathrm{out}}$ ; inside it $y_{\mathrm{out}}\to y_{\mathrm{match}}$ and it collapses to $y_{\mathrm{in}}$ . When $\varepsilon$ sits on the second derivative with no first-derivative term, the inner solution is no longer exponential but an Airy function — the case connecting this method to WKB.

\varepsilon^2 y''=Q(x)\,y,\quad Q(x_0)=0:\qquad y_{\mathrm{in}}\propto\mathrm{Ai}\!\left(\frac{x-x_0}{\varepsilon^{2/3}}\,[Q'(x_0)]^{1/3}\right) \qquad (8)

where: $\mathrm{Ai}$ — the Airy function, the inner solution near a simple turning point; $Q'(x_0)$ — the slope of $Q$ there, setting the $\varepsilon^{2/3}$ scaling. Reading it: where the WKB amplitude $Q^{-1/4}$ blows up, the Airy solution is the finite bridge across it.

Derivation AConstructing the composite for a convection–diffusion boundary layer

Our goal is a single formula, valid on the whole interval $0\le x\le1$ , for the constant-coefficient model $\varepsilon y''+y'+y=0$ with $y(0)=0$ and $y(1)=1$ . We solve the reduced outer problem, stretch the coordinate to resolve the wall layer, impose the wall condition, fix the last constant by matching to the outer limit, then add the two pieces while subtracting their shared overlap value so nothing is double-counted. Take $a=b=1$ ; since $a>0$ the drift sweeps information right, so the layer sits at $x=0$ and the outer solution carries the far condition at $x=1$ .

Step 1. Drop the $\varepsilon y''$ term. The reduced equation $Y_0'+Y_0=0$ has solution $Y_0=C\,e^{-x}$ . The far condition $Y_0(1)=1$ gives $C=e$ , so $Y_0=e^{\,1-x}$ . Why: the first-order reduced equation holds only one constant — enough for the far boundary, so the wall value $y(0)=0$ is deliberately left unmet.

Step 2. Resolve the wall region. With $X=x/\varepsilon$ the equation becomes $\tilde Y''+\tilde Y'+\varepsilon\tilde Y=0$ ; at leading order $\tilde Y_0''+\tilde Y_0'=0$ , solution $\tilde Y_0=d_1+d_2 e^{-X}$ . Why: $\delta=\varepsilon$ revives the second derivative at the order of the first; $b\tilde Y$ drops to $O(\varepsilon)$ .

Step 3. Impose the wall condition. At $X=0$ , $\tilde Y_0(0)=0$ gives $d_2=-d_1$ , so $\tilde Y_0=d_1(1-e^{-X})$ . Why: the inner meets the value the outer abandoned, using one constant and leaving one for matching.

Step 4. Match. As $X\to\infty$ , $\tilde Y_0\to d_1$ ; Prandtl's condition equates this to $\lim_{x\to0}Y_0=e$ . Hence $d_1=e$ , $d_2=-e$ , $\tilde Y_0=e\,(1-e^{-X})$ , and the common part is $y_{\mathrm{match}}=e$ . Why: the plateau the inner rises to must equal the value the outer descends to — the one number the two scales exchange.

Step 5. Assemble: $y_{\mathrm{comp}}=Y_0+\tilde Y_0-y_{\mathrm{match}}=e^{\,1-x}-e^{\,1-x/\varepsilon}$ . It obeys $y(0)=e-e=0$ and $y(1)=1-e^{\,1-1/\varepsilon}\to1$ as $\varepsilon\to0$ . Why: subtracting the common part removes the double-counted plateau, leaving a formula that reduces to the outer for $x\gg\varepsilon$ and the inner for $x\lesssim\varepsilon$ . *The exact solution has roots $r_+\to-1$ , $r_-\to-1/\varepsilon$ for small $\varepsilon$ , reproducing $e^{\,1-x}-e^{\,1-x/\varepsilon}$ to leading order. At $\varepsilon=0.05$ the composite matches to $\sim10^{-4}$ inside the layer and $O(\varepsilon)$ in the bulk.*

Go deeper · Derivation B · 1 minReading the layer thickness off dominant balance

Derivation BReading the layer thickness off dominant balance

The aim is to justify, rather than assert, why the stretched scale is $\delta\sim\varepsilon$ for a convection–diffusion layer and $\delta\sim\varepsilon^{2/3}$ for a turning point. The recipe: substitute a trial scale, tabulate how every term scales with it, and pick the exponent that brings the highest derivative into a genuine balance with at least one other term. A wrong exponent gives an inner equation that either loses the second derivative again or has no non-trivial layer solution.

Step 1. Substitute $X=x/\delta$ into $\varepsilon y''+a y'+b y=0$ ; the three terms scale as $\varepsilon/\delta^2$ , $a/\delta$ , and $b$ . Only the ratios matter, so list the balances between $\varepsilon/\delta^2$ and each rival. Why: dominant balance requires at least two terms of the same order with the rest smaller — and the second derivative must be one of them, or the singular character is lost.

Step 2. Balance diffusion against convection: $\varepsilon/\delta^2\sim a/\delta$ gives $\delta\sim\varepsilon$ , and then $b\sim O(1)$ is smaller by a factor $\varepsilon$ — consistent. The rival $\varepsilon/\delta^2\sim b$ gives $\delta\sim\sqrt\varepsilon$ , but then $a/\delta\sim a/\sqrt\varepsilon$ dominates both other terms — a contradiction. Why: a legitimate distinguished limit needs the neglected terms genuinely smaller; $\delta\sim\varepsilon$ passes, $\delta\sim\sqrt\varepsilon$ fails.

Step 3. For $\varepsilon^2 y''=Q(x)y$ with $Q(x_0)=0$ , $Q'(x_0)\ne0$ , near $x_0$ we have $Q\approx Q'(x_0)(x-x_0)$ , so with $X=(x-x_0)/\delta$ the balance $\varepsilon^2/\delta^2\sim Q'(x_0)\,\delta$ gives $\delta^3\sim\varepsilon^2$ , i.e. $\delta\sim\varepsilon^{2/3}$ . Why: since $Q$ vanishes linearly, the diffusion term balances a term that itself scales with $\delta$ , shifting the exponent from $1$ to $2/3$ and giving the Airy scale. *The scalings $\delta\sim\varepsilon$ and $\delta\sim\varepsilon^{2/3}$ are consistent with the limiting-cases table; substituting each back gives an inner equation with the second derivative present — $\tilde Y_0''+a(0)\tilde Y_0'=0$ for the convection–diffusion case and $y''=Xy$ (the Airy equation) for the turning-point case.*

Go deeper · Derivation C · 1 minMatched Airy solution and the WKB connection formula

Derivation CMatched Airy solution and the WKB connection formula

The purpose is to show that WKB is itself matched asymptotics — the two divergent outer (WKB) forms on either side of a classical turning point are stitched together by an inner Airy solution — and to extract the $\pi/4$ phase this forces. Take $\varepsilon^2 y''=Q(x)y$ with a simple turning point at $x_0$ , $Q'(x_0)>0$ , so $Q<0$ (allowed, oscillatory) for $x<x_0$ and $Q>0$ (forbidden) for $x>x_0$ , and follow the single Airy solution to both sides.

Step 1. Away from $x_0$ the outer solutions are the WKB forms $y_{\mathrm{out}}\sim Q^{-1/4}\exp(\pm\tfrac1\varepsilon\int\sqrt{Q}\,\mathrm{d}x)$ , valid only where $|Q|$ is bounded away from zero; the amplitude $Q^{-1/4}$ diverges as $x\to x_0$ — the signal an inner region is needed. Why: the breakdown of the naive expansion at a point is exactly what matched asymptotics repairs; the divergence marks the layer.

Step 2. Rescale with $X=(x-x_0)\varepsilon^{-2/3}[Q'(x_0)]^{1/3}$ . Linearising $Q\approx Q'(x_0)(x-x_0)$ turns the equation into $\mathrm{d}^2y/\mathrm{d}X^2=X\,y$ , the Airy equation, with decaying solution $y=\mathrm{Ai}(X)$ . Why: the $\varepsilon^{2/3}$ scale makes both sides order one, and the linearised potential makes the inner equation Airy rather than anything harder.

Step 3. Match on the forbidden side. For large $X$ , $\mathrm{Ai}(X)\sim\tfrac{1}{2\sqrt\pi}X^{-1/4}e^{-\frac23 X^{3/2}}$ , and since $\tfrac23 X^{3/2}=\tfrac1\varepsilon\int_{x_0}^{x}\sqrt{Q}\,\mathrm{d}x'$ , this is exactly the decaying WKB exponential. Why: the inner tail reproduces the outer solution it must join — matching on the exponential branch.

Step 4. Match on the allowed side. For large $X$ , $\mathrm{Ai}(-X)\sim\tfrac{1}{\sqrt\pi}X^{-1/4}\sin(\tfrac23 X^{3/2}+\tfrac\pi4)$ , so the oscillatory WKB solution inherits a $-\pi/4$ phase. The upshot is the connection formula $\tfrac{2}{|Q|^{1/4}}\cos(\tfrac1\varepsilon\int_x^{x_0}\sqrt{-Q}\,\mathrm{d}x'-\tfrac\pi4)\ \longleftrightarrow\ \tfrac{1}{Q^{1/4}}\exp(-\tfrac1\varepsilon\int_{x_0}^{x}\sqrt{Q}\,\mathrm{d}x')$ . Why: the single-valued Airy function cannot pick a different phase on each side, so it fixes the relative $\pi/4$ shift — later the $\tfrac12$ in Bohr–Sommerfeld. *The asymptotic forms of $\mathrm{Ai}$ give exactly the $\pi/4$ phase; applied at both turning points of a potential well this yields the Bohr–Sommerfeld quantisation condition $\oint p\,\mathrm{d}q=2\pi\hbar(n+\tfrac12)$ .*

Limiting cases

The structure of the equation dictates the inner scaling. Four balances cover almost every problem you meet:

Case	Result	Meaning
Boundary layer, $\varepsilon y''+a(x)y'+by=0$	$\delta(\varepsilon)=\varepsilon,\ \ X=\dfrac{x}{\varepsilon}$	Convection–diffusion balance $\varepsilon y''\sim a y'$ gives an $O(\varepsilon)$ layer at the boundary where $a$ points inward.
WKB / turning point, $\varepsilon^2 y''=Q(x)y$	$\delta(\varepsilon)=\varepsilon^{2/3},\ \ X=\dfrac{x-x_0}{\varepsilon^{2/3}}$	Near a simple zero $Q(x_0)=0$ the balance $\varepsilon^2 y''\sim Q'(x_0)(x-x_0)y$ forces the Airy scale $\varepsilon^{2/3}$ .
Quasi-steady (enzyme) kinetics	$\delta(\varepsilon)=\varepsilon,\ \ \tau=\dfrac{t}{\varepsilon}$	The fast initial transient of the intermediate lives on time $\sim\varepsilon$ ; the slow outer dynamics give Michaelis–Menten.
Regular limit (no layer)	$y(x;\varepsilon)=\sum_n\varepsilon^n y_n(x)$ uniformly	If $\varepsilon$ does not multiply the highest derivative, one regular expansion suffices — no stretching, no matching.

Common trapForcing the outer solution to satisfy every boundary condition

The commonest and most fatal mistake is to demand that the outer (reduced) solution obey both boundary conditions. It cannot — dropping $\varepsilon y''$ lowered the order, so the outer solution holds only one constant, and forcing two conditions into it yields nonsense. The discipline: decide where the layer sits (the sign of $a$ says which boundary the outer must abandon), stretch so the highest derivative survives, solve the inner problem with the abandoned condition, then match. A subtler trap: matching is not equating the two solutions at a point. It equates their overlapping limits — the outer limit of the inner and the inner limit of the outer — in the intermediate region where both are valid.

Worked examples

1 · The convection–diffusion layer, checked against the exact solution

Solve $\varepsilon y''+y'+y=0$ on $[0,1]$ with $y(0)=0$ , $y(1)=1$ at $\varepsilon=0.05$ by matched asymptotics, and compare the leading-order composite with the exact solution point by point.

Step 1. From the derivation the composite is $y_{\mathrm{comp}}(x)=e^{\,1-x}-e^{\,1-20x}$ . The layer thickness is $\delta=\varepsilon=0.05$ , so the plunge from the outer value $e\approx2.718$ down to $0$ happens over roughly $x\in[0,0.15]$ — about three layer widths.

Step 2. The exact solution is $y=(e^{r_+x}-e^{r_-x})/(e^{r_+}-e^{r_-})$ with $r_\pm=(-1\pm\sqrt{1-4\varepsilon})/(2\varepsilon)$ . At $\varepsilon=0.05$ , $r_+=-1.05573$ , $r_-=-18.94427$ — one root near $-1$ (outer decay), one near $-1/\varepsilon=-20$ (the layer). Evaluate both:

$x$	region	$y_{\mathrm{comp}}$	$y_{\mathrm{exact}}$	error
0.00	wall	0.0000	0.0000	0.0000
0.01	in layer	0.4657	0.4658	0.0001
0.025	in layer	1.0024	1.0094	0.0069
0.05	layer edge	1.5857	1.6117	0.0260
0.30	outer bulk	2.0070	2.0841	0.0771
0.50	outer bulk	1.6486	1.6951	0.0465
1.00	far wall	1.0000	1.0000	0.0000

Step 3. Read the numbers. Inside the layer the leading-order composite is essentially exact — the error at $x=0.01$ is $1\times10^{-4}$ . In the outer bulk the error rises to about $0.08$ , which is $O(\varepsilon)$ : precisely the first correction $\varepsilon Y_1(x)$ we chose not to compute. The composite captures the hard part (the layer) exactly and the easy part (the bulk) to within $\varepsilon$ , and one more order would shrink the bulk error to $O(\varepsilon^2)\approx0.003$ .

2 · Locating the layer when the drift reverses

Where does the boundary layer form for $\varepsilon y''-y'+y=0$ on $[0,1]$ , and what is the stretched variable?

Step 1. Here $a=-1<0$ , and the reduced equation $-Y_0'+Y_0=0$ gives a growing $Y_0=C\,e^{x}$ . With $a<0$ the drift points toward decreasing $x$ , sweeping disturbances to the left, so they pile up at the right boundary. The layer forms at $x=1$ , not $x=0$ .

Step 2. The outer solution therefore carries the condition at $x=0$ , and the inner variable stretches about the right wall: $X=(1-x)/\varepsilon$ . Had you placed the layer at $x=0$ , the inner exponential $e^{+X}$ would grow without bound and could not be matched — the equation flagging the wrong end. Layer at $x=1$ , stretched variable $X=(1-x)/\varepsilon$ , outer condition at $x=0$ .

3 · The turning-point scale in a physical estimate

For a quantum particle of energy $E$ in a linear potential $V(x)=Fx$ , the Schrödinger equation near the turning point is $-\tfrac{\hbar^2}{2m}\psi''+(Fx-E)\psi=0$ . Estimate the width of the region where WKB fails.

Step 1. Write it as $\psi''=\tfrac{2mF}{\hbar^2}(x-x_0)\psi$ with $x_0=E/F$ — the Airy form. By the turning-point balance the inner scale is $\ell=\big(\hbar^2/2mF\big)^{1/3}$ , the analogue of $\delta\sim\varepsilon^{2/3}$ in physical units.

Step 2. WKB holds where the action changes by many radians across a wavelength, i.e. outside a band of a few $\ell$ around $x_0$ . The failure region has width $\sim\ell=(\hbar^2/2mF)^{1/3}$ , and inside it the oscillatory and exponential branches are bridged by $\mathrm{Ai}((x-x_0)/\ell)$ .

Physics of the portrait

The method rests on a single physical insight — that a small viscosity, negligible in the bulk, becomes decisive in a thin skin against a wall — an insight due to Ludwig Prandtl that turned fluid mechanics into the engine of modern aerodynamics.

Origins · 1 minLudwig Prandtl and how this was discovered

Ludwig Prandtl

1875–1953 · founder of boundary-layer theory and modern aerodynamics

For a century after Euler, theoretical fluid dynamics had a scandal at its heart: ideal-fluid theory predicted that a body moving through a fluid feels no drag (d'Alembert's paradox), in flat contradiction with every ship and falling leaf. The viscous term in the Navier–Stokes equations, which could resolve it, was multiplied by a coefficient so small that everyone dropped it — reducing the equations' order and forcing the flow to slip along solid walls, which real fluids never do. In an eight-minute talk at the 1904 International Congress of Mathematicians in Heidelberg, Prandtl cut the knot. However small the viscosity, he argued, there must be a thin layer against the wall — a Grenzschicht, boundary layer — in which viscous forces are not negligible, because it is the only place the fluid can shed the velocity demanded by the no-slip condition. Outside it the ideal-fluid solution holds; inside it the reduced Prandtl boundary-layer equations govern a rapid adjustment; the two are matched across the layer's edge, whose thickness scales as $Re^{-1/2}$ with $Re=1/\varepsilon$ the Reynolds number — the very $\varepsilon\to0$ singular limit at the centre of this article. Prandtl thereby resolved d'Alembert's paradox, explained flow separation and stall, and founded aerodynamics at Göttingen, training a generation (Blasius, von Kármán, Tollmien) who turned the sketch into a theory. The apparatus that made his matching rigorous was assembled later by Kaplun, Lagerstrom, and above all Milton Van Dyke, whose 1964 Perturbation Methods in Fluid Mechanics codified the matching principle used today.

History

1904

— Prandtl introduces the boundary-layer concept at the Heidelberg ICM, resolving d'Alembert's paradox

1908

— Blasius computes the flat-plate boundary-layer profile from Prandtl's equations

1935

— Prandtl and collaborators develop the systematic large-Reynolds-number expansion

1957

— Kaplun and Lagerstrom formalise inner and outer limits and the overlap domain

1964

— Van Dyke publishes the matching rule in Perturbation Methods in Fluid Mechanics

1975

— Bender and Orszag give the method its standard graduate-textbook treatment

Going further

Texts. Bender & Orszag, Advanced Mathematical Methods for Scientists and Engineers, ch. 9–11 (the standard treatment, model BVP worked in full) · Hinch, Perturbation Methods (compact and rigorous) · Van Dyke, Perturbation Methods in Fluid Mechanics (source of the matching rule) · Kevorkian & Cole, Multiple Scale and Singular Perturbation Methods · Stone & Goldbart, Mathematics for Physics. The exact solution of the model problem is $y=(e^{r_+x}-e^{r_-x})/(e^{r_+}-e^{r_-})$ ; the small- $\varepsilon$ expansion of $r_\pm$ yields the composite $e^{\,1-x}-e^{\,1-x/\varepsilon}$ , confirmed against the exact solution at $\varepsilon=0.05$ in the comparison table above. The Airy asymptotics give the $\pi/4$ connection phase, and the Prandtl boundary-layer thickness scales as $\delta\sim Re^{-1/2}$ .

Prerequisites. Asymptotic Series and Poincaré Expansions · Order Symbols and Gauge Functions · WKB Approximation. Leads to. Connection Formulae and Turning Points · Bohr–Sommerfeld Quantization · Optimal Truncation and Superasymptotics.

Ask the tutor: Bring me any equation with a small parameter on its highest derivative and I will walk you through finding the layer, stretching the coordinate, matching, and writing down the uniformly valid composite.

How do I know which boundary the layer forms at?

Look at the sign of the first-derivative coefficient $a(x)$ . In $\varepsilon y''+a y'+by=0$ the convective term $a y'$ carries information in a definite direction, and the layer forms at the boundary it flows into: for constant $a>0$ the layer sits at $x=0$ , for $a<0$ at $x=1$ . If you guess wrong, the inner solution contains a growing exponential that cannot be matched to a finite outer limit — an unmatchable term is the equation telling you the layer is at the other end. When $a$ changes sign inside the domain you can get an interior layer needing its own local analysis.

Why are the inner gauge functions not always the powers ε^n?

Because the rescaled inner equation, not your expectations, decides them. Solving order by order, the equation can demand corrections at fractional powers or, notoriously, at logarithmic orders such as $\varepsilon\ln\varepsilon$ — these appear the moment you push low-Reynolds-number flow past leading order (the Stokes–Oseen problem). Forcing the sequence to be $1,\varepsilon,\varepsilon^2,\dots$ leaves terms that cannot be matched. The safe practice is to keep $\nu_n(\varepsilon)$ undetermined and let the matching at each order tell you what the next gauge must be.

Is the composite solution unique, or are there different valid forms?

The leading-order composite is unique up to terms smaller than the order worked. The additive form $y_{\mathrm{out}}+y_{\mathrm{in}}-y_{\mathrm{match}}$ is usual, but a multiplicative form $y_{\mathrm{out}}\cdot y_{\mathrm{in}}/y_{\mathrm{match}}$ can be more natural when the solution is a product of a slow envelope and a rapid factor. Both reduce to the outer in the bulk and the inner in the layer, differing only at higher order. Which you choose is convenience; the physics is identical.

How does this connect to the WKB method and to multiple scales?

They are three views of singular perturbation theory. WKB is matched asymptotics applied to $\varepsilon^2 y''=Q(x)y$ , the outer regions being the two WKB exponentials and the inner region the Airy layer at each turning point (Derivation 3). Multiple scales, by contrast, suits problems where rapid variation is distributed across the whole domain (a weakly damped oscillator) rather than confined to a layer — there you introduce fast and slow variables $t$ and $\varepsilon t$ and never match. Rule of thumb: a localized layer calls for matched asymptotics; a slowly modulated fast oscillation calls for multiple scales.

What is the overlap region, and does it always exist?

The overlap (intermediate) region is a range of $x$ simultaneously far from the wall on the layer's scale ( $X=x/\varepsilon\to\infty$ ) yet close to it on the domain's scale ( $x\to0$ ). Formally it lives on a scale $x\sim\varepsilon^\beta$ with $0<\beta<1$ , where both expansions are valid at once, and matching demands they agree there. For the well-behaved problems here the overlap exists and matching is straightforward. It can fail when the two expansions share no common region of validity (some triple-deck and nonlinear problems) — the signal that additional intermediate layers are required.

Can there be more than one boundary layer, or layers inside the domain?

Yes to both, and the machinery scales up. Layers at both ends occur when the reduced equation satisfies neither boundary condition — you build one outer and two inner solutions, matching each separately. Interior (or "shock") layers form where a coefficient changes sign inside the domain, so the solution jumps in the middle rather than at a wall, and you stretch about the interior point. Fluid-mechanics problems can even nest layers within layers — the triple-deck structure of trailing-edge and separation flows has three tiers, each matched to its neighbours. The principle never changes: find every region of rapid variation, expand in each, match across every interface.