Reading an Epsilon-Delta Definition

Read the Choices in Their Order

The epsilon-delta definition is not just a pair of inequalities. Its quantifiers specify who gets to choose each quantity, and when. The previous tutorial introduced the definition and used estimates to produce suitable values of \(\delta\). Here the focus is on reading the definition correctly: identifying what a continuity claim promises, what it does not promise, and how to recognize a failure.

Let \(E\subseteq\mathbb{R}\), let \(f:E\to\mathbb{R}\), and fix \(a\in E\). Continuity at \(a\) says that every positive output tolerance can be met by some positive input tolerance. Once \(\delta\) has been chosen, it must work for every \(x\) in the domain that lies within that input tolerance of \(a\).

Read the quantifiers from left to right. First, someone specifies \(\varepsilon>0\). In response, we choose \(\delta>0\), possibly using \(\varepsilon\), the function, and the point \(a\). Then the implication must hold for every \(x\in E\). The choice of \(\delta\) cannot depend on which particular \(x\) is later tested.

This wording highlights three common misreadings. First, continuity does not require one fixed \(\delta\) to work for every \(\varepsilon\). Second, it does not permit a new \(\delta\) for each \(x\). Third, “every \(x\)” means every point of \(E\), not every real number unless \(E=\mathbb{R}\). The point \(x=a\) is included, but causes no difficulty: its output difference is \(0<\varepsilon\).

Why the Order Cannot Be Reversed

The statement “for every \(\varepsilon\), there exists a \(\delta\)” is different from “there exists a \(\delta\) that works for every \(\varepsilon\).” The latter demands a single positive input tolerance that makes the output difference smaller than every positive number. That is usually too strong.

The universal quantifier over \(x\) matters just as much. A statement about one nearby input at a time might allow the input tolerance to shrink as that input changes. Continuity instead requires a single neighborhood that controls all domain points in it at once. In a proof, a choice made after \(x\) is known has the wrong order of dependence.

Reading the Negation: How Continuity Fails

To show that a function is not continuous at \(a\), negate the quantifiers and the implication carefully. The negation of “for every \(\varepsilon>0\), there exists \(\delta>0\)” begins “there exists \(\varepsilon_0>0\), such that for every \(\delta>0.” The negated implication requires a point that is close enough in the input but whose output difference is not smaller than \(\varepsilon_0\).

Proof. Continuity at \(a\) has the form \[ \forall\varepsilon>0\ \exists\delta>0\ \forall x\in E,\quad \bigl(|x-a|<\delta\Longrightarrow |f(x)-f(a)|<\varepsilon\bigr). \] Its negation is \[ \exists\varepsilon_0>0\ \forall\delta>0\ \exists x\in E,\quad \neg\bigl(|x-a|<\delta\Longrightarrow |f(x)-f(a)|<\varepsilon_0\bigr). \] For any propositions \(P,Q\), the negation of \(P\Longrightarrow Q\) is \(P\) and \(\neg Q\). Here those two conditions are \(|x-a|<\delta\) and \(|f(x)-f(a)|\geq\varepsilon_0\). This is exactly the stated criterion. Each logical step is reversible, so the criterion also implies that continuity fails. \(\square\)

The output tolerance \(\varepsilon_0\) in this criterion is fixed once and for all. No matter how small a positive input radius \(\delta\) is proposed, some domain point inside that radius still has an output discrepancy at least \(\varepsilon_0\). It is not enough for the discrepancy to remain nonzero; it must stay at least as large as one fixed positive number.

This method is often more efficient than trying to disprove continuity by guessing that a particular delta will fail. The criterion requires a witness for every proposed delta, while keeping the same \(\varepsilon_0\) throughout. A witness may depend on \(\delta\); the output tolerance may not.

A Countable Test for the Output Tolerances

Continuity asks for a valid delta for every positive real \(\varepsilon\), an uncountable collection of tolerances. For this definition, it is enough to check the particular tolerances \(1,1/2,1/3,\ldots\). This gives a useful way to organize a continuity argument as a sequence of increasingly stringent tests.

Proof. Suppose first that \(f\) is continuous at \(a\). For each positive integer \(n\), apply the definition with \(\varepsilon=1/n\). It supplies a \(\delta_n>0\) for which the displayed implication holds.

Conversely, suppose the stated condition holds for every positive integer \(n\), and let an arbitrary \(\varepsilon>0\) be given. By the Archimedean property of the real numbers, choose a positive integer \(n\) such that \(1/n<\varepsilon\). Use the condition for this \(n\) to choose \(\delta_n>0\). Whenever \(x\in E\) and \(|x-a|<\delta_n\), we then have \[ |f(x)-f(a)|<\frac1n<\varepsilon. \] Thus the epsilon-delta condition holds for the arbitrary \(\varepsilon>0\), so \(f\) is continuous at \(a\). \(\square\)

The \(\delta_n\) in this theorem may change with \(n\). The result does not claim that one delta works for all the tolerances \(1/n\). It says that passing every one of these tests is enough to handle any positive tolerance, because some \(1/n\) is smaller than that tolerance.

What to Keep in View

When reading or writing a continuity argument, it helps to identify the role of each quantity before manipulating inequalities. The given \(\varepsilon\) is the output tolerance. The chosen \(\delta\) is an input tolerance that may depend on it. The test point \(x\) is arbitrary within the domain and neighborhood. A valid argument must then establish the output bound for every such \(x\).

The failure criterion reverses this structure: one fixed positive output tolerance defeats every proposed input radius. The countable test offers a different perspective: to establish continuity, it suffices to meet a sequence of tolerances tending to zero. These are not competing definitions; both follow directly from the quantifier structure and can make different proofs easier to read.

If \(a\) is isolated in \(E\), the definition also remains exactly the same. In that case, there may be a \(\delta>0\) for which the only point of \(E\) satisfying \(|x-a|<\delta\) is \(a\). Since the output difference at \(a\) is zero, every positive \(\varepsilon\) is then satisfied. The domain condition is essential to interpreting this case correctly.