Asymptotic Expansions and Poincare Series

Asymptotic expansions are series approximations

f (x) \sim \sum a_{n} ϕ_{n} (x)

that capture the behavior of a function near a limit point (often

x \to \infty

x \to 0

) without necessarily converging. A Poincaré series is the prototypical case where

ϕ_{n} (x) = x^{- n}

, and the expansion is asymptotic in the sense of Poincaré: for each fixed truncation, the error is smaller than the last term retained. Applies throughout applied analysis, perturbation theory, special functions, and physics. Introduced rigorously by Henri Poincaré (1886) for celestial mechanics and divergent series.

An asymptotic expansion describes how a function behaves in a limiting regime through a (typically divergent) series whose partial sums approximate the function with controlled, ordered error. Unlike a convergent Taylor series — where adding terms always improves accuracy in the limit $N \to \infty$ for fixed $x$ — an asymptotic series improves accuracy as $x \to x_{0}$ for fixed $N$ , then generally worsens past an optimal truncation. This subject matters because the most important functions of analysis and physics (the gamma function, Bessel functions, exponential integral, solutions of ODEs with irregular singular points) admit asymptotic but divergent representations that nonetheless deliver extraordinary numerical accuracy. The surprising fact: a series that diverges for every $x$ can still represent a function to within an error of $1 0^{- 15}$ when truncated optimally — Stirling's series for $ln Γ (z)$ being the canonical example.

The Core Insight

The defining demand of an asymptotic expansion is not that the infinite series converge, but that each partial sum be a uniformly good approximation as the variable approaches its limit. Formally, $f (x) \sim \sum_{n = 0}^{\infty} a_{n} x^{- n}$ means that for every fixed $N$ ,

x^{N} [f (x) - n = 0 \sum N a_{n} x^{- n}] \to 0 as x \to \infty.

The error after

N

terms is of the order of the first omitted term. This is a fundamentally different mode of approximation than convergence, and it is exactly the mode in which divergent series become useful.

Formal Statement

Asymptotic Sequence and Asymptotic Expansion

A sequence of functions ${ϕ_{n} (x)}_{n \geq 0}$ is an asymptotic sequence (or gauge functions) as $x \to x_{0}$ if for each $n$ ,

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

⟳

ϕ_{n + 1} / ϕ_{n} \to 0

x \to x_{0}

by definition of little-o; each gauge function is asymptotically smaller than its predecessor.

i.e. $lim_{x \to x_{0}} ϕ_{n + 1} (x) / ϕ_{n} (x) = 0$ . Given such a sequence, a function $f$ has the asymptotic expansion

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

⟳this is shorthand for the family of remainder estimates in equation [3], one for each

N

; the "

\sim

" symbol encodes them all.

if for every nonnegative integer $N$ the remainder condition holds:

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

⟳subtract the

N

-term partial sum; the residual must vanish faster than the last gauge function retained — equivalently it is

O (ϕ_{N + 1})

when the next term exists.

Here the symbols are: $x$ the (real or complex) variable; $x_{0}$ the limit point (commonly $\infty$ ); $ϕ_{n}$ the gauge functions; $a_{n}$ the (uniquely determined) asymptotic coefficients.

Poincaré-Type (Power) Asymptotic Series

The Poincaré series is the special case $x_{0} = \infty$ with gauge functions $ϕ_{n} (x) = x^{- n}$ (possibly preceded by a factor such as $e^{x}$ or $x^{α}$ ). One writes

f (x) \sim n = 0 \sum \infty \frac{a _{n}}{x ^{n}} (x \to \infty),

⟳substitute

ϕ_{n} = x^{- n}

into [2]; the gauge property [1] holds since

x^{- (n + 1)} / x^{- n} = x^{- 1} \to 0

meaning that for every $N$ ,

x \to \infty lim x^{N} [f (x) - n = 0 \sum N \frac{a _{n}}{x ^{n}}] = 0.

⟳multiply [3] by

x^{N} = 1/ ϕ_{N}

; the remainder condition becomes the limit-zero statement above.

An expansion is said to be asymptotic in the sense of Poincaré when it satisfies [5]. Crucially, no claim is made about $\sum a_{n} x^{- n}$ converging.

Uniqueness of Asymptotic Coefficients

Given a fixed asymptotic sequence ${ϕ_{n}}$ , the coefficients $a_{n}$ of an asymptotic expansion of $f$ are uniquely determined by $f$ , and recovered by

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

⟳rearrange [3] with

N = n

, divide by

ϕ_{n}

, and take the limit; the

o (ϕ_{n})

remainder contributes nothing.

One function, many series — but each series, one function only

Uniqueness is one-directional. A given $f$ has unique coefficients relative to a chosen gauge sequence, but a single asymptotic series can represent infinitely many distinct functions: $f (x)$ and $f (x) + e^{- x}$ share the same Poincaré series as $x \to + \infty$ , because $e^{- x} = o (x^{- n})$ for every $n$ . Such subdominant (or recessive, beyond-all-orders) terms are invisible to the power series. This is the seed of exponential asymptotics and Stokes phenomena.

Why It Holds / Derivation

Existence and Construction of the Coefficients

We show that if the limits defining the $a_{n}$ exist, the remainder conditions [3] follow, and conversely.

Define $a_{0}$ . Suppose $f (x) / ϕ_{0} (x) \to a_{0}$ . Then $f (x) - a_{0} ϕ_{0} (x) = o (ϕ_{0} (x))$ , establishing [3] for $N = 0$ .

Inductive step. Assume $a_{0}, \dots, a_{n - 1}$ have been constructed so that

R_{n - 1} (x) := f (x) - k = 0 \sum n - 1 a_{k} ϕ_{k} (x) = o (ϕ_{n - 1} (x)) .

⟳this is exactly the remainder condition [3] at level

N = n - 1

, the induction hypothesis.

Extract the next coefficient. Define $a_{n} = lim_{x \to x_{0}} R_{n - 1} (x) / ϕ_{n} (x)$ , assumed to exist. Then

R_{n} (x) = R_{n - 1} (x) - a_{n} ϕ_{n} (x) = (a_{n} + o (1)) ϕ_{n} (x) - a_{n} ϕ_{n} (x) = o (ϕ_{n} (x)) .

⟳substitute

R_{n - 1} = (a_{n} + o (1)) ϕ_{n}

from the limit definition; the leading terms cancel, leaving

o (ϕ_{n})

Conclusion. By induction [3] holds for all $N$ , so $f \sim \sum a_{n} ϕ_{n}$ . Uniqueness follows because [6] determines each $a_{n}$ from $f$ and the previously fixed coefficients with no freedom.

No convergence is invoked. Nowhere did we let $N \to \infty$ ; the construction is purely term-by-term as $x \to x_{0}$ . This is why the series may diverge while every partial sum is sharp. $■$

Why divergent series are still optimal — error control

For the Watson–Borel model the remainder after $N$ terms of a Poincaré series is typically bounded by the first omitted term:

f (x) - n = 0 \sum N \frac{a _{n}}{x ^{n}} \leq \frac{∣ a _{N + 1} ∣}{x ^{N + 1}},

⟳for alternating/integral-representable series, integration by parts plus a sign-definite remainder integral yields a bound by the next term — the hallmark of well-behaved asymptotic series.

When $∣ a_{n} ∣ \sim n! / c^{n}$ (factorial divergence), the bound is minimized near $N^{*} \approx c x$ , giving an optimal truncation error $\sim e^{- c x}$ — exponentially small even though the full series diverges.

Henri Poincaré | 1854–1912

The French polymath Henri Poincaré placed divergent series on a rigorous footing in his 1886 Acta Mathematica memoir Sur les intégrales irrégulières des équations linéaires, motivated by the perturbation series of celestial mechanics that astronomers (notably the lunar theorists) had used for decades without justification. Poincaré recognized that these series, though divergent, gave the best numerical results when truncated, and he formalized this by the remainder condition [5]. His insight reconciled the working practice of mathematical astronomy with the demands of rigorous analysis, and the definition bearing his name remains the standard. He simultaneously developed much of the qualitative theory of differential equations and dynamical systems, in which such asymptotic methods are indispensable.

Building Intuition

Convergence vs. asymptoticity — two orthogonal limits

Picture a doubly-indexed error $E (N, x)$ — the discrepancy between $f$ and the $N$ -term partial sum at the point $x$ .

A convergent series controls the limit $N \to \infty$ at fixed $x$ : $E (N, x) \to 0$ .
An asymptotic series controls the limit $x \to x_{0}$ at fixed $N$ : $E (N, x) / ϕ_{N} (x) \to 0$ .

These are independent. A series can be both (Taylor series of $e^{x}$ ), neither, asymptotic-but-divergent (Stirling), or convergent-but-not-asymptotic (a slowly converging series may need many terms to be accurate near $x_{0}$ ). For divergent asymptotic series there is an optimal $N$ for each $x$ : error decreases, bottoms out, then grows. You "superasymptotically" follow the sum down to the smallest term and stop.

Error of a divergent asymptotic series versus number of terms retained, showing the optimal truncation valley

Worked Examples

The exponential integral — a textbook divergent Poincaré series

Given. The exponential integral

E_{1} (x) = \int_{x}^{\infty} \frac{e ^{- t}}{t} d t

for large real

x > 0

. We want its asymptotic expansion. Step 1 — integrate by parts repeatedly. With

u = 1/ t

d v = e^{- t} d t

E_{1} (x) = \frac{e ^{- x}}{x} - \int_{x}^{\infty} \frac{e ^{- t}}{t ^{2}} d t .

⟳integrate

e^{- t}

- e^{- t}

, differentiate

1/ t

- 1/ t^{2}

; the boundary term gives

e^{- x} / x

Step 2 — iterate $N$ times.

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

⟳each integration by parts pulls down a factor

- n / t

, generating the factorial and alternating sign; the leftover integral is

R_{N}

Step 3 — bound the remainder. Since

t \geq x

in the integral,

t^{- (N + 1)} \leq x^{- (N + 1)}

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

⟳pull the maximal

t^{- (N + 1)} = x^{- (N + 1)}

out of the integral; the remaining

\int_{x}^{\infty} e^{- t} = e^{- x}

which equals the first omitted term — confirming the Poincaré property [5]. The series $\sum (- 1)^{n} n! / x^{n + 1}$ diverges for every $x$ (terms eventually grow), yet truncating at $N^{*} \approx x$ gives error $\sim e^{- x} 2 π x /$ ... i.e. exponentially small.

Boxed answer.

E_{1} (x) \sim e^{- x} (\frac{1}{x} - \frac{1}{x ^{2}} + \frac{2}{x ^{3}} - \frac{6}{x ^{4}} + \dots), x \to \infty

x = 5

: optimal truncation (

N^{*} \approx 5

) gives

E_{1} (5) \approx 0.0011482955

versus exact

0.00114830

— error

< 1 0^{- 7}

from a divergent series.

Stirling's series for the log-gamma function

Given.

ln Γ (z)

for large

∣ z ∣

∣ ar g z ∣ < π

. Derive its asymptotic form. Step 1 — Binet / Euler–Maclaurin. Applying the Euler–Maclaurin formula to

\sum ln k

(or Binet's integral representation):

ln Γ (z) \sim (z - \frac{1}{2}) ln z - z + \frac{1}{2} ln (2 π) + n = 1 \sum \infty \frac{B _{2 n}}{2 n ( 2 n - 1 ) z ^{2 n - 1}},

⟳the Euler–Maclaurin correction terms involve Bernoulli numbers

B_{2 n}

; the leading

(z - \frac{1}{2}) ln z - z

comes from

\int ln t d t

and the trapezoidal endpoint.

Step 2 — display first coefficients. Using

B_{2} = \frac{1}{6}

B_{4} = - \frac{1}{30}

ln Γ (z) \sim (z - \frac{1}{2}) ln z - z + \frac{1}{2} ln (2 π) + \frac{1}{12 z} - \frac{1}{360 z ^{3}} + \frac{1}{1260 z ^{5}} - \dots

⟳

\frac{B _{2}}{2 \cdot 1} = \frac{1}{12}

\frac{B _{4}}{4 \cdot 3} = - \frac{1}{360}

\frac{B _{6}}{6 \cdot 5} = \frac{1}{1260}

Step 3 — divergence. Since

∣ B_{2 n} ∣ \sim \frac{2 ( 2 n )!}{( 2 π ) ^{2 n}}

, the term ratio grows like

(n / π z)^{2}

— the series diverges for all $z$ , yet is the gold standard for computing factorials. Boxed answer (Stirling for the factorial). Exponentiating,

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

n = 10

: leading factor gives

3598695.6

vs exact

10! = 3628800

(0.83% error); the

1/12 n

correction gives

3628685

— error reduced to

0.003%

, units: dimensionless count.

Common Misconceptions

Asymptotic ≠ convergent; "more terms" can be worse

The single most damaging error is treating an asymptotic series like a Taylor series. Adding terms beyond the optimal $N^{*}$ makes the approximation worse, not better. There is no "sum to infinity" to compute. Likewise, you cannot freely substitute $x \to 0$ or differentiate/integrate term-by-term without justification: differentiation of an asymptotic expansion is not automatically valid (though integration usually is, and differentiation is valid for analytic functions in a sector). Always retain explicit remainder bounds.

A function indistinguishable from zero by its Poincaré series

Let $f (x) = e^{- x}$ as $x \to + \infty$ . For every $n$ ,

x \to \infty lim x^{n} e^{- x} = 0,

⟳exponential decay beats any polynomial growth; L'Hôpital

n

times reduces

x^{n} / e^{x}

n! / e^{x} \to 0

so all asymptotic coefficients vanish: $e^{- x} \sim 0 + 0 \cdot x^{- 1} + 0 \cdot x^{- 2} + \dots$ . The zero function has the same Poincaré series. Hence the map {function} $\to$ {Poincaré series} is not injective: the entire ideal of subdominant terms ${e^{- x}, e^{- x} / x, \dots}$ is annihilated. Recovering these requires resurgence theory and the resummation of the divergent tail (Borel summation, optimal truncation, hyperasymptotics).

Connections

How this binds the analytic landscape together

Asymptotic expansions sit at a crossroads. They are produced by Watson's Lemma and the Method of Steepest Descent from integral representations, and by WKB Approximation and the theory of Irregular Singular Points from differential equations. The divergence of the typical Poincaré series is tamed by Borel Summation and the broader program of Resurgence Theory (Écalle's alien calculus), which restores the subdominant terms hidden by Stokes Phenomenon. The coefficients themselves are organized by Euler–Maclaurin Formula (for sums) and connect to Bernoulli Numbers. In algebra and number theory the word "Poincaré series" also denotes a generating function $\sum dim (V_{n}) t^{n}$ — a distinct meaning not to be confused with the asymptotic notion here. The framework underlies all of Perturbation Theory and singular Boundary Layer Theory.

Real-World Applications

Where Poincaré-type asymptotics decide the answer

Celestial mechanics & ephemerides. The original setting: perturbation series for planetary and lunar orbits are asymptotic, not convergent; truncation at the optimal order yields the precise tables used for navigation and space-mission planning (JPL ephemerides).
Quantum field theory. Perturbation series in the coupling $g$ (e.g. QED's expansion in $α \approx 1/137$ ) are divergent asymptotic series; the anomalous magnetic moment of the electron, computed to ~ $10$ digits, agrees with experiment by optimal truncation and Borel-type resummation.
Statistical mechanics. High- and low-temperature expansions and virial series are asymptotic; Stirling's formula governs partition-function combinatorics and entropy of large systems.
Numerical special functions. Libraries (Cephes, GNU Scientific Library, Boost) compute $Γ$ , Bessel $J_{ν}, Y_{ν}$ , Airy $Ai$ , and error functions for large argument via asymptotic series with built-in optimal-truncation logic.
Wave propagation & optics. WKB/eikonal asymptotics describe short-wavelength fields, caustics (Airy/Pearcey functions), and semiclassical quantization (Bohr–Sommerfeld).
Fluid dynamics — boundary layers. Prandtl's boundary-layer theory is matched asymptotic expansions in the inverse Reynolds number; drag and lift estimates rest on these.
Probability & statistics. Laplace's method and the saddle-point approximation give asymptotic expansions of tail probabilities, the Edgeworth series, and large-deviation rates.

Historical Development

From divergent scandal to rigorous tool

1730s–1750s — Euler & Stirling. Stirling (1730) and de Moivre derive the factorial asymptotics; Euler manipulates divergent series boldly and computes their "values," anticipating summation methods but without rigor.
1768 — Euler's $E_{1}$ series. Euler studies the divergent series for what is now the exponential integral, noting its remarkable numerical accuracy despite divergence.
1820s — Cauchy & Abel. The rise of rigor; Abel famously calls divergent series "an invention of the devil," and they fall into disrepute even as astronomers keep using them.
1886 — Poincaré. Sur les intégrales irrégulières des équations linéaires (Acta Math.) gives the precise definition [5], legitimizing divergent asymptotic series and tying them to ODEs with irregular singular points.
1886 — Stieltjes. Independently formalizes asymptotic series and pioneers the link to continued fractions and the moment problem (Stieltjes integral).
1902–1918 — Watson, Barnes. G. N. Watson proves Watson's Lemma (asymptotics of Laplace-type integrals) and develops the theory systematically; Barnes integrals give exact and asymptotic representations.
1899–1918 — Borel. Émile Borel develops Borel summation, the first systematic method to assign sums to divergent asymptotic series.
1857 / 1864 — Stokes (retrospectively). Stokes' discovery of the discontinuous appearance/disappearance of subdominant exponentials (the Stokes phenomenon) is later recognized as central.
1980s–1990s — Écalle, Berry, Dingle. Resurgence theory (Écalle), Dingle's interpretation of late terms, and Berry's smoothing of the Stokes phenomenon and hyperasymptotics push beyond all orders, recovering the exponentially small terms invisible to Poincaré.

Subtleties & Limits

Beyond all orders: resummation and uniqueness

The Poincaré definition deliberately discards exponentially small (beyond-all-orders) information, which is precisely the information that often controls physical phenomena: tunneling rates, separatrix splitting, instanton contributions, and Stokes transitions. The central open territory is exact resummation: given the divergent coefficients $a_{n}$ , can one uniquely reconstruct $f$ ? Borel summation works when the Borel transform $\sum a_{n} t^{n} / n!$ has analytic continuation with no singularities on the positive axis; otherwise lateral Borel sums differ by exponentially small amounts — the Stokes ambiguity. Écalle's resurgence organizes these ambiguities via alien derivatives and transseries ( $\sum_{k} e^{- A_{k} / x} x^{b_{k}} \sum_{n} a_{k, n} x^{- n}$ ), conjecturally rendering many physical series resurgent and resummable. Outstanding questions: which classes of nonlinear ODEs/PDEs and which QFT perturbation series are provably resurgent; sharp error bounds in hyperasymptotic (multi-level) expansions; and the rigorous status of resurgent resummation in genuinely strongly coupled regimes.