Asymptotic Expansion · aiwiki.page

An asymptotic expansion represents a function by successively smaller contributions as its argument or a parameter approaches a specified limit. Its meaning lies in the accuracy of finite truncations, not necessarily in convergence of the infinite series. An expansion may therefore diverge at every fixed nonzero parameter value while still providing useful approximations in the relevant limiting regime. (dlmf.nist.gov)

Definition and notation

An asymptotic scale is a sequence of functions $\phi_0,\phi_1,\ldots$ satisfying

\phi_{n+1}(x)=o(\phi_n(x)) \qquad (x\to x_0).

Here $g=o(h)$ means $g/h\to0$ . The notation

f(x)\sim\sum_{n=0}^{\infty}a_n\phi_n(x)

means that, for every fixed $N\ge0$ ,

f(x)-\sum_{n=0}^{N}a_n\phi_n(x) =o(\phi_N(x)).

The approach domain is part of the definition. (dlmf.nist.gov)

For an expansion in inverse powers,

f(x)\sim a_0+\frac{a_1}{x}+\frac{a_2}{x^2}+\cdots \qquad(x\to+\infty),

an equivalent formulation is

f(x)=\sum_{n=0}^{N-1}a_nx^{-n}+O(x^{-N})

for every fixed positive integer $N$ . The big- $O$ notation specifies a bounded ratio; its bound may depend on $N$ . Small-parameter expansions replace $x^{-1}$ by $\varepsilon\to0$ . (dlmf.nist.gov)

Convergence and divergence

The distinction from an ordinary infinite series concerns which quantity tends to a limit. Convergence fixes the argument and increases the number of terms indefinitely. Asymptotic approximation fixes a finite number of terms and moves the argument toward the specified limiting regime. A convergent power series is asymptotic to its sum near its expansion point, but asymptotic expansions need not converge. (dlmf.nist.gov)

A concrete example, obtained by expanding an integrand with an exact remainder, is

F(\varepsilon)=\int_0^\infty \frac{e^{-t}}{1+\varepsilon t}\,dt, \qquad \varepsilon>0.

The finite geometric identity gives

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

Consequently,

F(\varepsilon) =\sum_{n=0}^{N-1}(-1)^nn!\varepsilon^n +R_N(\varepsilon),

where

R_N(\varepsilon) =(-\varepsilon)^N \int_0^\infty\frac{e^{-t}t^N}{1+\varepsilon t}\,dt.

Since the denominator is at least one,

|R_N(\varepsilon)|\le N!\varepsilon^N.

This directly proves

F(\varepsilon)\sim 1-\varepsilon+2!\varepsilon^2-3!\varepsilon^3+\cdots.

Nevertheless, for any fixed $\varepsilon>0$ , successive term magnitudes have ratio $(n+1)\varepsilon$ , which eventually exceeds one. The terms fail to tend to zero, so the series diverges. This calculation illustrates the factorially divergent expansions arising from Laplace-type integrals. (dlmf.nist.gov)

Truncation and error control

For many divergent expansions, terms initially decrease and then increase in magnitude. Optimal truncation seeks the finite truncation with the smallest error; stopping near the least term often provides a useful approximation to that choice. In the example above, the least-term index is of order $1/\varepsilon$ . (dlmf.nist.gov)

This is not a universal error theorem. The asymptotic definition controls each fixed truncation as the parameter approaches its limit; it does not by itself control truncations whose order grows with the parameter. Nor does a small last retained term guarantee a small remainder. Rigorous bounds or independently justified remainder estimates are needed to certify numerical accuracy. (dlmf.nist.gov)

A classical example: Stirling’s expansion

The gamma function has the expansion

\Gamma(x)\sim \sqrt{2\pi}\,x^{x-\frac12}e^{-x} \left( 1+\frac{1}{12x} +\frac{1}{288x^2} -\frac{139}{51840x^3} +\cdots \right), \qquad x\to+\infty.

Its leading term is Stirling’s approximation. Higher terms supply successive corrections; the correction series is interpreted asymptotically rather than as a convergent representation. (dlmf.nist.gov)

Because $\Gamma(n+1)=n!$ , this expansion also approximates the factorial of a large integer. For complex arguments, the corresponding expansion requires restrictions on the argument’s angular sector, illustrating why a complete asymptotic statement includes its domain of validity. (dlmf.nist.gov)

Methods of construction

Several methods generate expansions while identifying the dominant part of a problem:

Watson’s lemma converts local behavior near an endpoint into an expansion of a Laplace-type integral. Under suitable integrability and growth assumptions,
$q(t)\sim\sum_{n=0}^{\infty}c_nt^{n+\alpha-1}, \qquad \alpha>0,$
leads to
$\int_0^\infty e^{-xt}q(t)\,dt \sim\sum_{n=0}^{\infty} c_n\Gamma(n+\alpha)x^{-n-\alpha}.$
Laplace’s method analyzes integrals dominated by neighborhoods where a real exponential attains its largest value.
Stationary phase analyzes oscillatory integrals, whose main contributions often arise near stationary points of the phase. (dlmf.nist.gov)

For complex contour integrals, the method of steepest descent deforms an integration contour through relevant saddle points, subject to analyticity and contour constraints. Expanding locally near these points produces asymptotic contributions; coalescing saddles may require different, uniform approximations. (dlmf.nist.gov)

In differential equations, substitution of a proposed expansion and comparison of successive orders yields equations for its coefficients. Such formal constructions require separate justification through remainder estimates or existence results. The WKB approximation is an important example of an asymptotic method for equations with a large parameter. (dlmf.nist.gov)

Uniqueness, operations, and uniformity

For a fixed scale and approach domain, coefficients are unique when the expansion exists:

a_n=\lim_{x\to x_0} \frac{f(x)-\sum_{k=0}^{n-1}a_k\phi_k(x)} {\phi_n(x)}.

The function is not generally unique: exponentially small additions can leave every coefficient unchanged. (dlmf.nist.gov)

Thus, on the positive real axis, adding $e^{-1/\varepsilon}$ to a function does not alter its power expansion as $\varepsilon\to0^+$ , because this exponential is smaller than every positive power of $\varepsilon$ . Such contributions are termed beyond all algebraic orders. (dlmf.nist.gov)

Addition and multiplication usually follow formal series rules; division requires a suitable nonvanishing leading term. Termwise differentiation requires additional hypotheses, since a small remainder need not have a small derivative. (dlmf.nist.gov)

When another parameter varies, a uniform expansion controls its remainder throughout a specified parameter set. This matters near transition regions where an otherwise valid approximation can fail. Differential-equation turning points and coalescing integral saddle points are standard examples requiring specially constructed uniform representations. (dlmf.nist.gov)

Exponentially small effects and the Stokes phenomenon

In complex analysis, the same function may require different asymptotic representations in different angular sectors. The Stokes phenomenon concerns changes in the representation of exponentially small contributions across particular rays or curves; it does not imply a discontinuity of the underlying analytic function. (dlmf.nist.gov)

Exponentially improved expansions analyze the remainder after near-optimal truncation and retain contributions invisible to an ordinary inverse-power expansion. Further re-expansion leads to hyperasymptotic methods, extending approximation beyond the accuracy available from a single truncated series. (dlmf.nist.gov)

References

DLMF: §2.1 Definitions and Elementary Propertiesdlmf.nist.gov
DLMF: §2.3 Integrals of a Real Variabledlmf.nist.gov
DLMF: §5.11 Asymptotic Expansionsdlmf.nist.gov
DLMF: §2.4 Contour Integralsdlmf.nist.gov
DLMF: §2.7 Differential Equationsdlmf.nist.gov
DLMF: §2.11 Remainder Terms; Stokes Phenomenondlmf.nist.gov