Convergence of Sequences and Series

Motivation: why this matters in quant finance

Every numerical method in quant finance produces a sequence of approximations that (hopefully) converges to the true answer. The binomial tree with

n

steps gives a price

C_n

that converges to the Black-Scholes price as

n \to \infty

. A Monte Carlo simulation with

N

samples gives an estimator

\hat{C}_N

that converges to the true expected value. Newton-Raphson iteration for implied volatility produces a sequence

\sigma_0, \sigma_1, \sigma_2, \ldots

that converges to

\sigma_{\text{impl}}

The questions a practitioner needs answered are: does the sequence converge?, how fast?, and when can I stop? Convergence tests answer the first question. Convergence rates answer the second and third. This page provides the mathematical tools for both.

Convergence of sequences

Definition

A sequence

\{a_n\}

converges to

L

if:

\forall\,\varepsilon > 0, \;\exists\,N: n > N \implies |a_n - L| < \varepsilon

We write

a_n \to L

\lim_{n \to \infty} a_n = L

. A sequence that does not converge diverges.

Monotone convergence theorem

A bounded, monotone sequence converges.

If $\{a_n\}$ is increasing and bounded above, it converges to $\sup_n a_n$ . If decreasing and bounded below, it converges to $\inf_n a_n$ .

Finance application: Many iterative calibration algorithms produce sequences that are monotonically decreasing in error (e.g., each step of a least-squares minimisation reduces the objective). If the error is bounded below by zero, the monotone convergence theorem guarantees the sequence converges.

Convergence rates

The rate of convergence describes how fast

a_n \to L

Rate	Condition	Example in finance
$O(1/n)$	$	a_n - L
$O(1/n^2)$	$	a_n - L
$O(1/\sqrt{n})$	$	a_n - L
Geometric / $O(r^n)$	$	a_n - L

The distinction matters enormously in practice. To halve the error of a Monte Carlo estimator ( $O(1/\sqrt{N})$ ), you need 4× the samples. To halve the error of a binomial tree ( $O(1/n)$ ), you need 2× the steps. Newton-Raphson for implied vol converges geometrically — each iteration roughly doubles the number of correct digits.

Cauchy sequences

A sequence

\{a_n\}

is a Cauchy sequence if:

\forall\,\varepsilon > 0, \;\exists\,N: m, n > N \implies |a_m - a_n| < \varepsilon

In $\mathbb{R}$ , a sequence converges if and only if it is Cauchy. The practical significance: you can check convergence without knowing the limit. In numerical algorithms, you monitor $|a_{n+1} - a_n|$ and stop when it falls below a tolerance — this is a Cauchy criterion applied in practice.

Convergence of series

Definition

A series $\sum_{n=1}^{\infty} a_n$ converges if the sequence of partial sums $S_N = \sum_{n=1}^{N} a_n$ converges: $S_N \to S$ as $N \to \infty$ .

Necessary condition (divergence test)

If $\sum a_n$ converges, then $a_n \to 0$ .

The contrapositive: if

a_n \not\to 0

, the series diverges. Warning:

a_n \to 0

is necessary but not sufficient — the harmonic series

\sum 1/n

diverges even though

1/n \to 0

Convergence tests

Geometric series:

\sum_{n=0}^{\infty} r^n = \frac{1}{1-r}

converges if

|r| < 1

. This is the basis of perpetuity valuation:

\text{PV} = \sum_{n=1}^{\infty} c/(1+r)^n = c/r

for

r > 0

. See Discounting.

$p$ -series:

\sum_{n=1}^{\infty} 1/n^p

converges if

p > 1

, diverges if

p \leq 1

. This benchmark is used in comparison tests.

Comparison test: If

0 \leq a_n \leq b_n

and

\sum b_n

converges, then

\sum a_n

converges. The continuous analogue is the comparison test for improper integrals.

Ratio test: If

L = \lim_{n \to \infty} |a_{n+1}/a_n|

exists, then the series converges absolutely if

L < 1

and diverges if

L > 1

. The ratio test is how you determine the radius of convergence of power series and Taylor series:

R = 1/L

Root test: If

L = \limsup_{n \to \infty} |a_n|^{1/n}

, then the series converges absolutely if

L < 1

and diverges if

L > 1

Integral test: If

f

is positive, continuous, and decreasing on

[1, \infty)

with

a_n = f(n)

, then

\sum a_n

converges if and only if

\int_1^{\infty} f(x)\,dx

converges. This connects series convergence to improper integral convergence.

Alternating series test (Leibniz): If

a_n > 0

a_n

is decreasing, and

a_n \to 0

, then

\sum (-1)^{n+1} a_n

converges. The error after

N

terms is bounded by

|a_{N+1}|

Absolute vs conditional convergence

A series converges absolutely if

\sum |a_n|

converges. It converges conditionally if

\sum a_n

converges but

\sum |a_n|

diverges. Absolutely convergent series can be rearranged in any order without changing the sum; conditionally convergent series cannot (Riemann rearrangement theorem).

In quant finance, the series you encounter (Taylor series within their radius, geometric series for PV calculations, Fourier series for characteristic functions) are typically absolutely convergent, so rearrangement is safe.

Examples and applications

Example 1: Newton-Raphson for implied volatility

Given a market call price $C_{\text{mkt}}$ , Newton-Raphson iterates:

\sigma_{n+1} = \sigma_n - \frac{C_{\text{BS}}(\sigma_n) - C_{\text{mkt}}}{\mathcal{V}(\sigma_n)}

where $\mathcal{V}$ is vega. Near the root, the error satisfies:

|\sigma_{n+1} - \sigma^*| \leq M \cdot |\sigma_n - \sigma^*|^2

This is quadratic convergence: each iteration doubles the number of correct digits. Starting from

\sigma_0 = 0.20

with true implied vol

\sigma^* = 0.2347

, typical convergence is:

Iteration	$\sigma_n$	Error
0	0.2000	$3.5 \times 10^{-2}$
1	0.2340	$7 \times 10^{-4}$
2	0.2347	$3 \times 10^{-7}$
3	0.2347	machine precision

Three iterations suffice. This geometric convergence rate is why Newton-Raphson is the standard method for implied vol computation. See Implicit Differentiation for the sensitivity analysis.

Example 2: Monte Carlo convergence rate

A Monte Carlo estimator

\hat{C}_N = \frac{1}{N}\sum_{i=1}^{N} f(S_T^{(i)})

estimates

\mathbb{E}[f(S_T)]

. By the Central Limit Theorem:

\hat{C}_N \approx \mathcal{N}\left(C, \frac{\text{Var}(f(S_T))}{N}\right)

The standard error is

\sigma/\sqrt{N}

— convergence rate

O(1/\sqrt{N})

. To reduce the error by a factor of 10, you need 100× more samples. This is slow compared to Newton-Raphson but dimension-independent, which is why Monte Carlo dominates for high-dimensional problems (basket options, path-dependent payoffs).

Example 3: convergence of the binomial tree to Black-Scholes

The CRR binomial tree price

C_n

satisfies:

|C_n - C_{\text{BS}}| = O(1/n)

as the number of steps

n \to \infty

. The convergence is not monotone —

C_n

oscillates around

C_{\text{BS}}

(depending on whether the strike falls between or on tree nodes). But

|C_n - C_{\text{BS}}|

is bounded by

M/n

for some constant

M

, so by the squeeze theorem,

C_n \to C_{\text{BS}}

Richardson extrapolation improves this to $O(1/n^2)$ by combining $C_n$ and $C_{2n}$ to cancel the leading error term — a standard trick for accelerating convergence.

Example 4: present value as a geometric series

A bond paying coupon $c$ semi-annually for $T$ years with yield $y$ :

P = \sum_{i=1}^{2T} \frac{c/2}{(1 + y/2)^i} + \frac{F}{(1 + y/2)^{2T}} = \frac{c/2}{y/2}\left(1 - \frac{1}{(1+y/2)^{2T}}\right) + \frac{F}{(1+y/2)^{2T}}

The coupon sum is a finite geometric series with ratio $r = 1/(1 + y/2) < 1$ . As $T \to \infty$ (perpetual bond), the partial sums converge to $c/y$ — the infinite geometric series. The ratio test confirms convergence: $|r| = 1/(1+y/2) < 1$ for $y > 0$ .

Common confusions and pitfalls

" $a_n \to 0$ implies $\sum a_n$ converges." No. The harmonic series

\sum 1/n

diverges. The terms shrinking to zero is necessary but not sufficient for the series to converge.

"Monte Carlo converges faster with more dimensions." No. The

O(1/\sqrt{N})

rate is dimension-independent, which is Monte Carlo's strength — but it does not improve with dimension. Grid-based methods converge faster in low dimensions (

O(1/N^{2/d})

for

d

dimensions) but degrade exponentially with dimension. The crossover is typically around

d = 4

–

6

"Newton-Raphson always converges." Not if the starting point is too far from the root or if the derivative (vega) is too small. For implied vol, starting with

\sigma_0 = 0.20

is usually safe, but for deep OTM or short-dated options where vega is tiny, Newton-Raphson can overshoot and diverge. Bisection (slower but guaranteed) is a safer fallback.

Confusing the rate of convergence with the number of iterations.

O(1/n^2)

convergence means 10× fewer steps for the same accuracy compared to

O(1/n)

. Quadratic (geometric) convergence like Newton-Raphson means 2–3 iterations often suffice regardless of the starting error, which is qualitatively different from polynomial rates.

Where this goes next

Convergence of sequences and series connects to:

Limits: Sequence convergence is the discrete version of the limit concept.
Squeeze Theorem and Bounds: Comparison and bounding arguments are the primary tools for proving convergence.
Power Series: The ratio and root tests determine the radius of convergence of power series.
Numerical Integration: Convergence rates of numerical methods determine computational cost.
Implicit Differentiation: Newton-Raphson convergence analysis for root-finding in calibration.
Central Limit Theorem: The CLT determines the $O(1/\sqrt{N})$ convergence rate of Monte Carlo.

References

Stewart, J. (2008). Single Variable Calculus: Early Transcendentals (6th ed.). Thomson Brooks/Cole. Ch. 11 Sections 11.1-11.6 (sequences, series, integral/comparison tests, alternating series).