Sequential Characterization of Continuity

The sequential characterization is the standard tool for proving that compositions and algebraic combinations of continuous functions remain continuous (foreshadowing Composition of Continuous Functions): one simply chases a sequence through each function in turn, using convergence of sequences at each stage instead of juggling nested $\varepsilon$'s and $\delta$'s. It underlies the proof that continuous images of sequentially compact sets are sequentially compact, which feeds directly into The Extreme Value Theorem — a continuous function on a closed bounded interval attains a maximum, proved by extracting convergent subsequences of a maximizing sequence and using sequential continuity to pass the limit through $f$. Beyond core analysis, the idea generalizes verbatim to metric spaces and (with care) to general topological spaces, where it becomes the basis for defining continuity via nets or filters when sequences alone are insufficient. In numerical analysis and scientific computing, the sequential viewpoint justifies why iterative approximation schemes (finite-difference or fixed-point methods) that converge to an input value also produce function values converging to the true output, provided the function being approximated is continuous. In probability theory, sequential continuity arguments justify the continuous mapping theorem, used to transfer convergence in distribution of random variables through continuous transformations. In machine learning, sequential characterizations of continuity underpin arguments about the stability of loss functions and activation functions under small perturbations of input data along training trajectories.

Ask the tutor: Can you show me how this same sequential trick proves that the composition of two continuous functions is continuous?

Think of $c$ as a target on the number line and $f(c)$ as the corresponding target in the range. The $\varepsilon$-$\delta$ definition says: shrink a window of radius $\varepsilon$ around $f(c)$, and you can always find a matching window of radius $\delta$ around $c$ whose entire image lands inside that $\varepsilon$-window. The sequential definition instead sends in scouts — sequences of points marching toward $c$ — and asks whether their images always march toward $f(c)$. The intuition for why these agree is that a sequence converging to $c$ is exactly a mechanism for entering every $\delta$-neighborhood of $c$ eventually, no matter how small $\delta$ is. So if $f$ has the $\delta$-neighborhood-mapping property, any sequence that eventually sits inside that neighborhood must have its image eventually sit inside the corresponding $\varepsilon$-neighborhood of $f(c)$. Conversely, if $f$ fails the $\varepsilon$-$\delta$ property, some $\varepsilon_0$ resists every $\delta$, and taking $\delta = 1/n$ produces witness points $x_n$ that converge to $c$ but whose images stubbornly stay $\varepsilon_0$ away from $f(c)$ — a concrete failing sequence.

Example 1: Confirming continuity of a piecewise-defined function using a specific sequence family. Let $D = \mathbb{R}$ and define

We claim $f$ is continuous at $c = 0$, and we verify this using the sequential characterization with an asymmetric, non-obvious test sequence rather than the trivial choice $x_n = 1/n$. Take $x_n = \dfrac{1}{(n+3)\pi + \pi/6}$ for $n = 1, 2, 3, \dots$ — chosen so that $\sin(1/x_n) = \sin\big((n+3)\pi + \pi/6\big) = \pm \sin(\pi/6) = \pm \tfrac12$, a deliberately non-simplifying, oscillating sequence of sine-values, rather than the sequence where $\sin(1/x_n)$ is constantly $0$ or $1$. First, $x_n \to 0$: as $n \to \infty$, the denominator $(n+3)\pi + \pi/6 \to \infty$, so $x_n \to 0$ by the standard fact that $1/n \to 0$ (formally, $0 < x_n < \frac{1}{n\pi}$ for all $n\ge 1$, and $\frac{1}{n\pi}\to 0$, so the Squeeze Theorem for sequences forces $x_n \to 0$). Now compute $f(x_n) = x_n^2 \sin(1/x_n)$. Since $|\sin(1/x_n)| \le 1$ always (regardless of the oscillating value $\pm\tfrac12$), we get

Since $x_n \to 0$, also $x_n^2 \to 0$ (product of convergent sequences), so by the Squeeze Theorem, $f(x_n) \to 0 = f(0)$. This single sequence does not prove continuity at $0$ (that requires checking all sequences, which the direct $\varepsilon$–$\delta$ argument does: given $\varepsilon>0$, take $\delta=\sqrt\varepsilon$, so $|x|<\delta \Rightarrow |f(x)| \le x^2 < \varepsilon$). But the exercise shows how the theorem's forward direction is used in practice: no matter which sequence approaching $0$ we pick — even an oscillating one built from an unusual arithmetic progression of angles — the images are forced to converge to $f(0) = 0$, exactly as the theorem guarantees.

Example 2: Detecting discontinuity by exhibiting one bad sequence. Let $D = \mathbb{R}$ and define the function

We use the sequential characterization to show $g$ is not continuous at $c = 2$. For $x \neq 2$, factor: $x^3 - 8 = (x-2)(x^2 + 2x + 4)$, so $g(x) = x^2 + 2x + 4$ for $x \ne 2$. Choose the sequence $x_n = 2 + \dfrac{1}{n^2+1}$ for $n = 1, 2, 3, \dots$ (an asymmetric, non-standard choice of step size, distinct from the generic $2 + 1/n$). Since $\frac{1}{n^2+1} \to 0$, we have $x_n \to 2$, and $x_n \neq 2$ for every $n$, so $x_n \in D \setminus \{2\}$. Compute the image sequence using $g(x_n) = x_n^2 + 2x_n + 4$ for $n\ge 1$. As $x_n \to 2$, and polynomials are continuous (a fact itself provable via this same theorem, since polynomials are built from sums and products of the identity and constant sequences), we get

But $g(2) = 5$ by definition. Since $12 \neq 5$, we have found a sequence $x_n \to 2$ with $g(x_n) \to 12 \ne g(2) = 5$. By the contrapositive of the forward direction of the theorem (equivalently, by the converse direction applied in reverse: sequential preservation fails, so continuity fails), $g$ is not continuous at $c = 2$. Notice we needed only one cleverly chosen sequence to kill continuity — this asymmetry (one sequence suffices to disprove, but all sequences are needed to prove) is the central practical payoff of the theorem.