Series: Convergence Tests and Absolute Convergence

Convergence tests are systematic criteria for deciding whether an infinite series

\sum a_{n}

converges (sums to a finite value) or diverges. Absolute convergence (

\sum ∣ a_{n} ∣

converges) is the strong form that guarantees rearrangements and manipulations behave; conditional convergence is the fragile case that does not. Applies throughout analysis, ODEs, Fourier theory, and numerical computation. Developed by Cauchy, d'Alembert, Abel, Dirichlet, and Riemann in the 18th–19th centuries.

An infinite series $\sum_{n = 1}^{\infty} a_{n}$ is defined as the limit of its sequence of partial sums $s_{N} = \sum_{n = 1}^{N} a_{n}$ ; if that limit exists and is finite, the series converges. The central problem is that we usually cannot compute the limit in closed form, so we need tests — comparisons, ratios, roots, and integral bounds — that certify convergence or divergence without summing. The single most important structural distinction is between absolute convergence (where $\sum ∣ a_{n} ∣$ converges) and conditional convergence (where $\sum a_{n}$ converges but $\sum ∣ a_{n} ∣$ diverges). The surprising fact, due to Riemann, is that a conditionally convergent series can be rearranged to sum to any real number whatsoever, or to diverge — order matters fundamentally. This makes absolute convergence the gold standard for safe manipulation.

The Core Insight

Convergence is a statement about the tail of a series: only the asymptotic behavior of $a_{n}$ matters, never the first finitely many terms. Almost every test reduces to comparing your series against a known benchmark — usually the geometric series $\sum r^{n}$ (geometric decay) or the p-series $\sum n^{- p}$ (polynomial decay). Absolute convergence is what separates "summation behaves like a finite sum" from "summation is a delicate limit that order can break."

Formal Statement

Series, Partial Sums, Convergence

Given a sequence $(a_{n})_{n \geq 1}$ of real (or complex) numbers, the $N$ -th partial sum is

s_{N} = n = 1 \sum N a_{n} .

⟳A finite sum; no limit is taken yet.

The series $\sum_{n = 1}^{\infty} a_{n}$ converges to $S$ if the sequence of partial sums has a finite limit:

S = N \to \infty lim s_{N}, S \in R .

⟳This is the definition of series convergence — convergence of the partial-sum sequence in the usual sense.

Otherwise the series diverges. Here $a_{n}$ is the general term, $S$ is the sum, and the tail is $\sum_{n = M}^{\infty} a_{n}$ for large $M$ .

Absolute and Conditional Convergence

The series $\sum a_{n}$ converges absolutely if the series of absolute values converges:

n = 1 \sum \infty ∣ a_{n} ∣ < \infty.

⟳Apply the convergence definition [2] to the nonnegative sequence

∣ a_{n} ∣

It converges conditionally if $\sum a_{n}$ converges but $\sum ∣ a_{n} ∣$ diverges.

The Divergence (n-th Term) Test

If $\sum a_{n}$ converges, then $lim_{n \to \infty} a_{n} = 0$ . Contrapositively:

n \to \infty lim a_{n} \neq = 0 ⟹ \sum a_{n} diverges.

⟳

a_{n} = s_{n} - s_{n - 1} \to S - S = 0

s_{n} \to S

Warning:

a_{n} \to 0

does not imply convergence (the harmonic series is the standard counterexample).

Absolute Convergence Implies Convergence

If $\sum ∣ a_{n} ∣$ converges, then $\sum a_{n}$ converges, and

n = 1 \sum \infty a_{n} \leq n = 1 \sum \infty ∣ a_{n} ∣.

⟳Use the triangle inequality on partial sums and Cauchy completeness of

R

The Standard Comparison Tests

Let $0 \leq a_{n} \leq b_{n}$ for all large $n$ . Direct comparison: if $\sum b_{n}$ converges then $\sum a_{n}$ converges; if $\sum a_{n}$ diverges then $\sum b_{n}$ diverges. Limit comparison: if $a_{n}, b_{n} > 0$ and

L = n \to \infty lim \frac{a _{n}}{b _{n}}, 0 < L < \infty,

⟳For large

n

\frac{L}{2} b_{n} < a_{n} < 2 L b_{n}

; sandwich and apply direct comparison both ways.

then $\sum a_{n}$ and $\sum b_{n}$ both converge or both diverge.

Ratio and Root Tests

Define (when the limits exist)

ρ = n \to \infty lim \frac{a _{n + 1}}{a _{n}}, σ = n \to \infty lim n ∣ a_{n} ∣ .

⟳Both compare

∣ a_{n} ∣

to a geometric series with ratio

ρ

(resp.

σ

If $ρ < 1$ (resp. $σ < 1$ ) the series converges absolutely; if $ρ > 1$ (resp. $σ > 1$ ) it diverges; if $= 1$ the test is inconclusive. The root test (using $lim sup$ ) is strictly stronger than the ratio test.

Integral Test

If $f : [1, \infty) \to [0, \infty)$ is continuous, positive, and decreasing with $f (n) = a_{n}$ , then

n = 1 \sum \infty a_{n} converges ⟺ \int_{1}^{\infty} f (x) d x converges,

⟳Bound the series above and below by integrals using monotone rectangles:

\int_{1}^{N + 1} f \leq s_{N} \leq a_{1} + \int_{1}^{N} f

with the error bound $0 \leq \sum_{n = N + 1}^{\infty} a_{n} \leq \int_{N}^{\infty} f (x) d x$ .

Alternating Series Test (Leibniz)

If $b_{n} > 0$ is monotonically decreasing with $lim_{n \to \infty} b_{n} = 0$ , then $\sum (- 1)^{n - 1} b_{n}$ converges, and the truncation error obeys

∣ S - s_{N} ∣ \leq b_{N + 1} .

⟳Partial sums of odd/even index are monotone nested intervals shrinking to a point because

b_{n} \to 0

Why It Holds / Derivation

:::proof[Absolute Convergence ⟹ Convergence (Theorem [5])] Step 1. Suppose $\sum ∣ a_{n} ∣$ converges. We use the Cauchy criterion: a series converges iff for every $ε > 0$ there is $N$ such that $∣ s_{m} - s_{n} ∣ < ε$ for all $m > n \geq N$ .

Step 2. By the triangle inequality, for

m > n

∣ s_{m} - s_{n} ∣ = k = n + 1 \sum m a_{k} \leq k = n + 1 \sum m ∣ a_{k} ∣.

Step 3. Because

\sum ∣ a_{n} ∣

converges, its partial sums are Cauchy: given

ε > 0

, choose

N

so that

\sum_{k = n + 1}^{m} ∣ a_{k} ∣ < ε

for

m > n \geq N

. Step 4. Combining Steps 2–3,

∣ s_{m} - s_{n} ∣ < ε

, so

(s_{N})

is Cauchy. Since

R

(and

C

) is complete,

(s_{N})

converges. Step 5. Taking

n \to \infty

∣ s_{n} ∣ \leq \sum_{k = 1}^{n} ∣ a_{k} ∣

gives the bound [5].

■

Ratio Test convergence case ($\rho<1$)

Step 1. Pick

r

with

ρ < r < 1

. By definition of the limit, there is

N

such that

∣ a_{n + 1} / a_{n} ∣ < r

for all

n \geq N

. Step 2. Inductively,

∣ a_{N + k} ∣ < r^{k} ∣ a_{N} ∣

for

k \geq 0

. Step 3. Then

\sum_{k \geq 0} ∣ a_{N + k} ∣ < ∣ a_{N} ∣ \sum_{k \geq 0} r^{k} = \frac{∣ a _{N} ∣}{1 - r} < \infty

, a convergent geometric series. Step 4. Hence the tail

\sum_{n \geq N} ∣ a_{n} ∣

converges, so

\sum ∣ a_{n} ∣

converges (finitely many extra terms cannot change this), giving absolute convergence.

■

Augustin-Louis Cauchy | 1789–1857

Cauchy gave the first rigorous definition of series convergence via partial sums and the Cauchy criterion in his Cours d'analyse (1821), replacing the loose 18th-century manipulation of divergent series. He introduced the root test and the condensation idea, and insisted that convergence be proved, not assumed. His insistence on rigor — that a series is its sequence of partial sums and nothing more — is the foundation every test in this article rests on.

Building Intuition

Think of a series as adding up an infinite list of "deposits." Convergence means the running total settles toward a finite number; the deposits must eventually shrink (that's the $n$ -th term test) — but shrinking is not enough, because they might shrink too slowly (the harmonic series shrinks like $1/ n$ and still grows without bound).

Two benchmark "speeds" anchor your intuition:

Geometric decay $r^{n}$ with $∣ r ∣ < 1$ converges because each term is a fixed fraction of the previous — exponentially fast. The ratio and root tests detect exactly this regime.
Polynomial decay $n^{- p}$ converges iff $p > 1$ . The cutoff is $p = 1$ (the harmonic series), the boundary between "fast enough" and "too slow." The integral test pins down this threshold by comparing the sum to $\int x^{- p} d x$ .

Absolute convergence is the difference between piling up and cancelling. If $\sum ∣ a_{n} ∣$ converges, the positive and negative parts each form convergent piles, and order is irrelevant. In conditional convergence the sum survives only through delicate cancellation — like balancing on a knife's edge — and reshuffling terms destroys the balance.

Comparison of decay rates: geometric vs p-series convergence behavior

Worked Examples

Ratio Test on a Factorial Series

Given: Determine convergence of

n = 1 \sum \infty \frac{3 ^{n}}{n !}

. Step 1. Form the ratio:

\frac{a _{n + 1}}{a _{n}} = \frac{3 ^{n + 1}}{( n + 1 )!} \cdot \frac{n !}{3 ^{n}} = \frac{3}{n + 1} .

Step 2. Take the limit:

ρ = n \to \infty lim \frac{3}{n + 1} = 0.

Step 3. Since

ρ = 0 < 1

, by the Ratio Test [7] the series converges absolutely. Step 4. (Bonus identification) In fact

\sum_{n = 0}^{\infty} \frac{3 ^{n}}{n !} = e^{3}

, so

\sum_{n = 1}^{\infty} \frac{3 ^{n}}{n !} = e^{3} - 1

Converges absolutely; sum = e^{3} - 1 \approx 19.086

Conditional Convergence and an Error Bound

Given: The alternating harmonic series

n = 1 \sum \infty \frac{( - 1 ) ^{n - 1}}{n}

. Decide its convergence type and bound the error of

s_{5}

. Step 1. Check absolute convergence:

\sum \frac{1}{n}

is the harmonic series, which diverges (integral test:

\int_{1}^{\infty} x^{- 1} d x = \infty

). So it is not absolutely convergent. Step 2. Apply the Alternating Series Test [9]:

b_{n} = 1/ n

is positive, decreasing, and

\to 0

. Therefore the series converges. Step 3. Conclusion: it converges conditionally. (Its sum is

ln 2 \approx 0.6931

.) Step 4. Error bound for the 5-term partial sum:

s_{5} = 1 - \frac{1}{2} + \frac{1}{3} - \frac{1}{4} + \frac{1}{5} = \frac{47}{60} \approx 0.7833.

By [9],

∣ S - s_{5} ∣ \leq b_{6} = \frac{1}{6} \approx 0.1667.

Conditionally convergent; S = ln 2, ∣ S - s_{5} ∣ \leq 0.1667

Common Misconceptions

" $a_{n} \to 0$ means the series converges." False. This is necessary but not sufficient. The harmonic series

\sum 1/ n

has

a_{n} \to 0

yet diverges. The

n

-th term test can only ever prove divergence, never convergence. "The ratio test with $ρ = 1$ means the series diverges." No —

ρ = 1

is inconclusive. Both

\sum 1/ n^{2}

(converges) and

\sum 1/ n

(diverges) give

ρ = 1

. Switch to a sharper tool (integral test, comparison, Raabe's test). "Convergent series can be rearranged and grouped freely." Only absolutely convergent series may be freely rearranged. Conditionally convergent ones cannot (see the counterexample below).

Riemann Rearrangement of the Alternating Harmonic Series

The series $\sum \frac{( - 1 ) ^{n - 1}}{n} = ln 2$ . By Riemann's Rearrangement Theorem, because it is only conditionally convergent, its terms can be reordered to sum to any target $L \in R$ (or to $\pm \infty$ ). For instance, grouping two positive terms before each negative term:

1 + \frac{1}{3} - \frac{1}{2} + \frac{1}{5} + \frac{1}{7} - \frac{1}{4} + \dots = \frac{3}{2} ln 2 \neq = ln 2.

The same numbers in a different order give a different sum. This is impossible for absolutely convergent series, where every rearrangement yields the identical sum. The mechanism: both

\sum (positive terms)

and

\sum ∣ negative terms ∣

diverge to

+ \infty

, so we can pull as many positives or negatives as needed to overshoot any target repeatedly.

Connections

Convergence tests connect to many neighbouring topics. The Geometric Series and p-series are the universal benchmarks behind comparison, ratio, and root tests. Power Series use the ratio/root test to compute their Radius of Convergence, on which Taylor and Maclaurin expansions depend. Improper Integrals are tied to series by the Integral Test — convergence questions are mirror images. The notion of absolute convergence generalizes to Uniform Convergence (via the Weierstrass M-test) and to normed-space convergence in Functional Analysis, where absolutely convergent series in a Banach space always converge — in fact this property characterizes completeness. Fourier Series convergence is a sophisticated extension where conditional/pointwise/uniform distinctions become central.

Real-World Applications

Numerical analysis / scientific computing. Absolute convergence guarantees that floating-point summation order does not change the limit (only roundoff), making algorithms like Kahan summation safe; conditionally convergent sums require careful ordering.
Physics — perturbation theory. Quantum-mechanical and field-theory expansions are often only asymptotic (divergent); convergence-test reasoning tells physicists where to truncate (optimal truncation) and when Borel resummation is needed.
Signal processing — Fourier analysis. Determining whether a signal's Fourier series converges (and absolutely vs. conditionally) governs whether reconstruction is order-independent and free of Gibbs-type pathologies.
Control theory / stability. The geometric-series convergence condition $∣ r ∣ < 1$ underlies discrete-time system stability (the Neumann series $\sum A^{n} = (I - A)^{- 1}$ converges when the spectral radius is below 1).
Probability — expected values. $E [X] = \sum x_{n} p_{n}$ must converge absolutely for the expectation to be well-defined and rearrangement-invariant; conditionally convergent "expectations" are undefined (e.g., the St. Petersburg paradox's divergence).
Finance — present value. Perpetuity and annuity valuations are geometric series; the discount factor $r < 1$ ensures convergence of total discounted cash flow.
Computer graphics — radiosity / global illumination. Light-transport operators are summed as Neumann series whose convergence (spectral radius < 1) guarantees a finite rendered solution.

Historical Development

1668 — Mercator and Newton derive the series for $ln (1 + x)$ , implicitly using the alternating harmonic series long before convergence is understood.
1715 — Brook Taylor publishes Taylor series; convergence questions are ignored.
1748 — Euler manipulates divergent series with brilliant but unrigorous formal methods (e.g., assigning $1 - 1 + 1 - \dots = \frac{1}{2}$ ).
1768 — Jean d'Alembert formulates the ratio test in studying the binomial series.
1821 — Cauchy's Cours d'analyse gives the rigorous definition of convergence, the Cauchy criterion, and the root test, founding modern analysis.
1827 — Niels Henrik Abel proves Abel's test and warns that "divergent series are the invention of the devil," demanding rigor.
1837 — Dirichlet introduces Dirichlet's test and explicitly distinguishes absolute from conditional convergence.
1853 — Bernhard Riemann proves the Rearrangement Theorem, showing conditionally convergent series can sum to any value.
1872 — Weierstrass formulates the M-test, linking absolute and uniform convergence and completing the rigorous edifice.

Subtleties & Limits

The classical tests leave a gray zone at the boundary $ρ = σ = p = 1$ , where increasingly refined criteria (Raabe's, Bertrand's, Kummer's, Gauss's tests) form an infinite hierarchy — but no single test resolves every series; one can always construct a series at the threshold of any given test. The deep statement is du Bois-Reymond's theorem: there is no universal "comparison scale" of series that decides all cases, so a complete algorithmic convergence test cannot exist for arbitrary series. Related open territory includes the convergence of random series $\sum \pm a_{n}$ (governed by the Kolmogorov three-series theorem) and the resummation of divergent series in physics (Borel, Padé, resurgence theory), where assigning meaningful finite "sums" to divergent expansions remains an active research frontier in mathematical physics.