Asymptotic Expansions

Motivation: why this matters in quant finance

An asymptotic expansion is an approximation that becomes increasingly accurate in a limiting regime (e.g., large

x

, small

\sigma

, short

T

) even though the series may diverge if you sum all its terms. This makes asymptotic expansions fundamentally different from convergent Taylor or power series — they are useful not because they converge, but because their first few terms give excellent approximations in the relevant regime.

In quant finance, asymptotic expansions are the formulas practitioners actually use for quick estimation:

Implied volatility approximations — closed-form approximations to the Black-Scholes implied vol as a function of moneyness and time (Brenner-Subrahmanyam, Gatheral's SVI parameterisation, small-time/large-time asymptotics).
Tail probabilities of the normal distribution — the Mills ratio $\Phi(-x) \approx \phi(x)/x$ for large $x$ , used in extreme quantile estimation and VaR calculations.
Greeks in extreme regimes — delta, gamma, vega behaviour for deep ITM/OTM options and near/far expiry, derived from asymptotic expansion of the Black-Scholes formula.
Saddle-point approximations — used in credit portfolio models (CreditRisk+, Vasicek) to approximate loss distributions without Monte Carlo.

Definition

Asymptotic expansion

A function

f(x)

has the asymptotic expansion

\sum_{n=0}^{N} a_n / x^n

x \to \infty

if:

f(x) - \sum_{n=0}^{N} \frac{a_n}{x^n} = o\left(\frac{1}{x^N}\right) \quad \text{as } x \to \infty

for each fixed $N$ . In words: the error from truncating at $N$ terms is smaller than the last term kept, as $x \to \infty$ .

We write:

f(x) \sim \sum_{n=0}^{\infty} \frac{a_n}{x^n} \quad (x \to \infty)

The symbol

\sim

means "asymptotically equal to" — it does not mean the series converges. The series may diverge for every

x

, yet the truncated sum at optimal

N

can give extraordinary accuracy.

Big-O and little-o notation

$f(x) = O(g(x))$ as $x \to a$ means $|f(x)| \leq C|g(x)|$ near $a$ for some constant $C$ .
$f(x) = o(g(x))$ as $x \to a$ means $f(x)/g(x) \to 0$ as $x \to a$ .

These notations describe the order of approximation:

e^{-rT} = 1 - rT + O(T^2) \quad \text{as } T \to 0

means the error from using $1 - rT$ as an approximation is at most proportional to $T^2$ .

Divergent but useful

The classic example is the asymptotic expansion of the exponential integral:

\int_x^{\infty} \frac{e^{-t}}{t}\,dt \sim \frac{e^{-x}}{x}\left(1 - \frac{1}{x} + \frac{2!}{x^2} - \frac{3!}{x^3} + \cdots\right)

The coefficients $n!$ grow factorially, so the series diverges for every $x$ . But for $x = 10$ , the first 5 terms give 10-digit accuracy. The optimal truncation point is near $N \approx x$ , after which the terms start growing.

Key asymptotic results in quant finance

Normal tail approximation (Mills ratio)

For large $x > 0$ :

\Phi(-x) = 1 - \Phi(x) \sim \frac{\phi(x)}{x}\left(1 - \frac{1}{x^2} + \frac{3}{x^4} - \frac{15}{x^6} + \cdots\right)

where $\phi(x) = \frac{1}{\sqrt{2\pi}}e^{-x^2/2}$ .

The leading term

\Phi(-x) \approx \phi(x)/x

is the Mills ratio. For

x = 3

: exact

\Phi(-3) = 0.00135

, Mills approximation

= 0.00144

, error

\approx 7\%

. For

x = 5

: exact

= 2.87 \times 10^{-7}

, Mills

= 2.97 \times 10^{-7}

, error

\approx 3\%

Finance application: VaR at extreme confidence levels (99.9%, 99.97% for Basel) requires

\Phi^{-1}(\alpha)

at very small

\alpha

. Inverting the Mills ratio gives:

\Phi^{-1}(\alpha) \approx -\sqrt{-2\ln(2\pi) - 2\ln\alpha + 2\ln(-\ln\alpha - \frac{1}{2}\ln(2\pi))}

which is much faster (and more numerically stable) than generic CDF inversion for extreme quantiles.

Black-Scholes asymptotics

The Black-Scholes call price $C = S\Phi(d_1) - Ke^{-rT}\Phi(d_2)$ has useful limiting behaviour:

Deep ITM ( $S \gg K$ , $d_1, d_2 \to +\infty$ ):

C \approx S - Ke^{-rT} + Ke^{-rT}\frac{\phi(d_2)}{d_2} + \cdots

The call price approaches intrinsic value plus a small correction (the time value), which vanishes exponentially.

Deep OTM ( $S \ll K$ , $d_1, d_2 \to -\infty$ ):

C \approx S\frac{\phi(d_1)}{|d_1|} + \cdots

The call price is exponentially small, dominated by the tail probability of finishing ITM.

Short expiry ( $T \to 0$ ):

C \approx (S - K)^+ + \frac{1}{2}S\sigma\sqrt{T}\,\phi\left(\frac{\ln(S/K)}{\sigma\sqrt{T}}\right)\left(\frac{1}{\sigma\sqrt{T}} + \cdots\right)

For ATM options ( $S = K$ ): $C \approx S\sigma\sqrt{T/(2\pi)} + O(T)$ . This gives the well-known ATM approximation $C_{\text{ATM}} \approx 0.4 \cdot S\sigma\sqrt{T}$ .

Implied volatility approximations

Brenner-Subrahmanyam (ATM):

\sigma_{\text{impl}} \approx \frac{C_{\text{ATM}}}{0.4 \cdot S\sqrt{T}} = \frac{C_{\text{ATM}}\sqrt{2\pi}}{S\sqrt{T}}

This inverts the short-expiry ATM approximation above and gives a quick estimate of implied vol from the ATM call price.

Small vol-of-vol expansion (Heston): When the vol-of-vol parameter

\xi

is small, the Heston implied vol can be expanded:

\sigma_{\text{impl}}(K, T) \approx \sigma_0 + \xi\,\sigma_1(K, T) + \xi^2\,\sigma_2(K, T) + \cdots

where $\sigma_0 = \sqrt{\theta}$ (the long-run vol) and the correction terms capture the skew and smile. This is an asymptotic expansion in $\xi$ — the series may diverge, but the first two or three terms give accurate approximations for realistic parameters.

Stirling's approximation

n! \sim \sqrt{2\pi n}\left(\frac{n}{e}\right)^n

Used in the binomial model when computing $\binom{n}{k}$ for large $n$ (De Moivre-Laplace theorem), and in combinatorial credit models.

Saddle-point approximation

For a random variable with cumulant generating function

K(t)

, the density can be approximated by:

f(x) \approx \frac{1}{\sqrt{2\pi K''(\hat{t})}}\,\exp\left(K(\hat{t}) - \hat{t}x\right)

where $\hat{t}$ is the saddle point satisfying $K'(\hat{t}) = x$ .

Finance application: In the Vasicek credit portfolio model, the loss distribution is the sum of many Bernoulli-like variables. The exact distribution requires a high-dimensional integral; the saddle-point approximation gives an accurate closed-form expression for the tail probabilities, which is what you need for economic capital calculations. This is faster than Monte Carlo by orders of magnitude.

Optimal truncation

Since asymptotic series typically diverge, you cannot sum all terms. The optimal truncation strategy:

Compute terms until they start growing in magnitude.
Stop just before the smallest term.
The error is approximately the size of the first omitted term.

For the Mills ratio at $x = 4$ : the terms in the asymptotic series are $1, -1/16, +3/256, -15/4096, \ldots$ . They decrease initially, reaching a minimum around the 8th term, then grow. Truncating at the 8th term gives $\sim 8$ digits of accuracy.

This "divergent series, optimal truncation" pattern is characteristic of asymptotic methods and is completely different from the "convergent series, take more terms for more accuracy" pattern of Taylor series within their radius of convergence.

Examples and applications

Example 1: quick VaR estimation

For a portfolio with daily return $R \sim \mathcal{N}(0, \sigma^2)$ , the 99.9% VaR:

\text{VaR}_{0.001} = \sigma \cdot \Phi^{-1}(0.999) = \sigma \cdot (-\Phi^{-1}(0.001))

Using the asymptotic inversion: $\Phi^{-1}(0.001) \approx -3.09$ , so $\text{VaR} \approx 3.09\sigma$ . For a portfolio with daily $\sigma = 1\%$ , the 99.9% daily VaR is approximately 3.09%.

Example 2: ATM option pricing shortcut

A 3-month ATM call on a stock at $S = 100$ with $\sigma = 25\%$ :

C \approx 0.4 \cdot S\sigma\sqrt{T} = 0.4 \cdot 100 \cdot 0.25 \cdot \sqrt{0.25} = 0.4 \cdot 100 \cdot 0.25 \cdot 0.5 = 5.00

Exact Black-Scholes (with $r = 0$ ): $C = 4.97$ . The asymptotic approximation gives an answer within 0.6% — useful for mental math and quick sanity checks on a trading desk.

Example 3: tail probability for stress testing

What is $\mathbb{P}(Z < -6)$ for a standard normal? Direct computation is numerically tricky at this extreme. The Mills ratio:

\Phi(-6) \approx \frac{\phi(6)}{6} = \frac{e^{-18}}{6\sqrt{2\pi}} \approx \frac{1.52 \times 10^{-8}}{15.03} \approx 1.01 \times 10^{-9}

Exact value: $9.87 \times 10^{-10}$ . The one-term approximation is within 3% — sufficient for stress testing where the uncertainty in the tail model itself dwarfs any approximation error.

Common confusions and pitfalls

"An asymptotic expansion converges." Usually not. The series typically diverges for every

x

. It is useful because the first few terms give accurate approximations in the appropriate regime, not because the series converges as a whole.

"More terms always improve accuracy." Only up to the optimal truncation point. Beyond that, additional terms make the approximation worse. This is the opposite of convergent series behaviour and is unintuitive.

"The ATM approximation $C \approx 0.4 S\sigma\sqrt{T}$ works for all options." It is an asymptotic result valid only for ATM options (

S \approx Ke^{-rT}

) in the short-expiry regime. For OTM options, ITM options, or long-dated options, different asymptotic regimes apply. Using the ATM formula for deep OTM options gives wildly wrong answers.

"Asymptotics replace exact formulas." No. Asymptotics complement exact formulas by providing intuition, quick estimates, and computational shortcuts in limiting regimes. When an exact formula exists (Black-Scholes for vanilla Europeans), use it. Asymptotics are most valuable when exact formulas do not exist (Heston, SABR) or when you need a quick sanity check.

Where this goes next

Asymptotic expansions connect to:

Taylor Series: Taylor series are the "small perturbation" asymptotic; general asymptotic expansions handle "large variable" and other regimes.
Power Series: Asymptotic series may look like power series but diverge — the radius of convergence is effectively zero.
Fourier Series and Transforms: Saddle-point and stationary-phase methods are asymptotic techniques applied to Fourier integrals.
Normal Distribution: The Mills ratio and tail asymptotics are the most-used normal distribution approximations in practice.
Numerical Integration: Asymptotic expansions provide error estimates for numerical methods and analytical alternatives to Monte Carlo in limiting regimes.
Implicit Differentiation: Implied vol approximations are often derived by asymptotic expansion of the implicit equation $C_{\text{BS}}(\sigma) = C_{\text{mkt}}$ .

References

Stewart, J. (2008). Single Variable Calculus: Early Transcendentals (6th ed.). Thomson Brooks/Cole. Ch. 11 Section 11.11 (Taylor Series applications) for approximation and remainder ideas; asymptotic notation extends beyond the core source.