Limits, Completeness, and Modes of Convergence

A convergence claim is incomplete until it names the objects, the ambient space and its metric or topology, the limiting parameter, and the mode of convergence. Completeness turns Cauchy control into existence of a limit inside that space. Pointwise and uniform convergence differ by the order of their quantifiers. Moving a limit through continuity, an integral, a derivative, an infinite sum, or a second limit is then a separate theorem, with separate hypotheses.

This page develops metric, pointwise, and uniform convergence and the elementary tests that distinguish them. Measure-theoretic, weak, and distributional convergence are named only at their handoffs.

Convergence data · Completeness · Pointwise versus uniform · Limit interchange · Series · Free-field example · Exercises

What a convergence claim must say

Before manipulating a limit, record seven pieces of data.

1. Objects: numbers, functions, vectors, operators, distributions, or something else.
2. Ambient space: the set in which every approximant and the proposed limit live.
3. Closeness: the metric, norm, seminorms, or other topology that defines convergence.
4. Parameter: for example $n\to\infty$, a cutoff $\Lambda\to\infty$, a regulator $\varepsilon\downarrow0$, or a volume $L\to\infty$.
5. Mode: pointwise, uniform, norm, weak, distributional, or another named mode.
6. Order: if several parameters vary, which limit is taken first, or whether a joint limit is claimed.
7. Operation: the map, sum, integral, derivative, expectation, or second limit through which the first limit is to pass.

The theorem that authorizes the last step must use the same data. A bound that is uniform in $x$ but deteriorates as $\varepsilon\downarrow0$, for example, is not uniform in $(x,\varepsilon)$.

Metric limits, Cauchy sequences, and completeness

Let $(X,d)$ be a metric space. A sequence $(x_n)$ converges to $x\in X$ when

$$x_n\longrightarrow x \quad\Longleftrightarrow\quad (\forall\varepsilon>0)(\exists N) (\forall n\geq N)\; d(x_n,x)<\varepsilon.$$

The metric and the ambient space are part of the assertion. If a sequence has two limits $x$ and $y$, choose $n$ large enough that both $d(x_n,x)<\varepsilon/2$ and $d(x_n,y)<\varepsilon/2$. The triangle inequality gives $d(x,y)<\varepsilon$ for every $\varepsilon>0$, hence $x=y$. Metric limits are unique.

A sequence is Cauchy when its late terms become mutually close:

$$(\forall\varepsilon>0)(\exists N) (\forall m,n\geq N)\; d(x_m,x_n)<\varepsilon.$$

Every convergent sequence is Cauchy. Indeed, if $x_n\to x$, choose $N$ so that $d(x_n,x)<\varepsilon/2$ for $n\geq N$. Then

$$d(x_m,x_n) \leq d(x_m,x)+d(x,x_n) <\varepsilon$$

whenever $m,n\geq N$. The converse is the extra property that matters.

A metric space is complete when every Cauchy sequence in it converges to a point of that same space. Completeness therefore converts an internal error estimate—late approximants are close to one another—into existence of the object they approximate; see Folland 1999, Chapter 0.

A missing limit

Give $\mathbb Q$ the metric inherited from $\mathbb R$ and set

$$q_n=\frac{\lfloor 10^n\sqrt2\rfloor}{10^n}\in\mathbb Q.$$

Since $0\leq\sqrt2-q_n<10^{-n}$, the sequence is Cauchy. In $\mathbb R$ it converges to $\sqrt2$, but no rational number can be its limit. Thus $\mathbb Q$ is not complete. The obstruction is not a failure of the approximants to settle down; the required point is absent from the chosen space.

Completeness is not boundedness. The sequence $(-1)^n$ is bounded in $\mathbb R$ but is not Cauchy. The interval $(0,1)$ is bounded but incomplete, while $\mathbb R$ is complete but unbounded. For orientation, every compact metric space is complete, but a complete metric space need not be compact. A subset of a complete metric space is complete in the inherited metric exactly when it is closed—that is, when it contains every ambient limit of a sequence from the subset.

Completeness also depends on the metric, not merely on the points or their open sets. On $\mathbb R$, the usual metric is complete, whereas

$$\rho(x,y)=|\arctan x-\arctan y|$$

induces the same topology—the same open sets—but is incomplete: $(n)$ is $\rho$-Cauchy and its $\arctan$-image satisfies $\arctan n\to\pi/2\notin\arctan(\mathbb R)$. Completion of normed spaces is developed later on the Banach and Hilbert spaces page.

Pointwise and uniform convergence

Let $E$ be a set, $(Y,d_Y)$ a metric space, and $f_n,f:E\to Y$. Pointwise and uniform convergence differ only in the order of two quantifiers, but that difference controls what survives the limit.

$$\begin{aligned} f_n\to f\text{ pointwise} \quad\Longleftrightarrow\quad &(\forall x\in E)(\forall\varepsilon>0) (\exists N=N(x,\varepsilon))\\ & (\forall n\geq N)\; d_Y(f_n(x),f(x))<\varepsilon, \end{aligned}$$

whereas

$$\begin{aligned} f_n\to f\text{ uniformly} \quad\Longleftrightarrow\quad &(\forall\varepsilon>0) (\exists N=N(\varepsilon))\\ & (\forall x\in E)(\forall n\geq N)\; d_Y(f_n(x),f(x))<\varepsilon. \end{aligned}$$

The uniform threshold cannot depend on $x$. Uniform convergence therefore implies pointwise convergence, but not conversely.

For bounded complex-valued functions, the supremum norm is

$$\lVert f\rVert_\infty = \sup_{x\in E}|f(x)|,$$

and uniform convergence is precisely $\lVert f_n-f\rVert_\infty\to0$. More generally, the sup metric

$$d_\infty(f,g) = \sup_{x\in E}d_Y(f(x),g(x))$$

is used on a class of maps for which the displayed supremum is finite.

The uniform Cauchy criterion

The sequence $(f_n)$ is uniformly Cauchy if

$$(\forall\varepsilon>0)(\exists N) (\forall m,n\geq N)(\forall x\in E)\; d_Y(f_m(x),f_n(x))<\varepsilon.$$

If $Y$ is complete, this criterion produces a uniform limit. For each fixed $x$, the sequence $(f_n(x))$ is Cauchy in $Y$, so define $f(x)=\lim_n f_n(x)$. Given $\varepsilon>0$, choose the uniform Cauchy threshold for $\varepsilon/2$. Hold $n\geq N$ fixed and let $m\to\infty$. The estimate gives

$$d_Y(f_n(x),f(x))\leq\frac{\varepsilon}{2}<\varepsilon \qquad\text{for every }x\in E,$$

so $f_n\to f$ uniformly. Consequently, bounded maps into a complete target form a complete metric space under the sup metric. If $E$ also has a topology, the uniform-limit theorem below shows that the continuous bounded maps form a closed subset and are therefore complete as well. In particular, write $C(K;\mathbb C)$ for the continuous complex-valued functions on a compact space $K$; this function space is complete in the sup norm. The uniform Cauchy criterion and this completeness argument are also treated in Lebl 2026, §§ 6.1–6.2, PDF.

The diagnostic example $x^n$

On $[0,1]$, let $f_n(x)=x^n$. The pointwise limit is

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

The convergence is not uniform because

$$\sup_{x\in[0,1]}|f_n(x)-f(x)|=1$$

for every $n$. Each $f_n$ is continuous, but $f$ is not. The same example shows that two existing iterated limits need not agree:

$$\lim_{x\to1^-}\lim_{n\to\infty}x^n=0, \qquad \lim_{n\to\infty}\lim_{x\to1^-}x^n=1.$$

Uniform convergence on a neighborhood of the second limiting point would rule out this behavior; pointwise convergence does not.

What a limit may pass through

For one fixed continuous map $F:X\to Y$, sequential continuity says

$$x_n\to x \quad\Longrightarrow\quad F(x_n)\to F(x).$$

This is already a limit-interchange theorem: continuity is the hypothesis that licenses $F(\lim_n x_n)=\lim_nF(x_n)$. Operations involving a changing function, an integral, a derivative, or another limiting parameter require different control.

Uniform limits preserve continuity

Suppose $f_n:E\to Y$ is continuous at $x_0$ for every $n$ and $f_n\to f$ uniformly. Given $\varepsilon>0$, choose $N$ such that

$$d_Y(f_N(x),f(x))<\frac{\varepsilon}{3} \qquad\text{for every }x\in E.$$

Continuity of $f_N$ at $x_0$ supplies a neighborhood in which $d_Y(f_N(x),f_N(x_0))<\varepsilon/3$. There,

$$\begin{aligned} d_Y(f(x),f(x_0)) \leq{}&d_Y(f(x),f_N(x)) +d_Y(f_N(x),f_N(x_0))\\ &+d_Y(f_N(x_0),f(x_0)) <\varepsilon. \end{aligned}$$

Thus $f$ is continuous at $x_0$. Equivalently, if $x\to x_0$ within $E$, then

$$\lim_{n\to\infty}\lim_{x\to x_0}f_n(x) = \lim_{x\to x_0}\lim_{n\to\infty}f_n(x) =f(x_0).$$

The $x^n$ example fails exactly because convergence is not uniform near $x=1$.

Integration: uniform convergence is sufficient

If Riemann-integrable functions $f_n:[a,b]\to\mathbb C$ converge uniformly to $f$, then $f$ is Riemann integrable and

$$\lim_{n\to\infty}\int_a^b f_n(x)\,\mathrm dx = \int_a^b f(x)\,\mathrm dx.$$

The decisive estimate is

$$\left| \int_a^b\bigl(f_n(x)-f(x)\bigr)\,\mathrm dx \right| \leq (b-a)\lVert f_n-f\rVert_\infty.$$

Uniform convergence is sufficient, not necessary. Later measure-theoretic theorems replace it with other hypotheses. Pointwise convergence alone is not enough: for $n\geq2$, define the continuous tent

$$h_n(x)= \begin{cases} n^2x,&0\leq x\leq 1/n,\\ n(2-nx),&1/n\leq x\leq2/n,\\ 0,&2/n\leq x\leq1. \end{cases}$$

For every fixed $x\in[0,1]$, $h_n(x)\to0$, but the triangle has base $2/n$ and height $n$, so

$$\int_0^1h_n(x)\,\mathrm dx=1.$$

Hence $\int\lim_n h_n=0$ while $\lim_n\int h_n=1$.

Differentiation: control the derivatives

Uniform convergence of the functions does not by itself authorize differentiation. A useful sufficient theorem is the following. Let $f_n\in C^1([a,b])$ and $x_0\in[a,b]$. If $f_n(x_0)\to L$ and $f_n'\to g$ uniformly, then $g$ is continuous, because it is a uniform limit of continuous functions, and

$$f_n\longrightarrow f \quad\text{uniformly}, \qquad f(x)=L+\int_{x_0}^{x}g(t)\,\mathrm dt, \qquad f'=g.$$

Indeed, the fundamental theorem of calculus gives the uniform estimate

$$|f_n(x)-f(x)| \leq |f_n(x_0)-L| +(b-a)\lVert f_n'-g\rVert_\infty.$$

The derivative hypothesis is essential. On $[-\pi,\pi]$,

$$f_n(x)=\frac{\sin(nx)}{n} \longrightarrow0$$

uniformly, because $\lVert f_n\rVert_\infty\leq1/n$. Yet $f_n'(x)=\cos(nx)$ does not even converge pointwise at $x=\pi$, since $f_n'(\pi)=(-1)^n$. It also satisfies $f_n'(0)=1$ for every $n$, whereas the derivative of the limiting zero function is $0$.

A limit-interchange checklist

Desired step	A sufficient hypothesis used on this page	What can fail without it
Pass a sequence limit through $F$	$F$ is continuous at the limiting point	The image sequence may approach a different value or no value
Preserve continuity of $f_n$	$f_n\to f$ uniformly	A pointwise limit can be discontinuous
Pass a limit through a finite-interval integral	Uniform convergence of Riemann-integrable functions	Concentrating spikes can retain nonzero area
Pass a limit through a derivative	One base value converges and the derivatives converge uniformly	Uniformly small functions can have nonconvergent derivatives
Swap two limits	A theorem providing joint or suitable uniform control	Both iterated limits may exist and disagree

These are sufficient conditions, not necessary ones. Failure of a sufficient test does not prove that a limit or interchange is impossible; it means that another theorem or a direct argument is needed.

Series and the M-test

An infinite series of functions is a sequence of partial sums

$$S_N(x)=\sum_{n=0}^{N}g_n(x).$$

All questions about continuity, integration, differentiation, or a second limit are therefore questions about the convergence mode of $(S_N)$.

The Weierstrass $M$-test gives uniform Cauchy control. If numbers $M_n\geq0$ satisfy

$$|g_n(x)|\leq M_n \quad\text{for every }x\in E, \qquad \sum_{n=0}^{\infty}M_n<\infty,$$

then $\sum_n g_n$ converges absolutely and uniformly. For $M>N$,

$$\sup_{x\in E}|S_M(x)-S_N(x)| \leq \sum_{n=N+1}^{M}M_n,$$

and the numerical tail tends to zero. Completeness of the scalar target then produces the uniform limit. Continuous summands have a continuous sum, and on a finite interval the sum may be integrated term by term. Differentiating term by term still requires control of the derivative series. The $M$-test is only sufficient: inability to find a summable majorant is not itself a proof of divergence. For the theorem and its limit-interchange consequences, see Colding 2025, Lectures 20–21, PDF.

A regulated free-field mode kernel

Consider a massive Euclidean free scalar kernel on the circle $S_L^1=\mathbb R/L\mathbb Z$, with $L>0$ and $m>0$. Declare the discrete Fourier convention

$$\widetilde f_n = \int_0^L\mathrm dx\, e^{+ik_nx}f(x), \qquad f(x) = \frac1L\sum_{n\in\mathbb Z} e^{-ik_nx}\widetilde f_n, \qquad k_n=\frac{2\pi n}{L}.$$

The finite-mode covariance is

$$G_N(x) = \frac1L \sum_{|n|\leq N} \frac{e^{-ik_nx}}{k_n^2+m^2}.$$

Here the regulator parameter is the integer cutoff $N$, the limit is $N\to\infty$ at fixed $L$ and $m$, and the claimed topology is the sup norm on $C(S_L^1;\mathbb C)$. For $M>N$,

$$\begin{aligned} \lVert G_M-G_N\rVert_\infty &\leq \frac2L\sum_{n=N+1}^{M} \frac1{(2\pi n/L)^2+m^2}. \end{aligned}$$

The numerical series is bounded by a constant multiple of $\sum_{n\geq1}n^{-2}$. The $M$-test therefore makes $(G_N)$ uniformly Cauchy. Since $C(S_L^1;\mathbb C)$ is complete, there is a continuous function $G$ such that

$$\lVert G_N-G\rVert_\infty\longrightarrow0.$$

Uniform convergence licenses, for example, passage through the circle integral. Only the zero mode survives, so

$$\int_0^L G(x)\,\mathrm dx = \lim_{N\to\infty}\int_0^L G_N(x)\,\mathrm dx = \frac1{m^2}.$$

Now test a different operation. Every $G_N$ is smooth, and termwise differentiation of the finite sum gives

$$\left(-\frac{\mathrm d^2}{\mathrm dx^2}+m^2\right)G_N(x) = \frac1L\sum_{|n|\leq N}e^{-ik_nx}.$$

At $x=0$, the right-hand side equals $(2N+1)/L$ and diverges. Thus uniform convergence of $G_N$ neither makes the displayed differential-operator sequence converge pointwise nor proves that $G$ is twice continuously differentiable. To identify its weaker limit, let $\varphi$ be a smooth periodic test function. Rapid decay of its Fourier coefficients gives

$$\begin{aligned} \int_0^L\mathrm dx\, \left(\frac1L\sum_{|n|\leq N}e^{-ik_nx}\right)\varphi(x) &= \frac1L\sum_{|n|\leq N}\widetilde\varphi_n\\ &\longrightarrow \frac1L\sum_{n\in\mathbb Z}\widetilde\varphi_n =\varphi(0). \end{aligned}$$

This is convergence to the periodic delta distribution. It is formulated by testing, not by assigning a finite value at every point.

This example establishes one precise result: removal of a finite-mode cutoff for this free scalar kernel in the sup norm. It does not establish pointwise convergence after differentiation, convergence of an operator-valued field, or the existence of an interacting continuum theory. Compare the free scalar mode analysis in Tong 2006–2007, §§ 2.2 and 2.3.2, PDF and the distributional status of free-field two-point functions in Zinn-Justin 2021, §§ 6.1.4 and 8.1. The test-function and distribution page defines the weaker limiting mode. Physical questions about regulator choice, observable convergence, error control, universality, and the order of continuum or infinite-volume limits continue in Regulators, Cutoffs, and Continuum Limits.

Common non-implications

Invalid shortcut	Counterexample or missing hypothesis
Bounded sequence $\Rightarrow$ Cauchy	$(-1)^n$ in $\mathbb R$
Cauchy $\Rightarrow$ convergent	Rational approximations to $\sqrt2$ in $\mathbb Q$
Pointwise $\Rightarrow$ uniform	$x^n$ on $[0,1]$
Pointwise limit of continuous functions is continuous	The same $x^n$ sequence
Pointwise convergence permits integral interchange	The shrinking tents $h_n$
Uniform convergence permits differentiation	$\sin(nx)/n$
Existence of both iterated limits makes them equal	$x^n$ as $n\to\infty$ and $x\to1^-$
Convergence at each fixed regulator proves regulator removal	A bound may deteriorate as the regulator is removed
One convergent regulated kernel establishes a continuum QFT	The object, topology, other cutoffs, and physical hypotheses remain separate

Where later convergence notions begin

• Almost-everywhere convergence, convergence in measure, monotone convergence, and dominated convergence require first developing measures and measurable functions, followed by Lebesgue integration and convergence theorems.
• Norm convergence in $L^p$, weak convergence, and the distinction between those modes belong to $L^p$ spaces, inequalities, and weak convergence.
• Completion of normed and inner-product spaces belongs to the Banach and Hilbert spaces page.
• Distributional convergence belongs to test-function spaces and distributions.
• Almost-sure convergence, convergence in probability, convergence in mean, and convergence in law belong to probabilistic convergence and limit theorems.

These are distinct convergence modes or limit theorems, governed by different hypotheses. Norm, weak, and distributional convergence arise from different topological structures in their own settings; they are not alternate names for pointwise or uniform convergence.

Exercises

Completeness check. Why does the estimate $|q_n-\sqrt2|<10^{-n}$ make $(q_n)$ Cauchy in $\mathbb Q$ without giving it a limit there?

Solution

For $m,n\geq N$,

$$|q_m-q_n| \leq |q_m-\sqrt2|+|\sqrt2-q_n| <2\cdot10^{-N}.$$

This proves the Cauchy property using the inherited metric. Its only possible real limit is $\sqrt2$, which is not an element of $\mathbb Q$; completeness is precisely the missing hypothesis.

Quantifier check. Negate uniform convergence of $f_n$ to $f$.

Solution

Uniform convergence fails exactly when there is an $\varepsilon_0>0$ such that for every $N$ there are $n\geq N$ and $x\in E$ with

$$d_Y(f_n(x),f(x))\geq\varepsilon_0.$$

The witness point may depend on $N$ and $n$. For $x^n$, points sufficiently close to $1$ provide such witnesses.

Integral check. Verify the pointwise limit and the integral of the tent functions $h_n$.

Solution

At $x=0$, every $h_n(0)=0$. For fixed $x>0$, eventually $2/n<x$, so $h_n(x)=0$. Hence $h_n\to0$ pointwise. Each graph is a triangle of base $2/n$ and height $n$, so its area is $\tfrac12(2/n)n=1$. Pointwise convergence alone supplies no theorem that would permit the interchange.

Derivative check. Which hypothesis in the differentiation theorem fails for $f_n(x)=\sin(nx)/n$?

Solution

The base-point values converge—for example $f_n(0)=0$—but the derivatives $f_n'(x)=\cos(nx)$ do not converge uniformly. They do not even converge pointwise at $x=\pi$, where $f_n'(\pi)=(-1)^n$. At $x=0$ every derivative equals $1$, while the derivative of the uniform limit is $0$.

Mode-sum check. Explain why the cutoff covariance converges uniformly but its image under $-\mathrm d^2/\mathrm dx^2+m^2$ does not converge pointwise at the origin.

Solution

The covariance coefficients decay like $n^{-2}$, so their absolute tails are bounded by a convergent numerical series independently of $x$. After applying the differential operator, each coefficient becomes $1$; at $x=0$, all $2N+1$ phases equal $1$. The value is therefore $(2N+1)/L$. The operation has destroyed the summable majorant that established uniform convergence.

References

• Tobias Holck Colding, 18.100B/18.1002 Real Analysis, Spring 2025, MIT OpenCourseWare PDF, Lectures 12 and 20–21. Metric convergence and completeness, the Weierstrass $M$-test, the uniform-limit theorem, completeness of $C([a,b])$, and uniform limit–integral interchange.
• Gerald B. Folland, Real Analysis: Modern Techniques and Their Applications, 2nd ed., Wiley (1999), publisher record, Chapter 0 and §§4.1–4.4. Metric-space and point-set-topology background for the chapter’s analysis framework.
• Jiří Lebl, Basic Analysis I, version 6.3 (2026), official PDF, §§ 6.1–6.2 and § 7.4. Pointwise and uniform convergence, uniform Cauchy criteria, continuity, integration and differentiation under limits, double-limit counterexamples, Cauchy sequences, and completeness.
• David Tong, Quantum Field Theory, Cambridge Part III lecture notes, University of Cambridge, 2006–2007, official PDF, §§2.2 and 2.3.2. Free scalar modes and the use and removal of ultraviolet regulators in a controlled free-field calculation.
• Jean Zinn-Justin, Quantum Field Theory and Critical Phenomena, 5th ed., Oxford University Press (2021), DOI:10.1093/oso/9780198834625.001.0001, §§6.1.4 and 8.1. Free-field two-point functions as distributions and the failure of ordinary coincident-point expressions in local field theory.