Uniform Convergence of Series

Uniform Convergence Is a Statement About Tails

The Weierstrass M-Test gives a powerful sufficient condition for uniform convergence: it bounds every term by a summable numerical sequence. But that condition is not necessary. To understand uniform convergence more generally, it helps to focus directly on the partial sums and on how small their tails become, uniformly across the domain.

Let $E$ be a nonempty set, and let $f_n:E\to\mathbb{R}$. The $N$th partial sum is $S_N(x)=\sum_{n=1}^{N}f_n(x)$. As with any sequence of functions, uniform convergence of the series means uniform convergence of its partial sums. The key practical reformulation is that the entire tail, beginning after a sufficiently large index, must be small for every $x\in E$.

Definition: The series $\sum_{n=1}^{\infty}f_n(x)$ converges uniformly on $E$ if its partial sums $S_N(x)=\sum_{n=1}^{N}f_n(x)$ converge uniformly on $E$ to a function $S:E\to\mathbb{R}$. In that case, $S(x)=\sum_{n=1}^{\infty}f_n(x)$ for every $x\in E$.

The pointwise series must converge at every $x$ for this definition to make sense. Uniform convergence adds a stronger requirement: one choice of how many terms to take must work for all inputs. The next criterion states that requirement in terms of finite portions of the tail.

Theorem (Uniform Cauchy Criterion for a Series): The series $\sum_{n=1}^{\infty}f_n(x)$ converges uniformly on $E$ if and only if, for every $\varepsilon>0$, there is an integer $N\geq0$ such that for all integers $q\geq p\geq N$ and all $x\in E$, $$ \left|\sum_{n=p+1}^{q}f_n(x)\right|<\varepsilon. $$

Proof. The partial sums $S_N$ satisfy $$ S_q(x)-S_p(x)=\sum_{n=p+1}^{q}f_n(x). $$ Thus the condition in the theorem says exactly that the sequence of partial sums is uniformly Cauchy. By the Uniform Cauchy Criterion for sequences established earlier in this course, a sequence of real-valued functions on $E$ converges uniformly if and only if it is uniformly Cauchy. Applying that criterion to $(S_N)$ proves both directions. $\square$

The criterion describes uniform control without first knowing the value of the sum $S$. It also explains why the M-Test works: its summable majorant bounds every finite tail by a numerical tail that tends to zero. But other forms of cancellation can make function-series tails uniformly small even when the absolute values do not have a summable majorant.

A Necessary Condition on the Terms

Uniform convergence controls tails, so it must also control individual terms. This gives a useful first check when uniform convergence is in question. The condition is necessary, but it is not sufficient: terms may become uniformly small while the partial sums still fail to converge uniformly.

Proposition (Terms of a Uniformly Convergent Series Vanish Uniformly): If $\sum_{n=1}^{\infty}f_n(x)$ converges uniformly on $E$, then $f_n\to0$ uniformly on $E$. Equivalently, $\sup_{x\in E}|f_n(x)|\to0$, with the supremum allowed to be infinite for individual early terms.

Proof. Let $S_N$ denote the partial sums and $S$ their uniform limit. Given $\varepsilon>0$, choose $N_0$ such that $$ |S_k(x)-S(x)|<\frac{\varepsilon}{2} \qquad\text{for every }k\geq N_0\text{ and }x\in E. $$ For $n\geq N_0+1$, we have $f_n=S_n-S_{n-1}$. Hence, for every $x\in E$, $$ |f_n(x)| \leq |S_n(x)-S(x)|+|S_{n-1}(x)-S(x)| <\frac{\varepsilon}{2}+\frac{\varepsilon}{2} =\varepsilon. $$ The same index $N_0+1$ works for every $x$, which is uniform convergence of $f_n$ to zero. $\square$

Worked Example: A Uniformly Convergent Geometric Series

On $E=[-2,2]$, consider $\sum_{n=1}^{\infty}x^{2n}/5^n$. Since $0\leq x^2/5\leq4/5$, $$ \left|\frac{x^{2n}}{5^n}\right| =\left(\frac{x^2}{5}\right)^n \leq\left(\frac45\right)^n \qquad(x\in[-2,2]). $$ The numerical series $\sum_{n=1}^{\infty}(4/5)^n$ converges, so the Weierstrass M-Test gives uniform convergence. More explicitly, after $N$ terms the remainder satisfies $$ \sup_{x\in[-2,2]}\left|\sum_{n=N+1}^{\infty}\frac{x^{2n}}{5^n}\right| \leq\sum_{n=N+1}^{\infty}\left(\frac45\right)^n =5\left(\frac45\right)^{N+1}. $$ This bound tends to zero independently of $x$, as required by the uniform Cauchy criterion. For each $x$, the geometric-series formula gives the sum $x^2/(5-x^2)$. The denominator is positive throughout the interval because $5-x^2\geq1$.

Uniform Convergence Without the M-Test

When terms alternate or otherwise cancel, estimating the absolute values separately can discard important information. Dirichlet’s test captures one common form of cancellation: the partial sums of one factor stay uniformly bounded, while the other factor decreases to zero uniformly.

Theorem (Uniform Dirichlet Test): Let $a_n,b_n:E\to\mathbb{R}$. Suppose that for each $x\in E$, the sequence $a_n(x)$ is nonnegative and nonincreasing in $n$, and suppose $\sup_{x\in E}a_n(x)\to0$. Suppose also that there is a finite $B\geq0$ such that $$ \left|\sum_{j=1}^{k}b_j(x)\right|\leq B \qquad\text{for every }k\geq1\text{ and }x\in E. $$ Then $\sum_{n=1}^{\infty}a_n(x)b_n(x)$ converges uniformly on $E$.

Proof. We verify the uniform Cauchy criterion for series. For fixed integers $q\geq p\geq1$ and $x\in E$, set $$ A_k(x)=\sum_{j=p}^{k}b_j(x)\qquad(p\leq k\leq q). $$ Each such sum is the difference of two initial partial sums (with the empty sum taken as zero), so $|A_k(x)|\leq2B$. Summation by parts gives $$ \sum_{n=p}^{q}a_n(x)b_n(x) =a_q(x)A_q(x)+\sum_{n=p}^{q-1}\bigl(a_n(x)-a_{n+1}(x)\bigr)A_n(x). $$ The coefficients $a_n(x)-a_{n+1}(x)$ are nonnegative. Taking absolute values and using $|A_k(x)|\leq2B$, we obtain $$ \begin{aligned} \left|\sum_{n=p}^{q}a_n(x)b_n(x)\right| &\leq2B\left(a_q(x)+\sum_{n=p}^{q-1}(a_n(x)-a_{n+1}(x))\right)\\ &=2B\,a_p(x) \leq2B\sup_{y\in E}a_p(y). \end{aligned} $$ The last bound tends to zero as $p\to\infty$, independently of $q\geq p$ and $x\in E$. If $B=0$, the same estimate says every finite tail is zero. In either case the Uniform Cauchy Criterion for a Series proves the claim. $\square$

Worked Example: An Alternating Power Series on a Closed Interval

Consider $\sum_{n=1}^{\infty}(-1)^{n-1}x^n/n$ on $[0,1]$. Set $a_n(x)=x^n/n$ and $b_n(x)=(-1)^{n-1}$. For every $x\in[0,1]$, $a_n(x)\geq0$, and $$ \frac{a_{n+1}(x)}{a_n(x)} =x\frac{n}{n+1}\leq1 $$ when $a_n(x)>0$; if $x=0$, all these terms are zero. Thus $a_n(x)$ is nonincreasing in $n$, and $$ \sup_{x\in[0,1]}a_n(x)=\frac1n\longrightarrow0. $$ The partial sums of $b_n$ alternate between $1$ and $0$, so their absolute values are at most $1$. The Uniform Dirichlet Test proves uniform convergence on $[0,1]$.

At $x=1$, the series is $\sum_{n=1}^{\infty}(-1)^{n-1}/n$, whose series of absolute values is the divergent harmonic series. Moreover, no summable sequence can serve as an M-Test majorant on $[0,1]$: any such majorant $M_n$ would have to satisfy $M_n\geq\sup_{x\in[0,1]}|(-1)^{n-1}x^n/n|=1/n$. Uniform convergence here comes from cancellation, not from a summable bound on the absolute values.

Worked Example: A Trigonometric Series Away from Its Boundary Points

Fix $0<\delta<\pi$ and let $E=[\delta,2\pi-\delta]$. Consider $\sum_{n=1}^{\infty}\sin(nx)/n$. For $x\in E$, the finite geometric-sum identity for $e^{ix}$ gives $$ \left|\sum_{j=1}^{k}e^{ijx}\right| =\left|\frac{e^{ix}(1-e^{ikx})}{1-e^{ix}}\right| \leq\frac{2}{|1-e^{ix}|} =\frac{1}{|\sin(x/2)|}. $$ Taking imaginary parts shows that $\left|\sum_{j=1}^{k}\sin(jx)\right|\leq1/|\sin(x/2)|$. On $E$, $\sin(x/2)\geq\sin(\delta/2)>0$. The partial sums of $b_n(x)=\sin(nx)$ are therefore bounded uniformly in both $k$ and $x$ by $1/\sin(\delta/2)$.

Now take $a_n(x)=1/n$. These factors are nonnegative, nonincreasing, and tend uniformly to zero. The Uniform Dirichlet Test proves that $\sum_{n=1}^{\infty}\sin(nx)/n$ converges uniformly on $E$. The interval stays away from $0$ and $2\pi$, where the bound on the trigonometric partial sums would cease to be uniform. This example shows why the domain matters when checking uniform convergence.

Small Terms Do Not Guarantee Uniform Convergence

The necessary condition that $f_n\to0$ uniformly is not a substitute for the uniform Cauchy criterion. The following telescoping series has terms that vanish uniformly, yet its partial sums do not converge uniformly on the whole domain.

Worked Example: Uniformly Vanishing Terms with Nonuniform Convergence

On $[0,1]$, set $f_1(x)=x$, and for $n\geq2$ set $f_n(x)=x^n-x^{n-1}$. The partial sums telescope: $$ S_N(x)=x+\sum_{n=2}^{N}(x^n-x^{n-1})=x^N. $$ For $0\leq x<1$, $S_N(x)\to0$, while $S_N(1)=1$ for every $N$. Thus the pointwise sum is $S(x)=0$ for $x<1$ and $S(1)=1$.

For every $N$, the supremum of $|S_N(x)-S(x)|$ on $[0,1]$ is $1$: for $x<1$, the error is $x^N$, which can be made arbitrarily close to $1$ by taking $x$ sufficiently close to $1$, while the error at $1$ is zero. Hence the partial sums do not converge uniformly. Nevertheless, for $n\geq2$, $$ \sup_{x\in[0,1]}|f_n(x)| =\sup_{x\in[0,1]}x^{n-1}(1-x) =\frac{(n-1)^{n-1}}{n^n} \leq\frac1n\longrightarrow0. $$ The maximum follows by differentiating $x^{n-1}(1-x)$, whose interior critical point is $x=(n-1)/n$; the endpoint values are zero. Thus the terms do vanish uniformly, but that fact alone does not make the series uniformly convergent.

How to Choose a Convergence Test

The uniform Cauchy criterion is the general test: it asks whether every sufficiently late finite tail is uniformly small. The M-Test answers this by bounding absolute values with a summable numerical sequence. The Uniform Dirichlet Test gives another route when cancellation is present. Neither convenient test is necessary for uniform convergence; failure to satisfy its hypotheses is not proof that a series fails to converge uniformly.

The condition $f_n\to0$ uniformly is a useful quick check. If it fails, the series cannot converge uniformly. If it holds, more work is needed: the telescoping example shows that uniform convergence can still fail. In practice, estimate the finite tails directly when possible, or identify structure—such as monotonicity and bounded partial sums—that makes a uniform tail estimate available.

Takeaway: Uniform convergence of a series is uniform convergence of its partial sums. Equivalently, every sufficiently late finite tail must be small uniformly over the domain. The terms must vanish uniformly, but that necessary condition is not sufficient; cancellation can also establish uniform convergence when the M-Test does not apply.

Check Your Understanding

Use the definitions, criteria, and examples in this tutorial to answer the following questions.

How does the Uniform Cauchy Criterion for a Series express control of the tails?
Why must the terms of a uniformly convergent series tend uniformly to zero?
Which two hypotheses in the Uniform Dirichlet Test control cancellation and decay, respectively?
Why does the alternating power series on $[0,1]$ converge uniformly even though its absolute series at $x=1$ diverges?
In the telescoping example, why does uniform convergence fail even though the terms vanish uniformly?

What You'll Learn