Asymptotic Expansions

Asymptotic expansions represent a function by a (often divergent) series whose partial sums approximate the function with controllable error in a limit (e.g.

x \to \infty

ε \to 0

). Key result: a divergent series

\sum a_{n} x^{- n}

can give superb numerical accuracy if truncated optimally, even though it never converges. Where it applies: singular perturbation theory, integral evaluation (Laplace, stationary phase, saddle point), special functions, WKB analysis, statistical mechanics. Who/when: Poincaré gave the rigorous definition (1886), building on Stirling, Euler, and Laplace; Stokes and others developed the analytic machinery.

An asymptotic expansion of a function $f$ with respect to an asymptotic sequence ${ϕ_{n}}$ is a formal series $\sum a_{n} ϕ_{n} (x)$ such that the remainder after $N$ terms is smaller than the last retained term as $x \to x_{0}$ . The defining feature is the order of limits: for fixed $N$ the error vanishes as $x \to x_{0}$ , but for fixed $x$ the error generally grows without bound as $N \to \infty$ . This makes asymptotic series fundamentally different from convergent Taylor series, and explains the surprising fact that a series summing to $+ \infty$ term-by-term can compute a function to ten decimal places. Asymptotics is the dominant practical tool whenever exact solutions are unavailable but a small or large parameter exists. Its rigorous control of error — captured by the notion of "least term" and exponentially small corrections — underlies most of applied analysis.

The Core Insight

A series need not converge to be useful. What matters is not the behavior of the partial sums as $N \to \infty$ at fixed $x$ , but the behavior of the remainder as $x \to x_{0}$ at fixed $N$ . Asymptotic expansions trade convergence for uniform smallness of the truncation error in a limiting regime, and the best accuracy is obtained by truncating at the smallest term (optimal truncation / superasymptotics).

Formal Statement

Asymptotic Sequence

A sequence of functions ${ϕ_{n} (x)}_{n = 0}^{\infty}$ is an asymptotic sequence (or scale) as $x \to x_{0}$ if for every $n$ ,

ϕ_{n + 1} (x) = o (ϕ_{n} (x)), x \to x_{0},

that is

lim_{x \to x_{0}} ϕ_{n + 1} (x) / ϕ_{n} (x) = 0

Common scales: ${x^{- n}}$ as $x \to \infty$ , ${(x - x_{0})^{n}}$ as $x \to x_{0}$ , ${ε^{n}}$ as $ε \to 0$ , and mixed scales like ${ε^{n} lo g^{m} ε}$ .

⟳The little-o relation is the minimal condition guaranteeing each successive term is genuinely a higher-order correction; without it the "expansion" cannot be ordered.

Asymptotic Expansion (Poincaré sense)

Let ${ϕ_{n}}$ be an asymptotic sequence as $x \to x_{0}$ . The formal series $\sum_{n = 0}^{\infty} a_{n} ϕ_{n} (x)$ is an asymptotic expansion of $f (x)$ , written

f (x) \sim n = 0 \sum \infty a_{n} ϕ_{n} (x), x \to x_{0},

if for every fixed

N \geq 0

f (x) - n = 0 \sum N a_{n} ϕ_{n} (x) = o (ϕ_{N} (x)), x \to x_{0} .

Equivalently the remainder is of the order of the first omitted term:

f (x) - n = 0 \sum N - 1 a_{n} ϕ_{n} (x) = O (ϕ_{N} (x)) .

⟳Equation [3] is the substantive definition; it says the truncation error is asymptotically negligible relative to the last kept term. Equation [4] follows by shifting

N \to N - 1

and using [1].

Uniqueness of Coefficients

Given an asymptotic sequence ${ϕ_{n}}$ , the coefficients $a_{n}$ in [2] are uniquely determined by $f$ via

a_{N} = x \to x_{0} lim \frac{f ( x ) - \sum _{n = 0}^{N - 1} a _{n} ϕ _{n} ( x )}{ϕ _{N} ( x )} .

However, the function is not uniquely determined by its expansion: two distinct functions may share the same asymptotic series.

⟳Subtract the first

N

terms, divide by

ϕ_{N}

, take the limit; the higher terms vanish by [1] and the remainder vanishes by [3]. Non-uniqueness follows from

e^{- 1/ x} \sim 0

in every power of

x

Operations on Asymptotic Series

If $f \sim \sum a_{n} x^{- n}$ and $g \sim \sum b_{n} x^{- n}$ as $x \to \infty$ , then:

Linearity: $α f + β g \sim \sum (α a_{n} + β b_{n}) x^{- n}$ .
Multiplication: $f g \sim \sum c_{n} x^{- n}$ with $c_{n} = \sum_{k = 0}^{n} a_{k} b_{n - k}$ (Cauchy product).
Integration: $\int_{x}^{\infty} (f (t) - a_{0} - a_{1} t^{- 1}) d t$ is asymptotic to the termwise integral.
Differentiation is NOT generally valid without extra hypotheses (e.g. $f^{'}$ must itself possess an asymptotic expansion).

⟳Linearity and multiplication follow directly from [3] applied to each function and Cauchy's product rule; integration converges because the integrand is

O (t^{- 2})

; differentiation fails because oscillatory terms like

e^{- x} sin (e^{x})

are

o (x^{- n})

yet have unbounded derivatives.

Henri Poincaré | 1854–1912

Poincaré supplied the rigorous definition (1886) in his study of celestial mechanics, where divergent series of perturbation theory demanded justification. He recognized that astronomers had been computing with divergent series for a century and that their utility lay in the order of the remainder, not convergence. His framing — now called the Poincaré definition — separates the analytic from the formal and remains the standard. The same insight powered his qualitative theory of differential equations and his work on the three-body problem.

Why It Holds / Derivation

The Asymptotic Series of an Integral — Watson/Laplace prototype

Consider the exponential integral and derive its expansion to show how [2]–[4] arise.

Step 1. Define, for

x > 0

f (x) = \int_{0}^{\infty} \frac{e ^{- x t}}{1 + t} d t .

Step 2. Repeated integration by parts (or expanding

1/ (1 + t)

with remainder) gives the geometric-with-remainder identity

\frac{1}{1 + t} = n = 0 \sum N - 1 (- 1)^{n} t^{n} + \frac{( - 1 ) ^{N} t ^{N}}{1 + t} .

Step 3. Insert [7] into [6] and use

\int_{0}^{\infty} t^{n} e^{- x t} d t = n! x^{- n - 1}

f (x) = n = 0 \sum N - 1 \frac{( - 1 ) ^{n} n !}{x ^{n + 1}} + R_{N} (x), R_{N} (x) = (- 1)^{N} \int_{0}^{\infty} \frac{t ^{N} e ^{- x t}}{1 + t} d t .

Step 4. Bound the remainder using

1/ (1 + t) \leq 1

∣ R_{N} (x) ∣ \leq \int_{0}^{\infty} t^{N} e^{- x t} d t = \frac{N !}{x ^{N + 1}} .

Thus

R_{N} = O (x^{- N - 1})

, which is the order of the first omitted term — exactly condition [4]. Hence

f (x) \sim n = 0 \sum \infty \frac{( - 1 ) ^{n} n !}{x ^{n + 1}}, x \to \infty.

Step 5. Convergence test: the ratio of consecutive terms is

∣ a_{n + 1} / a_{n} ∣ = (n + 1) / x \to \infty

, so the series diverges for every $x$ . Yet [9] shows the error is rigorously controlled. ∎

⟳The crucial trick is the finite expansion-with-exact-remainder [7], which converts an asymptotic claim into a hard integral bound [9]; this is the template for Watson's lemma.

Watson's Lemma

If $g (t) \sim \sum_{n = 0}^{\infty} a_{n} t^{(n + λ - μ) / μ}$ as $t \to 0^{+}$ (with $λ, μ > 0$ ) and the Laplace integral converges, then

\int_{0}^{\infty} e^{- x t} g (t) d t \sim n = 0 \sum \infty a_{n} \frac{Γ ( ( n + λ ) / μ )}{x ^{(n + λ) / μ}}, x \to \infty.

The dominant contribution comes entirely from a neighborhood of

t = 0

⟳Split the integral at

t = δ

; the tail is exponentially small (

O (e^{- x δ})

), while near

0

one substitutes the local expansion of

g

and integrates termwise, recovering Gamma functions exactly as in Step 3 above.

Building Intuition

Think of an asymptotic series as a sequence of successively finer rulers calibrated for a particular regime. Each term is much smaller than the last when $x$ is large, so adding terms improves the picture — up to a point. Because the coefficients $a_{n} = (- 1)^{n} n!$ grow factorially while the powers $x^{- n}$ only shrink geometrically, eventually $∣ a_{n} x^{- n} ∣$ stops decreasing and starts growing. The smallest term occurs near $n \approx x$ , and that minimal term sets the intrinsic accuracy floor: you cannot do better than $\sim e^{- x}$ (the size of the least term) using the series alone. This exponentially small residual is the realm of exponential asymptotics / resurgence, where the "missing" information is hidden across the Stokes lines of the complex plane.

Optimal truncation: error decreases then explodes; least term sets the accuracy floor

Worked Examples

Exponential Integral at x = 10

Given: Approximate

f (10) = \int_{0}^{\infty} \frac{e ^{- 10 t}}{1 + t} d t

using the divergent series [10]. The true value is

f (10) = 0.0915633 \dots

Steps. Terms

a_{n} = (- 1)^{n} n! /1 0^{n + 1}

$n = 0$ : $+ 0.100000$
$n = 1$ : $- 0.010000$ → partial $0.090000$
$n = 2$ : $+ 0.002000$ → $0.092000$
$n = 3$ : $- 0.000600$ → $0.091400$
$n = 4$ : $+ 0.000240$ → $0.091640$
$n = 5$ : $- 0.000120$ → $0.091520$
$n = 6$ : $+ 0.0000720$ → $0.091592$
$n = 7$ : $- 0.0000504$ → $0.091542$
...

Least term near $n \approx 10$ : $∣ a_{10} ∣ = 10! /1 0^{11} = 3.6 \times 1 0^{- 5}$ . Truncating just before it (around $n = 9$ , $∣ a_{9} ∣ = 3.6 \times 1 0^{- 5}$ ) gives the best accuracy.

Boxed answer.

f (10) \approx 0.09156, error \sim 3.6 \times 1 0^{- 5} (set by least term)

Adding more terms beyond

n \approx 10

worsens the result — the divergence asserts itself.

Stirling's Formula via Laplace's Method

Given: Find the leading asymptotics of

Γ (x + 1) = x! = \int_{0}^{\infty} t^{x} e^{- t} d t

x \to \infty

. Steps.

Write $t^{x} e^{- t} = e^{x l n t - t}$ and substitute $t = x s$ : integrand $= e^{x (l n (x s) - s)} = x^{x + 1} e^{x (l n s - s)} \cdot x^{- ...}$ . Cleaner: $Γ (x + 1) = x^{x + 1} \int_{0}^{\infty} e^{x (l n s - s)} d s$ .
The exponent $ϕ (s) = ln s - s$ has maximum at $s = 1$ , $ϕ (1) = - 1$ , $ϕ^{''} (1) = - 1$ .
Gaussian (Laplace) approximation around $s = 1$ :

\int_{0}^{\infty} e^{x ϕ (s)} d s \approx e^{- x} \frac{2 π}{x ∣ ϕ ^{''} ( 1 ) ∣} = e^{- x} \frac{2 π}{x} .

Multiply by $x^{x + 1}$ :

Γ (x + 1) \sim x^{x + 1} e^{- x} \frac{2 π}{x} = 2 π x x^{x} e^{- x} .

Numerical check, $x = 10$ : Stirling gives

20 π \cdot 1 0^{10} e^{- 10} = 3598695.6

; true

10! = 3628800

. Relative error

0.83% \approx 1/ (12 \cdot 10)

, consistent with the next correction factor

(1 + \frac{1}{12 x} + \dots)

. Boxed answer.

n! \sim 2 π n (\frac{n}{e})^{n} (1 + \frac{1}{12 n} + \frac{1}{288 n ^{2}} - \dots)

Derivation hint for [12]: expand $ϕ (s) = ϕ (1) - \frac{1}{2} (s - 1)^{2} + \dots$ , extend the Gaussian limits to $\pm \infty$ (error exponentially small), and use $\int e^{- x (s - 1)^{2} /2} d s = 2 π / x$ .

Common Misconceptions

"Asymptotic series converge." They usually do not. Convergence concerns

N \to \infty

at fixed

x

; asymptoticity concerns

x \to x_{0}

at fixed

N

. These are independent. A series can be convergent but not asymptotic, asymptotic but divergent (the common case), both, or neither. Never add "more terms" expecting indefinite improvement. "You can always differentiate an asymptotic expansion term by term." False without justification. The relation

f \sim \sum a_{n} ϕ_{n}

does not imply

f^{'} \sim \sum a_{n} ϕ_{n}^{'}

. Integration is safe (smoothing), differentiation is not (roughening). One must verify that

f^{'}

itself admits an expansion.

Same expansion, different functions

The function $g (x) = e^{- x}$ satisfies $e^{- x} = o (x^{- n})$ for every $n$ as $x \to \infty$ , so $e^{- x} \sim 0$ in the power scale ${x^{- n}}$ . Therefore $f (x)$ and $f (x) + e^{- x}$ have the identical asymptotic power series. The expansion cannot detect the exponentially small term — it lives "beyond all orders." This non-uniqueness (Theorem [5]) is precisely the gap that exponential asymptotics and resurgence theory aim to fill via Stokes phenomena.

Connections

Asymptotic expansions are the analytic engine of Perturbation Theory (regular and singular), where solutions are expanded in a small parameter $ε$ . In Method of Matched Asymptotic Expansions one stitches together inner and outer expansions valid in different regions of a boundary layer. The WKB Approximation is an asymptotic expansion of solutions to differential equations in $ℏ$ (or wavelength). The evaluation of Laplace's Method, Method of Steepest Descent (saddle point), and Stationary Phase Approximation all produce asymptotic series for integrals. The deep structure connecting divergent series across Stokes Phenomenon is formalized by Resurgence Theory and Borel summation, linking asymptotics to Analytic Continuation. The factorial growth of coefficients ties directly to Borel Summation.

Real-World Applications

Celestial mechanics & orbital perturbations: Poincaré's original motivation — series for planetary motion in the three-body problem are asymptotic, not convergent, yet accurate over astronomical timescales.
Quantum mechanics (WKB): Energy levels and tunneling rates from $ℏ \to 0$ expansions; the divergence of QED perturbation series ( $\sim n!$ growth) is a textbook asymptotic phenomenon.
Fluid dynamics / boundary layers: Prandtl's boundary-layer theory and high-Reynolds-number flows use matched asymptotic expansions to handle thin viscous regions.
Statistical mechanics: Stirling's approximation for entropy ( $ln N!$ ); virial and cluster expansions; large-deviation rates.
Special-function libraries: Bessel, Airy, error, and Gamma functions are computed via asymptotic series for large argument in numerical software (e.g. NIST DLMF, GSL).
Signal processing & optics: Stationary phase for diffraction integrals and Fourier-integral asymptotics.
Number theory: The asymptotic expansion of $π (x)$ and the prime-counting logarithmic integral; saddle-point analysis of partition asymptotics (Hardy–Ramanujan).

Historical Development

1730 (Stirling/De Moivre): Stirling's series for $n!$ — historically the first widely used divergent asymptotic series.
1754 (Euler): Manipulates divergent series ("the series $1 - 1! + 2! - \dots$ ") obtaining numerically sensible values; intuits summation methods.
1812 (Laplace): Develops the method of approximating integrals by the contribution of their dominant peak — Laplace's method.
1847–1857 (Stokes): Discovers the Stokes phenomenon: asymptotic representations change discontinuously across rays in the complex plane; studies the Airy function.
1863 (Riemann): Saddle-point / steepest-descent ideas in his work, later named after him and Debye.
1886 (Poincaré): Gives the rigorous definition of an asymptotic expansion and the notion of an asymptotic series, establishing the field.
1909 (Debye): Systematizes the method of steepest descent for Bessel-function asymptotics.
1918 (Watson): Proves Watson's lemma, the workhorse for Laplace-type integrals.
1920s–30s (Borel, Hardy): Theory of divergent series and Borel summation; Hardy's Divergent Series (1949).
1980s–present (Dingle, Écalle, Berry): Resurgence theory, hyperasymptotics, and smoothing of the Stokes phenomenon (Berry's error-function transition) — recovering exponentially small terms beyond all orders.

Subtleties & Limits

The exponentially small terms invisible to a power-series expansion ("beyond all orders") carry genuine physical and mathematical content — instantons, tunneling amplitudes, and selection rules in pattern formation. Resurgence theory (Écalle) provides a framework in which the divergence of a perturbative series encodes its own non-perturbative completion, but a fully general, constructive theory of resummation for arbitrary nonlinear PDEs remains incomplete. Open questions include: precise transseries representations for solutions of broad classes of differential equations; rigorous control of multi-instanton sectors in realistic quantum field theories; effective algorithms for hyperasymptotic summation past several Stokes lines; and the interplay between the Stokes phenomenon and singularities in the Borel plane for systems lacking known integral representations.