Pointwise Convergence of Sequences of Functions

Consider the functions $f_n(x) = x^n$ on the interval $[0,1]$ . For each fixed $x$ , the numerical sequence $f_n(x)$ has a limit. But that limit depends dramatically on which $x$ you fixed. The resulting limit function looks nothing like any single $f_n$ . This raises a foundational question: what does it even mean for a sequence of functions to "converge"? What properties of the functions survive the limiting process? Pointwise convergence is the most natural, first-pass answer. Check convergence one input at a time. Do it exactly the way you already check convergence of a sequence of real numbers. The idea matters because analysis is full of situations where a complicated function is built as a limit of simpler ones. Examples include Taylor series, Fourier series, iterative approximations to solutions of differential equations, and sequences of step functions approximating integrable functions. Before any of that machinery can be trusted, one must pin down precisely what "the functions approach a limit" means. Pointwise convergence is the baseline notion against which every stronger mode of convergence is measured. The big picture is that pointwise convergence is deceptively weak. It is defined by nothing more than Limits of Sequences of real numbers applied separately at each point. So it inherits none of the "uniform control" that lets continuity, differentiability, or integrability pass to the limit. This is not a defect to be apologized for. It is the entire reason the subject of sequences and series of functions has depth. Recognizing exactly what pointwise convergence fails to guarantee is the motivation for the stronger notion. This is the notion of Uniform Convergence, developed immediately afterward. This article gives the precise definition of pointwise convergence. It builds geometric intuition with graphs, and proves the basic uniqueness theorem for pointwise limits. It works through two asymmetric examples in full detail. It also exhibits a concrete counterexample showing why an extra hypothesis is needed. That extra hypothesis, uniformity, is needed to preserve continuity. It closes by connecting the notion to its role as the entry point into the analysis of function sequences.

Key takeaways

A sequence of functions $f_n : A \to \mathbb{R}$ converges pointwise to $f: A \to \mathbb{R}$ if $f_n(x) \to f(x)$ as a sequence of real numbers, for every fixed $x \in A$ .
Pointwise convergence is checked one point at a time; the rate of convergence may vary wildly from point to point.
The pointwise limit function, when it exists, is unique — this follows directly from uniqueness of limits of real sequences.
Pointwise limits can fail to preserve continuity, differentiability, and integrability, even when every $f_n$ has these properties.
The classic example $f_n(x) = x^n$ on $[0,1]$ converges pointwise to a discontinuous function, despite each $f_n$ being smooth.
Key idea: pointwise convergence is "convergence at each point separately," with no control on how the convergence behaves across points — that missing control is exactly what Uniform Convergence supplies.

Where this lives today

Pointwise convergence is the base case for essentially every construction of a function as a limit or sum of simpler functions. Fourier series of an $L^2$ or piecewise-smooth function converge pointwise (with subtleties at jump discontinuities, per the Dirichlet–Jordan tests) even where they fail to converge uniformly, which is central to signal processing and the analysis of periodic phenomena in physics and engineering. Probability theory uses pointwise convergence of distribution functions (weak convergence is built from a pointwise-at-continuity-points condition) in the study of the Central Limit Theorem and Markov chain limit theorems. In numerical analysis and computer science, iterative algorithms (e.g., value iteration in reinforcement learning, or fixed-point iterations for solving equations) often are proved to converge only pointwise at each state or input before a stronger, uniform rate estimate is established for complexity bounds. In measure theory and Lebesgue integration, pointwise (almost everywhere) convergence together with domination hypotheses is the hypothesis of the Dominated Convergence Theorem, which is exactly the machinery built to repair the integral-interchange failures that pointwise convergence alone cannot guarantee. In cryptography and complexity theory, sequences of functions (e.g., pseudo-random generators indexed by security parameter) are sometimes analyzed via pointwise behavior on fixed inputs before uniform-over-inputs statements are proved. In every one of these settings, the story is the same as the one told here: pointwise convergence is the natural first definition, and its known weaknesses are precisely what motivate the stronger machinery (uniform convergence, dominated/monotone convergence, equicontinuity) built on top of it.

:::quiz[For $f_n(x) = x^n$ on $[0,1]$ , the pointwise limit function $f$ is]

Continuous everywhere on $[0,1]$
Equal to $0$ on $[0,1)$ and $1$ at $x=1$ ✓
Equal to $1$ on $[0,1)$ and $0$ at $x=1$
Not a well-defined function because the limit fails to exist at $x=1$

:::quiz[In the worked Example 2, $f_n(x) = \dfrac{1}{1+n^2(x^2 - 1/4)^2}$ on $[-1,1]$ , the pointwise limit function $f$ is discontinuous at]

No points; $f$ is continuous on $[-1,1]$
Only $x = 0$
$x = 1$ and $x=-1$ only
$x = \tfrac12$ and $x = -\tfrac12$ ✓

Ask the tutor: Why does upgrading pointwise convergence to uniform convergence restore the preservation of continuity, and what exactly goes right in that proof that goes wrong here?

A \subseteq \mathbb{R}

Intuition

Think of each $f_n$ as a snapshot — a graph in the plane — and imagine flipping through the snapshots $f_1, f_2, f_3, \dots$ like a flipbook. Pointwise convergence says: pick any vertical line $x = x_0$ ; look at where each snapshot's graph crosses that vertical line; those crossing heights $f_n(x_0)$ form an ordinary sequence of real numbers, and we require that sequence to converge to $f(x_0)$ . Crucially, nothing is asked about how the whole graph moves at once. Different vertical lines are allowed to converge at wildly different speeds — some columns of the flipbook may snap into place almost immediately, others may take a very long time, with no uniform bound across the whole picture. The standard example makes this vivid. Let $f_n(x) = x^n$ on $[0,1]$ . At $x = 0$ : $f_n(0) = 0$ for all $n$ , so the sequence is constant, already at its limit. At $x = 1$ : $f_n(1) = 1$ for all $n$ . At any $x \in (0,1)$ : $x^n \to 0$ , but the speed of decay depends on $x$ — for $x$ close to $1$ , it takes enormously many terms before $x^n$ gets small, while for $x$ close to $0$ , convergence is almost instantaneous. The pointwise limit function is

f(x) = \begin{cases} 0 & 0 \le x < 1 \\ 1 & x = 1. \end{cases}

Every $f_n$ is continuous (indeed a polynomial), yet the limit function $f$ has a jump discontinuity at $x=1$ . This single example is the reason pointwise convergence, by itself, is too weak to transport continuity to the limit — motivating the sharper, uniform notion of convergence to come.

:::svg[The graphs of $x^n$ on $[0,1]$ pointwise-approach a discontinuous "step" limit]

The central theorem

A \subseteq \mathbb{R}

x \in A

f_n \to f

x \in A

Worked examples

Example 1 (a non-monotone, asymmetric rational sequence). Define $f_n : \mathbb{R} \to \mathbb{R}$ by

f_n(x) = \frac{n x^2 + \sin(nx)}{n x^2 + n}, \qquad n = 1,2,3,\dots

We determine the pointwise limit on all of $\mathbb{R}$ . Fix $x \in \mathbb{R}$ . Write

f_n(x) = \frac{nx^2 + \sin(nx)}{n(x^2+1)} = \frac{x^2}{x^2+1} + \frac{\sin(nx)}{n(x^2+1)}.

This splitting is valid for every $n \ge 1$ since $x^2 + 1 \ge 1 > 0$ , so no division by zero occurs. Now bound the second term: since $|\sin(nx)| \le 1$ for all $n, x$ ,

\left| \frac{\sin(nx)}{n(x^2+1)} \right| \le \frac{1}{n(x^2+1)} \le \frac{1}{n} \to 0 \quad \text{as } n \to \infty,

using $x^2 + 1 \ge 1$ . By the squeeze theorem for real sequences, $\dfrac{\sin(nx)}{n(x^2+1)} \to 0$ as $n \to \infty$ , for this fixed $x$ . The first term $\dfrac{x^2}{x^2+1}$ does not depend on $n$ at all, so trivially it "converges" to itself. By the sum rule for limits of real sequences,

f_n(x) = \frac{x^2}{x^2+1} + \frac{\sin(nx)}{n(x^2+1)} \;\longrightarrow\; \frac{x^2}{x^2+1} + 0 = \frac{x^2}{x^2+1}.

Since $x \in \mathbb{R}$ was arbitrary, we conclude $f_n \to f$ pointwise on $\mathbb{R}$ , where

f(x) = \frac{x^2}{x^2+1}.

Note the asymmetry deliberately built in: the oscillatory term $\sin(nx)$ has amplitude that does not decay uniformly in a simple way (its argument $nx$ sweeps through all phases as $n\to\infty$ whenever $x \ne 0$ ), yet the denominator's growth in $n$ still forces pointwise convergence to zero, at every fixed $x$ , regardless of $x$ 's sign or size. Also note $f$ is continuous and bounded ( $0 \le f(x) < 1$ ), even though the individual $f_n$ were not obviously so at first glance — this example does not break continuity, in contrast with our second example.

Example 2 (a sequence built from indicator-like bumps — non-uniform, non-monotone, and limit discontinuous at two points). Define $f_n : [-1, 1] \to \mathbb{R}$ by

f_n(x) = \frac{1}{1 + n^2 (x^2 - \tfrac14)^2}, \qquad n = 1, 2, 3, \dots

Each $f_n$ is a smooth (infinitely differentiable) function on $[-1,1]$ , being a composition of polynomials with the reciprocal function, and the denominator $1 + n^2(x^2 - 1/4)^2$ is always $\ge 1 > 0$ . We compute the pointwise limit. Fix $x \in [-1,1]$ and consider two cases. Case A: $x^2 = \tfrac14$ , i.e. $x = \tfrac12$ or $x = -\tfrac12$ . Then $(x^2 - \tfrac14)^2 = 0$ , so $f_n(x) = \dfrac{1}{1 + 0} = 1$ for every $n$ . Hence $f_n(x) \to 1$ . Case B: $x^2 \ne \tfrac14$ . Let $c = c(x) := (x^2 - \tfrac14)^2 > 0$ , a fixed positive constant depending on $x$ but not on $n$ . Then

f_n(x) = \frac{1}{1 + n^2 c}.

As $n \to \infty$ , $n^2 c \to \infty$ (since $c>0$ is fixed), so $1 + n^2 c \to \infty$ , and therefore $f_n(x) = \dfrac{1}{1+n^2 c} \to 0$ . (Formally: given $\varepsilon>0$ , choose $N$ with $N^2 c > \tfrac1\varepsilon - 1$ , i.e. $N > \sqrt{(1/\varepsilon - 1)/c}$ ; such $N$ exists since $\mathbb{R}$ is Archimedean; then for $n \ge N$ , $f_n(x) = 1/(1+n^2c) < \varepsilon$ .) Combining both cases, the pointwise limit function is

f(x) = \begin{cases} 1 & x = \tfrac12 \text{ or } x = -\tfrac12 \\ 0 & \text{otherwise.} \end{cases}

This $f$ is discontinuous at exactly two points, $x = \pm\tfrac12$ , despite every $f_n$ being smooth on all of $[-1,1]$ . Moreover the convergence is markedly non-uniform in speed: near $x = \pm\frac12$ the constant $c(x)$ is tiny, so an enormous $n$ is needed before $f_n(x)$ drops near $0$ , while for $x$ far from $\pm\frac12$ (e.g. $x=0$ , where $c = 1/16$ ) convergence to $0$ is comparatively fast. This is a second, structurally different illustration of the same phenomenon as Example 1's cousin $x^n$ : smooth building blocks, pointwise limit not even continuous.

f_n : A \to \mathbb{R}

Scope, edge cases & common traps

Situation	What happens
$f_n(x) \to f(x)$ but rate of convergence depends on $x$	Perfectly allowed; this is the generic case for pointwise convergence
Sequence $(f_n(x))$ diverges for some $x \in A$	Pointwise convergence on $A$ fails; one may restrict to the (possibly smaller) set of $x$ where it does converge
Each $f_n$ continuous/differentiable/integrable	Limit $f$ need not be continuous, differentiable, or (Riemann) integrable
$\lim_n \int f_n$ vs. $\int \lim_n f_n$	Need not agree under pointwise convergence alone (interchange of limit and integral can fail)
$f_n \to f$ and $g_n \to g$ pointwise	$f_n + g_n \to f+g$ , $f_n g_n \to fg$ pointwise — algebraic operations ARE safe (Corollary above)
Uniqueness of the limit function	Always holds, by the Uniqueness Theorem — no extra hypotheses needed

Common trapThe classic quantifier error

The defining condition of pointwise convergence is

\forall x \in A\ \forall \varepsilon>0\ \exists N(\varepsilon,x)\ \forall n \ge N:\ |f_n(x)-f(x)|<\varepsilon.

The single most common error is silently assuming $N$ can be chosen independently of $x$ — i.e. mentally swapping to $\forall \varepsilon>0\ \exists N\ \forall x \in A\ \forall n \ge N$ . That swapped statement is the definition of uniform convergence, a strictly stronger condition. Every "pointwise implies continuous/integrable/differentiable limit" claim that students are tempted to assert rests on exactly this illegitimate quantifier swap. Always ask: "does my choice of $N$ actually depend on $x$ ?" If it must, you only have pointwise convergence, and none of the preservation theorems for continuity, differentiability, or (naively) integrals are available without further hypotheses.

The mathematician behind it

Origins · 1 minKarl Weierstrass and how this was discovered

Karl Weierstrass

1815–1897 · Rigorization of analysis at the University of Berlin, mid-to-late 19th century

Karl Weierstrass is the figure most responsible for making precise the distinction between pointwise and uniform modes of convergence, a distinction that earlier analysts (including, at times, Cauchy) had blurred, leading to erroneous claims that limits of continuous functions must be continuous. Weierstrass's insistence on $\varepsilon$ – $N$ (and $\varepsilon$ – $\delta$ ) definitions, delivered in his lecture courses in Berlin from the 1850s onward, gave analysis the toolkit — precisely the toolkit used in the proofs above — to state exactly what is and is not preserved when passing to a limit of functions.

History

- 1734–1746 — Euler and contemporaries manipulate infinite series of functions termwise, often without justification, occasionally producing paradoxes.

- 1821 — Cauchy, in Cours d'Analyse, claims (incorrectly, without the uniformity hypothesis) that a convergent series of continuous functions has a continuous sum.

- 1826 — Abel publishes a counterexample-style remark pointing out cases where Cauchy's claimed theorem seems to fail, hinting at the missing hypothesis.

- 1841–1856 — Weierstrass, in unpublished lecture notes and later published work, isolates and names the distinction between pointwise and uniform convergence, supplying the correct hypothesis and correct proofs.

- 1861 onward — Weierstrass's Berlin lectures standardize the

\varepsilon

–

N

\varepsilon

–

\delta

formalism now universal in real analysis, cementing pointwise convergence as the baseline definition against which uniform convergence is contrasted.

Is pointwise convergence the same as "the graphs get close to each other"?

Not quite — that informal picture is closer to uniform convergence. Pointwise convergence only requires that at each individual input $x$ , the numbers $f_n(x)$ eventually get close to $f(x)$ ; it says nothing about the graphs looking close as whole curves. In Example 2 above, the graphs of $f_n$ develop a tall, narrow spike near $x=\pm\frac12$ that never disappears in height (it always equals $1$ exactly at $x=\pm\frac12$ ), yet the sequence still converges pointwise, because at each fixed $x$ the spike eventually moves away from that particular $x$ (unless $x$ is exactly $\pm\frac12$ ). The graphs as a whole do not uniformly hug the limit function.

Can a pointwise limit fail to exist even though every

f_n(x)

is a perfectly good real number?

Yes. Pointwise convergence requires that the sequence $(f_n(x))_n$ converges for each $x$ , not merely that each term is defined. For example, $f_n(x) = (-1)^n$ (constant in $x$ , alternating in $n$ ) has every value well-defined, but the sequence $(-1)^n$ diverges (oscillates between $-1$ and $1$ ) for every $x$ , so there is no pointwise limit function at all. Similarly $f_n(x) = \sin(nx)$ fails to converge pointwise at most points $x$ . One must always verify convergence of the real sequence at each $x$ before speaking of "the" pointwise limit.

Does pointwise convergence preserve boundedness?

No, not in general, and not even monotonic boundedness bounds survive uniformly. It is true that if each $f_n$ is bounded by some fixed constant $M$ independent of $n$ (i.e. $|f_n(x)| \le M$ for all $n,x$ ), then the pointwise limit $f$ also satisfies $|f(x)| \le M$ , by taking the limit of the inequality $|f_n(x)| \le M$ termwise at each fixed $x$ . But if the bound for $f_n$ is allowed to grow with $n$ , boundedness of the limit is not guaranteed at all without more care, and in fact even convergence itself may fail to occur in the first place; boundedness questions for the limit are typically handled via the stronger tool of uniform boundedness together with uniform convergence.

Why doesn't the "obvious" proof that continuity is preserved under pointwise convergence work?

The natural attempted proof tries to show $f$ is continuous at $a$ by writing $|f(x)-f(a)| \le |f(x)-f_n(x)| + |f_n(x)-f_n(a)| + |f_n(a)-f(a)|$ and making each piece small. The first and third pieces can be made small by choosing $n$ large — but the required "large enough $n$ " depends on which $x$ you're looking at (since it's pointwise convergence at $x$ , and separately at $a$ ), so you cannot fix one $n$ that works for all $x$ near $a$ before invoking continuity of $f_n$ at $a$ for the middle term. This gap is unavoidable under pointwise hypotheses alone, and is exactly plugged by the additional uniformity hypothesis in the theory of Uniform Convergence.

Is pointwise convergence a topology, i.e. can it be described by open sets, or a metric?

Pointwise convergence on an infinite domain $A$ is generally not metrizable — there is no single metric $d$ on the space of functions $A \to \mathbb{R}$ such that $d(f_n,f)\to 0$ is equivalent to pointwise convergence, when $A$ is uncountable (this is a standard fact from general topology/functional analysis, beyond this article's scope). It is, however, describable as the product topology on $\mathbb{R}^A$ (functions $A\to\mathbb{R}$ viewed as elements of a product of copies of $\mathbb{R}$ , one per point of $A$ ), where convergence in the product topology is by definition coordinatewise convergence — exactly pointwise convergence. This connection is usually developed later, once general topology is available.

How is pointwise convergence different for sequences versus series of functions?

A series of functions $\sum_{k=1}^\infty g_k(x)$ is said to converge pointwise on $A$ exactly when its sequence of partial sums $S_n(x) = \sum_{k=1}^n g_k(x)$ converges pointwise on $A$ , in the sense defined here. So series convergence is not a separate notion — it is pointwise (or uniform) convergence of the associated sequence of partial-sum functions. Every subtlety about pointwise convergence of sequences (non-uniform rates, failure to preserve continuity, etc.) therefore transfers directly to series, which is why term-by-term differentiation/integration of power series and Fourier series requires justification beyond bare pointwise convergence.

f_n \to f

pointwise, does

f_n' \to f'

pointwise (assuming all derivatives exist)?

Not in general, and this is one of the more dangerous traps. Even uniform convergence of $f_n \to f$ does not by itself guarantee convergence of derivatives; one needs the derivatives $f_n'$ to converge (uniformly, typically) to conclude $f' = \lim f_n'$ . A standard example is $f_n(x) = \frac{\sin(nx)}{\sqrt n}$ on $\mathbb{R}$ : here $f_n \to 0$ pointwise (in fact uniformly, since $|f_n(x)| \le 1/\sqrt n \to 0$ for all $x$ ), but $f_n'(x) = \sqrt n \cos(nx)$ , whose values do not converge to $0$ (or to anything) at most points. Differentiation is therefore one of the most fragile operations under any weak notion of convergence.